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 "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/SmallSet.h"
41 #include "llvm/ADT/StringExtras.h"
44 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
45 : TargetLowering(TM) {
46 Subtarget = &TM.getSubtarget<X86Subtarget>();
47 X86ScalarSSEf64 = Subtarget->hasSSE2();
48 X86ScalarSSEf32 = Subtarget->hasSSE1();
49 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
53 RegInfo = TM.getRegisterInfo();
55 // Set up the TargetLowering object.
57 // X86 is weird, it always uses i8 for shift amounts and setcc results.
58 setShiftAmountType(MVT::i8);
59 setSetCCResultContents(ZeroOrOneSetCCResult);
60 setSchedulingPreference(SchedulingForRegPressure);
61 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
62 setStackPointerRegisterToSaveRestore(X86StackPtr);
64 if (Subtarget->isTargetDarwin()) {
65 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
66 setUseUnderscoreSetJmp(false);
67 setUseUnderscoreLongJmp(false);
68 } else if (Subtarget->isTargetMingw()) {
69 // MS runtime is weird: it exports _setjmp, but longjmp!
70 setUseUnderscoreSetJmp(true);
71 setUseUnderscoreLongJmp(false);
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
77 // Set up the register classes.
78 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
79 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
80 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
81 if (Subtarget->is64Bit())
82 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
84 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
86 // We don't accept any truncstore of integer registers.
87 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
88 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
89 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
90 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
91 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
92 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
94 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
96 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
97 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
100 if (Subtarget->is64Bit()) {
101 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
102 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
105 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
106 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
108 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
111 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
113 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
114 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
115 // SSE has no i16 to fp conversion, only i32
116 if (X86ScalarSSEf32) {
117 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
118 // f32 and f64 cases are Legal, f80 case is not
119 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
121 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
122 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
125 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
126 // are Legal, f80 is custom lowered.
127 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
128 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
130 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
132 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
135 if (X86ScalarSSEf32) {
136 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
137 // f32 and f64 cases are Legal, f80 case is not
138 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
140 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
141 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
144 // Handle FP_TO_UINT by promoting the destination to a larger signed
146 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
147 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
148 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
150 if (Subtarget->is64Bit()) {
151 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
152 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
154 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
155 // Expand FP_TO_UINT into a select.
156 // FIXME: We would like to use a Custom expander here eventually to do
157 // the optimal thing for SSE vs. the default expansion in the legalizer.
158 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
160 // With SSE3 we can use fisttpll to convert to a signed i64.
161 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
164 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
165 if (!X86ScalarSSEf64) {
166 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
167 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
170 // Scalar integer divide and remainder are lowered to use operations that
171 // produce two results, to match the available instructions. This exposes
172 // the two-result form to trivial CSE, which is able to combine x/y and x%y
173 // into a single instruction.
175 // Scalar integer multiply-high is also lowered to use two-result
176 // operations, to match the available instructions. However, plain multiply
177 // (low) operations are left as Legal, as there are single-result
178 // instructions for this in x86. Using the two-result multiply instructions
179 // when both high and low results are needed must be arranged by dagcombine.
180 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
181 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
182 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
183 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
184 setOperationAction(ISD::SREM , MVT::i8 , Expand);
185 setOperationAction(ISD::UREM , MVT::i8 , Expand);
186 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
187 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
188 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
189 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
190 setOperationAction(ISD::SREM , MVT::i16 , Expand);
191 setOperationAction(ISD::UREM , MVT::i16 , Expand);
192 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
193 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
194 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
195 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
196 setOperationAction(ISD::SREM , MVT::i32 , Expand);
197 setOperationAction(ISD::UREM , MVT::i32 , Expand);
198 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
199 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
200 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
201 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
202 setOperationAction(ISD::SREM , MVT::i64 , Expand);
203 setOperationAction(ISD::UREM , MVT::i64 , Expand);
205 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
206 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
207 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
208 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
209 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
210 if (Subtarget->is64Bit())
211 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
212 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
215 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
216 setOperationAction(ISD::FREM , MVT::f32 , Expand);
217 setOperationAction(ISD::FREM , MVT::f64 , Expand);
218 setOperationAction(ISD::FREM , MVT::f80 , Expand);
219 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
221 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
222 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
223 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
224 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
225 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
226 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
227 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
228 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
229 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
230 if (Subtarget->is64Bit()) {
231 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
232 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
233 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
236 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
237 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
239 // These should be promoted to a larger select which is supported.
240 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
241 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
242 // X86 wants to expand cmov itself.
243 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
244 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
245 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
246 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
248 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
249 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
251 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
252 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
256 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
258 // X86 ret instruction may pop stack.
259 setOperationAction(ISD::RET , MVT::Other, Custom);
260 if (!Subtarget->is64Bit())
261 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
264 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
265 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
266 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
267 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
268 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
269 if (Subtarget->is64Bit()) {
270 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
271 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
272 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
273 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
275 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
276 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
277 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
278 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
279 if (Subtarget->is64Bit()) {
280 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
281 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
282 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
284 // X86 wants to expand memset / memcpy itself.
285 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
286 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
288 if (Subtarget->hasSSE1())
289 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
291 if (!Subtarget->hasSSE2())
292 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
294 setOperationAction(ISD::ATOMIC_LCS , MVT::i8, Custom);
295 setOperationAction(ISD::ATOMIC_LCS , MVT::i16, Custom);
296 setOperationAction(ISD::ATOMIC_LCS , MVT::i32, Custom);
297 setOperationAction(ISD::ATOMIC_LCS , MVT::i64, Custom);
299 // Use the default ISD::LOCATION, ISD::DECLARE expansion.
300 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
301 // FIXME - use subtarget debug flags
302 if (!Subtarget->isTargetDarwin() &&
303 !Subtarget->isTargetELF() &&
304 !Subtarget->isTargetCygMing())
305 setOperationAction(ISD::LABEL, MVT::Other, Expand);
307 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
308 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
309 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
310 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
311 if (Subtarget->is64Bit()) {
313 setExceptionPointerRegister(X86::RAX);
314 setExceptionSelectorRegister(X86::RDX);
316 setExceptionPointerRegister(X86::EAX);
317 setExceptionSelectorRegister(X86::EDX);
319 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
321 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
323 setOperationAction(ISD::TRAP, MVT::Other, Legal);
325 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
326 setOperationAction(ISD::VASTART , MVT::Other, Custom);
327 setOperationAction(ISD::VAARG , MVT::Other, Expand);
328 setOperationAction(ISD::VAEND , MVT::Other, Expand);
329 if (Subtarget->is64Bit())
330 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
332 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
334 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
335 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
338 if (Subtarget->isTargetCygMing())
339 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
341 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
343 if (X86ScalarSSEf64) {
344 // f32 and f64 use SSE.
345 // Set up the FP register classes.
346 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
347 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
349 // Use ANDPD to simulate FABS.
350 setOperationAction(ISD::FABS , MVT::f64, Custom);
351 setOperationAction(ISD::FABS , MVT::f32, Custom);
353 // Use XORP to simulate FNEG.
354 setOperationAction(ISD::FNEG , MVT::f64, Custom);
355 setOperationAction(ISD::FNEG , MVT::f32, Custom);
357 // Use ANDPD and ORPD to simulate FCOPYSIGN.
358 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
359 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
361 // We don't support sin/cos/fmod
362 setOperationAction(ISD::FSIN , MVT::f64, Expand);
363 setOperationAction(ISD::FCOS , MVT::f64, Expand);
364 setOperationAction(ISD::FSIN , MVT::f32, Expand);
365 setOperationAction(ISD::FCOS , MVT::f32, Expand);
367 // Expand FP immediates into loads from the stack, except for the special
369 addLegalFPImmediate(APFloat(+0.0)); // xorpd
370 addLegalFPImmediate(APFloat(+0.0f)); // xorps
372 // Floating truncations from f80 and extensions to f80 go through memory.
373 // If optimizing, we lie about this though and handle it in
374 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
376 setConvertAction(MVT::f32, MVT::f80, Expand);
377 setConvertAction(MVT::f64, MVT::f80, Expand);
378 setConvertAction(MVT::f80, MVT::f32, Expand);
379 setConvertAction(MVT::f80, MVT::f64, Expand);
381 } else if (X86ScalarSSEf32) {
382 // Use SSE for f32, x87 for f64.
383 // Set up the FP register classes.
384 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
385 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
387 // Use ANDPS to simulate FABS.
388 setOperationAction(ISD::FABS , MVT::f32, Custom);
390 // Use XORP to simulate FNEG.
391 setOperationAction(ISD::FNEG , MVT::f32, Custom);
393 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
395 // Use ANDPS and ORPS to simulate FCOPYSIGN.
396 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
397 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
399 // We don't support sin/cos/fmod
400 setOperationAction(ISD::FSIN , MVT::f32, Expand);
401 setOperationAction(ISD::FCOS , MVT::f32, Expand);
403 // Special cases we handle for FP constants.
404 addLegalFPImmediate(APFloat(+0.0f)); // xorps
405 addLegalFPImmediate(APFloat(+0.0)); // FLD0
406 addLegalFPImmediate(APFloat(+1.0)); // FLD1
407 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
408 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
410 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
411 // this though and handle it in InstructionSelectPreprocess so that
412 // dagcombine2 can hack on these.
414 setConvertAction(MVT::f32, MVT::f64, Expand);
415 setConvertAction(MVT::f32, MVT::f80, Expand);
416 setConvertAction(MVT::f80, MVT::f32, Expand);
417 setConvertAction(MVT::f64, MVT::f32, Expand);
418 // And x87->x87 truncations also.
419 setConvertAction(MVT::f80, MVT::f64, Expand);
423 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
424 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
427 // f32 and f64 in x87.
428 // Set up the FP register classes.
429 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
430 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
432 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
433 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
434 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
435 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
437 // Floating truncations go through memory. If optimizing, we lie about
438 // this though and handle it in InstructionSelectPreprocess so that
439 // dagcombine2 can hack on these.
441 setConvertAction(MVT::f80, MVT::f32, Expand);
442 setConvertAction(MVT::f64, MVT::f32, Expand);
443 setConvertAction(MVT::f80, MVT::f64, Expand);
447 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
448 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
450 addLegalFPImmediate(APFloat(+0.0)); // FLD0
451 addLegalFPImmediate(APFloat(+1.0)); // FLD1
452 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
453 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
454 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
455 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
456 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
457 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
460 // Long double always uses X87.
461 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
462 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
463 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
465 APFloat TmpFlt(+0.0);
466 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
467 addLegalFPImmediate(TmpFlt); // FLD0
469 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
470 APFloat TmpFlt2(+1.0);
471 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
472 addLegalFPImmediate(TmpFlt2); // FLD1
473 TmpFlt2.changeSign();
474 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
478 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
479 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
482 // Always use a library call for pow.
483 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
484 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
485 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
487 // First set operation action for all vector types to expand. Then we
488 // will selectively turn on ones that can be effectively codegen'd.
489 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
490 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
491 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
492 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
493 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
494 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
495 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
496 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
497 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
498 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
499 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
500 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
501 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
502 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
503 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
504 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
505 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
506 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
507 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
508 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
509 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
510 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
511 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
512 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
513 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
514 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
515 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
516 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
517 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
518 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
519 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
520 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
521 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
522 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
523 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
524 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
525 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
526 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
527 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
530 if (Subtarget->hasMMX()) {
531 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
532 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
533 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
534 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
536 // FIXME: add MMX packed arithmetics
538 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
539 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
540 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
541 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
543 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
544 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
545 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
546 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
548 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
549 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
551 setOperationAction(ISD::AND, MVT::v8i8, Promote);
552 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
553 setOperationAction(ISD::AND, MVT::v4i16, Promote);
554 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
555 setOperationAction(ISD::AND, MVT::v2i32, Promote);
556 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
557 setOperationAction(ISD::AND, MVT::v1i64, Legal);
559 setOperationAction(ISD::OR, MVT::v8i8, Promote);
560 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
561 setOperationAction(ISD::OR, MVT::v4i16, Promote);
562 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
563 setOperationAction(ISD::OR, MVT::v2i32, Promote);
564 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
565 setOperationAction(ISD::OR, MVT::v1i64, Legal);
567 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
568 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
569 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
570 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
571 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
572 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
573 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
575 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
576 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
577 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
578 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
579 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
580 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
581 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
583 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
584 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
585 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
586 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
588 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
589 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
590 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
591 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
593 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
594 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
595 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
598 if (Subtarget->hasSSE1()) {
599 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
601 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
602 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
603 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
604 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
605 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
606 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
607 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
608 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
609 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
610 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
611 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
614 if (Subtarget->hasSSE2()) {
615 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
616 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
617 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
618 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
619 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
621 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
625 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
629 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
630 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
631 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
632 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
633 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
634 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
635 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
637 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
638 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
639 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
640 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
641 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
643 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
644 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
645 // Do not attempt to custom lower non-power-of-2 vectors
646 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
648 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
649 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
650 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
652 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
653 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
656 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
657 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
658 if (Subtarget->is64Bit()) {
659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
660 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
663 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
664 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
665 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
666 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
667 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
668 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
669 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
670 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
671 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
672 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
673 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
674 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
677 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
679 // Custom lower v2i64 and v2f64 selects.
680 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
681 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
682 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
683 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
686 if (Subtarget->hasSSE41()) {
687 // FIXME: Do we need to handle scalar-to-vector here?
688 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
690 // i8 and i16 vectors are custom , because the source register and source
691 // source memory operand types are not the same width. f32 vectors are
692 // custom since the immediate controlling the insert encodes additional
694 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
696 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
697 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
699 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
701 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
702 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal);
704 if (Subtarget->is64Bit()) {
705 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
706 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
710 // We want to custom lower some of our intrinsics.
711 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
713 // We have target-specific dag combine patterns for the following nodes:
714 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
715 setTargetDAGCombine(ISD::SELECT);
716 setTargetDAGCombine(ISD::STORE);
718 computeRegisterProperties();
720 // FIXME: These should be based on subtarget info. Plus, the values should
721 // be smaller when we are in optimizing for size mode.
722 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
723 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
724 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
725 allowUnalignedMemoryAccesses = true; // x86 supports it!
726 setPrefLoopAlignment(16);
731 X86TargetLowering::getSetCCResultType(const SDOperand &) const {
736 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
737 /// the desired ByVal argument alignment.
738 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
741 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
742 if (VTy->getBitWidth() == 128)
744 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
745 unsigned EltAlign = 0;
746 getMaxByValAlign(ATy->getElementType(), EltAlign);
747 if (EltAlign > MaxAlign)
749 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
750 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
751 unsigned EltAlign = 0;
752 getMaxByValAlign(STy->getElementType(i), EltAlign);
753 if (EltAlign > MaxAlign)
762 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
763 /// function arguments in the caller parameter area. For X86, aggregates
764 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
765 /// are at 4-byte boundaries.
766 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
767 if (Subtarget->is64Bit())
768 return getTargetData()->getABITypeAlignment(Ty);
770 if (Subtarget->hasSSE1())
771 getMaxByValAlign(Ty, Align);
775 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
777 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
778 SelectionDAG &DAG) const {
779 if (usesGlobalOffsetTable())
780 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
781 if (!Subtarget->isPICStyleRIPRel())
782 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
786 //===----------------------------------------------------------------------===//
787 // Return Value Calling Convention Implementation
788 //===----------------------------------------------------------------------===//
790 #include "X86GenCallingConv.inc"
792 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
793 /// exists skip possible ISD:TokenFactor.
794 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
795 if (Chain.getOpcode() == X86ISD::TAILCALL) {
797 } else if (Chain.getOpcode() == ISD::TokenFactor) {
798 if (Chain.getNumOperands() &&
799 Chain.getOperand(0).getOpcode() == X86ISD::TAILCALL)
800 return Chain.getOperand(0);
805 /// LowerRET - Lower an ISD::RET node.
806 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
807 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
809 SmallVector<CCValAssign, 16> RVLocs;
810 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
811 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
812 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
813 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
815 // If this is the first return lowered for this function, add the regs to the
816 // liveout set for the function.
817 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
818 for (unsigned i = 0; i != RVLocs.size(); ++i)
819 if (RVLocs[i].isRegLoc())
820 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
822 SDOperand Chain = Op.getOperand(0);
824 // Handle tail call return.
825 Chain = GetPossiblePreceedingTailCall(Chain);
826 if (Chain.getOpcode() == X86ISD::TAILCALL) {
827 SDOperand TailCall = Chain;
828 SDOperand TargetAddress = TailCall.getOperand(1);
829 SDOperand StackAdjustment = TailCall.getOperand(2);
830 assert(((TargetAddress.getOpcode() == ISD::Register &&
831 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
832 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
833 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
834 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
835 "Expecting an global address, external symbol, or register");
836 assert(StackAdjustment.getOpcode() == ISD::Constant &&
837 "Expecting a const value");
839 SmallVector<SDOperand,8> Operands;
840 Operands.push_back(Chain.getOperand(0));
841 Operands.push_back(TargetAddress);
842 Operands.push_back(StackAdjustment);
843 // Copy registers used by the call. Last operand is a flag so it is not
845 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
846 Operands.push_back(Chain.getOperand(i));
848 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
855 SmallVector<SDOperand, 6> RetOps;
856 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
857 // Operand #1 = Bytes To Pop
858 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
860 // Copy the result values into the output registers.
861 for (unsigned i = 0; i != RVLocs.size(); ++i) {
862 CCValAssign &VA = RVLocs[i];
863 assert(VA.isRegLoc() && "Can only return in registers!");
864 SDOperand ValToCopy = Op.getOperand(i*2+1);
866 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
867 // the RET instruction and handled by the FP Stackifier.
868 if (RVLocs[i].getLocReg() == X86::ST0 ||
869 RVLocs[i].getLocReg() == X86::ST1) {
870 // If this is a copy from an xmm register to ST(0), use an FPExtend to
871 // change the value to the FP stack register class.
872 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
873 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
874 RetOps.push_back(ValToCopy);
875 // Don't emit a copytoreg.
879 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
880 Flag = Chain.getValue(1);
883 RetOps[0] = Chain; // Update chain.
885 // Add the flag if we have it.
887 RetOps.push_back(Flag);
889 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
893 /// LowerCallResult - Lower the result values of an ISD::CALL into the
894 /// appropriate copies out of appropriate physical registers. This assumes that
895 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
896 /// being lowered. The returns a SDNode with the same number of values as the
898 SDNode *X86TargetLowering::
899 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
900 unsigned CallingConv, SelectionDAG &DAG) {
902 // Assign locations to each value returned by this call.
903 SmallVector<CCValAssign, 16> RVLocs;
904 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
905 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
906 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
908 SmallVector<SDOperand, 8> ResultVals;
910 // Copy all of the result registers out of their specified physreg.
911 for (unsigned i = 0; i != RVLocs.size(); ++i) {
912 MVT::ValueType CopyVT = RVLocs[i].getValVT();
914 // If this is a call to a function that returns an fp value on the floating
915 // point stack, but where we prefer to use the value in xmm registers, copy
916 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
917 if (RVLocs[i].getLocReg() == X86::ST0 &&
918 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
922 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
923 CopyVT, InFlag).getValue(1);
924 SDOperand Val = Chain.getValue(0);
925 InFlag = Chain.getValue(2);
927 if (CopyVT != RVLocs[i].getValVT()) {
928 // Round the F80 the right size, which also moves to the appropriate xmm
930 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
931 // This truncation won't change the value.
932 DAG.getIntPtrConstant(1));
935 ResultVals.push_back(Val);
938 // Merge everything together with a MERGE_VALUES node.
939 ResultVals.push_back(Chain);
940 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
941 &ResultVals[0], ResultVals.size()).Val;
945 //===----------------------------------------------------------------------===//
946 // C & StdCall & Fast Calling Convention implementation
947 //===----------------------------------------------------------------------===//
948 // StdCall calling convention seems to be standard for many Windows' API
949 // routines and around. It differs from C calling convention just a little:
950 // callee should clean up the stack, not caller. Symbols should be also
951 // decorated in some fancy way :) It doesn't support any vector arguments.
952 // For info on fast calling convention see Fast Calling Convention (tail call)
953 // implementation LowerX86_32FastCCCallTo.
955 /// AddLiveIn - This helper function adds the specified physical register to the
956 /// MachineFunction as a live in value. It also creates a corresponding virtual
958 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
959 const TargetRegisterClass *RC) {
960 assert(RC->contains(PReg) && "Not the correct regclass!");
961 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
962 MF.getRegInfo().addLiveIn(PReg, VReg);
966 /// CallIsStructReturn - Determines whether a CALL node uses struct return
968 static bool CallIsStructReturn(SDOperand Op) {
969 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
973 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
976 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
977 /// return semantics.
978 static bool ArgsAreStructReturn(SDOperand Op) {
979 unsigned NumArgs = Op.Val->getNumValues() - 1;
983 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
986 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires the
987 /// callee to pop its own arguments. Callee pop is necessary to support tail
989 bool X86TargetLowering::IsCalleePop(SDOperand Op) {
990 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
994 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
997 case CallingConv::X86_StdCall:
998 return !Subtarget->is64Bit();
999 case CallingConv::X86_FastCall:
1000 return !Subtarget->is64Bit();
1001 case CallingConv::Fast:
1002 return PerformTailCallOpt;
1006 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1007 /// FORMAL_ARGUMENTS node.
1008 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDOperand Op) const {
1009 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1011 if (Subtarget->is64Bit()) {
1012 if (CC == CallingConv::Fast && PerformTailCallOpt)
1013 return CC_X86_64_TailCall;
1018 if (CC == CallingConv::X86_FastCall)
1019 return CC_X86_32_FastCall;
1020 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1021 return CC_X86_32_TailCall;
1026 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1027 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1029 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDOperand Op) {
1030 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1031 if (CC == CallingConv::X86_FastCall)
1033 else if (CC == CallingConv::X86_StdCall)
1038 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
1039 /// possibly be overwritten when lowering the outgoing arguments in a tail
1040 /// call. Currently the implementation of this call is very conservative and
1041 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
1042 /// virtual registers would be overwritten by direct lowering.
1043 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op,
1044 MachineFrameInfo * MFI) {
1045 RegisterSDNode * OpReg = NULL;
1046 FrameIndexSDNode * FrameIdxNode = NULL;
1048 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
1049 (Op.getOpcode()== ISD::CopyFromReg &&
1050 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
1051 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
1052 (Op.getOpcode() == ISD::LOAD &&
1053 (FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op.getOperand(1))) &&
1054 (MFI->isFixedObjectIndex((FrameIdx = FrameIdxNode->getIndex()))) &&
1055 (MFI->getObjectOffset(FrameIdx) >= 0)))
1060 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1061 /// in a register before calling.
1062 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1063 return !IsTailCall && !Is64Bit &&
1064 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1065 Subtarget->isPICStyleGOT();
1069 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1070 /// address to be loaded in a register.
1072 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1073 return !Is64Bit && IsTailCall &&
1074 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1075 Subtarget->isPICStyleGOT();
1078 /// CopyTailCallClobberedArgumentsToVRegs - Create virtual registers for all
1079 /// arguments to force loading and guarantee that arguments sourcing from
1080 /// incomming parameters are not overwriting each other.
1082 CopyTailCallClobberedArgumentsToVRegs(SDOperand Chain,
1083 SmallVector<std::pair<unsigned, SDOperand>, 8> &TailCallClobberedVRegs,
1085 MachineFunction &MF,
1086 const TargetLowering * TL) {
1089 for (unsigned i = 0, e = TailCallClobberedVRegs.size(); i != e; i++) {
1090 SDOperand Arg = TailCallClobberedVRegs[i].second;
1091 unsigned Idx = TailCallClobberedVRegs[i].first;
1094 createVirtualRegister(TL->getRegClassFor(Arg.getValueType()));
1095 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
1096 InFlag = Chain.getValue(1);
1097 Arg = DAG.getCopyFromReg(Chain, VReg, Arg.getValueType(), InFlag);
1098 TailCallClobberedVRegs[i] = std::make_pair(Idx, Arg);
1099 Chain = Arg.getValue(1);
1100 InFlag = Arg.getValue(2);
1105 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1106 /// by "Src" to address "Dst" with size and alignment information specified by
1107 /// the specific parameter attribute. The copy will be passed as a byval function
1110 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1111 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1112 SDOperand AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1113 SDOperand SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1114 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1115 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1118 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
1119 const CCValAssign &VA,
1120 MachineFrameInfo *MFI,
1122 SDOperand Root, unsigned i) {
1123 // Create the nodes corresponding to a load from this parameter slot.
1124 ISD::ArgFlagsTy Flags =
1125 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1126 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1127 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1129 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1130 // changed with more analysis.
1131 // In case of tail call optimization mark all arguments mutable. Since they
1132 // could be overwritten by lowering of arguments in case of a tail call.
1133 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1134 VA.getLocMemOffset(), isImmutable);
1135 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1136 if (Flags.isByVal())
1138 return DAG.getLoad(VA.getValVT(), Root, FIN,
1139 PseudoSourceValue::getFixedStack(), FI);
1143 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
1144 MachineFunction &MF = DAG.getMachineFunction();
1145 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1147 const Function* Fn = MF.getFunction();
1148 if (Fn->hasExternalLinkage() &&
1149 Subtarget->isTargetCygMing() &&
1150 Fn->getName() == "main")
1151 FuncInfo->setForceFramePointer(true);
1153 // Decorate the function name.
1154 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1156 MachineFrameInfo *MFI = MF.getFrameInfo();
1157 SDOperand Root = Op.getOperand(0);
1158 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1159 unsigned CC = MF.getFunction()->getCallingConv();
1160 bool Is64Bit = Subtarget->is64Bit();
1162 assert(!(isVarArg && CC == CallingConv::Fast) &&
1163 "Var args not supported with calling convention fastcc");
1165 // Assign locations to all of the incoming arguments.
1166 SmallVector<CCValAssign, 16> ArgLocs;
1167 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1168 CCInfo.AnalyzeFormalArguments(Op.Val, CCAssignFnForNode(Op));
1170 SmallVector<SDOperand, 8> ArgValues;
1171 unsigned LastVal = ~0U;
1172 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1173 CCValAssign &VA = ArgLocs[i];
1174 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1176 assert(VA.getValNo() != LastVal &&
1177 "Don't support value assigned to multiple locs yet");
1178 LastVal = VA.getValNo();
1180 if (VA.isRegLoc()) {
1181 MVT::ValueType RegVT = VA.getLocVT();
1182 TargetRegisterClass *RC;
1183 if (RegVT == MVT::i32)
1184 RC = X86::GR32RegisterClass;
1185 else if (Is64Bit && RegVT == MVT::i64)
1186 RC = X86::GR64RegisterClass;
1187 else if (RegVT == MVT::f32)
1188 RC = X86::FR32RegisterClass;
1189 else if (RegVT == MVT::f64)
1190 RC = X86::FR64RegisterClass;
1192 assert(MVT::isVector(RegVT));
1193 if (Is64Bit && MVT::getSizeInBits(RegVT) == 64) {
1194 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1197 RC = X86::VR128RegisterClass;
1200 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1201 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1203 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1204 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1206 if (VA.getLocInfo() == CCValAssign::SExt)
1207 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1208 DAG.getValueType(VA.getValVT()));
1209 else if (VA.getLocInfo() == CCValAssign::ZExt)
1210 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1211 DAG.getValueType(VA.getValVT()));
1213 if (VA.getLocInfo() != CCValAssign::Full)
1214 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1216 // Handle MMX values passed in GPRs.
1217 if (Is64Bit && RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1218 MVT::getSizeInBits(RegVT) == 64)
1219 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1221 ArgValues.push_back(ArgValue);
1223 assert(VA.isMemLoc());
1224 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1228 unsigned StackSize = CCInfo.getNextStackOffset();
1229 // align stack specially for tail calls
1230 if (CC == CallingConv::Fast)
1231 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1233 // If the function takes variable number of arguments, make a frame index for
1234 // the start of the first vararg value... for expansion of llvm.va_start.
1236 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1237 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1240 static const unsigned GPR64ArgRegs[] = {
1241 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1243 static const unsigned XMMArgRegs[] = {
1244 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1245 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1248 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1249 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1251 // For X86-64, if there are vararg parameters that are passed via
1252 // registers, then we must store them to their spots on the stack so they
1253 // may be loaded by deferencing the result of va_next.
1254 VarArgsGPOffset = NumIntRegs * 8;
1255 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1256 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1258 // Store the integer parameter registers.
1259 SmallVector<SDOperand, 8> MemOps;
1260 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1261 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1262 DAG.getIntPtrConstant(VarArgsGPOffset));
1263 for (; NumIntRegs != 6; ++NumIntRegs) {
1264 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1265 X86::GR64RegisterClass);
1266 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1268 DAG.getStore(Val.getValue(1), Val, FIN,
1269 PseudoSourceValue::getFixedStack(),
1271 MemOps.push_back(Store);
1272 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1273 DAG.getIntPtrConstant(8));
1276 // Now store the XMM (fp + vector) parameter registers.
1277 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1278 DAG.getIntPtrConstant(VarArgsFPOffset));
1279 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1280 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1281 X86::VR128RegisterClass);
1282 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1284 DAG.getStore(Val.getValue(1), Val, FIN,
1285 PseudoSourceValue::getFixedStack(),
1287 MemOps.push_back(Store);
1288 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1289 DAG.getIntPtrConstant(16));
1291 if (!MemOps.empty())
1292 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1293 &MemOps[0], MemOps.size());
1297 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1298 // arguments and the arguments after the retaddr has been pushed are
1300 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1301 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1302 (StackSize & 7) == 0)
1305 ArgValues.push_back(Root);
1307 // Some CCs need callee pop.
1308 if (IsCalleePop(Op)) {
1309 BytesToPopOnReturn = StackSize; // Callee pops everything.
1310 BytesCallerReserves = 0;
1312 BytesToPopOnReturn = 0; // Callee pops nothing.
1313 // If this is an sret function, the return should pop the hidden pointer.
1314 if (!Is64Bit && ArgsAreStructReturn(Op))
1315 BytesToPopOnReturn = 4;
1316 BytesCallerReserves = StackSize;
1320 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1321 if (CC == CallingConv::X86_FastCall)
1322 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1325 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1327 // Return the new list of results.
1328 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1329 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1333 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1334 const SDOperand &StackPtr,
1335 const CCValAssign &VA,
1338 unsigned LocMemOffset = VA.getLocMemOffset();
1339 SDOperand PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1340 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1341 ISD::ArgFlagsTy Flags =
1342 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1343 if (Flags.isByVal()) {
1344 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1346 return DAG.getStore(Chain, Arg, PtrOff,
1347 PseudoSourceValue::getStack(), LocMemOffset);
1351 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
1352 MachineFunction &MF = DAG.getMachineFunction();
1353 MachineFrameInfo * MFI = MF.getFrameInfo();
1354 SDOperand Chain = Op.getOperand(0);
1355 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1356 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1357 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1358 && CC == CallingConv::Fast && PerformTailCallOpt;
1359 SDOperand Callee = Op.getOperand(4);
1360 bool Is64Bit = Subtarget->is64Bit();
1361 bool IsStructRet = CallIsStructReturn(Op);
1363 assert(!(isVarArg && CC == CallingConv::Fast) &&
1364 "Var args not supported with calling convention fastcc");
1366 // Analyze operands of the call, assigning locations to each operand.
1367 SmallVector<CCValAssign, 16> ArgLocs;
1368 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1369 CCInfo.AnalyzeCallOperands(Op.Val, CCAssignFnForNode(Op));
1371 // Get a count of how many bytes are to be pushed on the stack.
1372 unsigned NumBytes = CCInfo.getNextStackOffset();
1373 if (CC == CallingConv::Fast)
1374 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1376 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1377 // arguments and the arguments after the retaddr has been pushed are aligned.
1378 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1379 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1380 (NumBytes & 7) == 0)
1385 // Lower arguments at fp - stackoffset + fpdiff.
1386 unsigned NumBytesCallerPushed =
1387 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1388 FPDiff = NumBytesCallerPushed - NumBytes;
1390 // Set the delta of movement of the returnaddr stackslot.
1391 // But only set if delta is greater than previous delta.
1392 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1393 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1396 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1398 SDOperand RetAddrFrIdx;
1400 // Adjust the Return address stack slot.
1402 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1403 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1404 // Load the "old" Return address.
1406 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1407 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1411 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1412 SmallVector<std::pair<unsigned, SDOperand>, 8> TailCallClobberedVRegs;
1413 SmallVector<SDOperand, 8> MemOpChains;
1417 // Walk the register/memloc assignments, inserting copies/loads. For tail
1418 // calls, remember all arguments for later special lowering.
1419 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1420 CCValAssign &VA = ArgLocs[i];
1421 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1423 // Promote the value if needed.
1424 switch (VA.getLocInfo()) {
1425 default: assert(0 && "Unknown loc info!");
1426 case CCValAssign::Full: break;
1427 case CCValAssign::SExt:
1428 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1430 case CCValAssign::ZExt:
1431 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1433 case CCValAssign::AExt:
1434 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1438 if (VA.isRegLoc()) {
1439 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1442 assert(VA.isMemLoc());
1443 if (StackPtr.Val == 0)
1444 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1446 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1448 } else if (IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
1449 TailCallClobberedVRegs.push_back(std::make_pair(i,Arg));
1454 if (!MemOpChains.empty())
1455 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1456 &MemOpChains[0], MemOpChains.size());
1458 // Build a sequence of copy-to-reg nodes chained together with token chain
1459 // and flag operands which copy the outgoing args into registers.
1461 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1462 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1464 InFlag = Chain.getValue(1);
1467 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1469 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1470 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1471 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1473 InFlag = Chain.getValue(1);
1475 // If we are tail calling and generating PIC/GOT style code load the address
1476 // of the callee into ecx. The value in ecx is used as target of the tail
1477 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1478 // calls on PIC/GOT architectures. Normally we would just put the address of
1479 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1480 // restored (since ebx is callee saved) before jumping to the target@PLT.
1481 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1482 // Note: The actual moving to ecx is done further down.
1483 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1484 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1485 !G->getGlobal()->hasProtectedVisibility())
1486 Callee = LowerGlobalAddress(Callee, DAG);
1487 else if (isa<ExternalSymbolSDNode>(Callee))
1488 Callee = LowerExternalSymbol(Callee,DAG);
1491 if (Is64Bit && isVarArg) {
1492 // From AMD64 ABI document:
1493 // For calls that may call functions that use varargs or stdargs
1494 // (prototype-less calls or calls to functions containing ellipsis (...) in
1495 // the declaration) %al is used as hidden argument to specify the number
1496 // of SSE registers used. The contents of %al do not need to match exactly
1497 // the number of registers, but must be an ubound on the number of SSE
1498 // registers used and is in the range 0 - 8 inclusive.
1500 // Count the number of XMM registers allocated.
1501 static const unsigned XMMArgRegs[] = {
1502 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1503 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1505 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1507 Chain = DAG.getCopyToReg(Chain, X86::AL,
1508 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1509 InFlag = Chain.getValue(1);
1513 // For tail calls lower the arguments to the 'real' stack slot.
1515 SmallVector<SDOperand, 8> MemOpChains2;
1518 // Do not flag preceeding copytoreg stuff together with the following stuff.
1519 InFlag = SDOperand();
1521 Chain = CopyTailCallClobberedArgumentsToVRegs(Chain, TailCallClobberedVRegs,
1524 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1525 CCValAssign &VA = ArgLocs[i];
1526 if (!VA.isRegLoc()) {
1527 assert(VA.isMemLoc());
1528 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1529 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1530 ISD::ArgFlagsTy Flags =
1531 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1532 // Create frame index.
1533 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1534 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1535 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1536 FIN = DAG.getFrameIndex(FI, MVT::i32);
1538 // Find virtual register for this argument.
1540 for (unsigned idx=0, e= TailCallClobberedVRegs.size(); idx < e; idx++)
1541 if (TailCallClobberedVRegs[idx].first==i) {
1542 Arg = TailCallClobberedVRegs[idx].second;
1546 assert(IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)==false ||
1547 (Found==true && "No corresponding Argument was found"));
1549 if (Flags.isByVal()) {
1550 // Copy relative to framepointer.
1551 MemOpChains2.push_back(CreateCopyOfByValArgument(Arg, FIN, Chain,
1554 // Store relative to framepointer.
1555 MemOpChains2.push_back(
1556 DAG.getStore(Chain, Arg, FIN,
1557 PseudoSourceValue::getFixedStack(), FI));
1562 if (!MemOpChains2.empty())
1563 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1564 &MemOpChains2[0], MemOpChains2.size());
1566 // Store the return address to the appropriate stack slot.
1568 // Calculate the new stack slot for the return address.
1569 int SlotSize = Is64Bit ? 8 : 4;
1570 int NewReturnAddrFI =
1571 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1572 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1573 SDOperand NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1574 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1575 PseudoSourceValue::getFixedStack(), NewReturnAddrFI);
1579 // If the callee is a GlobalAddress node (quite common, every direct call is)
1580 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1581 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1582 // We should use extra load for direct calls to dllimported functions in
1584 if ((IsTailCall || !Is64Bit ||
1585 getTargetMachine().getCodeModel() != CodeModel::Large)
1586 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1587 getTargetMachine(), true))
1588 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1589 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1590 if (IsTailCall || !Is64Bit ||
1591 getTargetMachine().getCodeModel() != CodeModel::Large)
1592 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1593 } else if (IsTailCall) {
1594 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1596 Chain = DAG.getCopyToReg(Chain,
1597 DAG.getRegister(Opc, getPointerTy()),
1599 Callee = DAG.getRegister(Opc, getPointerTy());
1600 // Add register as live out.
1601 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1604 // Returns a chain & a flag for retval copy to use.
1605 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1606 SmallVector<SDOperand, 8> Ops;
1609 Ops.push_back(Chain);
1610 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1611 Ops.push_back(DAG.getIntPtrConstant(0));
1613 Ops.push_back(InFlag);
1614 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1615 InFlag = Chain.getValue(1);
1617 // Returns a chain & a flag for retval copy to use.
1618 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1622 Ops.push_back(Chain);
1623 Ops.push_back(Callee);
1626 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1628 // Add argument registers to the end of the list so that they are known live
1630 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1631 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1632 RegsToPass[i].second.getValueType()));
1634 // Add an implicit use GOT pointer in EBX.
1635 if (!IsTailCall && !Is64Bit &&
1636 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1637 Subtarget->isPICStyleGOT())
1638 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1640 // Add an implicit use of AL for x86 vararg functions.
1641 if (Is64Bit && isVarArg)
1642 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1645 Ops.push_back(InFlag);
1648 assert(InFlag.Val &&
1649 "Flag must be set. Depend on flag being set in LowerRET");
1650 Chain = DAG.getNode(X86ISD::TAILCALL,
1651 Op.Val->getVTList(), &Ops[0], Ops.size());
1653 return SDOperand(Chain.Val, Op.ResNo);
1656 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1657 InFlag = Chain.getValue(1);
1659 // Create the CALLSEQ_END node.
1660 unsigned NumBytesForCalleeToPush;
1661 if (IsCalleePop(Op))
1662 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1663 else if (!Is64Bit && IsStructRet)
1664 // If this is is a call to a struct-return function, the callee
1665 // pops the hidden struct pointer, so we have to push it back.
1666 // This is common for Darwin/X86, Linux & Mingw32 targets.
1667 NumBytesForCalleeToPush = 4;
1669 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1671 // Returns a flag for retval copy to use.
1672 Chain = DAG.getCALLSEQ_END(Chain,
1673 DAG.getIntPtrConstant(NumBytes),
1674 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1676 InFlag = Chain.getValue(1);
1678 // Handle result values, copying them out of physregs into vregs that we
1680 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1684 //===----------------------------------------------------------------------===//
1685 // Fast Calling Convention (tail call) implementation
1686 //===----------------------------------------------------------------------===//
1688 // Like std call, callee cleans arguments, convention except that ECX is
1689 // reserved for storing the tail called function address. Only 2 registers are
1690 // free for argument passing (inreg). Tail call optimization is performed
1692 // * tailcallopt is enabled
1693 // * caller/callee are fastcc
1694 // On X86_64 architecture with GOT-style position independent code only local
1695 // (within module) calls are supported at the moment.
1696 // To keep the stack aligned according to platform abi the function
1697 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1698 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1699 // If a tail called function callee has more arguments than the caller the
1700 // caller needs to make sure that there is room to move the RETADDR to. This is
1701 // achieved by reserving an area the size of the argument delta right after the
1702 // original REtADDR, but before the saved framepointer or the spilled registers
1703 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1715 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1716 /// for a 16 byte align requirement.
1717 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1718 SelectionDAG& DAG) {
1719 if (PerformTailCallOpt) {
1720 MachineFunction &MF = DAG.getMachineFunction();
1721 const TargetMachine &TM = MF.getTarget();
1722 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1723 unsigned StackAlignment = TFI.getStackAlignment();
1724 uint64_t AlignMask = StackAlignment - 1;
1725 int64_t Offset = StackSize;
1726 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1727 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1728 // Number smaller than 12 so just add the difference.
1729 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1731 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1732 Offset = ((~AlignMask) & Offset) + StackAlignment +
1733 (StackAlignment-SlotSize);
1740 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1741 /// following the call is a return. A function is eligible if caller/callee
1742 /// calling conventions match, currently only fastcc supports tail calls, and
1743 /// the function CALL is immediatly followed by a RET.
1744 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1746 SelectionDAG& DAG) const {
1747 if (!PerformTailCallOpt)
1750 // Check whether CALL node immediatly preceeds the RET node and whether the
1751 // return uses the result of the node or is a void return.
1752 unsigned NumOps = Ret.getNumOperands();
1754 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1755 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1757 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1758 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1759 MachineFunction &MF = DAG.getMachineFunction();
1760 unsigned CallerCC = MF.getFunction()->getCallingConv();
1761 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1762 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1763 SDOperand Callee = Call.getOperand(4);
1764 // On x86/32Bit PIC/GOT tail calls are supported.
1765 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1766 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1769 // Can only do local tail calls (in same module, hidden or protected) on
1770 // x86_64 PIC/GOT at the moment.
1771 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1772 return G->getGlobal()->hasHiddenVisibility()
1773 || G->getGlobal()->hasProtectedVisibility();
1780 //===----------------------------------------------------------------------===//
1781 // Other Lowering Hooks
1782 //===----------------------------------------------------------------------===//
1785 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1786 MachineFunction &MF = DAG.getMachineFunction();
1787 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1788 int ReturnAddrIndex = FuncInfo->getRAIndex();
1790 if (ReturnAddrIndex == 0) {
1791 // Set up a frame object for the return address.
1792 if (Subtarget->is64Bit())
1793 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1795 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1797 FuncInfo->setRAIndex(ReturnAddrIndex);
1800 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1805 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1806 /// specific condition code. It returns a false if it cannot do a direct
1807 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1809 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1810 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1811 SelectionDAG &DAG) {
1812 X86CC = X86::COND_INVALID;
1814 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1815 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1816 // X > -1 -> X == 0, jump !sign.
1817 RHS = DAG.getConstant(0, RHS.getValueType());
1818 X86CC = X86::COND_NS;
1820 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1821 // X < 0 -> X == 0, jump on sign.
1822 X86CC = X86::COND_S;
1824 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1826 RHS = DAG.getConstant(0, RHS.getValueType());
1827 X86CC = X86::COND_LE;
1832 switch (SetCCOpcode) {
1834 case ISD::SETEQ: X86CC = X86::COND_E; break;
1835 case ISD::SETGT: X86CC = X86::COND_G; break;
1836 case ISD::SETGE: X86CC = X86::COND_GE; break;
1837 case ISD::SETLT: X86CC = X86::COND_L; break;
1838 case ISD::SETLE: X86CC = X86::COND_LE; break;
1839 case ISD::SETNE: X86CC = X86::COND_NE; break;
1840 case ISD::SETULT: X86CC = X86::COND_B; break;
1841 case ISD::SETUGT: X86CC = X86::COND_A; break;
1842 case ISD::SETULE: X86CC = X86::COND_BE; break;
1843 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1846 // On a floating point condition, the flags are set as follows:
1848 // 0 | 0 | 0 | X > Y
1849 // 0 | 0 | 1 | X < Y
1850 // 1 | 0 | 0 | X == Y
1851 // 1 | 1 | 1 | unordered
1853 switch (SetCCOpcode) {
1856 case ISD::SETEQ: X86CC = X86::COND_E; break;
1857 case ISD::SETOLT: Flip = true; // Fallthrough
1859 case ISD::SETGT: X86CC = X86::COND_A; break;
1860 case ISD::SETOLE: Flip = true; // Fallthrough
1862 case ISD::SETGE: X86CC = X86::COND_AE; break;
1863 case ISD::SETUGT: Flip = true; // Fallthrough
1865 case ISD::SETLT: X86CC = X86::COND_B; break;
1866 case ISD::SETUGE: Flip = true; // Fallthrough
1868 case ISD::SETLE: X86CC = X86::COND_BE; break;
1870 case ISD::SETNE: X86CC = X86::COND_NE; break;
1871 case ISD::SETUO: X86CC = X86::COND_P; break;
1872 case ISD::SETO: X86CC = X86::COND_NP; break;
1875 std::swap(LHS, RHS);
1878 return X86CC != X86::COND_INVALID;
1881 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1882 /// code. Current x86 isa includes the following FP cmov instructions:
1883 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1884 static bool hasFPCMov(unsigned X86CC) {
1900 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1901 /// true if Op is undef or if its value falls within the specified range (L, H].
1902 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1903 if (Op.getOpcode() == ISD::UNDEF)
1906 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1907 return (Val >= Low && Val < Hi);
1910 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1911 /// true if Op is undef or if its value equal to the specified value.
1912 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1913 if (Op.getOpcode() == ISD::UNDEF)
1915 return cast<ConstantSDNode>(Op)->getValue() == Val;
1918 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1919 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1920 bool X86::isPSHUFDMask(SDNode *N) {
1921 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1923 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
1926 // Check if the value doesn't reference the second vector.
1927 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1928 SDOperand Arg = N->getOperand(i);
1929 if (Arg.getOpcode() == ISD::UNDEF) continue;
1930 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1931 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
1938 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1939 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1940 bool X86::isPSHUFHWMask(SDNode *N) {
1941 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1943 if (N->getNumOperands() != 8)
1946 // Lower quadword copied in order.
1947 for (unsigned i = 0; i != 4; ++i) {
1948 SDOperand Arg = N->getOperand(i);
1949 if (Arg.getOpcode() == ISD::UNDEF) continue;
1950 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1951 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1955 // Upper quadword shuffled.
1956 for (unsigned i = 4; i != 8; ++i) {
1957 SDOperand Arg = N->getOperand(i);
1958 if (Arg.getOpcode() == ISD::UNDEF) continue;
1959 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1960 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1961 if (Val < 4 || Val > 7)
1968 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1969 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1970 bool X86::isPSHUFLWMask(SDNode *N) {
1971 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1973 if (N->getNumOperands() != 8)
1976 // Upper quadword copied in order.
1977 for (unsigned i = 4; i != 8; ++i)
1978 if (!isUndefOrEqual(N->getOperand(i), i))
1981 // Lower quadword shuffled.
1982 for (unsigned i = 0; i != 4; ++i)
1983 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1989 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1990 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1991 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1992 if (NumElems != 2 && NumElems != 4) return false;
1994 unsigned Half = NumElems / 2;
1995 for (unsigned i = 0; i < Half; ++i)
1996 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1998 for (unsigned i = Half; i < NumElems; ++i)
1999 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2005 bool X86::isSHUFPMask(SDNode *N) {
2006 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2007 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2010 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2011 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2012 /// half elements to come from vector 1 (which would equal the dest.) and
2013 /// the upper half to come from vector 2.
2014 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2015 if (NumOps != 2 && NumOps != 4) return false;
2017 unsigned Half = NumOps / 2;
2018 for (unsigned i = 0; i < Half; ++i)
2019 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2021 for (unsigned i = Half; i < NumOps; ++i)
2022 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2027 static bool isCommutedSHUFP(SDNode *N) {
2028 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2029 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2032 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2033 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2034 bool X86::isMOVHLPSMask(SDNode *N) {
2035 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2037 if (N->getNumOperands() != 4)
2040 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2041 return isUndefOrEqual(N->getOperand(0), 6) &&
2042 isUndefOrEqual(N->getOperand(1), 7) &&
2043 isUndefOrEqual(N->getOperand(2), 2) &&
2044 isUndefOrEqual(N->getOperand(3), 3);
2047 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2048 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2050 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2051 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2053 if (N->getNumOperands() != 4)
2056 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2057 return isUndefOrEqual(N->getOperand(0), 2) &&
2058 isUndefOrEqual(N->getOperand(1), 3) &&
2059 isUndefOrEqual(N->getOperand(2), 2) &&
2060 isUndefOrEqual(N->getOperand(3), 3);
2063 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2064 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2065 bool X86::isMOVLPMask(SDNode *N) {
2066 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2068 unsigned NumElems = N->getNumOperands();
2069 if (NumElems != 2 && NumElems != 4)
2072 for (unsigned i = 0; i < NumElems/2; ++i)
2073 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2076 for (unsigned i = NumElems/2; i < NumElems; ++i)
2077 if (!isUndefOrEqual(N->getOperand(i), i))
2083 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2084 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2086 bool X86::isMOVHPMask(SDNode *N) {
2087 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2089 unsigned NumElems = N->getNumOperands();
2090 if (NumElems != 2 && NumElems != 4)
2093 for (unsigned i = 0; i < NumElems/2; ++i)
2094 if (!isUndefOrEqual(N->getOperand(i), i))
2097 for (unsigned i = 0; i < NumElems/2; ++i) {
2098 SDOperand Arg = N->getOperand(i + NumElems/2);
2099 if (!isUndefOrEqual(Arg, i + NumElems))
2106 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2107 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2108 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2109 bool V2IsSplat = false) {
2110 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2113 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2114 SDOperand BitI = Elts[i];
2115 SDOperand BitI1 = Elts[i+1];
2116 if (!isUndefOrEqual(BitI, j))
2119 if (isUndefOrEqual(BitI1, NumElts))
2122 if (!isUndefOrEqual(BitI1, j + NumElts))
2130 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2131 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2132 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2135 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2136 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2137 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2138 bool V2IsSplat = false) {
2139 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2142 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2143 SDOperand BitI = Elts[i];
2144 SDOperand BitI1 = Elts[i+1];
2145 if (!isUndefOrEqual(BitI, j + NumElts/2))
2148 if (isUndefOrEqual(BitI1, NumElts))
2151 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2159 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2160 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2161 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2164 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2165 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2167 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2168 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2170 unsigned NumElems = N->getNumOperands();
2171 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2174 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2175 SDOperand BitI = N->getOperand(i);
2176 SDOperand BitI1 = N->getOperand(i+1);
2178 if (!isUndefOrEqual(BitI, j))
2180 if (!isUndefOrEqual(BitI1, j))
2187 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2188 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2190 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2191 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2193 unsigned NumElems = N->getNumOperands();
2194 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2197 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2198 SDOperand BitI = N->getOperand(i);
2199 SDOperand BitI1 = N->getOperand(i + 1);
2201 if (!isUndefOrEqual(BitI, j))
2203 if (!isUndefOrEqual(BitI1, j))
2210 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2211 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2212 /// MOVSD, and MOVD, i.e. setting the lowest element.
2213 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2214 if (NumElts != 2 && NumElts != 4)
2217 if (!isUndefOrEqual(Elts[0], NumElts))
2220 for (unsigned i = 1; i < NumElts; ++i) {
2221 if (!isUndefOrEqual(Elts[i], i))
2228 bool X86::isMOVLMask(SDNode *N) {
2229 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2230 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2233 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2234 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2235 /// element of vector 2 and the other elements to come from vector 1 in order.
2236 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2237 bool V2IsSplat = false,
2238 bool V2IsUndef = false) {
2239 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2242 if (!isUndefOrEqual(Ops[0], 0))
2245 for (unsigned i = 1; i < NumOps; ++i) {
2246 SDOperand Arg = Ops[i];
2247 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2248 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2249 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2256 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2257 bool V2IsUndef = false) {
2258 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2259 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2260 V2IsSplat, V2IsUndef);
2263 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2264 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2265 bool X86::isMOVSHDUPMask(SDNode *N) {
2266 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2268 if (N->getNumOperands() != 4)
2271 // Expect 1, 1, 3, 3
2272 for (unsigned i = 0; i < 2; ++i) {
2273 SDOperand Arg = N->getOperand(i);
2274 if (Arg.getOpcode() == ISD::UNDEF) continue;
2275 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2276 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2277 if (Val != 1) return false;
2281 for (unsigned i = 2; i < 4; ++i) {
2282 SDOperand Arg = N->getOperand(i);
2283 if (Arg.getOpcode() == ISD::UNDEF) continue;
2284 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2285 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2286 if (Val != 3) return false;
2290 // Don't use movshdup if it can be done with a shufps.
2294 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2295 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2296 bool X86::isMOVSLDUPMask(SDNode *N) {
2297 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2299 if (N->getNumOperands() != 4)
2302 // Expect 0, 0, 2, 2
2303 for (unsigned i = 0; i < 2; ++i) {
2304 SDOperand Arg = N->getOperand(i);
2305 if (Arg.getOpcode() == ISD::UNDEF) continue;
2306 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2307 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2308 if (Val != 0) return false;
2312 for (unsigned i = 2; i < 4; ++i) {
2313 SDOperand Arg = N->getOperand(i);
2314 if (Arg.getOpcode() == ISD::UNDEF) continue;
2315 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2316 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2317 if (Val != 2) return false;
2321 // Don't use movshdup if it can be done with a shufps.
2325 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2326 /// specifies a identity operation on the LHS or RHS.
2327 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2328 unsigned NumElems = N->getNumOperands();
2329 for (unsigned i = 0; i < NumElems; ++i)
2330 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2335 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2336 /// a splat of a single element.
2337 static bool isSplatMask(SDNode *N) {
2338 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2340 // This is a splat operation if each element of the permute is the same, and
2341 // if the value doesn't reference the second vector.
2342 unsigned NumElems = N->getNumOperands();
2343 SDOperand ElementBase;
2345 for (; i != NumElems; ++i) {
2346 SDOperand Elt = N->getOperand(i);
2347 if (isa<ConstantSDNode>(Elt)) {
2353 if (!ElementBase.Val)
2356 for (; i != NumElems; ++i) {
2357 SDOperand Arg = N->getOperand(i);
2358 if (Arg.getOpcode() == ISD::UNDEF) continue;
2359 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2360 if (Arg != ElementBase) return false;
2363 // Make sure it is a splat of the first vector operand.
2364 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2367 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2368 /// a splat of a single element and it's a 2 or 4 element mask.
2369 bool X86::isSplatMask(SDNode *N) {
2370 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2372 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2373 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2375 return ::isSplatMask(N);
2378 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2379 /// specifies a splat of zero element.
2380 bool X86::isSplatLoMask(SDNode *N) {
2381 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2383 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2384 if (!isUndefOrEqual(N->getOperand(i), 0))
2389 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2390 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2392 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2393 unsigned NumOperands = N->getNumOperands();
2394 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2396 for (unsigned i = 0; i < NumOperands; ++i) {
2398 SDOperand Arg = N->getOperand(NumOperands-i-1);
2399 if (Arg.getOpcode() != ISD::UNDEF)
2400 Val = cast<ConstantSDNode>(Arg)->getValue();
2401 if (Val >= NumOperands) Val -= NumOperands;
2403 if (i != NumOperands - 1)
2410 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2411 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2413 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2415 // 8 nodes, but we only care about the last 4.
2416 for (unsigned i = 7; i >= 4; --i) {
2418 SDOperand Arg = N->getOperand(i);
2419 if (Arg.getOpcode() != ISD::UNDEF)
2420 Val = cast<ConstantSDNode>(Arg)->getValue();
2429 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2430 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2432 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2434 // 8 nodes, but we only care about the first 4.
2435 for (int i = 3; i >= 0; --i) {
2437 SDOperand Arg = N->getOperand(i);
2438 if (Arg.getOpcode() != ISD::UNDEF)
2439 Val = cast<ConstantSDNode>(Arg)->getValue();
2448 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2449 /// specifies a 8 element shuffle that can be broken into a pair of
2450 /// PSHUFHW and PSHUFLW.
2451 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2452 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2454 if (N->getNumOperands() != 8)
2457 // Lower quadword shuffled.
2458 for (unsigned i = 0; i != 4; ++i) {
2459 SDOperand Arg = N->getOperand(i);
2460 if (Arg.getOpcode() == ISD::UNDEF) continue;
2461 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2462 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2467 // Upper quadword shuffled.
2468 for (unsigned i = 4; i != 8; ++i) {
2469 SDOperand Arg = N->getOperand(i);
2470 if (Arg.getOpcode() == ISD::UNDEF) continue;
2471 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2472 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2473 if (Val < 4 || Val > 7)
2480 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2481 /// values in ther permute mask.
2482 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2483 SDOperand &V2, SDOperand &Mask,
2484 SelectionDAG &DAG) {
2485 MVT::ValueType VT = Op.getValueType();
2486 MVT::ValueType MaskVT = Mask.getValueType();
2487 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2488 unsigned NumElems = Mask.getNumOperands();
2489 SmallVector<SDOperand, 8> MaskVec;
2491 for (unsigned i = 0; i != NumElems; ++i) {
2492 SDOperand Arg = Mask.getOperand(i);
2493 if (Arg.getOpcode() == ISD::UNDEF) {
2494 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2497 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2498 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2500 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2502 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2506 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2507 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2510 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2511 /// the two vector operands have swapped position.
2513 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2514 MVT::ValueType MaskVT = Mask.getValueType();
2515 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2516 unsigned NumElems = Mask.getNumOperands();
2517 SmallVector<SDOperand, 8> MaskVec;
2518 for (unsigned i = 0; i != NumElems; ++i) {
2519 SDOperand Arg = Mask.getOperand(i);
2520 if (Arg.getOpcode() == ISD::UNDEF) {
2521 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2524 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2525 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2527 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2529 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2531 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2535 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2536 /// match movhlps. The lower half elements should come from upper half of
2537 /// V1 (and in order), and the upper half elements should come from the upper
2538 /// half of V2 (and in order).
2539 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2540 unsigned NumElems = Mask->getNumOperands();
2543 for (unsigned i = 0, e = 2; i != e; ++i)
2544 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2546 for (unsigned i = 2; i != 4; ++i)
2547 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2552 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2553 /// is promoted to a vector.
2554 static inline bool isScalarLoadToVector(SDNode *N) {
2555 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2556 N = N->getOperand(0).Val;
2557 return ISD::isNON_EXTLoad(N);
2562 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2563 /// match movlp{s|d}. The lower half elements should come from lower half of
2564 /// V1 (and in order), and the upper half elements should come from the upper
2565 /// half of V2 (and in order). And since V1 will become the source of the
2566 /// MOVLP, it must be either a vector load or a scalar load to vector.
2567 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2568 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2570 // Is V2 is a vector load, don't do this transformation. We will try to use
2571 // load folding shufps op.
2572 if (ISD::isNON_EXTLoad(V2))
2575 unsigned NumElems = Mask->getNumOperands();
2576 if (NumElems != 2 && NumElems != 4)
2578 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2579 if (!isUndefOrEqual(Mask->getOperand(i), i))
2581 for (unsigned i = NumElems/2; i != NumElems; ++i)
2582 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2587 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2589 static bool isSplatVector(SDNode *N) {
2590 if (N->getOpcode() != ISD::BUILD_VECTOR)
2593 SDOperand SplatValue = N->getOperand(0);
2594 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2595 if (N->getOperand(i) != SplatValue)
2600 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2602 static bool isUndefShuffle(SDNode *N) {
2603 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2606 SDOperand V1 = N->getOperand(0);
2607 SDOperand V2 = N->getOperand(1);
2608 SDOperand Mask = N->getOperand(2);
2609 unsigned NumElems = Mask.getNumOperands();
2610 for (unsigned i = 0; i != NumElems; ++i) {
2611 SDOperand Arg = Mask.getOperand(i);
2612 if (Arg.getOpcode() != ISD::UNDEF) {
2613 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2614 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2616 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2623 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2625 static inline bool isZeroNode(SDOperand Elt) {
2626 return ((isa<ConstantSDNode>(Elt) &&
2627 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2628 (isa<ConstantFPSDNode>(Elt) &&
2629 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2632 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2633 /// to an zero vector.
2634 static bool isZeroShuffle(SDNode *N) {
2635 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2638 SDOperand V1 = N->getOperand(0);
2639 SDOperand V2 = N->getOperand(1);
2640 SDOperand Mask = N->getOperand(2);
2641 unsigned NumElems = Mask.getNumOperands();
2642 for (unsigned i = 0; i != NumElems; ++i) {
2643 SDOperand Arg = Mask.getOperand(i);
2644 if (Arg.getOpcode() == ISD::UNDEF)
2647 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2648 if (Idx < NumElems) {
2649 unsigned Opc = V1.Val->getOpcode();
2650 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2652 if (Opc != ISD::BUILD_VECTOR ||
2653 !isZeroNode(V1.Val->getOperand(Idx)))
2655 } else if (Idx >= NumElems) {
2656 unsigned Opc = V2.Val->getOpcode();
2657 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2659 if (Opc != ISD::BUILD_VECTOR ||
2660 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2667 /// getZeroVector - Returns a vector of specified type with all zero elements.
2669 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2670 assert(MVT::isVector(VT) && "Expected a vector type");
2672 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2673 // type. This ensures they get CSE'd.
2674 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2676 if (MVT::getSizeInBits(VT) == 64) // MMX
2677 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2679 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2680 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2683 /// getOnesVector - Returns a vector of specified type with all bits set.
2685 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2686 assert(MVT::isVector(VT) && "Expected a vector type");
2688 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2689 // type. This ensures they get CSE'd.
2690 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2692 if (MVT::getSizeInBits(VT) == 64) // MMX
2693 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2695 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2696 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2700 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2701 /// that point to V2 points to its first element.
2702 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2703 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2705 bool Changed = false;
2706 SmallVector<SDOperand, 8> MaskVec;
2707 unsigned NumElems = Mask.getNumOperands();
2708 for (unsigned i = 0; i != NumElems; ++i) {
2709 SDOperand Arg = Mask.getOperand(i);
2710 if (Arg.getOpcode() != ISD::UNDEF) {
2711 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2712 if (Val > NumElems) {
2713 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2717 MaskVec.push_back(Arg);
2721 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2722 &MaskVec[0], MaskVec.size());
2726 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2727 /// operation of specified width.
2728 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2729 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2730 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2732 SmallVector<SDOperand, 8> MaskVec;
2733 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2734 for (unsigned i = 1; i != NumElems; ++i)
2735 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2736 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2739 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2740 /// of specified width.
2741 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2742 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2743 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2744 SmallVector<SDOperand, 8> MaskVec;
2745 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2746 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2747 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2749 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2752 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2753 /// of specified width.
2754 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2755 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2756 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2757 unsigned Half = NumElems/2;
2758 SmallVector<SDOperand, 8> MaskVec;
2759 for (unsigned i = 0; i != Half; ++i) {
2760 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2761 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2763 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2766 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2767 /// element #0 of a vector with the specified index, leaving the rest of the
2768 /// elements in place.
2769 static SDOperand getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2770 SelectionDAG &DAG) {
2771 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2772 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2773 SmallVector<SDOperand, 8> MaskVec;
2774 // Element #0 of the result gets the elt we are replacing.
2775 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2776 for (unsigned i = 1; i != NumElems; ++i)
2777 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2778 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2781 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2783 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2784 SDOperand V1 = Op.getOperand(0);
2785 SDOperand Mask = Op.getOperand(2);
2786 MVT::ValueType VT = Op.getValueType();
2787 unsigned NumElems = Mask.getNumOperands();
2788 Mask = getUnpacklMask(NumElems, DAG);
2789 while (NumElems != 4) {
2790 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2793 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2795 Mask = getZeroVector(MVT::v4i32, DAG);
2796 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2797 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2798 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2801 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2802 /// vector of zero or undef vector. This produces a shuffle where the low
2803 /// element of V2 is swizzled into the zero/undef vector, landing at element
2804 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2805 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, unsigned Idx,
2806 bool isZero, SelectionDAG &DAG) {
2807 MVT::ValueType VT = V2.getValueType();
2808 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2809 unsigned NumElems = MVT::getVectorNumElements(V2.getValueType());
2810 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2811 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2812 SmallVector<SDOperand, 16> MaskVec;
2813 for (unsigned i = 0; i != NumElems; ++i)
2814 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2815 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2817 MaskVec.push_back(DAG.getConstant(i, EVT));
2818 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2819 &MaskVec[0], MaskVec.size());
2820 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2823 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2825 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2826 unsigned NumNonZero, unsigned NumZero,
2827 SelectionDAG &DAG, TargetLowering &TLI) {
2833 for (unsigned i = 0; i < 16; ++i) {
2834 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2835 if (ThisIsNonZero && First) {
2837 V = getZeroVector(MVT::v8i16, DAG);
2839 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2844 SDOperand ThisElt(0, 0), LastElt(0, 0);
2845 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2846 if (LastIsNonZero) {
2847 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2849 if (ThisIsNonZero) {
2850 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2851 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2852 ThisElt, DAG.getConstant(8, MVT::i8));
2854 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2859 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2860 DAG.getIntPtrConstant(i/2));
2864 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2867 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2869 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2870 unsigned NumNonZero, unsigned NumZero,
2871 SelectionDAG &DAG, TargetLowering &TLI) {
2877 for (unsigned i = 0; i < 8; ++i) {
2878 bool isNonZero = (NonZeros & (1 << i)) != 0;
2882 V = getZeroVector(MVT::v8i16, DAG);
2884 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2887 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2888 DAG.getIntPtrConstant(i));
2896 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2897 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
2898 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
2899 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
2900 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
2901 // eliminated on x86-32 hosts.
2902 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
2905 if (ISD::isBuildVectorAllOnes(Op.Val))
2906 return getOnesVector(Op.getValueType(), DAG);
2907 return getZeroVector(Op.getValueType(), DAG);
2910 MVT::ValueType VT = Op.getValueType();
2911 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2912 unsigned EVTBits = MVT::getSizeInBits(EVT);
2914 unsigned NumElems = Op.getNumOperands();
2915 unsigned NumZero = 0;
2916 unsigned NumNonZero = 0;
2917 unsigned NonZeros = 0;
2918 bool IsAllConstants = true;
2919 SmallSet<SDOperand, 8> Values;
2920 for (unsigned i = 0; i < NumElems; ++i) {
2921 SDOperand Elt = Op.getOperand(i);
2922 if (Elt.getOpcode() == ISD::UNDEF)
2925 if (Elt.getOpcode() != ISD::Constant &&
2926 Elt.getOpcode() != ISD::ConstantFP)
2927 IsAllConstants = false;
2928 if (isZeroNode(Elt))
2931 NonZeros |= (1 << i);
2936 if (NumNonZero == 0) {
2937 // All undef vector. Return an UNDEF. All zero vectors were handled above.
2938 return DAG.getNode(ISD::UNDEF, VT);
2941 // Special case for single non-zero, non-undef, element.
2942 if (NumNonZero == 1 && NumElems <= 4) {
2943 unsigned Idx = CountTrailingZeros_32(NonZeros);
2944 SDOperand Item = Op.getOperand(Idx);
2946 // If this is an insertion of an i64 value on x86-32, and if the top bits of
2947 // the value are obviously zero, truncate the value to i32 and do the
2948 // insertion that way. Only do this if the value is non-constant or if the
2949 // value is a constant being inserted into element 0. It is cheaper to do
2950 // a constant pool load than it is to do a movd + shuffle.
2951 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
2952 (!IsAllConstants || Idx == 0)) {
2953 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
2954 // Handle MMX and SSE both.
2955 MVT::ValueType VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
2956 MVT::ValueType VecElts = VT == MVT::v2i64 ? 4 : 2;
2958 // Truncate the value (which may itself be a constant) to i32, and
2959 // convert it to a vector with movd (S2V+shuffle to zero extend).
2960 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
2961 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
2962 Item = getShuffleVectorZeroOrUndef(Item, 0, true, DAG);
2964 // Now we have our 32-bit value zero extended in the low element of
2965 // a vector. If Idx != 0, swizzle it into place.
2968 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
2969 getSwapEltZeroMask(VecElts, Idx, DAG)
2971 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
2973 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
2977 // If we have a constant or non-constant insertion into the low element of
2978 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
2979 // the rest of the elements. This will be matched as movd/movq/movss/movsd
2980 // depending on what the source datatype is. Because we can only get here
2981 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
2983 // Don't do this for i64 values on x86-32.
2984 (EVT != MVT::i64 || Subtarget->is64Bit())) {
2985 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2986 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2987 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, DAG);
2990 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
2993 // Otherwise, if this is a vector with i32 or f32 elements, and the element
2994 // is a non-constant being inserted into an element other than the low one,
2995 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
2996 // movd/movss) to move this into the low element, then shuffle it into
2998 if (EVTBits == 32) {
2999 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3001 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3002 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, DAG);
3003 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3004 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3005 SmallVector<SDOperand, 8> MaskVec;
3006 for (unsigned i = 0; i < NumElems; i++)
3007 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3008 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3009 &MaskVec[0], MaskVec.size());
3010 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3011 DAG.getNode(ISD::UNDEF, VT), Mask);
3015 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3016 if (Values.size() == 1)
3019 // A vector full of immediates; various special cases are already
3020 // handled, so this is best done with a single constant-pool load.
3024 // Let legalizer expand 2-wide build_vectors.
3028 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3029 if (EVTBits == 8 && NumElems == 16) {
3030 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3032 if (V.Val) return V;
3035 if (EVTBits == 16 && NumElems == 8) {
3036 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3038 if (V.Val) return V;
3041 // If element VT is == 32 bits, turn it into a number of shuffles.
3042 SmallVector<SDOperand, 8> V;
3044 if (NumElems == 4 && NumZero > 0) {
3045 for (unsigned i = 0; i < 4; ++i) {
3046 bool isZero = !(NonZeros & (1 << i));
3048 V[i] = getZeroVector(VT, DAG);
3050 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3053 for (unsigned i = 0; i < 2; ++i) {
3054 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3057 V[i] = V[i*2]; // Must be a zero vector.
3060 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3061 getMOVLMask(NumElems, DAG));
3064 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3065 getMOVLMask(NumElems, DAG));
3068 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3069 getUnpacklMask(NumElems, DAG));
3074 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3075 // clears the upper bits.
3076 // FIXME: we can do the same for v4f32 case when we know both parts of
3077 // the lower half come from scalar_to_vector (loadf32). We should do
3078 // that in post legalizer dag combiner with target specific hooks.
3079 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3081 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3082 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3083 SmallVector<SDOperand, 8> MaskVec;
3084 bool Reverse = (NonZeros & 0x3) == 2;
3085 for (unsigned i = 0; i < 2; ++i)
3087 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3089 MaskVec.push_back(DAG.getConstant(i, EVT));
3090 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3091 for (unsigned i = 0; i < 2; ++i)
3093 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3095 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3096 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3097 &MaskVec[0], MaskVec.size());
3098 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3101 if (Values.size() > 2) {
3102 // Expand into a number of unpckl*.
3104 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3105 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3106 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3107 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3108 for (unsigned i = 0; i < NumElems; ++i)
3109 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3111 while (NumElems != 0) {
3112 for (unsigned i = 0; i < NumElems; ++i)
3113 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3124 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3125 SDOperand PermMask, SelectionDAG &DAG,
3126 TargetLowering &TLI) {
3128 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3129 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3130 MVT::ValueType PtrVT = TLI.getPointerTy();
3131 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3132 PermMask.Val->op_end());
3134 // First record which half of which vector the low elements come from.
3135 SmallVector<unsigned, 4> LowQuad(4);
3136 for (unsigned i = 0; i < 4; ++i) {
3137 SDOperand Elt = MaskElts[i];
3138 if (Elt.getOpcode() == ISD::UNDEF)
3140 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3141 int QuadIdx = EltIdx / 4;
3144 int BestLowQuad = -1;
3145 unsigned MaxQuad = 1;
3146 for (unsigned i = 0; i < 4; ++i) {
3147 if (LowQuad[i] > MaxQuad) {
3149 MaxQuad = LowQuad[i];
3153 // Record which half of which vector the high elements come from.
3154 SmallVector<unsigned, 4> HighQuad(4);
3155 for (unsigned i = 4; i < 8; ++i) {
3156 SDOperand Elt = MaskElts[i];
3157 if (Elt.getOpcode() == ISD::UNDEF)
3159 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3160 int QuadIdx = EltIdx / 4;
3161 ++HighQuad[QuadIdx];
3163 int BestHighQuad = -1;
3165 for (unsigned i = 0; i < 4; ++i) {
3166 if (HighQuad[i] > MaxQuad) {
3168 MaxQuad = HighQuad[i];
3172 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3173 if (BestLowQuad != -1 || BestHighQuad != -1) {
3174 // First sort the 4 chunks in order using shufpd.
3175 SmallVector<SDOperand, 8> MaskVec;
3176 if (BestLowQuad != -1)
3177 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3179 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3180 if (BestHighQuad != -1)
3181 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3183 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3184 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3185 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3186 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3187 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3188 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3190 // Now sort high and low parts separately.
3191 BitVector InOrder(8);
3192 if (BestLowQuad != -1) {
3193 // Sort lower half in order using PSHUFLW.
3195 bool AnyOutOrder = false;
3196 for (unsigned i = 0; i != 4; ++i) {
3197 SDOperand Elt = MaskElts[i];
3198 if (Elt.getOpcode() == ISD::UNDEF) {
3199 MaskVec.push_back(Elt);
3202 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3205 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3206 // If this element is in the right place after this shuffle, then
3208 if ((int)(EltIdx / 4) == BestLowQuad)
3213 for (unsigned i = 4; i != 8; ++i)
3214 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3215 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3216 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3220 if (BestHighQuad != -1) {
3221 // Sort high half in order using PSHUFHW if possible.
3223 for (unsigned i = 0; i != 4; ++i)
3224 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3225 bool AnyOutOrder = false;
3226 for (unsigned i = 4; i != 8; ++i) {
3227 SDOperand Elt = MaskElts[i];
3228 if (Elt.getOpcode() == ISD::UNDEF) {
3229 MaskVec.push_back(Elt);
3232 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3235 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3236 // If this element is in the right place after this shuffle, then
3238 if ((int)(EltIdx / 4) == BestHighQuad)
3243 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3244 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3248 // The other elements are put in the right place using pextrw and pinsrw.
3249 for (unsigned i = 0; i != 8; ++i) {
3252 SDOperand Elt = MaskElts[i];
3253 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3256 SDOperand ExtOp = (EltIdx < 8)
3257 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3258 DAG.getConstant(EltIdx, PtrVT))
3259 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3260 DAG.getConstant(EltIdx - 8, PtrVT));
3261 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3262 DAG.getConstant(i, PtrVT));
3267 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3268 ///as few as possible.
3269 // First, let's find out how many elements are already in the right order.
3270 unsigned V1InOrder = 0;
3271 unsigned V1FromV1 = 0;
3272 unsigned V2InOrder = 0;
3273 unsigned V2FromV2 = 0;
3274 SmallVector<SDOperand, 8> V1Elts;
3275 SmallVector<SDOperand, 8> V2Elts;
3276 for (unsigned i = 0; i < 8; ++i) {
3277 SDOperand Elt = MaskElts[i];
3278 if (Elt.getOpcode() == ISD::UNDEF) {
3279 V1Elts.push_back(Elt);
3280 V2Elts.push_back(Elt);
3285 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3287 V1Elts.push_back(Elt);
3288 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3290 } else if (EltIdx == i+8) {
3291 V1Elts.push_back(Elt);
3292 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3294 } else if (EltIdx < 8) {
3295 V1Elts.push_back(Elt);
3298 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3303 if (V2InOrder > V1InOrder) {
3304 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3306 std::swap(V1Elts, V2Elts);
3307 std::swap(V1FromV1, V2FromV2);
3310 if ((V1FromV1 + V1InOrder) != 8) {
3311 // Some elements are from V2.
3313 // If there are elements that are from V1 but out of place,
3314 // then first sort them in place
3315 SmallVector<SDOperand, 8> MaskVec;
3316 for (unsigned i = 0; i < 8; ++i) {
3317 SDOperand Elt = V1Elts[i];
3318 if (Elt.getOpcode() == ISD::UNDEF) {
3319 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3322 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3324 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3326 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3328 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3329 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3333 for (unsigned i = 0; i < 8; ++i) {
3334 SDOperand Elt = V1Elts[i];
3335 if (Elt.getOpcode() == ISD::UNDEF)
3337 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3340 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3341 DAG.getConstant(EltIdx - 8, PtrVT));
3342 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3343 DAG.getConstant(i, PtrVT));
3347 // All elements are from V1.
3349 for (unsigned i = 0; i < 8; ++i) {
3350 SDOperand Elt = V1Elts[i];
3351 if (Elt.getOpcode() == ISD::UNDEF)
3353 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3354 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3355 DAG.getConstant(EltIdx, PtrVT));
3356 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3357 DAG.getConstant(i, PtrVT));
3363 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3364 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3365 /// done when every pair / quad of shuffle mask elements point to elements in
3366 /// the right sequence. e.g.
3367 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3369 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3371 SDOperand PermMask, SelectionDAG &DAG,
3372 TargetLowering &TLI) {
3373 unsigned NumElems = PermMask.getNumOperands();
3374 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3375 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3376 MVT::ValueType NewVT = MaskVT;
3378 case MVT::v4f32: NewVT = MVT::v2f64; break;
3379 case MVT::v4i32: NewVT = MVT::v2i64; break;
3380 case MVT::v8i16: NewVT = MVT::v4i32; break;
3381 case MVT::v16i8: NewVT = MVT::v4i32; break;
3382 default: assert(false && "Unexpected!");
3385 if (NewWidth == 2) {
3386 if (MVT::isInteger(VT))
3391 unsigned Scale = NumElems / NewWidth;
3392 SmallVector<SDOperand, 8> MaskVec;
3393 for (unsigned i = 0; i < NumElems; i += Scale) {
3394 unsigned StartIdx = ~0U;
3395 for (unsigned j = 0; j < Scale; ++j) {
3396 SDOperand Elt = PermMask.getOperand(i+j);
3397 if (Elt.getOpcode() == ISD::UNDEF)
3399 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3400 if (StartIdx == ~0U)
3401 StartIdx = EltIdx - (EltIdx % Scale);
3402 if (EltIdx != StartIdx + j)
3405 if (StartIdx == ~0U)
3406 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3408 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3411 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3412 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3413 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3414 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3415 &MaskVec[0], MaskVec.size()));
3419 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3420 SDOperand V1 = Op.getOperand(0);
3421 SDOperand V2 = Op.getOperand(1);
3422 SDOperand PermMask = Op.getOperand(2);
3423 MVT::ValueType VT = Op.getValueType();
3424 unsigned NumElems = PermMask.getNumOperands();
3425 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3426 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3427 bool V1IsSplat = false;
3428 bool V2IsSplat = false;
3430 if (isUndefShuffle(Op.Val))
3431 return DAG.getNode(ISD::UNDEF, VT);
3433 if (isZeroShuffle(Op.Val))
3434 return getZeroVector(VT, DAG);
3436 if (isIdentityMask(PermMask.Val))
3438 else if (isIdentityMask(PermMask.Val, true))
3441 if (isSplatMask(PermMask.Val)) {
3442 if (NumElems <= 4) return Op;
3443 // Promote it to a v4i32 splat.
3444 return PromoteSplat(Op, DAG);
3447 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3449 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3450 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3452 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3453 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3454 // FIXME: Figure out a cleaner way to do this.
3455 // Try to make use of movq to zero out the top part.
3456 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3457 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3459 SDOperand NewV1 = NewOp.getOperand(0);
3460 SDOperand NewV2 = NewOp.getOperand(1);
3461 SDOperand NewMask = NewOp.getOperand(2);
3462 if (isCommutedMOVL(NewMask.Val, true, false)) {
3463 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3464 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3465 NewV1, NewV2, getMOVLMask(2, DAG));
3466 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3469 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3470 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3471 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3472 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3476 if (X86::isMOVLMask(PermMask.Val))
3477 return (V1IsUndef) ? V2 : Op;
3479 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3480 X86::isMOVSLDUPMask(PermMask.Val) ||
3481 X86::isMOVHLPSMask(PermMask.Val) ||
3482 X86::isMOVHPMask(PermMask.Val) ||
3483 X86::isMOVLPMask(PermMask.Val))
3486 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3487 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3488 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3490 bool Commuted = false;
3491 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3492 // 1,1,1,1 -> v8i16 though.
3493 V1IsSplat = isSplatVector(V1.Val);
3494 V2IsSplat = isSplatVector(V2.Val);
3496 // Canonicalize the splat or undef, if present, to be on the RHS.
3497 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3498 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3499 std::swap(V1IsSplat, V2IsSplat);
3500 std::swap(V1IsUndef, V2IsUndef);
3504 // FIXME: Figure out a cleaner way to do this.
3505 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3506 if (V2IsUndef) return V1;
3507 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3509 // V2 is a splat, so the mask may be malformed. That is, it may point
3510 // to any V2 element. The instruction selectior won't like this. Get
3511 // a corrected mask and commute to form a proper MOVS{S|D}.
3512 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3513 if (NewMask.Val != PermMask.Val)
3514 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3519 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3520 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3521 X86::isUNPCKLMask(PermMask.Val) ||
3522 X86::isUNPCKHMask(PermMask.Val))
3526 // Normalize mask so all entries that point to V2 points to its first
3527 // element then try to match unpck{h|l} again. If match, return a
3528 // new vector_shuffle with the corrected mask.
3529 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3530 if (NewMask.Val != PermMask.Val) {
3531 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3532 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3533 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3534 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3535 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3536 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3541 // Normalize the node to match x86 shuffle ops if needed
3542 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3543 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3546 // Commute is back and try unpck* again.
3547 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3548 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3549 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3550 X86::isUNPCKLMask(PermMask.Val) ||
3551 X86::isUNPCKHMask(PermMask.Val))
3555 // If VT is integer, try PSHUF* first, then SHUFP*.
3556 if (MVT::isInteger(VT)) {
3557 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3558 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3559 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3560 X86::isPSHUFDMask(PermMask.Val)) ||
3561 X86::isPSHUFHWMask(PermMask.Val) ||
3562 X86::isPSHUFLWMask(PermMask.Val)) {
3563 if (V2.getOpcode() != ISD::UNDEF)
3564 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3565 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3569 if (X86::isSHUFPMask(PermMask.Val) &&
3570 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3573 // Floating point cases in the other order.
3574 if (X86::isSHUFPMask(PermMask.Val))
3576 if (X86::isPSHUFDMask(PermMask.Val) ||
3577 X86::isPSHUFHWMask(PermMask.Val) ||
3578 X86::isPSHUFLWMask(PermMask.Val)) {
3579 if (V2.getOpcode() != ISD::UNDEF)
3580 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3581 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3586 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3587 if (VT == MVT::v8i16) {
3588 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3593 // Handle all 4 wide cases with a number of shuffles.
3594 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3595 // Don't do this for MMX.
3596 MVT::ValueType MaskVT = PermMask.getValueType();
3597 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3598 SmallVector<std::pair<int, int>, 8> Locs;
3599 Locs.reserve(NumElems);
3600 SmallVector<SDOperand, 8> Mask1(NumElems,
3601 DAG.getNode(ISD::UNDEF, MaskEVT));
3602 SmallVector<SDOperand, 8> Mask2(NumElems,
3603 DAG.getNode(ISD::UNDEF, MaskEVT));
3606 // If no more than two elements come from either vector. This can be
3607 // implemented with two shuffles. First shuffle gather the elements.
3608 // The second shuffle, which takes the first shuffle as both of its
3609 // vector operands, put the elements into the right order.
3610 for (unsigned i = 0; i != NumElems; ++i) {
3611 SDOperand Elt = PermMask.getOperand(i);
3612 if (Elt.getOpcode() == ISD::UNDEF) {
3613 Locs[i] = std::make_pair(-1, -1);
3615 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3616 if (Val < NumElems) {
3617 Locs[i] = std::make_pair(0, NumLo);
3621 Locs[i] = std::make_pair(1, NumHi);
3622 if (2+NumHi < NumElems)
3623 Mask1[2+NumHi] = Elt;
3628 if (NumLo <= 2 && NumHi <= 2) {
3629 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3630 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3631 &Mask1[0], Mask1.size()));
3632 for (unsigned i = 0; i != NumElems; ++i) {
3633 if (Locs[i].first == -1)
3636 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3637 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3638 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3642 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3643 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3644 &Mask2[0], Mask2.size()));
3647 // Break it into (shuffle shuffle_hi, shuffle_lo).
3649 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3650 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3651 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3652 unsigned MaskIdx = 0;
3654 unsigned HiIdx = NumElems/2;
3655 for (unsigned i = 0; i != NumElems; ++i) {
3656 if (i == NumElems/2) {
3662 SDOperand Elt = PermMask.getOperand(i);
3663 if (Elt.getOpcode() == ISD::UNDEF) {
3664 Locs[i] = std::make_pair(-1, -1);
3665 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3666 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3667 (*MaskPtr)[LoIdx] = Elt;
3670 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3671 (*MaskPtr)[HiIdx] = Elt;
3676 SDOperand LoShuffle =
3677 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3678 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3679 &LoMask[0], LoMask.size()));
3680 SDOperand HiShuffle =
3681 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3682 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3683 &HiMask[0], HiMask.size()));
3684 SmallVector<SDOperand, 8> MaskOps;
3685 for (unsigned i = 0; i != NumElems; ++i) {
3686 if (Locs[i].first == -1) {
3687 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3689 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3690 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3693 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3694 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3695 &MaskOps[0], MaskOps.size()));
3702 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDOperand Op,
3703 SelectionDAG &DAG) {
3704 MVT::ValueType VT = Op.getValueType();
3705 if (MVT::getSizeInBits(VT) == 8) {
3706 SDOperand Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
3707 Op.getOperand(0), Op.getOperand(1));
3708 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3709 DAG.getValueType(VT));
3710 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3711 } else if (MVT::getSizeInBits(VT) == 16) {
3712 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
3713 Op.getOperand(0), Op.getOperand(1));
3714 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3715 DAG.getValueType(VT));
3716 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3723 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3724 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3727 if (Subtarget->hasSSE41())
3728 return LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
3730 MVT::ValueType VT = Op.getValueType();
3731 // TODO: handle v16i8.
3732 if (MVT::getSizeInBits(VT) == 16) {
3733 SDOperand Vec = Op.getOperand(0);
3734 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3736 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3737 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3738 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3740 // Transform it so it match pextrw which produces a 32-bit result.
3741 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3742 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3743 Op.getOperand(0), Op.getOperand(1));
3744 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3745 DAG.getValueType(VT));
3746 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3747 } else if (MVT::getSizeInBits(VT) == 32) {
3748 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3751 // SHUFPS the element to the lowest double word, then movss.
3752 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3753 SmallVector<SDOperand, 8> IdxVec;
3755 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3757 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3759 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3761 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3762 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3763 &IdxVec[0], IdxVec.size());
3764 SDOperand Vec = Op.getOperand(0);
3765 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3766 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3767 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3768 DAG.getIntPtrConstant(0));
3769 } else if (MVT::getSizeInBits(VT) == 64) {
3770 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
3771 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
3772 // to match extract_elt for f64.
3773 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3777 // UNPCKHPD the element to the lowest double word, then movsd.
3778 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3779 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3780 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3781 SmallVector<SDOperand, 8> IdxVec;
3782 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3784 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3785 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3786 &IdxVec[0], IdxVec.size());
3787 SDOperand Vec = Op.getOperand(0);
3788 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3789 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3790 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3791 DAG.getIntPtrConstant(0));
3798 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDOperand Op, SelectionDAG &DAG){
3799 MVT::ValueType VT = Op.getValueType();
3800 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3802 SDOperand N0 = Op.getOperand(0);
3803 SDOperand N1 = Op.getOperand(1);
3804 SDOperand N2 = Op.getOperand(2);
3806 if ((MVT::getSizeInBits(EVT) == 8) || (MVT::getSizeInBits(EVT) == 16)) {
3807 unsigned Opc = (MVT::getSizeInBits(EVT) == 8) ? X86ISD::PINSRB
3809 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
3811 if (N1.getValueType() != MVT::i32)
3812 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3813 if (N2.getValueType() != MVT::i32)
3814 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3815 return DAG.getNode(Opc, VT, N0, N1, N2);
3816 } else if (EVT == MVT::f32) {
3817 // Bits [7:6] of the constant are the source select. This will always be
3818 // zero here. The DAG Combiner may combine an extract_elt index into these
3819 // bits. For example (insert (extract, 3), 2) could be matched by putting
3820 // the '3' into bits [7:6] of X86ISD::INSERTPS.
3821 // Bits [5:4] of the constant are the destination select. This is the
3822 // value of the incoming immediate.
3823 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
3824 // combine either bitwise AND or insert of float 0.0 to set these bits.
3825 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
3826 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
3832 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3833 MVT::ValueType VT = Op.getValueType();
3834 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3836 if (Subtarget->hasSSE41())
3837 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
3842 SDOperand N0 = Op.getOperand(0);
3843 SDOperand N1 = Op.getOperand(1);
3844 SDOperand N2 = Op.getOperand(2);
3846 if (MVT::getSizeInBits(EVT) == 16) {
3847 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3848 // as its second argument.
3849 if (N1.getValueType() != MVT::i32)
3850 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3851 if (N2.getValueType() != MVT::i32)
3852 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3853 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3859 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3860 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3861 MVT::ValueType VT = MVT::v2i32;
3862 switch (Op.getValueType()) {
3869 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
3870 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
3873 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3874 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3875 // one of the above mentioned nodes. It has to be wrapped because otherwise
3876 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3877 // be used to form addressing mode. These wrapped nodes will be selected
3880 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3881 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3882 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3884 CP->getAlignment());
3885 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3886 // With PIC, the address is actually $g + Offset.
3887 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3888 !Subtarget->isPICStyleRIPRel()) {
3889 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3890 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3898 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3899 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3900 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3901 // If it's a debug information descriptor, don't mess with it.
3902 if (DAG.isVerifiedDebugInfoDesc(Op))
3904 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3905 // With PIC, the address is actually $g + Offset.
3906 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3907 !Subtarget->isPICStyleRIPRel()) {
3908 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3909 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3913 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3914 // load the value at address GV, not the value of GV itself. This means that
3915 // the GlobalAddress must be in the base or index register of the address, not
3916 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3917 // The same applies for external symbols during PIC codegen
3918 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3919 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
3920 PseudoSourceValue::getGOT(), 0);
3925 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3927 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3928 const MVT::ValueType PtrVT) {
3930 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3931 DAG.getNode(X86ISD::GlobalBaseReg,
3933 InFlag = Chain.getValue(1);
3935 // emit leal symbol@TLSGD(,%ebx,1), %eax
3936 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3937 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3938 GA->getValueType(0),
3940 SDOperand Ops[] = { Chain, TGA, InFlag };
3941 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3942 InFlag = Result.getValue(2);
3943 Chain = Result.getValue(1);
3945 // call ___tls_get_addr. This function receives its argument in
3946 // the register EAX.
3947 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3948 InFlag = Chain.getValue(1);
3950 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3951 SDOperand Ops1[] = { Chain,
3952 DAG.getTargetExternalSymbol("___tls_get_addr",
3954 DAG.getRegister(X86::EAX, PtrVT),
3955 DAG.getRegister(X86::EBX, PtrVT),
3957 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3958 InFlag = Chain.getValue(1);
3960 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3963 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3964 // "local exec" model.
3966 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3967 const MVT::ValueType PtrVT) {
3968 // Get the Thread Pointer
3969 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3970 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3972 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3973 GA->getValueType(0),
3975 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3977 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3978 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
3979 PseudoSourceValue::getGOT(), 0);
3981 // The address of the thread local variable is the add of the thread
3982 // pointer with the offset of the variable.
3983 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3987 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3988 // TODO: implement the "local dynamic" model
3989 // TODO: implement the "initial exec"model for pic executables
3990 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3991 "TLS not implemented for non-ELF and 64-bit targets");
3992 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3993 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3994 // otherwise use the "Local Exec"TLS Model
3995 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3996 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3998 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4002 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
4003 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4004 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4005 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4006 // With PIC, the address is actually $g + Offset.
4007 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4008 !Subtarget->isPICStyleRIPRel()) {
4009 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4010 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4017 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
4018 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4019 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4020 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4021 // With PIC, the address is actually $g + Offset.
4022 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4023 !Subtarget->isPICStyleRIPRel()) {
4024 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4025 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4032 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4033 /// take a 2 x i32 value to shift plus a shift amount.
4034 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
4035 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4036 MVT::ValueType VT = Op.getValueType();
4037 unsigned VTBits = MVT::getSizeInBits(VT);
4038 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4039 SDOperand ShOpLo = Op.getOperand(0);
4040 SDOperand ShOpHi = Op.getOperand(1);
4041 SDOperand ShAmt = Op.getOperand(2);
4042 SDOperand Tmp1 = isSRA ?
4043 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4044 DAG.getConstant(0, VT);
4046 SDOperand Tmp2, Tmp3;
4047 if (Op.getOpcode() == ISD::SHL_PARTS) {
4048 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4049 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4051 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4052 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4055 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4056 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4057 DAG.getConstant(VTBits, MVT::i8));
4058 SDOperand Cond = DAG.getNode(X86ISD::CMP, VT,
4059 AndNode, DAG.getConstant(0, MVT::i8));
4062 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4063 VTs = DAG.getNodeValueTypes(VT, MVT::Flag);
4064 SmallVector<SDOperand, 4> Ops;
4065 if (Op.getOpcode() == ISD::SHL_PARTS) {
4066 Ops.push_back(Tmp2);
4067 Ops.push_back(Tmp3);
4069 Ops.push_back(Cond);
4070 Hi = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4073 Ops.push_back(Tmp3);
4074 Ops.push_back(Tmp1);
4076 Ops.push_back(Cond);
4077 Lo = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4079 Ops.push_back(Tmp2);
4080 Ops.push_back(Tmp3);
4082 Ops.push_back(Cond);
4083 Lo = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4086 Ops.push_back(Tmp3);
4087 Ops.push_back(Tmp1);
4089 Ops.push_back(Cond);
4090 Hi = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4093 VTs = DAG.getNodeValueTypes(VT, VT);
4097 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
4100 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
4101 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
4102 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
4103 "Unknown SINT_TO_FP to lower!");
4105 // These are really Legal; caller falls through into that case.
4106 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4108 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4109 Subtarget->is64Bit())
4112 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
4113 MachineFunction &MF = DAG.getMachineFunction();
4114 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4115 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4116 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4118 PseudoSourceValue::getFixedStack(),
4123 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4125 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4127 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4128 SmallVector<SDOperand, 8> Ops;
4129 Ops.push_back(Chain);
4130 Ops.push_back(StackSlot);
4131 Ops.push_back(DAG.getValueType(SrcVT));
4132 SDOperand Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4133 Tys, &Ops[0], Ops.size());
4136 Chain = Result.getValue(1);
4137 SDOperand InFlag = Result.getValue(2);
4139 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4140 // shouldn't be necessary except that RFP cannot be live across
4141 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4142 MachineFunction &MF = DAG.getMachineFunction();
4143 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4144 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4145 Tys = DAG.getVTList(MVT::Other);
4146 SmallVector<SDOperand, 8> Ops;
4147 Ops.push_back(Chain);
4148 Ops.push_back(Result);
4149 Ops.push_back(StackSlot);
4150 Ops.push_back(DAG.getValueType(Op.getValueType()));
4151 Ops.push_back(InFlag);
4152 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4153 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4154 PseudoSourceValue::getFixedStack(), SSFI);
4160 std::pair<SDOperand,SDOperand> X86TargetLowering::
4161 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4162 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4163 "Unknown FP_TO_SINT to lower!");
4165 // These are really Legal.
4166 if (Op.getValueType() == MVT::i32 &&
4167 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4168 return std::make_pair(SDOperand(), SDOperand());
4169 if (Subtarget->is64Bit() &&
4170 Op.getValueType() == MVT::i64 &&
4171 Op.getOperand(0).getValueType() != MVT::f80)
4172 return std::make_pair(SDOperand(), SDOperand());
4174 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4176 MachineFunction &MF = DAG.getMachineFunction();
4177 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4178 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4179 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4181 switch (Op.getValueType()) {
4182 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4183 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4184 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4185 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4188 SDOperand Chain = DAG.getEntryNode();
4189 SDOperand Value = Op.getOperand(0);
4190 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4191 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4192 Chain = DAG.getStore(Chain, Value, StackSlot,
4193 PseudoSourceValue::getFixedStack(), SSFI);
4194 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4196 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4198 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4199 Chain = Value.getValue(1);
4200 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4201 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4204 // Build the FP_TO_INT*_IN_MEM
4205 SDOperand Ops[] = { Chain, Value, StackSlot };
4206 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4208 return std::make_pair(FIST, StackSlot);
4211 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4212 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4213 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4214 if (FIST.Val == 0) return SDOperand();
4217 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4220 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4221 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4222 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4223 if (FIST.Val == 0) return 0;
4225 // Return an i64 load from the stack slot.
4226 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4228 // Use a MERGE_VALUES node to drop the chain result value.
4229 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4232 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4233 MVT::ValueType VT = Op.getValueType();
4234 MVT::ValueType EltVT = VT;
4235 if (MVT::isVector(VT))
4236 EltVT = MVT::getVectorElementType(VT);
4237 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4238 std::vector<Constant*> CV;
4239 if (EltVT == MVT::f64) {
4240 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4244 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4250 Constant *C = ConstantVector::get(CV);
4251 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4252 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4253 PseudoSourceValue::getConstantPool(), 0,
4255 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4258 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4259 MVT::ValueType VT = Op.getValueType();
4260 MVT::ValueType EltVT = VT;
4261 unsigned EltNum = 1;
4262 if (MVT::isVector(VT)) {
4263 EltVT = MVT::getVectorElementType(VT);
4264 EltNum = MVT::getVectorNumElements(VT);
4266 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4267 std::vector<Constant*> CV;
4268 if (EltVT == MVT::f64) {
4269 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4273 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4279 Constant *C = ConstantVector::get(CV);
4280 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4281 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4282 PseudoSourceValue::getConstantPool(), 0,
4284 if (MVT::isVector(VT)) {
4285 return DAG.getNode(ISD::BIT_CONVERT, VT,
4286 DAG.getNode(ISD::XOR, MVT::v2i64,
4287 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4288 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4290 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4294 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4295 SDOperand Op0 = Op.getOperand(0);
4296 SDOperand Op1 = Op.getOperand(1);
4297 MVT::ValueType VT = Op.getValueType();
4298 MVT::ValueType SrcVT = Op1.getValueType();
4299 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4301 // If second operand is smaller, extend it first.
4302 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4303 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4305 SrcTy = MVT::getTypeForValueType(SrcVT);
4307 // And if it is bigger, shrink it first.
4308 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4309 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4311 SrcTy = MVT::getTypeForValueType(SrcVT);
4314 // At this point the operands and the result should have the same
4315 // type, and that won't be f80 since that is not custom lowered.
4317 // First get the sign bit of second operand.
4318 std::vector<Constant*> CV;
4319 if (SrcVT == MVT::f64) {
4320 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4321 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4323 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4324 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4325 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4326 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4328 Constant *C = ConstantVector::get(CV);
4329 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4330 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4331 PseudoSourceValue::getConstantPool(), 0,
4333 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4335 // Shift sign bit right or left if the two operands have different types.
4336 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4337 // Op0 is MVT::f32, Op1 is MVT::f64.
4338 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4339 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4340 DAG.getConstant(32, MVT::i32));
4341 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4342 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4343 DAG.getIntPtrConstant(0));
4346 // Clear first operand sign bit.
4348 if (VT == MVT::f64) {
4349 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4350 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4352 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4353 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4354 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4355 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4357 C = ConstantVector::get(CV);
4358 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4359 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4360 PseudoSourceValue::getConstantPool(), 0,
4362 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4364 // Or the value with the sign bit.
4365 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4368 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4369 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4371 SDOperand Op0 = Op.getOperand(0);
4372 SDOperand Op1 = Op.getOperand(1);
4373 SDOperand CC = Op.getOperand(2);
4374 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4375 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4378 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4380 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4381 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4382 DAG.getConstant(X86CC, MVT::i8), Cond);
4385 assert(isFP && "Illegal integer SetCC!");
4387 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4388 switch (SetCCOpcode) {
4389 default: assert(false && "Illegal floating point SetCC!");
4390 case ISD::SETOEQ: { // !PF & ZF
4391 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4392 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4393 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4394 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4395 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4397 case ISD::SETUNE: { // PF | !ZF
4398 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4399 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4400 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4401 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4402 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4408 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4409 bool addTest = true;
4410 SDOperand Cond = Op.getOperand(0);
4413 if (Cond.getOpcode() == ISD::SETCC)
4414 Cond = LowerSETCC(Cond, DAG);
4416 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4417 // setting operand in place of the X86ISD::SETCC.
4418 if (Cond.getOpcode() == X86ISD::SETCC) {
4419 CC = Cond.getOperand(0);
4421 SDOperand Cmp = Cond.getOperand(1);
4422 unsigned Opc = Cmp.getOpcode();
4423 MVT::ValueType VT = Op.getValueType();
4425 bool IllegalFPCMov = false;
4426 if (MVT::isFloatingPoint(VT) && !MVT::isVector(VT) &&
4427 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4428 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4430 if ((Opc == X86ISD::CMP ||
4431 Opc == X86ISD::COMI ||
4432 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4439 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4440 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4443 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4445 SmallVector<SDOperand, 4> Ops;
4446 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4447 // condition is true.
4448 Ops.push_back(Op.getOperand(2));
4449 Ops.push_back(Op.getOperand(1));
4451 Ops.push_back(Cond);
4452 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4455 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4456 bool addTest = true;
4457 SDOperand Chain = Op.getOperand(0);
4458 SDOperand Cond = Op.getOperand(1);
4459 SDOperand Dest = Op.getOperand(2);
4462 if (Cond.getOpcode() == ISD::SETCC)
4463 Cond = LowerSETCC(Cond, DAG);
4465 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4466 // setting operand in place of the X86ISD::SETCC.
4467 if (Cond.getOpcode() == X86ISD::SETCC) {
4468 CC = Cond.getOperand(0);
4470 SDOperand Cmp = Cond.getOperand(1);
4471 unsigned Opc = Cmp.getOpcode();
4472 if (Opc == X86ISD::CMP ||
4473 Opc == X86ISD::COMI ||
4474 Opc == X86ISD::UCOMI) {
4481 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4482 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4484 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4485 Chain, Op.getOperand(2), CC, Cond);
4489 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4490 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4491 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4492 // that the guard pages used by the OS virtual memory manager are allocated in
4493 // correct sequence.
4495 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4496 SelectionDAG &DAG) {
4497 assert(Subtarget->isTargetCygMing() &&
4498 "This should be used only on Cygwin/Mingw targets");
4501 SDOperand Chain = Op.getOperand(0);
4502 SDOperand Size = Op.getOperand(1);
4503 // FIXME: Ensure alignment here
4507 MVT::ValueType IntPtr = getPointerTy();
4508 MVT::ValueType SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
4510 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4511 Flag = Chain.getValue(1);
4513 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4514 SDOperand Ops[] = { Chain,
4515 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4516 DAG.getRegister(X86::EAX, IntPtr),
4518 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4519 Flag = Chain.getValue(1);
4521 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4523 std::vector<MVT::ValueType> Tys;
4524 Tys.push_back(SPTy);
4525 Tys.push_back(MVT::Other);
4526 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4527 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4530 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4531 SDOperand InFlag(0, 0);
4532 SDOperand Chain = Op.getOperand(0);
4534 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4535 if (Align == 0) Align = 1;
4537 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4538 // If not DWORD aligned or size is more than the threshold, call memset.
4539 // The libc version is likely to be faster for these cases. It can use the
4540 // address value and run time information about the CPU.
4541 if ((Align & 3) != 0 ||
4542 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4543 MVT::ValueType IntPtr = getPointerTy();
4544 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4545 TargetLowering::ArgListTy Args;
4546 TargetLowering::ArgListEntry Entry;
4547 Entry.Node = Op.getOperand(1);
4548 Entry.Ty = IntPtrTy;
4549 Args.push_back(Entry);
4550 // Extend the unsigned i8 argument to be an int value for the call.
4551 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4552 Entry.Ty = IntPtrTy;
4553 Args.push_back(Entry);
4554 Entry.Node = Op.getOperand(3);
4555 Args.push_back(Entry);
4556 std::pair<SDOperand,SDOperand> CallResult =
4557 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
4558 false, DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4559 return CallResult.second;
4564 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4565 unsigned BytesLeft = 0;
4566 bool TwoRepStos = false;
4569 uint64_t Val = ValC->getValue() & 255;
4571 // If the value is a constant, then we can potentially use larger sets.
4572 switch (Align & 3) {
4573 case 2: // WORD aligned
4576 Val = (Val << 8) | Val;
4578 case 0: // DWORD aligned
4581 Val = (Val << 8) | Val;
4582 Val = (Val << 16) | Val;
4583 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4586 Val = (Val << 32) | Val;
4589 default: // Byte aligned
4592 Count = Op.getOperand(3);
4596 if (AVT > MVT::i8) {
4598 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4599 Count = DAG.getIntPtrConstant(I->getValue() / UBytes);
4600 BytesLeft = I->getValue() % UBytes;
4602 assert(AVT >= MVT::i32 &&
4603 "Do not use rep;stos if not at least DWORD aligned");
4604 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4605 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4610 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4612 InFlag = Chain.getValue(1);
4615 Count = Op.getOperand(3);
4616 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4617 InFlag = Chain.getValue(1);
4620 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4622 InFlag = Chain.getValue(1);
4623 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4624 Op.getOperand(1), InFlag);
4625 InFlag = Chain.getValue(1);
4627 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4628 SmallVector<SDOperand, 8> Ops;
4629 Ops.push_back(Chain);
4630 Ops.push_back(DAG.getValueType(AVT));
4631 Ops.push_back(InFlag);
4632 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4635 InFlag = Chain.getValue(1);
4636 Count = Op.getOperand(3);
4637 MVT::ValueType CVT = Count.getValueType();
4638 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4639 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4640 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4642 InFlag = Chain.getValue(1);
4643 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4645 Ops.push_back(Chain);
4646 Ops.push_back(DAG.getValueType(MVT::i8));
4647 Ops.push_back(InFlag);
4648 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4649 } else if (BytesLeft) {
4650 // Issue stores for the last 1 - 7 bytes.
4652 unsigned Val = ValC->getValue() & 255;
4653 unsigned Offset = I->getValue() - BytesLeft;
4654 SDOperand DstAddr = Op.getOperand(1);
4655 MVT::ValueType AddrVT = DstAddr.getValueType();
4656 if (BytesLeft >= 4) {
4657 Val = (Val << 8) | Val;
4658 Val = (Val << 16) | Val;
4659 Value = DAG.getConstant(Val, MVT::i32);
4660 Chain = DAG.getStore(Chain, Value,
4661 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4662 DAG.getConstant(Offset, AddrVT)),
4667 if (BytesLeft >= 2) {
4668 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4669 Chain = DAG.getStore(Chain, Value,
4670 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4671 DAG.getConstant(Offset, AddrVT)),
4676 if (BytesLeft == 1) {
4677 Value = DAG.getConstant(Val, MVT::i8);
4678 Chain = DAG.getStore(Chain, Value,
4679 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4680 DAG.getConstant(Offset, AddrVT)),
4688 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4693 SelectionDAG &DAG) {
4695 unsigned BytesLeft = 0;
4696 switch (Align & 3) {
4697 case 2: // WORD aligned
4700 case 0: // DWORD aligned
4702 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4705 default: // Byte aligned
4710 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4711 SDOperand Count = DAG.getIntPtrConstant(Size / UBytes);
4712 BytesLeft = Size % UBytes;
4714 SDOperand InFlag(0, 0);
4715 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4717 InFlag = Chain.getValue(1);
4718 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4720 InFlag = Chain.getValue(1);
4721 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4723 InFlag = Chain.getValue(1);
4725 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4726 SmallVector<SDOperand, 8> Ops;
4727 Ops.push_back(Chain);
4728 Ops.push_back(DAG.getValueType(AVT));
4729 Ops.push_back(InFlag);
4730 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4733 // Issue loads and stores for the last 1 - 7 bytes.
4734 unsigned Offset = Size - BytesLeft;
4735 SDOperand DstAddr = Dest;
4736 MVT::ValueType DstVT = DstAddr.getValueType();
4737 SDOperand SrcAddr = Source;
4738 MVT::ValueType SrcVT = SrcAddr.getValueType();
4740 if (BytesLeft >= 4) {
4741 Value = DAG.getLoad(MVT::i32, Chain,
4742 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4743 DAG.getConstant(Offset, SrcVT)),
4745 Chain = Value.getValue(1);
4746 Chain = DAG.getStore(Chain, Value,
4747 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4748 DAG.getConstant(Offset, DstVT)),
4753 if (BytesLeft >= 2) {
4754 Value = DAG.getLoad(MVT::i16, Chain,
4755 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4756 DAG.getConstant(Offset, SrcVT)),
4758 Chain = Value.getValue(1);
4759 Chain = DAG.getStore(Chain, Value,
4760 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4761 DAG.getConstant(Offset, DstVT)),
4767 if (BytesLeft == 1) {
4768 Value = DAG.getLoad(MVT::i8, Chain,
4769 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4770 DAG.getConstant(Offset, SrcVT)),
4772 Chain = Value.getValue(1);
4773 Chain = DAG.getStore(Chain, Value,
4774 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4775 DAG.getConstant(Offset, DstVT)),
4783 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4784 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4785 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4786 SDOperand TheChain = N->getOperand(0);
4787 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4788 if (Subtarget->is64Bit()) {
4789 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4790 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4791 MVT::i64, rax.getValue(2));
4792 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4793 DAG.getConstant(32, MVT::i8));
4795 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4798 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4799 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4802 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4803 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4804 MVT::i32, eax.getValue(2));
4805 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4806 SDOperand Ops[] = { eax, edx };
4807 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
4809 // Use a MERGE_VALUES to return the value and chain.
4810 Ops[1] = edx.getValue(1);
4811 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4812 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4815 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4816 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4818 if (!Subtarget->is64Bit()) {
4819 // vastart just stores the address of the VarArgsFrameIndex slot into the
4820 // memory location argument.
4821 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4822 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
4826 // gp_offset (0 - 6 * 8)
4827 // fp_offset (48 - 48 + 8 * 16)
4828 // overflow_arg_area (point to parameters coming in memory).
4830 SmallVector<SDOperand, 8> MemOps;
4831 SDOperand FIN = Op.getOperand(1);
4833 SDOperand Store = DAG.getStore(Op.getOperand(0),
4834 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4836 MemOps.push_back(Store);
4839 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4840 Store = DAG.getStore(Op.getOperand(0),
4841 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4843 MemOps.push_back(Store);
4845 // Store ptr to overflow_arg_area
4846 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4847 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4848 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
4849 MemOps.push_back(Store);
4851 // Store ptr to reg_save_area.
4852 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
4853 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4854 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
4855 MemOps.push_back(Store);
4856 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4859 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4860 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4861 SDOperand Chain = Op.getOperand(0);
4862 SDOperand DstPtr = Op.getOperand(1);
4863 SDOperand SrcPtr = Op.getOperand(2);
4864 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4865 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4867 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr, SrcSV, 0);
4868 Chain = SrcPtr.getValue(1);
4869 for (unsigned i = 0; i < 3; ++i) {
4870 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr, SrcSV, 0);
4871 Chain = Val.getValue(1);
4872 Chain = DAG.getStore(Chain, Val, DstPtr, DstSV, 0);
4875 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4876 DAG.getIntPtrConstant(8));
4877 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4878 DAG.getIntPtrConstant(8));
4884 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4885 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4887 default: return SDOperand(); // Don't custom lower most intrinsics.
4888 // Comparison intrinsics.
4889 case Intrinsic::x86_sse_comieq_ss:
4890 case Intrinsic::x86_sse_comilt_ss:
4891 case Intrinsic::x86_sse_comile_ss:
4892 case Intrinsic::x86_sse_comigt_ss:
4893 case Intrinsic::x86_sse_comige_ss:
4894 case Intrinsic::x86_sse_comineq_ss:
4895 case Intrinsic::x86_sse_ucomieq_ss:
4896 case Intrinsic::x86_sse_ucomilt_ss:
4897 case Intrinsic::x86_sse_ucomile_ss:
4898 case Intrinsic::x86_sse_ucomigt_ss:
4899 case Intrinsic::x86_sse_ucomige_ss:
4900 case Intrinsic::x86_sse_ucomineq_ss:
4901 case Intrinsic::x86_sse2_comieq_sd:
4902 case Intrinsic::x86_sse2_comilt_sd:
4903 case Intrinsic::x86_sse2_comile_sd:
4904 case Intrinsic::x86_sse2_comigt_sd:
4905 case Intrinsic::x86_sse2_comige_sd:
4906 case Intrinsic::x86_sse2_comineq_sd:
4907 case Intrinsic::x86_sse2_ucomieq_sd:
4908 case Intrinsic::x86_sse2_ucomilt_sd:
4909 case Intrinsic::x86_sse2_ucomile_sd:
4910 case Intrinsic::x86_sse2_ucomigt_sd:
4911 case Intrinsic::x86_sse2_ucomige_sd:
4912 case Intrinsic::x86_sse2_ucomineq_sd: {
4914 ISD::CondCode CC = ISD::SETCC_INVALID;
4917 case Intrinsic::x86_sse_comieq_ss:
4918 case Intrinsic::x86_sse2_comieq_sd:
4922 case Intrinsic::x86_sse_comilt_ss:
4923 case Intrinsic::x86_sse2_comilt_sd:
4927 case Intrinsic::x86_sse_comile_ss:
4928 case Intrinsic::x86_sse2_comile_sd:
4932 case Intrinsic::x86_sse_comigt_ss:
4933 case Intrinsic::x86_sse2_comigt_sd:
4937 case Intrinsic::x86_sse_comige_ss:
4938 case Intrinsic::x86_sse2_comige_sd:
4942 case Intrinsic::x86_sse_comineq_ss:
4943 case Intrinsic::x86_sse2_comineq_sd:
4947 case Intrinsic::x86_sse_ucomieq_ss:
4948 case Intrinsic::x86_sse2_ucomieq_sd:
4949 Opc = X86ISD::UCOMI;
4952 case Intrinsic::x86_sse_ucomilt_ss:
4953 case Intrinsic::x86_sse2_ucomilt_sd:
4954 Opc = X86ISD::UCOMI;
4957 case Intrinsic::x86_sse_ucomile_ss:
4958 case Intrinsic::x86_sse2_ucomile_sd:
4959 Opc = X86ISD::UCOMI;
4962 case Intrinsic::x86_sse_ucomigt_ss:
4963 case Intrinsic::x86_sse2_ucomigt_sd:
4964 Opc = X86ISD::UCOMI;
4967 case Intrinsic::x86_sse_ucomige_ss:
4968 case Intrinsic::x86_sse2_ucomige_sd:
4969 Opc = X86ISD::UCOMI;
4972 case Intrinsic::x86_sse_ucomineq_ss:
4973 case Intrinsic::x86_sse2_ucomineq_sd:
4974 Opc = X86ISD::UCOMI;
4980 SDOperand LHS = Op.getOperand(1);
4981 SDOperand RHS = Op.getOperand(2);
4982 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4984 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4985 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4986 DAG.getConstant(X86CC, MVT::i8), Cond);
4987 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4992 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4993 // Depths > 0 not supported yet!
4994 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4997 // Just load the return address
4998 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4999 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5002 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
5003 // Depths > 0 not supported yet!
5004 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5007 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5008 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5009 DAG.getIntPtrConstant(4));
5012 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
5013 SelectionDAG &DAG) {
5014 // Is not yet supported on x86-64
5015 if (Subtarget->is64Bit())
5018 return DAG.getIntPtrConstant(8);
5021 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
5023 assert(!Subtarget->is64Bit() &&
5024 "Lowering of eh_return builtin is not supported yet on x86-64");
5026 MachineFunction &MF = DAG.getMachineFunction();
5027 SDOperand Chain = Op.getOperand(0);
5028 SDOperand Offset = Op.getOperand(1);
5029 SDOperand Handler = Op.getOperand(2);
5031 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
5034 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5035 DAG.getIntPtrConstant(-4UL));
5036 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5037 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5038 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5039 MF.getRegInfo().addLiveOut(X86::ECX);
5041 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5042 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5045 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
5046 SelectionDAG &DAG) {
5047 SDOperand Root = Op.getOperand(0);
5048 SDOperand Trmp = Op.getOperand(1); // trampoline
5049 SDOperand FPtr = Op.getOperand(2); // nested function
5050 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
5052 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5054 const X86InstrInfo *TII =
5055 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5057 if (Subtarget->is64Bit()) {
5058 SDOperand OutChains[6];
5060 // Large code-model.
5062 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5063 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5065 const unsigned char N86R10 =
5066 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R10);
5067 const unsigned char N86R11 =
5068 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R11);
5070 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5072 // Load the pointer to the nested function into R11.
5073 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5074 SDOperand Addr = Trmp;
5075 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5078 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5079 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5081 // Load the 'nest' parameter value into R10.
5082 // R10 is specified in X86CallingConv.td
5083 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5084 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5085 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5088 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5089 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5091 // Jump to the nested function.
5092 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5093 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5094 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5097 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5098 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5099 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5103 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5104 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5106 const Function *Func =
5107 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5108 unsigned CC = Func->getCallingConv();
5113 assert(0 && "Unsupported calling convention");
5114 case CallingConv::C:
5115 case CallingConv::X86_StdCall: {
5116 // Pass 'nest' parameter in ECX.
5117 // Must be kept in sync with X86CallingConv.td
5120 // Check that ECX wasn't needed by an 'inreg' parameter.
5121 const FunctionType *FTy = Func->getFunctionType();
5122 const PAListPtr &Attrs = Func->getParamAttrs();
5124 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5125 unsigned InRegCount = 0;
5128 for (FunctionType::param_iterator I = FTy->param_begin(),
5129 E = FTy->param_end(); I != E; ++I, ++Idx)
5130 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5131 // FIXME: should only count parameters that are lowered to integers.
5132 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5134 if (InRegCount > 2) {
5135 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5141 case CallingConv::X86_FastCall:
5142 // Pass 'nest' parameter in EAX.
5143 // Must be kept in sync with X86CallingConv.td
5148 SDOperand OutChains[4];
5149 SDOperand Addr, Disp;
5151 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5152 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5154 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5155 const unsigned char N86Reg =
5156 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5157 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5160 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5161 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5163 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5164 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5165 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5166 TrmpAddr, 5, false, 1);
5168 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5169 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5172 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5173 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5177 SDOperand X86TargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
5179 The rounding mode is in bits 11:10 of FPSR, and has the following
5186 FLT_ROUNDS, on the other hand, expects the following:
5193 To perform the conversion, we do:
5194 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5197 MachineFunction &MF = DAG.getMachineFunction();
5198 const TargetMachine &TM = MF.getTarget();
5199 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5200 unsigned StackAlignment = TFI.getStackAlignment();
5201 MVT::ValueType VT = Op.getValueType();
5203 // Save FP Control Word to stack slot
5204 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5205 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5207 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5208 DAG.getEntryNode(), StackSlot);
5210 // Load FP Control Word from stack slot
5211 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5213 // Transform as necessary
5215 DAG.getNode(ISD::SRL, MVT::i16,
5216 DAG.getNode(ISD::AND, MVT::i16,
5217 CWD, DAG.getConstant(0x800, MVT::i16)),
5218 DAG.getConstant(11, MVT::i8));
5220 DAG.getNode(ISD::SRL, MVT::i16,
5221 DAG.getNode(ISD::AND, MVT::i16,
5222 CWD, DAG.getConstant(0x400, MVT::i16)),
5223 DAG.getConstant(9, MVT::i8));
5226 DAG.getNode(ISD::AND, MVT::i16,
5227 DAG.getNode(ISD::ADD, MVT::i16,
5228 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5229 DAG.getConstant(1, MVT::i16)),
5230 DAG.getConstant(3, MVT::i16));
5233 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5234 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5237 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5238 MVT::ValueType VT = Op.getValueType();
5239 MVT::ValueType OpVT = VT;
5240 unsigned NumBits = MVT::getSizeInBits(VT);
5242 Op = Op.getOperand(0);
5243 if (VT == MVT::i8) {
5244 // Zero extend to i32 since there is not an i8 bsr.
5246 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5249 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5250 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5251 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5253 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5254 SmallVector<SDOperand, 4> Ops;
5256 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5257 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5258 Ops.push_back(Op.getValue(1));
5259 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5261 // Finally xor with NumBits-1.
5262 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5265 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5269 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5270 MVT::ValueType VT = Op.getValueType();
5271 MVT::ValueType OpVT = VT;
5272 unsigned NumBits = MVT::getSizeInBits(VT);
5274 Op = Op.getOperand(0);
5275 if (VT == MVT::i8) {
5277 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5280 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5281 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5282 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5284 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5285 SmallVector<SDOperand, 4> Ops;
5287 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5288 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5289 Ops.push_back(Op.getValue(1));
5290 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5293 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5297 SDOperand X86TargetLowering::LowerLCS(SDOperand Op, SelectionDAG &DAG) {
5298 MVT::ValueType T = cast<AtomicSDNode>(Op.Val)->getVT();
5302 case MVT::i8: Reg = X86::AL; size = 1; break;
5303 case MVT::i16: Reg = X86::AX; size = 2; break;
5304 case MVT::i32: Reg = X86::EAX; size = 4; break;
5306 if (Subtarget->is64Bit()) {
5307 Reg = X86::RAX; size = 8;
5308 } else //Should go away when LowerType stuff lands
5309 return SDOperand(ExpandATOMIC_LCS(Op.Val, DAG), 0);
5312 SDOperand cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5313 Op.getOperand(3), SDOperand());
5314 SDOperand Ops[] = { cpIn.getValue(0),
5317 DAG.getTargetConstant(size, MVT::i8),
5319 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5320 SDOperand Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5322 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5326 SDNode* X86TargetLowering::ExpandATOMIC_LCS(SDNode* Op, SelectionDAG &DAG) {
5327 MVT::ValueType T = cast<AtomicSDNode>(Op)->getVT();
5328 assert (T == MVT::i64 && "Only know how to expand i64 CAS");
5329 SDOperand cpInL, cpInH;
5330 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5331 DAG.getConstant(0, MVT::i32));
5332 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5333 DAG.getConstant(1, MVT::i32));
5334 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5335 cpInL, SDOperand());
5336 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5337 cpInH, cpInL.getValue(1));
5338 SDOperand swapInL, swapInH;
5339 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5340 DAG.getConstant(0, MVT::i32));
5341 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5342 DAG.getConstant(1, MVT::i32));
5343 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5344 swapInL, cpInH.getValue(1));
5345 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5346 swapInH, swapInL.getValue(1));
5347 SDOperand Ops[] = { swapInH.getValue(0),
5349 swapInH.getValue(1)};
5350 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5351 SDOperand Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5352 SDOperand cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5353 Result.getValue(1));
5354 SDOperand cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5355 cpOutL.getValue(2));
5356 SDOperand OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5357 SDOperand ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5358 Tys = DAG.getVTList(MVT::i64, MVT::Other);
5359 return DAG.getNode(ISD::MERGE_VALUES, Tys, ResultVal, cpOutH.getValue(1)).Val;
5362 /// LowerOperation - Provide custom lowering hooks for some operations.
5364 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5365 switch (Op.getOpcode()) {
5366 default: assert(0 && "Should not custom lower this!");
5367 case ISD::ATOMIC_LCS: return LowerLCS(Op,DAG);
5368 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5369 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5370 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5371 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5372 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5373 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5374 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5375 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5376 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5377 case ISD::SHL_PARTS:
5378 case ISD::SRA_PARTS:
5379 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5380 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5381 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5382 case ISD::FABS: return LowerFABS(Op, DAG);
5383 case ISD::FNEG: return LowerFNEG(Op, DAG);
5384 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5385 case ISD::SETCC: return LowerSETCC(Op, DAG);
5386 case ISD::SELECT: return LowerSELECT(Op, DAG);
5387 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5388 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5389 case ISD::CALL: return LowerCALL(Op, DAG);
5390 case ISD::RET: return LowerRET(Op, DAG);
5391 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5392 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5393 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5394 case ISD::VASTART: return LowerVASTART(Op, DAG);
5395 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5396 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5397 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5398 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5399 case ISD::FRAME_TO_ARGS_OFFSET:
5400 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5401 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5402 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5403 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5404 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5405 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5406 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5408 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5409 case ISD::READCYCLECOUNTER:
5410 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5414 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5415 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5416 switch (N->getOpcode()) {
5417 default: assert(0 && "Should not custom lower this!");
5418 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5419 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5420 case ISD::ATOMIC_LCS: return ExpandATOMIC_LCS(N, DAG);
5424 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5426 default: return NULL;
5427 case X86ISD::BSF: return "X86ISD::BSF";
5428 case X86ISD::BSR: return "X86ISD::BSR";
5429 case X86ISD::SHLD: return "X86ISD::SHLD";
5430 case X86ISD::SHRD: return "X86ISD::SHRD";
5431 case X86ISD::FAND: return "X86ISD::FAND";
5432 case X86ISD::FOR: return "X86ISD::FOR";
5433 case X86ISD::FXOR: return "X86ISD::FXOR";
5434 case X86ISD::FSRL: return "X86ISD::FSRL";
5435 case X86ISD::FILD: return "X86ISD::FILD";
5436 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5437 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5438 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5439 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5440 case X86ISD::FLD: return "X86ISD::FLD";
5441 case X86ISD::FST: return "X86ISD::FST";
5442 case X86ISD::CALL: return "X86ISD::CALL";
5443 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5444 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5445 case X86ISD::CMP: return "X86ISD::CMP";
5446 case X86ISD::COMI: return "X86ISD::COMI";
5447 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5448 case X86ISD::SETCC: return "X86ISD::SETCC";
5449 case X86ISD::CMOV: return "X86ISD::CMOV";
5450 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5451 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5452 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5453 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5454 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5455 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5456 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
5457 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5458 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
5459 case X86ISD::PINSRB: return "X86ISD::PINSRB";
5460 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5461 case X86ISD::FMAX: return "X86ISD::FMAX";
5462 case X86ISD::FMIN: return "X86ISD::FMIN";
5463 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5464 case X86ISD::FRCP: return "X86ISD::FRCP";
5465 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5466 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5467 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5468 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5469 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5470 case X86ISD::LCMPXCHG_DAG: return "x86ISD::LCMPXCHG_DAG";
5471 case X86ISD::LCMPXCHG8_DAG: return "x86ISD::LCMPXCHG8_DAG";
5475 // isLegalAddressingMode - Return true if the addressing mode represented
5476 // by AM is legal for this target, for a load/store of the specified type.
5477 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5478 const Type *Ty) const {
5479 // X86 supports extremely general addressing modes.
5481 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5482 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5486 // We can only fold this if we don't need an extra load.
5487 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5490 // X86-64 only supports addr of globals in small code model.
5491 if (Subtarget->is64Bit()) {
5492 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5494 // If lower 4G is not available, then we must use rip-relative addressing.
5495 if (AM.BaseOffs || AM.Scale > 1)
5506 // These scales always work.
5511 // These scales are formed with basereg+scalereg. Only accept if there is
5516 default: // Other stuff never works.
5524 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5525 if (!Ty1->isInteger() || !Ty2->isInteger())
5527 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5528 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5529 if (NumBits1 <= NumBits2)
5531 return Subtarget->is64Bit() || NumBits1 < 64;
5534 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5535 MVT::ValueType VT2) const {
5536 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5538 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5539 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5540 if (NumBits1 <= NumBits2)
5542 return Subtarget->is64Bit() || NumBits1 < 64;
5545 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5546 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5547 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5548 /// are assumed to be legal.
5550 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5551 // Only do shuffles on 128-bit vector types for now.
5552 if (MVT::getSizeInBits(VT) == 64) return false;
5553 return (Mask.Val->getNumOperands() <= 4 ||
5554 isIdentityMask(Mask.Val) ||
5555 isIdentityMask(Mask.Val, true) ||
5556 isSplatMask(Mask.Val) ||
5557 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5558 X86::isUNPCKLMask(Mask.Val) ||
5559 X86::isUNPCKHMask(Mask.Val) ||
5560 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5561 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5564 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5566 SelectionDAG &DAG) const {
5567 unsigned NumElts = BVOps.size();
5568 // Only do shuffles on 128-bit vector types for now.
5569 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5570 if (NumElts == 2) return true;
5572 return (isMOVLMask(&BVOps[0], 4) ||
5573 isCommutedMOVL(&BVOps[0], 4, true) ||
5574 isSHUFPMask(&BVOps[0], 4) ||
5575 isCommutedSHUFP(&BVOps[0], 4));
5580 //===----------------------------------------------------------------------===//
5581 // X86 Scheduler Hooks
5582 //===----------------------------------------------------------------------===//
5585 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5586 MachineBasicBlock *BB) {
5587 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5588 switch (MI->getOpcode()) {
5589 default: assert(false && "Unexpected instr type to insert");
5590 case X86::CMOV_FR32:
5591 case X86::CMOV_FR64:
5592 case X86::CMOV_V4F32:
5593 case X86::CMOV_V2F64:
5594 case X86::CMOV_V2I64: {
5595 // To "insert" a SELECT_CC instruction, we actually have to insert the
5596 // diamond control-flow pattern. The incoming instruction knows the
5597 // destination vreg to set, the condition code register to branch on, the
5598 // true/false values to select between, and a branch opcode to use.
5599 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5600 ilist<MachineBasicBlock>::iterator It = BB;
5606 // cmpTY ccX, r1, r2
5608 // fallthrough --> copy0MBB
5609 MachineBasicBlock *thisMBB = BB;
5610 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5611 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5613 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5614 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5615 MachineFunction *F = BB->getParent();
5616 F->getBasicBlockList().insert(It, copy0MBB);
5617 F->getBasicBlockList().insert(It, sinkMBB);
5618 // Update machine-CFG edges by first adding all successors of the current
5619 // block to the new block which will contain the Phi node for the select.
5620 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5621 e = BB->succ_end(); i != e; ++i)
5622 sinkMBB->addSuccessor(*i);
5623 // Next, remove all successors of the current block, and add the true
5624 // and fallthrough blocks as its successors.
5625 while(!BB->succ_empty())
5626 BB->removeSuccessor(BB->succ_begin());
5627 BB->addSuccessor(copy0MBB);
5628 BB->addSuccessor(sinkMBB);
5631 // %FalseValue = ...
5632 // # fallthrough to sinkMBB
5635 // Update machine-CFG edges
5636 BB->addSuccessor(sinkMBB);
5639 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5642 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5643 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5644 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5646 delete MI; // The pseudo instruction is gone now.
5650 case X86::FP32_TO_INT16_IN_MEM:
5651 case X86::FP32_TO_INT32_IN_MEM:
5652 case X86::FP32_TO_INT64_IN_MEM:
5653 case X86::FP64_TO_INT16_IN_MEM:
5654 case X86::FP64_TO_INT32_IN_MEM:
5655 case X86::FP64_TO_INT64_IN_MEM:
5656 case X86::FP80_TO_INT16_IN_MEM:
5657 case X86::FP80_TO_INT32_IN_MEM:
5658 case X86::FP80_TO_INT64_IN_MEM: {
5659 // Change the floating point control register to use "round towards zero"
5660 // mode when truncating to an integer value.
5661 MachineFunction *F = BB->getParent();
5662 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5663 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5665 // Load the old value of the high byte of the control word...
5667 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
5668 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5670 // Set the high part to be round to zero...
5671 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5674 // Reload the modified control word now...
5675 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5677 // Restore the memory image of control word to original value
5678 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5681 // Get the X86 opcode to use.
5683 switch (MI->getOpcode()) {
5684 default: assert(0 && "illegal opcode!");
5685 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5686 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5687 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5688 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5689 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5690 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5691 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5692 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5693 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5697 MachineOperand &Op = MI->getOperand(0);
5698 if (Op.isRegister()) {
5699 AM.BaseType = X86AddressMode::RegBase;
5700 AM.Base.Reg = Op.getReg();
5702 AM.BaseType = X86AddressMode::FrameIndexBase;
5703 AM.Base.FrameIndex = Op.getIndex();
5705 Op = MI->getOperand(1);
5706 if (Op.isImmediate())
5707 AM.Scale = Op.getImm();
5708 Op = MI->getOperand(2);
5709 if (Op.isImmediate())
5710 AM.IndexReg = Op.getImm();
5711 Op = MI->getOperand(3);
5712 if (Op.isGlobalAddress()) {
5713 AM.GV = Op.getGlobal();
5715 AM.Disp = Op.getImm();
5717 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5718 .addReg(MI->getOperand(4).getReg());
5720 // Reload the original control word now.
5721 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5723 delete MI; // The pseudo instruction is gone now.
5729 //===----------------------------------------------------------------------===//
5730 // X86 Optimization Hooks
5731 //===----------------------------------------------------------------------===//
5733 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5737 const SelectionDAG &DAG,
5738 unsigned Depth) const {
5739 unsigned Opc = Op.getOpcode();
5740 assert((Opc >= ISD::BUILTIN_OP_END ||
5741 Opc == ISD::INTRINSIC_WO_CHAIN ||
5742 Opc == ISD::INTRINSIC_W_CHAIN ||
5743 Opc == ISD::INTRINSIC_VOID) &&
5744 "Should use MaskedValueIsZero if you don't know whether Op"
5745 " is a target node!");
5747 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
5751 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
5752 Mask.getBitWidth() - 1);
5757 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5758 /// element of the result of the vector shuffle.
5759 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5760 MVT::ValueType VT = N->getValueType(0);
5761 SDOperand PermMask = N->getOperand(2);
5762 unsigned NumElems = PermMask.getNumOperands();
5763 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5765 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5767 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5768 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5769 SDOperand Idx = PermMask.getOperand(i);
5770 if (Idx.getOpcode() == ISD::UNDEF)
5771 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5772 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5777 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5778 /// node is a GlobalAddress + an offset.
5779 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5780 unsigned Opc = N->getOpcode();
5781 if (Opc == X86ISD::Wrapper) {
5782 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5783 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5786 } else if (Opc == ISD::ADD) {
5787 SDOperand N1 = N->getOperand(0);
5788 SDOperand N2 = N->getOperand(1);
5789 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5790 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5792 Offset += V->getSignExtended();
5795 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5796 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5798 Offset += V->getSignExtended();
5806 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5808 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5809 MachineFrameInfo *MFI) {
5810 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5813 SDOperand Loc = N->getOperand(1);
5814 SDOperand BaseLoc = Base->getOperand(1);
5815 if (Loc.getOpcode() == ISD::FrameIndex) {
5816 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5818 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5819 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5820 int FS = MFI->getObjectSize(FI);
5821 int BFS = MFI->getObjectSize(BFI);
5822 if (FS != BFS || FS != Size) return false;
5823 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5825 GlobalValue *GV1 = NULL;
5826 GlobalValue *GV2 = NULL;
5827 int64_t Offset1 = 0;
5828 int64_t Offset2 = 0;
5829 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5830 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5831 if (isGA1 && isGA2 && GV1 == GV2)
5832 return Offset1 == (Offset2 + Dist*Size);
5838 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5839 const X86Subtarget *Subtarget) {
5842 if (isGAPlusOffset(Base, GV, Offset))
5843 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5844 // DAG combine handles the stack object case.
5849 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5850 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5851 /// if the load addresses are consecutive, non-overlapping, and in the right
5853 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5854 const X86Subtarget *Subtarget) {
5855 MachineFunction &MF = DAG.getMachineFunction();
5856 MachineFrameInfo *MFI = MF.getFrameInfo();
5857 MVT::ValueType VT = N->getValueType(0);
5858 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5859 SDOperand PermMask = N->getOperand(2);
5860 int NumElems = (int)PermMask.getNumOperands();
5861 SDNode *Base = NULL;
5862 for (int i = 0; i < NumElems; ++i) {
5863 SDOperand Idx = PermMask.getOperand(i);
5864 if (Idx.getOpcode() == ISD::UNDEF) {
5865 if (!Base) return SDOperand();
5868 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5869 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5873 else if (!isConsecutiveLoad(Arg.Val, Base,
5874 i, MVT::getSizeInBits(EVT)/8,MFI))
5879 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5880 LoadSDNode *LD = cast<LoadSDNode>(Base);
5882 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5883 LD->getSrcValueOffset(), LD->isVolatile());
5885 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5886 LD->getSrcValueOffset(), LD->isVolatile(),
5887 LD->getAlignment());
5891 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5892 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5893 const X86Subtarget *Subtarget) {
5894 SDOperand Cond = N->getOperand(0);
5896 // If we have SSE[12] support, try to form min/max nodes.
5897 if (Subtarget->hasSSE2() &&
5898 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5899 if (Cond.getOpcode() == ISD::SETCC) {
5900 // Get the LHS/RHS of the select.
5901 SDOperand LHS = N->getOperand(1);
5902 SDOperand RHS = N->getOperand(2);
5903 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5905 unsigned Opcode = 0;
5906 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5909 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5912 if (!UnsafeFPMath) break;
5914 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5916 Opcode = X86ISD::FMIN;
5919 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5922 if (!UnsafeFPMath) break;
5924 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5926 Opcode = X86ISD::FMAX;
5929 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5932 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5935 if (!UnsafeFPMath) break;
5937 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5939 Opcode = X86ISD::FMIN;
5942 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5945 if (!UnsafeFPMath) break;
5947 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5949 Opcode = X86ISD::FMAX;
5955 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5963 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
5964 static SDOperand PerformSTORECombine(StoreSDNode *St, SelectionDAG &DAG,
5965 const X86Subtarget *Subtarget) {
5966 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
5967 // the FP state in cases where an emms may be missing.
5968 // A preferable solution to the general problem is to figure out the right
5969 // places to insert EMMS. This qualifies as a quick hack.
5970 if (MVT::isVector(St->getValue().getValueType()) &&
5971 MVT::getSizeInBits(St->getValue().getValueType()) == 64 &&
5972 isa<LoadSDNode>(St->getValue()) &&
5973 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
5974 St->getChain().hasOneUse() && !St->isVolatile()) {
5975 SDNode* LdVal = St->getValue().Val;
5977 int TokenFactorIndex = -1;
5978 SmallVector<SDOperand, 8> Ops;
5979 SDNode* ChainVal = St->getChain().Val;
5980 // Must be a store of a load. We currently handle two cases: the load
5981 // is a direct child, and it's under an intervening TokenFactor. It is
5982 // possible to dig deeper under nested TokenFactors.
5983 if (ChainVal == LdVal)
5984 Ld = cast<LoadSDNode>(St->getChain());
5985 else if (St->getValue().hasOneUse() &&
5986 ChainVal->getOpcode() == ISD::TokenFactor) {
5987 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
5988 if (ChainVal->getOperand(i).Val == LdVal) {
5989 TokenFactorIndex = i;
5990 Ld = cast<LoadSDNode>(St->getValue());
5992 Ops.push_back(ChainVal->getOperand(i));
5996 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
5997 if (Subtarget->is64Bit()) {
5998 SDOperand NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
5999 Ld->getBasePtr(), Ld->getSrcValue(),
6000 Ld->getSrcValueOffset(), Ld->isVolatile(),
6001 Ld->getAlignment());
6002 SDOperand NewChain = NewLd.getValue(1);
6003 if (TokenFactorIndex != -1) {
6004 Ops.push_back(NewLd);
6005 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6008 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6009 St->getSrcValue(), St->getSrcValueOffset(),
6010 St->isVolatile(), St->getAlignment());
6013 // Otherwise, lower to two 32-bit copies.
6014 SDOperand LoAddr = Ld->getBasePtr();
6015 SDOperand HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6016 DAG.getConstant(MVT::i32, 4));
6018 SDOperand LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6019 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6020 Ld->isVolatile(), Ld->getAlignment());
6021 SDOperand HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6022 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6024 MinAlign(Ld->getAlignment(), 4));
6026 SDOperand NewChain = LoLd.getValue(1);
6027 if (TokenFactorIndex != -1) {
6028 Ops.push_back(LoLd);
6029 Ops.push_back(HiLd);
6030 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6034 LoAddr = St->getBasePtr();
6035 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6036 DAG.getConstant(MVT::i32, 4));
6038 SDOperand LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6039 St->getSrcValue(), St->getSrcValueOffset(),
6040 St->isVolatile(), St->getAlignment());
6041 SDOperand HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6042 St->getSrcValue(), St->getSrcValueOffset()+4,
6044 MinAlign(St->getAlignment(), 4));
6045 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6051 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6052 /// X86ISD::FXOR nodes.
6053 static SDOperand PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6054 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6055 // F[X]OR(0.0, x) -> x
6056 // F[X]OR(x, 0.0) -> x
6057 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6058 if (C->getValueAPF().isPosZero())
6059 return N->getOperand(1);
6060 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6061 if (C->getValueAPF().isPosZero())
6062 return N->getOperand(0);
6066 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6067 static SDOperand PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6068 // FAND(0.0, x) -> 0.0
6069 // FAND(x, 0.0) -> 0.0
6070 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6071 if (C->getValueAPF().isPosZero())
6072 return N->getOperand(0);
6073 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6074 if (C->getValueAPF().isPosZero())
6075 return N->getOperand(1);
6080 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
6081 DAGCombinerInfo &DCI) const {
6082 SelectionDAG &DAG = DCI.DAG;
6083 switch (N->getOpcode()) {
6085 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, Subtarget);
6086 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
6088 return PerformSTORECombine(cast<StoreSDNode>(N), DAG, Subtarget);
6090 case X86ISD::FOR: return PerformFORCombine(N, DAG);
6091 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
6097 //===----------------------------------------------------------------------===//
6098 // X86 Inline Assembly Support
6099 //===----------------------------------------------------------------------===//
6101 /// getConstraintType - Given a constraint letter, return the type of
6102 /// constraint it is for this target.
6103 X86TargetLowering::ConstraintType
6104 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
6105 if (Constraint.size() == 1) {
6106 switch (Constraint[0]) {
6116 return C_RegisterClass;
6121 return TargetLowering::getConstraintType(Constraint);
6124 /// LowerXConstraint - try to replace an X constraint, which matches anything,
6125 /// with another that has more specific requirements based on the type of the
6126 /// corresponding operand.
6127 void X86TargetLowering::lowerXConstraint(MVT::ValueType ConstraintVT,
6128 std::string& s) const {
6129 if (MVT::isFloatingPoint(ConstraintVT)) {
6130 if (Subtarget->hasSSE2())
6132 else if (Subtarget->hasSSE1())
6137 return TargetLowering::lowerXConstraint(ConstraintVT, s);
6140 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6141 /// vector. If it is invalid, don't add anything to Ops.
6142 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
6144 std::vector<SDOperand>&Ops,
6145 SelectionDAG &DAG) {
6146 SDOperand Result(0, 0);
6148 switch (Constraint) {
6151 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6152 if (C->getValue() <= 31) {
6153 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6159 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6160 if (C->getValue() <= 255) {
6161 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6167 // Literal immediates are always ok.
6168 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
6169 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
6173 // If we are in non-pic codegen mode, we allow the address of a global (with
6174 // an optional displacement) to be used with 'i'.
6175 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
6178 // Match either (GA) or (GA+C)
6180 Offset = GA->getOffset();
6181 } else if (Op.getOpcode() == ISD::ADD) {
6182 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6183 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6185 Offset = GA->getOffset()+C->getValue();
6187 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6188 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6190 Offset = GA->getOffset()+C->getValue();
6197 // If addressing this global requires a load (e.g. in PIC mode), we can't
6199 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
6203 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
6209 // Otherwise, not valid for this mode.
6215 Ops.push_back(Result);
6218 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6221 std::vector<unsigned> X86TargetLowering::
6222 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6223 MVT::ValueType VT) const {
6224 if (Constraint.size() == 1) {
6225 // FIXME: not handling fp-stack yet!
6226 switch (Constraint[0]) { // GCC X86 Constraint Letters
6227 default: break; // Unknown constraint letter
6228 case 'A': // EAX/EDX
6229 if (VT == MVT::i32 || VT == MVT::i64)
6230 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
6232 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
6235 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
6236 else if (VT == MVT::i16)
6237 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
6238 else if (VT == MVT::i8)
6239 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
6240 else if (VT == MVT::i64)
6241 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
6246 return std::vector<unsigned>();
6249 std::pair<unsigned, const TargetRegisterClass*>
6250 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6251 MVT::ValueType VT) const {
6252 // First, see if this is a constraint that directly corresponds to an LLVM
6254 if (Constraint.size() == 1) {
6255 // GCC Constraint Letters
6256 switch (Constraint[0]) {
6258 case 'r': // GENERAL_REGS
6259 case 'R': // LEGACY_REGS
6260 case 'l': // INDEX_REGS
6261 if (VT == MVT::i64 && Subtarget->is64Bit())
6262 return std::make_pair(0U, X86::GR64RegisterClass);
6264 return std::make_pair(0U, X86::GR32RegisterClass);
6265 else if (VT == MVT::i16)
6266 return std::make_pair(0U, X86::GR16RegisterClass);
6267 else if (VT == MVT::i8)
6268 return std::make_pair(0U, X86::GR8RegisterClass);
6270 case 'f': // FP Stack registers.
6271 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
6272 // value to the correct fpstack register class.
6273 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
6274 return std::make_pair(0U, X86::RFP32RegisterClass);
6275 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
6276 return std::make_pair(0U, X86::RFP64RegisterClass);
6277 return std::make_pair(0U, X86::RFP80RegisterClass);
6278 case 'y': // MMX_REGS if MMX allowed.
6279 if (!Subtarget->hasMMX()) break;
6280 return std::make_pair(0U, X86::VR64RegisterClass);
6282 case 'Y': // SSE_REGS if SSE2 allowed
6283 if (!Subtarget->hasSSE2()) break;
6285 case 'x': // SSE_REGS if SSE1 allowed
6286 if (!Subtarget->hasSSE1()) break;
6290 // Scalar SSE types.
6293 return std::make_pair(0U, X86::FR32RegisterClass);
6296 return std::make_pair(0U, X86::FR64RegisterClass);
6304 return std::make_pair(0U, X86::VR128RegisterClass);
6310 // Use the default implementation in TargetLowering to convert the register
6311 // constraint into a member of a register class.
6312 std::pair<unsigned, const TargetRegisterClass*> Res;
6313 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6315 // Not found as a standard register?
6316 if (Res.second == 0) {
6317 // GCC calls "st(0)" just plain "st".
6318 if (StringsEqualNoCase("{st}", Constraint)) {
6319 Res.first = X86::ST0;
6320 Res.second = X86::RFP80RegisterClass;
6326 // Otherwise, check to see if this is a register class of the wrong value
6327 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6328 // turn into {ax},{dx}.
6329 if (Res.second->hasType(VT))
6330 return Res; // Correct type already, nothing to do.
6332 // All of the single-register GCC register classes map their values onto
6333 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6334 // really want an 8-bit or 32-bit register, map to the appropriate register
6335 // class and return the appropriate register.
6336 if (Res.second != X86::GR16RegisterClass)
6339 if (VT == MVT::i8) {
6340 unsigned DestReg = 0;
6341 switch (Res.first) {
6343 case X86::AX: DestReg = X86::AL; break;
6344 case X86::DX: DestReg = X86::DL; break;
6345 case X86::CX: DestReg = X86::CL; break;
6346 case X86::BX: DestReg = X86::BL; break;
6349 Res.first = DestReg;
6350 Res.second = Res.second = X86::GR8RegisterClass;
6352 } else if (VT == MVT::i32) {
6353 unsigned DestReg = 0;
6354 switch (Res.first) {
6356 case X86::AX: DestReg = X86::EAX; break;
6357 case X86::DX: DestReg = X86::EDX; break;
6358 case X86::CX: DestReg = X86::ECX; break;
6359 case X86::BX: DestReg = X86::EBX; break;
6360 case X86::SI: DestReg = X86::ESI; break;
6361 case X86::DI: DestReg = X86::EDI; break;
6362 case X86::BP: DestReg = X86::EBP; break;
6363 case X86::SP: DestReg = X86::ESP; break;
6366 Res.first = DestReg;
6367 Res.second = Res.second = X86::GR32RegisterClass;
6369 } else if (VT == MVT::i64) {
6370 unsigned DestReg = 0;
6371 switch (Res.first) {
6373 case X86::AX: DestReg = X86::RAX; break;
6374 case X86::DX: DestReg = X86::RDX; break;
6375 case X86::CX: DestReg = X86::RCX; break;
6376 case X86::BX: DestReg = X86::RBX; break;
6377 case X86::SI: DestReg = X86::RSI; break;
6378 case X86::DI: DestReg = X86::RDI; break;
6379 case X86::BP: DestReg = X86::RBP; break;
6380 case X86::SP: DestReg = X86::RSP; break;
6383 Res.first = DestReg;
6384 Res.second = Res.second = X86::GR64RegisterClass;