1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "X86TargetObjectFile.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/MC/MCAsmInfo.h"
37 #include "llvm/MC/MCContext.h"
38 #include "llvm/MC/MCExpr.h"
39 #include "llvm/MC/MCSymbol.h"
40 #include "llvm/ADT/BitVector.h"
41 #include "llvm/ADT/SmallSet.h"
42 #include "llvm/ADT/StringExtras.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
54 // Disable16Bit - 16-bit operations typically have a larger encoding than
55 // corresponding 32-bit instructions, and 16-bit code is slow on some
56 // processors. This is an experimental flag to disable 16-bit operations
57 // (which forces them to be Legalized to 32-bit operations).
59 Disable16Bit("disable-16bit", cl::Hidden,
60 cl::desc("Disable use of 16-bit instructions"));
62 // Forward declarations.
63 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
66 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
67 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
68 default: llvm_unreachable("unknown subtarget type");
69 case X86Subtarget::isDarwin:
70 if (TM.getSubtarget<X86Subtarget>().is64Bit())
71 return new X8664_MachoTargetObjectFile();
72 return new X8632_MachoTargetObjectFile();
73 case X86Subtarget::isELF:
74 return new TargetLoweringObjectFileELF();
75 case X86Subtarget::isMingw:
76 case X86Subtarget::isCygwin:
77 case X86Subtarget::isWindows:
78 return new TargetLoweringObjectFileCOFF();
83 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
84 : TargetLowering(TM, createTLOF(TM)) {
85 Subtarget = &TM.getSubtarget<X86Subtarget>();
86 X86ScalarSSEf64 = Subtarget->hasSSE2();
87 X86ScalarSSEf32 = Subtarget->hasSSE1();
88 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
90 RegInfo = TM.getRegisterInfo();
93 // Set up the TargetLowering object.
95 // X86 is weird, it always uses i8 for shift amounts and setcc results.
96 setShiftAmountType(MVT::i8);
97 setBooleanContents(ZeroOrOneBooleanContent);
98 setSchedulingPreference(SchedulingForRegPressure);
99 setStackPointerRegisterToSaveRestore(X86StackPtr);
101 if (Subtarget->isTargetDarwin()) {
102 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
103 setUseUnderscoreSetJmp(false);
104 setUseUnderscoreLongJmp(false);
105 } else if (Subtarget->isTargetMingw()) {
106 // MS runtime is weird: it exports _setjmp, but longjmp!
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(false);
110 setUseUnderscoreSetJmp(true);
111 setUseUnderscoreLongJmp(true);
114 // Set up the register classes.
115 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
117 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
118 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
119 if (Subtarget->is64Bit())
120 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
122 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
124 // We don't accept any truncstore of integer registers.
125 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
127 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
128 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
130 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
131 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
132 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
134 // SETOEQ and SETUNE require checking two conditions.
135 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
136 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
137 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
138 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
139 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
140 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
142 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
144 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
145 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
148 if (Subtarget->is64Bit()) {
149 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
151 } else if (!UseSoftFloat) {
152 if (X86ScalarSSEf64) {
153 // We have an impenetrably clever algorithm for ui64->double only.
154 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
156 // We have an algorithm for SSE2, and we turn this into a 64-bit
157 // FILD for other targets.
158 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
161 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
163 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
164 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
167 // SSE has no i16 to fp conversion, only i32
168 if (X86ScalarSSEf32) {
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
170 // f32 and f64 cases are Legal, f80 case is not
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
181 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
182 // are Legal, f80 is custom lowered.
183 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
186 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
188 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
191 if (X86ScalarSSEf32) {
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
193 // f32 and f64 cases are Legal, f80 case is not
194 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
197 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
200 // Handle FP_TO_UINT by promoting the destination to a larger signed
202 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
203 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
206 if (Subtarget->is64Bit()) {
207 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
209 } else if (!UseSoftFloat) {
210 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
211 // Expand FP_TO_UINT into a select.
212 // FIXME: We would like to use a Custom expander here eventually to do
213 // the optimal thing for SSE vs. the default expansion in the legalizer.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
216 // With SSE3 we can use fisttpll to convert to a signed i64; without
217 // SSE, we're stuck with a fistpll.
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
221 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
222 if (!X86ScalarSSEf64) {
223 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
227 // Scalar integer divide and remainder are lowered to use operations that
228 // produce two results, to match the available instructions. This exposes
229 // the two-result form to trivial CSE, which is able to combine x/y and x%y
230 // into a single instruction.
232 // Scalar integer multiply-high is also lowered to use two-result
233 // operations, to match the available instructions. However, plain multiply
234 // (low) operations are left as Legal, as there are single-result
235 // instructions for this in x86. Using the two-result multiply instructions
236 // when both high and low results are needed must be arranged by dagcombine.
237 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
238 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
239 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
240 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::SREM , MVT::i8 , Expand);
242 setOperationAction(ISD::UREM , MVT::i8 , Expand);
243 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
244 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
245 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
246 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::SREM , MVT::i16 , Expand);
248 setOperationAction(ISD::UREM , MVT::i16 , Expand);
249 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
250 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
251 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
252 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::SREM , MVT::i32 , Expand);
254 setOperationAction(ISD::UREM , MVT::i32 , Expand);
255 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
256 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
257 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
258 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::SREM , MVT::i64 , Expand);
260 setOperationAction(ISD::UREM , MVT::i64 , Expand);
262 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
263 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
264 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
265 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
266 if (Subtarget->is64Bit())
267 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
271 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
272 setOperationAction(ISD::FREM , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f64 , Expand);
274 setOperationAction(ISD::FREM , MVT::f80 , Expand);
275 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
277 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
278 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
279 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
283 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
288 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
291 if (Subtarget->is64Bit()) {
292 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
293 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
294 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
297 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
298 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
300 // These should be promoted to a larger select which is supported.
301 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
302 // X86 wants to expand cmov itself.
303 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
305 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
307 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
308 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
309 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
310 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
311 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
312 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
314 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
316 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
317 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
318 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
319 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
320 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
321 if (Subtarget->is64Bit()) {
322 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
323 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
325 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
328 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
329 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
330 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
331 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
332 if (Subtarget->is64Bit())
333 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
334 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
335 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
336 if (Subtarget->is64Bit()) {
337 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
338 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
339 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
340 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
341 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
343 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
344 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
345 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
346 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
347 if (Subtarget->is64Bit()) {
348 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
349 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
350 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
353 if (Subtarget->hasSSE1())
354 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
356 if (!Subtarget->hasSSE2())
357 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
359 // Expand certain atomics
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
361 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
362 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
363 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
370 if (!Subtarget->is64Bit()) {
371 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
380 // FIXME - use subtarget debug flags
381 if (!Subtarget->isTargetDarwin() &&
382 !Subtarget->isTargetELF() &&
383 !Subtarget->isTargetCygMing()) {
384 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
387 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
388 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
389 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
390 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
391 if (Subtarget->is64Bit()) {
392 setExceptionPointerRegister(X86::RAX);
393 setExceptionSelectorRegister(X86::RDX);
395 setExceptionPointerRegister(X86::EAX);
396 setExceptionSelectorRegister(X86::EDX);
398 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
399 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
401 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
403 setOperationAction(ISD::TRAP, MVT::Other, Legal);
405 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
406 setOperationAction(ISD::VASTART , MVT::Other, Custom);
407 setOperationAction(ISD::VAEND , MVT::Other, Expand);
408 if (Subtarget->is64Bit()) {
409 setOperationAction(ISD::VAARG , MVT::Other, Custom);
410 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
412 setOperationAction(ISD::VAARG , MVT::Other, Expand);
413 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
416 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
417 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
418 if (Subtarget->is64Bit())
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
420 if (Subtarget->isTargetCygMing())
421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
423 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
425 if (!UseSoftFloat && X86ScalarSSEf64) {
426 // f32 and f64 use SSE.
427 // Set up the FP register classes.
428 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
429 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
431 // Use ANDPD to simulate FABS.
432 setOperationAction(ISD::FABS , MVT::f64, Custom);
433 setOperationAction(ISD::FABS , MVT::f32, Custom);
435 // Use XORP to simulate FNEG.
436 setOperationAction(ISD::FNEG , MVT::f64, Custom);
437 setOperationAction(ISD::FNEG , MVT::f32, Custom);
439 // Use ANDPD and ORPD to simulate FCOPYSIGN.
440 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
441 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
443 // We don't support sin/cos/fmod
444 setOperationAction(ISD::FSIN , MVT::f64, Expand);
445 setOperationAction(ISD::FCOS , MVT::f64, Expand);
446 setOperationAction(ISD::FSIN , MVT::f32, Expand);
447 setOperationAction(ISD::FCOS , MVT::f32, Expand);
449 // Expand FP immediates into loads from the stack, except for the special
451 addLegalFPImmediate(APFloat(+0.0)); // xorpd
452 addLegalFPImmediate(APFloat(+0.0f)); // xorps
453 } else if (!UseSoftFloat && X86ScalarSSEf32) {
454 // Use SSE for f32, x87 for f64.
455 // Set up the FP register classes.
456 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
457 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
459 // Use ANDPS to simulate FABS.
460 setOperationAction(ISD::FABS , MVT::f32, Custom);
462 // Use XORP to simulate FNEG.
463 setOperationAction(ISD::FNEG , MVT::f32, Custom);
465 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
467 // Use ANDPS and ORPS to simulate FCOPYSIGN.
468 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
471 // We don't support sin/cos/fmod
472 setOperationAction(ISD::FSIN , MVT::f32, Expand);
473 setOperationAction(ISD::FCOS , MVT::f32, Expand);
475 // Special cases we handle for FP constants.
476 addLegalFPImmediate(APFloat(+0.0f)); // xorps
477 addLegalFPImmediate(APFloat(+0.0)); // FLD0
478 addLegalFPImmediate(APFloat(+1.0)); // FLD1
479 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
480 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
483 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
484 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
486 } else if (!UseSoftFloat) {
487 // f32 and f64 in x87.
488 // Set up the FP register classes.
489 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
490 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
492 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
493 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
494 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
495 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
498 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
499 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
501 addLegalFPImmediate(APFloat(+0.0)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
505 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
506 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
507 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
508 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
511 // Long double always uses X87.
513 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
514 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
515 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
518 APFloat TmpFlt(+0.0);
519 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
521 addLegalFPImmediate(TmpFlt); // FLD0
523 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
524 APFloat TmpFlt2(+1.0);
525 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
527 addLegalFPImmediate(TmpFlt2); // FLD1
528 TmpFlt2.changeSign();
529 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
533 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
534 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
538 // Always use a library call for pow.
539 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
540 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
541 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
543 setOperationAction(ISD::FLOG, MVT::f80, Expand);
544 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
545 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
546 setOperationAction(ISD::FEXP, MVT::f80, Expand);
547 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
549 // First set operation action for all vector types to either promote
550 // (for widening) or expand (for scalarization). Then we will selectively
551 // turn on ones that can be effectively codegen'd.
552 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
553 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
554 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
569 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
570 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
603 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
607 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
608 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
609 setTruncStoreAction((MVT::SimpleValueType)VT,
610 (MVT::SimpleValueType)InnerVT, Expand);
611 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
616 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
617 // with -msoft-float, disable use of MMX as well.
618 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
619 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
620 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
621 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
622 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
623 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
625 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
626 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
627 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
628 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
630 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
631 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
632 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
633 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
635 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
636 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
638 setOperationAction(ISD::AND, MVT::v8i8, Promote);
639 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v4i16, Promote);
641 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
642 setOperationAction(ISD::AND, MVT::v2i32, Promote);
643 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
644 setOperationAction(ISD::AND, MVT::v1i64, Legal);
646 setOperationAction(ISD::OR, MVT::v8i8, Promote);
647 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v4i16, Promote);
649 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
650 setOperationAction(ISD::OR, MVT::v2i32, Promote);
651 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
652 setOperationAction(ISD::OR, MVT::v1i64, Legal);
654 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
657 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
658 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
659 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
660 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
662 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
667 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
668 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
669 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
670 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
673 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
674 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
675 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
676 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
679 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
680 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
681 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
683 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
684 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
685 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
686 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
688 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
690 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
691 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
692 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
693 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
694 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
695 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
696 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector()) {
798 setOperationAction(ISD::AND, SVT, Promote);
799 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
800 setOperationAction(ISD::OR, SVT, Promote);
801 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
802 setOperationAction(ISD::XOR, SVT, Promote);
803 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
804 setOperationAction(ISD::LOAD, SVT, Promote);
805 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
806 setOperationAction(ISD::SELECT, SVT, Promote);
807 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
810 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
812 // Custom lower v2i64 and v2f64 selects.
813 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
814 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
815 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
816 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
818 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
819 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
820 if (!DisableMMX && Subtarget->hasMMX()) {
821 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
822 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
826 if (Subtarget->hasSSE41()) {
827 // FIXME: Do we need to handle scalar-to-vector here?
828 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
830 // i8 and i16 vectors are custom , because the source register and source
831 // source memory operand types are not the same width. f32 vectors are
832 // custom since the immediate controlling the insert encodes additional
834 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
835 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
837 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
839 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
840 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
841 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
842 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
844 if (Subtarget->is64Bit()) {
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
850 if (Subtarget->hasSSE42()) {
851 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
854 if (!UseSoftFloat && Subtarget->hasAVX()) {
855 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
856 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
857 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
858 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
860 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
861 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
862 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
863 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
864 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
865 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
866 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
867 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
868 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
869 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
870 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
871 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
872 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
873 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
874 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
876 // Operations to consider commented out -v16i16 v32i8
877 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
878 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
879 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
880 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
881 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
882 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
883 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
884 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
885 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
886 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
887 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
888 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
889 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
890 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
892 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
893 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
894 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
895 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
897 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
898 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
899 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
900 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
901 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
903 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
904 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
905 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
906 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
907 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
908 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
911 // Not sure we want to do this since there are no 256-bit integer
914 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
915 // This includes 256-bit vectors
916 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
917 EVT VT = (MVT::SimpleValueType)i;
919 // Do not attempt to custom lower non-power-of-2 vectors
920 if (!isPowerOf2_32(VT.getVectorNumElements()))
923 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
924 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
925 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
928 if (Subtarget->is64Bit()) {
929 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
930 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
935 // Not sure we want to do this since there are no 256-bit integer
938 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
939 // Including 256-bit vectors
940 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
941 EVT VT = (MVT::SimpleValueType)i;
943 if (!VT.is256BitVector()) {
946 setOperationAction(ISD::AND, VT, Promote);
947 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
948 setOperationAction(ISD::OR, VT, Promote);
949 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
950 setOperationAction(ISD::XOR, VT, Promote);
951 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
952 setOperationAction(ISD::LOAD, VT, Promote);
953 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
954 setOperationAction(ISD::SELECT, VT, Promote);
955 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
958 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
962 // We want to custom lower some of our intrinsics.
963 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
965 // Add/Sub/Mul with overflow operations are custom lowered.
966 setOperationAction(ISD::SADDO, MVT::i32, Custom);
967 setOperationAction(ISD::SADDO, MVT::i64, Custom);
968 setOperationAction(ISD::UADDO, MVT::i32, Custom);
969 setOperationAction(ISD::UADDO, MVT::i64, Custom);
970 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
971 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
972 setOperationAction(ISD::USUBO, MVT::i32, Custom);
973 setOperationAction(ISD::USUBO, MVT::i64, Custom);
974 setOperationAction(ISD::SMULO, MVT::i32, Custom);
975 setOperationAction(ISD::SMULO, MVT::i64, Custom);
977 if (!Subtarget->is64Bit()) {
978 // These libcalls are not available in 32-bit.
979 setLibcallName(RTLIB::SHL_I128, 0);
980 setLibcallName(RTLIB::SRL_I128, 0);
981 setLibcallName(RTLIB::SRA_I128, 0);
984 // We have target-specific dag combine patterns for the following nodes:
985 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
986 setTargetDAGCombine(ISD::BUILD_VECTOR);
987 setTargetDAGCombine(ISD::SELECT);
988 setTargetDAGCombine(ISD::SHL);
989 setTargetDAGCombine(ISD::SRA);
990 setTargetDAGCombine(ISD::SRL);
991 setTargetDAGCombine(ISD::OR);
992 setTargetDAGCombine(ISD::STORE);
993 setTargetDAGCombine(ISD::MEMBARRIER);
994 setTargetDAGCombine(ISD::ZERO_EXTEND);
995 if (Subtarget->is64Bit())
996 setTargetDAGCombine(ISD::MUL);
998 computeRegisterProperties();
1000 // Divide and reminder operations have no vector equivalent and can
1001 // trap. Do a custom widening for these operations in which we never
1002 // generate more divides/remainder than the original vector width.
1003 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1004 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
1005 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
1006 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
1007 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
1008 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
1009 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1013 // FIXME: These should be based on subtarget info. Plus, the values should
1014 // be smaller when we are in optimizing for size mode.
1015 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1016 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1017 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1018 setPrefLoopAlignment(16);
1019 benefitFromCodePlacementOpt = true;
1023 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1028 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1029 /// the desired ByVal argument alignment.
1030 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1033 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1034 if (VTy->getBitWidth() == 128)
1036 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1037 unsigned EltAlign = 0;
1038 getMaxByValAlign(ATy->getElementType(), EltAlign);
1039 if (EltAlign > MaxAlign)
1040 MaxAlign = EltAlign;
1041 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1042 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1043 unsigned EltAlign = 0;
1044 getMaxByValAlign(STy->getElementType(i), EltAlign);
1045 if (EltAlign > MaxAlign)
1046 MaxAlign = EltAlign;
1054 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1055 /// function arguments in the caller parameter area. For X86, aggregates
1056 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1057 /// are at 4-byte boundaries.
1058 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1059 if (Subtarget->is64Bit()) {
1060 // Max of 8 and alignment of type.
1061 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1068 if (Subtarget->hasSSE1())
1069 getMaxByValAlign(Ty, Align);
1073 /// getOptimalMemOpType - Returns the target specific optimal type for load
1074 /// and store operations as a result of memset, memcpy, and memmove
1075 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1078 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1079 bool isSrcConst, bool isSrcStr,
1080 SelectionDAG &DAG) const {
1081 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1082 // linux. This is because the stack realignment code can't handle certain
1083 // cases like PR2962. This should be removed when PR2962 is fixed.
1084 const Function *F = DAG.getMachineFunction().getFunction();
1085 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1086 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1087 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1089 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1092 if (Subtarget->is64Bit() && Size >= 8)
1097 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1098 /// current function. The returned value is a member of the
1099 /// MachineJumpTableInfo::JTEntryKind enum.
1100 unsigned X86TargetLowering::getJumpTableEncoding() const {
1101 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1103 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1104 Subtarget->isPICStyleGOT())
1105 return MachineJumpTableInfo::EK_Custom32;
1107 // Otherwise, use the normal jump table encoding heuristics.
1108 return TargetLowering::getJumpTableEncoding();
1111 /// getPICBaseSymbol - Return the X86-32 PIC base.
1113 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1114 MCContext &Ctx) const {
1115 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1116 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1117 Twine(MF->getFunctionNumber())+"$pb");
1122 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1123 const MachineBasicBlock *MBB,
1124 unsigned uid,MCContext &Ctx) const{
1125 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1126 Subtarget->isPICStyleGOT());
1127 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1130 // FIXME: @GOTOFF should be a property of MCSymbolRefExpr not in the MCSymbol.
1131 std::string Name = MBB->getSymbol(Ctx)->getName() + "@GOTOFF";
1132 return MCSymbolRefExpr::Create(Ctx.GetOrCreateSymbol(StringRef(Name)), Ctx);
1135 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1137 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1138 SelectionDAG &DAG) const {
1139 if (!Subtarget->is64Bit())
1140 // This doesn't have DebugLoc associated with it, but is not really the
1141 // same as a Register.
1142 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1147 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1148 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1150 const MCExpr *X86TargetLowering::
1151 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1152 MCContext &Ctx) const {
1153 // X86-64 uses RIP relative addressing based on the jump table label.
1154 if (Subtarget->isPICStyleRIPRel())
1155 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1157 // Otherwise, the reference is relative to the PIC base.
1158 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1161 /// getFunctionAlignment - Return the Log2 alignment of this function.
1162 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1163 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1166 //===----------------------------------------------------------------------===//
1167 // Return Value Calling Convention Implementation
1168 //===----------------------------------------------------------------------===//
1170 #include "X86GenCallingConv.inc"
1173 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1174 const SmallVectorImpl<EVT> &OutTys,
1175 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1176 SelectionDAG &DAG) {
1177 SmallVector<CCValAssign, 16> RVLocs;
1178 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1179 RVLocs, *DAG.getContext());
1180 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1184 X86TargetLowering::LowerReturn(SDValue Chain,
1185 CallingConv::ID CallConv, bool isVarArg,
1186 const SmallVectorImpl<ISD::OutputArg> &Outs,
1187 DebugLoc dl, SelectionDAG &DAG) {
1189 SmallVector<CCValAssign, 16> RVLocs;
1190 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1191 RVLocs, *DAG.getContext());
1192 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1194 // If this is the first return lowered for this function, add the regs to the
1195 // liveout set for the function.
1196 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1197 for (unsigned i = 0; i != RVLocs.size(); ++i)
1198 if (RVLocs[i].isRegLoc())
1199 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1204 SmallVector<SDValue, 6> RetOps;
1205 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1206 // Operand #1 = Bytes To Pop
1207 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1209 // Copy the result values into the output registers.
1210 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1211 CCValAssign &VA = RVLocs[i];
1212 assert(VA.isRegLoc() && "Can only return in registers!");
1213 SDValue ValToCopy = Outs[i].Val;
1215 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1216 // the RET instruction and handled by the FP Stackifier.
1217 if (VA.getLocReg() == X86::ST0 ||
1218 VA.getLocReg() == X86::ST1) {
1219 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1220 // change the value to the FP stack register class.
1221 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1222 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1223 RetOps.push_back(ValToCopy);
1224 // Don't emit a copytoreg.
1228 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1229 // which is returned in RAX / RDX.
1230 if (Subtarget->is64Bit()) {
1231 EVT ValVT = ValToCopy.getValueType();
1232 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1233 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1234 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1235 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1239 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1240 Flag = Chain.getValue(1);
1243 // The x86-64 ABI for returning structs by value requires that we copy
1244 // the sret argument into %rax for the return. We saved the argument into
1245 // a virtual register in the entry block, so now we copy the value out
1247 if (Subtarget->is64Bit() &&
1248 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1249 MachineFunction &MF = DAG.getMachineFunction();
1250 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1251 unsigned Reg = FuncInfo->getSRetReturnReg();
1253 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1254 FuncInfo->setSRetReturnReg(Reg);
1256 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1258 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1259 Flag = Chain.getValue(1);
1261 // RAX now acts like a return value.
1262 MF.getRegInfo().addLiveOut(X86::RAX);
1265 RetOps[0] = Chain; // Update chain.
1267 // Add the flag if we have it.
1269 RetOps.push_back(Flag);
1271 return DAG.getNode(X86ISD::RET_FLAG, dl,
1272 MVT::Other, &RetOps[0], RetOps.size());
1275 /// LowerCallResult - Lower the result values of a call into the
1276 /// appropriate copies out of appropriate physical registers.
1279 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1280 CallingConv::ID CallConv, bool isVarArg,
1281 const SmallVectorImpl<ISD::InputArg> &Ins,
1282 DebugLoc dl, SelectionDAG &DAG,
1283 SmallVectorImpl<SDValue> &InVals) {
1285 // Assign locations to each value returned by this call.
1286 SmallVector<CCValAssign, 16> RVLocs;
1287 bool Is64Bit = Subtarget->is64Bit();
1288 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1289 RVLocs, *DAG.getContext());
1290 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1292 // Copy all of the result registers out of their specified physreg.
1293 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1294 CCValAssign &VA = RVLocs[i];
1295 EVT CopyVT = VA.getValVT();
1297 // If this is x86-64, and we disabled SSE, we can't return FP values
1298 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1299 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1300 llvm_report_error("SSE register return with SSE disabled");
1303 // If this is a call to a function that returns an fp value on the floating
1304 // point stack, but where we prefer to use the value in xmm registers, copy
1305 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1306 if ((VA.getLocReg() == X86::ST0 ||
1307 VA.getLocReg() == X86::ST1) &&
1308 isScalarFPTypeInSSEReg(VA.getValVT())) {
1313 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1314 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1315 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1316 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1317 MVT::v2i64, InFlag).getValue(1);
1318 Val = Chain.getValue(0);
1319 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1320 Val, DAG.getConstant(0, MVT::i64));
1322 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1323 MVT::i64, InFlag).getValue(1);
1324 Val = Chain.getValue(0);
1326 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1328 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1329 CopyVT, InFlag).getValue(1);
1330 Val = Chain.getValue(0);
1332 InFlag = Chain.getValue(2);
1334 if (CopyVT != VA.getValVT()) {
1335 // Round the F80 the right size, which also moves to the appropriate xmm
1337 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1338 // This truncation won't change the value.
1339 DAG.getIntPtrConstant(1));
1342 InVals.push_back(Val);
1349 //===----------------------------------------------------------------------===//
1350 // C & StdCall & Fast Calling Convention implementation
1351 //===----------------------------------------------------------------------===//
1352 // StdCall calling convention seems to be standard for many Windows' API
1353 // routines and around. It differs from C calling convention just a little:
1354 // callee should clean up the stack, not caller. Symbols should be also
1355 // decorated in some fancy way :) It doesn't support any vector arguments.
1356 // For info on fast calling convention see Fast Calling Convention (tail call)
1357 // implementation LowerX86_32FastCCCallTo.
1359 /// CallIsStructReturn - Determines whether a call uses struct return
1361 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1365 return Outs[0].Flags.isSRet();
1368 /// ArgsAreStructReturn - Determines whether a function uses struct
1369 /// return semantics.
1371 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1375 return Ins[0].Flags.isSRet();
1378 /// IsCalleePop - Determines whether the callee is required to pop its
1379 /// own arguments. Callee pop is necessary to support tail calls.
1380 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1384 switch (CallingConv) {
1387 case CallingConv::X86_StdCall:
1388 return !Subtarget->is64Bit();
1389 case CallingConv::X86_FastCall:
1390 return !Subtarget->is64Bit();
1391 case CallingConv::Fast:
1392 return PerformTailCallOpt;
1396 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1397 /// given CallingConvention value.
1398 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1399 if (Subtarget->is64Bit()) {
1400 if (Subtarget->isTargetWin64())
1401 return CC_X86_Win64_C;
1406 if (CC == CallingConv::X86_FastCall)
1407 return CC_X86_32_FastCall;
1408 else if (CC == CallingConv::Fast)
1409 return CC_X86_32_FastCC;
1414 /// NameDecorationForCallConv - Selects the appropriate decoration to
1415 /// apply to a MachineFunction containing a given calling convention.
1417 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1418 if (CallConv == CallingConv::X86_FastCall)
1420 else if (CallConv == CallingConv::X86_StdCall)
1426 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1427 /// by "Src" to address "Dst" with size and alignment information specified by
1428 /// the specific parameter attribute. The copy will be passed as a byval
1429 /// function parameter.
1431 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1432 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1434 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1435 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1436 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1440 X86TargetLowering::LowerMemArgument(SDValue Chain,
1441 CallingConv::ID CallConv,
1442 const SmallVectorImpl<ISD::InputArg> &Ins,
1443 DebugLoc dl, SelectionDAG &DAG,
1444 const CCValAssign &VA,
1445 MachineFrameInfo *MFI,
1448 // Create the nodes corresponding to a load from this parameter slot.
1449 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1450 bool AlwaysUseMutable = X86::IsEligibleForTailCallOpt(CallConv);
1451 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1454 // If value is passed by pointer we have address passed instead of the value
1456 if (VA.getLocInfo() == CCValAssign::Indirect)
1457 ValVT = VA.getLocVT();
1459 ValVT = VA.getValVT();
1461 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1462 // changed with more analysis.
1463 // In case of tail call optimization mark all arguments mutable. Since they
1464 // could be overwritten by lowering of arguments in case of a tail call.
1465 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1466 VA.getLocMemOffset(), isImmutable, false);
1467 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1468 if (Flags.isByVal())
1470 return DAG.getLoad(ValVT, dl, Chain, FIN,
1471 PseudoSourceValue::getFixedStack(FI), 0);
1475 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1476 CallingConv::ID CallConv,
1478 const SmallVectorImpl<ISD::InputArg> &Ins,
1481 SmallVectorImpl<SDValue> &InVals) {
1483 MachineFunction &MF = DAG.getMachineFunction();
1484 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1486 const Function* Fn = MF.getFunction();
1487 if (Fn->hasExternalLinkage() &&
1488 Subtarget->isTargetCygMing() &&
1489 Fn->getName() == "main")
1490 FuncInfo->setForceFramePointer(true);
1492 // Decorate the function name.
1493 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1495 MachineFrameInfo *MFI = MF.getFrameInfo();
1496 bool Is64Bit = Subtarget->is64Bit();
1497 bool IsWin64 = Subtarget->isTargetWin64();
1499 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1500 "Var args not supported with calling convention fastcc");
1502 // Assign locations to all of the incoming arguments.
1503 SmallVector<CCValAssign, 16> ArgLocs;
1504 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1505 ArgLocs, *DAG.getContext());
1506 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1508 unsigned LastVal = ~0U;
1510 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1511 CCValAssign &VA = ArgLocs[i];
1512 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1514 assert(VA.getValNo() != LastVal &&
1515 "Don't support value assigned to multiple locs yet");
1516 LastVal = VA.getValNo();
1518 if (VA.isRegLoc()) {
1519 EVT RegVT = VA.getLocVT();
1520 TargetRegisterClass *RC = NULL;
1521 if (RegVT == MVT::i32)
1522 RC = X86::GR32RegisterClass;
1523 else if (Is64Bit && RegVT == MVT::i64)
1524 RC = X86::GR64RegisterClass;
1525 else if (RegVT == MVT::f32)
1526 RC = X86::FR32RegisterClass;
1527 else if (RegVT == MVT::f64)
1528 RC = X86::FR64RegisterClass;
1529 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1530 RC = X86::VR128RegisterClass;
1531 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1532 RC = X86::VR64RegisterClass;
1534 llvm_unreachable("Unknown argument type!");
1536 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1537 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1539 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1540 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1542 if (VA.getLocInfo() == CCValAssign::SExt)
1543 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1544 DAG.getValueType(VA.getValVT()));
1545 else if (VA.getLocInfo() == CCValAssign::ZExt)
1546 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1547 DAG.getValueType(VA.getValVT()));
1548 else if (VA.getLocInfo() == CCValAssign::BCvt)
1549 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1551 if (VA.isExtInLoc()) {
1552 // Handle MMX values passed in XMM regs.
1553 if (RegVT.isVector()) {
1554 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1555 ArgValue, DAG.getConstant(0, MVT::i64));
1556 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1558 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1561 assert(VA.isMemLoc());
1562 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1565 // If value is passed via pointer - do a load.
1566 if (VA.getLocInfo() == CCValAssign::Indirect)
1567 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1569 InVals.push_back(ArgValue);
1572 // The x86-64 ABI for returning structs by value requires that we copy
1573 // the sret argument into %rax for the return. Save the argument into
1574 // a virtual register so that we can access it from the return points.
1575 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1576 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1577 unsigned Reg = FuncInfo->getSRetReturnReg();
1579 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1580 FuncInfo->setSRetReturnReg(Reg);
1582 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1583 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1586 unsigned StackSize = CCInfo.getNextStackOffset();
1587 // align stack specially for tail calls
1588 if (X86::IsEligibleForTailCallOpt(CallConv))
1589 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1591 // If the function takes variable number of arguments, make a frame index for
1592 // the start of the first vararg value... for expansion of llvm.va_start.
1594 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1595 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1598 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1600 // FIXME: We should really autogenerate these arrays
1601 static const unsigned GPR64ArgRegsWin64[] = {
1602 X86::RCX, X86::RDX, X86::R8, X86::R9
1604 static const unsigned XMMArgRegsWin64[] = {
1605 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1607 static const unsigned GPR64ArgRegs64Bit[] = {
1608 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1610 static const unsigned XMMArgRegs64Bit[] = {
1611 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1612 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1614 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1617 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1618 GPR64ArgRegs = GPR64ArgRegsWin64;
1619 XMMArgRegs = XMMArgRegsWin64;
1621 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1622 GPR64ArgRegs = GPR64ArgRegs64Bit;
1623 XMMArgRegs = XMMArgRegs64Bit;
1625 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1627 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1630 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1631 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1632 "SSE register cannot be used when SSE is disabled!");
1633 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1634 "SSE register cannot be used when SSE is disabled!");
1635 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1636 // Kernel mode asks for SSE to be disabled, so don't push them
1638 TotalNumXMMRegs = 0;
1640 // For X86-64, if there are vararg parameters that are passed via
1641 // registers, then we must store them to their spots on the stack so they
1642 // may be loaded by deferencing the result of va_next.
1643 VarArgsGPOffset = NumIntRegs * 8;
1644 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1645 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1646 TotalNumXMMRegs * 16, 16,
1649 // Store the integer parameter registers.
1650 SmallVector<SDValue, 8> MemOps;
1651 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1652 unsigned Offset = VarArgsGPOffset;
1653 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1654 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1655 DAG.getIntPtrConstant(Offset));
1656 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1657 X86::GR64RegisterClass);
1658 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1660 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1661 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1663 MemOps.push_back(Store);
1667 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1668 // Now store the XMM (fp + vector) parameter registers.
1669 SmallVector<SDValue, 11> SaveXMMOps;
1670 SaveXMMOps.push_back(Chain);
1672 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1673 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1674 SaveXMMOps.push_back(ALVal);
1676 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1677 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1679 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1680 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1681 X86::VR128RegisterClass);
1682 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1683 SaveXMMOps.push_back(Val);
1685 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1687 &SaveXMMOps[0], SaveXMMOps.size()));
1690 if (!MemOps.empty())
1691 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1692 &MemOps[0], MemOps.size());
1696 // Some CCs need callee pop.
1697 if (IsCalleePop(isVarArg, CallConv)) {
1698 BytesToPopOnReturn = StackSize; // Callee pops everything.
1699 BytesCallerReserves = 0;
1701 BytesToPopOnReturn = 0; // Callee pops nothing.
1702 // If this is an sret function, the return should pop the hidden pointer.
1703 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1704 BytesToPopOnReturn = 4;
1705 BytesCallerReserves = StackSize;
1709 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1710 if (CallConv == CallingConv::X86_FastCall)
1711 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1714 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1720 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1721 SDValue StackPtr, SDValue Arg,
1722 DebugLoc dl, SelectionDAG &DAG,
1723 const CCValAssign &VA,
1724 ISD::ArgFlagsTy Flags) {
1725 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1726 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1727 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1728 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1729 if (Flags.isByVal()) {
1730 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1732 return DAG.getStore(Chain, dl, Arg, PtrOff,
1733 PseudoSourceValue::getStack(), LocMemOffset);
1736 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1737 /// optimization is performed and it is required.
1739 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1740 SDValue &OutRetAddr, SDValue Chain,
1741 bool IsTailCall, bool Is64Bit,
1742 int FPDiff, DebugLoc dl) {
1743 if (!IsTailCall || FPDiff==0) return Chain;
1745 // Adjust the Return address stack slot.
1746 EVT VT = getPointerTy();
1747 OutRetAddr = getReturnAddressFrameIndex(DAG);
1749 // Load the "old" Return address.
1750 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1751 return SDValue(OutRetAddr.getNode(), 1);
1754 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1755 /// optimization is performed and it is required (FPDiff!=0).
1757 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1758 SDValue Chain, SDValue RetAddrFrIdx,
1759 bool Is64Bit, int FPDiff, DebugLoc dl) {
1760 // Store the return address to the appropriate stack slot.
1761 if (!FPDiff) return Chain;
1762 // Calculate the new stack slot for the return address.
1763 int SlotSize = Is64Bit ? 8 : 4;
1764 int NewReturnAddrFI =
1765 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1766 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1767 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1768 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1769 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1774 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1775 CallingConv::ID CallConv, bool isVarArg,
1777 const SmallVectorImpl<ISD::OutputArg> &Outs,
1778 const SmallVectorImpl<ISD::InputArg> &Ins,
1779 DebugLoc dl, SelectionDAG &DAG,
1780 SmallVectorImpl<SDValue> &InVals) {
1782 MachineFunction &MF = DAG.getMachineFunction();
1783 bool Is64Bit = Subtarget->is64Bit();
1784 bool IsStructRet = CallIsStructReturn(Outs);
1786 assert((!isTailCall || X86::IsEligibleForTailCallOpt(CallConv)) &&
1787 "Call is not eligible for tail call optimization!");
1788 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1789 "Var args not supported with calling convention fastcc");
1791 // Analyze operands of the call, assigning locations to each operand.
1792 SmallVector<CCValAssign, 16> ArgLocs;
1793 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1794 ArgLocs, *DAG.getContext());
1795 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1797 // Get a count of how many bytes are to be pushed on the stack.
1798 unsigned NumBytes = CCInfo.getNextStackOffset();
1799 if (X86::IsEligibleForTailCallOpt(CallConv))
1800 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1804 // Lower arguments at fp - stackoffset + fpdiff.
1805 unsigned NumBytesCallerPushed =
1806 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1807 FPDiff = NumBytesCallerPushed - NumBytes;
1809 // Set the delta of movement of the returnaddr stackslot.
1810 // But only set if delta is greater than previous delta.
1811 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1812 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1815 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1817 SDValue RetAddrFrIdx;
1818 // Load return adress for tail calls.
1819 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1822 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1823 SmallVector<SDValue, 8> MemOpChains;
1826 // Walk the register/memloc assignments, inserting copies/loads. In the case
1827 // of tail call optimization arguments are handle later.
1828 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1829 CCValAssign &VA = ArgLocs[i];
1830 EVT RegVT = VA.getLocVT();
1831 SDValue Arg = Outs[i].Val;
1832 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1833 bool isByVal = Flags.isByVal();
1835 // Promote the value if needed.
1836 switch (VA.getLocInfo()) {
1837 default: llvm_unreachable("Unknown loc info!");
1838 case CCValAssign::Full: break;
1839 case CCValAssign::SExt:
1840 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1842 case CCValAssign::ZExt:
1843 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1845 case CCValAssign::AExt:
1846 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1847 // Special case: passing MMX values in XMM registers.
1848 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1849 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1850 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1852 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1854 case CCValAssign::BCvt:
1855 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1857 case CCValAssign::Indirect: {
1858 // Store the argument.
1859 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1860 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1861 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1862 PseudoSourceValue::getFixedStack(FI), 0);
1868 if (VA.isRegLoc()) {
1869 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1871 if (!isTailCall || (isTailCall && isByVal)) {
1872 assert(VA.isMemLoc());
1873 if (StackPtr.getNode() == 0)
1874 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1876 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1877 dl, DAG, VA, Flags));
1882 if (!MemOpChains.empty())
1883 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1884 &MemOpChains[0], MemOpChains.size());
1886 // Build a sequence of copy-to-reg nodes chained together with token chain
1887 // and flag operands which copy the outgoing args into registers.
1889 // Tail call byval lowering might overwrite argument registers so in case of
1890 // tail call optimization the copies to registers are lowered later.
1892 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1893 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1894 RegsToPass[i].second, InFlag);
1895 InFlag = Chain.getValue(1);
1899 if (Subtarget->isPICStyleGOT()) {
1900 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1903 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1904 DAG.getNode(X86ISD::GlobalBaseReg,
1905 DebugLoc::getUnknownLoc(),
1908 InFlag = Chain.getValue(1);
1910 // If we are tail calling and generating PIC/GOT style code load the
1911 // address of the callee into ECX. The value in ecx is used as target of
1912 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1913 // for tail calls on PIC/GOT architectures. Normally we would just put the
1914 // address of GOT into ebx and then call target@PLT. But for tail calls
1915 // ebx would be restored (since ebx is callee saved) before jumping to the
1918 // Note: The actual moving to ECX is done further down.
1919 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1920 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1921 !G->getGlobal()->hasProtectedVisibility())
1922 Callee = LowerGlobalAddress(Callee, DAG);
1923 else if (isa<ExternalSymbolSDNode>(Callee))
1924 Callee = LowerExternalSymbol(Callee, DAG);
1928 if (Is64Bit && isVarArg) {
1929 // From AMD64 ABI document:
1930 // For calls that may call functions that use varargs or stdargs
1931 // (prototype-less calls or calls to functions containing ellipsis (...) in
1932 // the declaration) %al is used as hidden argument to specify the number
1933 // of SSE registers used. The contents of %al do not need to match exactly
1934 // the number of registers, but must be an ubound on the number of SSE
1935 // registers used and is in the range 0 - 8 inclusive.
1937 // FIXME: Verify this on Win64
1938 // Count the number of XMM registers allocated.
1939 static const unsigned XMMArgRegs[] = {
1940 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1941 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1943 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1944 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1945 && "SSE registers cannot be used when SSE is disabled");
1947 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1948 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1949 InFlag = Chain.getValue(1);
1953 // For tail calls lower the arguments to the 'real' stack slot.
1955 // Force all the incoming stack arguments to be loaded from the stack
1956 // before any new outgoing arguments are stored to the stack, because the
1957 // outgoing stack slots may alias the incoming argument stack slots, and
1958 // the alias isn't otherwise explicit. This is slightly more conservative
1959 // than necessary, because it means that each store effectively depends
1960 // on every argument instead of just those arguments it would clobber.
1961 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1963 SmallVector<SDValue, 8> MemOpChains2;
1966 // Do not flag preceeding copytoreg stuff together with the following stuff.
1968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1969 CCValAssign &VA = ArgLocs[i];
1970 if (!VA.isRegLoc()) {
1971 assert(VA.isMemLoc());
1972 SDValue Arg = Outs[i].Val;
1973 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1974 // Create frame index.
1975 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1976 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1977 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1978 FIN = DAG.getFrameIndex(FI, getPointerTy());
1980 if (Flags.isByVal()) {
1981 // Copy relative to framepointer.
1982 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1983 if (StackPtr.getNode() == 0)
1984 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1986 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1988 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1992 // Store relative to framepointer.
1993 MemOpChains2.push_back(
1994 DAG.getStore(ArgChain, dl, Arg, FIN,
1995 PseudoSourceValue::getFixedStack(FI), 0));
2000 if (!MemOpChains2.empty())
2001 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2002 &MemOpChains2[0], MemOpChains2.size());
2004 // Copy arguments to their registers.
2005 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2006 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2007 RegsToPass[i].second, InFlag);
2008 InFlag = Chain.getValue(1);
2012 // Store the return address to the appropriate stack slot.
2013 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2017 bool WasGlobalOrExternal = false;
2018 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2019 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2020 // In the 64-bit large code model, we have to make all calls
2021 // through a register, since the call instruction's 32-bit
2022 // pc-relative offset may not be large enough to hold the whole
2024 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2025 WasGlobalOrExternal = true;
2026 // If the callee is a GlobalAddress node (quite common, every direct call
2027 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2030 // We should use extra load for direct calls to dllimported functions in
2032 GlobalValue *GV = G->getGlobal();
2033 if (!GV->hasDLLImportLinkage()) {
2034 unsigned char OpFlags = 0;
2036 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2037 // external symbols most go through the PLT in PIC mode. If the symbol
2038 // has hidden or protected visibility, or if it is static or local, then
2039 // we don't need to use the PLT - we can directly call it.
2040 if (Subtarget->isTargetELF() &&
2041 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2042 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2043 OpFlags = X86II::MO_PLT;
2044 } else if (Subtarget->isPICStyleStubAny() &&
2045 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2046 Subtarget->getDarwinVers() < 9) {
2047 // PC-relative references to external symbols should go through $stub,
2048 // unless we're building with the leopard linker or later, which
2049 // automatically synthesizes these stubs.
2050 OpFlags = X86II::MO_DARWIN_STUB;
2053 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2054 G->getOffset(), OpFlags);
2056 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2057 WasGlobalOrExternal = true;
2058 unsigned char OpFlags = 0;
2060 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2061 // symbols should go through the PLT.
2062 if (Subtarget->isTargetELF() &&
2063 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2064 OpFlags = X86II::MO_PLT;
2065 } else if (Subtarget->isPICStyleStubAny() &&
2066 Subtarget->getDarwinVers() < 9) {
2067 // PC-relative references to external symbols should go through $stub,
2068 // unless we're building with the leopard linker or later, which
2069 // automatically synthesizes these stubs.
2070 OpFlags = X86II::MO_DARWIN_STUB;
2073 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2077 if (isTailCall && !WasGlobalOrExternal) {
2078 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2080 Chain = DAG.getCopyToReg(Chain, dl,
2081 DAG.getRegister(Opc, getPointerTy()),
2083 Callee = DAG.getRegister(Opc, getPointerTy());
2084 // Add register as live out.
2085 MF.getRegInfo().addLiveOut(Opc);
2088 // Returns a chain & a flag for retval copy to use.
2089 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2090 SmallVector<SDValue, 8> Ops;
2093 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2094 DAG.getIntPtrConstant(0, true), InFlag);
2095 InFlag = Chain.getValue(1);
2098 Ops.push_back(Chain);
2099 Ops.push_back(Callee);
2102 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2104 // Add argument registers to the end of the list so that they are known live
2106 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2107 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2108 RegsToPass[i].second.getValueType()));
2110 // Add an implicit use GOT pointer in EBX.
2111 if (!isTailCall && Subtarget->isPICStyleGOT())
2112 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2114 // Add an implicit use of AL for x86 vararg functions.
2115 if (Is64Bit && isVarArg)
2116 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2118 if (InFlag.getNode())
2119 Ops.push_back(InFlag);
2122 // If this is the first return lowered for this function, add the regs
2123 // to the liveout set for the function.
2124 if (MF.getRegInfo().liveout_empty()) {
2125 SmallVector<CCValAssign, 16> RVLocs;
2126 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2128 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2129 for (unsigned i = 0; i != RVLocs.size(); ++i)
2130 if (RVLocs[i].isRegLoc())
2131 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2134 assert(((Callee.getOpcode() == ISD::Register &&
2135 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2136 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2137 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2138 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2139 "Expecting a global address, external symbol, or scratch register");
2141 return DAG.getNode(X86ISD::TC_RETURN, dl,
2142 NodeTys, &Ops[0], Ops.size());
2145 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2146 InFlag = Chain.getValue(1);
2148 // Create the CALLSEQ_END node.
2149 unsigned NumBytesForCalleeToPush;
2150 if (IsCalleePop(isVarArg, CallConv))
2151 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2152 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2153 // If this is is a call to a struct-return function, the callee
2154 // pops the hidden struct pointer, so we have to push it back.
2155 // This is common for Darwin/X86, Linux & Mingw32 targets.
2156 NumBytesForCalleeToPush = 4;
2158 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2160 // Returns a flag for retval copy to use.
2161 Chain = DAG.getCALLSEQ_END(Chain,
2162 DAG.getIntPtrConstant(NumBytes, true),
2163 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2166 InFlag = Chain.getValue(1);
2168 // Handle result values, copying them out of physregs into vregs that we
2170 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2171 Ins, dl, DAG, InVals);
2175 //===----------------------------------------------------------------------===//
2176 // Fast Calling Convention (tail call) implementation
2177 //===----------------------------------------------------------------------===//
2179 // Like std call, callee cleans arguments, convention except that ECX is
2180 // reserved for storing the tail called function address. Only 2 registers are
2181 // free for argument passing (inreg). Tail call optimization is performed
2183 // * tailcallopt is enabled
2184 // * caller/callee are fastcc
2185 // On X86_64 architecture with GOT-style position independent code only local
2186 // (within module) calls are supported at the moment.
2187 // To keep the stack aligned according to platform abi the function
2188 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2189 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2190 // If a tail called function callee has more arguments than the caller the
2191 // caller needs to make sure that there is room to move the RETADDR to. This is
2192 // achieved by reserving an area the size of the argument delta right after the
2193 // original REtADDR, but before the saved framepointer or the spilled registers
2194 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2206 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2207 /// for a 16 byte align requirement.
2208 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2209 SelectionDAG& DAG) {
2210 MachineFunction &MF = DAG.getMachineFunction();
2211 const TargetMachine &TM = MF.getTarget();
2212 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2213 unsigned StackAlignment = TFI.getStackAlignment();
2214 uint64_t AlignMask = StackAlignment - 1;
2215 int64_t Offset = StackSize;
2216 uint64_t SlotSize = TD->getPointerSize();
2217 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2218 // Number smaller than 12 so just add the difference.
2219 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2221 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2222 Offset = ((~AlignMask) & Offset) + StackAlignment +
2223 (StackAlignment-SlotSize);
2228 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2229 /// for tail call optimization. Targets which want to do tail call
2230 /// optimization should implement this function.
2232 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2233 CallingConv::ID CalleeCC,
2235 const SmallVectorImpl<ISD::InputArg> &Ins,
2236 SelectionDAG& DAG) const {
2237 return X86::IsEligibleForTailCallOpt(CalleeCC) &&
2238 DAG.getMachineFunction().getFunction()->getCallingConv() == CalleeCC;
2242 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2244 DenseMap<const Value *, unsigned> &vm,
2245 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2246 DenseMap<const AllocaInst *, int> &am
2248 , SmallSet<Instruction*, 8> &cil
2251 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2259 //===----------------------------------------------------------------------===//
2260 // Other Lowering Hooks
2261 //===----------------------------------------------------------------------===//
2264 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2265 MachineFunction &MF = DAG.getMachineFunction();
2266 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2267 int ReturnAddrIndex = FuncInfo->getRAIndex();
2269 if (ReturnAddrIndex == 0) {
2270 // Set up a frame object for the return address.
2271 uint64_t SlotSize = TD->getPointerSize();
2272 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2274 FuncInfo->setRAIndex(ReturnAddrIndex);
2277 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2281 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2282 bool hasSymbolicDisplacement) {
2283 // Offset should fit into 32 bit immediate field.
2284 if (!isInt32(Offset))
2287 // If we don't have a symbolic displacement - we don't have any extra
2289 if (!hasSymbolicDisplacement)
2292 // FIXME: Some tweaks might be needed for medium code model.
2293 if (M != CodeModel::Small && M != CodeModel::Kernel)
2296 // For small code model we assume that latest object is 16MB before end of 31
2297 // bits boundary. We may also accept pretty large negative constants knowing
2298 // that all objects are in the positive half of address space.
2299 if (M == CodeModel::Small && Offset < 16*1024*1024)
2302 // For kernel code model we know that all object resist in the negative half
2303 // of 32bits address space. We may not accept negative offsets, since they may
2304 // be just off and we may accept pretty large positive ones.
2305 if (M == CodeModel::Kernel && Offset > 0)
2311 bool X86::IsEligibleForTailCallOpt(CallingConv::ID CC) {
2312 return PerformTailCallOpt && CC == CallingConv::Fast;
2315 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2316 /// specific condition code, returning the condition code and the LHS/RHS of the
2317 /// comparison to make.
2318 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2319 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2321 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2322 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2323 // X > -1 -> X == 0, jump !sign.
2324 RHS = DAG.getConstant(0, RHS.getValueType());
2325 return X86::COND_NS;
2326 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2327 // X < 0 -> X == 0, jump on sign.
2329 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2331 RHS = DAG.getConstant(0, RHS.getValueType());
2332 return X86::COND_LE;
2336 switch (SetCCOpcode) {
2337 default: llvm_unreachable("Invalid integer condition!");
2338 case ISD::SETEQ: return X86::COND_E;
2339 case ISD::SETGT: return X86::COND_G;
2340 case ISD::SETGE: return X86::COND_GE;
2341 case ISD::SETLT: return X86::COND_L;
2342 case ISD::SETLE: return X86::COND_LE;
2343 case ISD::SETNE: return X86::COND_NE;
2344 case ISD::SETULT: return X86::COND_B;
2345 case ISD::SETUGT: return X86::COND_A;
2346 case ISD::SETULE: return X86::COND_BE;
2347 case ISD::SETUGE: return X86::COND_AE;
2351 // First determine if it is required or is profitable to flip the operands.
2353 // If LHS is a foldable load, but RHS is not, flip the condition.
2354 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2355 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2356 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2357 std::swap(LHS, RHS);
2360 switch (SetCCOpcode) {
2366 std::swap(LHS, RHS);
2370 // On a floating point condition, the flags are set as follows:
2372 // 0 | 0 | 0 | X > Y
2373 // 0 | 0 | 1 | X < Y
2374 // 1 | 0 | 0 | X == Y
2375 // 1 | 1 | 1 | unordered
2376 switch (SetCCOpcode) {
2377 default: llvm_unreachable("Condcode should be pre-legalized away");
2379 case ISD::SETEQ: return X86::COND_E;
2380 case ISD::SETOLT: // flipped
2382 case ISD::SETGT: return X86::COND_A;
2383 case ISD::SETOLE: // flipped
2385 case ISD::SETGE: return X86::COND_AE;
2386 case ISD::SETUGT: // flipped
2388 case ISD::SETLT: return X86::COND_B;
2389 case ISD::SETUGE: // flipped
2391 case ISD::SETLE: return X86::COND_BE;
2393 case ISD::SETNE: return X86::COND_NE;
2394 case ISD::SETUO: return X86::COND_P;
2395 case ISD::SETO: return X86::COND_NP;
2397 case ISD::SETUNE: return X86::COND_INVALID;
2401 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2402 /// code. Current x86 isa includes the following FP cmov instructions:
2403 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2404 static bool hasFPCMov(unsigned X86CC) {
2420 /// isFPImmLegal - Returns true if the target can instruction select the
2421 /// specified FP immediate natively. If false, the legalizer will
2422 /// materialize the FP immediate as a load from a constant pool.
2423 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2424 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2425 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2431 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2432 /// the specified range (L, H].
2433 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2434 return (Val < 0) || (Val >= Low && Val < Hi);
2437 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2438 /// specified value.
2439 static bool isUndefOrEqual(int Val, int CmpVal) {
2440 if (Val < 0 || Val == CmpVal)
2445 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2446 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2447 /// the second operand.
2448 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2449 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2450 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2451 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2452 return (Mask[0] < 2 && Mask[1] < 2);
2456 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2457 SmallVector<int, 8> M;
2459 return ::isPSHUFDMask(M, N->getValueType(0));
2462 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2463 /// is suitable for input to PSHUFHW.
2464 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2465 if (VT != MVT::v8i16)
2468 // Lower quadword copied in order or undef.
2469 for (int i = 0; i != 4; ++i)
2470 if (Mask[i] >= 0 && Mask[i] != i)
2473 // Upper quadword shuffled.
2474 for (int i = 4; i != 8; ++i)
2475 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2481 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2482 SmallVector<int, 8> M;
2484 return ::isPSHUFHWMask(M, N->getValueType(0));
2487 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2488 /// is suitable for input to PSHUFLW.
2489 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2490 if (VT != MVT::v8i16)
2493 // Upper quadword copied in order.
2494 for (int i = 4; i != 8; ++i)
2495 if (Mask[i] >= 0 && Mask[i] != i)
2498 // Lower quadword shuffled.
2499 for (int i = 0; i != 4; ++i)
2506 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2507 SmallVector<int, 8> M;
2509 return ::isPSHUFLWMask(M, N->getValueType(0));
2512 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2513 /// is suitable for input to PALIGNR.
2514 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2516 int i, e = VT.getVectorNumElements();
2518 // Do not handle v2i64 / v2f64 shuffles with palignr.
2519 if (e < 4 || !hasSSSE3)
2522 for (i = 0; i != e; ++i)
2526 // All undef, not a palignr.
2530 // Determine if it's ok to perform a palignr with only the LHS, since we
2531 // don't have access to the actual shuffle elements to see if RHS is undef.
2532 bool Unary = Mask[i] < (int)e;
2533 bool NeedsUnary = false;
2535 int s = Mask[i] - i;
2537 // Check the rest of the elements to see if they are consecutive.
2538 for (++i; i != e; ++i) {
2543 Unary = Unary && (m < (int)e);
2544 NeedsUnary = NeedsUnary || (m < s);
2546 if (NeedsUnary && !Unary)
2548 if (Unary && m != ((s+i) & (e-1)))
2550 if (!Unary && m != (s+i))
2556 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2557 SmallVector<int, 8> M;
2559 return ::isPALIGNRMask(M, N->getValueType(0), true);
2562 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2563 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2564 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2565 int NumElems = VT.getVectorNumElements();
2566 if (NumElems != 2 && NumElems != 4)
2569 int Half = NumElems / 2;
2570 for (int i = 0; i < Half; ++i)
2571 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2573 for (int i = Half; i < NumElems; ++i)
2574 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2580 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2581 SmallVector<int, 8> M;
2583 return ::isSHUFPMask(M, N->getValueType(0));
2586 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2587 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2588 /// half elements to come from vector 1 (which would equal the dest.) and
2589 /// the upper half to come from vector 2.
2590 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2591 int NumElems = VT.getVectorNumElements();
2593 if (NumElems != 2 && NumElems != 4)
2596 int Half = NumElems / 2;
2597 for (int i = 0; i < Half; ++i)
2598 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2600 for (int i = Half; i < NumElems; ++i)
2601 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2606 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2607 SmallVector<int, 8> M;
2609 return isCommutedSHUFPMask(M, N->getValueType(0));
2612 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2613 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2614 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2615 if (N->getValueType(0).getVectorNumElements() != 4)
2618 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2619 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2620 isUndefOrEqual(N->getMaskElt(1), 7) &&
2621 isUndefOrEqual(N->getMaskElt(2), 2) &&
2622 isUndefOrEqual(N->getMaskElt(3), 3);
2625 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2626 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2628 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2629 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2634 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2635 isUndefOrEqual(N->getMaskElt(1), 3) &&
2636 isUndefOrEqual(N->getMaskElt(2), 2) &&
2637 isUndefOrEqual(N->getMaskElt(3), 3);
2640 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2641 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2642 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2643 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2645 if (NumElems != 2 && NumElems != 4)
2648 for (unsigned i = 0; i < NumElems/2; ++i)
2649 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2652 for (unsigned i = NumElems/2; i < NumElems; ++i)
2653 if (!isUndefOrEqual(N->getMaskElt(i), i))
2659 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2660 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2661 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2662 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2664 if (NumElems != 2 && NumElems != 4)
2667 for (unsigned i = 0; i < NumElems/2; ++i)
2668 if (!isUndefOrEqual(N->getMaskElt(i), i))
2671 for (unsigned i = 0; i < NumElems/2; ++i)
2672 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2678 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2679 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2680 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2681 bool V2IsSplat = false) {
2682 int NumElts = VT.getVectorNumElements();
2683 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2686 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2688 int BitI1 = Mask[i+1];
2689 if (!isUndefOrEqual(BitI, j))
2692 if (!isUndefOrEqual(BitI1, NumElts))
2695 if (!isUndefOrEqual(BitI1, j + NumElts))
2702 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2703 SmallVector<int, 8> M;
2705 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2708 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2709 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2710 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2711 bool V2IsSplat = false) {
2712 int NumElts = VT.getVectorNumElements();
2713 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2716 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2718 int BitI1 = Mask[i+1];
2719 if (!isUndefOrEqual(BitI, j + NumElts/2))
2722 if (isUndefOrEqual(BitI1, NumElts))
2725 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2732 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2733 SmallVector<int, 8> M;
2735 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2738 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2739 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2741 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2742 int NumElems = VT.getVectorNumElements();
2743 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2746 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2748 int BitI1 = Mask[i+1];
2749 if (!isUndefOrEqual(BitI, j))
2751 if (!isUndefOrEqual(BitI1, j))
2757 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2758 SmallVector<int, 8> M;
2760 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2763 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2764 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2766 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2767 int NumElems = VT.getVectorNumElements();
2768 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2771 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2773 int BitI1 = Mask[i+1];
2774 if (!isUndefOrEqual(BitI, j))
2776 if (!isUndefOrEqual(BitI1, j))
2782 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2783 SmallVector<int, 8> M;
2785 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2788 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2789 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2790 /// MOVSD, and MOVD, i.e. setting the lowest element.
2791 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2792 if (VT.getVectorElementType().getSizeInBits() < 32)
2795 int NumElts = VT.getVectorNumElements();
2797 if (!isUndefOrEqual(Mask[0], NumElts))
2800 for (int i = 1; i < NumElts; ++i)
2801 if (!isUndefOrEqual(Mask[i], i))
2807 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2808 SmallVector<int, 8> M;
2810 return ::isMOVLMask(M, N->getValueType(0));
2813 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2814 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2815 /// element of vector 2 and the other elements to come from vector 1 in order.
2816 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2817 bool V2IsSplat = false, bool V2IsUndef = false) {
2818 int NumOps = VT.getVectorNumElements();
2819 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2822 if (!isUndefOrEqual(Mask[0], 0))
2825 for (int i = 1; i < NumOps; ++i)
2826 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2827 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2828 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2834 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2835 bool V2IsUndef = false) {
2836 SmallVector<int, 8> M;
2838 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2841 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2842 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2843 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2844 if (N->getValueType(0).getVectorNumElements() != 4)
2847 // Expect 1, 1, 3, 3
2848 for (unsigned i = 0; i < 2; ++i) {
2849 int Elt = N->getMaskElt(i);
2850 if (Elt >= 0 && Elt != 1)
2855 for (unsigned i = 2; i < 4; ++i) {
2856 int Elt = N->getMaskElt(i);
2857 if (Elt >= 0 && Elt != 3)
2862 // Don't use movshdup if it can be done with a shufps.
2863 // FIXME: verify that matching u, u, 3, 3 is what we want.
2867 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2868 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2869 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2870 if (N->getValueType(0).getVectorNumElements() != 4)
2873 // Expect 0, 0, 2, 2
2874 for (unsigned i = 0; i < 2; ++i)
2875 if (N->getMaskElt(i) > 0)
2879 for (unsigned i = 2; i < 4; ++i) {
2880 int Elt = N->getMaskElt(i);
2881 if (Elt >= 0 && Elt != 2)
2886 // Don't use movsldup if it can be done with a shufps.
2890 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2891 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2892 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2893 int e = N->getValueType(0).getVectorNumElements() / 2;
2895 for (int i = 0; i < e; ++i)
2896 if (!isUndefOrEqual(N->getMaskElt(i), i))
2898 for (int i = 0; i < e; ++i)
2899 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2904 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2905 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2906 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2907 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2908 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2910 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2912 for (int i = 0; i < NumOperands; ++i) {
2913 int Val = SVOp->getMaskElt(NumOperands-i-1);
2914 if (Val < 0) Val = 0;
2915 if (Val >= NumOperands) Val -= NumOperands;
2917 if (i != NumOperands - 1)
2923 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2924 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2925 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2926 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2928 // 8 nodes, but we only care about the last 4.
2929 for (unsigned i = 7; i >= 4; --i) {
2930 int Val = SVOp->getMaskElt(i);
2939 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2940 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2941 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2942 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2944 // 8 nodes, but we only care about the first 4.
2945 for (int i = 3; i >= 0; --i) {
2946 int Val = SVOp->getMaskElt(i);
2955 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2956 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2957 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2958 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2959 EVT VVT = N->getValueType(0);
2960 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2964 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2965 Val = SVOp->getMaskElt(i);
2969 return (Val - i) * EltSize;
2972 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2974 bool X86::isZeroNode(SDValue Elt) {
2975 return ((isa<ConstantSDNode>(Elt) &&
2976 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2977 (isa<ConstantFPSDNode>(Elt) &&
2978 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2981 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2982 /// their permute mask.
2983 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2984 SelectionDAG &DAG) {
2985 EVT VT = SVOp->getValueType(0);
2986 unsigned NumElems = VT.getVectorNumElements();
2987 SmallVector<int, 8> MaskVec;
2989 for (unsigned i = 0; i != NumElems; ++i) {
2990 int idx = SVOp->getMaskElt(i);
2992 MaskVec.push_back(idx);
2993 else if (idx < (int)NumElems)
2994 MaskVec.push_back(idx + NumElems);
2996 MaskVec.push_back(idx - NumElems);
2998 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2999 SVOp->getOperand(0), &MaskVec[0]);
3002 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3003 /// the two vector operands have swapped position.
3004 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3005 unsigned NumElems = VT.getVectorNumElements();
3006 for (unsigned i = 0; i != NumElems; ++i) {
3010 else if (idx < (int)NumElems)
3011 Mask[i] = idx + NumElems;
3013 Mask[i] = idx - NumElems;
3017 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3018 /// match movhlps. The lower half elements should come from upper half of
3019 /// V1 (and in order), and the upper half elements should come from the upper
3020 /// half of V2 (and in order).
3021 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3022 if (Op->getValueType(0).getVectorNumElements() != 4)
3024 for (unsigned i = 0, e = 2; i != e; ++i)
3025 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3027 for (unsigned i = 2; i != 4; ++i)
3028 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3033 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3034 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3036 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3037 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3039 N = N->getOperand(0).getNode();
3040 if (!ISD::isNON_EXTLoad(N))
3043 *LD = cast<LoadSDNode>(N);
3047 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3048 /// match movlp{s|d}. The lower half elements should come from lower half of
3049 /// V1 (and in order), and the upper half elements should come from the upper
3050 /// half of V2 (and in order). And since V1 will become the source of the
3051 /// MOVLP, it must be either a vector load or a scalar load to vector.
3052 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3053 ShuffleVectorSDNode *Op) {
3054 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3056 // Is V2 is a vector load, don't do this transformation. We will try to use
3057 // load folding shufps op.
3058 if (ISD::isNON_EXTLoad(V2))
3061 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3063 if (NumElems != 2 && NumElems != 4)
3065 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3066 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3068 for (unsigned i = NumElems/2; i != NumElems; ++i)
3069 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3074 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3076 static bool isSplatVector(SDNode *N) {
3077 if (N->getOpcode() != ISD::BUILD_VECTOR)
3080 SDValue SplatValue = N->getOperand(0);
3081 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3082 if (N->getOperand(i) != SplatValue)
3087 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3088 /// to an zero vector.
3089 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3090 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3091 SDValue V1 = N->getOperand(0);
3092 SDValue V2 = N->getOperand(1);
3093 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3094 for (unsigned i = 0; i != NumElems; ++i) {
3095 int Idx = N->getMaskElt(i);
3096 if (Idx >= (int)NumElems) {
3097 unsigned Opc = V2.getOpcode();
3098 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3100 if (Opc != ISD::BUILD_VECTOR ||
3101 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3103 } else if (Idx >= 0) {
3104 unsigned Opc = V1.getOpcode();
3105 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3107 if (Opc != ISD::BUILD_VECTOR ||
3108 !X86::isZeroNode(V1.getOperand(Idx)))
3115 /// getZeroVector - Returns a vector of specified type with all zero elements.
3117 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3119 assert(VT.isVector() && "Expected a vector type");
3121 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3122 // type. This ensures they get CSE'd.
3124 if (VT.getSizeInBits() == 64) { // MMX
3125 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3126 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3127 } else if (HasSSE2) { // SSE2
3128 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3129 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3131 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3132 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3134 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3137 /// getOnesVector - Returns a vector of specified type with all bits set.
3139 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3140 assert(VT.isVector() && "Expected a vector type");
3142 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3143 // type. This ensures they get CSE'd.
3144 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3146 if (VT.getSizeInBits() == 64) // MMX
3147 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3149 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3150 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3154 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3155 /// that point to V2 points to its first element.
3156 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3157 EVT VT = SVOp->getValueType(0);
3158 unsigned NumElems = VT.getVectorNumElements();
3160 bool Changed = false;
3161 SmallVector<int, 8> MaskVec;
3162 SVOp->getMask(MaskVec);
3164 for (unsigned i = 0; i != NumElems; ++i) {
3165 if (MaskVec[i] > (int)NumElems) {
3166 MaskVec[i] = NumElems;
3171 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3172 SVOp->getOperand(1), &MaskVec[0]);
3173 return SDValue(SVOp, 0);
3176 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3177 /// operation of specified width.
3178 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3180 unsigned NumElems = VT.getVectorNumElements();
3181 SmallVector<int, 8> Mask;
3182 Mask.push_back(NumElems);
3183 for (unsigned i = 1; i != NumElems; ++i)
3185 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3188 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3189 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3191 unsigned NumElems = VT.getVectorNumElements();
3192 SmallVector<int, 8> Mask;
3193 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3195 Mask.push_back(i + NumElems);
3197 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3200 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3201 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3203 unsigned NumElems = VT.getVectorNumElements();
3204 unsigned Half = NumElems/2;
3205 SmallVector<int, 8> Mask;
3206 for (unsigned i = 0; i != Half; ++i) {
3207 Mask.push_back(i + Half);
3208 Mask.push_back(i + NumElems + Half);
3210 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3213 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3214 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3216 if (SV->getValueType(0).getVectorNumElements() <= 4)
3217 return SDValue(SV, 0);
3219 EVT PVT = MVT::v4f32;
3220 EVT VT = SV->getValueType(0);
3221 DebugLoc dl = SV->getDebugLoc();
3222 SDValue V1 = SV->getOperand(0);
3223 int NumElems = VT.getVectorNumElements();
3224 int EltNo = SV->getSplatIndex();
3226 // unpack elements to the correct location
3227 while (NumElems > 4) {
3228 if (EltNo < NumElems/2) {
3229 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3231 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3232 EltNo -= NumElems/2;
3237 // Perform the splat.
3238 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3239 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3240 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3241 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3244 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3245 /// vector of zero or undef vector. This produces a shuffle where the low
3246 /// element of V2 is swizzled into the zero/undef vector, landing at element
3247 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3248 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3249 bool isZero, bool HasSSE2,
3250 SelectionDAG &DAG) {
3251 EVT VT = V2.getValueType();
3253 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3254 unsigned NumElems = VT.getVectorNumElements();
3255 SmallVector<int, 16> MaskVec;
3256 for (unsigned i = 0; i != NumElems; ++i)
3257 // If this is the insertion idx, put the low elt of V2 here.
3258 MaskVec.push_back(i == Idx ? NumElems : i);
3259 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3262 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3263 /// a shuffle that is zero.
3265 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3266 bool Low, SelectionDAG &DAG) {
3267 unsigned NumZeros = 0;
3268 for (int i = 0; i < NumElems; ++i) {
3269 unsigned Index = Low ? i : NumElems-i-1;
3270 int Idx = SVOp->getMaskElt(Index);
3275 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3276 if (Elt.getNode() && X86::isZeroNode(Elt))
3284 /// isVectorShift - Returns true if the shuffle can be implemented as a
3285 /// logical left or right shift of a vector.
3286 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3287 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3288 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3289 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3292 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3295 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3299 bool SeenV1 = false;
3300 bool SeenV2 = false;
3301 for (int i = NumZeros; i < NumElems; ++i) {
3302 int Val = isLeft ? (i - NumZeros) : i;
3303 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3315 if (SeenV1 && SeenV2)
3318 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3324 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3326 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3327 unsigned NumNonZero, unsigned NumZero,
3328 SelectionDAG &DAG, TargetLowering &TLI) {
3332 DebugLoc dl = Op.getDebugLoc();
3335 for (unsigned i = 0; i < 16; ++i) {
3336 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3337 if (ThisIsNonZero && First) {
3339 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3341 V = DAG.getUNDEF(MVT::v8i16);
3346 SDValue ThisElt(0, 0), LastElt(0, 0);
3347 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3348 if (LastIsNonZero) {
3349 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3350 MVT::i16, Op.getOperand(i-1));
3352 if (ThisIsNonZero) {
3353 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3354 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3355 ThisElt, DAG.getConstant(8, MVT::i8));
3357 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3361 if (ThisElt.getNode())
3362 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3363 DAG.getIntPtrConstant(i/2));
3367 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3370 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3372 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3373 unsigned NumNonZero, unsigned NumZero,
3374 SelectionDAG &DAG, TargetLowering &TLI) {
3378 DebugLoc dl = Op.getDebugLoc();
3381 for (unsigned i = 0; i < 8; ++i) {
3382 bool isNonZero = (NonZeros & (1 << i)) != 0;
3386 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3388 V = DAG.getUNDEF(MVT::v8i16);
3391 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3392 MVT::v8i16, V, Op.getOperand(i),
3393 DAG.getIntPtrConstant(i));
3400 /// getVShift - Return a vector logical shift node.
3402 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3403 unsigned NumBits, SelectionDAG &DAG,
3404 const TargetLowering &TLI, DebugLoc dl) {
3405 bool isMMX = VT.getSizeInBits() == 64;
3406 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3407 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3408 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3409 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3410 DAG.getNode(Opc, dl, ShVT, SrcOp,
3411 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3415 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3416 SelectionDAG &DAG) {
3418 // Check if the scalar load can be widened into a vector load. And if
3419 // the address is "base + cst" see if the cst can be "absorbed" into
3420 // the shuffle mask.
3421 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3422 SDValue Ptr = LD->getBasePtr();
3423 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3425 EVT PVT = LD->getValueType(0);
3426 if (PVT != MVT::i32 && PVT != MVT::f32)
3431 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3432 FI = FINode->getIndex();
3434 } else if (Ptr.getOpcode() == ISD::ADD &&
3435 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3436 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3437 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3438 Offset = Ptr.getConstantOperandVal(1);
3439 Ptr = Ptr.getOperand(0);
3444 SDValue Chain = LD->getChain();
3445 // Make sure the stack object alignment is at least 16.
3446 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3447 if (DAG.InferPtrAlignment(Ptr) < 16) {
3448 if (MFI->isFixedObjectIndex(FI)) {
3449 // Can't change the alignment. FIXME: It's possible to compute
3450 // the exact stack offset and reference FI + adjust offset instead.
3451 // If someone *really* cares about this. That's the way to implement it.
3454 MFI->setObjectAlignment(FI, 16);
3458 // (Offset % 16) must be multiple of 4. Then address is then
3459 // Ptr + (Offset & ~15).
3462 if ((Offset % 16) & 3)
3464 int64_t StartOffset = Offset & ~15;
3466 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3467 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3469 int EltNo = (Offset - StartOffset) >> 2;
3470 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3471 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3472 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3473 // Canonicalize it to a v4i32 shuffle.
3474 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3475 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3476 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3477 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3484 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3485 DebugLoc dl = Op.getDebugLoc();
3486 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3487 if (ISD::isBuildVectorAllZeros(Op.getNode())
3488 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3489 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3490 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3491 // eliminated on x86-32 hosts.
3492 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3495 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3496 return getOnesVector(Op.getValueType(), DAG, dl);
3497 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3500 EVT VT = Op.getValueType();
3501 EVT ExtVT = VT.getVectorElementType();
3502 unsigned EVTBits = ExtVT.getSizeInBits();
3504 unsigned NumElems = Op.getNumOperands();
3505 unsigned NumZero = 0;
3506 unsigned NumNonZero = 0;
3507 unsigned NonZeros = 0;
3508 bool IsAllConstants = true;
3509 SmallSet<SDValue, 8> Values;
3510 for (unsigned i = 0; i < NumElems; ++i) {
3511 SDValue Elt = Op.getOperand(i);
3512 if (Elt.getOpcode() == ISD::UNDEF)
3515 if (Elt.getOpcode() != ISD::Constant &&
3516 Elt.getOpcode() != ISD::ConstantFP)
3517 IsAllConstants = false;
3518 if (X86::isZeroNode(Elt))
3521 NonZeros |= (1 << i);
3526 if (NumNonZero == 0) {
3527 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3528 return DAG.getUNDEF(VT);
3531 // Special case for single non-zero, non-undef, element.
3532 if (NumNonZero == 1) {
3533 unsigned Idx = CountTrailingZeros_32(NonZeros);
3534 SDValue Item = Op.getOperand(Idx);
3536 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3537 // the value are obviously zero, truncate the value to i32 and do the
3538 // insertion that way. Only do this if the value is non-constant or if the
3539 // value is a constant being inserted into element 0. It is cheaper to do
3540 // a constant pool load than it is to do a movd + shuffle.
3541 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3542 (!IsAllConstants || Idx == 0)) {
3543 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3544 // Handle MMX and SSE both.
3545 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3546 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3548 // Truncate the value (which may itself be a constant) to i32, and
3549 // convert it to a vector with movd (S2V+shuffle to zero extend).
3550 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3551 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3552 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3553 Subtarget->hasSSE2(), DAG);
3555 // Now we have our 32-bit value zero extended in the low element of
3556 // a vector. If Idx != 0, swizzle it into place.
3558 SmallVector<int, 4> Mask;
3559 Mask.push_back(Idx);
3560 for (unsigned i = 1; i != VecElts; ++i)
3562 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3563 DAG.getUNDEF(Item.getValueType()),
3566 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3570 // If we have a constant or non-constant insertion into the low element of
3571 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3572 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3573 // depending on what the source datatype is.
3576 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3577 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3578 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3579 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3580 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3581 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3583 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3584 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3585 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3586 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3587 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3588 Subtarget->hasSSE2(), DAG);
3589 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3593 // Is it a vector logical left shift?
3594 if (NumElems == 2 && Idx == 1 &&
3595 X86::isZeroNode(Op.getOperand(0)) &&
3596 !X86::isZeroNode(Op.getOperand(1))) {
3597 unsigned NumBits = VT.getSizeInBits();
3598 return getVShift(true, VT,
3599 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3600 VT, Op.getOperand(1)),
3601 NumBits/2, DAG, *this, dl);
3604 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3607 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3608 // is a non-constant being inserted into an element other than the low one,
3609 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3610 // movd/movss) to move this into the low element, then shuffle it into
3612 if (EVTBits == 32) {
3613 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3615 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3616 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3617 Subtarget->hasSSE2(), DAG);
3618 SmallVector<int, 8> MaskVec;
3619 for (unsigned i = 0; i < NumElems; i++)
3620 MaskVec.push_back(i == Idx ? 0 : 1);
3621 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3625 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3626 if (Values.size() == 1) {
3627 if (EVTBits == 32) {
3628 // Instead of a shuffle like this:
3629 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3630 // Check if it's possible to issue this instead.
3631 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3632 unsigned Idx = CountTrailingZeros_32(NonZeros);
3633 SDValue Item = Op.getOperand(Idx);
3634 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3635 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3640 // A vector full of immediates; various special cases are already
3641 // handled, so this is best done with a single constant-pool load.
3645 // Let legalizer expand 2-wide build_vectors.
3646 if (EVTBits == 64) {
3647 if (NumNonZero == 1) {
3648 // One half is zero or undef.
3649 unsigned Idx = CountTrailingZeros_32(NonZeros);
3650 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3651 Op.getOperand(Idx));
3652 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3653 Subtarget->hasSSE2(), DAG);
3658 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3659 if (EVTBits == 8 && NumElems == 16) {
3660 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3662 if (V.getNode()) return V;
3665 if (EVTBits == 16 && NumElems == 8) {
3666 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3668 if (V.getNode()) return V;
3671 // If element VT is == 32 bits, turn it into a number of shuffles.
3672 SmallVector<SDValue, 8> V;
3674 if (NumElems == 4 && NumZero > 0) {
3675 for (unsigned i = 0; i < 4; ++i) {
3676 bool isZero = !(NonZeros & (1 << i));
3678 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3680 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3683 for (unsigned i = 0; i < 2; ++i) {
3684 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3687 V[i] = V[i*2]; // Must be a zero vector.
3690 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3693 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3696 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3701 SmallVector<int, 8> MaskVec;
3702 bool Reverse = (NonZeros & 0x3) == 2;
3703 for (unsigned i = 0; i < 2; ++i)
3704 MaskVec.push_back(Reverse ? 1-i : i);
3705 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3706 for (unsigned i = 0; i < 2; ++i)
3707 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3708 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3711 if (Values.size() > 2) {
3712 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3713 // values to be inserted is equal to the number of elements, in which case
3714 // use the unpack code below in the hopes of matching the consecutive elts
3715 // load merge pattern for shuffles.
3716 // FIXME: We could probably just check that here directly.
3717 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3718 getSubtarget()->hasSSE41()) {
3719 V[0] = DAG.getUNDEF(VT);
3720 for (unsigned i = 0; i < NumElems; ++i)
3721 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3722 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3723 Op.getOperand(i), DAG.getIntPtrConstant(i));
3726 // Expand into a number of unpckl*.
3728 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3729 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3730 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3731 for (unsigned i = 0; i < NumElems; ++i)
3732 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3734 while (NumElems != 0) {
3735 for (unsigned i = 0; i < NumElems; ++i)
3736 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3746 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3747 // We support concatenate two MMX registers and place them in a MMX
3748 // register. This is better than doing a stack convert.
3749 DebugLoc dl = Op.getDebugLoc();
3750 EVT ResVT = Op.getValueType();
3751 assert(Op.getNumOperands() == 2);
3752 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3753 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3755 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3756 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3757 InVec = Op.getOperand(1);
3758 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3759 unsigned NumElts = ResVT.getVectorNumElements();
3760 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3761 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3762 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3764 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3765 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3766 Mask[0] = 0; Mask[1] = 2;
3767 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3769 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3772 // v8i16 shuffles - Prefer shuffles in the following order:
3773 // 1. [all] pshuflw, pshufhw, optional move
3774 // 2. [ssse3] 1 x pshufb
3775 // 3. [ssse3] 2 x pshufb + 1 x por
3776 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3778 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3779 SelectionDAG &DAG, X86TargetLowering &TLI) {
3780 SDValue V1 = SVOp->getOperand(0);
3781 SDValue V2 = SVOp->getOperand(1);
3782 DebugLoc dl = SVOp->getDebugLoc();
3783 SmallVector<int, 8> MaskVals;
3785 // Determine if more than 1 of the words in each of the low and high quadwords
3786 // of the result come from the same quadword of one of the two inputs. Undef
3787 // mask values count as coming from any quadword, for better codegen.
3788 SmallVector<unsigned, 4> LoQuad(4);
3789 SmallVector<unsigned, 4> HiQuad(4);
3790 BitVector InputQuads(4);
3791 for (unsigned i = 0; i < 8; ++i) {
3792 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3793 int EltIdx = SVOp->getMaskElt(i);
3794 MaskVals.push_back(EltIdx);
3803 InputQuads.set(EltIdx / 4);
3806 int BestLoQuad = -1;
3807 unsigned MaxQuad = 1;
3808 for (unsigned i = 0; i < 4; ++i) {
3809 if (LoQuad[i] > MaxQuad) {
3811 MaxQuad = LoQuad[i];
3815 int BestHiQuad = -1;
3817 for (unsigned i = 0; i < 4; ++i) {
3818 if (HiQuad[i] > MaxQuad) {
3820 MaxQuad = HiQuad[i];
3824 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3825 // of the two input vectors, shuffle them into one input vector so only a
3826 // single pshufb instruction is necessary. If There are more than 2 input
3827 // quads, disable the next transformation since it does not help SSSE3.
3828 bool V1Used = InputQuads[0] || InputQuads[1];
3829 bool V2Used = InputQuads[2] || InputQuads[3];
3830 if (TLI.getSubtarget()->hasSSSE3()) {
3831 if (InputQuads.count() == 2 && V1Used && V2Used) {
3832 BestLoQuad = InputQuads.find_first();
3833 BestHiQuad = InputQuads.find_next(BestLoQuad);
3835 if (InputQuads.count() > 2) {
3841 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3842 // the shuffle mask. If a quad is scored as -1, that means that it contains
3843 // words from all 4 input quadwords.
3845 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3846 SmallVector<int, 8> MaskV;
3847 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3848 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3849 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3850 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3851 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3852 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3854 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3855 // source words for the shuffle, to aid later transformations.
3856 bool AllWordsInNewV = true;
3857 bool InOrder[2] = { true, true };
3858 for (unsigned i = 0; i != 8; ++i) {
3859 int idx = MaskVals[i];
3861 InOrder[i/4] = false;
3862 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3864 AllWordsInNewV = false;
3868 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3869 if (AllWordsInNewV) {
3870 for (int i = 0; i != 8; ++i) {
3871 int idx = MaskVals[i];
3874 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3875 if ((idx != i) && idx < 4)
3877 if ((idx != i) && idx > 3)
3886 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3887 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3888 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3889 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3890 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3894 // If we have SSSE3, and all words of the result are from 1 input vector,
3895 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3896 // is present, fall back to case 4.
3897 if (TLI.getSubtarget()->hasSSSE3()) {
3898 SmallVector<SDValue,16> pshufbMask;
3900 // If we have elements from both input vectors, set the high bit of the
3901 // shuffle mask element to zero out elements that come from V2 in the V1
3902 // mask, and elements that come from V1 in the V2 mask, so that the two
3903 // results can be OR'd together.
3904 bool TwoInputs = V1Used && V2Used;
3905 for (unsigned i = 0; i != 8; ++i) {
3906 int EltIdx = MaskVals[i] * 2;
3907 if (TwoInputs && (EltIdx >= 16)) {
3908 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3909 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3912 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3913 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3915 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3916 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3917 DAG.getNode(ISD::BUILD_VECTOR, dl,
3918 MVT::v16i8, &pshufbMask[0], 16));
3920 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3922 // Calculate the shuffle mask for the second input, shuffle it, and
3923 // OR it with the first shuffled input.
3925 for (unsigned i = 0; i != 8; ++i) {
3926 int EltIdx = MaskVals[i] * 2;
3928 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3929 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3932 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3933 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3935 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3936 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3937 DAG.getNode(ISD::BUILD_VECTOR, dl,
3938 MVT::v16i8, &pshufbMask[0], 16));
3939 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3940 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3943 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3944 // and update MaskVals with new element order.
3945 BitVector InOrder(8);
3946 if (BestLoQuad >= 0) {
3947 SmallVector<int, 8> MaskV;
3948 for (int i = 0; i != 4; ++i) {
3949 int idx = MaskVals[i];
3951 MaskV.push_back(-1);
3953 } else if ((idx / 4) == BestLoQuad) {
3954 MaskV.push_back(idx & 3);
3957 MaskV.push_back(-1);
3960 for (unsigned i = 4; i != 8; ++i)
3962 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3966 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3967 // and update MaskVals with the new element order.
3968 if (BestHiQuad >= 0) {
3969 SmallVector<int, 8> MaskV;
3970 for (unsigned i = 0; i != 4; ++i)
3972 for (unsigned i = 4; i != 8; ++i) {
3973 int idx = MaskVals[i];
3975 MaskV.push_back(-1);
3977 } else if ((idx / 4) == BestHiQuad) {
3978 MaskV.push_back((idx & 3) + 4);
3981 MaskV.push_back(-1);
3984 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3988 // In case BestHi & BestLo were both -1, which means each quadword has a word
3989 // from each of the four input quadwords, calculate the InOrder bitvector now
3990 // before falling through to the insert/extract cleanup.
3991 if (BestLoQuad == -1 && BestHiQuad == -1) {
3993 for (int i = 0; i != 8; ++i)
3994 if (MaskVals[i] < 0 || MaskVals[i] == i)
3998 // The other elements are put in the right place using pextrw and pinsrw.
3999 for (unsigned i = 0; i != 8; ++i) {
4002 int EltIdx = MaskVals[i];
4005 SDValue ExtOp = (EltIdx < 8)
4006 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4007 DAG.getIntPtrConstant(EltIdx))
4008 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4009 DAG.getIntPtrConstant(EltIdx - 8));
4010 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4011 DAG.getIntPtrConstant(i));
4016 // v16i8 shuffles - Prefer shuffles in the following order:
4017 // 1. [ssse3] 1 x pshufb
4018 // 2. [ssse3] 2 x pshufb + 1 x por
4019 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4021 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4022 SelectionDAG &DAG, X86TargetLowering &TLI) {
4023 SDValue V1 = SVOp->getOperand(0);
4024 SDValue V2 = SVOp->getOperand(1);
4025 DebugLoc dl = SVOp->getDebugLoc();
4026 SmallVector<int, 16> MaskVals;
4027 SVOp->getMask(MaskVals);
4029 // If we have SSSE3, case 1 is generated when all result bytes come from
4030 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4031 // present, fall back to case 3.
4032 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4035 for (unsigned i = 0; i < 16; ++i) {
4036 int EltIdx = MaskVals[i];
4045 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4046 if (TLI.getSubtarget()->hasSSSE3()) {
4047 SmallVector<SDValue,16> pshufbMask;
4049 // If all result elements are from one input vector, then only translate
4050 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4052 // Otherwise, we have elements from both input vectors, and must zero out
4053 // elements that come from V2 in the first mask, and V1 in the second mask
4054 // so that we can OR them together.
4055 bool TwoInputs = !(V1Only || V2Only);
4056 for (unsigned i = 0; i != 16; ++i) {
4057 int EltIdx = MaskVals[i];
4058 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4059 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4062 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4064 // If all the elements are from V2, assign it to V1 and return after
4065 // building the first pshufb.
4068 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4069 DAG.getNode(ISD::BUILD_VECTOR, dl,
4070 MVT::v16i8, &pshufbMask[0], 16));
4074 // Calculate the shuffle mask for the second input, shuffle it, and
4075 // OR it with the first shuffled input.
4077 for (unsigned i = 0; i != 16; ++i) {
4078 int EltIdx = MaskVals[i];
4080 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4083 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4085 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4086 DAG.getNode(ISD::BUILD_VECTOR, dl,
4087 MVT::v16i8, &pshufbMask[0], 16));
4088 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4091 // No SSSE3 - Calculate in place words and then fix all out of place words
4092 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4093 // the 16 different words that comprise the two doublequadword input vectors.
4094 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4095 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4096 SDValue NewV = V2Only ? V2 : V1;
4097 for (int i = 0; i != 8; ++i) {
4098 int Elt0 = MaskVals[i*2];
4099 int Elt1 = MaskVals[i*2+1];
4101 // This word of the result is all undef, skip it.
4102 if (Elt0 < 0 && Elt1 < 0)
4105 // This word of the result is already in the correct place, skip it.
4106 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4108 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4111 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4112 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4115 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4116 // using a single extract together, load it and store it.
4117 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4118 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4119 DAG.getIntPtrConstant(Elt1 / 2));
4120 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4121 DAG.getIntPtrConstant(i));
4125 // If Elt1 is defined, extract it from the appropriate source. If the
4126 // source byte is not also odd, shift the extracted word left 8 bits
4127 // otherwise clear the bottom 8 bits if we need to do an or.
4129 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4130 DAG.getIntPtrConstant(Elt1 / 2));
4131 if ((Elt1 & 1) == 0)
4132 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4133 DAG.getConstant(8, TLI.getShiftAmountTy()));
4135 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4136 DAG.getConstant(0xFF00, MVT::i16));
4138 // If Elt0 is defined, extract it from the appropriate source. If the
4139 // source byte is not also even, shift the extracted word right 8 bits. If
4140 // Elt1 was also defined, OR the extracted values together before
4141 // inserting them in the result.
4143 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4144 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4145 if ((Elt0 & 1) != 0)
4146 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4147 DAG.getConstant(8, TLI.getShiftAmountTy()));
4149 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4150 DAG.getConstant(0x00FF, MVT::i16));
4151 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4154 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4155 DAG.getIntPtrConstant(i));
4157 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4160 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4161 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4162 /// done when every pair / quad of shuffle mask elements point to elements in
4163 /// the right sequence. e.g.
4164 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4166 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4168 TargetLowering &TLI, DebugLoc dl) {
4169 EVT VT = SVOp->getValueType(0);
4170 SDValue V1 = SVOp->getOperand(0);
4171 SDValue V2 = SVOp->getOperand(1);
4172 unsigned NumElems = VT.getVectorNumElements();
4173 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4174 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4175 EVT MaskEltVT = MaskVT.getVectorElementType();
4177 switch (VT.getSimpleVT().SimpleTy) {
4178 default: assert(false && "Unexpected!");
4179 case MVT::v4f32: NewVT = MVT::v2f64; break;
4180 case MVT::v4i32: NewVT = MVT::v2i64; break;
4181 case MVT::v8i16: NewVT = MVT::v4i32; break;
4182 case MVT::v16i8: NewVT = MVT::v4i32; break;
4185 if (NewWidth == 2) {
4191 int Scale = NumElems / NewWidth;
4192 SmallVector<int, 8> MaskVec;
4193 for (unsigned i = 0; i < NumElems; i += Scale) {
4195 for (int j = 0; j < Scale; ++j) {
4196 int EltIdx = SVOp->getMaskElt(i+j);
4200 StartIdx = EltIdx - (EltIdx % Scale);
4201 if (EltIdx != StartIdx + j)
4205 MaskVec.push_back(-1);
4207 MaskVec.push_back(StartIdx / Scale);
4210 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4211 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4212 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4215 /// getVZextMovL - Return a zero-extending vector move low node.
4217 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4218 SDValue SrcOp, SelectionDAG &DAG,
4219 const X86Subtarget *Subtarget, DebugLoc dl) {
4220 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4221 LoadSDNode *LD = NULL;
4222 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4223 LD = dyn_cast<LoadSDNode>(SrcOp);
4225 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4227 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4228 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4229 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4230 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4231 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4233 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4234 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4235 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4236 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4244 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4245 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4246 DAG.getNode(ISD::BIT_CONVERT, dl,
4250 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4253 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4254 SDValue V1 = SVOp->getOperand(0);
4255 SDValue V2 = SVOp->getOperand(1);
4256 DebugLoc dl = SVOp->getDebugLoc();
4257 EVT VT = SVOp->getValueType(0);
4259 SmallVector<std::pair<int, int>, 8> Locs;
4261 SmallVector<int, 8> Mask1(4U, -1);
4262 SmallVector<int, 8> PermMask;
4263 SVOp->getMask(PermMask);
4267 for (unsigned i = 0; i != 4; ++i) {
4268 int Idx = PermMask[i];
4270 Locs[i] = std::make_pair(-1, -1);
4272 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4274 Locs[i] = std::make_pair(0, NumLo);
4278 Locs[i] = std::make_pair(1, NumHi);
4280 Mask1[2+NumHi] = Idx;
4286 if (NumLo <= 2 && NumHi <= 2) {
4287 // If no more than two elements come from either vector. This can be
4288 // implemented with two shuffles. First shuffle gather the elements.
4289 // The second shuffle, which takes the first shuffle as both of its
4290 // vector operands, put the elements into the right order.
4291 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4293 SmallVector<int, 8> Mask2(4U, -1);
4295 for (unsigned i = 0; i != 4; ++i) {
4296 if (Locs[i].first == -1)
4299 unsigned Idx = (i < 2) ? 0 : 4;
4300 Idx += Locs[i].first * 2 + Locs[i].second;
4305 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4306 } else if (NumLo == 3 || NumHi == 3) {
4307 // Otherwise, we must have three elements from one vector, call it X, and
4308 // one element from the other, call it Y. First, use a shufps to build an
4309 // intermediate vector with the one element from Y and the element from X
4310 // that will be in the same half in the final destination (the indexes don't
4311 // matter). Then, use a shufps to build the final vector, taking the half
4312 // containing the element from Y from the intermediate, and the other half
4315 // Normalize it so the 3 elements come from V1.
4316 CommuteVectorShuffleMask(PermMask, VT);
4320 // Find the element from V2.
4322 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4323 int Val = PermMask[HiIndex];
4330 Mask1[0] = PermMask[HiIndex];
4332 Mask1[2] = PermMask[HiIndex^1];
4334 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4337 Mask1[0] = PermMask[0];
4338 Mask1[1] = PermMask[1];
4339 Mask1[2] = HiIndex & 1 ? 6 : 4;
4340 Mask1[3] = HiIndex & 1 ? 4 : 6;
4341 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4343 Mask1[0] = HiIndex & 1 ? 2 : 0;
4344 Mask1[1] = HiIndex & 1 ? 0 : 2;
4345 Mask1[2] = PermMask[2];
4346 Mask1[3] = PermMask[3];
4351 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4355 // Break it into (shuffle shuffle_hi, shuffle_lo).
4357 SmallVector<int,8> LoMask(4U, -1);
4358 SmallVector<int,8> HiMask(4U, -1);
4360 SmallVector<int,8> *MaskPtr = &LoMask;
4361 unsigned MaskIdx = 0;
4364 for (unsigned i = 0; i != 4; ++i) {
4371 int Idx = PermMask[i];
4373 Locs[i] = std::make_pair(-1, -1);
4374 } else if (Idx < 4) {
4375 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4376 (*MaskPtr)[LoIdx] = Idx;
4379 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4380 (*MaskPtr)[HiIdx] = Idx;
4385 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4386 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4387 SmallVector<int, 8> MaskOps;
4388 for (unsigned i = 0; i != 4; ++i) {
4389 if (Locs[i].first == -1) {
4390 MaskOps.push_back(-1);
4392 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4393 MaskOps.push_back(Idx);
4396 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4400 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4401 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4402 SDValue V1 = Op.getOperand(0);
4403 SDValue V2 = Op.getOperand(1);
4404 EVT VT = Op.getValueType();
4405 DebugLoc dl = Op.getDebugLoc();
4406 unsigned NumElems = VT.getVectorNumElements();
4407 bool isMMX = VT.getSizeInBits() == 64;
4408 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4409 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4410 bool V1IsSplat = false;
4411 bool V2IsSplat = false;
4413 if (isZeroShuffle(SVOp))
4414 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4416 // Promote splats to v4f32.
4417 if (SVOp->isSplat()) {
4418 if (isMMX || NumElems < 4)
4420 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4423 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4425 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4426 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4427 if (NewOp.getNode())
4428 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4429 LowerVECTOR_SHUFFLE(NewOp, DAG));
4430 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4431 // FIXME: Figure out a cleaner way to do this.
4432 // Try to make use of movq to zero out the top part.
4433 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4434 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4435 if (NewOp.getNode()) {
4436 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4437 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4438 DAG, Subtarget, dl);
4440 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4441 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4442 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4443 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4444 DAG, Subtarget, dl);
4448 if (X86::isPSHUFDMask(SVOp))
4451 // Check if this can be converted into a logical shift.
4452 bool isLeft = false;
4455 bool isShift = getSubtarget()->hasSSE2() &&
4456 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4457 if (isShift && ShVal.hasOneUse()) {
4458 // If the shifted value has multiple uses, it may be cheaper to use
4459 // v_set0 + movlhps or movhlps, etc.
4460 EVT EltVT = VT.getVectorElementType();
4461 ShAmt *= EltVT.getSizeInBits();
4462 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4465 if (X86::isMOVLMask(SVOp)) {
4468 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4469 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4474 // FIXME: fold these into legal mask.
4475 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4476 X86::isMOVSLDUPMask(SVOp) ||
4477 X86::isMOVHLPSMask(SVOp) ||
4478 X86::isMOVLHPSMask(SVOp) ||
4479 X86::isMOVLPMask(SVOp)))
4482 if (ShouldXformToMOVHLPS(SVOp) ||
4483 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4484 return CommuteVectorShuffle(SVOp, DAG);
4487 // No better options. Use a vshl / vsrl.
4488 EVT EltVT = VT.getVectorElementType();
4489 ShAmt *= EltVT.getSizeInBits();
4490 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4493 bool Commuted = false;
4494 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4495 // 1,1,1,1 -> v8i16 though.
4496 V1IsSplat = isSplatVector(V1.getNode());
4497 V2IsSplat = isSplatVector(V2.getNode());
4499 // Canonicalize the splat or undef, if present, to be on the RHS.
4500 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4501 Op = CommuteVectorShuffle(SVOp, DAG);
4502 SVOp = cast<ShuffleVectorSDNode>(Op);
4503 V1 = SVOp->getOperand(0);
4504 V2 = SVOp->getOperand(1);
4505 std::swap(V1IsSplat, V2IsSplat);
4506 std::swap(V1IsUndef, V2IsUndef);
4510 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4511 // Shuffling low element of v1 into undef, just return v1.
4514 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4515 // the instruction selector will not match, so get a canonical MOVL with
4516 // swapped operands to undo the commute.
4517 return getMOVL(DAG, dl, VT, V2, V1);
4520 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4521 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4522 X86::isUNPCKLMask(SVOp) ||
4523 X86::isUNPCKHMask(SVOp))
4527 // Normalize mask so all entries that point to V2 points to its first
4528 // element then try to match unpck{h|l} again. If match, return a
4529 // new vector_shuffle with the corrected mask.
4530 SDValue NewMask = NormalizeMask(SVOp, DAG);
4531 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4532 if (NSVOp != SVOp) {
4533 if (X86::isUNPCKLMask(NSVOp, true)) {
4535 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4542 // Commute is back and try unpck* again.
4543 // FIXME: this seems wrong.
4544 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4545 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4546 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4547 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4548 X86::isUNPCKLMask(NewSVOp) ||
4549 X86::isUNPCKHMask(NewSVOp))
4553 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4555 // Normalize the node to match x86 shuffle ops if needed
4556 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4557 return CommuteVectorShuffle(SVOp, DAG);
4559 // Check for legal shuffle and return?
4560 SmallVector<int, 16> PermMask;
4561 SVOp->getMask(PermMask);
4562 if (isShuffleMaskLegal(PermMask, VT))
4565 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4566 if (VT == MVT::v8i16) {
4567 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4568 if (NewOp.getNode())
4572 if (VT == MVT::v16i8) {
4573 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4574 if (NewOp.getNode())
4578 // Handle all 4 wide cases with a number of shuffles except for MMX.
4579 if (NumElems == 4 && !isMMX)
4580 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4586 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4587 SelectionDAG &DAG) {
4588 EVT VT = Op.getValueType();
4589 DebugLoc dl = Op.getDebugLoc();
4590 if (VT.getSizeInBits() == 8) {
4591 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4592 Op.getOperand(0), Op.getOperand(1));
4593 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4594 DAG.getValueType(VT));
4595 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4596 } else if (VT.getSizeInBits() == 16) {
4597 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4598 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4600 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4601 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4602 DAG.getNode(ISD::BIT_CONVERT, dl,
4606 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4607 Op.getOperand(0), Op.getOperand(1));
4608 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4609 DAG.getValueType(VT));
4610 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4611 } else if (VT == MVT::f32) {
4612 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4613 // the result back to FR32 register. It's only worth matching if the
4614 // result has a single use which is a store or a bitcast to i32. And in
4615 // the case of a store, it's not worth it if the index is a constant 0,
4616 // because a MOVSSmr can be used instead, which is smaller and faster.
4617 if (!Op.hasOneUse())
4619 SDNode *User = *Op.getNode()->use_begin();
4620 if ((User->getOpcode() != ISD::STORE ||
4621 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4622 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4623 (User->getOpcode() != ISD::BIT_CONVERT ||
4624 User->getValueType(0) != MVT::i32))
4626 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4627 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4630 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4631 } else if (VT == MVT::i32) {
4632 // ExtractPS works with constant index.
4633 if (isa<ConstantSDNode>(Op.getOperand(1)))
4641 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4642 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4645 if (Subtarget->hasSSE41()) {
4646 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4651 EVT VT = Op.getValueType();
4652 DebugLoc dl = Op.getDebugLoc();
4653 // TODO: handle v16i8.
4654 if (VT.getSizeInBits() == 16) {
4655 SDValue Vec = Op.getOperand(0);
4656 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4658 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4659 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4660 DAG.getNode(ISD::BIT_CONVERT, dl,
4663 // Transform it so it match pextrw which produces a 32-bit result.
4664 EVT EltVT = MVT::i32;
4665 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4666 Op.getOperand(0), Op.getOperand(1));
4667 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4668 DAG.getValueType(VT));
4669 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4670 } else if (VT.getSizeInBits() == 32) {
4671 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4675 // SHUFPS the element to the lowest double word, then movss.
4676 int Mask[4] = { Idx, -1, -1, -1 };
4677 EVT VVT = Op.getOperand(0).getValueType();
4678 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4679 DAG.getUNDEF(VVT), Mask);
4680 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4681 DAG.getIntPtrConstant(0));
4682 } else if (VT.getSizeInBits() == 64) {
4683 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4684 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4685 // to match extract_elt for f64.
4686 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4690 // UNPCKHPD the element to the lowest double word, then movsd.
4691 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4692 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4693 int Mask[2] = { 1, -1 };
4694 EVT VVT = Op.getOperand(0).getValueType();
4695 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4696 DAG.getUNDEF(VVT), Mask);
4697 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4698 DAG.getIntPtrConstant(0));
4705 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4706 EVT VT = Op.getValueType();
4707 EVT EltVT = VT.getVectorElementType();
4708 DebugLoc dl = Op.getDebugLoc();
4710 SDValue N0 = Op.getOperand(0);
4711 SDValue N1 = Op.getOperand(1);
4712 SDValue N2 = Op.getOperand(2);
4714 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4715 isa<ConstantSDNode>(N2)) {
4716 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4718 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4720 if (N1.getValueType() != MVT::i32)
4721 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4722 if (N2.getValueType() != MVT::i32)
4723 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4724 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4725 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4726 // Bits [7:6] of the constant are the source select. This will always be
4727 // zero here. The DAG Combiner may combine an extract_elt index into these
4728 // bits. For example (insert (extract, 3), 2) could be matched by putting
4729 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4730 // Bits [5:4] of the constant are the destination select. This is the
4731 // value of the incoming immediate.
4732 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4733 // combine either bitwise AND or insert of float 0.0 to set these bits.
4734 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4735 // Create this as a scalar to vector..
4736 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4737 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4738 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4739 // PINSR* works with constant index.
4746 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4747 EVT VT = Op.getValueType();
4748 EVT EltVT = VT.getVectorElementType();
4750 if (Subtarget->hasSSE41())
4751 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4753 if (EltVT == MVT::i8)
4756 DebugLoc dl = Op.getDebugLoc();
4757 SDValue N0 = Op.getOperand(0);
4758 SDValue N1 = Op.getOperand(1);
4759 SDValue N2 = Op.getOperand(2);
4761 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4762 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4763 // as its second argument.
4764 if (N1.getValueType() != MVT::i32)
4765 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4766 if (N2.getValueType() != MVT::i32)
4767 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4768 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4774 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4775 DebugLoc dl = Op.getDebugLoc();
4776 if (Op.getValueType() == MVT::v2f32)
4777 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4778 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4779 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4780 Op.getOperand(0))));
4782 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4783 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4785 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4786 EVT VT = MVT::v2i32;
4787 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4794 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4795 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4798 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4799 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4800 // one of the above mentioned nodes. It has to be wrapped because otherwise
4801 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4802 // be used to form addressing mode. These wrapped nodes will be selected
4805 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4806 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4808 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4810 unsigned char OpFlag = 0;
4811 unsigned WrapperKind = X86ISD::Wrapper;
4812 CodeModel::Model M = getTargetMachine().getCodeModel();
4814 if (Subtarget->isPICStyleRIPRel() &&
4815 (M == CodeModel::Small || M == CodeModel::Kernel))
4816 WrapperKind = X86ISD::WrapperRIP;
4817 else if (Subtarget->isPICStyleGOT())
4818 OpFlag = X86II::MO_GOTOFF;
4819 else if (Subtarget->isPICStyleStubPIC())
4820 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4822 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4824 CP->getOffset(), OpFlag);
4825 DebugLoc DL = CP->getDebugLoc();
4826 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4827 // With PIC, the address is actually $g + Offset.
4829 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4830 DAG.getNode(X86ISD::GlobalBaseReg,
4831 DebugLoc::getUnknownLoc(), getPointerTy()),
4838 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4839 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4841 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4843 unsigned char OpFlag = 0;
4844 unsigned WrapperKind = X86ISD::Wrapper;
4845 CodeModel::Model M = getTargetMachine().getCodeModel();
4847 if (Subtarget->isPICStyleRIPRel() &&
4848 (M == CodeModel::Small || M == CodeModel::Kernel))
4849 WrapperKind = X86ISD::WrapperRIP;
4850 else if (Subtarget->isPICStyleGOT())
4851 OpFlag = X86II::MO_GOTOFF;
4852 else if (Subtarget->isPICStyleStubPIC())
4853 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4855 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4857 DebugLoc DL = JT->getDebugLoc();
4858 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4860 // With PIC, the address is actually $g + Offset.
4862 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4863 DAG.getNode(X86ISD::GlobalBaseReg,
4864 DebugLoc::getUnknownLoc(), getPointerTy()),
4872 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4873 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4875 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4877 unsigned char OpFlag = 0;
4878 unsigned WrapperKind = X86ISD::Wrapper;
4879 CodeModel::Model M = getTargetMachine().getCodeModel();
4881 if (Subtarget->isPICStyleRIPRel() &&
4882 (M == CodeModel::Small || M == CodeModel::Kernel))
4883 WrapperKind = X86ISD::WrapperRIP;
4884 else if (Subtarget->isPICStyleGOT())
4885 OpFlag = X86II::MO_GOTOFF;
4886 else if (Subtarget->isPICStyleStubPIC())
4887 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4889 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4891 DebugLoc DL = Op.getDebugLoc();
4892 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4895 // With PIC, the address is actually $g + Offset.
4896 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4897 !Subtarget->is64Bit()) {
4898 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4899 DAG.getNode(X86ISD::GlobalBaseReg,
4900 DebugLoc::getUnknownLoc(),
4909 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4910 // Create the TargetBlockAddressAddress node.
4911 unsigned char OpFlags =
4912 Subtarget->ClassifyBlockAddressReference();
4913 CodeModel::Model M = getTargetMachine().getCodeModel();
4914 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4915 DebugLoc dl = Op.getDebugLoc();
4916 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4917 /*isTarget=*/true, OpFlags);
4919 if (Subtarget->isPICStyleRIPRel() &&
4920 (M == CodeModel::Small || M == CodeModel::Kernel))
4921 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4923 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4925 // With PIC, the address is actually $g + Offset.
4926 if (isGlobalRelativeToPICBase(OpFlags)) {
4927 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4928 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4936 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4938 SelectionDAG &DAG) const {
4939 // Create the TargetGlobalAddress node, folding in the constant
4940 // offset if it is legal.
4941 unsigned char OpFlags =
4942 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4943 CodeModel::Model M = getTargetMachine().getCodeModel();
4945 if (OpFlags == X86II::MO_NO_FLAG &&
4946 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4947 // A direct static reference to a global.
4948 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4951 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4954 if (Subtarget->isPICStyleRIPRel() &&
4955 (M == CodeModel::Small || M == CodeModel::Kernel))
4956 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4958 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4960 // With PIC, the address is actually $g + Offset.
4961 if (isGlobalRelativeToPICBase(OpFlags)) {
4962 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4963 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4967 // For globals that require a load from a stub to get the address, emit the
4969 if (isGlobalStubReference(OpFlags))
4970 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4971 PseudoSourceValue::getGOT(), 0);
4973 // If there was a non-zero offset that we didn't fold, create an explicit
4976 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4977 DAG.getConstant(Offset, getPointerTy()));
4983 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4984 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4985 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4986 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4990 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4991 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4992 unsigned char OperandFlags) {
4993 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4994 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4995 DebugLoc dl = GA->getDebugLoc();
4996 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4997 GA->getValueType(0),
5001 SDValue Ops[] = { Chain, TGA, *InFlag };
5002 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5004 SDValue Ops[] = { Chain, TGA };
5005 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5008 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5009 MFI->setHasCalls(true);
5011 SDValue Flag = Chain.getValue(1);
5012 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5015 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5017 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5020 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5021 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5022 DAG.getNode(X86ISD::GlobalBaseReg,
5023 DebugLoc::getUnknownLoc(),
5025 InFlag = Chain.getValue(1);
5027 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5030 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5032 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5034 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5035 X86::RAX, X86II::MO_TLSGD);
5038 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5039 // "local exec" model.
5040 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5041 const EVT PtrVT, TLSModel::Model model,
5043 DebugLoc dl = GA->getDebugLoc();
5044 // Get the Thread Pointer
5045 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5046 DebugLoc::getUnknownLoc(), PtrVT,
5047 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5050 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5053 unsigned char OperandFlags = 0;
5054 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5056 unsigned WrapperKind = X86ISD::Wrapper;
5057 if (model == TLSModel::LocalExec) {
5058 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5059 } else if (is64Bit) {
5060 assert(model == TLSModel::InitialExec);
5061 OperandFlags = X86II::MO_GOTTPOFF;
5062 WrapperKind = X86ISD::WrapperRIP;
5064 assert(model == TLSModel::InitialExec);
5065 OperandFlags = X86II::MO_INDNTPOFF;
5068 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5070 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5071 GA->getOffset(), OperandFlags);
5072 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5074 if (model == TLSModel::InitialExec)
5075 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5076 PseudoSourceValue::getGOT(), 0);
5078 // The address of the thread local variable is the add of the thread
5079 // pointer with the offset of the variable.
5080 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5084 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5085 // TODO: implement the "local dynamic" model
5086 // TODO: implement the "initial exec"model for pic executables
5087 assert(Subtarget->isTargetELF() &&
5088 "TLS not implemented for non-ELF targets");
5089 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5090 const GlobalValue *GV = GA->getGlobal();
5092 // If GV is an alias then use the aliasee for determining
5093 // thread-localness.
5094 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5095 GV = GA->resolveAliasedGlobal(false);
5097 TLSModel::Model model = getTLSModel(GV,
5098 getTargetMachine().getRelocationModel());
5101 case TLSModel::GeneralDynamic:
5102 case TLSModel::LocalDynamic: // not implemented
5103 if (Subtarget->is64Bit())
5104 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5105 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5107 case TLSModel::InitialExec:
5108 case TLSModel::LocalExec:
5109 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5110 Subtarget->is64Bit());
5113 llvm_unreachable("Unreachable");
5118 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5119 /// take a 2 x i32 value to shift plus a shift amount.
5120 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5121 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5122 EVT VT = Op.getValueType();
5123 unsigned VTBits = VT.getSizeInBits();
5124 DebugLoc dl = Op.getDebugLoc();
5125 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5126 SDValue ShOpLo = Op.getOperand(0);
5127 SDValue ShOpHi = Op.getOperand(1);
5128 SDValue ShAmt = Op.getOperand(2);
5129 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5130 DAG.getConstant(VTBits - 1, MVT::i8))
5131 : DAG.getConstant(0, VT);
5134 if (Op.getOpcode() == ISD::SHL_PARTS) {
5135 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5136 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5138 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5139 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5142 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5143 DAG.getConstant(VTBits, MVT::i8));
5144 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5145 AndNode, DAG.getConstant(0, MVT::i8));
5148 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5149 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5150 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5152 if (Op.getOpcode() == ISD::SHL_PARTS) {
5153 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5154 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5156 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5157 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5160 SDValue Ops[2] = { Lo, Hi };
5161 return DAG.getMergeValues(Ops, 2, dl);
5164 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5165 EVT SrcVT = Op.getOperand(0).getValueType();
5167 if (SrcVT.isVector()) {
5168 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5174 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5175 "Unknown SINT_TO_FP to lower!");
5177 // These are really Legal; return the operand so the caller accepts it as
5179 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5181 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5182 Subtarget->is64Bit()) {
5186 DebugLoc dl = Op.getDebugLoc();
5187 unsigned Size = SrcVT.getSizeInBits()/8;
5188 MachineFunction &MF = DAG.getMachineFunction();
5189 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5190 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5191 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5193 PseudoSourceValue::getFixedStack(SSFI), 0);
5194 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5197 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5199 SelectionDAG &DAG) {
5201 DebugLoc dl = Op.getDebugLoc();
5203 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5205 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5207 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5208 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5209 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5210 Tys, Ops, array_lengthof(Ops));
5213 Chain = Result.getValue(1);
5214 SDValue InFlag = Result.getValue(2);
5216 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5217 // shouldn't be necessary except that RFP cannot be live across
5218 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5219 MachineFunction &MF = DAG.getMachineFunction();
5220 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5221 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5222 Tys = DAG.getVTList(MVT::Other);
5224 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5226 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5227 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5228 PseudoSourceValue::getFixedStack(SSFI), 0);
5234 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5235 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5236 // This algorithm is not obvious. Here it is in C code, more or less:
5238 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5239 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5240 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5242 // Copy ints to xmm registers.
5243 __m128i xh = _mm_cvtsi32_si128( hi );
5244 __m128i xl = _mm_cvtsi32_si128( lo );
5246 // Combine into low half of a single xmm register.
5247 __m128i x = _mm_unpacklo_epi32( xh, xl );
5251 // Merge in appropriate exponents to give the integer bits the right
5253 x = _mm_unpacklo_epi32( x, exp );
5255 // Subtract away the biases to deal with the IEEE-754 double precision
5257 d = _mm_sub_pd( (__m128d) x, bias );
5259 // All conversions up to here are exact. The correctly rounded result is
5260 // calculated using the current rounding mode using the following
5262 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5263 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5264 // store doesn't really need to be here (except
5265 // maybe to zero the other double)
5270 DebugLoc dl = Op.getDebugLoc();
5271 LLVMContext *Context = DAG.getContext();
5273 // Build some magic constants.
5274 std::vector<Constant*> CV0;
5275 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5276 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5277 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5278 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5279 Constant *C0 = ConstantVector::get(CV0);
5280 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5282 std::vector<Constant*> CV1;
5284 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5286 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5287 Constant *C1 = ConstantVector::get(CV1);
5288 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5290 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5291 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5293 DAG.getIntPtrConstant(1)));
5294 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5295 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5297 DAG.getIntPtrConstant(0)));
5298 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5299 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5300 PseudoSourceValue::getConstantPool(), 0,
5302 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5303 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5304 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5305 PseudoSourceValue::getConstantPool(), 0,
5307 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5309 // Add the halves; easiest way is to swap them into another reg first.
5310 int ShufMask[2] = { 1, -1 };
5311 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5312 DAG.getUNDEF(MVT::v2f64), ShufMask);
5313 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5314 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5315 DAG.getIntPtrConstant(0));
5318 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5319 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5320 DebugLoc dl = Op.getDebugLoc();
5321 // FP constant to bias correct the final result.
5322 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5325 // Load the 32-bit value into an XMM register.
5326 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5327 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5329 DAG.getIntPtrConstant(0)));
5331 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5332 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5333 DAG.getIntPtrConstant(0));
5335 // Or the load with the bias.
5336 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5337 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5338 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5340 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5341 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5342 MVT::v2f64, Bias)));
5343 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5344 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5345 DAG.getIntPtrConstant(0));
5347 // Subtract the bias.
5348 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5350 // Handle final rounding.
5351 EVT DestVT = Op.getValueType();
5353 if (DestVT.bitsLT(MVT::f64)) {
5354 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5355 DAG.getIntPtrConstant(0));
5356 } else if (DestVT.bitsGT(MVT::f64)) {
5357 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5360 // Handle final rounding.
5364 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5365 SDValue N0 = Op.getOperand(0);
5366 DebugLoc dl = Op.getDebugLoc();
5368 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5369 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5370 // the optimization here.
5371 if (DAG.SignBitIsZero(N0))
5372 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5374 EVT SrcVT = N0.getValueType();
5375 if (SrcVT == MVT::i64) {
5376 // We only handle SSE2 f64 target here; caller can expand the rest.
5377 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5380 return LowerUINT_TO_FP_i64(Op, DAG);
5381 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5382 return LowerUINT_TO_FP_i32(Op, DAG);
5385 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5387 // Make a 64-bit buffer, and use it to build an FILD.
5388 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5389 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5390 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5391 getPointerTy(), StackSlot, WordOff);
5392 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5393 StackSlot, NULL, 0);
5394 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5395 OffsetSlot, NULL, 0);
5396 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5399 std::pair<SDValue,SDValue> X86TargetLowering::
5400 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5401 DebugLoc dl = Op.getDebugLoc();
5403 EVT DstTy = Op.getValueType();
5406 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5410 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5411 DstTy.getSimpleVT() >= MVT::i16 &&
5412 "Unknown FP_TO_SINT to lower!");
5414 // These are really Legal.
5415 if (DstTy == MVT::i32 &&
5416 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5417 return std::make_pair(SDValue(), SDValue());
5418 if (Subtarget->is64Bit() &&
5419 DstTy == MVT::i64 &&
5420 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5421 return std::make_pair(SDValue(), SDValue());
5423 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5425 MachineFunction &MF = DAG.getMachineFunction();
5426 unsigned MemSize = DstTy.getSizeInBits()/8;
5427 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5428 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5431 switch (DstTy.getSimpleVT().SimpleTy) {
5432 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5433 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5434 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5435 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5438 SDValue Chain = DAG.getEntryNode();
5439 SDValue Value = Op.getOperand(0);
5440 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5441 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5442 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5443 PseudoSourceValue::getFixedStack(SSFI), 0);
5444 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5446 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5448 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5449 Chain = Value.getValue(1);
5450 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5451 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5454 // Build the FP_TO_INT*_IN_MEM
5455 SDValue Ops[] = { Chain, Value, StackSlot };
5456 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5458 return std::make_pair(FIST, StackSlot);
5461 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5462 if (Op.getValueType().isVector()) {
5463 if (Op.getValueType() == MVT::v2i32 &&
5464 Op.getOperand(0).getValueType() == MVT::v2f64) {
5470 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5471 SDValue FIST = Vals.first, StackSlot = Vals.second;
5472 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5473 if (FIST.getNode() == 0) return Op;
5476 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5477 FIST, StackSlot, NULL, 0);
5480 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5481 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5482 SDValue FIST = Vals.first, StackSlot = Vals.second;
5483 assert(FIST.getNode() && "Unexpected failure");
5486 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5487 FIST, StackSlot, NULL, 0);
5490 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5491 LLVMContext *Context = DAG.getContext();
5492 DebugLoc dl = Op.getDebugLoc();
5493 EVT VT = Op.getValueType();
5496 EltVT = VT.getVectorElementType();
5497 std::vector<Constant*> CV;
5498 if (EltVT == MVT::f64) {
5499 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5503 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5509 Constant *C = ConstantVector::get(CV);
5510 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5511 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5512 PseudoSourceValue::getConstantPool(), 0,
5514 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5517 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5518 LLVMContext *Context = DAG.getContext();
5519 DebugLoc dl = Op.getDebugLoc();
5520 EVT VT = Op.getValueType();
5523 EltVT = VT.getVectorElementType();
5524 std::vector<Constant*> CV;
5525 if (EltVT == MVT::f64) {
5526 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5530 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5536 Constant *C = ConstantVector::get(CV);
5537 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5538 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5539 PseudoSourceValue::getConstantPool(), 0,
5541 if (VT.isVector()) {
5542 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5543 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5544 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5546 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5548 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5552 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5553 LLVMContext *Context = DAG.getContext();
5554 SDValue Op0 = Op.getOperand(0);
5555 SDValue Op1 = Op.getOperand(1);
5556 DebugLoc dl = Op.getDebugLoc();
5557 EVT VT = Op.getValueType();
5558 EVT SrcVT = Op1.getValueType();
5560 // If second operand is smaller, extend it first.
5561 if (SrcVT.bitsLT(VT)) {
5562 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5565 // And if it is bigger, shrink it first.
5566 if (SrcVT.bitsGT(VT)) {
5567 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5571 // At this point the operands and the result should have the same
5572 // type, and that won't be f80 since that is not custom lowered.
5574 // First get the sign bit of second operand.
5575 std::vector<Constant*> CV;
5576 if (SrcVT == MVT::f64) {
5577 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5578 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5580 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5581 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5582 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5583 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5585 Constant *C = ConstantVector::get(CV);
5586 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5587 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5588 PseudoSourceValue::getConstantPool(), 0,
5590 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5592 // Shift sign bit right or left if the two operands have different types.
5593 if (SrcVT.bitsGT(VT)) {
5594 // Op0 is MVT::f32, Op1 is MVT::f64.
5595 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5596 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5597 DAG.getConstant(32, MVT::i32));
5598 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5599 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5600 DAG.getIntPtrConstant(0));
5603 // Clear first operand sign bit.
5605 if (VT == MVT::f64) {
5606 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5607 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5609 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5610 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5611 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5612 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5614 C = ConstantVector::get(CV);
5615 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5616 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5617 PseudoSourceValue::getConstantPool(), 0,
5619 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5621 // Or the value with the sign bit.
5622 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5625 /// Emit nodes that will be selected as "test Op0,Op0", or something
5627 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5628 SelectionDAG &DAG) {
5629 DebugLoc dl = Op.getDebugLoc();
5631 // CF and OF aren't always set the way we want. Determine which
5632 // of these we need.
5633 bool NeedCF = false;
5634 bool NeedOF = false;
5636 case X86::COND_A: case X86::COND_AE:
5637 case X86::COND_B: case X86::COND_BE:
5640 case X86::COND_G: case X86::COND_GE:
5641 case X86::COND_L: case X86::COND_LE:
5642 case X86::COND_O: case X86::COND_NO:
5648 // See if we can use the EFLAGS value from the operand instead of
5649 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5650 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5651 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5652 unsigned Opcode = 0;
5653 unsigned NumOperands = 0;
5654 switch (Op.getNode()->getOpcode()) {
5656 // Due to an isel shortcoming, be conservative if this add is likely to
5657 // be selected as part of a load-modify-store instruction. When the root
5658 // node in a match is a store, isel doesn't know how to remap non-chain
5659 // non-flag uses of other nodes in the match, such as the ADD in this
5660 // case. This leads to the ADD being left around and reselected, with
5661 // the result being two adds in the output.
5662 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5663 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5664 if (UI->getOpcode() == ISD::STORE)
5666 if (ConstantSDNode *C =
5667 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5668 // An add of one will be selected as an INC.
5669 if (C->getAPIntValue() == 1) {
5670 Opcode = X86ISD::INC;
5674 // An add of negative one (subtract of one) will be selected as a DEC.
5675 if (C->getAPIntValue().isAllOnesValue()) {
5676 Opcode = X86ISD::DEC;
5681 // Otherwise use a regular EFLAGS-setting add.
5682 Opcode = X86ISD::ADD;
5686 // If the primary and result isn't used, don't bother using X86ISD::AND,
5687 // because a TEST instruction will be better.
5688 bool NonFlagUse = false;
5689 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5690 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5692 unsigned UOpNo = UI.getOperandNo();
5693 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5694 // Look pass truncate.
5695 UOpNo = User->use_begin().getOperandNo();
5696 User = *User->use_begin();
5698 if (User->getOpcode() != ISD::BRCOND &&
5699 User->getOpcode() != ISD::SETCC &&
5700 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5712 // Due to the ISEL shortcoming noted above, be conservative if this op is
5713 // likely to be selected as part of a load-modify-store instruction.
5714 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5715 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5716 if (UI->getOpcode() == ISD::STORE)
5718 // Otherwise use a regular EFLAGS-setting instruction.
5719 switch (Op.getNode()->getOpcode()) {
5720 case ISD::SUB: Opcode = X86ISD::SUB; break;
5721 case ISD::OR: Opcode = X86ISD::OR; break;
5722 case ISD::XOR: Opcode = X86ISD::XOR; break;
5723 case ISD::AND: Opcode = X86ISD::AND; break;
5724 default: llvm_unreachable("unexpected operator!");
5735 return SDValue(Op.getNode(), 1);
5741 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5742 SmallVector<SDValue, 4> Ops;
5743 for (unsigned i = 0; i != NumOperands; ++i)
5744 Ops.push_back(Op.getOperand(i));
5745 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5746 DAG.ReplaceAllUsesWith(Op, New);
5747 return SDValue(New.getNode(), 1);
5751 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5752 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5753 DAG.getConstant(0, Op.getValueType()));
5756 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5758 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5759 SelectionDAG &DAG) {
5760 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5761 if (C->getAPIntValue() == 0)
5762 return EmitTest(Op0, X86CC, DAG);
5764 DebugLoc dl = Op0.getDebugLoc();
5765 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5768 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5769 /// if it's possible.
5770 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5771 DebugLoc dl, SelectionDAG &DAG) {
5773 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5774 if (ConstantSDNode *Op010C =
5775 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5776 if (Op010C->getZExtValue() == 1) {
5777 LHS = Op0.getOperand(0);
5778 RHS = Op0.getOperand(1).getOperand(1);
5780 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5781 if (ConstantSDNode *Op000C =
5782 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5783 if (Op000C->getZExtValue() == 1) {
5784 LHS = Op0.getOperand(1);
5785 RHS = Op0.getOperand(0).getOperand(1);
5787 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5788 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5789 SDValue AndLHS = Op0.getOperand(0);
5790 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5791 LHS = AndLHS.getOperand(0);
5792 RHS = AndLHS.getOperand(1);
5796 if (LHS.getNode()) {
5797 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5798 // instruction. Since the shift amount is in-range-or-undefined, we know
5799 // that doing a bittest on the i16 value is ok. We extend to i32 because
5800 // the encoding for the i16 version is larger than the i32 version.
5801 if (LHS.getValueType() == MVT::i8)
5802 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5804 // If the operand types disagree, extend the shift amount to match. Since
5805 // BT ignores high bits (like shifts) we can use anyextend.
5806 if (LHS.getValueType() != RHS.getValueType())
5807 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5809 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5810 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5811 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5812 DAG.getConstant(Cond, MVT::i8), BT);
5818 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5819 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5820 SDValue Op0 = Op.getOperand(0);
5821 SDValue Op1 = Op.getOperand(1);
5822 DebugLoc dl = Op.getDebugLoc();
5823 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5825 // Optimize to BT if possible.
5826 // Lower (X & (1 << N)) == 0 to BT(X, N).
5827 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5828 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5829 if (Op0.getOpcode() == ISD::AND &&
5831 Op1.getOpcode() == ISD::Constant &&
5832 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5833 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5834 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5835 if (NewSetCC.getNode())
5839 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5840 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5841 if (X86CC == X86::COND_INVALID)
5844 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5846 // Use sbb x, x to materialize carry bit into a GPR.
5847 if (X86CC == X86::COND_B)
5848 return DAG.getNode(ISD::AND, dl, MVT::i8,
5849 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5850 DAG.getConstant(X86CC, MVT::i8), Cond),
5851 DAG.getConstant(1, MVT::i8));
5853 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5854 DAG.getConstant(X86CC, MVT::i8), Cond);
5857 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5859 SDValue Op0 = Op.getOperand(0);
5860 SDValue Op1 = Op.getOperand(1);
5861 SDValue CC = Op.getOperand(2);
5862 EVT VT = Op.getValueType();
5863 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5864 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5865 DebugLoc dl = Op.getDebugLoc();
5869 EVT VT0 = Op0.getValueType();
5870 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5871 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5874 switch (SetCCOpcode) {
5877 case ISD::SETEQ: SSECC = 0; break;
5879 case ISD::SETGT: Swap = true; // Fallthrough
5881 case ISD::SETOLT: SSECC = 1; break;
5883 case ISD::SETGE: Swap = true; // Fallthrough
5885 case ISD::SETOLE: SSECC = 2; break;
5886 case ISD::SETUO: SSECC = 3; break;
5888 case ISD::SETNE: SSECC = 4; break;
5889 case ISD::SETULE: Swap = true;
5890 case ISD::SETUGE: SSECC = 5; break;
5891 case ISD::SETULT: Swap = true;
5892 case ISD::SETUGT: SSECC = 6; break;
5893 case ISD::SETO: SSECC = 7; break;
5896 std::swap(Op0, Op1);
5898 // In the two special cases we can't handle, emit two comparisons.
5900 if (SetCCOpcode == ISD::SETUEQ) {
5902 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5903 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5904 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5906 else if (SetCCOpcode == ISD::SETONE) {
5908 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5909 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5910 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5912 llvm_unreachable("Illegal FP comparison");
5914 // Handle all other FP comparisons here.
5915 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5918 // We are handling one of the integer comparisons here. Since SSE only has
5919 // GT and EQ comparisons for integer, swapping operands and multiple
5920 // operations may be required for some comparisons.
5921 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5922 bool Swap = false, Invert = false, FlipSigns = false;
5924 switch (VT.getSimpleVT().SimpleTy) {
5927 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5929 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5931 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5932 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5935 switch (SetCCOpcode) {
5937 case ISD::SETNE: Invert = true;
5938 case ISD::SETEQ: Opc = EQOpc; break;
5939 case ISD::SETLT: Swap = true;
5940 case ISD::SETGT: Opc = GTOpc; break;
5941 case ISD::SETGE: Swap = true;
5942 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5943 case ISD::SETULT: Swap = true;
5944 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5945 case ISD::SETUGE: Swap = true;
5946 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5949 std::swap(Op0, Op1);
5951 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5952 // bits of the inputs before performing those operations.
5954 EVT EltVT = VT.getVectorElementType();
5955 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5957 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5958 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5960 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5961 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5964 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5966 // If the logical-not of the result is required, perform that now.
5968 Result = DAG.getNOT(dl, Result, VT);
5973 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5974 static bool isX86LogicalCmp(SDValue Op) {
5975 unsigned Opc = Op.getNode()->getOpcode();
5976 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5978 if (Op.getResNo() == 1 &&
5979 (Opc == X86ISD::ADD ||
5980 Opc == X86ISD::SUB ||
5981 Opc == X86ISD::SMUL ||
5982 Opc == X86ISD::UMUL ||
5983 Opc == X86ISD::INC ||
5984 Opc == X86ISD::DEC ||
5985 Opc == X86ISD::OR ||
5986 Opc == X86ISD::XOR ||
5987 Opc == X86ISD::AND))
5993 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5994 bool addTest = true;
5995 SDValue Cond = Op.getOperand(0);
5996 DebugLoc dl = Op.getDebugLoc();
5999 if (Cond.getOpcode() == ISD::SETCC) {
6000 SDValue NewCond = LowerSETCC(Cond, DAG);
6001 if (NewCond.getNode())
6005 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6006 SDValue Op1 = Op.getOperand(1);
6007 SDValue Op2 = Op.getOperand(2);
6008 if (Cond.getOpcode() == X86ISD::SETCC &&
6009 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6010 SDValue Cmp = Cond.getOperand(1);
6011 if (Cmp.getOpcode() == X86ISD::CMP) {
6012 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6013 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6014 ConstantSDNode *RHSC =
6015 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6016 if (N1C && N1C->isAllOnesValue() &&
6017 N2C && N2C->isNullValue() &&
6018 RHSC && RHSC->isNullValue()) {
6019 SDValue CmpOp0 = Cmp.getOperand(0);
6020 Cmp = DAG.getNode(X86ISD::CMP, dl, Op.getValueType(),
6021 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6022 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6023 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6028 // Look pass (and (setcc_carry (cmp ...)), 1).
6029 if (Cond.getOpcode() == ISD::AND &&
6030 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6031 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6032 if (C && C->getAPIntValue() == 1)
6033 Cond = Cond.getOperand(0);
6036 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6037 // setting operand in place of the X86ISD::SETCC.
6038 if (Cond.getOpcode() == X86ISD::SETCC ||
6039 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6040 CC = Cond.getOperand(0);
6042 SDValue Cmp = Cond.getOperand(1);
6043 unsigned Opc = Cmp.getOpcode();
6044 EVT VT = Op.getValueType();
6046 bool IllegalFPCMov = false;
6047 if (VT.isFloatingPoint() && !VT.isVector() &&
6048 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6049 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6051 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6052 Opc == X86ISD::BT) { // FIXME
6059 // Look pass the truncate.
6060 if (Cond.getOpcode() == ISD::TRUNCATE)
6061 Cond = Cond.getOperand(0);
6063 // We know the result of AND is compared against zero. Try to match
6065 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6066 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6067 if (NewSetCC.getNode()) {
6068 CC = NewSetCC.getOperand(0);
6069 Cond = NewSetCC.getOperand(1);
6076 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6077 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6080 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6081 // condition is true.
6082 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6083 SDValue Ops[] = { Op2, Op1, CC, Cond };
6084 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6087 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6088 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6089 // from the AND / OR.
6090 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6091 Opc = Op.getOpcode();
6092 if (Opc != ISD::OR && Opc != ISD::AND)
6094 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6095 Op.getOperand(0).hasOneUse() &&
6096 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6097 Op.getOperand(1).hasOneUse());
6100 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6101 // 1 and that the SETCC node has a single use.
6102 static bool isXor1OfSetCC(SDValue Op) {
6103 if (Op.getOpcode() != ISD::XOR)
6105 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6106 if (N1C && N1C->getAPIntValue() == 1) {
6107 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6108 Op.getOperand(0).hasOneUse();
6113 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6114 bool addTest = true;
6115 SDValue Chain = Op.getOperand(0);
6116 SDValue Cond = Op.getOperand(1);
6117 SDValue Dest = Op.getOperand(2);
6118 DebugLoc dl = Op.getDebugLoc();
6121 if (Cond.getOpcode() == ISD::SETCC) {
6122 SDValue NewCond = LowerSETCC(Cond, DAG);
6123 if (NewCond.getNode())
6127 // FIXME: LowerXALUO doesn't handle these!!
6128 else if (Cond.getOpcode() == X86ISD::ADD ||
6129 Cond.getOpcode() == X86ISD::SUB ||
6130 Cond.getOpcode() == X86ISD::SMUL ||
6131 Cond.getOpcode() == X86ISD::UMUL)
6132 Cond = LowerXALUO(Cond, DAG);
6135 // Look pass (and (setcc_carry (cmp ...)), 1).
6136 if (Cond.getOpcode() == ISD::AND &&
6137 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6138 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6139 if (C && C->getAPIntValue() == 1)
6140 Cond = Cond.getOperand(0);
6143 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6144 // setting operand in place of the X86ISD::SETCC.
6145 if (Cond.getOpcode() == X86ISD::SETCC ||
6146 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6147 CC = Cond.getOperand(0);
6149 SDValue Cmp = Cond.getOperand(1);
6150 unsigned Opc = Cmp.getOpcode();
6151 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6152 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6156 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6160 // These can only come from an arithmetic instruction with overflow,
6161 // e.g. SADDO, UADDO.
6162 Cond = Cond.getNode()->getOperand(1);
6169 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6170 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6171 if (CondOpc == ISD::OR) {
6172 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6173 // two branches instead of an explicit OR instruction with a
6175 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6176 isX86LogicalCmp(Cmp)) {
6177 CC = Cond.getOperand(0).getOperand(0);
6178 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6179 Chain, Dest, CC, Cmp);
6180 CC = Cond.getOperand(1).getOperand(0);
6184 } else { // ISD::AND
6185 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6186 // two branches instead of an explicit AND instruction with a
6187 // separate test. However, we only do this if this block doesn't
6188 // have a fall-through edge, because this requires an explicit
6189 // jmp when the condition is false.
6190 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6191 isX86LogicalCmp(Cmp) &&
6192 Op.getNode()->hasOneUse()) {
6193 X86::CondCode CCode =
6194 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6195 CCode = X86::GetOppositeBranchCondition(CCode);
6196 CC = DAG.getConstant(CCode, MVT::i8);
6197 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6198 // Look for an unconditional branch following this conditional branch.
6199 // We need this because we need to reverse the successors in order
6200 // to implement FCMP_OEQ.
6201 if (User.getOpcode() == ISD::BR) {
6202 SDValue FalseBB = User.getOperand(1);
6204 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6205 assert(NewBR == User);
6208 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6209 Chain, Dest, CC, Cmp);
6210 X86::CondCode CCode =
6211 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6212 CCode = X86::GetOppositeBranchCondition(CCode);
6213 CC = DAG.getConstant(CCode, MVT::i8);
6219 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6220 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6221 // It should be transformed during dag combiner except when the condition
6222 // is set by a arithmetics with overflow node.
6223 X86::CondCode CCode =
6224 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6225 CCode = X86::GetOppositeBranchCondition(CCode);
6226 CC = DAG.getConstant(CCode, MVT::i8);
6227 Cond = Cond.getOperand(0).getOperand(1);
6233 // Look pass the truncate.
6234 if (Cond.getOpcode() == ISD::TRUNCATE)
6235 Cond = Cond.getOperand(0);
6237 // We know the result of AND is compared against zero. Try to match
6239 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6240 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6241 if (NewSetCC.getNode()) {
6242 CC = NewSetCC.getOperand(0);
6243 Cond = NewSetCC.getOperand(1);
6250 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6251 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6253 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6254 Chain, Dest, CC, Cond);
6258 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6259 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6260 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6261 // that the guard pages used by the OS virtual memory manager are allocated in
6262 // correct sequence.
6264 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6265 SelectionDAG &DAG) {
6266 assert(Subtarget->isTargetCygMing() &&
6267 "This should be used only on Cygwin/Mingw targets");
6268 DebugLoc dl = Op.getDebugLoc();
6271 SDValue Chain = Op.getOperand(0);
6272 SDValue Size = Op.getOperand(1);
6273 // FIXME: Ensure alignment here
6277 EVT IntPtr = getPointerTy();
6278 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6280 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6282 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6283 Flag = Chain.getValue(1);
6285 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6286 SDValue Ops[] = { Chain,
6287 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6288 DAG.getRegister(X86::EAX, IntPtr),
6289 DAG.getRegister(X86StackPtr, SPTy),
6291 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6292 Flag = Chain.getValue(1);
6294 Chain = DAG.getCALLSEQ_END(Chain,
6295 DAG.getIntPtrConstant(0, true),
6296 DAG.getIntPtrConstant(0, true),
6299 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6301 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6302 return DAG.getMergeValues(Ops1, 2, dl);
6306 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6308 SDValue Dst, SDValue Src,
6309 SDValue Size, unsigned Align,
6311 uint64_t DstSVOff) {
6312 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6314 // If not DWORD aligned or size is more than the threshold, call the library.
6315 // The libc version is likely to be faster for these cases. It can use the
6316 // address value and run time information about the CPU.
6317 if ((Align & 3) != 0 ||
6319 ConstantSize->getZExtValue() >
6320 getSubtarget()->getMaxInlineSizeThreshold()) {
6321 SDValue InFlag(0, 0);
6323 // Check to see if there is a specialized entry-point for memory zeroing.
6324 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6326 if (const char *bzeroEntry = V &&
6327 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6328 EVT IntPtr = getPointerTy();
6329 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6330 TargetLowering::ArgListTy Args;
6331 TargetLowering::ArgListEntry Entry;
6333 Entry.Ty = IntPtrTy;
6334 Args.push_back(Entry);
6336 Args.push_back(Entry);
6337 std::pair<SDValue,SDValue> CallResult =
6338 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6339 false, false, false, false,
6340 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6341 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6342 DAG.GetOrdering(Chain.getNode()));
6343 return CallResult.second;
6346 // Otherwise have the target-independent code call memset.
6350 uint64_t SizeVal = ConstantSize->getZExtValue();
6351 SDValue InFlag(0, 0);
6354 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6355 unsigned BytesLeft = 0;
6356 bool TwoRepStos = false;
6359 uint64_t Val = ValC->getZExtValue() & 255;
6361 // If the value is a constant, then we can potentially use larger sets.
6362 switch (Align & 3) {
6363 case 2: // WORD aligned
6366 Val = (Val << 8) | Val;
6368 case 0: // DWORD aligned
6371 Val = (Val << 8) | Val;
6372 Val = (Val << 16) | Val;
6373 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6376 Val = (Val << 32) | Val;
6379 default: // Byte aligned
6382 Count = DAG.getIntPtrConstant(SizeVal);
6386 if (AVT.bitsGT(MVT::i8)) {
6387 unsigned UBytes = AVT.getSizeInBits() / 8;
6388 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6389 BytesLeft = SizeVal % UBytes;
6392 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6394 InFlag = Chain.getValue(1);
6397 Count = DAG.getIntPtrConstant(SizeVal);
6398 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6399 InFlag = Chain.getValue(1);
6402 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6405 InFlag = Chain.getValue(1);
6406 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6409 InFlag = Chain.getValue(1);
6411 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6412 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6413 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6416 InFlag = Chain.getValue(1);
6418 EVT CVT = Count.getValueType();
6419 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6420 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6421 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6424 InFlag = Chain.getValue(1);
6425 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6426 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6427 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6428 } else if (BytesLeft) {
6429 // Handle the last 1 - 7 bytes.
6430 unsigned Offset = SizeVal - BytesLeft;
6431 EVT AddrVT = Dst.getValueType();
6432 EVT SizeVT = Size.getValueType();
6434 Chain = DAG.getMemset(Chain, dl,
6435 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6436 DAG.getConstant(Offset, AddrVT)),
6438 DAG.getConstant(BytesLeft, SizeVT),
6439 Align, DstSV, DstSVOff + Offset);
6442 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6447 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6448 SDValue Chain, SDValue Dst, SDValue Src,
6449 SDValue Size, unsigned Align,
6451 const Value *DstSV, uint64_t DstSVOff,
6452 const Value *SrcSV, uint64_t SrcSVOff) {
6453 // This requires the copy size to be a constant, preferrably
6454 // within a subtarget-specific limit.
6455 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6458 uint64_t SizeVal = ConstantSize->getZExtValue();
6459 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6462 /// If not DWORD aligned, call the library.
6463 if ((Align & 3) != 0)
6468 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6471 unsigned UBytes = AVT.getSizeInBits() / 8;
6472 unsigned CountVal = SizeVal / UBytes;
6473 SDValue Count = DAG.getIntPtrConstant(CountVal);
6474 unsigned BytesLeft = SizeVal % UBytes;
6476 SDValue InFlag(0, 0);
6477 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6480 InFlag = Chain.getValue(1);
6481 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6484 InFlag = Chain.getValue(1);
6485 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6488 InFlag = Chain.getValue(1);
6490 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6491 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6492 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6493 array_lengthof(Ops));
6495 SmallVector<SDValue, 4> Results;
6496 Results.push_back(RepMovs);
6498 // Handle the last 1 - 7 bytes.
6499 unsigned Offset = SizeVal - BytesLeft;
6500 EVT DstVT = Dst.getValueType();
6501 EVT SrcVT = Src.getValueType();
6502 EVT SizeVT = Size.getValueType();
6503 Results.push_back(DAG.getMemcpy(Chain, dl,
6504 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6505 DAG.getConstant(Offset, DstVT)),
6506 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6507 DAG.getConstant(Offset, SrcVT)),
6508 DAG.getConstant(BytesLeft, SizeVT),
6509 Align, AlwaysInline,
6510 DstSV, DstSVOff + Offset,
6511 SrcSV, SrcSVOff + Offset));
6514 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6515 &Results[0], Results.size());
6518 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6519 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6520 DebugLoc dl = Op.getDebugLoc();
6522 if (!Subtarget->is64Bit()) {
6523 // vastart just stores the address of the VarArgsFrameIndex slot into the
6524 // memory location argument.
6525 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6526 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6530 // gp_offset (0 - 6 * 8)
6531 // fp_offset (48 - 48 + 8 * 16)
6532 // overflow_arg_area (point to parameters coming in memory).
6534 SmallVector<SDValue, 8> MemOps;
6535 SDValue FIN = Op.getOperand(1);
6537 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6538 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6540 MemOps.push_back(Store);
6543 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6544 FIN, DAG.getIntPtrConstant(4));
6545 Store = DAG.getStore(Op.getOperand(0), dl,
6546 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6548 MemOps.push_back(Store);
6550 // Store ptr to overflow_arg_area
6551 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6552 FIN, DAG.getIntPtrConstant(4));
6553 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6554 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6555 MemOps.push_back(Store);
6557 // Store ptr to reg_save_area.
6558 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6559 FIN, DAG.getIntPtrConstant(8));
6560 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6561 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6562 MemOps.push_back(Store);
6563 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6564 &MemOps[0], MemOps.size());
6567 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6568 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6569 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6570 SDValue Chain = Op.getOperand(0);
6571 SDValue SrcPtr = Op.getOperand(1);
6572 SDValue SrcSV = Op.getOperand(2);
6574 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6578 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6579 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6580 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6581 SDValue Chain = Op.getOperand(0);
6582 SDValue DstPtr = Op.getOperand(1);
6583 SDValue SrcPtr = Op.getOperand(2);
6584 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6585 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6586 DebugLoc dl = Op.getDebugLoc();
6588 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6589 DAG.getIntPtrConstant(24), 8, false,
6590 DstSV, 0, SrcSV, 0);
6594 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6595 DebugLoc dl = Op.getDebugLoc();
6596 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6598 default: return SDValue(); // Don't custom lower most intrinsics.
6599 // Comparison intrinsics.
6600 case Intrinsic::x86_sse_comieq_ss:
6601 case Intrinsic::x86_sse_comilt_ss:
6602 case Intrinsic::x86_sse_comile_ss:
6603 case Intrinsic::x86_sse_comigt_ss:
6604 case Intrinsic::x86_sse_comige_ss:
6605 case Intrinsic::x86_sse_comineq_ss:
6606 case Intrinsic::x86_sse_ucomieq_ss:
6607 case Intrinsic::x86_sse_ucomilt_ss:
6608 case Intrinsic::x86_sse_ucomile_ss:
6609 case Intrinsic::x86_sse_ucomigt_ss:
6610 case Intrinsic::x86_sse_ucomige_ss:
6611 case Intrinsic::x86_sse_ucomineq_ss:
6612 case Intrinsic::x86_sse2_comieq_sd:
6613 case Intrinsic::x86_sse2_comilt_sd:
6614 case Intrinsic::x86_sse2_comile_sd:
6615 case Intrinsic::x86_sse2_comigt_sd:
6616 case Intrinsic::x86_sse2_comige_sd:
6617 case Intrinsic::x86_sse2_comineq_sd:
6618 case Intrinsic::x86_sse2_ucomieq_sd:
6619 case Intrinsic::x86_sse2_ucomilt_sd:
6620 case Intrinsic::x86_sse2_ucomile_sd:
6621 case Intrinsic::x86_sse2_ucomigt_sd:
6622 case Intrinsic::x86_sse2_ucomige_sd:
6623 case Intrinsic::x86_sse2_ucomineq_sd: {
6625 ISD::CondCode CC = ISD::SETCC_INVALID;
6628 case Intrinsic::x86_sse_comieq_ss:
6629 case Intrinsic::x86_sse2_comieq_sd:
6633 case Intrinsic::x86_sse_comilt_ss:
6634 case Intrinsic::x86_sse2_comilt_sd:
6638 case Intrinsic::x86_sse_comile_ss:
6639 case Intrinsic::x86_sse2_comile_sd:
6643 case Intrinsic::x86_sse_comigt_ss:
6644 case Intrinsic::x86_sse2_comigt_sd:
6648 case Intrinsic::x86_sse_comige_ss:
6649 case Intrinsic::x86_sse2_comige_sd:
6653 case Intrinsic::x86_sse_comineq_ss:
6654 case Intrinsic::x86_sse2_comineq_sd:
6658 case Intrinsic::x86_sse_ucomieq_ss:
6659 case Intrinsic::x86_sse2_ucomieq_sd:
6660 Opc = X86ISD::UCOMI;
6663 case Intrinsic::x86_sse_ucomilt_ss:
6664 case Intrinsic::x86_sse2_ucomilt_sd:
6665 Opc = X86ISD::UCOMI;
6668 case Intrinsic::x86_sse_ucomile_ss:
6669 case Intrinsic::x86_sse2_ucomile_sd:
6670 Opc = X86ISD::UCOMI;
6673 case Intrinsic::x86_sse_ucomigt_ss:
6674 case Intrinsic::x86_sse2_ucomigt_sd:
6675 Opc = X86ISD::UCOMI;
6678 case Intrinsic::x86_sse_ucomige_ss:
6679 case Intrinsic::x86_sse2_ucomige_sd:
6680 Opc = X86ISD::UCOMI;
6683 case Intrinsic::x86_sse_ucomineq_ss:
6684 case Intrinsic::x86_sse2_ucomineq_sd:
6685 Opc = X86ISD::UCOMI;
6690 SDValue LHS = Op.getOperand(1);
6691 SDValue RHS = Op.getOperand(2);
6692 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6693 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6694 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6695 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6696 DAG.getConstant(X86CC, MVT::i8), Cond);
6697 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6699 // ptest intrinsics. The intrinsic these come from are designed to return
6700 // an integer value, not just an instruction so lower it to the ptest
6701 // pattern and a setcc for the result.
6702 case Intrinsic::x86_sse41_ptestz:
6703 case Intrinsic::x86_sse41_ptestc:
6704 case Intrinsic::x86_sse41_ptestnzc:{
6707 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6708 case Intrinsic::x86_sse41_ptestz:
6710 X86CC = X86::COND_E;
6712 case Intrinsic::x86_sse41_ptestc:
6714 X86CC = X86::COND_B;
6716 case Intrinsic::x86_sse41_ptestnzc:
6718 X86CC = X86::COND_A;
6722 SDValue LHS = Op.getOperand(1);
6723 SDValue RHS = Op.getOperand(2);
6724 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6725 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6726 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6727 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6730 // Fix vector shift instructions where the last operand is a non-immediate
6732 case Intrinsic::x86_sse2_pslli_w:
6733 case Intrinsic::x86_sse2_pslli_d:
6734 case Intrinsic::x86_sse2_pslli_q:
6735 case Intrinsic::x86_sse2_psrli_w:
6736 case Intrinsic::x86_sse2_psrli_d:
6737 case Intrinsic::x86_sse2_psrli_q:
6738 case Intrinsic::x86_sse2_psrai_w:
6739 case Intrinsic::x86_sse2_psrai_d:
6740 case Intrinsic::x86_mmx_pslli_w:
6741 case Intrinsic::x86_mmx_pslli_d:
6742 case Intrinsic::x86_mmx_pslli_q:
6743 case Intrinsic::x86_mmx_psrli_w:
6744 case Intrinsic::x86_mmx_psrli_d:
6745 case Intrinsic::x86_mmx_psrli_q:
6746 case Intrinsic::x86_mmx_psrai_w:
6747 case Intrinsic::x86_mmx_psrai_d: {
6748 SDValue ShAmt = Op.getOperand(2);
6749 if (isa<ConstantSDNode>(ShAmt))
6752 unsigned NewIntNo = 0;
6753 EVT ShAmtVT = MVT::v4i32;
6755 case Intrinsic::x86_sse2_pslli_w:
6756 NewIntNo = Intrinsic::x86_sse2_psll_w;
6758 case Intrinsic::x86_sse2_pslli_d:
6759 NewIntNo = Intrinsic::x86_sse2_psll_d;
6761 case Intrinsic::x86_sse2_pslli_q:
6762 NewIntNo = Intrinsic::x86_sse2_psll_q;
6764 case Intrinsic::x86_sse2_psrli_w:
6765 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6767 case Intrinsic::x86_sse2_psrli_d:
6768 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6770 case Intrinsic::x86_sse2_psrli_q:
6771 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6773 case Intrinsic::x86_sse2_psrai_w:
6774 NewIntNo = Intrinsic::x86_sse2_psra_w;
6776 case Intrinsic::x86_sse2_psrai_d:
6777 NewIntNo = Intrinsic::x86_sse2_psra_d;
6780 ShAmtVT = MVT::v2i32;
6782 case Intrinsic::x86_mmx_pslli_w:
6783 NewIntNo = Intrinsic::x86_mmx_psll_w;
6785 case Intrinsic::x86_mmx_pslli_d:
6786 NewIntNo = Intrinsic::x86_mmx_psll_d;
6788 case Intrinsic::x86_mmx_pslli_q:
6789 NewIntNo = Intrinsic::x86_mmx_psll_q;
6791 case Intrinsic::x86_mmx_psrli_w:
6792 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6794 case Intrinsic::x86_mmx_psrli_d:
6795 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6797 case Intrinsic::x86_mmx_psrli_q:
6798 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6800 case Intrinsic::x86_mmx_psrai_w:
6801 NewIntNo = Intrinsic::x86_mmx_psra_w;
6803 case Intrinsic::x86_mmx_psrai_d:
6804 NewIntNo = Intrinsic::x86_mmx_psra_d;
6806 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6812 // The vector shift intrinsics with scalars uses 32b shift amounts but
6813 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6817 ShOps[1] = DAG.getConstant(0, MVT::i32);
6818 if (ShAmtVT == MVT::v4i32) {
6819 ShOps[2] = DAG.getUNDEF(MVT::i32);
6820 ShOps[3] = DAG.getUNDEF(MVT::i32);
6821 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6823 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6826 EVT VT = Op.getValueType();
6827 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6828 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6829 DAG.getConstant(NewIntNo, MVT::i32),
6830 Op.getOperand(1), ShAmt);
6835 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6836 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6837 DebugLoc dl = Op.getDebugLoc();
6840 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6842 DAG.getConstant(TD->getPointerSize(),
6843 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6844 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6845 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6850 // Just load the return address.
6851 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6852 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6853 RetAddrFI, NULL, 0);
6856 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6857 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6858 MFI->setFrameAddressIsTaken(true);
6859 EVT VT = Op.getValueType();
6860 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6861 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6862 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6863 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6865 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6869 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6870 SelectionDAG &DAG) {
6871 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6874 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6876 MachineFunction &MF = DAG.getMachineFunction();
6877 SDValue Chain = Op.getOperand(0);
6878 SDValue Offset = Op.getOperand(1);
6879 SDValue Handler = Op.getOperand(2);
6880 DebugLoc dl = Op.getDebugLoc();
6882 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6884 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6886 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6887 DAG.getIntPtrConstant(-TD->getPointerSize()));
6888 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6889 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6890 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6891 MF.getRegInfo().addLiveOut(StoreAddrReg);
6893 return DAG.getNode(X86ISD::EH_RETURN, dl,
6895 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6898 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6899 SelectionDAG &DAG) {
6900 SDValue Root = Op.getOperand(0);
6901 SDValue Trmp = Op.getOperand(1); // trampoline
6902 SDValue FPtr = Op.getOperand(2); // nested function
6903 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6904 DebugLoc dl = Op.getDebugLoc();
6906 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6908 const X86InstrInfo *TII =
6909 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6911 if (Subtarget->is64Bit()) {
6912 SDValue OutChains[6];
6914 // Large code-model.
6916 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6917 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6919 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6920 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6922 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6924 // Load the pointer to the nested function into R11.
6925 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6926 SDValue Addr = Trmp;
6927 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6930 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6931 DAG.getConstant(2, MVT::i64));
6932 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6934 // Load the 'nest' parameter value into R10.
6935 // R10 is specified in X86CallingConv.td
6936 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6937 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6938 DAG.getConstant(10, MVT::i64));
6939 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6940 Addr, TrmpAddr, 10);
6942 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6943 DAG.getConstant(12, MVT::i64));
6944 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6946 // Jump to the nested function.
6947 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6948 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6949 DAG.getConstant(20, MVT::i64));
6950 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6951 Addr, TrmpAddr, 20);
6953 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6954 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6955 DAG.getConstant(22, MVT::i64));
6956 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6960 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6961 return DAG.getMergeValues(Ops, 2, dl);
6963 const Function *Func =
6964 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6965 CallingConv::ID CC = Func->getCallingConv();
6970 llvm_unreachable("Unsupported calling convention");
6971 case CallingConv::C:
6972 case CallingConv::X86_StdCall: {
6973 // Pass 'nest' parameter in ECX.
6974 // Must be kept in sync with X86CallingConv.td
6977 // Check that ECX wasn't needed by an 'inreg' parameter.
6978 const FunctionType *FTy = Func->getFunctionType();
6979 const AttrListPtr &Attrs = Func->getAttributes();
6981 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6982 unsigned InRegCount = 0;
6985 for (FunctionType::param_iterator I = FTy->param_begin(),
6986 E = FTy->param_end(); I != E; ++I, ++Idx)
6987 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6988 // FIXME: should only count parameters that are lowered to integers.
6989 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6991 if (InRegCount > 2) {
6992 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6997 case CallingConv::X86_FastCall:
6998 case CallingConv::Fast:
6999 // Pass 'nest' parameter in EAX.
7000 // Must be kept in sync with X86CallingConv.td
7005 SDValue OutChains[4];
7008 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7009 DAG.getConstant(10, MVT::i32));
7010 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7012 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7013 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7014 OutChains[0] = DAG.getStore(Root, dl,
7015 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7018 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7019 DAG.getConstant(1, MVT::i32));
7020 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7022 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7023 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7024 DAG.getConstant(5, MVT::i32));
7025 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7026 TrmpAddr, 5, false, 1);
7028 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7029 DAG.getConstant(6, MVT::i32));
7030 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7033 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7034 return DAG.getMergeValues(Ops, 2, dl);
7038 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7040 The rounding mode is in bits 11:10 of FPSR, and has the following
7047 FLT_ROUNDS, on the other hand, expects the following:
7054 To perform the conversion, we do:
7055 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7058 MachineFunction &MF = DAG.getMachineFunction();
7059 const TargetMachine &TM = MF.getTarget();
7060 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7061 unsigned StackAlignment = TFI.getStackAlignment();
7062 EVT VT = Op.getValueType();
7063 DebugLoc dl = Op.getDebugLoc();
7065 // Save FP Control Word to stack slot
7066 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7067 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7069 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7070 DAG.getEntryNode(), StackSlot);
7072 // Load FP Control Word from stack slot
7073 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7075 // Transform as necessary
7077 DAG.getNode(ISD::SRL, dl, MVT::i16,
7078 DAG.getNode(ISD::AND, dl, MVT::i16,
7079 CWD, DAG.getConstant(0x800, MVT::i16)),
7080 DAG.getConstant(11, MVT::i8));
7082 DAG.getNode(ISD::SRL, dl, MVT::i16,
7083 DAG.getNode(ISD::AND, dl, MVT::i16,
7084 CWD, DAG.getConstant(0x400, MVT::i16)),
7085 DAG.getConstant(9, MVT::i8));
7088 DAG.getNode(ISD::AND, dl, MVT::i16,
7089 DAG.getNode(ISD::ADD, dl, MVT::i16,
7090 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7091 DAG.getConstant(1, MVT::i16)),
7092 DAG.getConstant(3, MVT::i16));
7095 return DAG.getNode((VT.getSizeInBits() < 16 ?
7096 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7099 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7100 EVT VT = Op.getValueType();
7102 unsigned NumBits = VT.getSizeInBits();
7103 DebugLoc dl = Op.getDebugLoc();
7105 Op = Op.getOperand(0);
7106 if (VT == MVT::i8) {
7107 // Zero extend to i32 since there is not an i8 bsr.
7109 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7112 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7113 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7114 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7116 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7119 DAG.getConstant(NumBits+NumBits-1, OpVT),
7120 DAG.getConstant(X86::COND_E, MVT::i8),
7123 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7125 // Finally xor with NumBits-1.
7126 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7129 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7133 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7134 EVT VT = Op.getValueType();
7136 unsigned NumBits = VT.getSizeInBits();
7137 DebugLoc dl = Op.getDebugLoc();
7139 Op = Op.getOperand(0);
7140 if (VT == MVT::i8) {
7142 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7145 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7146 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7147 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7149 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7152 DAG.getConstant(NumBits, OpVT),
7153 DAG.getConstant(X86::COND_E, MVT::i8),
7156 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7159 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7163 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7164 EVT VT = Op.getValueType();
7165 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7166 DebugLoc dl = Op.getDebugLoc();
7168 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7169 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7170 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7171 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7172 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7174 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7175 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7176 // return AloBlo + AloBhi + AhiBlo;
7178 SDValue A = Op.getOperand(0);
7179 SDValue B = Op.getOperand(1);
7181 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7182 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7183 A, DAG.getConstant(32, MVT::i32));
7184 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7185 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7186 B, DAG.getConstant(32, MVT::i32));
7187 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7188 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7190 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7191 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7193 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7194 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7196 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7197 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7198 AloBhi, DAG.getConstant(32, MVT::i32));
7199 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7200 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7201 AhiBlo, DAG.getConstant(32, MVT::i32));
7202 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7203 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7208 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7209 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7210 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7211 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7212 // has only one use.
7213 SDNode *N = Op.getNode();
7214 SDValue LHS = N->getOperand(0);
7215 SDValue RHS = N->getOperand(1);
7216 unsigned BaseOp = 0;
7218 DebugLoc dl = Op.getDebugLoc();
7220 switch (Op.getOpcode()) {
7221 default: llvm_unreachable("Unknown ovf instruction!");
7223 // A subtract of one will be selected as a INC. Note that INC doesn't
7224 // set CF, so we can't do this for UADDO.
7225 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7226 if (C->getAPIntValue() == 1) {
7227 BaseOp = X86ISD::INC;
7231 BaseOp = X86ISD::ADD;
7235 BaseOp = X86ISD::ADD;
7239 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7240 // set CF, so we can't do this for USUBO.
7241 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7242 if (C->getAPIntValue() == 1) {
7243 BaseOp = X86ISD::DEC;
7247 BaseOp = X86ISD::SUB;
7251 BaseOp = X86ISD::SUB;
7255 BaseOp = X86ISD::SMUL;
7259 BaseOp = X86ISD::UMUL;
7264 // Also sets EFLAGS.
7265 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7266 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7269 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7270 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7272 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7276 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7277 EVT T = Op.getValueType();
7278 DebugLoc dl = Op.getDebugLoc();
7281 switch(T.getSimpleVT().SimpleTy) {
7283 assert(false && "Invalid value type!");
7284 case MVT::i8: Reg = X86::AL; size = 1; break;
7285 case MVT::i16: Reg = X86::AX; size = 2; break;
7286 case MVT::i32: Reg = X86::EAX; size = 4; break;
7288 assert(Subtarget->is64Bit() && "Node not type legal!");
7289 Reg = X86::RAX; size = 8;
7292 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7293 Op.getOperand(2), SDValue());
7294 SDValue Ops[] = { cpIn.getValue(0),
7297 DAG.getTargetConstant(size, MVT::i8),
7299 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7300 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7302 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7306 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7307 SelectionDAG &DAG) {
7308 assert(Subtarget->is64Bit() && "Result not type legalized?");
7309 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7310 SDValue TheChain = Op.getOperand(0);
7311 DebugLoc dl = Op.getDebugLoc();
7312 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7313 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7314 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7316 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7317 DAG.getConstant(32, MVT::i8));
7319 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7322 return DAG.getMergeValues(Ops, 2, dl);
7325 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7326 SDNode *Node = Op.getNode();
7327 DebugLoc dl = Node->getDebugLoc();
7328 EVT T = Node->getValueType(0);
7329 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7330 DAG.getConstant(0, T), Node->getOperand(2));
7331 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7332 cast<AtomicSDNode>(Node)->getMemoryVT(),
7333 Node->getOperand(0),
7334 Node->getOperand(1), negOp,
7335 cast<AtomicSDNode>(Node)->getSrcValue(),
7336 cast<AtomicSDNode>(Node)->getAlignment());
7339 /// LowerOperation - Provide custom lowering hooks for some operations.
7341 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7342 switch (Op.getOpcode()) {
7343 default: llvm_unreachable("Should not custom lower this!");
7344 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7345 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7346 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7347 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7348 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7349 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7350 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7351 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7352 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7353 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7354 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7355 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7356 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7357 case ISD::SHL_PARTS:
7358 case ISD::SRA_PARTS:
7359 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7360 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7361 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7362 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7363 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7364 case ISD::FABS: return LowerFABS(Op, DAG);
7365 case ISD::FNEG: return LowerFNEG(Op, DAG);
7366 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7367 case ISD::SETCC: return LowerSETCC(Op, DAG);
7368 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7369 case ISD::SELECT: return LowerSELECT(Op, DAG);
7370 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7371 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7372 case ISD::VASTART: return LowerVASTART(Op, DAG);
7373 case ISD::VAARG: return LowerVAARG(Op, DAG);
7374 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7375 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7376 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7377 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7378 case ISD::FRAME_TO_ARGS_OFFSET:
7379 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7380 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7381 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7382 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7383 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7384 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7385 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7386 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7392 case ISD::UMULO: return LowerXALUO(Op, DAG);
7393 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7397 void X86TargetLowering::
7398 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7399 SelectionDAG &DAG, unsigned NewOp) {
7400 EVT T = Node->getValueType(0);
7401 DebugLoc dl = Node->getDebugLoc();
7402 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7404 SDValue Chain = Node->getOperand(0);
7405 SDValue In1 = Node->getOperand(1);
7406 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7407 Node->getOperand(2), DAG.getIntPtrConstant(0));
7408 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7409 Node->getOperand(2), DAG.getIntPtrConstant(1));
7410 SDValue Ops[] = { Chain, In1, In2L, In2H };
7411 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7413 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7414 cast<MemSDNode>(Node)->getMemOperand());
7415 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7416 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7417 Results.push_back(Result.getValue(2));
7420 /// ReplaceNodeResults - Replace a node with an illegal result type
7421 /// with a new node built out of custom code.
7422 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7423 SmallVectorImpl<SDValue>&Results,
7424 SelectionDAG &DAG) {
7425 DebugLoc dl = N->getDebugLoc();
7426 switch (N->getOpcode()) {
7428 assert(false && "Do not know how to custom type legalize this operation!");
7430 case ISD::FP_TO_SINT: {
7431 std::pair<SDValue,SDValue> Vals =
7432 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7433 SDValue FIST = Vals.first, StackSlot = Vals.second;
7434 if (FIST.getNode() != 0) {
7435 EVT VT = N->getValueType(0);
7436 // Return a load from the stack slot.
7437 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7441 case ISD::READCYCLECOUNTER: {
7442 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7443 SDValue TheChain = N->getOperand(0);
7444 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7445 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7447 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7449 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7450 SDValue Ops[] = { eax, edx };
7451 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7452 Results.push_back(edx.getValue(1));
7459 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7460 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7463 case ISD::ATOMIC_CMP_SWAP: {
7464 EVT T = N->getValueType(0);
7465 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7466 SDValue cpInL, cpInH;
7467 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7468 DAG.getConstant(0, MVT::i32));
7469 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7470 DAG.getConstant(1, MVT::i32));
7471 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7472 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7474 SDValue swapInL, swapInH;
7475 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7476 DAG.getConstant(0, MVT::i32));
7477 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7478 DAG.getConstant(1, MVT::i32));
7479 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7481 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7482 swapInL.getValue(1));
7483 SDValue Ops[] = { swapInH.getValue(0),
7485 swapInH.getValue(1) };
7486 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7487 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7488 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7489 MVT::i32, Result.getValue(1));
7490 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7491 MVT::i32, cpOutL.getValue(2));
7492 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7493 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7494 Results.push_back(cpOutH.getValue(1));
7497 case ISD::ATOMIC_LOAD_ADD:
7498 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7500 case ISD::ATOMIC_LOAD_AND:
7501 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7503 case ISD::ATOMIC_LOAD_NAND:
7504 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7506 case ISD::ATOMIC_LOAD_OR:
7507 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7509 case ISD::ATOMIC_LOAD_SUB:
7510 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7512 case ISD::ATOMIC_LOAD_XOR:
7513 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7515 case ISD::ATOMIC_SWAP:
7516 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7521 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7523 default: return NULL;
7524 case X86ISD::BSF: return "X86ISD::BSF";
7525 case X86ISD::BSR: return "X86ISD::BSR";
7526 case X86ISD::SHLD: return "X86ISD::SHLD";
7527 case X86ISD::SHRD: return "X86ISD::SHRD";
7528 case X86ISD::FAND: return "X86ISD::FAND";
7529 case X86ISD::FOR: return "X86ISD::FOR";
7530 case X86ISD::FXOR: return "X86ISD::FXOR";
7531 case X86ISD::FSRL: return "X86ISD::FSRL";
7532 case X86ISD::FILD: return "X86ISD::FILD";
7533 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7534 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7535 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7536 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7537 case X86ISD::FLD: return "X86ISD::FLD";
7538 case X86ISD::FST: return "X86ISD::FST";
7539 case X86ISD::CALL: return "X86ISD::CALL";
7540 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7541 case X86ISD::BT: return "X86ISD::BT";
7542 case X86ISD::CMP: return "X86ISD::CMP";
7543 case X86ISD::COMI: return "X86ISD::COMI";
7544 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7545 case X86ISD::SETCC: return "X86ISD::SETCC";
7546 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7547 case X86ISD::CMOV: return "X86ISD::CMOV";
7548 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7549 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7550 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7551 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7552 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7553 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7554 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7555 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7556 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7557 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7558 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7559 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7560 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7561 case X86ISD::FMAX: return "X86ISD::FMAX";
7562 case X86ISD::FMIN: return "X86ISD::FMIN";
7563 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7564 case X86ISD::FRCP: return "X86ISD::FRCP";
7565 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7566 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7567 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7568 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7569 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7570 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7571 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7572 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7573 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7574 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7575 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7576 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7577 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7578 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7579 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7580 case X86ISD::VSHL: return "X86ISD::VSHL";
7581 case X86ISD::VSRL: return "X86ISD::VSRL";
7582 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7583 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7584 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7585 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7586 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7587 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7588 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7589 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7590 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7591 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7592 case X86ISD::ADD: return "X86ISD::ADD";
7593 case X86ISD::SUB: return "X86ISD::SUB";
7594 case X86ISD::SMUL: return "X86ISD::SMUL";
7595 case X86ISD::UMUL: return "X86ISD::UMUL";
7596 case X86ISD::INC: return "X86ISD::INC";
7597 case X86ISD::DEC: return "X86ISD::DEC";
7598 case X86ISD::OR: return "X86ISD::OR";
7599 case X86ISD::XOR: return "X86ISD::XOR";
7600 case X86ISD::AND: return "X86ISD::AND";
7601 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7602 case X86ISD::PTEST: return "X86ISD::PTEST";
7603 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7607 // isLegalAddressingMode - Return true if the addressing mode represented
7608 // by AM is legal for this target, for a load/store of the specified type.
7609 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7610 const Type *Ty) const {
7611 // X86 supports extremely general addressing modes.
7612 CodeModel::Model M = getTargetMachine().getCodeModel();
7614 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7615 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7620 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7622 // If a reference to this global requires an extra load, we can't fold it.
7623 if (isGlobalStubReference(GVFlags))
7626 // If BaseGV requires a register for the PIC base, we cannot also have a
7627 // BaseReg specified.
7628 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7631 // If lower 4G is not available, then we must use rip-relative addressing.
7632 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7642 // These scales always work.
7647 // These scales are formed with basereg+scalereg. Only accept if there is
7652 default: // Other stuff never works.
7660 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7661 if (!Ty1->isInteger() || !Ty2->isInteger())
7663 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7664 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7665 if (NumBits1 <= NumBits2)
7667 return Subtarget->is64Bit() || NumBits1 < 64;
7670 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7671 if (!VT1.isInteger() || !VT2.isInteger())
7673 unsigned NumBits1 = VT1.getSizeInBits();
7674 unsigned NumBits2 = VT2.getSizeInBits();
7675 if (NumBits1 <= NumBits2)
7677 return Subtarget->is64Bit() || NumBits1 < 64;
7680 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7681 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7682 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7685 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7686 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7687 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7690 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7691 // i16 instructions are longer (0x66 prefix) and potentially slower.
7692 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7695 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7696 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7697 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7698 /// are assumed to be legal.
7700 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7702 // Only do shuffles on 128-bit vector types for now.
7703 if (VT.getSizeInBits() == 64)
7706 // FIXME: pshufb, blends, shifts.
7707 return (VT.getVectorNumElements() == 2 ||
7708 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7709 isMOVLMask(M, VT) ||
7710 isSHUFPMask(M, VT) ||
7711 isPSHUFDMask(M, VT) ||
7712 isPSHUFHWMask(M, VT) ||
7713 isPSHUFLWMask(M, VT) ||
7714 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7715 isUNPCKLMask(M, VT) ||
7716 isUNPCKHMask(M, VT) ||
7717 isUNPCKL_v_undef_Mask(M, VT) ||
7718 isUNPCKH_v_undef_Mask(M, VT));
7722 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7724 unsigned NumElts = VT.getVectorNumElements();
7725 // FIXME: This collection of masks seems suspect.
7728 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7729 return (isMOVLMask(Mask, VT) ||
7730 isCommutedMOVLMask(Mask, VT, true) ||
7731 isSHUFPMask(Mask, VT) ||
7732 isCommutedSHUFPMask(Mask, VT));
7737 //===----------------------------------------------------------------------===//
7738 // X86 Scheduler Hooks
7739 //===----------------------------------------------------------------------===//
7741 // private utility function
7743 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7744 MachineBasicBlock *MBB,
7752 TargetRegisterClass *RC,
7753 bool invSrc) const {
7754 // For the atomic bitwise operator, we generate
7757 // ld t1 = [bitinstr.addr]
7758 // op t2 = t1, [bitinstr.val]
7760 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7762 // fallthrough -->nextMBB
7763 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7764 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7765 MachineFunction::iterator MBBIter = MBB;
7768 /// First build the CFG
7769 MachineFunction *F = MBB->getParent();
7770 MachineBasicBlock *thisMBB = MBB;
7771 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7772 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7773 F->insert(MBBIter, newMBB);
7774 F->insert(MBBIter, nextMBB);
7776 // Move all successors to thisMBB to nextMBB
7777 nextMBB->transferSuccessors(thisMBB);
7779 // Update thisMBB to fall through to newMBB
7780 thisMBB->addSuccessor(newMBB);
7782 // newMBB jumps to itself and fall through to nextMBB
7783 newMBB->addSuccessor(nextMBB);
7784 newMBB->addSuccessor(newMBB);
7786 // Insert instructions into newMBB based on incoming instruction
7787 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7788 "unexpected number of operands");
7789 DebugLoc dl = bInstr->getDebugLoc();
7790 MachineOperand& destOper = bInstr->getOperand(0);
7791 MachineOperand* argOpers[2 + X86AddrNumOperands];
7792 int numArgs = bInstr->getNumOperands() - 1;
7793 for (int i=0; i < numArgs; ++i)
7794 argOpers[i] = &bInstr->getOperand(i+1);
7796 // x86 address has 4 operands: base, index, scale, and displacement
7797 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7798 int valArgIndx = lastAddrIndx + 1;
7800 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7801 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7802 for (int i=0; i <= lastAddrIndx; ++i)
7803 (*MIB).addOperand(*argOpers[i]);
7805 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7807 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7812 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7813 assert((argOpers[valArgIndx]->isReg() ||
7814 argOpers[valArgIndx]->isImm()) &&
7816 if (argOpers[valArgIndx]->isReg())
7817 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7819 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7821 (*MIB).addOperand(*argOpers[valArgIndx]);
7823 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7826 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7827 for (int i=0; i <= lastAddrIndx; ++i)
7828 (*MIB).addOperand(*argOpers[i]);
7830 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7831 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7832 bInstr->memoperands_end());
7834 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7838 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7840 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7844 // private utility function: 64 bit atomics on 32 bit host.
7846 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7847 MachineBasicBlock *MBB,
7852 bool invSrc) const {
7853 // For the atomic bitwise operator, we generate
7854 // thisMBB (instructions are in pairs, except cmpxchg8b)
7855 // ld t1,t2 = [bitinstr.addr]
7857 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7858 // op t5, t6 <- out1, out2, [bitinstr.val]
7859 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7860 // mov ECX, EBX <- t5, t6
7861 // mov EAX, EDX <- t1, t2
7862 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7863 // mov t3, t4 <- EAX, EDX
7865 // result in out1, out2
7866 // fallthrough -->nextMBB
7868 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7869 const unsigned LoadOpc = X86::MOV32rm;
7870 const unsigned copyOpc = X86::MOV32rr;
7871 const unsigned NotOpc = X86::NOT32r;
7872 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7873 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7874 MachineFunction::iterator MBBIter = MBB;
7877 /// First build the CFG
7878 MachineFunction *F = MBB->getParent();
7879 MachineBasicBlock *thisMBB = MBB;
7880 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7881 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7882 F->insert(MBBIter, newMBB);
7883 F->insert(MBBIter, nextMBB);
7885 // Move all successors to thisMBB to nextMBB
7886 nextMBB->transferSuccessors(thisMBB);
7888 // Update thisMBB to fall through to newMBB
7889 thisMBB->addSuccessor(newMBB);
7891 // newMBB jumps to itself and fall through to nextMBB
7892 newMBB->addSuccessor(nextMBB);
7893 newMBB->addSuccessor(newMBB);
7895 DebugLoc dl = bInstr->getDebugLoc();
7896 // Insert instructions into newMBB based on incoming instruction
7897 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7898 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7899 "unexpected number of operands");
7900 MachineOperand& dest1Oper = bInstr->getOperand(0);
7901 MachineOperand& dest2Oper = bInstr->getOperand(1);
7902 MachineOperand* argOpers[2 + X86AddrNumOperands];
7903 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7904 argOpers[i] = &bInstr->getOperand(i+2);
7906 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7907 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7909 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7910 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7911 for (int i=0; i <= lastAddrIndx; ++i)
7912 (*MIB).addOperand(*argOpers[i]);
7913 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7914 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7915 // add 4 to displacement.
7916 for (int i=0; i <= lastAddrIndx-2; ++i)
7917 (*MIB).addOperand(*argOpers[i]);
7918 MachineOperand newOp3 = *(argOpers[3]);
7920 newOp3.setImm(newOp3.getImm()+4);
7922 newOp3.setOffset(newOp3.getOffset()+4);
7923 (*MIB).addOperand(newOp3);
7924 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7926 // t3/4 are defined later, at the bottom of the loop
7927 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7928 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7929 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7930 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7931 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7932 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7934 // The subsequent operations should be using the destination registers of
7935 //the PHI instructions.
7937 t1 = F->getRegInfo().createVirtualRegister(RC);
7938 t2 = F->getRegInfo().createVirtualRegister(RC);
7939 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7940 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7942 t1 = dest1Oper.getReg();
7943 t2 = dest2Oper.getReg();
7946 int valArgIndx = lastAddrIndx + 1;
7947 assert((argOpers[valArgIndx]->isReg() ||
7948 argOpers[valArgIndx]->isImm()) &&
7950 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7951 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7952 if (argOpers[valArgIndx]->isReg())
7953 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7955 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7956 if (regOpcL != X86::MOV32rr)
7958 (*MIB).addOperand(*argOpers[valArgIndx]);
7959 assert(argOpers[valArgIndx + 1]->isReg() ==
7960 argOpers[valArgIndx]->isReg());
7961 assert(argOpers[valArgIndx + 1]->isImm() ==
7962 argOpers[valArgIndx]->isImm());
7963 if (argOpers[valArgIndx + 1]->isReg())
7964 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7966 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7967 if (regOpcH != X86::MOV32rr)
7969 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7971 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7973 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7976 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7978 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7981 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7982 for (int i=0; i <= lastAddrIndx; ++i)
7983 (*MIB).addOperand(*argOpers[i]);
7985 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7986 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7987 bInstr->memoperands_end());
7989 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7990 MIB.addReg(X86::EAX);
7991 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7992 MIB.addReg(X86::EDX);
7995 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7997 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8001 // private utility function
8003 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8004 MachineBasicBlock *MBB,
8005 unsigned cmovOpc) const {
8006 // For the atomic min/max operator, we generate
8009 // ld t1 = [min/max.addr]
8010 // mov t2 = [min/max.val]
8012 // cmov[cond] t2 = t1
8014 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8016 // fallthrough -->nextMBB
8018 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8019 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8020 MachineFunction::iterator MBBIter = MBB;
8023 /// First build the CFG
8024 MachineFunction *F = MBB->getParent();
8025 MachineBasicBlock *thisMBB = MBB;
8026 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8027 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8028 F->insert(MBBIter, newMBB);
8029 F->insert(MBBIter, nextMBB);
8031 // Move all successors of thisMBB to nextMBB
8032 nextMBB->transferSuccessors(thisMBB);
8034 // Update thisMBB to fall through to newMBB
8035 thisMBB->addSuccessor(newMBB);
8037 // newMBB jumps to newMBB and fall through to nextMBB
8038 newMBB->addSuccessor(nextMBB);
8039 newMBB->addSuccessor(newMBB);
8041 DebugLoc dl = mInstr->getDebugLoc();
8042 // Insert instructions into newMBB based on incoming instruction
8043 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8044 "unexpected number of operands");
8045 MachineOperand& destOper = mInstr->getOperand(0);
8046 MachineOperand* argOpers[2 + X86AddrNumOperands];
8047 int numArgs = mInstr->getNumOperands() - 1;
8048 for (int i=0; i < numArgs; ++i)
8049 argOpers[i] = &mInstr->getOperand(i+1);
8051 // x86 address has 4 operands: base, index, scale, and displacement
8052 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8053 int valArgIndx = lastAddrIndx + 1;
8055 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8056 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8057 for (int i=0; i <= lastAddrIndx; ++i)
8058 (*MIB).addOperand(*argOpers[i]);
8060 // We only support register and immediate values
8061 assert((argOpers[valArgIndx]->isReg() ||
8062 argOpers[valArgIndx]->isImm()) &&
8065 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8066 if (argOpers[valArgIndx]->isReg())
8067 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8069 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8070 (*MIB).addOperand(*argOpers[valArgIndx]);
8072 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8075 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8080 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8081 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8085 // Cmp and exchange if none has modified the memory location
8086 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8087 for (int i=0; i <= lastAddrIndx; ++i)
8088 (*MIB).addOperand(*argOpers[i]);
8090 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8091 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8092 mInstr->memoperands_end());
8094 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8095 MIB.addReg(X86::EAX);
8098 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8100 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8104 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8105 // all of this code can be replaced with that in the .td file.
8107 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8108 unsigned numArgs, bool memArg) const {
8110 MachineFunction *F = BB->getParent();
8111 DebugLoc dl = MI->getDebugLoc();
8112 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8116 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8118 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8120 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8122 for (unsigned i = 0; i < numArgs; ++i) {
8123 MachineOperand &Op = MI->getOperand(i+1);
8125 if (!(Op.isReg() && Op.isImplicit()))
8129 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8132 F->DeleteMachineInstr(MI);
8138 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8140 MachineBasicBlock *MBB) const {
8141 // Emit code to save XMM registers to the stack. The ABI says that the
8142 // number of registers to save is given in %al, so it's theoretically
8143 // possible to do an indirect jump trick to avoid saving all of them,
8144 // however this code takes a simpler approach and just executes all
8145 // of the stores if %al is non-zero. It's less code, and it's probably
8146 // easier on the hardware branch predictor, and stores aren't all that
8147 // expensive anyway.
8149 // Create the new basic blocks. One block contains all the XMM stores,
8150 // and one block is the final destination regardless of whether any
8151 // stores were performed.
8152 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8153 MachineFunction *F = MBB->getParent();
8154 MachineFunction::iterator MBBIter = MBB;
8156 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8157 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8158 F->insert(MBBIter, XMMSaveMBB);
8159 F->insert(MBBIter, EndMBB);
8162 // Move any original successors of MBB to the end block.
8163 EndMBB->transferSuccessors(MBB);
8164 // The original block will now fall through to the XMM save block.
8165 MBB->addSuccessor(XMMSaveMBB);
8166 // The XMMSaveMBB will fall through to the end block.
8167 XMMSaveMBB->addSuccessor(EndMBB);
8169 // Now add the instructions.
8170 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8171 DebugLoc DL = MI->getDebugLoc();
8173 unsigned CountReg = MI->getOperand(0).getReg();
8174 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8175 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8177 if (!Subtarget->isTargetWin64()) {
8178 // If %al is 0, branch around the XMM save block.
8179 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8180 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8181 MBB->addSuccessor(EndMBB);
8184 // In the XMM save block, save all the XMM argument registers.
8185 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8186 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8187 MachineMemOperand *MMO =
8188 F->getMachineMemOperand(
8189 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8190 MachineMemOperand::MOStore, Offset,
8191 /*Size=*/16, /*Align=*/16);
8192 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8193 .addFrameIndex(RegSaveFrameIndex)
8194 .addImm(/*Scale=*/1)
8195 .addReg(/*IndexReg=*/0)
8196 .addImm(/*Disp=*/Offset)
8197 .addReg(/*Segment=*/0)
8198 .addReg(MI->getOperand(i).getReg())
8199 .addMemOperand(MMO);
8202 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8208 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8209 MachineBasicBlock *BB,
8210 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8211 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8212 DebugLoc DL = MI->getDebugLoc();
8214 // To "insert" a SELECT_CC instruction, we actually have to insert the
8215 // diamond control-flow pattern. The incoming instruction knows the
8216 // destination vreg to set, the condition code register to branch on, the
8217 // true/false values to select between, and a branch opcode to use.
8218 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8219 MachineFunction::iterator It = BB;
8225 // cmpTY ccX, r1, r2
8227 // fallthrough --> copy0MBB
8228 MachineBasicBlock *thisMBB = BB;
8229 MachineFunction *F = BB->getParent();
8230 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8231 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8233 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8234 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8235 F->insert(It, copy0MBB);
8236 F->insert(It, sinkMBB);
8237 // Update machine-CFG edges by first adding all successors of the current
8238 // block to the new block which will contain the Phi node for the select.
8239 // Also inform sdisel of the edge changes.
8240 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8241 E = BB->succ_end(); I != E; ++I) {
8242 EM->insert(std::make_pair(*I, sinkMBB));
8243 sinkMBB->addSuccessor(*I);
8245 // Next, remove all successors of the current block, and add the true
8246 // and fallthrough blocks as its successors.
8247 while (!BB->succ_empty())
8248 BB->removeSuccessor(BB->succ_begin());
8249 // Add the true and fallthrough blocks as its successors.
8250 BB->addSuccessor(copy0MBB);
8251 BB->addSuccessor(sinkMBB);
8254 // %FalseValue = ...
8255 // # fallthrough to sinkMBB
8258 // Update machine-CFG edges
8259 BB->addSuccessor(sinkMBB);
8262 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8265 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8266 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8267 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8269 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8275 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8276 MachineBasicBlock *BB,
8277 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8278 switch (MI->getOpcode()) {
8279 default: assert(false && "Unexpected instr type to insert");
8281 case X86::CMOV_V1I64:
8282 case X86::CMOV_FR32:
8283 case X86::CMOV_FR64:
8284 case X86::CMOV_V4F32:
8285 case X86::CMOV_V2F64:
8286 case X86::CMOV_V2I64:
8287 return EmitLoweredSelect(MI, BB, EM);
8289 case X86::FP32_TO_INT16_IN_MEM:
8290 case X86::FP32_TO_INT32_IN_MEM:
8291 case X86::FP32_TO_INT64_IN_MEM:
8292 case X86::FP64_TO_INT16_IN_MEM:
8293 case X86::FP64_TO_INT32_IN_MEM:
8294 case X86::FP64_TO_INT64_IN_MEM:
8295 case X86::FP80_TO_INT16_IN_MEM:
8296 case X86::FP80_TO_INT32_IN_MEM:
8297 case X86::FP80_TO_INT64_IN_MEM: {
8298 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8299 DebugLoc DL = MI->getDebugLoc();
8301 // Change the floating point control register to use "round towards zero"
8302 // mode when truncating to an integer value.
8303 MachineFunction *F = BB->getParent();
8304 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8305 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8307 // Load the old value of the high byte of the control word...
8309 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8310 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8313 // Set the high part to be round to zero...
8314 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8317 // Reload the modified control word now...
8318 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8320 // Restore the memory image of control word to original value
8321 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8324 // Get the X86 opcode to use.
8326 switch (MI->getOpcode()) {
8327 default: llvm_unreachable("illegal opcode!");
8328 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8329 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8330 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8331 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8332 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8333 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8334 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8335 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8336 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8340 MachineOperand &Op = MI->getOperand(0);
8342 AM.BaseType = X86AddressMode::RegBase;
8343 AM.Base.Reg = Op.getReg();
8345 AM.BaseType = X86AddressMode::FrameIndexBase;
8346 AM.Base.FrameIndex = Op.getIndex();
8348 Op = MI->getOperand(1);
8350 AM.Scale = Op.getImm();
8351 Op = MI->getOperand(2);
8353 AM.IndexReg = Op.getImm();
8354 Op = MI->getOperand(3);
8355 if (Op.isGlobal()) {
8356 AM.GV = Op.getGlobal();
8358 AM.Disp = Op.getImm();
8360 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8361 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8363 // Reload the original control word now.
8364 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8366 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8369 // String/text processing lowering.
8370 case X86::PCMPISTRM128REG:
8371 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8372 case X86::PCMPISTRM128MEM:
8373 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8374 case X86::PCMPESTRM128REG:
8375 return EmitPCMP(MI, BB, 5, false /* in mem */);
8376 case X86::PCMPESTRM128MEM:
8377 return EmitPCMP(MI, BB, 5, true /* in mem */);
8380 case X86::ATOMAND32:
8381 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8382 X86::AND32ri, X86::MOV32rm,
8383 X86::LCMPXCHG32, X86::MOV32rr,
8384 X86::NOT32r, X86::EAX,
8385 X86::GR32RegisterClass);
8387 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8388 X86::OR32ri, X86::MOV32rm,
8389 X86::LCMPXCHG32, X86::MOV32rr,
8390 X86::NOT32r, X86::EAX,
8391 X86::GR32RegisterClass);
8392 case X86::ATOMXOR32:
8393 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8394 X86::XOR32ri, X86::MOV32rm,
8395 X86::LCMPXCHG32, X86::MOV32rr,
8396 X86::NOT32r, X86::EAX,
8397 X86::GR32RegisterClass);
8398 case X86::ATOMNAND32:
8399 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8400 X86::AND32ri, X86::MOV32rm,
8401 X86::LCMPXCHG32, X86::MOV32rr,
8402 X86::NOT32r, X86::EAX,
8403 X86::GR32RegisterClass, true);
8404 case X86::ATOMMIN32:
8405 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8406 case X86::ATOMMAX32:
8407 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8408 case X86::ATOMUMIN32:
8409 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8410 case X86::ATOMUMAX32:
8411 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8413 case X86::ATOMAND16:
8414 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8415 X86::AND16ri, X86::MOV16rm,
8416 X86::LCMPXCHG16, X86::MOV16rr,
8417 X86::NOT16r, X86::AX,
8418 X86::GR16RegisterClass);
8420 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8421 X86::OR16ri, X86::MOV16rm,
8422 X86::LCMPXCHG16, X86::MOV16rr,
8423 X86::NOT16r, X86::AX,
8424 X86::GR16RegisterClass);
8425 case X86::ATOMXOR16:
8426 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8427 X86::XOR16ri, X86::MOV16rm,
8428 X86::LCMPXCHG16, X86::MOV16rr,
8429 X86::NOT16r, X86::AX,
8430 X86::GR16RegisterClass);
8431 case X86::ATOMNAND16:
8432 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8433 X86::AND16ri, X86::MOV16rm,
8434 X86::LCMPXCHG16, X86::MOV16rr,
8435 X86::NOT16r, X86::AX,
8436 X86::GR16RegisterClass, true);
8437 case X86::ATOMMIN16:
8438 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8439 case X86::ATOMMAX16:
8440 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8441 case X86::ATOMUMIN16:
8442 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8443 case X86::ATOMUMAX16:
8444 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8447 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8448 X86::AND8ri, X86::MOV8rm,
8449 X86::LCMPXCHG8, X86::MOV8rr,
8450 X86::NOT8r, X86::AL,
8451 X86::GR8RegisterClass);
8453 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8454 X86::OR8ri, X86::MOV8rm,
8455 X86::LCMPXCHG8, X86::MOV8rr,
8456 X86::NOT8r, X86::AL,
8457 X86::GR8RegisterClass);
8459 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8460 X86::XOR8ri, X86::MOV8rm,
8461 X86::LCMPXCHG8, X86::MOV8rr,
8462 X86::NOT8r, X86::AL,
8463 X86::GR8RegisterClass);
8464 case X86::ATOMNAND8:
8465 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8466 X86::AND8ri, X86::MOV8rm,
8467 X86::LCMPXCHG8, X86::MOV8rr,
8468 X86::NOT8r, X86::AL,
8469 X86::GR8RegisterClass, true);
8470 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8471 // This group is for 64-bit host.
8472 case X86::ATOMAND64:
8473 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8474 X86::AND64ri32, X86::MOV64rm,
8475 X86::LCMPXCHG64, X86::MOV64rr,
8476 X86::NOT64r, X86::RAX,
8477 X86::GR64RegisterClass);
8479 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8480 X86::OR64ri32, X86::MOV64rm,
8481 X86::LCMPXCHG64, X86::MOV64rr,
8482 X86::NOT64r, X86::RAX,
8483 X86::GR64RegisterClass);
8484 case X86::ATOMXOR64:
8485 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8486 X86::XOR64ri32, X86::MOV64rm,
8487 X86::LCMPXCHG64, X86::MOV64rr,
8488 X86::NOT64r, X86::RAX,
8489 X86::GR64RegisterClass);
8490 case X86::ATOMNAND64:
8491 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8492 X86::AND64ri32, X86::MOV64rm,
8493 X86::LCMPXCHG64, X86::MOV64rr,
8494 X86::NOT64r, X86::RAX,
8495 X86::GR64RegisterClass, true);
8496 case X86::ATOMMIN64:
8497 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8498 case X86::ATOMMAX64:
8499 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8500 case X86::ATOMUMIN64:
8501 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8502 case X86::ATOMUMAX64:
8503 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8505 // This group does 64-bit operations on a 32-bit host.
8506 case X86::ATOMAND6432:
8507 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8508 X86::AND32rr, X86::AND32rr,
8509 X86::AND32ri, X86::AND32ri,
8511 case X86::ATOMOR6432:
8512 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8513 X86::OR32rr, X86::OR32rr,
8514 X86::OR32ri, X86::OR32ri,
8516 case X86::ATOMXOR6432:
8517 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8518 X86::XOR32rr, X86::XOR32rr,
8519 X86::XOR32ri, X86::XOR32ri,
8521 case X86::ATOMNAND6432:
8522 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8523 X86::AND32rr, X86::AND32rr,
8524 X86::AND32ri, X86::AND32ri,
8526 case X86::ATOMADD6432:
8527 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8528 X86::ADD32rr, X86::ADC32rr,
8529 X86::ADD32ri, X86::ADC32ri,
8531 case X86::ATOMSUB6432:
8532 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8533 X86::SUB32rr, X86::SBB32rr,
8534 X86::SUB32ri, X86::SBB32ri,
8536 case X86::ATOMSWAP6432:
8537 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8538 X86::MOV32rr, X86::MOV32rr,
8539 X86::MOV32ri, X86::MOV32ri,
8541 case X86::VASTART_SAVE_XMM_REGS:
8542 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8546 //===----------------------------------------------------------------------===//
8547 // X86 Optimization Hooks
8548 //===----------------------------------------------------------------------===//
8550 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8554 const SelectionDAG &DAG,
8555 unsigned Depth) const {
8556 unsigned Opc = Op.getOpcode();
8557 assert((Opc >= ISD::BUILTIN_OP_END ||
8558 Opc == ISD::INTRINSIC_WO_CHAIN ||
8559 Opc == ISD::INTRINSIC_W_CHAIN ||
8560 Opc == ISD::INTRINSIC_VOID) &&
8561 "Should use MaskedValueIsZero if you don't know whether Op"
8562 " is a target node!");
8564 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8576 // These nodes' second result is a boolean.
8577 if (Op.getResNo() == 0)
8581 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8582 Mask.getBitWidth() - 1);
8587 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8588 /// node is a GlobalAddress + offset.
8589 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8590 GlobalValue* &GA, int64_t &Offset) const{
8591 if (N->getOpcode() == X86ISD::Wrapper) {
8592 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8593 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8594 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8598 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8601 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8602 EVT EltVT, LoadSDNode *&LDBase,
8603 unsigned &LastLoadedElt,
8604 SelectionDAG &DAG, MachineFrameInfo *MFI,
8605 const TargetLowering &TLI) {
8607 LastLoadedElt = -1U;
8608 for (unsigned i = 0; i < NumElems; ++i) {
8609 if (N->getMaskElt(i) < 0) {
8615 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8616 if (!Elt.getNode() ||
8617 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8620 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8622 LDBase = cast<LoadSDNode>(Elt.getNode());
8626 if (Elt.getOpcode() == ISD::UNDEF)
8629 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8630 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8637 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8638 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8639 /// if the load addresses are consecutive, non-overlapping, and in the right
8640 /// order. In the case of v2i64, it will see if it can rewrite the
8641 /// shuffle to be an appropriate build vector so it can take advantage of
8642 // performBuildVectorCombine.
8643 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8644 const TargetLowering &TLI) {
8645 DebugLoc dl = N->getDebugLoc();
8646 EVT VT = N->getValueType(0);
8647 EVT EltVT = VT.getVectorElementType();
8648 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8649 unsigned NumElems = VT.getVectorNumElements();
8651 if (VT.getSizeInBits() != 128)
8654 // Try to combine a vector_shuffle into a 128-bit load.
8655 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8656 LoadSDNode *LD = NULL;
8657 unsigned LastLoadedElt;
8658 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8662 if (LastLoadedElt == NumElems - 1) {
8663 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8664 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8665 LD->getSrcValue(), LD->getSrcValueOffset(),
8667 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8668 LD->getSrcValue(), LD->getSrcValueOffset(),
8669 LD->isVolatile(), LD->getAlignment());
8670 } else if (NumElems == 4 && LastLoadedElt == 1) {
8671 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8672 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8673 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8674 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8679 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8680 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8681 const X86Subtarget *Subtarget) {
8682 DebugLoc DL = N->getDebugLoc();
8683 SDValue Cond = N->getOperand(0);
8684 // Get the LHS/RHS of the select.
8685 SDValue LHS = N->getOperand(1);
8686 SDValue RHS = N->getOperand(2);
8688 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8689 // instructions have the peculiarity that if either operand is a NaN,
8690 // they chose what we call the RHS operand (and as such are not symmetric).
8691 // It happens that this matches the semantics of the common C idiom
8692 // x<y?x:y and related forms, so we can recognize these cases.
8693 if (Subtarget->hasSSE2() &&
8694 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8695 Cond.getOpcode() == ISD::SETCC) {
8696 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8698 unsigned Opcode = 0;
8699 // Check for x CC y ? x : y.
8700 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8704 // This can be a min if we can prove that at least one of the operands
8706 if (!FiniteOnlyFPMath()) {
8707 if (DAG.isKnownNeverNaN(RHS)) {
8708 // Put the potential NaN in the RHS so that SSE will preserve it.
8709 std::swap(LHS, RHS);
8710 } else if (!DAG.isKnownNeverNaN(LHS))
8713 Opcode = X86ISD::FMIN;
8716 // This can be a min if we can prove that at least one of the operands
8718 if (!FiniteOnlyFPMath()) {
8719 if (DAG.isKnownNeverNaN(LHS)) {
8720 // Put the potential NaN in the RHS so that SSE will preserve it.
8721 std::swap(LHS, RHS);
8722 } else if (!DAG.isKnownNeverNaN(RHS))
8725 Opcode = X86ISD::FMIN;
8728 // This can be a min, but if either operand is a NaN we need it to
8729 // preserve the original LHS.
8730 std::swap(LHS, RHS);
8734 Opcode = X86ISD::FMIN;
8738 // This can be a max if we can prove that at least one of the operands
8740 if (!FiniteOnlyFPMath()) {
8741 if (DAG.isKnownNeverNaN(LHS)) {
8742 // Put the potential NaN in the RHS so that SSE will preserve it.
8743 std::swap(LHS, RHS);
8744 } else if (!DAG.isKnownNeverNaN(RHS))
8747 Opcode = X86ISD::FMAX;
8750 // This can be a max if we can prove that at least one of the operands
8752 if (!FiniteOnlyFPMath()) {
8753 if (DAG.isKnownNeverNaN(RHS)) {
8754 // Put the potential NaN in the RHS so that SSE will preserve it.
8755 std::swap(LHS, RHS);
8756 } else if (!DAG.isKnownNeverNaN(LHS))
8759 Opcode = X86ISD::FMAX;
8762 // This can be a max, but if either operand is a NaN we need it to
8763 // preserve the original LHS.
8764 std::swap(LHS, RHS);
8768 Opcode = X86ISD::FMAX;
8771 // Check for x CC y ? y : x -- a min/max with reversed arms.
8772 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8776 // This can be a min if we can prove that at least one of the operands
8778 if (!FiniteOnlyFPMath()) {
8779 if (DAG.isKnownNeverNaN(RHS)) {
8780 // Put the potential NaN in the RHS so that SSE will preserve it.
8781 std::swap(LHS, RHS);
8782 } else if (!DAG.isKnownNeverNaN(LHS))
8785 Opcode = X86ISD::FMIN;
8788 // This can be a min if we can prove that at least one of the operands
8790 if (!FiniteOnlyFPMath()) {
8791 if (DAG.isKnownNeverNaN(LHS)) {
8792 // Put the potential NaN in the RHS so that SSE will preserve it.
8793 std::swap(LHS, RHS);
8794 } else if (!DAG.isKnownNeverNaN(RHS))
8797 Opcode = X86ISD::FMIN;
8800 // This can be a min, but if either operand is a NaN we need it to
8801 // preserve the original LHS.
8802 std::swap(LHS, RHS);
8806 Opcode = X86ISD::FMIN;
8810 // This can be a max if we can prove that at least one of the operands
8812 if (!FiniteOnlyFPMath()) {
8813 if (DAG.isKnownNeverNaN(LHS)) {
8814 // Put the potential NaN in the RHS so that SSE will preserve it.
8815 std::swap(LHS, RHS);
8816 } else if (!DAG.isKnownNeverNaN(RHS))
8819 Opcode = X86ISD::FMAX;
8822 // This can be a max if we can prove that at least one of the operands
8824 if (!FiniteOnlyFPMath()) {
8825 if (DAG.isKnownNeverNaN(RHS)) {
8826 // Put the potential NaN in the RHS so that SSE will preserve it.
8827 std::swap(LHS, RHS);
8828 } else if (!DAG.isKnownNeverNaN(LHS))
8831 Opcode = X86ISD::FMAX;
8834 // This can be a max, but if either operand is a NaN we need it to
8835 // preserve the original LHS.
8836 std::swap(LHS, RHS);
8840 Opcode = X86ISD::FMAX;
8846 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8849 // If this is a select between two integer constants, try to do some
8851 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8852 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8853 // Don't do this for crazy integer types.
8854 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8855 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8856 // so that TrueC (the true value) is larger than FalseC.
8857 bool NeedsCondInvert = false;
8859 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8860 // Efficiently invertible.
8861 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8862 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8863 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8864 NeedsCondInvert = true;
8865 std::swap(TrueC, FalseC);
8868 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8869 if (FalseC->getAPIntValue() == 0 &&
8870 TrueC->getAPIntValue().isPowerOf2()) {
8871 if (NeedsCondInvert) // Invert the condition if needed.
8872 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8873 DAG.getConstant(1, Cond.getValueType()));
8875 // Zero extend the condition if needed.
8876 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8878 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8879 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8880 DAG.getConstant(ShAmt, MVT::i8));
8883 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8884 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8885 if (NeedsCondInvert) // Invert the condition if needed.
8886 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8887 DAG.getConstant(1, Cond.getValueType()));
8889 // Zero extend the condition if needed.
8890 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8891 FalseC->getValueType(0), Cond);
8892 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8893 SDValue(FalseC, 0));
8896 // Optimize cases that will turn into an LEA instruction. This requires
8897 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8898 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8899 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8900 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8902 bool isFastMultiplier = false;
8904 switch ((unsigned char)Diff) {
8906 case 1: // result = add base, cond
8907 case 2: // result = lea base( , cond*2)
8908 case 3: // result = lea base(cond, cond*2)
8909 case 4: // result = lea base( , cond*4)
8910 case 5: // result = lea base(cond, cond*4)
8911 case 8: // result = lea base( , cond*8)
8912 case 9: // result = lea base(cond, cond*8)
8913 isFastMultiplier = true;
8918 if (isFastMultiplier) {
8919 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8920 if (NeedsCondInvert) // Invert the condition if needed.
8921 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8922 DAG.getConstant(1, Cond.getValueType()));
8924 // Zero extend the condition if needed.
8925 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8927 // Scale the condition by the difference.
8929 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8930 DAG.getConstant(Diff, Cond.getValueType()));
8932 // Add the base if non-zero.
8933 if (FalseC->getAPIntValue() != 0)
8934 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8935 SDValue(FalseC, 0));
8945 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8946 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8947 TargetLowering::DAGCombinerInfo &DCI) {
8948 DebugLoc DL = N->getDebugLoc();
8950 // If the flag operand isn't dead, don't touch this CMOV.
8951 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8954 // If this is a select between two integer constants, try to do some
8955 // optimizations. Note that the operands are ordered the opposite of SELECT
8957 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8958 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8959 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8960 // larger than FalseC (the false value).
8961 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8963 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8964 CC = X86::GetOppositeBranchCondition(CC);
8965 std::swap(TrueC, FalseC);
8968 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8969 // This is efficient for any integer data type (including i8/i16) and
8971 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8972 SDValue Cond = N->getOperand(3);
8973 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8974 DAG.getConstant(CC, MVT::i8), Cond);
8976 // Zero extend the condition if needed.
8977 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8979 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8980 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8981 DAG.getConstant(ShAmt, MVT::i8));
8982 if (N->getNumValues() == 2) // Dead flag value?
8983 return DCI.CombineTo(N, Cond, SDValue());
8987 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8988 // for any integer data type, including i8/i16.
8989 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8990 SDValue Cond = N->getOperand(3);
8991 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8992 DAG.getConstant(CC, MVT::i8), Cond);
8994 // Zero extend the condition if needed.
8995 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8996 FalseC->getValueType(0), Cond);
8997 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8998 SDValue(FalseC, 0));
9000 if (N->getNumValues() == 2) // Dead flag value?
9001 return DCI.CombineTo(N, Cond, SDValue());
9005 // Optimize cases that will turn into an LEA instruction. This requires
9006 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9007 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9008 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9009 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9011 bool isFastMultiplier = false;
9013 switch ((unsigned char)Diff) {
9015 case 1: // result = add base, cond
9016 case 2: // result = lea base( , cond*2)
9017 case 3: // result = lea base(cond, cond*2)
9018 case 4: // result = lea base( , cond*4)
9019 case 5: // result = lea base(cond, cond*4)
9020 case 8: // result = lea base( , cond*8)
9021 case 9: // result = lea base(cond, cond*8)
9022 isFastMultiplier = true;
9027 if (isFastMultiplier) {
9028 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9029 SDValue Cond = N->getOperand(3);
9030 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9031 DAG.getConstant(CC, MVT::i8), Cond);
9032 // Zero extend the condition if needed.
9033 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9035 // Scale the condition by the difference.
9037 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9038 DAG.getConstant(Diff, Cond.getValueType()));
9040 // Add the base if non-zero.
9041 if (FalseC->getAPIntValue() != 0)
9042 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9043 SDValue(FalseC, 0));
9044 if (N->getNumValues() == 2) // Dead flag value?
9045 return DCI.CombineTo(N, Cond, SDValue());
9055 /// PerformMulCombine - Optimize a single multiply with constant into two
9056 /// in order to implement it with two cheaper instructions, e.g.
9057 /// LEA + SHL, LEA + LEA.
9058 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9059 TargetLowering::DAGCombinerInfo &DCI) {
9060 if (DAG.getMachineFunction().
9061 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9064 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9067 EVT VT = N->getValueType(0);
9071 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9074 uint64_t MulAmt = C->getZExtValue();
9075 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9078 uint64_t MulAmt1 = 0;
9079 uint64_t MulAmt2 = 0;
9080 if ((MulAmt % 9) == 0) {
9082 MulAmt2 = MulAmt / 9;
9083 } else if ((MulAmt % 5) == 0) {
9085 MulAmt2 = MulAmt / 5;
9086 } else if ((MulAmt % 3) == 0) {
9088 MulAmt2 = MulAmt / 3;
9091 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9092 DebugLoc DL = N->getDebugLoc();
9094 if (isPowerOf2_64(MulAmt2) &&
9095 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9096 // If second multiplifer is pow2, issue it first. We want the multiply by
9097 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9099 std::swap(MulAmt1, MulAmt2);
9102 if (isPowerOf2_64(MulAmt1))
9103 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9104 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9106 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9107 DAG.getConstant(MulAmt1, VT));
9109 if (isPowerOf2_64(MulAmt2))
9110 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9111 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9113 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9114 DAG.getConstant(MulAmt2, VT));
9116 // Do not add new nodes to DAG combiner worklist.
9117 DCI.CombineTo(N, NewMul, false);
9122 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9123 SDValue N0 = N->getOperand(0);
9124 SDValue N1 = N->getOperand(1);
9125 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9126 EVT VT = N0.getValueType();
9128 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9129 // since the result of setcc_c is all zero's or all ones.
9130 if (N1C && N0.getOpcode() == ISD::AND &&
9131 N0.getOperand(1).getOpcode() == ISD::Constant) {
9132 SDValue N00 = N0.getOperand(0);
9133 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9134 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9135 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9136 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9137 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9138 APInt ShAmt = N1C->getAPIntValue();
9139 Mask = Mask.shl(ShAmt);
9141 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9142 N00, DAG.getConstant(Mask, VT));
9149 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9151 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9152 const X86Subtarget *Subtarget) {
9153 EVT VT = N->getValueType(0);
9154 if (!VT.isVector() && VT.isInteger() &&
9155 N->getOpcode() == ISD::SHL)
9156 return PerformSHLCombine(N, DAG);
9158 // On X86 with SSE2 support, we can transform this to a vector shift if
9159 // all elements are shifted by the same amount. We can't do this in legalize
9160 // because the a constant vector is typically transformed to a constant pool
9161 // so we have no knowledge of the shift amount.
9162 if (!Subtarget->hasSSE2())
9165 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9168 SDValue ShAmtOp = N->getOperand(1);
9169 EVT EltVT = VT.getVectorElementType();
9170 DebugLoc DL = N->getDebugLoc();
9171 SDValue BaseShAmt = SDValue();
9172 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9173 unsigned NumElts = VT.getVectorNumElements();
9175 for (; i != NumElts; ++i) {
9176 SDValue Arg = ShAmtOp.getOperand(i);
9177 if (Arg.getOpcode() == ISD::UNDEF) continue;
9181 for (; i != NumElts; ++i) {
9182 SDValue Arg = ShAmtOp.getOperand(i);
9183 if (Arg.getOpcode() == ISD::UNDEF) continue;
9184 if (Arg != BaseShAmt) {
9188 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9189 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9190 SDValue InVec = ShAmtOp.getOperand(0);
9191 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9192 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9194 for (; i != NumElts; ++i) {
9195 SDValue Arg = InVec.getOperand(i);
9196 if (Arg.getOpcode() == ISD::UNDEF) continue;
9200 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9201 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9202 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9203 if (C->getZExtValue() == SplatIdx)
9204 BaseShAmt = InVec.getOperand(1);
9207 if (BaseShAmt.getNode() == 0)
9208 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9209 DAG.getIntPtrConstant(0));
9213 // The shift amount is an i32.
9214 if (EltVT.bitsGT(MVT::i32))
9215 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9216 else if (EltVT.bitsLT(MVT::i32))
9217 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9219 // The shift amount is identical so we can do a vector shift.
9220 SDValue ValOp = N->getOperand(0);
9221 switch (N->getOpcode()) {
9223 llvm_unreachable("Unknown shift opcode!");
9226 if (VT == MVT::v2i64)
9227 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9228 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9230 if (VT == MVT::v4i32)
9231 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9232 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9234 if (VT == MVT::v8i16)
9235 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9236 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9240 if (VT == MVT::v4i32)
9241 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9242 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9244 if (VT == MVT::v8i16)
9245 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9246 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9250 if (VT == MVT::v2i64)
9251 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9252 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9254 if (VT == MVT::v4i32)
9255 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9256 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9258 if (VT == MVT::v8i16)
9259 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9260 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9267 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9268 const X86Subtarget *Subtarget) {
9269 EVT VT = N->getValueType(0);
9270 if (VT != MVT::i64 || !Subtarget->is64Bit())
9273 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9274 SDValue N0 = N->getOperand(0);
9275 SDValue N1 = N->getOperand(1);
9276 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9278 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9281 SDValue ShAmt0 = N0.getOperand(1);
9282 if (ShAmt0.getValueType() != MVT::i8)
9284 SDValue ShAmt1 = N1.getOperand(1);
9285 if (ShAmt1.getValueType() != MVT::i8)
9287 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9288 ShAmt0 = ShAmt0.getOperand(0);
9289 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9290 ShAmt1 = ShAmt1.getOperand(0);
9292 DebugLoc DL = N->getDebugLoc();
9293 unsigned Opc = X86ISD::SHLD;
9294 SDValue Op0 = N0.getOperand(0);
9295 SDValue Op1 = N1.getOperand(0);
9296 if (ShAmt0.getOpcode() == ISD::SUB) {
9298 std::swap(Op0, Op1);
9299 std::swap(ShAmt0, ShAmt1);
9302 if (ShAmt1.getOpcode() == ISD::SUB) {
9303 SDValue Sum = ShAmt1.getOperand(0);
9304 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9305 if (SumC->getSExtValue() == 64 &&
9306 ShAmt1.getOperand(1) == ShAmt0)
9307 return DAG.getNode(Opc, DL, VT,
9309 DAG.getNode(ISD::TRUNCATE, DL,
9312 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9313 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9315 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9316 return DAG.getNode(Opc, DL, VT,
9317 N0.getOperand(0), N1.getOperand(0),
9318 DAG.getNode(ISD::TRUNCATE, DL,
9325 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9326 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9327 const X86Subtarget *Subtarget) {
9328 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9329 // the FP state in cases where an emms may be missing.
9330 // A preferable solution to the general problem is to figure out the right
9331 // places to insert EMMS. This qualifies as a quick hack.
9333 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9334 StoreSDNode *St = cast<StoreSDNode>(N);
9335 EVT VT = St->getValue().getValueType();
9336 if (VT.getSizeInBits() != 64)
9339 const Function *F = DAG.getMachineFunction().getFunction();
9340 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9341 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9342 && Subtarget->hasSSE2();
9343 if ((VT.isVector() ||
9344 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9345 isa<LoadSDNode>(St->getValue()) &&
9346 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9347 St->getChain().hasOneUse() && !St->isVolatile()) {
9348 SDNode* LdVal = St->getValue().getNode();
9350 int TokenFactorIndex = -1;
9351 SmallVector<SDValue, 8> Ops;
9352 SDNode* ChainVal = St->getChain().getNode();
9353 // Must be a store of a load. We currently handle two cases: the load
9354 // is a direct child, and it's under an intervening TokenFactor. It is
9355 // possible to dig deeper under nested TokenFactors.
9356 if (ChainVal == LdVal)
9357 Ld = cast<LoadSDNode>(St->getChain());
9358 else if (St->getValue().hasOneUse() &&
9359 ChainVal->getOpcode() == ISD::TokenFactor) {
9360 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9361 if (ChainVal->getOperand(i).getNode() == LdVal) {
9362 TokenFactorIndex = i;
9363 Ld = cast<LoadSDNode>(St->getValue());
9365 Ops.push_back(ChainVal->getOperand(i));
9369 if (!Ld || !ISD::isNormalLoad(Ld))
9372 // If this is not the MMX case, i.e. we are just turning i64 load/store
9373 // into f64 load/store, avoid the transformation if there are multiple
9374 // uses of the loaded value.
9375 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9378 DebugLoc LdDL = Ld->getDebugLoc();
9379 DebugLoc StDL = N->getDebugLoc();
9380 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9381 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9383 if (Subtarget->is64Bit() || F64IsLegal) {
9384 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9385 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9386 Ld->getBasePtr(), Ld->getSrcValue(),
9387 Ld->getSrcValueOffset(), Ld->isVolatile(),
9388 Ld->getAlignment());
9389 SDValue NewChain = NewLd.getValue(1);
9390 if (TokenFactorIndex != -1) {
9391 Ops.push_back(NewChain);
9392 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9395 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9396 St->getSrcValue(), St->getSrcValueOffset(),
9397 St->isVolatile(), St->getAlignment());
9400 // Otherwise, lower to two pairs of 32-bit loads / stores.
9401 SDValue LoAddr = Ld->getBasePtr();
9402 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9403 DAG.getConstant(4, MVT::i32));
9405 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9406 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9407 Ld->isVolatile(), Ld->getAlignment());
9408 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9409 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9411 MinAlign(Ld->getAlignment(), 4));
9413 SDValue NewChain = LoLd.getValue(1);
9414 if (TokenFactorIndex != -1) {
9415 Ops.push_back(LoLd);
9416 Ops.push_back(HiLd);
9417 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9421 LoAddr = St->getBasePtr();
9422 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9423 DAG.getConstant(4, MVT::i32));
9425 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9426 St->getSrcValue(), St->getSrcValueOffset(),
9427 St->isVolatile(), St->getAlignment());
9428 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9430 St->getSrcValueOffset() + 4,
9432 MinAlign(St->getAlignment(), 4));
9433 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9438 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9439 /// X86ISD::FXOR nodes.
9440 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9441 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9442 // F[X]OR(0.0, x) -> x
9443 // F[X]OR(x, 0.0) -> x
9444 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9445 if (C->getValueAPF().isPosZero())
9446 return N->getOperand(1);
9447 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9448 if (C->getValueAPF().isPosZero())
9449 return N->getOperand(0);
9453 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9454 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9455 // FAND(0.0, x) -> 0.0
9456 // FAND(x, 0.0) -> 0.0
9457 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9458 if (C->getValueAPF().isPosZero())
9459 return N->getOperand(0);
9460 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9461 if (C->getValueAPF().isPosZero())
9462 return N->getOperand(1);
9466 static SDValue PerformBTCombine(SDNode *N,
9468 TargetLowering::DAGCombinerInfo &DCI) {
9469 // BT ignores high bits in the bit index operand.
9470 SDValue Op1 = N->getOperand(1);
9471 if (Op1.hasOneUse()) {
9472 unsigned BitWidth = Op1.getValueSizeInBits();
9473 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9474 APInt KnownZero, KnownOne;
9475 TargetLowering::TargetLoweringOpt TLO(DAG);
9476 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9477 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9478 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9479 DCI.CommitTargetLoweringOpt(TLO);
9484 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9485 SDValue Op = N->getOperand(0);
9486 if (Op.getOpcode() == ISD::BIT_CONVERT)
9487 Op = Op.getOperand(0);
9488 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9489 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9490 VT.getVectorElementType().getSizeInBits() ==
9491 OpVT.getVectorElementType().getSizeInBits()) {
9492 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9497 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9498 // Locked instructions, in turn, have implicit fence semantics (all memory
9499 // operations are flushed before issuing the locked instruction, and the
9500 // are not buffered), so we can fold away the common pattern of
9501 // fence-atomic-fence.
9502 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9503 SDValue atomic = N->getOperand(0);
9504 switch (atomic.getOpcode()) {
9505 case ISD::ATOMIC_CMP_SWAP:
9506 case ISD::ATOMIC_SWAP:
9507 case ISD::ATOMIC_LOAD_ADD:
9508 case ISD::ATOMIC_LOAD_SUB:
9509 case ISD::ATOMIC_LOAD_AND:
9510 case ISD::ATOMIC_LOAD_OR:
9511 case ISD::ATOMIC_LOAD_XOR:
9512 case ISD::ATOMIC_LOAD_NAND:
9513 case ISD::ATOMIC_LOAD_MIN:
9514 case ISD::ATOMIC_LOAD_MAX:
9515 case ISD::ATOMIC_LOAD_UMIN:
9516 case ISD::ATOMIC_LOAD_UMAX:
9522 SDValue fence = atomic.getOperand(0);
9523 if (fence.getOpcode() != ISD::MEMBARRIER)
9526 switch (atomic.getOpcode()) {
9527 case ISD::ATOMIC_CMP_SWAP:
9528 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9529 atomic.getOperand(1), atomic.getOperand(2),
9530 atomic.getOperand(3));
9531 case ISD::ATOMIC_SWAP:
9532 case ISD::ATOMIC_LOAD_ADD:
9533 case ISD::ATOMIC_LOAD_SUB:
9534 case ISD::ATOMIC_LOAD_AND:
9535 case ISD::ATOMIC_LOAD_OR:
9536 case ISD::ATOMIC_LOAD_XOR:
9537 case ISD::ATOMIC_LOAD_NAND:
9538 case ISD::ATOMIC_LOAD_MIN:
9539 case ISD::ATOMIC_LOAD_MAX:
9540 case ISD::ATOMIC_LOAD_UMIN:
9541 case ISD::ATOMIC_LOAD_UMAX:
9542 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9543 atomic.getOperand(1), atomic.getOperand(2));
9549 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9550 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9551 // (and (i32 x86isd::setcc_carry), 1)
9552 // This eliminates the zext. This transformation is necessary because
9553 // ISD::SETCC is always legalized to i8.
9554 DebugLoc dl = N->getDebugLoc();
9555 SDValue N0 = N->getOperand(0);
9556 EVT VT = N->getValueType(0);
9557 if (N0.getOpcode() == ISD::AND &&
9559 N0.getOperand(0).hasOneUse()) {
9560 SDValue N00 = N0.getOperand(0);
9561 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9563 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9564 if (!C || C->getZExtValue() != 1)
9566 return DAG.getNode(ISD::AND, dl, VT,
9567 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9568 N00.getOperand(0), N00.getOperand(1)),
9569 DAG.getConstant(1, VT));
9575 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9576 DAGCombinerInfo &DCI) const {
9577 SelectionDAG &DAG = DCI.DAG;
9578 switch (N->getOpcode()) {
9580 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9581 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9582 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9583 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9586 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9587 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9588 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9590 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9591 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9592 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9593 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9594 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9595 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9601 //===----------------------------------------------------------------------===//
9602 // X86 Inline Assembly Support
9603 //===----------------------------------------------------------------------===//
9605 static bool LowerToBSwap(CallInst *CI) {
9606 // FIXME: this should verify that we are targetting a 486 or better. If not,
9607 // we will turn this bswap into something that will be lowered to logical ops
9608 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9609 // so don't worry about this.
9611 // Verify this is a simple bswap.
9612 if (CI->getNumOperands() != 2 ||
9613 CI->getType() != CI->getOperand(1)->getType() ||
9614 !CI->getType()->isInteger())
9617 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9618 if (!Ty || Ty->getBitWidth() % 16 != 0)
9621 // Okay, we can do this xform, do so now.
9622 const Type *Tys[] = { Ty };
9623 Module *M = CI->getParent()->getParent()->getParent();
9624 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9626 Value *Op = CI->getOperand(1);
9627 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9629 CI->replaceAllUsesWith(Op);
9630 CI->eraseFromParent();
9634 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9635 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9636 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9638 std::string AsmStr = IA->getAsmString();
9640 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9641 SmallVector<StringRef, 4> AsmPieces;
9642 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9644 switch (AsmPieces.size()) {
9645 default: return false;
9647 AsmStr = AsmPieces[0];
9649 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9652 if (AsmPieces.size() == 2 &&
9653 (AsmPieces[0] == "bswap" ||
9654 AsmPieces[0] == "bswapq" ||
9655 AsmPieces[0] == "bswapl") &&
9656 (AsmPieces[1] == "$0" ||
9657 AsmPieces[1] == "${0:q}")) {
9658 // No need to check constraints, nothing other than the equivalent of
9659 // "=r,0" would be valid here.
9660 return LowerToBSwap(CI);
9662 // rorw $$8, ${0:w} --> llvm.bswap.i16
9663 if (CI->getType()->isInteger(16) &&
9664 AsmPieces.size() == 3 &&
9665 AsmPieces[0] == "rorw" &&
9666 AsmPieces[1] == "$$8," &&
9667 AsmPieces[2] == "${0:w}" &&
9668 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9669 return LowerToBSwap(CI);
9673 if (CI->getType()->isInteger(64) &&
9674 Constraints.size() >= 2 &&
9675 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9676 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9677 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9678 SmallVector<StringRef, 4> Words;
9679 SplitString(AsmPieces[0], Words, " \t");
9680 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9682 SplitString(AsmPieces[1], Words, " \t");
9683 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9685 SplitString(AsmPieces[2], Words, " \t,");
9686 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9687 Words[2] == "%edx") {
9688 return LowerToBSwap(CI);
9700 /// getConstraintType - Given a constraint letter, return the type of
9701 /// constraint it is for this target.
9702 X86TargetLowering::ConstraintType
9703 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9704 if (Constraint.size() == 1) {
9705 switch (Constraint[0]) {
9717 return C_RegisterClass;
9725 return TargetLowering::getConstraintType(Constraint);
9728 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9729 /// with another that has more specific requirements based on the type of the
9730 /// corresponding operand.
9731 const char *X86TargetLowering::
9732 LowerXConstraint(EVT ConstraintVT) const {
9733 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9734 // 'f' like normal targets.
9735 if (ConstraintVT.isFloatingPoint()) {
9736 if (Subtarget->hasSSE2())
9738 if (Subtarget->hasSSE1())
9742 return TargetLowering::LowerXConstraint(ConstraintVT);
9745 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9746 /// vector. If it is invalid, don't add anything to Ops.
9747 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9750 std::vector<SDValue>&Ops,
9751 SelectionDAG &DAG) const {
9752 SDValue Result(0, 0);
9754 switch (Constraint) {
9757 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9758 if (C->getZExtValue() <= 31) {
9759 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9765 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9766 if (C->getZExtValue() <= 63) {
9767 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9773 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9774 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9775 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9781 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9782 if (C->getZExtValue() <= 255) {
9783 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9789 // 32-bit signed value
9790 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9791 const ConstantInt *CI = C->getConstantIntValue();
9792 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9793 C->getSExtValue())) {
9794 // Widen to 64 bits here to get it sign extended.
9795 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9798 // FIXME gcc accepts some relocatable values here too, but only in certain
9799 // memory models; it's complicated.
9804 // 32-bit unsigned value
9805 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9806 const ConstantInt *CI = C->getConstantIntValue();
9807 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9808 C->getZExtValue())) {
9809 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9813 // FIXME gcc accepts some relocatable values here too, but only in certain
9814 // memory models; it's complicated.
9818 // Literal immediates are always ok.
9819 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9820 // Widen to 64 bits here to get it sign extended.
9821 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9825 // If we are in non-pic codegen mode, we allow the address of a global (with
9826 // an optional displacement) to be used with 'i'.
9827 GlobalAddressSDNode *GA = 0;
9830 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9832 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9833 Offset += GA->getOffset();
9835 } else if (Op.getOpcode() == ISD::ADD) {
9836 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9837 Offset += C->getZExtValue();
9838 Op = Op.getOperand(0);
9841 } else if (Op.getOpcode() == ISD::SUB) {
9842 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9843 Offset += -C->getZExtValue();
9844 Op = Op.getOperand(0);
9849 // Otherwise, this isn't something we can handle, reject it.
9853 GlobalValue *GV = GA->getGlobal();
9854 // If we require an extra load to get this address, as in PIC mode, we
9856 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9857 getTargetMachine())))
9861 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9863 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9869 if (Result.getNode()) {
9870 Ops.push_back(Result);
9873 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9877 std::vector<unsigned> X86TargetLowering::
9878 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9880 if (Constraint.size() == 1) {
9881 // FIXME: not handling fp-stack yet!
9882 switch (Constraint[0]) { // GCC X86 Constraint Letters
9883 default: break; // Unknown constraint letter
9884 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9885 if (Subtarget->is64Bit()) {
9887 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9888 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9889 X86::R10D,X86::R11D,X86::R12D,
9890 X86::R13D,X86::R14D,X86::R15D,
9891 X86::EBP, X86::ESP, 0);
9892 else if (VT == MVT::i16)
9893 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9894 X86::SI, X86::DI, X86::R8W,X86::R9W,
9895 X86::R10W,X86::R11W,X86::R12W,
9896 X86::R13W,X86::R14W,X86::R15W,
9897 X86::BP, X86::SP, 0);
9898 else if (VT == MVT::i8)
9899 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9900 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9901 X86::R10B,X86::R11B,X86::R12B,
9902 X86::R13B,X86::R14B,X86::R15B,
9903 X86::BPL, X86::SPL, 0);
9905 else if (VT == MVT::i64)
9906 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9907 X86::RSI, X86::RDI, X86::R8, X86::R9,
9908 X86::R10, X86::R11, X86::R12,
9909 X86::R13, X86::R14, X86::R15,
9910 X86::RBP, X86::RSP, 0);
9914 // 32-bit fallthrough
9917 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9918 else if (VT == MVT::i16)
9919 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9920 else if (VT == MVT::i8)
9921 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9922 else if (VT == MVT::i64)
9923 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9928 return std::vector<unsigned>();
9931 std::pair<unsigned, const TargetRegisterClass*>
9932 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9934 // First, see if this is a constraint that directly corresponds to an LLVM
9936 if (Constraint.size() == 1) {
9937 // GCC Constraint Letters
9938 switch (Constraint[0]) {
9940 case 'r': // GENERAL_REGS
9941 case 'l': // INDEX_REGS
9943 return std::make_pair(0U, X86::GR8RegisterClass);
9945 return std::make_pair(0U, X86::GR16RegisterClass);
9946 if (VT == MVT::i32 || !Subtarget->is64Bit())
9947 return std::make_pair(0U, X86::GR32RegisterClass);
9948 return std::make_pair(0U, X86::GR64RegisterClass);
9949 case 'R': // LEGACY_REGS
9951 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9953 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9954 if (VT == MVT::i32 || !Subtarget->is64Bit())
9955 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9956 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9957 case 'f': // FP Stack registers.
9958 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9959 // value to the correct fpstack register class.
9960 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9961 return std::make_pair(0U, X86::RFP32RegisterClass);
9962 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9963 return std::make_pair(0U, X86::RFP64RegisterClass);
9964 return std::make_pair(0U, X86::RFP80RegisterClass);
9965 case 'y': // MMX_REGS if MMX allowed.
9966 if (!Subtarget->hasMMX()) break;
9967 return std::make_pair(0U, X86::VR64RegisterClass);
9968 case 'Y': // SSE_REGS if SSE2 allowed
9969 if (!Subtarget->hasSSE2()) break;
9971 case 'x': // SSE_REGS if SSE1 allowed
9972 if (!Subtarget->hasSSE1()) break;
9974 switch (VT.getSimpleVT().SimpleTy) {
9976 // Scalar SSE types.
9979 return std::make_pair(0U, X86::FR32RegisterClass);
9982 return std::make_pair(0U, X86::FR64RegisterClass);
9990 return std::make_pair(0U, X86::VR128RegisterClass);
9996 // Use the default implementation in TargetLowering to convert the register
9997 // constraint into a member of a register class.
9998 std::pair<unsigned, const TargetRegisterClass*> Res;
9999 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10001 // Not found as a standard register?
10002 if (Res.second == 0) {
10003 // Map st(0) -> st(7) -> ST0
10004 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10005 tolower(Constraint[1]) == 's' &&
10006 tolower(Constraint[2]) == 't' &&
10007 Constraint[3] == '(' &&
10008 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10009 Constraint[5] == ')' &&
10010 Constraint[6] == '}') {
10012 Res.first = X86::ST0+Constraint[4]-'0';
10013 Res.second = X86::RFP80RegisterClass;
10017 // GCC allows "st(0)" to be called just plain "st".
10018 if (StringRef("{st}").equals_lower(Constraint)) {
10019 Res.first = X86::ST0;
10020 Res.second = X86::RFP80RegisterClass;
10025 if (StringRef("{flags}").equals_lower(Constraint)) {
10026 Res.first = X86::EFLAGS;
10027 Res.second = X86::CCRRegisterClass;
10031 // 'A' means EAX + EDX.
10032 if (Constraint == "A") {
10033 Res.first = X86::EAX;
10034 Res.second = X86::GR32_ADRegisterClass;
10040 // Otherwise, check to see if this is a register class of the wrong value
10041 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10042 // turn into {ax},{dx}.
10043 if (Res.second->hasType(VT))
10044 return Res; // Correct type already, nothing to do.
10046 // All of the single-register GCC register classes map their values onto
10047 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10048 // really want an 8-bit or 32-bit register, map to the appropriate register
10049 // class and return the appropriate register.
10050 if (Res.second == X86::GR16RegisterClass) {
10051 if (VT == MVT::i8) {
10052 unsigned DestReg = 0;
10053 switch (Res.first) {
10055 case X86::AX: DestReg = X86::AL; break;
10056 case X86::DX: DestReg = X86::DL; break;
10057 case X86::CX: DestReg = X86::CL; break;
10058 case X86::BX: DestReg = X86::BL; break;
10061 Res.first = DestReg;
10062 Res.second = X86::GR8RegisterClass;
10064 } else if (VT == MVT::i32) {
10065 unsigned DestReg = 0;
10066 switch (Res.first) {
10068 case X86::AX: DestReg = X86::EAX; break;
10069 case X86::DX: DestReg = X86::EDX; break;
10070 case X86::CX: DestReg = X86::ECX; break;
10071 case X86::BX: DestReg = X86::EBX; break;
10072 case X86::SI: DestReg = X86::ESI; break;
10073 case X86::DI: DestReg = X86::EDI; break;
10074 case X86::BP: DestReg = X86::EBP; break;
10075 case X86::SP: DestReg = X86::ESP; break;
10078 Res.first = DestReg;
10079 Res.second = X86::GR32RegisterClass;
10081 } else if (VT == MVT::i64) {
10082 unsigned DestReg = 0;
10083 switch (Res.first) {
10085 case X86::AX: DestReg = X86::RAX; break;
10086 case X86::DX: DestReg = X86::RDX; break;
10087 case X86::CX: DestReg = X86::RCX; break;
10088 case X86::BX: DestReg = X86::RBX; break;
10089 case X86::SI: DestReg = X86::RSI; break;
10090 case X86::DI: DestReg = X86::RDI; break;
10091 case X86::BP: DestReg = X86::RBP; break;
10092 case X86::SP: DestReg = X86::RSP; break;
10095 Res.first = DestReg;
10096 Res.second = X86::GR64RegisterClass;
10099 } else if (Res.second == X86::FR32RegisterClass ||
10100 Res.second == X86::FR64RegisterClass ||
10101 Res.second == X86::VR128RegisterClass) {
10102 // Handle references to XMM physical registers that got mapped into the
10103 // wrong class. This can happen with constraints like {xmm0} where the
10104 // target independent register mapper will just pick the first match it can
10105 // find, ignoring the required type.
10106 if (VT == MVT::f32)
10107 Res.second = X86::FR32RegisterClass;
10108 else if (VT == MVT::f64)
10109 Res.second = X86::FR64RegisterClass;
10110 else if (X86::VR128RegisterClass->hasType(VT))
10111 Res.second = X86::VR128RegisterClass;
10117 //===----------------------------------------------------------------------===//
10118 // X86 Widen vector type
10119 //===----------------------------------------------------------------------===//
10121 /// getWidenVectorType: given a vector type, returns the type to widen
10122 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10123 /// If there is no vector type that we want to widen to, returns MVT::Other
10124 /// When and where to widen is target dependent based on the cost of
10125 /// scalarizing vs using the wider vector type.
10127 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10128 assert(VT.isVector());
10129 if (isTypeLegal(VT))
10132 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10133 // type based on element type. This would speed up our search (though
10134 // it may not be worth it since the size of the list is relatively
10136 EVT EltVT = VT.getVectorElementType();
10137 unsigned NElts = VT.getVectorNumElements();
10139 // On X86, it make sense to widen any vector wider than 1
10143 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10144 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10145 EVT SVT = (MVT::SimpleValueType)nVT;
10147 if (isTypeLegal(SVT) &&
10148 SVT.getVectorElementType() == EltVT &&
10149 SVT.getVectorNumElements() > NElts)