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 /// getPreferredLSDADataFormat - Return the preferred exception handling data
1059 /// format for the LSDA.
1060 unsigned X86TargetLowering::getPreferredLSDADataFormat() const {
1061 if (Subtarget->isTargetDarwin())
1062 return dwarf::DW_EH_PE_pcrel;
1064 CodeModel::Model M = getTargetMachine().getCodeModel();
1066 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1067 if (!Subtarget->is64Bit() || M == CodeModel::Small)
1068 return dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4;
1070 return dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata8;
1073 if (M == CodeModel::Small)
1074 return dwarf::DW_EH_PE_sdata4;
1076 return dwarf::DW_EH_PE_absptr;
1079 /// getPreferredFDEDataFormat - Return the preferred exception handling data
1080 /// format for the FDE.
1081 unsigned X86TargetLowering::getPreferredFDEDataFormat() const {
1082 if (Subtarget->isTargetDarwin())
1083 return dwarf::DW_EH_PE_pcrel;
1085 CodeModel::Model M = getTargetMachine().getCodeModel();
1087 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1088 if (!Subtarget->is64Bit() ||
1089 M == CodeModel::Small || M == CodeModel::Medium)
1090 return dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4;
1092 return dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata8;
1095 if (M == CodeModel::Small || M == CodeModel::Medium)
1096 return dwarf::DW_EH_PE_sdata4;
1098 return dwarf::DW_EH_PE_absptr;
1101 //===----------------------------------------------------------------------===//
1102 // Return Value Calling Convention Implementation
1103 //===----------------------------------------------------------------------===//
1105 #include "X86GenCallingConv.inc"
1108 X86TargetLowering::LowerReturn(SDValue Chain,
1109 unsigned CallConv, bool isVarArg,
1110 const SmallVectorImpl<ISD::OutputArg> &Outs,
1111 DebugLoc dl, SelectionDAG &DAG) {
1113 SmallVector<CCValAssign, 16> RVLocs;
1114 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1115 RVLocs, *DAG.getContext());
1116 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1118 // If this is the first return lowered for this function, add the regs to the
1119 // liveout set for the function.
1120 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1121 for (unsigned i = 0; i != RVLocs.size(); ++i)
1122 if (RVLocs[i].isRegLoc())
1123 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1128 SmallVector<SDValue, 6> RetOps;
1129 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1130 // Operand #1 = Bytes To Pop
1131 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1133 // Copy the result values into the output registers.
1134 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1135 CCValAssign &VA = RVLocs[i];
1136 assert(VA.isRegLoc() && "Can only return in registers!");
1137 SDValue ValToCopy = Outs[i].Val;
1139 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1140 // the RET instruction and handled by the FP Stackifier.
1141 if (VA.getLocReg() == X86::ST0 ||
1142 VA.getLocReg() == X86::ST1) {
1143 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1144 // change the value to the FP stack register class.
1145 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1146 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1147 RetOps.push_back(ValToCopy);
1148 // Don't emit a copytoreg.
1152 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1153 // which is returned in RAX / RDX.
1154 if (Subtarget->is64Bit()) {
1155 EVT ValVT = ValToCopy.getValueType();
1156 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1157 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1158 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1159 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1163 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1164 Flag = Chain.getValue(1);
1167 // The x86-64 ABI for returning structs by value requires that we copy
1168 // the sret argument into %rax for the return. We saved the argument into
1169 // a virtual register in the entry block, so now we copy the value out
1171 if (Subtarget->is64Bit() &&
1172 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1173 MachineFunction &MF = DAG.getMachineFunction();
1174 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1175 unsigned Reg = FuncInfo->getSRetReturnReg();
1177 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1178 FuncInfo->setSRetReturnReg(Reg);
1180 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1182 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1183 Flag = Chain.getValue(1);
1186 RetOps[0] = Chain; // Update chain.
1188 // Add the flag if we have it.
1190 RetOps.push_back(Flag);
1192 return DAG.getNode(X86ISD::RET_FLAG, dl,
1193 MVT::Other, &RetOps[0], RetOps.size());
1196 /// LowerCallResult - Lower the result values of a call into the
1197 /// appropriate copies out of appropriate physical registers.
1200 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1201 unsigned CallConv, bool isVarArg,
1202 const SmallVectorImpl<ISD::InputArg> &Ins,
1203 DebugLoc dl, SelectionDAG &DAG,
1204 SmallVectorImpl<SDValue> &InVals) {
1206 // Assign locations to each value returned by this call.
1207 SmallVector<CCValAssign, 16> RVLocs;
1208 bool Is64Bit = Subtarget->is64Bit();
1209 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1210 RVLocs, *DAG.getContext());
1211 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1213 // Copy all of the result registers out of their specified physreg.
1214 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1215 CCValAssign &VA = RVLocs[i];
1216 EVT CopyVT = VA.getValVT();
1218 // If this is x86-64, and we disabled SSE, we can't return FP values
1219 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1220 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1221 llvm_report_error("SSE register return with SSE disabled");
1224 // If this is a call to a function that returns an fp value on the floating
1225 // point stack, but where we prefer to use the value in xmm registers, copy
1226 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1227 if ((VA.getLocReg() == X86::ST0 ||
1228 VA.getLocReg() == X86::ST1) &&
1229 isScalarFPTypeInSSEReg(VA.getValVT())) {
1234 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1235 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1236 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1237 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1238 MVT::v2i64, InFlag).getValue(1);
1239 Val = Chain.getValue(0);
1240 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1241 Val, DAG.getConstant(0, MVT::i64));
1243 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1244 MVT::i64, InFlag).getValue(1);
1245 Val = Chain.getValue(0);
1247 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1249 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1250 CopyVT, InFlag).getValue(1);
1251 Val = Chain.getValue(0);
1253 InFlag = Chain.getValue(2);
1255 if (CopyVT != VA.getValVT()) {
1256 // Round the F80 the right size, which also moves to the appropriate xmm
1258 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1259 // This truncation won't change the value.
1260 DAG.getIntPtrConstant(1));
1263 InVals.push_back(Val);
1270 //===----------------------------------------------------------------------===//
1271 // C & StdCall & Fast Calling Convention implementation
1272 //===----------------------------------------------------------------------===//
1273 // StdCall calling convention seems to be standard for many Windows' API
1274 // routines and around. It differs from C calling convention just a little:
1275 // callee should clean up the stack, not caller. Symbols should be also
1276 // decorated in some fancy way :) It doesn't support any vector arguments.
1277 // For info on fast calling convention see Fast Calling Convention (tail call)
1278 // implementation LowerX86_32FastCCCallTo.
1280 /// CallIsStructReturn - Determines whether a call uses struct return
1282 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1286 return Outs[0].Flags.isSRet();
1289 /// ArgsAreStructReturn - Determines whether a function uses struct
1290 /// return semantics.
1292 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1296 return Ins[0].Flags.isSRet();
1299 /// IsCalleePop - Determines whether the callee is required to pop its
1300 /// own arguments. Callee pop is necessary to support tail calls.
1301 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1305 switch (CallingConv) {
1308 case CallingConv::X86_StdCall:
1309 return !Subtarget->is64Bit();
1310 case CallingConv::X86_FastCall:
1311 return !Subtarget->is64Bit();
1312 case CallingConv::Fast:
1313 return PerformTailCallOpt;
1317 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1318 /// given CallingConvention value.
1319 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1320 if (Subtarget->is64Bit()) {
1321 if (Subtarget->isTargetWin64())
1322 return CC_X86_Win64_C;
1327 if (CC == CallingConv::X86_FastCall)
1328 return CC_X86_32_FastCall;
1329 else if (CC == CallingConv::Fast)
1330 return CC_X86_32_FastCC;
1335 /// NameDecorationForCallConv - Selects the appropriate decoration to
1336 /// apply to a MachineFunction containing a given calling convention.
1338 X86TargetLowering::NameDecorationForCallConv(unsigned CallConv) {
1339 if (CallConv == CallingConv::X86_FastCall)
1341 else if (CallConv == CallingConv::X86_StdCall)
1347 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1348 /// by "Src" to address "Dst" with size and alignment information specified by
1349 /// the specific parameter attribute. The copy will be passed as a byval
1350 /// function parameter.
1352 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1353 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1355 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1356 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1357 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1361 X86TargetLowering::LowerMemArgument(SDValue Chain,
1363 const SmallVectorImpl<ISD::InputArg> &Ins,
1364 DebugLoc dl, SelectionDAG &DAG,
1365 const CCValAssign &VA,
1366 MachineFrameInfo *MFI,
1369 // Create the nodes corresponding to a load from this parameter slot.
1370 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1371 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1372 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1375 // If value is passed by pointer we have address passed instead of the value
1377 if (VA.getLocInfo() == CCValAssign::Indirect)
1378 ValVT = VA.getLocVT();
1380 ValVT = VA.getValVT();
1382 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1383 // changed with more analysis.
1384 // In case of tail call optimization mark all arguments mutable. Since they
1385 // could be overwritten by lowering of arguments in case of a tail call.
1386 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1387 VA.getLocMemOffset(), isImmutable);
1388 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1389 if (Flags.isByVal())
1391 return DAG.getLoad(ValVT, dl, Chain, FIN,
1392 PseudoSourceValue::getFixedStack(FI), 0);
1396 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1399 const SmallVectorImpl<ISD::InputArg> &Ins,
1402 SmallVectorImpl<SDValue> &InVals) {
1404 MachineFunction &MF = DAG.getMachineFunction();
1405 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1407 const Function* Fn = MF.getFunction();
1408 if (Fn->hasExternalLinkage() &&
1409 Subtarget->isTargetCygMing() &&
1410 Fn->getName() == "main")
1411 FuncInfo->setForceFramePointer(true);
1413 // Decorate the function name.
1414 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1416 MachineFrameInfo *MFI = MF.getFrameInfo();
1417 bool Is64Bit = Subtarget->is64Bit();
1418 bool IsWin64 = Subtarget->isTargetWin64();
1420 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1421 "Var args not supported with calling convention fastcc");
1423 // Assign locations to all of the incoming arguments.
1424 SmallVector<CCValAssign, 16> ArgLocs;
1425 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1426 ArgLocs, *DAG.getContext());
1427 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1429 unsigned LastVal = ~0U;
1431 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1432 CCValAssign &VA = ArgLocs[i];
1433 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1435 assert(VA.getValNo() != LastVal &&
1436 "Don't support value assigned to multiple locs yet");
1437 LastVal = VA.getValNo();
1439 if (VA.isRegLoc()) {
1440 EVT RegVT = VA.getLocVT();
1441 TargetRegisterClass *RC = NULL;
1442 if (RegVT == MVT::i32)
1443 RC = X86::GR32RegisterClass;
1444 else if (Is64Bit && RegVT == MVT::i64)
1445 RC = X86::GR64RegisterClass;
1446 else if (RegVT == MVT::f32)
1447 RC = X86::FR32RegisterClass;
1448 else if (RegVT == MVT::f64)
1449 RC = X86::FR64RegisterClass;
1450 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1451 RC = X86::VR128RegisterClass;
1452 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1453 RC = X86::VR64RegisterClass;
1455 llvm_unreachable("Unknown argument type!");
1457 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1458 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1460 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1461 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1463 if (VA.getLocInfo() == CCValAssign::SExt)
1464 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1465 DAG.getValueType(VA.getValVT()));
1466 else if (VA.getLocInfo() == CCValAssign::ZExt)
1467 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1468 DAG.getValueType(VA.getValVT()));
1469 else if (VA.getLocInfo() == CCValAssign::BCvt)
1470 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1472 if (VA.isExtInLoc()) {
1473 // Handle MMX values passed in XMM regs.
1474 if (RegVT.isVector()) {
1475 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1476 ArgValue, DAG.getConstant(0, MVT::i64));
1477 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1479 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1482 assert(VA.isMemLoc());
1483 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1486 // If value is passed via pointer - do a load.
1487 if (VA.getLocInfo() == CCValAssign::Indirect)
1488 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1490 InVals.push_back(ArgValue);
1493 // The x86-64 ABI for returning structs by value requires that we copy
1494 // the sret argument into %rax for the return. Save the argument into
1495 // a virtual register so that we can access it from the return points.
1496 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1497 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1498 unsigned Reg = FuncInfo->getSRetReturnReg();
1500 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1501 FuncInfo->setSRetReturnReg(Reg);
1503 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1504 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1507 unsigned StackSize = CCInfo.getNextStackOffset();
1508 // align stack specially for tail calls
1509 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1510 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1512 // If the function takes variable number of arguments, make a frame index for
1513 // the start of the first vararg value... for expansion of llvm.va_start.
1515 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1516 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1519 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1521 // FIXME: We should really autogenerate these arrays
1522 static const unsigned GPR64ArgRegsWin64[] = {
1523 X86::RCX, X86::RDX, X86::R8, X86::R9
1525 static const unsigned XMMArgRegsWin64[] = {
1526 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1528 static const unsigned GPR64ArgRegs64Bit[] = {
1529 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1531 static const unsigned XMMArgRegs64Bit[] = {
1532 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1533 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1535 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1538 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1539 GPR64ArgRegs = GPR64ArgRegsWin64;
1540 XMMArgRegs = XMMArgRegsWin64;
1542 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1543 GPR64ArgRegs = GPR64ArgRegs64Bit;
1544 XMMArgRegs = XMMArgRegs64Bit;
1546 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1548 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1551 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1552 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1553 "SSE register cannot be used when SSE is disabled!");
1554 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1555 "SSE register cannot be used when SSE is disabled!");
1556 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1557 // Kernel mode asks for SSE to be disabled, so don't push them
1559 TotalNumXMMRegs = 0;
1561 // For X86-64, if there are vararg parameters that are passed via
1562 // registers, then we must store them to their spots on the stack so they
1563 // may be loaded by deferencing the result of va_next.
1564 VarArgsGPOffset = NumIntRegs * 8;
1565 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1566 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1567 TotalNumXMMRegs * 16, 16);
1569 // Store the integer parameter registers.
1570 SmallVector<SDValue, 8> MemOps;
1571 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1572 unsigned Offset = VarArgsGPOffset;
1573 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1574 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1575 DAG.getIntPtrConstant(Offset));
1576 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1577 X86::GR64RegisterClass);
1578 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1580 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1581 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1583 MemOps.push_back(Store);
1587 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1588 // Now store the XMM (fp + vector) parameter registers.
1589 SmallVector<SDValue, 11> SaveXMMOps;
1590 SaveXMMOps.push_back(Chain);
1592 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1593 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1594 SaveXMMOps.push_back(ALVal);
1596 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1597 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1599 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1600 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1601 X86::VR128RegisterClass);
1602 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1603 SaveXMMOps.push_back(Val);
1605 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1607 &SaveXMMOps[0], SaveXMMOps.size()));
1610 if (!MemOps.empty())
1611 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1612 &MemOps[0], MemOps.size());
1616 // Some CCs need callee pop.
1617 if (IsCalleePop(isVarArg, CallConv)) {
1618 BytesToPopOnReturn = StackSize; // Callee pops everything.
1619 BytesCallerReserves = 0;
1621 BytesToPopOnReturn = 0; // Callee pops nothing.
1622 // If this is an sret function, the return should pop the hidden pointer.
1623 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1624 BytesToPopOnReturn = 4;
1625 BytesCallerReserves = StackSize;
1629 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1630 if (CallConv == CallingConv::X86_FastCall)
1631 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1634 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1640 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1641 SDValue StackPtr, SDValue Arg,
1642 DebugLoc dl, SelectionDAG &DAG,
1643 const CCValAssign &VA,
1644 ISD::ArgFlagsTy Flags) {
1645 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1646 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1647 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1648 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1649 if (Flags.isByVal()) {
1650 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1652 return DAG.getStore(Chain, dl, Arg, PtrOff,
1653 PseudoSourceValue::getStack(), LocMemOffset);
1656 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1657 /// optimization is performed and it is required.
1659 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1660 SDValue &OutRetAddr,
1666 if (!IsTailCall || FPDiff==0) return Chain;
1668 // Adjust the Return address stack slot.
1669 EVT VT = getPointerTy();
1670 OutRetAddr = getReturnAddressFrameIndex(DAG);
1672 // Load the "old" Return address.
1673 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1674 return SDValue(OutRetAddr.getNode(), 1);
1677 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1678 /// optimization is performed and it is required (FPDiff!=0).
1680 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1681 SDValue Chain, SDValue RetAddrFrIdx,
1682 bool Is64Bit, int FPDiff, DebugLoc dl) {
1683 // Store the return address to the appropriate stack slot.
1684 if (!FPDiff) return Chain;
1685 // Calculate the new stack slot for the return address.
1686 int SlotSize = Is64Bit ? 8 : 4;
1687 int NewReturnAddrFI =
1688 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1689 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1690 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1691 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1692 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1697 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1698 unsigned CallConv, bool isVarArg, bool isTailCall,
1699 const SmallVectorImpl<ISD::OutputArg> &Outs,
1700 const SmallVectorImpl<ISD::InputArg> &Ins,
1701 DebugLoc dl, SelectionDAG &DAG,
1702 SmallVectorImpl<SDValue> &InVals) {
1704 MachineFunction &MF = DAG.getMachineFunction();
1705 bool Is64Bit = Subtarget->is64Bit();
1706 bool IsStructRet = CallIsStructReturn(Outs);
1708 assert((!isTailCall ||
1709 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1710 "IsEligibleForTailCallOptimization missed a case!");
1711 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1712 "Var args not supported with calling convention fastcc");
1714 // Analyze operands of the call, assigning locations to each operand.
1715 SmallVector<CCValAssign, 16> ArgLocs;
1716 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1717 ArgLocs, *DAG.getContext());
1718 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1720 // Get a count of how many bytes are to be pushed on the stack.
1721 unsigned NumBytes = CCInfo.getNextStackOffset();
1722 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1723 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1727 // Lower arguments at fp - stackoffset + fpdiff.
1728 unsigned NumBytesCallerPushed =
1729 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1730 FPDiff = NumBytesCallerPushed - NumBytes;
1732 // Set the delta of movement of the returnaddr stackslot.
1733 // But only set if delta is greater than previous delta.
1734 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1735 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1738 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1740 SDValue RetAddrFrIdx;
1741 // Load return adress for tail calls.
1742 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1745 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1746 SmallVector<SDValue, 8> MemOpChains;
1749 // Walk the register/memloc assignments, inserting copies/loads. In the case
1750 // of tail call optimization arguments are handle later.
1751 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1752 CCValAssign &VA = ArgLocs[i];
1753 EVT RegVT = VA.getLocVT();
1754 SDValue Arg = Outs[i].Val;
1755 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1756 bool isByVal = Flags.isByVal();
1758 // Promote the value if needed.
1759 switch (VA.getLocInfo()) {
1760 default: llvm_unreachable("Unknown loc info!");
1761 case CCValAssign::Full: break;
1762 case CCValAssign::SExt:
1763 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1765 case CCValAssign::ZExt:
1766 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1768 case CCValAssign::AExt:
1769 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1770 // Special case: passing MMX values in XMM registers.
1771 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1772 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1773 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1775 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1777 case CCValAssign::BCvt:
1778 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1780 case CCValAssign::Indirect: {
1781 // Store the argument.
1782 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1783 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1784 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1785 PseudoSourceValue::getFixedStack(FI), 0);
1791 if (VA.isRegLoc()) {
1792 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1794 if (!isTailCall || (isTailCall && isByVal)) {
1795 assert(VA.isMemLoc());
1796 if (StackPtr.getNode() == 0)
1797 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1799 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1800 dl, DAG, VA, Flags));
1805 if (!MemOpChains.empty())
1806 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1807 &MemOpChains[0], MemOpChains.size());
1809 // Build a sequence of copy-to-reg nodes chained together with token chain
1810 // and flag operands which copy the outgoing args into registers.
1812 // Tail call byval lowering might overwrite argument registers so in case of
1813 // tail call optimization the copies to registers are lowered later.
1815 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1816 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1817 RegsToPass[i].second, InFlag);
1818 InFlag = Chain.getValue(1);
1822 if (Subtarget->isPICStyleGOT()) {
1823 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1826 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1827 DAG.getNode(X86ISD::GlobalBaseReg,
1828 DebugLoc::getUnknownLoc(),
1831 InFlag = Chain.getValue(1);
1833 // If we are tail calling and generating PIC/GOT style code load the
1834 // address of the callee into ECX. The value in ecx is used as target of
1835 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1836 // for tail calls on PIC/GOT architectures. Normally we would just put the
1837 // address of GOT into ebx and then call target@PLT. But for tail calls
1838 // ebx would be restored (since ebx is callee saved) before jumping to the
1841 // Note: The actual moving to ECX is done further down.
1842 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1843 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1844 !G->getGlobal()->hasProtectedVisibility())
1845 Callee = LowerGlobalAddress(Callee, DAG);
1846 else if (isa<ExternalSymbolSDNode>(Callee))
1847 Callee = LowerExternalSymbol(Callee, DAG);
1851 if (Is64Bit && isVarArg) {
1852 // From AMD64 ABI document:
1853 // For calls that may call functions that use varargs or stdargs
1854 // (prototype-less calls or calls to functions containing ellipsis (...) in
1855 // the declaration) %al is used as hidden argument to specify the number
1856 // of SSE registers used. The contents of %al do not need to match exactly
1857 // the number of registers, but must be an ubound on the number of SSE
1858 // registers used and is in the range 0 - 8 inclusive.
1860 // FIXME: Verify this on Win64
1861 // Count the number of XMM registers allocated.
1862 static const unsigned XMMArgRegs[] = {
1863 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1864 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1866 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1867 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1868 && "SSE registers cannot be used when SSE is disabled");
1870 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1871 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1872 InFlag = Chain.getValue(1);
1876 // For tail calls lower the arguments to the 'real' stack slot.
1878 // Force all the incoming stack arguments to be loaded from the stack
1879 // before any new outgoing arguments are stored to the stack, because the
1880 // outgoing stack slots may alias the incoming argument stack slots, and
1881 // the alias isn't otherwise explicit. This is slightly more conservative
1882 // than necessary, because it means that each store effectively depends
1883 // on every argument instead of just those arguments it would clobber.
1884 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1886 SmallVector<SDValue, 8> MemOpChains2;
1889 // Do not flag preceeding copytoreg stuff together with the following stuff.
1891 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1892 CCValAssign &VA = ArgLocs[i];
1893 if (!VA.isRegLoc()) {
1894 assert(VA.isMemLoc());
1895 SDValue Arg = Outs[i].Val;
1896 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1897 // Create frame index.
1898 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1899 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1900 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1901 FIN = DAG.getFrameIndex(FI, getPointerTy());
1903 if (Flags.isByVal()) {
1904 // Copy relative to framepointer.
1905 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1906 if (StackPtr.getNode() == 0)
1907 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1909 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1911 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1915 // Store relative to framepointer.
1916 MemOpChains2.push_back(
1917 DAG.getStore(ArgChain, dl, Arg, FIN,
1918 PseudoSourceValue::getFixedStack(FI), 0));
1923 if (!MemOpChains2.empty())
1924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1925 &MemOpChains2[0], MemOpChains2.size());
1927 // Copy arguments to their registers.
1928 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1929 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1930 RegsToPass[i].second, InFlag);
1931 InFlag = Chain.getValue(1);
1935 // Store the return address to the appropriate stack slot.
1936 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1940 // If the callee is a GlobalAddress node (quite common, every direct call is)
1941 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1942 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1943 // We should use extra load for direct calls to dllimported functions in
1945 GlobalValue *GV = G->getGlobal();
1946 if (!GV->hasDLLImportLinkage()) {
1947 unsigned char OpFlags = 0;
1949 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1950 // external symbols most go through the PLT in PIC mode. If the symbol
1951 // has hidden or protected visibility, or if it is static or local, then
1952 // we don't need to use the PLT - we can directly call it.
1953 if (Subtarget->isTargetELF() &&
1954 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1955 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1956 OpFlags = X86II::MO_PLT;
1957 } else if (Subtarget->isPICStyleStubAny() &&
1958 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1959 Subtarget->getDarwinVers() < 9) {
1960 // PC-relative references to external symbols should go through $stub,
1961 // unless we're building with the leopard linker or later, which
1962 // automatically synthesizes these stubs.
1963 OpFlags = X86II::MO_DARWIN_STUB;
1966 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1967 G->getOffset(), OpFlags);
1969 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1970 unsigned char OpFlags = 0;
1972 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1973 // symbols should go through the PLT.
1974 if (Subtarget->isTargetELF() &&
1975 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1976 OpFlags = X86II::MO_PLT;
1977 } else if (Subtarget->isPICStyleStubAny() &&
1978 Subtarget->getDarwinVers() < 9) {
1979 // PC-relative references to external symbols should go through $stub,
1980 // unless we're building with the leopard linker or later, which
1981 // automatically synthesizes these stubs.
1982 OpFlags = X86II::MO_DARWIN_STUB;
1985 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1987 } else if (isTailCall) {
1988 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1990 Chain = DAG.getCopyToReg(Chain, dl,
1991 DAG.getRegister(Opc, getPointerTy()),
1993 Callee = DAG.getRegister(Opc, getPointerTy());
1994 // Add register as live out.
1995 MF.getRegInfo().addLiveOut(Opc);
1998 // Returns a chain & a flag for retval copy to use.
1999 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2000 SmallVector<SDValue, 8> Ops;
2003 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2004 DAG.getIntPtrConstant(0, true), InFlag);
2005 InFlag = Chain.getValue(1);
2008 Ops.push_back(Chain);
2009 Ops.push_back(Callee);
2012 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2014 // Add argument registers to the end of the list so that they are known live
2016 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2017 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2018 RegsToPass[i].second.getValueType()));
2020 // Add an implicit use GOT pointer in EBX.
2021 if (!isTailCall && Subtarget->isPICStyleGOT())
2022 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2024 // Add an implicit use of AL for x86 vararg functions.
2025 if (Is64Bit && isVarArg)
2026 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2028 if (InFlag.getNode())
2029 Ops.push_back(InFlag);
2032 // If this is the first return lowered for this function, add the regs
2033 // to the liveout set for the function.
2034 if (MF.getRegInfo().liveout_empty()) {
2035 SmallVector<CCValAssign, 16> RVLocs;
2036 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2038 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2039 for (unsigned i = 0; i != RVLocs.size(); ++i)
2040 if (RVLocs[i].isRegLoc())
2041 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2044 assert(((Callee.getOpcode() == ISD::Register &&
2045 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2046 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2047 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2048 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2049 "Expecting an global address, external symbol, or register");
2051 return DAG.getNode(X86ISD::TC_RETURN, dl,
2052 NodeTys, &Ops[0], Ops.size());
2055 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2056 InFlag = Chain.getValue(1);
2058 // Create the CALLSEQ_END node.
2059 unsigned NumBytesForCalleeToPush;
2060 if (IsCalleePop(isVarArg, CallConv))
2061 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2062 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2063 // If this is is a call to a struct-return function, the callee
2064 // pops the hidden struct pointer, so we have to push it back.
2065 // This is common for Darwin/X86, Linux & Mingw32 targets.
2066 NumBytesForCalleeToPush = 4;
2068 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2070 // Returns a flag for retval copy to use.
2071 Chain = DAG.getCALLSEQ_END(Chain,
2072 DAG.getIntPtrConstant(NumBytes, true),
2073 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2076 InFlag = Chain.getValue(1);
2078 // Handle result values, copying them out of physregs into vregs that we
2080 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2081 Ins, dl, DAG, InVals);
2085 //===----------------------------------------------------------------------===//
2086 // Fast Calling Convention (tail call) implementation
2087 //===----------------------------------------------------------------------===//
2089 // Like std call, callee cleans arguments, convention except that ECX is
2090 // reserved for storing the tail called function address. Only 2 registers are
2091 // free for argument passing (inreg). Tail call optimization is performed
2093 // * tailcallopt is enabled
2094 // * caller/callee are fastcc
2095 // On X86_64 architecture with GOT-style position independent code only local
2096 // (within module) calls are supported at the moment.
2097 // To keep the stack aligned according to platform abi the function
2098 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2099 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2100 // If a tail called function callee has more arguments than the caller the
2101 // caller needs to make sure that there is room to move the RETADDR to. This is
2102 // achieved by reserving an area the size of the argument delta right after the
2103 // original REtADDR, but before the saved framepointer or the spilled registers
2104 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2116 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2117 /// for a 16 byte align requirement.
2118 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2119 SelectionDAG& DAG) {
2120 MachineFunction &MF = DAG.getMachineFunction();
2121 const TargetMachine &TM = MF.getTarget();
2122 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2123 unsigned StackAlignment = TFI.getStackAlignment();
2124 uint64_t AlignMask = StackAlignment - 1;
2125 int64_t Offset = StackSize;
2126 uint64_t SlotSize = TD->getPointerSize();
2127 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2128 // Number smaller than 12 so just add the difference.
2129 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2131 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2132 Offset = ((~AlignMask) & Offset) + StackAlignment +
2133 (StackAlignment-SlotSize);
2138 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2139 /// for tail call optimization. Targets which want to do tail call
2140 /// optimization should implement this function.
2142 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2145 const SmallVectorImpl<ISD::InputArg> &Ins,
2146 SelectionDAG& DAG) const {
2147 MachineFunction &MF = DAG.getMachineFunction();
2148 unsigned CallerCC = MF.getFunction()->getCallingConv();
2149 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2153 X86TargetLowering::createFastISel(MachineFunction &mf,
2154 MachineModuleInfo *mmo,
2156 DenseMap<const Value *, unsigned> &vm,
2157 DenseMap<const BasicBlock *,
2158 MachineBasicBlock *> &bm,
2159 DenseMap<const AllocaInst *, int> &am
2161 , SmallSet<Instruction*, 8> &cil
2164 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2172 //===----------------------------------------------------------------------===//
2173 // Other Lowering Hooks
2174 //===----------------------------------------------------------------------===//
2177 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2178 MachineFunction &MF = DAG.getMachineFunction();
2179 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2180 int ReturnAddrIndex = FuncInfo->getRAIndex();
2182 if (ReturnAddrIndex == 0) {
2183 // Set up a frame object for the return address.
2184 uint64_t SlotSize = TD->getPointerSize();
2185 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2186 FuncInfo->setRAIndex(ReturnAddrIndex);
2189 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2193 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2194 bool hasSymbolicDisplacement) {
2195 // Offset should fit into 32 bit immediate field.
2196 if (!isInt32(Offset))
2199 // If we don't have a symbolic displacement - we don't have any extra
2201 if (!hasSymbolicDisplacement)
2204 // FIXME: Some tweaks might be needed for medium code model.
2205 if (M != CodeModel::Small && M != CodeModel::Kernel)
2208 // For small code model we assume that latest object is 16MB before end of 31
2209 // bits boundary. We may also accept pretty large negative constants knowing
2210 // that all objects are in the positive half of address space.
2211 if (M == CodeModel::Small && Offset < 16*1024*1024)
2214 // For kernel code model we know that all object resist in the negative half
2215 // of 32bits address space. We may not accept negative offsets, since they may
2216 // be just off and we may accept pretty large positive ones.
2217 if (M == CodeModel::Kernel && Offset > 0)
2223 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2224 /// specific condition code, returning the condition code and the LHS/RHS of the
2225 /// comparison to make.
2226 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2227 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2229 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2230 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2231 // X > -1 -> X == 0, jump !sign.
2232 RHS = DAG.getConstant(0, RHS.getValueType());
2233 return X86::COND_NS;
2234 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2235 // X < 0 -> X == 0, jump on sign.
2237 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2239 RHS = DAG.getConstant(0, RHS.getValueType());
2240 return X86::COND_LE;
2244 switch (SetCCOpcode) {
2245 default: llvm_unreachable("Invalid integer condition!");
2246 case ISD::SETEQ: return X86::COND_E;
2247 case ISD::SETGT: return X86::COND_G;
2248 case ISD::SETGE: return X86::COND_GE;
2249 case ISD::SETLT: return X86::COND_L;
2250 case ISD::SETLE: return X86::COND_LE;
2251 case ISD::SETNE: return X86::COND_NE;
2252 case ISD::SETULT: return X86::COND_B;
2253 case ISD::SETUGT: return X86::COND_A;
2254 case ISD::SETULE: return X86::COND_BE;
2255 case ISD::SETUGE: return X86::COND_AE;
2259 // First determine if it is required or is profitable to flip the operands.
2261 // If LHS is a foldable load, but RHS is not, flip the condition.
2262 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2263 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2264 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2265 std::swap(LHS, RHS);
2268 switch (SetCCOpcode) {
2274 std::swap(LHS, RHS);
2278 // On a floating point condition, the flags are set as follows:
2280 // 0 | 0 | 0 | X > Y
2281 // 0 | 0 | 1 | X < Y
2282 // 1 | 0 | 0 | X == Y
2283 // 1 | 1 | 1 | unordered
2284 switch (SetCCOpcode) {
2285 default: llvm_unreachable("Condcode should be pre-legalized away");
2287 case ISD::SETEQ: return X86::COND_E;
2288 case ISD::SETOLT: // flipped
2290 case ISD::SETGT: return X86::COND_A;
2291 case ISD::SETOLE: // flipped
2293 case ISD::SETGE: return X86::COND_AE;
2294 case ISD::SETUGT: // flipped
2296 case ISD::SETLT: return X86::COND_B;
2297 case ISD::SETUGE: // flipped
2299 case ISD::SETLE: return X86::COND_BE;
2301 case ISD::SETNE: return X86::COND_NE;
2302 case ISD::SETUO: return X86::COND_P;
2303 case ISD::SETO: return X86::COND_NP;
2307 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2308 /// code. Current x86 isa includes the following FP cmov instructions:
2309 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2310 static bool hasFPCMov(unsigned X86CC) {
2326 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2327 /// the specified range (L, H].
2328 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2329 return (Val < 0) || (Val >= Low && Val < Hi);
2332 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2333 /// specified value.
2334 static bool isUndefOrEqual(int Val, int CmpVal) {
2335 if (Val < 0 || Val == CmpVal)
2340 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2341 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2342 /// the second operand.
2343 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2344 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2345 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2346 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2347 return (Mask[0] < 2 && Mask[1] < 2);
2351 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2352 SmallVector<int, 8> M;
2354 return ::isPSHUFDMask(M, N->getValueType(0));
2357 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2358 /// is suitable for input to PSHUFHW.
2359 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2360 if (VT != MVT::v8i16)
2363 // Lower quadword copied in order or undef.
2364 for (int i = 0; i != 4; ++i)
2365 if (Mask[i] >= 0 && Mask[i] != i)
2368 // Upper quadword shuffled.
2369 for (int i = 4; i != 8; ++i)
2370 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2376 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2377 SmallVector<int, 8> M;
2379 return ::isPSHUFHWMask(M, N->getValueType(0));
2382 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2383 /// is suitable for input to PSHUFLW.
2384 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2385 if (VT != MVT::v8i16)
2388 // Upper quadword copied in order.
2389 for (int i = 4; i != 8; ++i)
2390 if (Mask[i] >= 0 && Mask[i] != i)
2393 // Lower quadword shuffled.
2394 for (int i = 0; i != 4; ++i)
2401 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2402 SmallVector<int, 8> M;
2404 return ::isPSHUFLWMask(M, N->getValueType(0));
2407 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2408 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2409 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2410 int NumElems = VT.getVectorNumElements();
2411 if (NumElems != 2 && NumElems != 4)
2414 int Half = NumElems / 2;
2415 for (int i = 0; i < Half; ++i)
2416 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2418 for (int i = Half; i < NumElems; ++i)
2419 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2425 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2426 SmallVector<int, 8> M;
2428 return ::isSHUFPMask(M, N->getValueType(0));
2431 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2432 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2433 /// half elements to come from vector 1 (which would equal the dest.) and
2434 /// the upper half to come from vector 2.
2435 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2436 int NumElems = VT.getVectorNumElements();
2438 if (NumElems != 2 && NumElems != 4)
2441 int Half = NumElems / 2;
2442 for (int i = 0; i < Half; ++i)
2443 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2445 for (int i = Half; i < NumElems; ++i)
2446 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2451 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2452 SmallVector<int, 8> M;
2454 return isCommutedSHUFPMask(M, N->getValueType(0));
2457 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2458 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2459 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2460 if (N->getValueType(0).getVectorNumElements() != 4)
2463 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2464 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2465 isUndefOrEqual(N->getMaskElt(1), 7) &&
2466 isUndefOrEqual(N->getMaskElt(2), 2) &&
2467 isUndefOrEqual(N->getMaskElt(3), 3);
2470 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2471 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2472 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2473 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2475 if (NumElems != 2 && NumElems != 4)
2478 for (unsigned i = 0; i < NumElems/2; ++i)
2479 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2482 for (unsigned i = NumElems/2; i < NumElems; ++i)
2483 if (!isUndefOrEqual(N->getMaskElt(i), i))
2489 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2490 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2492 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2493 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2495 if (NumElems != 2 && NumElems != 4)
2498 for (unsigned i = 0; i < NumElems/2; ++i)
2499 if (!isUndefOrEqual(N->getMaskElt(i), i))
2502 for (unsigned i = 0; i < NumElems/2; ++i)
2503 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2509 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2510 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2512 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2513 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2518 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2519 isUndefOrEqual(N->getMaskElt(1), 3) &&
2520 isUndefOrEqual(N->getMaskElt(2), 2) &&
2521 isUndefOrEqual(N->getMaskElt(3), 3);
2524 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2525 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2526 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2527 bool V2IsSplat = false) {
2528 int NumElts = VT.getVectorNumElements();
2529 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2532 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2534 int BitI1 = Mask[i+1];
2535 if (!isUndefOrEqual(BitI, j))
2538 if (!isUndefOrEqual(BitI1, NumElts))
2541 if (!isUndefOrEqual(BitI1, j + NumElts))
2548 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2549 SmallVector<int, 8> M;
2551 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2554 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2555 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2556 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2557 bool V2IsSplat = false) {
2558 int NumElts = VT.getVectorNumElements();
2559 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2562 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2564 int BitI1 = Mask[i+1];
2565 if (!isUndefOrEqual(BitI, j + NumElts/2))
2568 if (isUndefOrEqual(BitI1, NumElts))
2571 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2578 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2579 SmallVector<int, 8> M;
2581 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2584 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2585 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2587 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2588 int NumElems = VT.getVectorNumElements();
2589 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2592 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2594 int BitI1 = Mask[i+1];
2595 if (!isUndefOrEqual(BitI, j))
2597 if (!isUndefOrEqual(BitI1, j))
2603 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2604 SmallVector<int, 8> M;
2606 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2609 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2610 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2612 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2613 int NumElems = VT.getVectorNumElements();
2614 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2617 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2619 int BitI1 = Mask[i+1];
2620 if (!isUndefOrEqual(BitI, j))
2622 if (!isUndefOrEqual(BitI1, j))
2628 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2629 SmallVector<int, 8> M;
2631 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2634 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2635 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2636 /// MOVSD, and MOVD, i.e. setting the lowest element.
2637 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2638 if (VT.getVectorElementType().getSizeInBits() < 32)
2641 int NumElts = VT.getVectorNumElements();
2643 if (!isUndefOrEqual(Mask[0], NumElts))
2646 for (int i = 1; i < NumElts; ++i)
2647 if (!isUndefOrEqual(Mask[i], i))
2653 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2654 SmallVector<int, 8> M;
2656 return ::isMOVLMask(M, N->getValueType(0));
2659 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2660 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2661 /// element of vector 2 and the other elements to come from vector 1 in order.
2662 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2663 bool V2IsSplat = false, bool V2IsUndef = false) {
2664 int NumOps = VT.getVectorNumElements();
2665 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2668 if (!isUndefOrEqual(Mask[0], 0))
2671 for (int i = 1; i < NumOps; ++i)
2672 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2673 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2674 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2680 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2681 bool V2IsUndef = false) {
2682 SmallVector<int, 8> M;
2684 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2687 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2688 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2689 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2690 if (N->getValueType(0).getVectorNumElements() != 4)
2693 // Expect 1, 1, 3, 3
2694 for (unsigned i = 0; i < 2; ++i) {
2695 int Elt = N->getMaskElt(i);
2696 if (Elt >= 0 && Elt != 1)
2701 for (unsigned i = 2; i < 4; ++i) {
2702 int Elt = N->getMaskElt(i);
2703 if (Elt >= 0 && Elt != 3)
2708 // Don't use movshdup if it can be done with a shufps.
2709 // FIXME: verify that matching u, u, 3, 3 is what we want.
2713 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2714 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2715 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2716 if (N->getValueType(0).getVectorNumElements() != 4)
2719 // Expect 0, 0, 2, 2
2720 for (unsigned i = 0; i < 2; ++i)
2721 if (N->getMaskElt(i) > 0)
2725 for (unsigned i = 2; i < 4; ++i) {
2726 int Elt = N->getMaskElt(i);
2727 if (Elt >= 0 && Elt != 2)
2732 // Don't use movsldup if it can be done with a shufps.
2736 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2737 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2738 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2739 int e = N->getValueType(0).getVectorNumElements() / 2;
2741 for (int i = 0; i < e; ++i)
2742 if (!isUndefOrEqual(N->getMaskElt(i), i))
2744 for (int i = 0; i < e; ++i)
2745 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2750 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2751 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2753 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2754 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2755 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2757 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2759 for (int i = 0; i < NumOperands; ++i) {
2760 int Val = SVOp->getMaskElt(NumOperands-i-1);
2761 if (Val < 0) Val = 0;
2762 if (Val >= NumOperands) Val -= NumOperands;
2764 if (i != NumOperands - 1)
2770 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2771 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2773 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2774 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2776 // 8 nodes, but we only care about the last 4.
2777 for (unsigned i = 7; i >= 4; --i) {
2778 int Val = SVOp->getMaskElt(i);
2787 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2788 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2790 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2791 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2793 // 8 nodes, but we only care about the first 4.
2794 for (int i = 3; i >= 0; --i) {
2795 int Val = SVOp->getMaskElt(i);
2804 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2806 bool X86::isZeroNode(SDValue Elt) {
2807 return ((isa<ConstantSDNode>(Elt) &&
2808 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2809 (isa<ConstantFPSDNode>(Elt) &&
2810 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2813 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2814 /// their permute mask.
2815 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2816 SelectionDAG &DAG) {
2817 EVT VT = SVOp->getValueType(0);
2818 unsigned NumElems = VT.getVectorNumElements();
2819 SmallVector<int, 8> MaskVec;
2821 for (unsigned i = 0; i != NumElems; ++i) {
2822 int idx = SVOp->getMaskElt(i);
2824 MaskVec.push_back(idx);
2825 else if (idx < (int)NumElems)
2826 MaskVec.push_back(idx + NumElems);
2828 MaskVec.push_back(idx - NumElems);
2830 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2831 SVOp->getOperand(0), &MaskVec[0]);
2834 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2835 /// the two vector operands have swapped position.
2836 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2837 unsigned NumElems = VT.getVectorNumElements();
2838 for (unsigned i = 0; i != NumElems; ++i) {
2842 else if (idx < (int)NumElems)
2843 Mask[i] = idx + NumElems;
2845 Mask[i] = idx - NumElems;
2849 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2850 /// match movhlps. The lower half elements should come from upper half of
2851 /// V1 (and in order), and the upper half elements should come from the upper
2852 /// half of V2 (and in order).
2853 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2854 if (Op->getValueType(0).getVectorNumElements() != 4)
2856 for (unsigned i = 0, e = 2; i != e; ++i)
2857 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2859 for (unsigned i = 2; i != 4; ++i)
2860 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2865 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2866 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2868 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2869 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2871 N = N->getOperand(0).getNode();
2872 if (!ISD::isNON_EXTLoad(N))
2875 *LD = cast<LoadSDNode>(N);
2879 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2880 /// match movlp{s|d}. The lower half elements should come from lower half of
2881 /// V1 (and in order), and the upper half elements should come from the upper
2882 /// half of V2 (and in order). And since V1 will become the source of the
2883 /// MOVLP, it must be either a vector load or a scalar load to vector.
2884 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2885 ShuffleVectorSDNode *Op) {
2886 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2888 // Is V2 is a vector load, don't do this transformation. We will try to use
2889 // load folding shufps op.
2890 if (ISD::isNON_EXTLoad(V2))
2893 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2895 if (NumElems != 2 && NumElems != 4)
2897 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2898 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2900 for (unsigned i = NumElems/2; i != NumElems; ++i)
2901 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2906 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2908 static bool isSplatVector(SDNode *N) {
2909 if (N->getOpcode() != ISD::BUILD_VECTOR)
2912 SDValue SplatValue = N->getOperand(0);
2913 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2914 if (N->getOperand(i) != SplatValue)
2919 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2920 /// to an zero vector.
2921 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2922 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2923 SDValue V1 = N->getOperand(0);
2924 SDValue V2 = N->getOperand(1);
2925 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2926 for (unsigned i = 0; i != NumElems; ++i) {
2927 int Idx = N->getMaskElt(i);
2928 if (Idx >= (int)NumElems) {
2929 unsigned Opc = V2.getOpcode();
2930 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2932 if (Opc != ISD::BUILD_VECTOR ||
2933 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2935 } else if (Idx >= 0) {
2936 unsigned Opc = V1.getOpcode();
2937 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2939 if (Opc != ISD::BUILD_VECTOR ||
2940 !X86::isZeroNode(V1.getOperand(Idx)))
2947 /// getZeroVector - Returns a vector of specified type with all zero elements.
2949 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
2951 assert(VT.isVector() && "Expected a vector type");
2953 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2954 // type. This ensures they get CSE'd.
2956 if (VT.getSizeInBits() == 64) { // MMX
2957 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2958 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2959 } else if (HasSSE2) { // SSE2
2960 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2961 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2963 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2964 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2966 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2969 /// getOnesVector - Returns a vector of specified type with all bits set.
2971 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2972 assert(VT.isVector() && "Expected a vector type");
2974 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2975 // type. This ensures they get CSE'd.
2976 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2978 if (VT.getSizeInBits() == 64) // MMX
2979 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2981 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2982 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2986 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2987 /// that point to V2 points to its first element.
2988 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2989 EVT VT = SVOp->getValueType(0);
2990 unsigned NumElems = VT.getVectorNumElements();
2992 bool Changed = false;
2993 SmallVector<int, 8> MaskVec;
2994 SVOp->getMask(MaskVec);
2996 for (unsigned i = 0; i != NumElems; ++i) {
2997 if (MaskVec[i] > (int)NumElems) {
2998 MaskVec[i] = NumElems;
3003 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3004 SVOp->getOperand(1), &MaskVec[0]);
3005 return SDValue(SVOp, 0);
3008 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3009 /// operation of specified width.
3010 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3012 unsigned NumElems = VT.getVectorNumElements();
3013 SmallVector<int, 8> Mask;
3014 Mask.push_back(NumElems);
3015 for (unsigned i = 1; i != NumElems; ++i)
3017 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3020 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3021 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3023 unsigned NumElems = VT.getVectorNumElements();
3024 SmallVector<int, 8> Mask;
3025 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3027 Mask.push_back(i + NumElems);
3029 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3032 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3033 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3035 unsigned NumElems = VT.getVectorNumElements();
3036 unsigned Half = NumElems/2;
3037 SmallVector<int, 8> Mask;
3038 for (unsigned i = 0; i != Half; ++i) {
3039 Mask.push_back(i + Half);
3040 Mask.push_back(i + NumElems + Half);
3042 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3045 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3046 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3048 if (SV->getValueType(0).getVectorNumElements() <= 4)
3049 return SDValue(SV, 0);
3051 EVT PVT = MVT::v4f32;
3052 EVT VT = SV->getValueType(0);
3053 DebugLoc dl = SV->getDebugLoc();
3054 SDValue V1 = SV->getOperand(0);
3055 int NumElems = VT.getVectorNumElements();
3056 int EltNo = SV->getSplatIndex();
3058 // unpack elements to the correct location
3059 while (NumElems > 4) {
3060 if (EltNo < NumElems/2) {
3061 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3063 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3064 EltNo -= NumElems/2;
3069 // Perform the splat.
3070 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3071 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3072 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3073 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3076 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3077 /// vector of zero or undef vector. This produces a shuffle where the low
3078 /// element of V2 is swizzled into the zero/undef vector, landing at element
3079 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3080 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3081 bool isZero, bool HasSSE2,
3082 SelectionDAG &DAG) {
3083 EVT VT = V2.getValueType();
3085 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3086 unsigned NumElems = VT.getVectorNumElements();
3087 SmallVector<int, 16> MaskVec;
3088 for (unsigned i = 0; i != NumElems; ++i)
3089 // If this is the insertion idx, put the low elt of V2 here.
3090 MaskVec.push_back(i == Idx ? NumElems : i);
3091 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3094 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3095 /// a shuffle that is zero.
3097 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3098 bool Low, SelectionDAG &DAG) {
3099 unsigned NumZeros = 0;
3100 for (int i = 0; i < NumElems; ++i) {
3101 unsigned Index = Low ? i : NumElems-i-1;
3102 int Idx = SVOp->getMaskElt(Index);
3107 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3108 if (Elt.getNode() && X86::isZeroNode(Elt))
3116 /// isVectorShift - Returns true if the shuffle can be implemented as a
3117 /// logical left or right shift of a vector.
3118 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3119 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3120 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3121 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3124 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3127 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3131 bool SeenV1 = false;
3132 bool SeenV2 = false;
3133 for (int i = NumZeros; i < NumElems; ++i) {
3134 int Val = isLeft ? (i - NumZeros) : i;
3135 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3147 if (SeenV1 && SeenV2)
3150 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3156 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3158 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3159 unsigned NumNonZero, unsigned NumZero,
3160 SelectionDAG &DAG, TargetLowering &TLI) {
3164 DebugLoc dl = Op.getDebugLoc();
3167 for (unsigned i = 0; i < 16; ++i) {
3168 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3169 if (ThisIsNonZero && First) {
3171 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3173 V = DAG.getUNDEF(MVT::v8i16);
3178 SDValue ThisElt(0, 0), LastElt(0, 0);
3179 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3180 if (LastIsNonZero) {
3181 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3182 MVT::i16, Op.getOperand(i-1));
3184 if (ThisIsNonZero) {
3185 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3186 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3187 ThisElt, DAG.getConstant(8, MVT::i8));
3189 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3193 if (ThisElt.getNode())
3194 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3195 DAG.getIntPtrConstant(i/2));
3199 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3202 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3204 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3205 unsigned NumNonZero, unsigned NumZero,
3206 SelectionDAG &DAG, TargetLowering &TLI) {
3210 DebugLoc dl = Op.getDebugLoc();
3213 for (unsigned i = 0; i < 8; ++i) {
3214 bool isNonZero = (NonZeros & (1 << i)) != 0;
3218 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3220 V = DAG.getUNDEF(MVT::v8i16);
3223 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3224 MVT::v8i16, V, Op.getOperand(i),
3225 DAG.getIntPtrConstant(i));
3232 /// getVShift - Return a vector logical shift node.
3234 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3235 unsigned NumBits, SelectionDAG &DAG,
3236 const TargetLowering &TLI, DebugLoc dl) {
3237 bool isMMX = VT.getSizeInBits() == 64;
3238 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3239 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3240 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3241 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3242 DAG.getNode(Opc, dl, ShVT, SrcOp,
3243 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3247 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3248 DebugLoc dl = Op.getDebugLoc();
3249 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3250 if (ISD::isBuildVectorAllZeros(Op.getNode())
3251 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3252 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3253 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3254 // eliminated on x86-32 hosts.
3255 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3258 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3259 return getOnesVector(Op.getValueType(), DAG, dl);
3260 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3263 EVT VT = Op.getValueType();
3264 EVT ExtVT = VT.getVectorElementType();
3265 unsigned EVTBits = ExtVT.getSizeInBits();
3267 unsigned NumElems = Op.getNumOperands();
3268 unsigned NumZero = 0;
3269 unsigned NumNonZero = 0;
3270 unsigned NonZeros = 0;
3271 bool IsAllConstants = true;
3272 SmallSet<SDValue, 8> Values;
3273 for (unsigned i = 0; i < NumElems; ++i) {
3274 SDValue Elt = Op.getOperand(i);
3275 if (Elt.getOpcode() == ISD::UNDEF)
3278 if (Elt.getOpcode() != ISD::Constant &&
3279 Elt.getOpcode() != ISD::ConstantFP)
3280 IsAllConstants = false;
3281 if (X86::isZeroNode(Elt))
3284 NonZeros |= (1 << i);
3289 if (NumNonZero == 0) {
3290 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3291 return DAG.getUNDEF(VT);
3294 // Special case for single non-zero, non-undef, element.
3295 if (NumNonZero == 1) {
3296 unsigned Idx = CountTrailingZeros_32(NonZeros);
3297 SDValue Item = Op.getOperand(Idx);
3299 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3300 // the value are obviously zero, truncate the value to i32 and do the
3301 // insertion that way. Only do this if the value is non-constant or if the
3302 // value is a constant being inserted into element 0. It is cheaper to do
3303 // a constant pool load than it is to do a movd + shuffle.
3304 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3305 (!IsAllConstants || Idx == 0)) {
3306 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3307 // Handle MMX and SSE both.
3308 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3309 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3311 // Truncate the value (which may itself be a constant) to i32, and
3312 // convert it to a vector with movd (S2V+shuffle to zero extend).
3313 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3314 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3315 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3316 Subtarget->hasSSE2(), DAG);
3318 // Now we have our 32-bit value zero extended in the low element of
3319 // a vector. If Idx != 0, swizzle it into place.
3321 SmallVector<int, 4> Mask;
3322 Mask.push_back(Idx);
3323 for (unsigned i = 1; i != VecElts; ++i)
3325 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3326 DAG.getUNDEF(Item.getValueType()),
3329 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3333 // If we have a constant or non-constant insertion into the low element of
3334 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3335 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3336 // depending on what the source datatype is.
3339 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3340 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3341 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3342 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3343 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3344 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3346 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3347 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3348 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3349 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3350 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3351 Subtarget->hasSSE2(), DAG);
3352 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3356 // Is it a vector logical left shift?
3357 if (NumElems == 2 && Idx == 1 &&
3358 X86::isZeroNode(Op.getOperand(0)) &&
3359 !X86::isZeroNode(Op.getOperand(1))) {
3360 unsigned NumBits = VT.getSizeInBits();
3361 return getVShift(true, VT,
3362 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3363 VT, Op.getOperand(1)),
3364 NumBits/2, DAG, *this, dl);
3367 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3370 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3371 // is a non-constant being inserted into an element other than the low one,
3372 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3373 // movd/movss) to move this into the low element, then shuffle it into
3375 if (EVTBits == 32) {
3376 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3378 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3379 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3380 Subtarget->hasSSE2(), DAG);
3381 SmallVector<int, 8> MaskVec;
3382 for (unsigned i = 0; i < NumElems; i++)
3383 MaskVec.push_back(i == Idx ? 0 : 1);
3384 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3388 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3389 if (Values.size() == 1)
3392 // A vector full of immediates; various special cases are already
3393 // handled, so this is best done with a single constant-pool load.
3397 // Let legalizer expand 2-wide build_vectors.
3398 if (EVTBits == 64) {
3399 if (NumNonZero == 1) {
3400 // One half is zero or undef.
3401 unsigned Idx = CountTrailingZeros_32(NonZeros);
3402 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3403 Op.getOperand(Idx));
3404 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3405 Subtarget->hasSSE2(), DAG);
3410 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3411 if (EVTBits == 8 && NumElems == 16) {
3412 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3414 if (V.getNode()) return V;
3417 if (EVTBits == 16 && NumElems == 8) {
3418 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3420 if (V.getNode()) return V;
3423 // If element VT is == 32 bits, turn it into a number of shuffles.
3424 SmallVector<SDValue, 8> V;
3426 if (NumElems == 4 && NumZero > 0) {
3427 for (unsigned i = 0; i < 4; ++i) {
3428 bool isZero = !(NonZeros & (1 << i));
3430 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3432 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3435 for (unsigned i = 0; i < 2; ++i) {
3436 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3439 V[i] = V[i*2]; // Must be a zero vector.
3442 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3445 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3448 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3453 SmallVector<int, 8> MaskVec;
3454 bool Reverse = (NonZeros & 0x3) == 2;
3455 for (unsigned i = 0; i < 2; ++i)
3456 MaskVec.push_back(Reverse ? 1-i : i);
3457 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3458 for (unsigned i = 0; i < 2; ++i)
3459 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3460 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3463 if (Values.size() > 2) {
3464 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3465 // values to be inserted is equal to the number of elements, in which case
3466 // use the unpack code below in the hopes of matching the consecutive elts
3467 // load merge pattern for shuffles.
3468 // FIXME: We could probably just check that here directly.
3469 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3470 getSubtarget()->hasSSE41()) {
3471 V[0] = DAG.getUNDEF(VT);
3472 for (unsigned i = 0; i < NumElems; ++i)
3473 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3474 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3475 Op.getOperand(i), DAG.getIntPtrConstant(i));
3478 // Expand into a number of unpckl*.
3480 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3481 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3482 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3483 for (unsigned i = 0; i < NumElems; ++i)
3484 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3486 while (NumElems != 0) {
3487 for (unsigned i = 0; i < NumElems; ++i)
3488 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3497 // v8i16 shuffles - Prefer shuffles in the following order:
3498 // 1. [all] pshuflw, pshufhw, optional move
3499 // 2. [ssse3] 1 x pshufb
3500 // 3. [ssse3] 2 x pshufb + 1 x por
3501 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3503 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3504 SelectionDAG &DAG, X86TargetLowering &TLI) {
3505 SDValue V1 = SVOp->getOperand(0);
3506 SDValue V2 = SVOp->getOperand(1);
3507 DebugLoc dl = SVOp->getDebugLoc();
3508 SmallVector<int, 8> MaskVals;
3510 // Determine if more than 1 of the words in each of the low and high quadwords
3511 // of the result come from the same quadword of one of the two inputs. Undef
3512 // mask values count as coming from any quadword, for better codegen.
3513 SmallVector<unsigned, 4> LoQuad(4);
3514 SmallVector<unsigned, 4> HiQuad(4);
3515 BitVector InputQuads(4);
3516 for (unsigned i = 0; i < 8; ++i) {
3517 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3518 int EltIdx = SVOp->getMaskElt(i);
3519 MaskVals.push_back(EltIdx);
3528 InputQuads.set(EltIdx / 4);
3531 int BestLoQuad = -1;
3532 unsigned MaxQuad = 1;
3533 for (unsigned i = 0; i < 4; ++i) {
3534 if (LoQuad[i] > MaxQuad) {
3536 MaxQuad = LoQuad[i];
3540 int BestHiQuad = -1;
3542 for (unsigned i = 0; i < 4; ++i) {
3543 if (HiQuad[i] > MaxQuad) {
3545 MaxQuad = HiQuad[i];
3549 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3550 // of the two input vectors, shuffle them into one input vector so only a
3551 // single pshufb instruction is necessary. If There are more than 2 input
3552 // quads, disable the next transformation since it does not help SSSE3.
3553 bool V1Used = InputQuads[0] || InputQuads[1];
3554 bool V2Used = InputQuads[2] || InputQuads[3];
3555 if (TLI.getSubtarget()->hasSSSE3()) {
3556 if (InputQuads.count() == 2 && V1Used && V2Used) {
3557 BestLoQuad = InputQuads.find_first();
3558 BestHiQuad = InputQuads.find_next(BestLoQuad);
3560 if (InputQuads.count() > 2) {
3566 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3567 // the shuffle mask. If a quad is scored as -1, that means that it contains
3568 // words from all 4 input quadwords.
3570 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3571 SmallVector<int, 8> MaskV;
3572 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3573 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3574 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3575 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3576 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3577 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3579 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3580 // source words for the shuffle, to aid later transformations.
3581 bool AllWordsInNewV = true;
3582 bool InOrder[2] = { true, true };
3583 for (unsigned i = 0; i != 8; ++i) {
3584 int idx = MaskVals[i];
3586 InOrder[i/4] = false;
3587 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3589 AllWordsInNewV = false;
3593 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3594 if (AllWordsInNewV) {
3595 for (int i = 0; i != 8; ++i) {
3596 int idx = MaskVals[i];
3599 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3600 if ((idx != i) && idx < 4)
3602 if ((idx != i) && idx > 3)
3611 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3612 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3613 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3614 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3615 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3619 // If we have SSSE3, and all words of the result are from 1 input vector,
3620 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3621 // is present, fall back to case 4.
3622 if (TLI.getSubtarget()->hasSSSE3()) {
3623 SmallVector<SDValue,16> pshufbMask;
3625 // If we have elements from both input vectors, set the high bit of the
3626 // shuffle mask element to zero out elements that come from V2 in the V1
3627 // mask, and elements that come from V1 in the V2 mask, so that the two
3628 // results can be OR'd together.
3629 bool TwoInputs = V1Used && V2Used;
3630 for (unsigned i = 0; i != 8; ++i) {
3631 int EltIdx = MaskVals[i] * 2;
3632 if (TwoInputs && (EltIdx >= 16)) {
3633 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3634 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3637 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3638 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3640 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3641 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3642 DAG.getNode(ISD::BUILD_VECTOR, dl,
3643 MVT::v16i8, &pshufbMask[0], 16));
3645 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3647 // Calculate the shuffle mask for the second input, shuffle it, and
3648 // OR it with the first shuffled input.
3650 for (unsigned i = 0; i != 8; ++i) {
3651 int EltIdx = MaskVals[i] * 2;
3653 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3654 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3657 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3658 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3660 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3661 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3662 DAG.getNode(ISD::BUILD_VECTOR, dl,
3663 MVT::v16i8, &pshufbMask[0], 16));
3664 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3665 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3668 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3669 // and update MaskVals with new element order.
3670 BitVector InOrder(8);
3671 if (BestLoQuad >= 0) {
3672 SmallVector<int, 8> MaskV;
3673 for (int i = 0; i != 4; ++i) {
3674 int idx = MaskVals[i];
3676 MaskV.push_back(-1);
3678 } else if ((idx / 4) == BestLoQuad) {
3679 MaskV.push_back(idx & 3);
3682 MaskV.push_back(-1);
3685 for (unsigned i = 4; i != 8; ++i)
3687 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3691 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3692 // and update MaskVals with the new element order.
3693 if (BestHiQuad >= 0) {
3694 SmallVector<int, 8> MaskV;
3695 for (unsigned i = 0; i != 4; ++i)
3697 for (unsigned i = 4; i != 8; ++i) {
3698 int idx = MaskVals[i];
3700 MaskV.push_back(-1);
3702 } else if ((idx / 4) == BestHiQuad) {
3703 MaskV.push_back((idx & 3) + 4);
3706 MaskV.push_back(-1);
3709 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3713 // In case BestHi & BestLo were both -1, which means each quadword has a word
3714 // from each of the four input quadwords, calculate the InOrder bitvector now
3715 // before falling through to the insert/extract cleanup.
3716 if (BestLoQuad == -1 && BestHiQuad == -1) {
3718 for (int i = 0; i != 8; ++i)
3719 if (MaskVals[i] < 0 || MaskVals[i] == i)
3723 // The other elements are put in the right place using pextrw and pinsrw.
3724 for (unsigned i = 0; i != 8; ++i) {
3727 int EltIdx = MaskVals[i];
3730 SDValue ExtOp = (EltIdx < 8)
3731 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3732 DAG.getIntPtrConstant(EltIdx))
3733 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3734 DAG.getIntPtrConstant(EltIdx - 8));
3735 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3736 DAG.getIntPtrConstant(i));
3741 // v16i8 shuffles - Prefer shuffles in the following order:
3742 // 1. [ssse3] 1 x pshufb
3743 // 2. [ssse3] 2 x pshufb + 1 x por
3744 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3746 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3747 SelectionDAG &DAG, X86TargetLowering &TLI) {
3748 SDValue V1 = SVOp->getOperand(0);
3749 SDValue V2 = SVOp->getOperand(1);
3750 DebugLoc dl = SVOp->getDebugLoc();
3751 SmallVector<int, 16> MaskVals;
3752 SVOp->getMask(MaskVals);
3754 // If we have SSSE3, case 1 is generated when all result bytes come from
3755 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3756 // present, fall back to case 3.
3757 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3760 for (unsigned i = 0; i < 16; ++i) {
3761 int EltIdx = MaskVals[i];
3770 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3771 if (TLI.getSubtarget()->hasSSSE3()) {
3772 SmallVector<SDValue,16> pshufbMask;
3774 // If all result elements are from one input vector, then only translate
3775 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3777 // Otherwise, we have elements from both input vectors, and must zero out
3778 // elements that come from V2 in the first mask, and V1 in the second mask
3779 // so that we can OR them together.
3780 bool TwoInputs = !(V1Only || V2Only);
3781 for (unsigned i = 0; i != 16; ++i) {
3782 int EltIdx = MaskVals[i];
3783 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3784 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3787 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3789 // If all the elements are from V2, assign it to V1 and return after
3790 // building the first pshufb.
3793 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3794 DAG.getNode(ISD::BUILD_VECTOR, dl,
3795 MVT::v16i8, &pshufbMask[0], 16));
3799 // Calculate the shuffle mask for the second input, shuffle it, and
3800 // OR it with the first shuffled input.
3802 for (unsigned i = 0; i != 16; ++i) {
3803 int EltIdx = MaskVals[i];
3805 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3808 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3810 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3811 DAG.getNode(ISD::BUILD_VECTOR, dl,
3812 MVT::v16i8, &pshufbMask[0], 16));
3813 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3816 // No SSSE3 - Calculate in place words and then fix all out of place words
3817 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3818 // the 16 different words that comprise the two doublequadword input vectors.
3819 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3820 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3821 SDValue NewV = V2Only ? V2 : V1;
3822 for (int i = 0; i != 8; ++i) {
3823 int Elt0 = MaskVals[i*2];
3824 int Elt1 = MaskVals[i*2+1];
3826 // This word of the result is all undef, skip it.
3827 if (Elt0 < 0 && Elt1 < 0)
3830 // This word of the result is already in the correct place, skip it.
3831 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3833 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3836 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3837 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3840 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3841 // using a single extract together, load it and store it.
3842 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3843 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3844 DAG.getIntPtrConstant(Elt1 / 2));
3845 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3846 DAG.getIntPtrConstant(i));
3850 // If Elt1 is defined, extract it from the appropriate source. If the
3851 // source byte is not also odd, shift the extracted word left 8 bits
3852 // otherwise clear the bottom 8 bits if we need to do an or.
3854 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3855 DAG.getIntPtrConstant(Elt1 / 2));
3856 if ((Elt1 & 1) == 0)
3857 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3858 DAG.getConstant(8, TLI.getShiftAmountTy()));
3860 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3861 DAG.getConstant(0xFF00, MVT::i16));
3863 // If Elt0 is defined, extract it from the appropriate source. If the
3864 // source byte is not also even, shift the extracted word right 8 bits. If
3865 // Elt1 was also defined, OR the extracted values together before
3866 // inserting them in the result.
3868 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3869 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3870 if ((Elt0 & 1) != 0)
3871 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3872 DAG.getConstant(8, TLI.getShiftAmountTy()));
3874 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3875 DAG.getConstant(0x00FF, MVT::i16));
3876 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3879 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3880 DAG.getIntPtrConstant(i));
3882 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3885 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3886 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3887 /// done when every pair / quad of shuffle mask elements point to elements in
3888 /// the right sequence. e.g.
3889 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3891 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3893 TargetLowering &TLI, DebugLoc dl) {
3894 EVT VT = SVOp->getValueType(0);
3895 SDValue V1 = SVOp->getOperand(0);
3896 SDValue V2 = SVOp->getOperand(1);
3897 unsigned NumElems = VT.getVectorNumElements();
3898 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3899 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3900 EVT MaskEltVT = MaskVT.getVectorElementType();
3902 switch (VT.getSimpleVT().SimpleTy) {
3903 default: assert(false && "Unexpected!");
3904 case MVT::v4f32: NewVT = MVT::v2f64; break;
3905 case MVT::v4i32: NewVT = MVT::v2i64; break;
3906 case MVT::v8i16: NewVT = MVT::v4i32; break;
3907 case MVT::v16i8: NewVT = MVT::v4i32; break;
3910 if (NewWidth == 2) {
3916 int Scale = NumElems / NewWidth;
3917 SmallVector<int, 8> MaskVec;
3918 for (unsigned i = 0; i < NumElems; i += Scale) {
3920 for (int j = 0; j < Scale; ++j) {
3921 int EltIdx = SVOp->getMaskElt(i+j);
3925 StartIdx = EltIdx - (EltIdx % Scale);
3926 if (EltIdx != StartIdx + j)
3930 MaskVec.push_back(-1);
3932 MaskVec.push_back(StartIdx / Scale);
3935 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3936 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3937 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3940 /// getVZextMovL - Return a zero-extending vector move low node.
3942 static SDValue getVZextMovL(EVT VT, EVT OpVT,
3943 SDValue SrcOp, SelectionDAG &DAG,
3944 const X86Subtarget *Subtarget, DebugLoc dl) {
3945 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3946 LoadSDNode *LD = NULL;
3947 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3948 LD = dyn_cast<LoadSDNode>(SrcOp);
3950 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3952 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3953 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
3954 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3955 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3956 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
3958 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3959 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3960 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3961 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3969 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3970 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3971 DAG.getNode(ISD::BIT_CONVERT, dl,
3975 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3978 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3979 SDValue V1 = SVOp->getOperand(0);
3980 SDValue V2 = SVOp->getOperand(1);
3981 DebugLoc dl = SVOp->getDebugLoc();
3982 EVT VT = SVOp->getValueType(0);
3984 SmallVector<std::pair<int, int>, 8> Locs;
3986 SmallVector<int, 8> Mask1(4U, -1);
3987 SmallVector<int, 8> PermMask;
3988 SVOp->getMask(PermMask);
3992 for (unsigned i = 0; i != 4; ++i) {
3993 int Idx = PermMask[i];
3995 Locs[i] = std::make_pair(-1, -1);
3997 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3999 Locs[i] = std::make_pair(0, NumLo);
4003 Locs[i] = std::make_pair(1, NumHi);
4005 Mask1[2+NumHi] = Idx;
4011 if (NumLo <= 2 && NumHi <= 2) {
4012 // If no more than two elements come from either vector. This can be
4013 // implemented with two shuffles. First shuffle gather the elements.
4014 // The second shuffle, which takes the first shuffle as both of its
4015 // vector operands, put the elements into the right order.
4016 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4018 SmallVector<int, 8> Mask2(4U, -1);
4020 for (unsigned i = 0; i != 4; ++i) {
4021 if (Locs[i].first == -1)
4024 unsigned Idx = (i < 2) ? 0 : 4;
4025 Idx += Locs[i].first * 2 + Locs[i].second;
4030 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4031 } else if (NumLo == 3 || NumHi == 3) {
4032 // Otherwise, we must have three elements from one vector, call it X, and
4033 // one element from the other, call it Y. First, use a shufps to build an
4034 // intermediate vector with the one element from Y and the element from X
4035 // that will be in the same half in the final destination (the indexes don't
4036 // matter). Then, use a shufps to build the final vector, taking the half
4037 // containing the element from Y from the intermediate, and the other half
4040 // Normalize it so the 3 elements come from V1.
4041 CommuteVectorShuffleMask(PermMask, VT);
4045 // Find the element from V2.
4047 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4048 int Val = PermMask[HiIndex];
4055 Mask1[0] = PermMask[HiIndex];
4057 Mask1[2] = PermMask[HiIndex^1];
4059 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4062 Mask1[0] = PermMask[0];
4063 Mask1[1] = PermMask[1];
4064 Mask1[2] = HiIndex & 1 ? 6 : 4;
4065 Mask1[3] = HiIndex & 1 ? 4 : 6;
4066 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4068 Mask1[0] = HiIndex & 1 ? 2 : 0;
4069 Mask1[1] = HiIndex & 1 ? 0 : 2;
4070 Mask1[2] = PermMask[2];
4071 Mask1[3] = PermMask[3];
4076 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4080 // Break it into (shuffle shuffle_hi, shuffle_lo).
4082 SmallVector<int,8> LoMask(4U, -1);
4083 SmallVector<int,8> HiMask(4U, -1);
4085 SmallVector<int,8> *MaskPtr = &LoMask;
4086 unsigned MaskIdx = 0;
4089 for (unsigned i = 0; i != 4; ++i) {
4096 int Idx = PermMask[i];
4098 Locs[i] = std::make_pair(-1, -1);
4099 } else if (Idx < 4) {
4100 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4101 (*MaskPtr)[LoIdx] = Idx;
4104 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4105 (*MaskPtr)[HiIdx] = Idx;
4110 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4111 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4112 SmallVector<int, 8> MaskOps;
4113 for (unsigned i = 0; i != 4; ++i) {
4114 if (Locs[i].first == -1) {
4115 MaskOps.push_back(-1);
4117 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4118 MaskOps.push_back(Idx);
4121 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4125 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4126 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4127 SDValue V1 = Op.getOperand(0);
4128 SDValue V2 = Op.getOperand(1);
4129 EVT VT = Op.getValueType();
4130 DebugLoc dl = Op.getDebugLoc();
4131 unsigned NumElems = VT.getVectorNumElements();
4132 bool isMMX = VT.getSizeInBits() == 64;
4133 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4134 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4135 bool V1IsSplat = false;
4136 bool V2IsSplat = false;
4138 if (isZeroShuffle(SVOp))
4139 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4141 // Promote splats to v4f32.
4142 if (SVOp->isSplat()) {
4143 if (isMMX || NumElems < 4)
4145 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4148 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4150 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4151 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4152 if (NewOp.getNode())
4153 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4154 LowerVECTOR_SHUFFLE(NewOp, DAG));
4155 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4156 // FIXME: Figure out a cleaner way to do this.
4157 // Try to make use of movq to zero out the top part.
4158 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4159 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4160 if (NewOp.getNode()) {
4161 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4162 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4163 DAG, Subtarget, dl);
4165 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4166 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4167 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4168 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4169 DAG, Subtarget, dl);
4173 if (X86::isPSHUFDMask(SVOp))
4176 // Check if this can be converted into a logical shift.
4177 bool isLeft = false;
4180 bool isShift = getSubtarget()->hasSSE2() &&
4181 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4182 if (isShift && ShVal.hasOneUse()) {
4183 // If the shifted value has multiple uses, it may be cheaper to use
4184 // v_set0 + movlhps or movhlps, etc.
4185 EVT EVT = VT.getVectorElementType();
4186 ShAmt *= EVT.getSizeInBits();
4187 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4190 if (X86::isMOVLMask(SVOp)) {
4193 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4194 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4199 // FIXME: fold these into legal mask.
4200 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4201 X86::isMOVSLDUPMask(SVOp) ||
4202 X86::isMOVHLPSMask(SVOp) ||
4203 X86::isMOVHPMask(SVOp) ||
4204 X86::isMOVLPMask(SVOp)))
4207 if (ShouldXformToMOVHLPS(SVOp) ||
4208 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4209 return CommuteVectorShuffle(SVOp, DAG);
4212 // No better options. Use a vshl / vsrl.
4213 EVT EVT = VT.getVectorElementType();
4214 ShAmt *= EVT.getSizeInBits();
4215 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4218 bool Commuted = false;
4219 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4220 // 1,1,1,1 -> v8i16 though.
4221 V1IsSplat = isSplatVector(V1.getNode());
4222 V2IsSplat = isSplatVector(V2.getNode());
4224 // Canonicalize the splat or undef, if present, to be on the RHS.
4225 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4226 Op = CommuteVectorShuffle(SVOp, DAG);
4227 SVOp = cast<ShuffleVectorSDNode>(Op);
4228 V1 = SVOp->getOperand(0);
4229 V2 = SVOp->getOperand(1);
4230 std::swap(V1IsSplat, V2IsSplat);
4231 std::swap(V1IsUndef, V2IsUndef);
4235 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4236 // Shuffling low element of v1 into undef, just return v1.
4239 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4240 // the instruction selector will not match, so get a canonical MOVL with
4241 // swapped operands to undo the commute.
4242 return getMOVL(DAG, dl, VT, V2, V1);
4245 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4246 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4247 X86::isUNPCKLMask(SVOp) ||
4248 X86::isUNPCKHMask(SVOp))
4252 // Normalize mask so all entries that point to V2 points to its first
4253 // element then try to match unpck{h|l} again. If match, return a
4254 // new vector_shuffle with the corrected mask.
4255 SDValue NewMask = NormalizeMask(SVOp, DAG);
4256 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4257 if (NSVOp != SVOp) {
4258 if (X86::isUNPCKLMask(NSVOp, true)) {
4260 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4267 // Commute is back and try unpck* again.
4268 // FIXME: this seems wrong.
4269 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4270 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4271 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4272 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4273 X86::isUNPCKLMask(NewSVOp) ||
4274 X86::isUNPCKHMask(NewSVOp))
4278 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4280 // Normalize the node to match x86 shuffle ops if needed
4281 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4282 return CommuteVectorShuffle(SVOp, DAG);
4284 // Check for legal shuffle and return?
4285 SmallVector<int, 16> PermMask;
4286 SVOp->getMask(PermMask);
4287 if (isShuffleMaskLegal(PermMask, VT))
4290 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4291 if (VT == MVT::v8i16) {
4292 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4293 if (NewOp.getNode())
4297 if (VT == MVT::v16i8) {
4298 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4299 if (NewOp.getNode())
4303 // Handle all 4 wide cases with a number of shuffles except for MMX.
4304 if (NumElems == 4 && !isMMX)
4305 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4311 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4312 SelectionDAG &DAG) {
4313 EVT VT = Op.getValueType();
4314 DebugLoc dl = Op.getDebugLoc();
4315 if (VT.getSizeInBits() == 8) {
4316 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4317 Op.getOperand(0), Op.getOperand(1));
4318 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4319 DAG.getValueType(VT));
4320 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4321 } else if (VT.getSizeInBits() == 16) {
4322 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4323 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4325 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4326 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4327 DAG.getNode(ISD::BIT_CONVERT, dl,
4331 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4332 Op.getOperand(0), Op.getOperand(1));
4333 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4334 DAG.getValueType(VT));
4335 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4336 } else if (VT == MVT::f32) {
4337 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4338 // the result back to FR32 register. It's only worth matching if the
4339 // result has a single use which is a store or a bitcast to i32. And in
4340 // the case of a store, it's not worth it if the index is a constant 0,
4341 // because a MOVSSmr can be used instead, which is smaller and faster.
4342 if (!Op.hasOneUse())
4344 SDNode *User = *Op.getNode()->use_begin();
4345 if ((User->getOpcode() != ISD::STORE ||
4346 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4347 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4348 (User->getOpcode() != ISD::BIT_CONVERT ||
4349 User->getValueType(0) != MVT::i32))
4351 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4352 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4355 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4356 } else if (VT == MVT::i32) {
4357 // ExtractPS works with constant index.
4358 if (isa<ConstantSDNode>(Op.getOperand(1)))
4366 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4367 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4370 if (Subtarget->hasSSE41()) {
4371 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4376 EVT VT = Op.getValueType();
4377 DebugLoc dl = Op.getDebugLoc();
4378 // TODO: handle v16i8.
4379 if (VT.getSizeInBits() == 16) {
4380 SDValue Vec = Op.getOperand(0);
4381 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4383 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4384 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4385 DAG.getNode(ISD::BIT_CONVERT, dl,
4388 // Transform it so it match pextrw which produces a 32-bit result.
4389 EVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4390 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4391 Op.getOperand(0), Op.getOperand(1));
4392 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4393 DAG.getValueType(VT));
4394 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4395 } else if (VT.getSizeInBits() == 32) {
4396 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4400 // SHUFPS the element to the lowest double word, then movss.
4401 int Mask[4] = { Idx, -1, -1, -1 };
4402 EVT VVT = Op.getOperand(0).getValueType();
4403 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4404 DAG.getUNDEF(VVT), Mask);
4405 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4406 DAG.getIntPtrConstant(0));
4407 } else if (VT.getSizeInBits() == 64) {
4408 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4409 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4410 // to match extract_elt for f64.
4411 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4415 // UNPCKHPD the element to the lowest double word, then movsd.
4416 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4417 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4418 int Mask[2] = { 1, -1 };
4419 EVT VVT = Op.getOperand(0).getValueType();
4420 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4421 DAG.getUNDEF(VVT), Mask);
4422 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4423 DAG.getIntPtrConstant(0));
4430 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4431 EVT VT = Op.getValueType();
4432 EVT EVT = VT.getVectorElementType();
4433 DebugLoc dl = Op.getDebugLoc();
4435 SDValue N0 = Op.getOperand(0);
4436 SDValue N1 = Op.getOperand(1);
4437 SDValue N2 = Op.getOperand(2);
4439 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4440 isa<ConstantSDNode>(N2)) {
4441 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4443 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4445 if (N1.getValueType() != MVT::i32)
4446 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4447 if (N2.getValueType() != MVT::i32)
4448 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4449 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4450 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4451 // Bits [7:6] of the constant are the source select. This will always be
4452 // zero here. The DAG Combiner may combine an extract_elt index into these
4453 // bits. For example (insert (extract, 3), 2) could be matched by putting
4454 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4455 // Bits [5:4] of the constant are the destination select. This is the
4456 // value of the incoming immediate.
4457 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4458 // combine either bitwise AND or insert of float 0.0 to set these bits.
4459 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4460 // Create this as a scalar to vector..
4461 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4462 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4463 } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4464 // PINSR* works with constant index.
4471 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4472 EVT VT = Op.getValueType();
4473 EVT EVT = VT.getVectorElementType();
4475 if (Subtarget->hasSSE41())
4476 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4481 DebugLoc dl = Op.getDebugLoc();
4482 SDValue N0 = Op.getOperand(0);
4483 SDValue N1 = Op.getOperand(1);
4484 SDValue N2 = Op.getOperand(2);
4486 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4487 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4488 // as its second argument.
4489 if (N1.getValueType() != MVT::i32)
4490 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4491 if (N2.getValueType() != MVT::i32)
4492 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4493 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4499 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4500 DebugLoc dl = Op.getDebugLoc();
4501 if (Op.getValueType() == MVT::v2f32)
4502 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4503 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4504 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4505 Op.getOperand(0))));
4507 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4508 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4510 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4511 EVT VT = MVT::v2i32;
4512 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4519 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4520 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4523 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4524 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4525 // one of the above mentioned nodes. It has to be wrapped because otherwise
4526 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4527 // be used to form addressing mode. These wrapped nodes will be selected
4530 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4531 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4533 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4535 unsigned char OpFlag = 0;
4536 unsigned WrapperKind = X86ISD::Wrapper;
4537 CodeModel::Model M = getTargetMachine().getCodeModel();
4539 if (Subtarget->isPICStyleRIPRel() &&
4540 (M == CodeModel::Small || M == CodeModel::Kernel))
4541 WrapperKind = X86ISD::WrapperRIP;
4542 else if (Subtarget->isPICStyleGOT())
4543 OpFlag = X86II::MO_GOTOFF;
4544 else if (Subtarget->isPICStyleStubPIC())
4545 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4547 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4549 CP->getOffset(), OpFlag);
4550 DebugLoc DL = CP->getDebugLoc();
4551 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4552 // With PIC, the address is actually $g + Offset.
4554 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4555 DAG.getNode(X86ISD::GlobalBaseReg,
4556 DebugLoc::getUnknownLoc(), getPointerTy()),
4563 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4564 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4566 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4568 unsigned char OpFlag = 0;
4569 unsigned WrapperKind = X86ISD::Wrapper;
4570 CodeModel::Model M = getTargetMachine().getCodeModel();
4572 if (Subtarget->isPICStyleRIPRel() &&
4573 (M == CodeModel::Small || M == CodeModel::Kernel))
4574 WrapperKind = X86ISD::WrapperRIP;
4575 else if (Subtarget->isPICStyleGOT())
4576 OpFlag = X86II::MO_GOTOFF;
4577 else if (Subtarget->isPICStyleStubPIC())
4578 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4580 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4582 DebugLoc DL = JT->getDebugLoc();
4583 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4585 // With PIC, the address is actually $g + Offset.
4587 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4588 DAG.getNode(X86ISD::GlobalBaseReg,
4589 DebugLoc::getUnknownLoc(), getPointerTy()),
4597 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4598 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4600 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4602 unsigned char OpFlag = 0;
4603 unsigned WrapperKind = X86ISD::Wrapper;
4604 CodeModel::Model M = getTargetMachine().getCodeModel();
4606 if (Subtarget->isPICStyleRIPRel() &&
4607 (M == CodeModel::Small || M == CodeModel::Kernel))
4608 WrapperKind = X86ISD::WrapperRIP;
4609 else if (Subtarget->isPICStyleGOT())
4610 OpFlag = X86II::MO_GOTOFF;
4611 else if (Subtarget->isPICStyleStubPIC())
4612 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4614 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4616 DebugLoc DL = Op.getDebugLoc();
4617 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4620 // With PIC, the address is actually $g + Offset.
4621 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4622 !Subtarget->is64Bit()) {
4623 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4624 DAG.getNode(X86ISD::GlobalBaseReg,
4625 DebugLoc::getUnknownLoc(),
4634 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4636 SelectionDAG &DAG) const {
4637 // Create the TargetGlobalAddress node, folding in the constant
4638 // offset if it is legal.
4639 unsigned char OpFlags =
4640 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4641 CodeModel::Model M = getTargetMachine().getCodeModel();
4643 if (OpFlags == X86II::MO_NO_FLAG &&
4644 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4645 // A direct static reference to a global.
4646 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4649 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4652 if (Subtarget->isPICStyleRIPRel() &&
4653 (M == CodeModel::Small || M == CodeModel::Kernel))
4654 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4656 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4658 // With PIC, the address is actually $g + Offset.
4659 if (isGlobalRelativeToPICBase(OpFlags)) {
4660 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4661 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4665 // For globals that require a load from a stub to get the address, emit the
4667 if (isGlobalStubReference(OpFlags))
4668 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4669 PseudoSourceValue::getGOT(), 0);
4671 // If there was a non-zero offset that we didn't fold, create an explicit
4674 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4675 DAG.getConstant(Offset, getPointerTy()));
4681 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4682 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4683 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4684 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4688 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4689 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4690 unsigned char OperandFlags) {
4691 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4692 DebugLoc dl = GA->getDebugLoc();
4693 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4694 GA->getValueType(0),
4698 SDValue Ops[] = { Chain, TGA, *InFlag };
4699 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4701 SDValue Ops[] = { Chain, TGA };
4702 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4704 SDValue Flag = Chain.getValue(1);
4705 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4708 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4710 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4713 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4714 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4715 DAG.getNode(X86ISD::GlobalBaseReg,
4716 DebugLoc::getUnknownLoc(),
4718 InFlag = Chain.getValue(1);
4720 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4723 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4725 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4727 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4728 X86::RAX, X86II::MO_TLSGD);
4731 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4732 // "local exec" model.
4733 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4734 const EVT PtrVT, TLSModel::Model model,
4736 DebugLoc dl = GA->getDebugLoc();
4737 // Get the Thread Pointer
4738 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4739 DebugLoc::getUnknownLoc(), PtrVT,
4740 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4743 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4746 unsigned char OperandFlags = 0;
4747 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4749 unsigned WrapperKind = X86ISD::Wrapper;
4750 if (model == TLSModel::LocalExec) {
4751 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4752 } else if (is64Bit) {
4753 assert(model == TLSModel::InitialExec);
4754 OperandFlags = X86II::MO_GOTTPOFF;
4755 WrapperKind = X86ISD::WrapperRIP;
4757 assert(model == TLSModel::InitialExec);
4758 OperandFlags = X86II::MO_INDNTPOFF;
4761 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4763 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4764 GA->getOffset(), OperandFlags);
4765 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4767 if (model == TLSModel::InitialExec)
4768 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4769 PseudoSourceValue::getGOT(), 0);
4771 // The address of the thread local variable is the add of the thread
4772 // pointer with the offset of the variable.
4773 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4777 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4778 // TODO: implement the "local dynamic" model
4779 // TODO: implement the "initial exec"model for pic executables
4780 assert(Subtarget->isTargetELF() &&
4781 "TLS not implemented for non-ELF targets");
4782 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4783 const GlobalValue *GV = GA->getGlobal();
4785 // If GV is an alias then use the aliasee for determining
4786 // thread-localness.
4787 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4788 GV = GA->resolveAliasedGlobal(false);
4790 TLSModel::Model model = getTLSModel(GV,
4791 getTargetMachine().getRelocationModel());
4794 case TLSModel::GeneralDynamic:
4795 case TLSModel::LocalDynamic: // not implemented
4796 if (Subtarget->is64Bit())
4797 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4798 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4800 case TLSModel::InitialExec:
4801 case TLSModel::LocalExec:
4802 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4803 Subtarget->is64Bit());
4806 llvm_unreachable("Unreachable");
4811 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4812 /// take a 2 x i32 value to shift plus a shift amount.
4813 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4814 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4815 EVT VT = Op.getValueType();
4816 unsigned VTBits = VT.getSizeInBits();
4817 DebugLoc dl = Op.getDebugLoc();
4818 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4819 SDValue ShOpLo = Op.getOperand(0);
4820 SDValue ShOpHi = Op.getOperand(1);
4821 SDValue ShAmt = Op.getOperand(2);
4822 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4823 DAG.getConstant(VTBits - 1, MVT::i8))
4824 : DAG.getConstant(0, VT);
4827 if (Op.getOpcode() == ISD::SHL_PARTS) {
4828 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4829 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4831 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4832 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4835 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4836 DAG.getConstant(VTBits, MVT::i8));
4837 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4838 AndNode, DAG.getConstant(0, MVT::i8));
4841 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4842 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4843 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4845 if (Op.getOpcode() == ISD::SHL_PARTS) {
4846 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4847 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4849 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4850 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4853 SDValue Ops[2] = { Lo, Hi };
4854 return DAG.getMergeValues(Ops, 2, dl);
4857 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4858 EVT SrcVT = Op.getOperand(0).getValueType();
4860 if (SrcVT.isVector()) {
4861 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4867 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4868 "Unknown SINT_TO_FP to lower!");
4870 // These are really Legal; return the operand so the caller accepts it as
4872 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4874 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4875 Subtarget->is64Bit()) {
4879 DebugLoc dl = Op.getDebugLoc();
4880 unsigned Size = SrcVT.getSizeInBits()/8;
4881 MachineFunction &MF = DAG.getMachineFunction();
4882 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4883 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4884 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4886 PseudoSourceValue::getFixedStack(SSFI), 0);
4887 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4890 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
4892 SelectionDAG &DAG) {
4894 DebugLoc dl = Op.getDebugLoc();
4896 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4898 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4900 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4901 SmallVector<SDValue, 8> Ops;
4902 Ops.push_back(Chain);
4903 Ops.push_back(StackSlot);
4904 Ops.push_back(DAG.getValueType(SrcVT));
4905 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4906 Tys, &Ops[0], Ops.size());
4909 Chain = Result.getValue(1);
4910 SDValue InFlag = Result.getValue(2);
4912 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4913 // shouldn't be necessary except that RFP cannot be live across
4914 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4915 MachineFunction &MF = DAG.getMachineFunction();
4916 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4917 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4918 Tys = DAG.getVTList(MVT::Other);
4919 SmallVector<SDValue, 8> Ops;
4920 Ops.push_back(Chain);
4921 Ops.push_back(Result);
4922 Ops.push_back(StackSlot);
4923 Ops.push_back(DAG.getValueType(Op.getValueType()));
4924 Ops.push_back(InFlag);
4925 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4926 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4927 PseudoSourceValue::getFixedStack(SSFI), 0);
4933 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4934 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4935 // This algorithm is not obvious. Here it is in C code, more or less:
4937 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4938 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4939 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4941 // Copy ints to xmm registers.
4942 __m128i xh = _mm_cvtsi32_si128( hi );
4943 __m128i xl = _mm_cvtsi32_si128( lo );
4945 // Combine into low half of a single xmm register.
4946 __m128i x = _mm_unpacklo_epi32( xh, xl );
4950 // Merge in appropriate exponents to give the integer bits the right
4952 x = _mm_unpacklo_epi32( x, exp );
4954 // Subtract away the biases to deal with the IEEE-754 double precision
4956 d = _mm_sub_pd( (__m128d) x, bias );
4958 // All conversions up to here are exact. The correctly rounded result is
4959 // calculated using the current rounding mode using the following
4961 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4962 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4963 // store doesn't really need to be here (except
4964 // maybe to zero the other double)
4969 DebugLoc dl = Op.getDebugLoc();
4970 LLVMContext *Context = DAG.getContext();
4972 // Build some magic constants.
4973 std::vector<Constant*> CV0;
4974 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
4975 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
4976 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4977 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4978 Constant *C0 = ConstantVector::get(CV0);
4979 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4981 std::vector<Constant*> CV1;
4983 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
4985 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
4986 Constant *C1 = ConstantVector::get(CV1);
4987 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4989 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4990 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4992 DAG.getIntPtrConstant(1)));
4993 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4994 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4996 DAG.getIntPtrConstant(0)));
4997 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4998 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4999 PseudoSourceValue::getConstantPool(), 0,
5001 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5002 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5003 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5004 PseudoSourceValue::getConstantPool(), 0,
5006 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5008 // Add the halves; easiest way is to swap them into another reg first.
5009 int ShufMask[2] = { 1, -1 };
5010 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5011 DAG.getUNDEF(MVT::v2f64), ShufMask);
5012 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5013 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5014 DAG.getIntPtrConstant(0));
5017 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5018 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5019 DebugLoc dl = Op.getDebugLoc();
5020 // FP constant to bias correct the final result.
5021 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5024 // Load the 32-bit value into an XMM register.
5025 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5026 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5028 DAG.getIntPtrConstant(0)));
5030 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5031 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5032 DAG.getIntPtrConstant(0));
5034 // Or the load with the bias.
5035 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5036 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5037 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5039 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5040 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5041 MVT::v2f64, Bias)));
5042 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5043 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5044 DAG.getIntPtrConstant(0));
5046 // Subtract the bias.
5047 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5049 // Handle final rounding.
5050 EVT DestVT = Op.getValueType();
5052 if (DestVT.bitsLT(MVT::f64)) {
5053 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5054 DAG.getIntPtrConstant(0));
5055 } else if (DestVT.bitsGT(MVT::f64)) {
5056 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5059 // Handle final rounding.
5063 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5064 SDValue N0 = Op.getOperand(0);
5065 DebugLoc dl = Op.getDebugLoc();
5067 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5068 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5069 // the optimization here.
5070 if (DAG.SignBitIsZero(N0))
5071 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5073 EVT SrcVT = N0.getValueType();
5074 if (SrcVT == MVT::i64) {
5075 // We only handle SSE2 f64 target here; caller can expand the rest.
5076 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5079 return LowerUINT_TO_FP_i64(Op, DAG);
5080 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5081 return LowerUINT_TO_FP_i32(Op, DAG);
5084 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5086 // Make a 64-bit buffer, and use it to build an FILD.
5087 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5088 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5089 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5090 getPointerTy(), StackSlot, WordOff);
5091 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5092 StackSlot, NULL, 0);
5093 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5094 OffsetSlot, NULL, 0);
5095 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5098 std::pair<SDValue,SDValue> X86TargetLowering::
5099 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5100 DebugLoc dl = Op.getDebugLoc();
5102 EVT DstTy = Op.getValueType();
5105 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5109 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5110 DstTy.getSimpleVT() >= MVT::i16 &&
5111 "Unknown FP_TO_SINT to lower!");
5113 // These are really Legal.
5114 if (DstTy == MVT::i32 &&
5115 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5116 return std::make_pair(SDValue(), SDValue());
5117 if (Subtarget->is64Bit() &&
5118 DstTy == MVT::i64 &&
5119 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5120 return std::make_pair(SDValue(), SDValue());
5122 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5124 MachineFunction &MF = DAG.getMachineFunction();
5125 unsigned MemSize = DstTy.getSizeInBits()/8;
5126 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5127 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5130 switch (DstTy.getSimpleVT().SimpleTy) {
5131 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5132 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5133 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5134 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5137 SDValue Chain = DAG.getEntryNode();
5138 SDValue Value = Op.getOperand(0);
5139 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5140 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5141 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5142 PseudoSourceValue::getFixedStack(SSFI), 0);
5143 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5145 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5147 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5148 Chain = Value.getValue(1);
5149 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5150 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5153 // Build the FP_TO_INT*_IN_MEM
5154 SDValue Ops[] = { Chain, Value, StackSlot };
5155 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5157 return std::make_pair(FIST, StackSlot);
5160 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5161 if (Op.getValueType().isVector()) {
5162 if (Op.getValueType() == MVT::v2i32 &&
5163 Op.getOperand(0).getValueType() == MVT::v2f64) {
5169 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5170 SDValue FIST = Vals.first, StackSlot = Vals.second;
5171 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5172 if (FIST.getNode() == 0) return Op;
5175 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5176 FIST, StackSlot, NULL, 0);
5179 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5180 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5181 SDValue FIST = Vals.first, StackSlot = Vals.second;
5182 assert(FIST.getNode() && "Unexpected failure");
5185 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5186 FIST, StackSlot, NULL, 0);
5189 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5190 LLVMContext *Context = DAG.getContext();
5191 DebugLoc dl = Op.getDebugLoc();
5192 EVT VT = Op.getValueType();
5195 EltVT = VT.getVectorElementType();
5196 std::vector<Constant*> CV;
5197 if (EltVT == MVT::f64) {
5198 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5202 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5208 Constant *C = ConstantVector::get(CV);
5209 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5210 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5211 PseudoSourceValue::getConstantPool(), 0,
5213 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5216 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5217 LLVMContext *Context = DAG.getContext();
5218 DebugLoc dl = Op.getDebugLoc();
5219 EVT VT = Op.getValueType();
5221 unsigned EltNum = 1;
5222 if (VT.isVector()) {
5223 EltVT = VT.getVectorElementType();
5224 EltNum = VT.getVectorNumElements();
5226 std::vector<Constant*> CV;
5227 if (EltVT == MVT::f64) {
5228 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5232 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5238 Constant *C = ConstantVector::get(CV);
5239 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5240 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5241 PseudoSourceValue::getConstantPool(), 0,
5243 if (VT.isVector()) {
5244 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5245 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5246 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5248 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5250 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5254 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5255 LLVMContext *Context = DAG.getContext();
5256 SDValue Op0 = Op.getOperand(0);
5257 SDValue Op1 = Op.getOperand(1);
5258 DebugLoc dl = Op.getDebugLoc();
5259 EVT VT = Op.getValueType();
5260 EVT SrcVT = Op1.getValueType();
5262 // If second operand is smaller, extend it first.
5263 if (SrcVT.bitsLT(VT)) {
5264 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5267 // And if it is bigger, shrink it first.
5268 if (SrcVT.bitsGT(VT)) {
5269 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5273 // At this point the operands and the result should have the same
5274 // type, and that won't be f80 since that is not custom lowered.
5276 // First get the sign bit of second operand.
5277 std::vector<Constant*> CV;
5278 if (SrcVT == MVT::f64) {
5279 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5280 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5282 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5283 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5284 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5285 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5287 Constant *C = ConstantVector::get(CV);
5288 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5289 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5290 PseudoSourceValue::getConstantPool(), 0,
5292 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5294 // Shift sign bit right or left if the two operands have different types.
5295 if (SrcVT.bitsGT(VT)) {
5296 // Op0 is MVT::f32, Op1 is MVT::f64.
5297 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5298 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5299 DAG.getConstant(32, MVT::i32));
5300 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5301 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5302 DAG.getIntPtrConstant(0));
5305 // Clear first operand sign bit.
5307 if (VT == MVT::f64) {
5308 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5309 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5311 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5312 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5313 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5314 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5316 C = ConstantVector::get(CV);
5317 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5318 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5319 PseudoSourceValue::getConstantPool(), 0,
5321 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5323 // Or the value with the sign bit.
5324 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5327 /// Emit nodes that will be selected as "test Op0,Op0", or something
5329 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5330 SelectionDAG &DAG) {
5331 DebugLoc dl = Op.getDebugLoc();
5333 // CF and OF aren't always set the way we want. Determine which
5334 // of these we need.
5335 bool NeedCF = false;
5336 bool NeedOF = false;
5338 case X86::COND_A: case X86::COND_AE:
5339 case X86::COND_B: case X86::COND_BE:
5342 case X86::COND_G: case X86::COND_GE:
5343 case X86::COND_L: case X86::COND_LE:
5344 case X86::COND_O: case X86::COND_NO:
5350 // See if we can use the EFLAGS value from the operand instead of
5351 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5352 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5353 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5354 unsigned Opcode = 0;
5355 unsigned NumOperands = 0;
5356 switch (Op.getNode()->getOpcode()) {
5358 // Due to an isel shortcoming, be conservative if this add is likely to
5359 // be selected as part of a load-modify-store instruction. When the root
5360 // node in a match is a store, isel doesn't know how to remap non-chain
5361 // non-flag uses of other nodes in the match, such as the ADD in this
5362 // case. This leads to the ADD being left around and reselected, with
5363 // the result being two adds in the output.
5364 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5365 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5366 if (UI->getOpcode() == ISD::STORE)
5368 if (ConstantSDNode *C =
5369 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5370 // An add of one will be selected as an INC.
5371 if (C->getAPIntValue() == 1) {
5372 Opcode = X86ISD::INC;
5376 // An add of negative one (subtract of one) will be selected as a DEC.
5377 if (C->getAPIntValue().isAllOnesValue()) {
5378 Opcode = X86ISD::DEC;
5383 // Otherwise use a regular EFLAGS-setting add.
5384 Opcode = X86ISD::ADD;
5388 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5389 // likely to be selected as part of a load-modify-store instruction.
5390 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5391 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5392 if (UI->getOpcode() == ISD::STORE)
5394 // Otherwise use a regular EFLAGS-setting sub.
5395 Opcode = X86ISD::SUB;
5402 return SDValue(Op.getNode(), 1);
5408 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5409 SmallVector<SDValue, 4> Ops;
5410 for (unsigned i = 0; i != NumOperands; ++i)
5411 Ops.push_back(Op.getOperand(i));
5412 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5413 DAG.ReplaceAllUsesWith(Op, New);
5414 return SDValue(New.getNode(), 1);
5418 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5419 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5420 DAG.getConstant(0, Op.getValueType()));
5423 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5425 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5426 SelectionDAG &DAG) {
5427 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5428 if (C->getAPIntValue() == 0)
5429 return EmitTest(Op0, X86CC, DAG);
5431 DebugLoc dl = Op0.getDebugLoc();
5432 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5435 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5436 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5437 SDValue Op0 = Op.getOperand(0);
5438 SDValue Op1 = Op.getOperand(1);
5439 DebugLoc dl = Op.getDebugLoc();
5440 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5442 // Lower (X & (1 << N)) == 0 to BT(X, N).
5443 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5444 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5445 if (Op0.getOpcode() == ISD::AND &&
5447 Op1.getOpcode() == ISD::Constant &&
5448 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5449 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5451 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5452 if (ConstantSDNode *Op010C =
5453 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5454 if (Op010C->getZExtValue() == 1) {
5455 LHS = Op0.getOperand(0);
5456 RHS = Op0.getOperand(1).getOperand(1);
5458 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5459 if (ConstantSDNode *Op000C =
5460 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5461 if (Op000C->getZExtValue() == 1) {
5462 LHS = Op0.getOperand(1);
5463 RHS = Op0.getOperand(0).getOperand(1);
5465 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5466 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5467 SDValue AndLHS = Op0.getOperand(0);
5468 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5469 LHS = AndLHS.getOperand(0);
5470 RHS = AndLHS.getOperand(1);
5474 if (LHS.getNode()) {
5475 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5476 // instruction. Since the shift amount is in-range-or-undefined, we know
5477 // that doing a bittest on the i16 value is ok. We extend to i32 because
5478 // the encoding for the i16 version is larger than the i32 version.
5479 if (LHS.getValueType() == MVT::i8)
5480 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5482 // If the operand types disagree, extend the shift amount to match. Since
5483 // BT ignores high bits (like shifts) we can use anyextend.
5484 if (LHS.getValueType() != RHS.getValueType())
5485 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5487 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5488 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5489 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5490 DAG.getConstant(Cond, MVT::i8), BT);
5494 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5495 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5497 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5498 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5499 DAG.getConstant(X86CC, MVT::i8), Cond);
5502 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5504 SDValue Op0 = Op.getOperand(0);
5505 SDValue Op1 = Op.getOperand(1);
5506 SDValue CC = Op.getOperand(2);
5507 EVT VT = Op.getValueType();
5508 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5509 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5510 DebugLoc dl = Op.getDebugLoc();
5514 EVT VT0 = Op0.getValueType();
5515 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5516 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5519 switch (SetCCOpcode) {
5522 case ISD::SETEQ: SSECC = 0; break;
5524 case ISD::SETGT: Swap = true; // Fallthrough
5526 case ISD::SETOLT: SSECC = 1; break;
5528 case ISD::SETGE: Swap = true; // Fallthrough
5530 case ISD::SETOLE: SSECC = 2; break;
5531 case ISD::SETUO: SSECC = 3; break;
5533 case ISD::SETNE: SSECC = 4; break;
5534 case ISD::SETULE: Swap = true;
5535 case ISD::SETUGE: SSECC = 5; break;
5536 case ISD::SETULT: Swap = true;
5537 case ISD::SETUGT: SSECC = 6; break;
5538 case ISD::SETO: SSECC = 7; break;
5541 std::swap(Op0, Op1);
5543 // In the two special cases we can't handle, emit two comparisons.
5545 if (SetCCOpcode == ISD::SETUEQ) {
5547 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5548 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5549 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5551 else if (SetCCOpcode == ISD::SETONE) {
5553 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5554 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5555 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5557 llvm_unreachable("Illegal FP comparison");
5559 // Handle all other FP comparisons here.
5560 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5563 // We are handling one of the integer comparisons here. Since SSE only has
5564 // GT and EQ comparisons for integer, swapping operands and multiple
5565 // operations may be required for some comparisons.
5566 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5567 bool Swap = false, Invert = false, FlipSigns = false;
5569 switch (VT.getSimpleVT().SimpleTy) {
5572 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5574 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5576 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5577 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5580 switch (SetCCOpcode) {
5582 case ISD::SETNE: Invert = true;
5583 case ISD::SETEQ: Opc = EQOpc; break;
5584 case ISD::SETLT: Swap = true;
5585 case ISD::SETGT: Opc = GTOpc; break;
5586 case ISD::SETGE: Swap = true;
5587 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5588 case ISD::SETULT: Swap = true;
5589 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5590 case ISD::SETUGE: Swap = true;
5591 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5594 std::swap(Op0, Op1);
5596 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5597 // bits of the inputs before performing those operations.
5599 EVT EltVT = VT.getVectorElementType();
5600 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5602 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5603 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5605 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5606 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5609 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5611 // If the logical-not of the result is required, perform that now.
5613 Result = DAG.getNOT(dl, Result, VT);
5618 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5619 static bool isX86LogicalCmp(SDValue Op) {
5620 unsigned Opc = Op.getNode()->getOpcode();
5621 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5623 if (Op.getResNo() == 1 &&
5624 (Opc == X86ISD::ADD ||
5625 Opc == X86ISD::SUB ||
5626 Opc == X86ISD::SMUL ||
5627 Opc == X86ISD::UMUL ||
5628 Opc == X86ISD::INC ||
5629 Opc == X86ISD::DEC))
5635 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5636 bool addTest = true;
5637 SDValue Cond = Op.getOperand(0);
5638 DebugLoc dl = Op.getDebugLoc();
5641 if (Cond.getOpcode() == ISD::SETCC)
5642 Cond = LowerSETCC(Cond, DAG);
5644 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5645 // setting operand in place of the X86ISD::SETCC.
5646 if (Cond.getOpcode() == X86ISD::SETCC) {
5647 CC = Cond.getOperand(0);
5649 SDValue Cmp = Cond.getOperand(1);
5650 unsigned Opc = Cmp.getOpcode();
5651 EVT VT = Op.getValueType();
5653 bool IllegalFPCMov = false;
5654 if (VT.isFloatingPoint() && !VT.isVector() &&
5655 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5656 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5658 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5659 Opc == X86ISD::BT) { // FIXME
5666 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5667 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5670 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5671 SmallVector<SDValue, 4> Ops;
5672 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5673 // condition is true.
5674 Ops.push_back(Op.getOperand(2));
5675 Ops.push_back(Op.getOperand(1));
5677 Ops.push_back(Cond);
5678 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5681 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5682 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5683 // from the AND / OR.
5684 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5685 Opc = Op.getOpcode();
5686 if (Opc != ISD::OR && Opc != ISD::AND)
5688 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5689 Op.getOperand(0).hasOneUse() &&
5690 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5691 Op.getOperand(1).hasOneUse());
5694 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5695 // 1 and that the SETCC node has a single use.
5696 static bool isXor1OfSetCC(SDValue Op) {
5697 if (Op.getOpcode() != ISD::XOR)
5699 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5700 if (N1C && N1C->getAPIntValue() == 1) {
5701 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5702 Op.getOperand(0).hasOneUse();
5707 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5708 bool addTest = true;
5709 SDValue Chain = Op.getOperand(0);
5710 SDValue Cond = Op.getOperand(1);
5711 SDValue Dest = Op.getOperand(2);
5712 DebugLoc dl = Op.getDebugLoc();
5715 if (Cond.getOpcode() == ISD::SETCC)
5716 Cond = LowerSETCC(Cond, DAG);
5718 // FIXME: LowerXALUO doesn't handle these!!
5719 else if (Cond.getOpcode() == X86ISD::ADD ||
5720 Cond.getOpcode() == X86ISD::SUB ||
5721 Cond.getOpcode() == X86ISD::SMUL ||
5722 Cond.getOpcode() == X86ISD::UMUL)
5723 Cond = LowerXALUO(Cond, DAG);
5726 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5727 // setting operand in place of the X86ISD::SETCC.
5728 if (Cond.getOpcode() == X86ISD::SETCC) {
5729 CC = Cond.getOperand(0);
5731 SDValue Cmp = Cond.getOperand(1);
5732 unsigned Opc = Cmp.getOpcode();
5733 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5734 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5738 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5742 // These can only come from an arithmetic instruction with overflow,
5743 // e.g. SADDO, UADDO.
5744 Cond = Cond.getNode()->getOperand(1);
5751 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5752 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5753 if (CondOpc == ISD::OR) {
5754 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5755 // two branches instead of an explicit OR instruction with a
5757 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5758 isX86LogicalCmp(Cmp)) {
5759 CC = Cond.getOperand(0).getOperand(0);
5760 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5761 Chain, Dest, CC, Cmp);
5762 CC = Cond.getOperand(1).getOperand(0);
5766 } else { // ISD::AND
5767 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5768 // two branches instead of an explicit AND instruction with a
5769 // separate test. However, we only do this if this block doesn't
5770 // have a fall-through edge, because this requires an explicit
5771 // jmp when the condition is false.
5772 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5773 isX86LogicalCmp(Cmp) &&
5774 Op.getNode()->hasOneUse()) {
5775 X86::CondCode CCode =
5776 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5777 CCode = X86::GetOppositeBranchCondition(CCode);
5778 CC = DAG.getConstant(CCode, MVT::i8);
5779 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5780 // Look for an unconditional branch following this conditional branch.
5781 // We need this because we need to reverse the successors in order
5782 // to implement FCMP_OEQ.
5783 if (User.getOpcode() == ISD::BR) {
5784 SDValue FalseBB = User.getOperand(1);
5786 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5787 assert(NewBR == User);
5790 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5791 Chain, Dest, CC, Cmp);
5792 X86::CondCode CCode =
5793 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5794 CCode = X86::GetOppositeBranchCondition(CCode);
5795 CC = DAG.getConstant(CCode, MVT::i8);
5801 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5802 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5803 // It should be transformed during dag combiner except when the condition
5804 // is set by a arithmetics with overflow node.
5805 X86::CondCode CCode =
5806 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5807 CCode = X86::GetOppositeBranchCondition(CCode);
5808 CC = DAG.getConstant(CCode, MVT::i8);
5809 Cond = Cond.getOperand(0).getOperand(1);
5815 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5816 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5818 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5819 Chain, Dest, CC, Cond);
5823 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5824 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5825 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5826 // that the guard pages used by the OS virtual memory manager are allocated in
5827 // correct sequence.
5829 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5830 SelectionDAG &DAG) {
5831 assert(Subtarget->isTargetCygMing() &&
5832 "This should be used only on Cygwin/Mingw targets");
5833 DebugLoc dl = Op.getDebugLoc();
5836 SDValue Chain = Op.getOperand(0);
5837 SDValue Size = Op.getOperand(1);
5838 // FIXME: Ensure alignment here
5842 EVT IntPtr = getPointerTy();
5843 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5845 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5847 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5848 Flag = Chain.getValue(1);
5850 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5851 SDValue Ops[] = { Chain,
5852 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5853 DAG.getRegister(X86::EAX, IntPtr),
5854 DAG.getRegister(X86StackPtr, SPTy),
5856 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5857 Flag = Chain.getValue(1);
5859 Chain = DAG.getCALLSEQ_END(Chain,
5860 DAG.getIntPtrConstant(0, true),
5861 DAG.getIntPtrConstant(0, true),
5864 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5866 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5867 return DAG.getMergeValues(Ops1, 2, dl);
5871 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5873 SDValue Dst, SDValue Src,
5874 SDValue Size, unsigned Align,
5876 uint64_t DstSVOff) {
5877 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5879 // If not DWORD aligned or size is more than the threshold, call the library.
5880 // The libc version is likely to be faster for these cases. It can use the
5881 // address value and run time information about the CPU.
5882 if ((Align & 3) != 0 ||
5884 ConstantSize->getZExtValue() >
5885 getSubtarget()->getMaxInlineSizeThreshold()) {
5886 SDValue InFlag(0, 0);
5888 // Check to see if there is a specialized entry-point for memory zeroing.
5889 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5891 if (const char *bzeroEntry = V &&
5892 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5893 EVT IntPtr = getPointerTy();
5894 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
5895 TargetLowering::ArgListTy Args;
5896 TargetLowering::ArgListEntry Entry;
5898 Entry.Ty = IntPtrTy;
5899 Args.push_back(Entry);
5901 Args.push_back(Entry);
5902 std::pair<SDValue,SDValue> CallResult =
5903 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
5904 false, false, false, false,
5905 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
5906 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5907 return CallResult.second;
5910 // Otherwise have the target-independent code call memset.
5914 uint64_t SizeVal = ConstantSize->getZExtValue();
5915 SDValue InFlag(0, 0);
5918 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5919 unsigned BytesLeft = 0;
5920 bool TwoRepStos = false;
5923 uint64_t Val = ValC->getZExtValue() & 255;
5925 // If the value is a constant, then we can potentially use larger sets.
5926 switch (Align & 3) {
5927 case 2: // WORD aligned
5930 Val = (Val << 8) | Val;
5932 case 0: // DWORD aligned
5935 Val = (Val << 8) | Val;
5936 Val = (Val << 16) | Val;
5937 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5940 Val = (Val << 32) | Val;
5943 default: // Byte aligned
5946 Count = DAG.getIntPtrConstant(SizeVal);
5950 if (AVT.bitsGT(MVT::i8)) {
5951 unsigned UBytes = AVT.getSizeInBits() / 8;
5952 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5953 BytesLeft = SizeVal % UBytes;
5956 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5958 InFlag = Chain.getValue(1);
5961 Count = DAG.getIntPtrConstant(SizeVal);
5962 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5963 InFlag = Chain.getValue(1);
5966 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5969 InFlag = Chain.getValue(1);
5970 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5973 InFlag = Chain.getValue(1);
5975 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5976 SmallVector<SDValue, 8> Ops;
5977 Ops.push_back(Chain);
5978 Ops.push_back(DAG.getValueType(AVT));
5979 Ops.push_back(InFlag);
5980 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5983 InFlag = Chain.getValue(1);
5985 EVT CVT = Count.getValueType();
5986 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5987 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5988 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5991 InFlag = Chain.getValue(1);
5992 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5994 Ops.push_back(Chain);
5995 Ops.push_back(DAG.getValueType(MVT::i8));
5996 Ops.push_back(InFlag);
5997 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5998 } else if (BytesLeft) {
5999 // Handle the last 1 - 7 bytes.
6000 unsigned Offset = SizeVal - BytesLeft;
6001 EVT AddrVT = Dst.getValueType();
6002 EVT SizeVT = Size.getValueType();
6004 Chain = DAG.getMemset(Chain, dl,
6005 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6006 DAG.getConstant(Offset, AddrVT)),
6008 DAG.getConstant(BytesLeft, SizeVT),
6009 Align, DstSV, DstSVOff + Offset);
6012 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6017 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6018 SDValue Chain, SDValue Dst, SDValue Src,
6019 SDValue Size, unsigned Align,
6021 const Value *DstSV, uint64_t DstSVOff,
6022 const Value *SrcSV, uint64_t SrcSVOff) {
6023 // This requires the copy size to be a constant, preferrably
6024 // within a subtarget-specific limit.
6025 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6028 uint64_t SizeVal = ConstantSize->getZExtValue();
6029 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6032 /// If not DWORD aligned, call the library.
6033 if ((Align & 3) != 0)
6038 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6041 unsigned UBytes = AVT.getSizeInBits() / 8;
6042 unsigned CountVal = SizeVal / UBytes;
6043 SDValue Count = DAG.getIntPtrConstant(CountVal);
6044 unsigned BytesLeft = SizeVal % UBytes;
6046 SDValue InFlag(0, 0);
6047 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6050 InFlag = Chain.getValue(1);
6051 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6054 InFlag = Chain.getValue(1);
6055 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6058 InFlag = Chain.getValue(1);
6060 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6061 SmallVector<SDValue, 8> Ops;
6062 Ops.push_back(Chain);
6063 Ops.push_back(DAG.getValueType(AVT));
6064 Ops.push_back(InFlag);
6065 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6067 SmallVector<SDValue, 4> Results;
6068 Results.push_back(RepMovs);
6070 // Handle the last 1 - 7 bytes.
6071 unsigned Offset = SizeVal - BytesLeft;
6072 EVT DstVT = Dst.getValueType();
6073 EVT SrcVT = Src.getValueType();
6074 EVT SizeVT = Size.getValueType();
6075 Results.push_back(DAG.getMemcpy(Chain, dl,
6076 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6077 DAG.getConstant(Offset, DstVT)),
6078 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6079 DAG.getConstant(Offset, SrcVT)),
6080 DAG.getConstant(BytesLeft, SizeVT),
6081 Align, AlwaysInline,
6082 DstSV, DstSVOff + Offset,
6083 SrcSV, SrcSVOff + Offset));
6086 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6087 &Results[0], Results.size());
6090 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6091 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6092 DebugLoc dl = Op.getDebugLoc();
6094 if (!Subtarget->is64Bit()) {
6095 // vastart just stores the address of the VarArgsFrameIndex slot into the
6096 // memory location argument.
6097 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6098 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6102 // gp_offset (0 - 6 * 8)
6103 // fp_offset (48 - 48 + 8 * 16)
6104 // overflow_arg_area (point to parameters coming in memory).
6106 SmallVector<SDValue, 8> MemOps;
6107 SDValue FIN = Op.getOperand(1);
6109 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6110 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6112 MemOps.push_back(Store);
6115 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6116 FIN, DAG.getIntPtrConstant(4));
6117 Store = DAG.getStore(Op.getOperand(0), dl,
6118 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6120 MemOps.push_back(Store);
6122 // Store ptr to overflow_arg_area
6123 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6124 FIN, DAG.getIntPtrConstant(4));
6125 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6126 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6127 MemOps.push_back(Store);
6129 // Store ptr to reg_save_area.
6130 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6131 FIN, DAG.getIntPtrConstant(8));
6132 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6133 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6134 MemOps.push_back(Store);
6135 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6136 &MemOps[0], MemOps.size());
6139 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6140 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6141 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6142 SDValue Chain = Op.getOperand(0);
6143 SDValue SrcPtr = Op.getOperand(1);
6144 SDValue SrcSV = Op.getOperand(2);
6146 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6150 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6151 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6152 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6153 SDValue Chain = Op.getOperand(0);
6154 SDValue DstPtr = Op.getOperand(1);
6155 SDValue SrcPtr = Op.getOperand(2);
6156 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6157 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6158 DebugLoc dl = Op.getDebugLoc();
6160 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6161 DAG.getIntPtrConstant(24), 8, false,
6162 DstSV, 0, SrcSV, 0);
6166 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6167 DebugLoc dl = Op.getDebugLoc();
6168 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6170 default: return SDValue(); // Don't custom lower most intrinsics.
6171 // Comparison intrinsics.
6172 case Intrinsic::x86_sse_comieq_ss:
6173 case Intrinsic::x86_sse_comilt_ss:
6174 case Intrinsic::x86_sse_comile_ss:
6175 case Intrinsic::x86_sse_comigt_ss:
6176 case Intrinsic::x86_sse_comige_ss:
6177 case Intrinsic::x86_sse_comineq_ss:
6178 case Intrinsic::x86_sse_ucomieq_ss:
6179 case Intrinsic::x86_sse_ucomilt_ss:
6180 case Intrinsic::x86_sse_ucomile_ss:
6181 case Intrinsic::x86_sse_ucomigt_ss:
6182 case Intrinsic::x86_sse_ucomige_ss:
6183 case Intrinsic::x86_sse_ucomineq_ss:
6184 case Intrinsic::x86_sse2_comieq_sd:
6185 case Intrinsic::x86_sse2_comilt_sd:
6186 case Intrinsic::x86_sse2_comile_sd:
6187 case Intrinsic::x86_sse2_comigt_sd:
6188 case Intrinsic::x86_sse2_comige_sd:
6189 case Intrinsic::x86_sse2_comineq_sd:
6190 case Intrinsic::x86_sse2_ucomieq_sd:
6191 case Intrinsic::x86_sse2_ucomilt_sd:
6192 case Intrinsic::x86_sse2_ucomile_sd:
6193 case Intrinsic::x86_sse2_ucomigt_sd:
6194 case Intrinsic::x86_sse2_ucomige_sd:
6195 case Intrinsic::x86_sse2_ucomineq_sd: {
6197 ISD::CondCode CC = ISD::SETCC_INVALID;
6200 case Intrinsic::x86_sse_comieq_ss:
6201 case Intrinsic::x86_sse2_comieq_sd:
6205 case Intrinsic::x86_sse_comilt_ss:
6206 case Intrinsic::x86_sse2_comilt_sd:
6210 case Intrinsic::x86_sse_comile_ss:
6211 case Intrinsic::x86_sse2_comile_sd:
6215 case Intrinsic::x86_sse_comigt_ss:
6216 case Intrinsic::x86_sse2_comigt_sd:
6220 case Intrinsic::x86_sse_comige_ss:
6221 case Intrinsic::x86_sse2_comige_sd:
6225 case Intrinsic::x86_sse_comineq_ss:
6226 case Intrinsic::x86_sse2_comineq_sd:
6230 case Intrinsic::x86_sse_ucomieq_ss:
6231 case Intrinsic::x86_sse2_ucomieq_sd:
6232 Opc = X86ISD::UCOMI;
6235 case Intrinsic::x86_sse_ucomilt_ss:
6236 case Intrinsic::x86_sse2_ucomilt_sd:
6237 Opc = X86ISD::UCOMI;
6240 case Intrinsic::x86_sse_ucomile_ss:
6241 case Intrinsic::x86_sse2_ucomile_sd:
6242 Opc = X86ISD::UCOMI;
6245 case Intrinsic::x86_sse_ucomigt_ss:
6246 case Intrinsic::x86_sse2_ucomigt_sd:
6247 Opc = X86ISD::UCOMI;
6250 case Intrinsic::x86_sse_ucomige_ss:
6251 case Intrinsic::x86_sse2_ucomige_sd:
6252 Opc = X86ISD::UCOMI;
6255 case Intrinsic::x86_sse_ucomineq_ss:
6256 case Intrinsic::x86_sse2_ucomineq_sd:
6257 Opc = X86ISD::UCOMI;
6262 SDValue LHS = Op.getOperand(1);
6263 SDValue RHS = Op.getOperand(2);
6264 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6265 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6266 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6267 DAG.getConstant(X86CC, MVT::i8), Cond);
6268 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6270 // ptest intrinsics. The intrinsic these come from are designed to return
6271 // an integer value, not just an instruction so lower it to the ptest
6272 // pattern and a setcc for the result.
6273 case Intrinsic::x86_sse41_ptestz:
6274 case Intrinsic::x86_sse41_ptestc:
6275 case Intrinsic::x86_sse41_ptestnzc:{
6278 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6279 case Intrinsic::x86_sse41_ptestz:
6281 X86CC = X86::COND_E;
6283 case Intrinsic::x86_sse41_ptestc:
6285 X86CC = X86::COND_B;
6287 case Intrinsic::x86_sse41_ptestnzc:
6289 X86CC = X86::COND_A;
6293 SDValue LHS = Op.getOperand(1);
6294 SDValue RHS = Op.getOperand(2);
6295 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6296 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6297 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6298 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6301 // Fix vector shift instructions where the last operand is a non-immediate
6303 case Intrinsic::x86_sse2_pslli_w:
6304 case Intrinsic::x86_sse2_pslli_d:
6305 case Intrinsic::x86_sse2_pslli_q:
6306 case Intrinsic::x86_sse2_psrli_w:
6307 case Intrinsic::x86_sse2_psrli_d:
6308 case Intrinsic::x86_sse2_psrli_q:
6309 case Intrinsic::x86_sse2_psrai_w:
6310 case Intrinsic::x86_sse2_psrai_d:
6311 case Intrinsic::x86_mmx_pslli_w:
6312 case Intrinsic::x86_mmx_pslli_d:
6313 case Intrinsic::x86_mmx_pslli_q:
6314 case Intrinsic::x86_mmx_psrli_w:
6315 case Intrinsic::x86_mmx_psrli_d:
6316 case Intrinsic::x86_mmx_psrli_q:
6317 case Intrinsic::x86_mmx_psrai_w:
6318 case Intrinsic::x86_mmx_psrai_d: {
6319 SDValue ShAmt = Op.getOperand(2);
6320 if (isa<ConstantSDNode>(ShAmt))
6323 unsigned NewIntNo = 0;
6324 EVT ShAmtVT = MVT::v4i32;
6326 case Intrinsic::x86_sse2_pslli_w:
6327 NewIntNo = Intrinsic::x86_sse2_psll_w;
6329 case Intrinsic::x86_sse2_pslli_d:
6330 NewIntNo = Intrinsic::x86_sse2_psll_d;
6332 case Intrinsic::x86_sse2_pslli_q:
6333 NewIntNo = Intrinsic::x86_sse2_psll_q;
6335 case Intrinsic::x86_sse2_psrli_w:
6336 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6338 case Intrinsic::x86_sse2_psrli_d:
6339 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6341 case Intrinsic::x86_sse2_psrli_q:
6342 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6344 case Intrinsic::x86_sse2_psrai_w:
6345 NewIntNo = Intrinsic::x86_sse2_psra_w;
6347 case Intrinsic::x86_sse2_psrai_d:
6348 NewIntNo = Intrinsic::x86_sse2_psra_d;
6351 ShAmtVT = MVT::v2i32;
6353 case Intrinsic::x86_mmx_pslli_w:
6354 NewIntNo = Intrinsic::x86_mmx_psll_w;
6356 case Intrinsic::x86_mmx_pslli_d:
6357 NewIntNo = Intrinsic::x86_mmx_psll_d;
6359 case Intrinsic::x86_mmx_pslli_q:
6360 NewIntNo = Intrinsic::x86_mmx_psll_q;
6362 case Intrinsic::x86_mmx_psrli_w:
6363 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6365 case Intrinsic::x86_mmx_psrli_d:
6366 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6368 case Intrinsic::x86_mmx_psrli_q:
6369 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6371 case Intrinsic::x86_mmx_psrai_w:
6372 NewIntNo = Intrinsic::x86_mmx_psra_w;
6374 case Intrinsic::x86_mmx_psrai_d:
6375 NewIntNo = Intrinsic::x86_mmx_psra_d;
6377 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6382 EVT VT = Op.getValueType();
6383 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6384 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6385 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6386 DAG.getConstant(NewIntNo, MVT::i32),
6387 Op.getOperand(1), ShAmt);
6392 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6393 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6394 DebugLoc dl = Op.getDebugLoc();
6397 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6399 DAG.getConstant(TD->getPointerSize(),
6400 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6401 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6402 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6407 // Just load the return address.
6408 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6409 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6410 RetAddrFI, NULL, 0);
6413 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6414 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6415 MFI->setFrameAddressIsTaken(true);
6416 EVT VT = Op.getValueType();
6417 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6418 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6419 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6420 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6422 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6426 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6427 SelectionDAG &DAG) {
6428 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6431 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6433 MachineFunction &MF = DAG.getMachineFunction();
6434 SDValue Chain = Op.getOperand(0);
6435 SDValue Offset = Op.getOperand(1);
6436 SDValue Handler = Op.getOperand(2);
6437 DebugLoc dl = Op.getDebugLoc();
6439 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6441 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6443 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6444 DAG.getIntPtrConstant(-TD->getPointerSize()));
6445 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6446 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6447 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6448 MF.getRegInfo().addLiveOut(StoreAddrReg);
6450 return DAG.getNode(X86ISD::EH_RETURN, dl,
6452 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6455 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6456 SelectionDAG &DAG) {
6457 SDValue Root = Op.getOperand(0);
6458 SDValue Trmp = Op.getOperand(1); // trampoline
6459 SDValue FPtr = Op.getOperand(2); // nested function
6460 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6461 DebugLoc dl = Op.getDebugLoc();
6463 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6465 const X86InstrInfo *TII =
6466 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6468 if (Subtarget->is64Bit()) {
6469 SDValue OutChains[6];
6471 // Large code-model.
6473 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6474 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6476 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6477 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6479 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6481 // Load the pointer to the nested function into R11.
6482 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6483 SDValue Addr = Trmp;
6484 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6487 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6488 DAG.getConstant(2, MVT::i64));
6489 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6491 // Load the 'nest' parameter value into R10.
6492 // R10 is specified in X86CallingConv.td
6493 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6494 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6495 DAG.getConstant(10, MVT::i64));
6496 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6497 Addr, TrmpAddr, 10);
6499 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6500 DAG.getConstant(12, MVT::i64));
6501 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6503 // Jump to the nested function.
6504 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6505 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6506 DAG.getConstant(20, MVT::i64));
6507 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6508 Addr, TrmpAddr, 20);
6510 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6511 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6512 DAG.getConstant(22, MVT::i64));
6513 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6517 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6518 return DAG.getMergeValues(Ops, 2, dl);
6520 const Function *Func =
6521 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6522 unsigned CC = Func->getCallingConv();
6527 llvm_unreachable("Unsupported calling convention");
6528 case CallingConv::C:
6529 case CallingConv::X86_StdCall: {
6530 // Pass 'nest' parameter in ECX.
6531 // Must be kept in sync with X86CallingConv.td
6534 // Check that ECX wasn't needed by an 'inreg' parameter.
6535 const FunctionType *FTy = Func->getFunctionType();
6536 const AttrListPtr &Attrs = Func->getAttributes();
6538 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6539 unsigned InRegCount = 0;
6542 for (FunctionType::param_iterator I = FTy->param_begin(),
6543 E = FTy->param_end(); I != E; ++I, ++Idx)
6544 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6545 // FIXME: should only count parameters that are lowered to integers.
6546 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6548 if (InRegCount > 2) {
6549 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6554 case CallingConv::X86_FastCall:
6555 case CallingConv::Fast:
6556 // Pass 'nest' parameter in EAX.
6557 // Must be kept in sync with X86CallingConv.td
6562 SDValue OutChains[4];
6565 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6566 DAG.getConstant(10, MVT::i32));
6567 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6569 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6570 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6571 OutChains[0] = DAG.getStore(Root, dl,
6572 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6575 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6576 DAG.getConstant(1, MVT::i32));
6577 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6579 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6580 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6581 DAG.getConstant(5, MVT::i32));
6582 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6583 TrmpAddr, 5, false, 1);
6585 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6586 DAG.getConstant(6, MVT::i32));
6587 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6590 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6591 return DAG.getMergeValues(Ops, 2, dl);
6595 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6597 The rounding mode is in bits 11:10 of FPSR, and has the following
6604 FLT_ROUNDS, on the other hand, expects the following:
6611 To perform the conversion, we do:
6612 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6615 MachineFunction &MF = DAG.getMachineFunction();
6616 const TargetMachine &TM = MF.getTarget();
6617 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6618 unsigned StackAlignment = TFI.getStackAlignment();
6619 EVT VT = Op.getValueType();
6620 DebugLoc dl = Op.getDebugLoc();
6622 // Save FP Control Word to stack slot
6623 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6624 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6626 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6627 DAG.getEntryNode(), StackSlot);
6629 // Load FP Control Word from stack slot
6630 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6632 // Transform as necessary
6634 DAG.getNode(ISD::SRL, dl, MVT::i16,
6635 DAG.getNode(ISD::AND, dl, MVT::i16,
6636 CWD, DAG.getConstant(0x800, MVT::i16)),
6637 DAG.getConstant(11, MVT::i8));
6639 DAG.getNode(ISD::SRL, dl, MVT::i16,
6640 DAG.getNode(ISD::AND, dl, MVT::i16,
6641 CWD, DAG.getConstant(0x400, MVT::i16)),
6642 DAG.getConstant(9, MVT::i8));
6645 DAG.getNode(ISD::AND, dl, MVT::i16,
6646 DAG.getNode(ISD::ADD, dl, MVT::i16,
6647 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6648 DAG.getConstant(1, MVT::i16)),
6649 DAG.getConstant(3, MVT::i16));
6652 return DAG.getNode((VT.getSizeInBits() < 16 ?
6653 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6656 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6657 EVT VT = Op.getValueType();
6659 unsigned NumBits = VT.getSizeInBits();
6660 DebugLoc dl = Op.getDebugLoc();
6662 Op = Op.getOperand(0);
6663 if (VT == MVT::i8) {
6664 // Zero extend to i32 since there is not an i8 bsr.
6666 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6669 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6670 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6671 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6673 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6674 SmallVector<SDValue, 4> Ops;
6676 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6677 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6678 Ops.push_back(Op.getValue(1));
6679 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6681 // Finally xor with NumBits-1.
6682 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6685 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6689 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6690 EVT VT = Op.getValueType();
6692 unsigned NumBits = VT.getSizeInBits();
6693 DebugLoc dl = Op.getDebugLoc();
6695 Op = Op.getOperand(0);
6696 if (VT == MVT::i8) {
6698 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6701 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6702 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6703 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6705 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6706 SmallVector<SDValue, 4> Ops;
6708 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6709 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6710 Ops.push_back(Op.getValue(1));
6711 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6714 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6718 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6719 EVT VT = Op.getValueType();
6720 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6721 DebugLoc dl = Op.getDebugLoc();
6723 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6724 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6725 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6726 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6727 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6729 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6730 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6731 // return AloBlo + AloBhi + AhiBlo;
6733 SDValue A = Op.getOperand(0);
6734 SDValue B = Op.getOperand(1);
6736 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6737 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6738 A, DAG.getConstant(32, MVT::i32));
6739 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6740 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6741 B, DAG.getConstant(32, MVT::i32));
6742 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6743 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6745 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6746 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6748 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6749 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6751 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6752 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6753 AloBhi, DAG.getConstant(32, MVT::i32));
6754 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6755 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6756 AhiBlo, DAG.getConstant(32, MVT::i32));
6757 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6758 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6763 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6764 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6765 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6766 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6767 // has only one use.
6768 SDNode *N = Op.getNode();
6769 SDValue LHS = N->getOperand(0);
6770 SDValue RHS = N->getOperand(1);
6771 unsigned BaseOp = 0;
6773 DebugLoc dl = Op.getDebugLoc();
6775 switch (Op.getOpcode()) {
6776 default: llvm_unreachable("Unknown ovf instruction!");
6778 // A subtract of one will be selected as a INC. Note that INC doesn't
6779 // set CF, so we can't do this for UADDO.
6780 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6781 if (C->getAPIntValue() == 1) {
6782 BaseOp = X86ISD::INC;
6786 BaseOp = X86ISD::ADD;
6790 BaseOp = X86ISD::ADD;
6794 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6795 // set CF, so we can't do this for USUBO.
6796 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6797 if (C->getAPIntValue() == 1) {
6798 BaseOp = X86ISD::DEC;
6802 BaseOp = X86ISD::SUB;
6806 BaseOp = X86ISD::SUB;
6810 BaseOp = X86ISD::SMUL;
6814 BaseOp = X86ISD::UMUL;
6819 // Also sets EFLAGS.
6820 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6821 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6824 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6825 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6827 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6831 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6832 EVT T = Op.getValueType();
6833 DebugLoc dl = Op.getDebugLoc();
6836 switch(T.getSimpleVT().SimpleTy) {
6838 assert(false && "Invalid value type!");
6839 case MVT::i8: Reg = X86::AL; size = 1; break;
6840 case MVT::i16: Reg = X86::AX; size = 2; break;
6841 case MVT::i32: Reg = X86::EAX; size = 4; break;
6843 assert(Subtarget->is64Bit() && "Node not type legal!");
6844 Reg = X86::RAX; size = 8;
6847 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6848 Op.getOperand(2), SDValue());
6849 SDValue Ops[] = { cpIn.getValue(0),
6852 DAG.getTargetConstant(size, MVT::i8),
6854 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6855 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6857 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6861 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6862 SelectionDAG &DAG) {
6863 assert(Subtarget->is64Bit() && "Result not type legalized?");
6864 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6865 SDValue TheChain = Op.getOperand(0);
6866 DebugLoc dl = Op.getDebugLoc();
6867 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6868 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6869 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6871 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6872 DAG.getConstant(32, MVT::i8));
6874 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6877 return DAG.getMergeValues(Ops, 2, dl);
6880 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6881 SDNode *Node = Op.getNode();
6882 DebugLoc dl = Node->getDebugLoc();
6883 EVT T = Node->getValueType(0);
6884 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6885 DAG.getConstant(0, T), Node->getOperand(2));
6886 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6887 cast<AtomicSDNode>(Node)->getMemoryVT(),
6888 Node->getOperand(0),
6889 Node->getOperand(1), negOp,
6890 cast<AtomicSDNode>(Node)->getSrcValue(),
6891 cast<AtomicSDNode>(Node)->getAlignment());
6894 /// LowerOperation - Provide custom lowering hooks for some operations.
6896 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6897 switch (Op.getOpcode()) {
6898 default: llvm_unreachable("Should not custom lower this!");
6899 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6900 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6901 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6902 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6903 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6904 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6905 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6906 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6907 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6908 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6909 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6910 case ISD::SHL_PARTS:
6911 case ISD::SRA_PARTS:
6912 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6913 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6914 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6915 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6916 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6917 case ISD::FABS: return LowerFABS(Op, DAG);
6918 case ISD::FNEG: return LowerFNEG(Op, DAG);
6919 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6920 case ISD::SETCC: return LowerSETCC(Op, DAG);
6921 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6922 case ISD::SELECT: return LowerSELECT(Op, DAG);
6923 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6924 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6925 case ISD::VASTART: return LowerVASTART(Op, DAG);
6926 case ISD::VAARG: return LowerVAARG(Op, DAG);
6927 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6928 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6929 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6930 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6931 case ISD::FRAME_TO_ARGS_OFFSET:
6932 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6933 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6934 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6935 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6936 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6937 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6938 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6939 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6945 case ISD::UMULO: return LowerXALUO(Op, DAG);
6946 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6950 void X86TargetLowering::
6951 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6952 SelectionDAG &DAG, unsigned NewOp) {
6953 EVT T = Node->getValueType(0);
6954 DebugLoc dl = Node->getDebugLoc();
6955 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6957 SDValue Chain = Node->getOperand(0);
6958 SDValue In1 = Node->getOperand(1);
6959 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6960 Node->getOperand(2), DAG.getIntPtrConstant(0));
6961 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6962 Node->getOperand(2), DAG.getIntPtrConstant(1));
6963 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6964 // have a MemOperand. Pass the info through as a normal operand.
6965 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6966 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6967 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6968 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6969 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6970 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6971 Results.push_back(Result.getValue(2));
6974 /// ReplaceNodeResults - Replace a node with an illegal result type
6975 /// with a new node built out of custom code.
6976 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6977 SmallVectorImpl<SDValue>&Results,
6978 SelectionDAG &DAG) {
6979 DebugLoc dl = N->getDebugLoc();
6980 switch (N->getOpcode()) {
6982 assert(false && "Do not know how to custom type legalize this operation!");
6984 case ISD::FP_TO_SINT: {
6985 std::pair<SDValue,SDValue> Vals =
6986 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6987 SDValue FIST = Vals.first, StackSlot = Vals.second;
6988 if (FIST.getNode() != 0) {
6989 EVT VT = N->getValueType(0);
6990 // Return a load from the stack slot.
6991 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6995 case ISD::READCYCLECOUNTER: {
6996 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6997 SDValue TheChain = N->getOperand(0);
6998 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6999 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7001 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7003 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7004 SDValue Ops[] = { eax, edx };
7005 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7006 Results.push_back(edx.getValue(1));
7009 case ISD::ATOMIC_CMP_SWAP: {
7010 EVT T = N->getValueType(0);
7011 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7012 SDValue cpInL, cpInH;
7013 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7014 DAG.getConstant(0, MVT::i32));
7015 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7016 DAG.getConstant(1, MVT::i32));
7017 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7018 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7020 SDValue swapInL, swapInH;
7021 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7022 DAG.getConstant(0, MVT::i32));
7023 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7024 DAG.getConstant(1, MVT::i32));
7025 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7027 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7028 swapInL.getValue(1));
7029 SDValue Ops[] = { swapInH.getValue(0),
7031 swapInH.getValue(1) };
7032 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7033 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7034 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7035 MVT::i32, Result.getValue(1));
7036 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7037 MVT::i32, cpOutL.getValue(2));
7038 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7039 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7040 Results.push_back(cpOutH.getValue(1));
7043 case ISD::ATOMIC_LOAD_ADD:
7044 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7046 case ISD::ATOMIC_LOAD_AND:
7047 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7049 case ISD::ATOMIC_LOAD_NAND:
7050 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7052 case ISD::ATOMIC_LOAD_OR:
7053 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7055 case ISD::ATOMIC_LOAD_SUB:
7056 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7058 case ISD::ATOMIC_LOAD_XOR:
7059 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7061 case ISD::ATOMIC_SWAP:
7062 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7067 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7069 default: return NULL;
7070 case X86ISD::BSF: return "X86ISD::BSF";
7071 case X86ISD::BSR: return "X86ISD::BSR";
7072 case X86ISD::SHLD: return "X86ISD::SHLD";
7073 case X86ISD::SHRD: return "X86ISD::SHRD";
7074 case X86ISD::FAND: return "X86ISD::FAND";
7075 case X86ISD::FOR: return "X86ISD::FOR";
7076 case X86ISD::FXOR: return "X86ISD::FXOR";
7077 case X86ISD::FSRL: return "X86ISD::FSRL";
7078 case X86ISD::FILD: return "X86ISD::FILD";
7079 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7080 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7081 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7082 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7083 case X86ISD::FLD: return "X86ISD::FLD";
7084 case X86ISD::FST: return "X86ISD::FST";
7085 case X86ISD::CALL: return "X86ISD::CALL";
7086 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7087 case X86ISD::BT: return "X86ISD::BT";
7088 case X86ISD::CMP: return "X86ISD::CMP";
7089 case X86ISD::COMI: return "X86ISD::COMI";
7090 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7091 case X86ISD::SETCC: return "X86ISD::SETCC";
7092 case X86ISD::CMOV: return "X86ISD::CMOV";
7093 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7094 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7095 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7096 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7097 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7098 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7099 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7100 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7101 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7102 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7103 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7104 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7105 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7106 case X86ISD::FMAX: return "X86ISD::FMAX";
7107 case X86ISD::FMIN: return "X86ISD::FMIN";
7108 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7109 case X86ISD::FRCP: return "X86ISD::FRCP";
7110 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7111 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7112 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7113 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7114 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7115 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7116 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7117 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7118 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7119 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7120 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7121 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7122 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7123 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7124 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7125 case X86ISD::VSHL: return "X86ISD::VSHL";
7126 case X86ISD::VSRL: return "X86ISD::VSRL";
7127 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7128 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7129 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7130 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7131 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7132 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7133 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7134 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7135 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7136 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7137 case X86ISD::ADD: return "X86ISD::ADD";
7138 case X86ISD::SUB: return "X86ISD::SUB";
7139 case X86ISD::SMUL: return "X86ISD::SMUL";
7140 case X86ISD::UMUL: return "X86ISD::UMUL";
7141 case X86ISD::INC: return "X86ISD::INC";
7142 case X86ISD::DEC: return "X86ISD::DEC";
7143 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7144 case X86ISD::PTEST: return "X86ISD::PTEST";
7145 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7149 // isLegalAddressingMode - Return true if the addressing mode represented
7150 // by AM is legal for this target, for a load/store of the specified type.
7151 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7152 const Type *Ty) const {
7153 // X86 supports extremely general addressing modes.
7154 CodeModel::Model M = getTargetMachine().getCodeModel();
7156 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7157 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7162 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7164 // If a reference to this global requires an extra load, we can't fold it.
7165 if (isGlobalStubReference(GVFlags))
7168 // If BaseGV requires a register for the PIC base, we cannot also have a
7169 // BaseReg specified.
7170 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7173 // If lower 4G is not available, then we must use rip-relative addressing.
7174 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7184 // These scales always work.
7189 // These scales are formed with basereg+scalereg. Only accept if there is
7194 default: // Other stuff never works.
7202 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7203 if (!Ty1->isInteger() || !Ty2->isInteger())
7205 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7206 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7207 if (NumBits1 <= NumBits2)
7209 return Subtarget->is64Bit() || NumBits1 < 64;
7212 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7213 if (!VT1.isInteger() || !VT2.isInteger())
7215 unsigned NumBits1 = VT1.getSizeInBits();
7216 unsigned NumBits2 = VT2.getSizeInBits();
7217 if (NumBits1 <= NumBits2)
7219 return Subtarget->is64Bit() || NumBits1 < 64;
7222 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7223 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7224 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7225 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7228 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7229 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7230 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7233 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7234 // i16 instructions are longer (0x66 prefix) and potentially slower.
7235 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7238 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7239 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7240 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7241 /// are assumed to be legal.
7243 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7245 // Only do shuffles on 128-bit vector types for now.
7246 if (VT.getSizeInBits() == 64)
7249 // FIXME: pshufb, blends, palignr, shifts.
7250 return (VT.getVectorNumElements() == 2 ||
7251 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7252 isMOVLMask(M, VT) ||
7253 isSHUFPMask(M, VT) ||
7254 isPSHUFDMask(M, VT) ||
7255 isPSHUFHWMask(M, VT) ||
7256 isPSHUFLWMask(M, VT) ||
7257 isUNPCKLMask(M, VT) ||
7258 isUNPCKHMask(M, VT) ||
7259 isUNPCKL_v_undef_Mask(M, VT) ||
7260 isUNPCKH_v_undef_Mask(M, VT));
7264 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7266 unsigned NumElts = VT.getVectorNumElements();
7267 // FIXME: This collection of masks seems suspect.
7270 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7271 return (isMOVLMask(Mask, VT) ||
7272 isCommutedMOVLMask(Mask, VT, true) ||
7273 isSHUFPMask(Mask, VT) ||
7274 isCommutedSHUFPMask(Mask, VT));
7279 //===----------------------------------------------------------------------===//
7280 // X86 Scheduler Hooks
7281 //===----------------------------------------------------------------------===//
7283 // private utility function
7285 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7286 MachineBasicBlock *MBB,
7294 TargetRegisterClass *RC,
7295 bool invSrc) const {
7296 // For the atomic bitwise operator, we generate
7299 // ld t1 = [bitinstr.addr]
7300 // op t2 = t1, [bitinstr.val]
7302 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7304 // fallthrough -->nextMBB
7305 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7306 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7307 MachineFunction::iterator MBBIter = MBB;
7310 /// First build the CFG
7311 MachineFunction *F = MBB->getParent();
7312 MachineBasicBlock *thisMBB = MBB;
7313 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7314 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7315 F->insert(MBBIter, newMBB);
7316 F->insert(MBBIter, nextMBB);
7318 // Move all successors to thisMBB to nextMBB
7319 nextMBB->transferSuccessors(thisMBB);
7321 // Update thisMBB to fall through to newMBB
7322 thisMBB->addSuccessor(newMBB);
7324 // newMBB jumps to itself and fall through to nextMBB
7325 newMBB->addSuccessor(nextMBB);
7326 newMBB->addSuccessor(newMBB);
7328 // Insert instructions into newMBB based on incoming instruction
7329 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7330 "unexpected number of operands");
7331 DebugLoc dl = bInstr->getDebugLoc();
7332 MachineOperand& destOper = bInstr->getOperand(0);
7333 MachineOperand* argOpers[2 + X86AddrNumOperands];
7334 int numArgs = bInstr->getNumOperands() - 1;
7335 for (int i=0; i < numArgs; ++i)
7336 argOpers[i] = &bInstr->getOperand(i+1);
7338 // x86 address has 4 operands: base, index, scale, and displacement
7339 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7340 int valArgIndx = lastAddrIndx + 1;
7342 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7343 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7344 for (int i=0; i <= lastAddrIndx; ++i)
7345 (*MIB).addOperand(*argOpers[i]);
7347 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7349 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7354 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7355 assert((argOpers[valArgIndx]->isReg() ||
7356 argOpers[valArgIndx]->isImm()) &&
7358 if (argOpers[valArgIndx]->isReg())
7359 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7361 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7363 (*MIB).addOperand(*argOpers[valArgIndx]);
7365 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7368 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7369 for (int i=0; i <= lastAddrIndx; ++i)
7370 (*MIB).addOperand(*argOpers[i]);
7372 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7373 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7375 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7379 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7381 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7385 // private utility function: 64 bit atomics on 32 bit host.
7387 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7388 MachineBasicBlock *MBB,
7393 bool invSrc) const {
7394 // For the atomic bitwise operator, we generate
7395 // thisMBB (instructions are in pairs, except cmpxchg8b)
7396 // ld t1,t2 = [bitinstr.addr]
7398 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7399 // op t5, t6 <- out1, out2, [bitinstr.val]
7400 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7401 // mov ECX, EBX <- t5, t6
7402 // mov EAX, EDX <- t1, t2
7403 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7404 // mov t3, t4 <- EAX, EDX
7406 // result in out1, out2
7407 // fallthrough -->nextMBB
7409 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7410 const unsigned LoadOpc = X86::MOV32rm;
7411 const unsigned copyOpc = X86::MOV32rr;
7412 const unsigned NotOpc = X86::NOT32r;
7413 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7414 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7415 MachineFunction::iterator MBBIter = MBB;
7418 /// First build the CFG
7419 MachineFunction *F = MBB->getParent();
7420 MachineBasicBlock *thisMBB = MBB;
7421 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7422 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7423 F->insert(MBBIter, newMBB);
7424 F->insert(MBBIter, nextMBB);
7426 // Move all successors to thisMBB to nextMBB
7427 nextMBB->transferSuccessors(thisMBB);
7429 // Update thisMBB to fall through to newMBB
7430 thisMBB->addSuccessor(newMBB);
7432 // newMBB jumps to itself and fall through to nextMBB
7433 newMBB->addSuccessor(nextMBB);
7434 newMBB->addSuccessor(newMBB);
7436 DebugLoc dl = bInstr->getDebugLoc();
7437 // Insert instructions into newMBB based on incoming instruction
7438 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7439 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7440 "unexpected number of operands");
7441 MachineOperand& dest1Oper = bInstr->getOperand(0);
7442 MachineOperand& dest2Oper = bInstr->getOperand(1);
7443 MachineOperand* argOpers[2 + X86AddrNumOperands];
7444 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7445 argOpers[i] = &bInstr->getOperand(i+2);
7447 // x86 address has 4 operands: base, index, scale, and displacement
7448 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7450 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7451 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7452 for (int i=0; i <= lastAddrIndx; ++i)
7453 (*MIB).addOperand(*argOpers[i]);
7454 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7455 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7456 // add 4 to displacement.
7457 for (int i=0; i <= lastAddrIndx-2; ++i)
7458 (*MIB).addOperand(*argOpers[i]);
7459 MachineOperand newOp3 = *(argOpers[3]);
7461 newOp3.setImm(newOp3.getImm()+4);
7463 newOp3.setOffset(newOp3.getOffset()+4);
7464 (*MIB).addOperand(newOp3);
7465 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7467 // t3/4 are defined later, at the bottom of the loop
7468 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7469 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7470 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7471 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7472 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7473 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7475 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7476 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7478 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7479 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7485 int valArgIndx = lastAddrIndx + 1;
7486 assert((argOpers[valArgIndx]->isReg() ||
7487 argOpers[valArgIndx]->isImm()) &&
7489 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7490 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7491 if (argOpers[valArgIndx]->isReg())
7492 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7494 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7495 if (regOpcL != X86::MOV32rr)
7497 (*MIB).addOperand(*argOpers[valArgIndx]);
7498 assert(argOpers[valArgIndx + 1]->isReg() ==
7499 argOpers[valArgIndx]->isReg());
7500 assert(argOpers[valArgIndx + 1]->isImm() ==
7501 argOpers[valArgIndx]->isImm());
7502 if (argOpers[valArgIndx + 1]->isReg())
7503 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7505 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7506 if (regOpcH != X86::MOV32rr)
7508 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7510 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7512 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7515 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7517 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7520 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7521 for (int i=0; i <= lastAddrIndx; ++i)
7522 (*MIB).addOperand(*argOpers[i]);
7524 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7525 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7527 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7528 MIB.addReg(X86::EAX);
7529 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7530 MIB.addReg(X86::EDX);
7533 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7535 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7539 // private utility function
7541 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7542 MachineBasicBlock *MBB,
7543 unsigned cmovOpc) const {
7544 // For the atomic min/max operator, we generate
7547 // ld t1 = [min/max.addr]
7548 // mov t2 = [min/max.val]
7550 // cmov[cond] t2 = t1
7552 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7554 // fallthrough -->nextMBB
7556 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7557 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7558 MachineFunction::iterator MBBIter = MBB;
7561 /// First build the CFG
7562 MachineFunction *F = MBB->getParent();
7563 MachineBasicBlock *thisMBB = MBB;
7564 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7565 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7566 F->insert(MBBIter, newMBB);
7567 F->insert(MBBIter, nextMBB);
7569 // Move all successors of thisMBB to nextMBB
7570 nextMBB->transferSuccessors(thisMBB);
7572 // Update thisMBB to fall through to newMBB
7573 thisMBB->addSuccessor(newMBB);
7575 // newMBB jumps to newMBB and fall through to nextMBB
7576 newMBB->addSuccessor(nextMBB);
7577 newMBB->addSuccessor(newMBB);
7579 DebugLoc dl = mInstr->getDebugLoc();
7580 // Insert instructions into newMBB based on incoming instruction
7581 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7582 "unexpected number of operands");
7583 MachineOperand& destOper = mInstr->getOperand(0);
7584 MachineOperand* argOpers[2 + X86AddrNumOperands];
7585 int numArgs = mInstr->getNumOperands() - 1;
7586 for (int i=0; i < numArgs; ++i)
7587 argOpers[i] = &mInstr->getOperand(i+1);
7589 // x86 address has 4 operands: base, index, scale, and displacement
7590 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7591 int valArgIndx = lastAddrIndx + 1;
7593 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7594 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7595 for (int i=0; i <= lastAddrIndx; ++i)
7596 (*MIB).addOperand(*argOpers[i]);
7598 // We only support register and immediate values
7599 assert((argOpers[valArgIndx]->isReg() ||
7600 argOpers[valArgIndx]->isImm()) &&
7603 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7604 if (argOpers[valArgIndx]->isReg())
7605 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7607 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7608 (*MIB).addOperand(*argOpers[valArgIndx]);
7610 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7613 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7618 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7619 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7623 // Cmp and exchange if none has modified the memory location
7624 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7625 for (int i=0; i <= lastAddrIndx; ++i)
7626 (*MIB).addOperand(*argOpers[i]);
7628 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7629 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7631 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7632 MIB.addReg(X86::EAX);
7635 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7637 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7641 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7642 // all of this code can be replaced with that in the .td file.
7644 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7645 unsigned numArgs, bool memArg) const {
7647 MachineFunction *F = BB->getParent();
7648 DebugLoc dl = MI->getDebugLoc();
7649 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7654 Opc = numArgs == 3 ?
7655 X86::PCMPISTRM128rm :
7656 X86::PCMPESTRM128rm;
7658 Opc = numArgs == 3 ?
7659 X86::PCMPISTRM128rr :
7660 X86::PCMPESTRM128rr;
7663 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7665 for (unsigned i = 0; i < numArgs; ++i) {
7666 MachineOperand &Op = MI->getOperand(i+1);
7668 if (!(Op.isReg() && Op.isImplicit()))
7672 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7675 F->DeleteMachineInstr(MI);
7681 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7683 MachineBasicBlock *MBB) const {
7684 // Emit code to save XMM registers to the stack. The ABI says that the
7685 // number of registers to save is given in %al, so it's theoretically
7686 // possible to do an indirect jump trick to avoid saving all of them,
7687 // however this code takes a simpler approach and just executes all
7688 // of the stores if %al is non-zero. It's less code, and it's probably
7689 // easier on the hardware branch predictor, and stores aren't all that
7690 // expensive anyway.
7692 // Create the new basic blocks. One block contains all the XMM stores,
7693 // and one block is the final destination regardless of whether any
7694 // stores were performed.
7695 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7696 MachineFunction *F = MBB->getParent();
7697 MachineFunction::iterator MBBIter = MBB;
7699 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7700 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7701 F->insert(MBBIter, XMMSaveMBB);
7702 F->insert(MBBIter, EndMBB);
7705 // Move any original successors of MBB to the end block.
7706 EndMBB->transferSuccessors(MBB);
7707 // The original block will now fall through to the XMM save block.
7708 MBB->addSuccessor(XMMSaveMBB);
7709 // The XMMSaveMBB will fall through to the end block.
7710 XMMSaveMBB->addSuccessor(EndMBB);
7712 // Now add the instructions.
7713 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7714 DebugLoc DL = MI->getDebugLoc();
7716 unsigned CountReg = MI->getOperand(0).getReg();
7717 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7718 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7720 if (!Subtarget->isTargetWin64()) {
7721 // If %al is 0, branch around the XMM save block.
7722 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7723 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7724 MBB->addSuccessor(EndMBB);
7727 // In the XMM save block, save all the XMM argument registers.
7728 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7729 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7730 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7731 .addFrameIndex(RegSaveFrameIndex)
7732 .addImm(/*Scale=*/1)
7733 .addReg(/*IndexReg=*/0)
7734 .addImm(/*Disp=*/Offset)
7735 .addReg(/*Segment=*/0)
7736 .addReg(MI->getOperand(i).getReg())
7737 .addMemOperand(MachineMemOperand(
7738 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
7739 MachineMemOperand::MOStore, Offset,
7740 /*Size=*/16, /*Align=*/16));
7743 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7749 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7750 MachineBasicBlock *BB) const {
7751 DebugLoc dl = MI->getDebugLoc();
7752 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7753 switch (MI->getOpcode()) {
7754 default: assert(false && "Unexpected instr type to insert");
7756 case X86::CMOV_V1I64:
7757 case X86::CMOV_FR32:
7758 case X86::CMOV_FR64:
7759 case X86::CMOV_V4F32:
7760 case X86::CMOV_V2F64:
7761 case X86::CMOV_V2I64: {
7762 // To "insert" a SELECT_CC instruction, we actually have to insert the
7763 // diamond control-flow pattern. The incoming instruction knows the
7764 // destination vreg to set, the condition code register to branch on, the
7765 // true/false values to select between, and a branch opcode to use.
7766 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7767 MachineFunction::iterator It = BB;
7773 // cmpTY ccX, r1, r2
7775 // fallthrough --> copy0MBB
7776 MachineBasicBlock *thisMBB = BB;
7777 MachineFunction *F = BB->getParent();
7778 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7779 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7781 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7782 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7783 F->insert(It, copy0MBB);
7784 F->insert(It, sinkMBB);
7785 // Update machine-CFG edges by transferring all successors of the current
7786 // block to the new block which will contain the Phi node for the select.
7787 sinkMBB->transferSuccessors(BB);
7789 // Add the true and fallthrough blocks as its successors.
7790 BB->addSuccessor(copy0MBB);
7791 BB->addSuccessor(sinkMBB);
7794 // %FalseValue = ...
7795 // # fallthrough to sinkMBB
7798 // Update machine-CFG edges
7799 BB->addSuccessor(sinkMBB);
7802 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7805 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7806 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7807 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7809 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7813 case X86::FP32_TO_INT16_IN_MEM:
7814 case X86::FP32_TO_INT32_IN_MEM:
7815 case X86::FP32_TO_INT64_IN_MEM:
7816 case X86::FP64_TO_INT16_IN_MEM:
7817 case X86::FP64_TO_INT32_IN_MEM:
7818 case X86::FP64_TO_INT64_IN_MEM:
7819 case X86::FP80_TO_INT16_IN_MEM:
7820 case X86::FP80_TO_INT32_IN_MEM:
7821 case X86::FP80_TO_INT64_IN_MEM: {
7822 // Change the floating point control register to use "round towards zero"
7823 // mode when truncating to an integer value.
7824 MachineFunction *F = BB->getParent();
7825 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7826 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7828 // Load the old value of the high byte of the control word...
7830 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7831 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7834 // Set the high part to be round to zero...
7835 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7838 // Reload the modified control word now...
7839 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7841 // Restore the memory image of control word to original value
7842 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7845 // Get the X86 opcode to use.
7847 switch (MI->getOpcode()) {
7848 default: llvm_unreachable("illegal opcode!");
7849 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7850 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7851 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7852 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7853 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7854 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7855 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7856 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7857 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7861 MachineOperand &Op = MI->getOperand(0);
7863 AM.BaseType = X86AddressMode::RegBase;
7864 AM.Base.Reg = Op.getReg();
7866 AM.BaseType = X86AddressMode::FrameIndexBase;
7867 AM.Base.FrameIndex = Op.getIndex();
7869 Op = MI->getOperand(1);
7871 AM.Scale = Op.getImm();
7872 Op = MI->getOperand(2);
7874 AM.IndexReg = Op.getImm();
7875 Op = MI->getOperand(3);
7876 if (Op.isGlobal()) {
7877 AM.GV = Op.getGlobal();
7879 AM.Disp = Op.getImm();
7881 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7882 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7884 // Reload the original control word now.
7885 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7887 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7890 // String/text processing lowering.
7891 case X86::PCMPISTRM128REG:
7892 return EmitPCMP(MI, BB, 3, false /* in-mem */);
7893 case X86::PCMPISTRM128MEM:
7894 return EmitPCMP(MI, BB, 3, true /* in-mem */);
7895 case X86::PCMPESTRM128REG:
7896 return EmitPCMP(MI, BB, 5, false /* in mem */);
7897 case X86::PCMPESTRM128MEM:
7898 return EmitPCMP(MI, BB, 5, true /* in mem */);
7901 case X86::ATOMAND32:
7902 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7903 X86::AND32ri, X86::MOV32rm,
7904 X86::LCMPXCHG32, X86::MOV32rr,
7905 X86::NOT32r, X86::EAX,
7906 X86::GR32RegisterClass);
7908 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7909 X86::OR32ri, X86::MOV32rm,
7910 X86::LCMPXCHG32, X86::MOV32rr,
7911 X86::NOT32r, X86::EAX,
7912 X86::GR32RegisterClass);
7913 case X86::ATOMXOR32:
7914 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7915 X86::XOR32ri, X86::MOV32rm,
7916 X86::LCMPXCHG32, X86::MOV32rr,
7917 X86::NOT32r, X86::EAX,
7918 X86::GR32RegisterClass);
7919 case X86::ATOMNAND32:
7920 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7921 X86::AND32ri, X86::MOV32rm,
7922 X86::LCMPXCHG32, X86::MOV32rr,
7923 X86::NOT32r, X86::EAX,
7924 X86::GR32RegisterClass, true);
7925 case X86::ATOMMIN32:
7926 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7927 case X86::ATOMMAX32:
7928 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7929 case X86::ATOMUMIN32:
7930 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7931 case X86::ATOMUMAX32:
7932 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7934 case X86::ATOMAND16:
7935 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7936 X86::AND16ri, X86::MOV16rm,
7937 X86::LCMPXCHG16, X86::MOV16rr,
7938 X86::NOT16r, X86::AX,
7939 X86::GR16RegisterClass);
7941 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7942 X86::OR16ri, X86::MOV16rm,
7943 X86::LCMPXCHG16, X86::MOV16rr,
7944 X86::NOT16r, X86::AX,
7945 X86::GR16RegisterClass);
7946 case X86::ATOMXOR16:
7947 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7948 X86::XOR16ri, X86::MOV16rm,
7949 X86::LCMPXCHG16, X86::MOV16rr,
7950 X86::NOT16r, X86::AX,
7951 X86::GR16RegisterClass);
7952 case X86::ATOMNAND16:
7953 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7954 X86::AND16ri, X86::MOV16rm,
7955 X86::LCMPXCHG16, X86::MOV16rr,
7956 X86::NOT16r, X86::AX,
7957 X86::GR16RegisterClass, true);
7958 case X86::ATOMMIN16:
7959 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7960 case X86::ATOMMAX16:
7961 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7962 case X86::ATOMUMIN16:
7963 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7964 case X86::ATOMUMAX16:
7965 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7968 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7969 X86::AND8ri, X86::MOV8rm,
7970 X86::LCMPXCHG8, X86::MOV8rr,
7971 X86::NOT8r, X86::AL,
7972 X86::GR8RegisterClass);
7974 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7975 X86::OR8ri, X86::MOV8rm,
7976 X86::LCMPXCHG8, X86::MOV8rr,
7977 X86::NOT8r, X86::AL,
7978 X86::GR8RegisterClass);
7980 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7981 X86::XOR8ri, X86::MOV8rm,
7982 X86::LCMPXCHG8, X86::MOV8rr,
7983 X86::NOT8r, X86::AL,
7984 X86::GR8RegisterClass);
7985 case X86::ATOMNAND8:
7986 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7987 X86::AND8ri, X86::MOV8rm,
7988 X86::LCMPXCHG8, X86::MOV8rr,
7989 X86::NOT8r, X86::AL,
7990 X86::GR8RegisterClass, true);
7991 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7992 // This group is for 64-bit host.
7993 case X86::ATOMAND64:
7994 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7995 X86::AND64ri32, X86::MOV64rm,
7996 X86::LCMPXCHG64, X86::MOV64rr,
7997 X86::NOT64r, X86::RAX,
7998 X86::GR64RegisterClass);
8000 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8001 X86::OR64ri32, X86::MOV64rm,
8002 X86::LCMPXCHG64, X86::MOV64rr,
8003 X86::NOT64r, X86::RAX,
8004 X86::GR64RegisterClass);
8005 case X86::ATOMXOR64:
8006 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8007 X86::XOR64ri32, X86::MOV64rm,
8008 X86::LCMPXCHG64, X86::MOV64rr,
8009 X86::NOT64r, X86::RAX,
8010 X86::GR64RegisterClass);
8011 case X86::ATOMNAND64:
8012 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8013 X86::AND64ri32, X86::MOV64rm,
8014 X86::LCMPXCHG64, X86::MOV64rr,
8015 X86::NOT64r, X86::RAX,
8016 X86::GR64RegisterClass, true);
8017 case X86::ATOMMIN64:
8018 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8019 case X86::ATOMMAX64:
8020 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8021 case X86::ATOMUMIN64:
8022 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8023 case X86::ATOMUMAX64:
8024 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8026 // This group does 64-bit operations on a 32-bit host.
8027 case X86::ATOMAND6432:
8028 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8029 X86::AND32rr, X86::AND32rr,
8030 X86::AND32ri, X86::AND32ri,
8032 case X86::ATOMOR6432:
8033 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8034 X86::OR32rr, X86::OR32rr,
8035 X86::OR32ri, X86::OR32ri,
8037 case X86::ATOMXOR6432:
8038 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8039 X86::XOR32rr, X86::XOR32rr,
8040 X86::XOR32ri, X86::XOR32ri,
8042 case X86::ATOMNAND6432:
8043 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8044 X86::AND32rr, X86::AND32rr,
8045 X86::AND32ri, X86::AND32ri,
8047 case X86::ATOMADD6432:
8048 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8049 X86::ADD32rr, X86::ADC32rr,
8050 X86::ADD32ri, X86::ADC32ri,
8052 case X86::ATOMSUB6432:
8053 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8054 X86::SUB32rr, X86::SBB32rr,
8055 X86::SUB32ri, X86::SBB32ri,
8057 case X86::ATOMSWAP6432:
8058 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8059 X86::MOV32rr, X86::MOV32rr,
8060 X86::MOV32ri, X86::MOV32ri,
8062 case X86::VASTART_SAVE_XMM_REGS:
8063 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8067 //===----------------------------------------------------------------------===//
8068 // X86 Optimization Hooks
8069 //===----------------------------------------------------------------------===//
8071 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8075 const SelectionDAG &DAG,
8076 unsigned Depth) const {
8077 unsigned Opc = Op.getOpcode();
8078 assert((Opc >= ISD::BUILTIN_OP_END ||
8079 Opc == ISD::INTRINSIC_WO_CHAIN ||
8080 Opc == ISD::INTRINSIC_W_CHAIN ||
8081 Opc == ISD::INTRINSIC_VOID) &&
8082 "Should use MaskedValueIsZero if you don't know whether Op"
8083 " is a target node!");
8085 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8094 // These nodes' second result is a boolean.
8095 if (Op.getResNo() == 0)
8099 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8100 Mask.getBitWidth() - 1);
8105 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8106 /// node is a GlobalAddress + offset.
8107 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8108 GlobalValue* &GA, int64_t &Offset) const{
8109 if (N->getOpcode() == X86ISD::Wrapper) {
8110 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8111 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8112 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8116 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8119 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8120 const TargetLowering &TLI) {
8123 if (TLI.isGAPlusOffset(Base, GV, Offset))
8124 return (GV->getAlignment() >= N && (Offset % N) == 0);
8125 // DAG combine handles the stack object case.
8129 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8130 EVT EVT, LoadSDNode *&LDBase,
8131 unsigned &LastLoadedElt,
8132 SelectionDAG &DAG, MachineFrameInfo *MFI,
8133 const TargetLowering &TLI) {
8135 LastLoadedElt = -1U;
8136 for (unsigned i = 0; i < NumElems; ++i) {
8137 if (N->getMaskElt(i) < 0) {
8143 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8144 if (!Elt.getNode() ||
8145 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8148 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8150 LDBase = cast<LoadSDNode>(Elt.getNode());
8154 if (Elt.getOpcode() == ISD::UNDEF)
8157 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8158 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
8165 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8166 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8167 /// if the load addresses are consecutive, non-overlapping, and in the right
8168 /// order. In the case of v2i64, it will see if it can rewrite the
8169 /// shuffle to be an appropriate build vector so it can take advantage of
8170 // performBuildVectorCombine.
8171 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8172 const TargetLowering &TLI) {
8173 DebugLoc dl = N->getDebugLoc();
8174 EVT VT = N->getValueType(0);
8175 EVT EVT = VT.getVectorElementType();
8176 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8177 unsigned NumElems = VT.getVectorNumElements();
8179 if (VT.getSizeInBits() != 128)
8182 // Try to combine a vector_shuffle into a 128-bit load.
8183 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8184 LoadSDNode *LD = NULL;
8185 unsigned LastLoadedElt;
8186 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
8190 if (LastLoadedElt == NumElems - 1) {
8191 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8192 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8193 LD->getSrcValue(), LD->getSrcValueOffset(),
8195 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8196 LD->getSrcValue(), LD->getSrcValueOffset(),
8197 LD->isVolatile(), LD->getAlignment());
8198 } else if (NumElems == 4 && LastLoadedElt == 1) {
8199 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8200 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8201 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8202 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8207 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8208 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8209 const X86Subtarget *Subtarget) {
8210 DebugLoc DL = N->getDebugLoc();
8211 SDValue Cond = N->getOperand(0);
8212 // Get the LHS/RHS of the select.
8213 SDValue LHS = N->getOperand(1);
8214 SDValue RHS = N->getOperand(2);
8216 // If we have SSE[12] support, try to form min/max nodes.
8217 if (Subtarget->hasSSE2() &&
8218 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8219 Cond.getOpcode() == ISD::SETCC) {
8220 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8222 unsigned Opcode = 0;
8223 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8226 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8229 if (!UnsafeFPMath) break;
8231 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8233 Opcode = X86ISD::FMIN;
8236 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8239 if (!UnsafeFPMath) break;
8241 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8243 Opcode = X86ISD::FMAX;
8246 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8249 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8252 if (!UnsafeFPMath) break;
8254 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8256 Opcode = X86ISD::FMIN;
8259 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8262 if (!UnsafeFPMath) break;
8264 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8266 Opcode = X86ISD::FMAX;
8272 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8275 // If this is a select between two integer constants, try to do some
8277 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8278 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8279 // Don't do this for crazy integer types.
8280 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8281 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8282 // so that TrueC (the true value) is larger than FalseC.
8283 bool NeedsCondInvert = false;
8285 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8286 // Efficiently invertible.
8287 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8288 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8289 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8290 NeedsCondInvert = true;
8291 std::swap(TrueC, FalseC);
8294 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8295 if (FalseC->getAPIntValue() == 0 &&
8296 TrueC->getAPIntValue().isPowerOf2()) {
8297 if (NeedsCondInvert) // Invert the condition if needed.
8298 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8299 DAG.getConstant(1, Cond.getValueType()));
8301 // Zero extend the condition if needed.
8302 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8304 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8305 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8306 DAG.getConstant(ShAmt, MVT::i8));
8309 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8310 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8311 if (NeedsCondInvert) // Invert the condition if needed.
8312 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8313 DAG.getConstant(1, Cond.getValueType()));
8315 // Zero extend the condition if needed.
8316 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8317 FalseC->getValueType(0), Cond);
8318 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8319 SDValue(FalseC, 0));
8322 // Optimize cases that will turn into an LEA instruction. This requires
8323 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8324 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8325 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8326 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8328 bool isFastMultiplier = false;
8330 switch ((unsigned char)Diff) {
8332 case 1: // result = add base, cond
8333 case 2: // result = lea base( , cond*2)
8334 case 3: // result = lea base(cond, cond*2)
8335 case 4: // result = lea base( , cond*4)
8336 case 5: // result = lea base(cond, cond*4)
8337 case 8: // result = lea base( , cond*8)
8338 case 9: // result = lea base(cond, cond*8)
8339 isFastMultiplier = true;
8344 if (isFastMultiplier) {
8345 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8346 if (NeedsCondInvert) // Invert the condition if needed.
8347 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8348 DAG.getConstant(1, Cond.getValueType()));
8350 // Zero extend the condition if needed.
8351 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8353 // Scale the condition by the difference.
8355 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8356 DAG.getConstant(Diff, Cond.getValueType()));
8358 // Add the base if non-zero.
8359 if (FalseC->getAPIntValue() != 0)
8360 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8361 SDValue(FalseC, 0));
8371 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8372 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8373 TargetLowering::DAGCombinerInfo &DCI) {
8374 DebugLoc DL = N->getDebugLoc();
8376 // If the flag operand isn't dead, don't touch this CMOV.
8377 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8380 // If this is a select between two integer constants, try to do some
8381 // optimizations. Note that the operands are ordered the opposite of SELECT
8383 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8384 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8385 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8386 // larger than FalseC (the false value).
8387 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8389 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8390 CC = X86::GetOppositeBranchCondition(CC);
8391 std::swap(TrueC, FalseC);
8394 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8395 // This is efficient for any integer data type (including i8/i16) and
8397 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8398 SDValue Cond = N->getOperand(3);
8399 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8400 DAG.getConstant(CC, MVT::i8), Cond);
8402 // Zero extend the condition if needed.
8403 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8405 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8406 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8407 DAG.getConstant(ShAmt, MVT::i8));
8408 if (N->getNumValues() == 2) // Dead flag value?
8409 return DCI.CombineTo(N, Cond, SDValue());
8413 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8414 // for any integer data type, including i8/i16.
8415 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8416 SDValue Cond = N->getOperand(3);
8417 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8418 DAG.getConstant(CC, MVT::i8), Cond);
8420 // Zero extend the condition if needed.
8421 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8422 FalseC->getValueType(0), Cond);
8423 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8424 SDValue(FalseC, 0));
8426 if (N->getNumValues() == 2) // Dead flag value?
8427 return DCI.CombineTo(N, Cond, SDValue());
8431 // Optimize cases that will turn into an LEA instruction. This requires
8432 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8433 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8434 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8435 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8437 bool isFastMultiplier = false;
8439 switch ((unsigned char)Diff) {
8441 case 1: // result = add base, cond
8442 case 2: // result = lea base( , cond*2)
8443 case 3: // result = lea base(cond, cond*2)
8444 case 4: // result = lea base( , cond*4)
8445 case 5: // result = lea base(cond, cond*4)
8446 case 8: // result = lea base( , cond*8)
8447 case 9: // result = lea base(cond, cond*8)
8448 isFastMultiplier = true;
8453 if (isFastMultiplier) {
8454 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8455 SDValue Cond = N->getOperand(3);
8456 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8457 DAG.getConstant(CC, MVT::i8), Cond);
8458 // Zero extend the condition if needed.
8459 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8461 // Scale the condition by the difference.
8463 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8464 DAG.getConstant(Diff, Cond.getValueType()));
8466 // Add the base if non-zero.
8467 if (FalseC->getAPIntValue() != 0)
8468 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8469 SDValue(FalseC, 0));
8470 if (N->getNumValues() == 2) // Dead flag value?
8471 return DCI.CombineTo(N, Cond, SDValue());
8481 /// PerformMulCombine - Optimize a single multiply with constant into two
8482 /// in order to implement it with two cheaper instructions, e.g.
8483 /// LEA + SHL, LEA + LEA.
8484 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8485 TargetLowering::DAGCombinerInfo &DCI) {
8486 if (DAG.getMachineFunction().
8487 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8490 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8493 EVT VT = N->getValueType(0);
8497 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8500 uint64_t MulAmt = C->getZExtValue();
8501 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8504 uint64_t MulAmt1 = 0;
8505 uint64_t MulAmt2 = 0;
8506 if ((MulAmt % 9) == 0) {
8508 MulAmt2 = MulAmt / 9;
8509 } else if ((MulAmt % 5) == 0) {
8511 MulAmt2 = MulAmt / 5;
8512 } else if ((MulAmt % 3) == 0) {
8514 MulAmt2 = MulAmt / 3;
8517 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8518 DebugLoc DL = N->getDebugLoc();
8520 if (isPowerOf2_64(MulAmt2) &&
8521 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8522 // If second multiplifer is pow2, issue it first. We want the multiply by
8523 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8525 std::swap(MulAmt1, MulAmt2);
8528 if (isPowerOf2_64(MulAmt1))
8529 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8530 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8532 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8533 DAG.getConstant(MulAmt1, VT));
8535 if (isPowerOf2_64(MulAmt2))
8536 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8537 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8539 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8540 DAG.getConstant(MulAmt2, VT));
8542 // Do not add new nodes to DAG combiner worklist.
8543 DCI.CombineTo(N, NewMul, false);
8549 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8551 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8552 const X86Subtarget *Subtarget) {
8553 // On X86 with SSE2 support, we can transform this to a vector shift if
8554 // all elements are shifted by the same amount. We can't do this in legalize
8555 // because the a constant vector is typically transformed to a constant pool
8556 // so we have no knowledge of the shift amount.
8557 if (!Subtarget->hasSSE2())
8560 EVT VT = N->getValueType(0);
8561 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8564 SDValue ShAmtOp = N->getOperand(1);
8565 EVT EltVT = VT.getVectorElementType();
8566 DebugLoc DL = N->getDebugLoc();
8568 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8569 unsigned NumElts = VT.getVectorNumElements();
8571 for (; i != NumElts; ++i) {
8572 SDValue Arg = ShAmtOp.getOperand(i);
8573 if (Arg.getOpcode() == ISD::UNDEF) continue;
8577 for (; i != NumElts; ++i) {
8578 SDValue Arg = ShAmtOp.getOperand(i);
8579 if (Arg.getOpcode() == ISD::UNDEF) continue;
8580 if (Arg != BaseShAmt) {
8584 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8585 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8586 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8587 DAG.getIntPtrConstant(0));
8591 if (EltVT.bitsGT(MVT::i32))
8592 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8593 else if (EltVT.bitsLT(MVT::i32))
8594 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8596 // The shift amount is identical so we can do a vector shift.
8597 SDValue ValOp = N->getOperand(0);
8598 switch (N->getOpcode()) {
8600 llvm_unreachable("Unknown shift opcode!");
8603 if (VT == MVT::v2i64)
8604 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8605 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8607 if (VT == MVT::v4i32)
8608 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8609 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8611 if (VT == MVT::v8i16)
8612 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8613 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8617 if (VT == MVT::v4i32)
8618 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8619 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8621 if (VT == MVT::v8i16)
8622 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8623 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8627 if (VT == MVT::v2i64)
8628 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8629 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8631 if (VT == MVT::v4i32)
8632 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8633 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8635 if (VT == MVT::v8i16)
8636 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8637 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8644 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8645 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8646 const X86Subtarget *Subtarget) {
8647 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8648 // the FP state in cases where an emms may be missing.
8649 // A preferable solution to the general problem is to figure out the right
8650 // places to insert EMMS. This qualifies as a quick hack.
8652 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8653 StoreSDNode *St = cast<StoreSDNode>(N);
8654 EVT VT = St->getValue().getValueType();
8655 if (VT.getSizeInBits() != 64)
8658 const Function *F = DAG.getMachineFunction().getFunction();
8659 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8660 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8661 && Subtarget->hasSSE2();
8662 if ((VT.isVector() ||
8663 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8664 isa<LoadSDNode>(St->getValue()) &&
8665 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8666 St->getChain().hasOneUse() && !St->isVolatile()) {
8667 SDNode* LdVal = St->getValue().getNode();
8669 int TokenFactorIndex = -1;
8670 SmallVector<SDValue, 8> Ops;
8671 SDNode* ChainVal = St->getChain().getNode();
8672 // Must be a store of a load. We currently handle two cases: the load
8673 // is a direct child, and it's under an intervening TokenFactor. It is
8674 // possible to dig deeper under nested TokenFactors.
8675 if (ChainVal == LdVal)
8676 Ld = cast<LoadSDNode>(St->getChain());
8677 else if (St->getValue().hasOneUse() &&
8678 ChainVal->getOpcode() == ISD::TokenFactor) {
8679 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8680 if (ChainVal->getOperand(i).getNode() == LdVal) {
8681 TokenFactorIndex = i;
8682 Ld = cast<LoadSDNode>(St->getValue());
8684 Ops.push_back(ChainVal->getOperand(i));
8688 if (!Ld || !ISD::isNormalLoad(Ld))
8691 // If this is not the MMX case, i.e. we are just turning i64 load/store
8692 // into f64 load/store, avoid the transformation if there are multiple
8693 // uses of the loaded value.
8694 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8697 DebugLoc LdDL = Ld->getDebugLoc();
8698 DebugLoc StDL = N->getDebugLoc();
8699 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8700 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8702 if (Subtarget->is64Bit() || F64IsLegal) {
8703 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8704 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8705 Ld->getBasePtr(), Ld->getSrcValue(),
8706 Ld->getSrcValueOffset(), Ld->isVolatile(),
8707 Ld->getAlignment());
8708 SDValue NewChain = NewLd.getValue(1);
8709 if (TokenFactorIndex != -1) {
8710 Ops.push_back(NewChain);
8711 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8714 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8715 St->getSrcValue(), St->getSrcValueOffset(),
8716 St->isVolatile(), St->getAlignment());
8719 // Otherwise, lower to two pairs of 32-bit loads / stores.
8720 SDValue LoAddr = Ld->getBasePtr();
8721 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8722 DAG.getConstant(4, MVT::i32));
8724 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8725 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8726 Ld->isVolatile(), Ld->getAlignment());
8727 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8728 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8730 MinAlign(Ld->getAlignment(), 4));
8732 SDValue NewChain = LoLd.getValue(1);
8733 if (TokenFactorIndex != -1) {
8734 Ops.push_back(LoLd);
8735 Ops.push_back(HiLd);
8736 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8740 LoAddr = St->getBasePtr();
8741 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8742 DAG.getConstant(4, MVT::i32));
8744 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8745 St->getSrcValue(), St->getSrcValueOffset(),
8746 St->isVolatile(), St->getAlignment());
8747 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8749 St->getSrcValueOffset() + 4,
8751 MinAlign(St->getAlignment(), 4));
8752 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8757 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8758 /// X86ISD::FXOR nodes.
8759 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8760 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8761 // F[X]OR(0.0, x) -> x
8762 // F[X]OR(x, 0.0) -> x
8763 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8764 if (C->getValueAPF().isPosZero())
8765 return N->getOperand(1);
8766 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8767 if (C->getValueAPF().isPosZero())
8768 return N->getOperand(0);
8772 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8773 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8774 // FAND(0.0, x) -> 0.0
8775 // FAND(x, 0.0) -> 0.0
8776 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8777 if (C->getValueAPF().isPosZero())
8778 return N->getOperand(0);
8779 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8780 if (C->getValueAPF().isPosZero())
8781 return N->getOperand(1);
8785 static SDValue PerformBTCombine(SDNode *N,
8787 TargetLowering::DAGCombinerInfo &DCI) {
8788 // BT ignores high bits in the bit index operand.
8789 SDValue Op1 = N->getOperand(1);
8790 if (Op1.hasOneUse()) {
8791 unsigned BitWidth = Op1.getValueSizeInBits();
8792 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8793 APInt KnownZero, KnownOne;
8794 TargetLowering::TargetLoweringOpt TLO(DAG);
8795 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8796 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8797 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8798 DCI.CommitTargetLoweringOpt(TLO);
8803 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8804 SDValue Op = N->getOperand(0);
8805 if (Op.getOpcode() == ISD::BIT_CONVERT)
8806 Op = Op.getOperand(0);
8807 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
8808 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8809 VT.getVectorElementType().getSizeInBits() ==
8810 OpVT.getVectorElementType().getSizeInBits()) {
8811 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8816 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8817 // Locked instructions, in turn, have implicit fence semantics (all memory
8818 // operations are flushed before issuing the locked instruction, and the
8819 // are not buffered), so we can fold away the common pattern of
8820 // fence-atomic-fence.
8821 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8822 SDValue atomic = N->getOperand(0);
8823 switch (atomic.getOpcode()) {
8824 case ISD::ATOMIC_CMP_SWAP:
8825 case ISD::ATOMIC_SWAP:
8826 case ISD::ATOMIC_LOAD_ADD:
8827 case ISD::ATOMIC_LOAD_SUB:
8828 case ISD::ATOMIC_LOAD_AND:
8829 case ISD::ATOMIC_LOAD_OR:
8830 case ISD::ATOMIC_LOAD_XOR:
8831 case ISD::ATOMIC_LOAD_NAND:
8832 case ISD::ATOMIC_LOAD_MIN:
8833 case ISD::ATOMIC_LOAD_MAX:
8834 case ISD::ATOMIC_LOAD_UMIN:
8835 case ISD::ATOMIC_LOAD_UMAX:
8841 SDValue fence = atomic.getOperand(0);
8842 if (fence.getOpcode() != ISD::MEMBARRIER)
8845 switch (atomic.getOpcode()) {
8846 case ISD::ATOMIC_CMP_SWAP:
8847 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8848 atomic.getOperand(1), atomic.getOperand(2),
8849 atomic.getOperand(3));
8850 case ISD::ATOMIC_SWAP:
8851 case ISD::ATOMIC_LOAD_ADD:
8852 case ISD::ATOMIC_LOAD_SUB:
8853 case ISD::ATOMIC_LOAD_AND:
8854 case ISD::ATOMIC_LOAD_OR:
8855 case ISD::ATOMIC_LOAD_XOR:
8856 case ISD::ATOMIC_LOAD_NAND:
8857 case ISD::ATOMIC_LOAD_MIN:
8858 case ISD::ATOMIC_LOAD_MAX:
8859 case ISD::ATOMIC_LOAD_UMIN:
8860 case ISD::ATOMIC_LOAD_UMAX:
8861 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8862 atomic.getOperand(1), atomic.getOperand(2));
8868 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8869 DAGCombinerInfo &DCI) const {
8870 SelectionDAG &DAG = DCI.DAG;
8871 switch (N->getOpcode()) {
8873 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8874 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8875 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8876 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8879 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8880 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8882 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8883 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8884 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8885 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8886 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8892 //===----------------------------------------------------------------------===//
8893 // X86 Inline Assembly Support
8894 //===----------------------------------------------------------------------===//
8896 static bool LowerToBSwap(CallInst *CI) {
8897 // FIXME: this should verify that we are targetting a 486 or better. If not,
8898 // we will turn this bswap into something that will be lowered to logical ops
8899 // instead of emitting the bswap asm. For now, we don't support 486 or lower
8900 // so don't worry about this.
8902 // Verify this is a simple bswap.
8903 if (CI->getNumOperands() != 2 ||
8904 CI->getType() != CI->getOperand(1)->getType() ||
8905 !CI->getType()->isInteger())
8908 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8909 if (!Ty || Ty->getBitWidth() % 16 != 0)
8912 // Okay, we can do this xform, do so now.
8913 const Type *Tys[] = { Ty };
8914 Module *M = CI->getParent()->getParent()->getParent();
8915 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
8917 Value *Op = CI->getOperand(1);
8918 Op = CallInst::Create(Int, Op, CI->getName(), CI);
8920 CI->replaceAllUsesWith(Op);
8921 CI->eraseFromParent();
8925 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
8926 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8927 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
8929 std::string AsmStr = IA->getAsmString();
8931 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
8932 std::vector<std::string> AsmPieces;
8933 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
8935 switch (AsmPieces.size()) {
8936 default: return false;
8938 AsmStr = AsmPieces[0];
8940 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
8943 if (AsmPieces.size() == 2 &&
8944 (AsmPieces[0] == "bswap" ||
8945 AsmPieces[0] == "bswapq" ||
8946 AsmPieces[0] == "bswapl") &&
8947 (AsmPieces[1] == "$0" ||
8948 AsmPieces[1] == "${0:q}")) {
8949 // No need to check constraints, nothing other than the equivalent of
8950 // "=r,0" would be valid here.
8951 return LowerToBSwap(CI);
8953 // rorw $$8, ${0:w} --> llvm.bswap.i16
8954 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
8955 AsmPieces.size() == 3 &&
8956 AsmPieces[0] == "rorw" &&
8957 AsmPieces[1] == "$$8," &&
8958 AsmPieces[2] == "${0:w}" &&
8959 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
8960 return LowerToBSwap(CI);
8964 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
8965 Constraints.size() >= 2 &&
8966 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
8967 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
8968 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
8969 std::vector<std::string> Words;
8970 SplitString(AsmPieces[0], Words, " \t");
8971 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
8973 SplitString(AsmPieces[1], Words, " \t");
8974 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
8976 SplitString(AsmPieces[2], Words, " \t,");
8977 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
8978 Words[2] == "%edx") {
8979 return LowerToBSwap(CI);
8991 /// getConstraintType - Given a constraint letter, return the type of
8992 /// constraint it is for this target.
8993 X86TargetLowering::ConstraintType
8994 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8995 if (Constraint.size() == 1) {
8996 switch (Constraint[0]) {
9008 return C_RegisterClass;
9016 return TargetLowering::getConstraintType(Constraint);
9019 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9020 /// with another that has more specific requirements based on the type of the
9021 /// corresponding operand.
9022 const char *X86TargetLowering::
9023 LowerXConstraint(EVT ConstraintVT) const {
9024 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9025 // 'f' like normal targets.
9026 if (ConstraintVT.isFloatingPoint()) {
9027 if (Subtarget->hasSSE2())
9029 if (Subtarget->hasSSE1())
9033 return TargetLowering::LowerXConstraint(ConstraintVT);
9036 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9037 /// vector. If it is invalid, don't add anything to Ops.
9038 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9041 std::vector<SDValue>&Ops,
9042 SelectionDAG &DAG) const {
9043 SDValue Result(0, 0);
9045 switch (Constraint) {
9048 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9049 if (C->getZExtValue() <= 31) {
9050 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9056 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9057 if (C->getZExtValue() <= 63) {
9058 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9064 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9065 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9066 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9072 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9073 if (C->getZExtValue() <= 255) {
9074 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9080 // 32-bit signed value
9081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9082 const ConstantInt *CI = C->getConstantIntValue();
9083 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9084 C->getSExtValue())) {
9085 // Widen to 64 bits here to get it sign extended.
9086 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9089 // FIXME gcc accepts some relocatable values here too, but only in certain
9090 // memory models; it's complicated.
9095 // 32-bit unsigned value
9096 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9097 const ConstantInt *CI = C->getConstantIntValue();
9098 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9099 C->getZExtValue())) {
9100 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9104 // FIXME gcc accepts some relocatable values here too, but only in certain
9105 // memory models; it's complicated.
9109 // Literal immediates are always ok.
9110 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9111 // Widen to 64 bits here to get it sign extended.
9112 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9116 // If we are in non-pic codegen mode, we allow the address of a global (with
9117 // an optional displacement) to be used with 'i'.
9118 GlobalAddressSDNode *GA = 0;
9121 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9123 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9124 Offset += GA->getOffset();
9126 } else if (Op.getOpcode() == ISD::ADD) {
9127 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9128 Offset += C->getZExtValue();
9129 Op = Op.getOperand(0);
9132 } else if (Op.getOpcode() == ISD::SUB) {
9133 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9134 Offset += -C->getZExtValue();
9135 Op = Op.getOperand(0);
9140 // Otherwise, this isn't something we can handle, reject it.
9144 GlobalValue *GV = GA->getGlobal();
9145 // If we require an extra load to get this address, as in PIC mode, we
9147 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9148 getTargetMachine())))
9152 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9154 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9160 if (Result.getNode()) {
9161 Ops.push_back(Result);
9164 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9168 std::vector<unsigned> X86TargetLowering::
9169 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9171 if (Constraint.size() == 1) {
9172 // FIXME: not handling fp-stack yet!
9173 switch (Constraint[0]) { // GCC X86 Constraint Letters
9174 default: break; // Unknown constraint letter
9175 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9176 if (Subtarget->is64Bit()) {
9178 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9179 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9180 X86::R10D,X86::R11D,X86::R12D,
9181 X86::R13D,X86::R14D,X86::R15D,
9182 X86::EBP, X86::ESP, 0);
9183 else if (VT == MVT::i16)
9184 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9185 X86::SI, X86::DI, X86::R8W,X86::R9W,
9186 X86::R10W,X86::R11W,X86::R12W,
9187 X86::R13W,X86::R14W,X86::R15W,
9188 X86::BP, X86::SP, 0);
9189 else if (VT == MVT::i8)
9190 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9191 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9192 X86::R10B,X86::R11B,X86::R12B,
9193 X86::R13B,X86::R14B,X86::R15B,
9194 X86::BPL, X86::SPL, 0);
9196 else if (VT == MVT::i64)
9197 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9198 X86::RSI, X86::RDI, X86::R8, X86::R9,
9199 X86::R10, X86::R11, X86::R12,
9200 X86::R13, X86::R14, X86::R15,
9201 X86::RBP, X86::RSP, 0);
9205 // 32-bit fallthrough
9208 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9209 else if (VT == MVT::i16)
9210 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9211 else if (VT == MVT::i8)
9212 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9213 else if (VT == MVT::i64)
9214 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9219 return std::vector<unsigned>();
9222 std::pair<unsigned, const TargetRegisterClass*>
9223 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9225 // First, see if this is a constraint that directly corresponds to an LLVM
9227 if (Constraint.size() == 1) {
9228 // GCC Constraint Letters
9229 switch (Constraint[0]) {
9231 case 'r': // GENERAL_REGS
9232 case 'R': // LEGACY_REGS
9233 case 'l': // INDEX_REGS
9235 return std::make_pair(0U, X86::GR8RegisterClass);
9237 return std::make_pair(0U, X86::GR16RegisterClass);
9238 if (VT == MVT::i32 || !Subtarget->is64Bit())
9239 return std::make_pair(0U, X86::GR32RegisterClass);
9240 return std::make_pair(0U, X86::GR64RegisterClass);
9241 case 'f': // FP Stack registers.
9242 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9243 // value to the correct fpstack register class.
9244 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9245 return std::make_pair(0U, X86::RFP32RegisterClass);
9246 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9247 return std::make_pair(0U, X86::RFP64RegisterClass);
9248 return std::make_pair(0U, X86::RFP80RegisterClass);
9249 case 'y': // MMX_REGS if MMX allowed.
9250 if (!Subtarget->hasMMX()) break;
9251 return std::make_pair(0U, X86::VR64RegisterClass);
9252 case 'Y': // SSE_REGS if SSE2 allowed
9253 if (!Subtarget->hasSSE2()) break;
9255 case 'x': // SSE_REGS if SSE1 allowed
9256 if (!Subtarget->hasSSE1()) break;
9258 switch (VT.getSimpleVT().SimpleTy) {
9260 // Scalar SSE types.
9263 return std::make_pair(0U, X86::FR32RegisterClass);
9266 return std::make_pair(0U, X86::FR64RegisterClass);
9274 return std::make_pair(0U, X86::VR128RegisterClass);
9280 // Use the default implementation in TargetLowering to convert the register
9281 // constraint into a member of a register class.
9282 std::pair<unsigned, const TargetRegisterClass*> Res;
9283 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9285 // Not found as a standard register?
9286 if (Res.second == 0) {
9287 // GCC calls "st(0)" just plain "st".
9288 if (StringsEqualNoCase("{st}", Constraint)) {
9289 Res.first = X86::ST0;
9290 Res.second = X86::RFP80RegisterClass;
9292 // 'A' means EAX + EDX.
9293 if (Constraint == "A") {
9294 Res.first = X86::EAX;
9295 Res.second = X86::GR32_ADRegisterClass;
9300 // Otherwise, check to see if this is a register class of the wrong value
9301 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9302 // turn into {ax},{dx}.
9303 if (Res.second->hasType(VT))
9304 return Res; // Correct type already, nothing to do.
9306 // All of the single-register GCC register classes map their values onto
9307 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9308 // really want an 8-bit or 32-bit register, map to the appropriate register
9309 // class and return the appropriate register.
9310 if (Res.second == X86::GR16RegisterClass) {
9311 if (VT == MVT::i8) {
9312 unsigned DestReg = 0;
9313 switch (Res.first) {
9315 case X86::AX: DestReg = X86::AL; break;
9316 case X86::DX: DestReg = X86::DL; break;
9317 case X86::CX: DestReg = X86::CL; break;
9318 case X86::BX: DestReg = X86::BL; break;
9321 Res.first = DestReg;
9322 Res.second = X86::GR8RegisterClass;
9324 } else if (VT == MVT::i32) {
9325 unsigned DestReg = 0;
9326 switch (Res.first) {
9328 case X86::AX: DestReg = X86::EAX; break;
9329 case X86::DX: DestReg = X86::EDX; break;
9330 case X86::CX: DestReg = X86::ECX; break;
9331 case X86::BX: DestReg = X86::EBX; break;
9332 case X86::SI: DestReg = X86::ESI; break;
9333 case X86::DI: DestReg = X86::EDI; break;
9334 case X86::BP: DestReg = X86::EBP; break;
9335 case X86::SP: DestReg = X86::ESP; break;
9338 Res.first = DestReg;
9339 Res.second = X86::GR32RegisterClass;
9341 } else if (VT == MVT::i64) {
9342 unsigned DestReg = 0;
9343 switch (Res.first) {
9345 case X86::AX: DestReg = X86::RAX; break;
9346 case X86::DX: DestReg = X86::RDX; break;
9347 case X86::CX: DestReg = X86::RCX; break;
9348 case X86::BX: DestReg = X86::RBX; break;
9349 case X86::SI: DestReg = X86::RSI; break;
9350 case X86::DI: DestReg = X86::RDI; break;
9351 case X86::BP: DestReg = X86::RBP; break;
9352 case X86::SP: DestReg = X86::RSP; break;
9355 Res.first = DestReg;
9356 Res.second = X86::GR64RegisterClass;
9359 } else if (Res.second == X86::FR32RegisterClass ||
9360 Res.second == X86::FR64RegisterClass ||
9361 Res.second == X86::VR128RegisterClass) {
9362 // Handle references to XMM physical registers that got mapped into the
9363 // wrong class. This can happen with constraints like {xmm0} where the
9364 // target independent register mapper will just pick the first match it can
9365 // find, ignoring the required type.
9367 Res.second = X86::FR32RegisterClass;
9368 else if (VT == MVT::f64)
9369 Res.second = X86::FR64RegisterClass;
9370 else if (X86::VR128RegisterClass->hasType(VT))
9371 Res.second = X86::VR128RegisterClass;
9377 //===----------------------------------------------------------------------===//
9378 // X86 Widen vector type
9379 //===----------------------------------------------------------------------===//
9381 /// getWidenVectorType: given a vector type, returns the type to widen
9382 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9383 /// If there is no vector type that we want to widen to, returns MVT::Other
9384 /// When and where to widen is target dependent based on the cost of
9385 /// scalarizing vs using the wider vector type.
9387 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9388 assert(VT.isVector());
9389 if (isTypeLegal(VT))
9392 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9393 // type based on element type. This would speed up our search (though
9394 // it may not be worth it since the size of the list is relatively
9396 EVT EltVT = VT.getVectorElementType();
9397 unsigned NElts = VT.getVectorNumElements();
9399 // On X86, it make sense to widen any vector wider than 1
9403 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9404 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9405 EVT SVT = (MVT::SimpleValueType)nVT;
9407 if (isTypeLegal(SVT) &&
9408 SVT.getVectorElementType() == EltVT &&
9409 SVT.getVectorNumElements() > NElts)