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"
42 #include "llvm/ParamAttrsList.h"
45 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
46 : TargetLowering(TM) {
47 Subtarget = &TM.getSubtarget<X86Subtarget>();
48 X86ScalarSSEf64 = Subtarget->hasSSE2();
49 X86ScalarSSEf32 = Subtarget->hasSSE1();
50 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
54 RegInfo = TM.getRegisterInfo();
56 // Set up the TargetLowering object.
58 // X86 is weird, it always uses i8 for shift amounts and setcc results.
59 setShiftAmountType(MVT::i8);
60 setSetCCResultType(MVT::i8);
61 setSetCCResultContents(ZeroOrOneSetCCResult);
62 setSchedulingPreference(SchedulingForRegPressure);
63 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
64 setStackPointerRegisterToSaveRestore(X86StackPtr);
66 if (Subtarget->isTargetDarwin()) {
67 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
68 setUseUnderscoreSetJmp(false);
69 setUseUnderscoreLongJmp(false);
70 } else if (Subtarget->isTargetMingw()) {
71 // MS runtime is weird: it exports _setjmp, but longjmp!
72 setUseUnderscoreSetJmp(true);
73 setUseUnderscoreLongJmp(false);
75 setUseUnderscoreSetJmp(true);
76 setUseUnderscoreLongJmp(true);
79 // Set up the register classes.
80 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
81 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
82 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
83 if (Subtarget->is64Bit())
84 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
86 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
88 // We don't accept any truncstore of integer registers.
89 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
90 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
92 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
93 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
94 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
96 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
98 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
99 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
102 if (Subtarget->is64Bit()) {
103 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
104 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
107 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
108 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
110 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
113 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
115 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
116 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
117 // SSE has no i16 to fp conversion, only i32
118 if (X86ScalarSSEf32) {
119 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
120 // f32 and f64 cases are Legal, f80 case is not
121 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
123 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
124 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
127 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
128 // are Legal, f80 is custom lowered.
129 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
130 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
132 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
134 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
135 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
137 if (X86ScalarSSEf32) {
138 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
139 // f32 and f64 cases are Legal, f80 case is not
140 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
142 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
143 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
146 // Handle FP_TO_UINT by promoting the destination to a larger signed
148 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
149 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
150 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
152 if (Subtarget->is64Bit()) {
153 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
154 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
156 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
157 // Expand FP_TO_UINT into a select.
158 // FIXME: We would like to use a Custom expander here eventually to do
159 // the optimal thing for SSE vs. the default expansion in the legalizer.
160 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
162 // With SSE3 we can use fisttpll to convert to a signed i64.
163 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
166 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
167 if (!X86ScalarSSEf64) {
168 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
169 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
172 // Scalar integer divide and remainder are lowered to use operations that
173 // produce two results, to match the available instructions. This exposes
174 // the two-result form to trivial CSE, which is able to combine x/y and x%y
175 // into a single instruction.
177 // Scalar integer multiply-high is also lowered to use two-result
178 // operations, to match the available instructions. However, plain multiply
179 // (low) operations are left as Legal, as there are single-result
180 // instructions for this in x86. Using the two-result multiply instructions
181 // when both high and low results are needed must be arranged by dagcombine.
182 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
183 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
184 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
185 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
186 setOperationAction(ISD::SREM , MVT::i8 , Expand);
187 setOperationAction(ISD::UREM , MVT::i8 , Expand);
188 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
189 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
190 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
191 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
192 setOperationAction(ISD::SREM , MVT::i16 , Expand);
193 setOperationAction(ISD::UREM , MVT::i16 , Expand);
194 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
195 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
196 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
197 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
198 setOperationAction(ISD::SREM , MVT::i32 , Expand);
199 setOperationAction(ISD::UREM , MVT::i32 , Expand);
200 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
201 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
202 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
203 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::SREM , MVT::i64 , Expand);
205 setOperationAction(ISD::UREM , MVT::i64 , Expand);
207 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
208 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
209 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
210 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
211 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
212 if (Subtarget->is64Bit())
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
216 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
217 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
218 setOperationAction(ISD::FREM , MVT::f64 , 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 // X86 wants to expand memset / memcpy itself.
280 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
281 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
283 if (!Subtarget->hasSSE2())
284 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
286 setOperationAction(ISD::ATOMIC_LCS , MVT::i8, Custom);
287 setOperationAction(ISD::ATOMIC_LCS , MVT::i16, Custom);
288 setOperationAction(ISD::ATOMIC_LCS , MVT::i32, Custom);
290 // Use the default ISD::LOCATION, ISD::DECLARE expansion.
291 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
292 // FIXME - use subtarget debug flags
293 if (!Subtarget->isTargetDarwin() &&
294 !Subtarget->isTargetELF() &&
295 !Subtarget->isTargetCygMing())
296 setOperationAction(ISD::LABEL, MVT::Other, Expand);
298 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
299 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
300 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
301 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
302 if (Subtarget->is64Bit()) {
304 setExceptionPointerRegister(X86::RAX);
305 setExceptionSelectorRegister(X86::RDX);
307 setExceptionPointerRegister(X86::EAX);
308 setExceptionSelectorRegister(X86::EDX);
310 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
312 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
314 setOperationAction(ISD::TRAP, MVT::Other, Legal);
316 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
317 setOperationAction(ISD::VASTART , MVT::Other, Custom);
318 setOperationAction(ISD::VAARG , MVT::Other, Expand);
319 setOperationAction(ISD::VAEND , MVT::Other, Expand);
320 if (Subtarget->is64Bit())
321 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
323 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
325 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
326 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
327 if (Subtarget->is64Bit())
328 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
329 if (Subtarget->isTargetCygMing())
330 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
332 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
334 if (X86ScalarSSEf64) {
335 // f32 and f64 use SSE.
336 // Set up the FP register classes.
337 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
338 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
340 // Use ANDPD to simulate FABS.
341 setOperationAction(ISD::FABS , MVT::f64, Custom);
342 setOperationAction(ISD::FABS , MVT::f32, Custom);
344 // Use XORP to simulate FNEG.
345 setOperationAction(ISD::FNEG , MVT::f64, Custom);
346 setOperationAction(ISD::FNEG , MVT::f32, Custom);
348 // Use ANDPD and ORPD to simulate FCOPYSIGN.
349 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
350 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
352 // We don't support sin/cos/fmod
353 setOperationAction(ISD::FSIN , MVT::f64, Expand);
354 setOperationAction(ISD::FCOS , MVT::f64, Expand);
355 setOperationAction(ISD::FREM , MVT::f64, Expand);
356 setOperationAction(ISD::FSIN , MVT::f32, Expand);
357 setOperationAction(ISD::FCOS , MVT::f32, Expand);
358 setOperationAction(ISD::FREM , MVT::f32, Expand);
360 // Expand FP immediates into loads from the stack, except for the special
362 addLegalFPImmediate(APFloat(+0.0)); // xorpd
363 addLegalFPImmediate(APFloat(+0.0f)); // xorps
365 // Floating truncations from f80 and extensions to f80 go through memory.
366 // If optimizing, we lie about this though and handle it in
367 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
369 setConvertAction(MVT::f32, MVT::f80, Expand);
370 setConvertAction(MVT::f64, MVT::f80, Expand);
371 setConvertAction(MVT::f80, MVT::f32, Expand);
372 setConvertAction(MVT::f80, MVT::f64, Expand);
374 } else if (X86ScalarSSEf32) {
375 // Use SSE for f32, x87 for f64.
376 // Set up the FP register classes.
377 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
378 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
380 // Use ANDPS to simulate FABS.
381 setOperationAction(ISD::FABS , MVT::f32, Custom);
383 // Use XORP to simulate FNEG.
384 setOperationAction(ISD::FNEG , MVT::f32, Custom);
386 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
388 // Use ANDPS and ORPS to simulate FCOPYSIGN.
389 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
390 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
392 // We don't support sin/cos/fmod
393 setOperationAction(ISD::FSIN , MVT::f32, Expand);
394 setOperationAction(ISD::FCOS , MVT::f32, Expand);
395 setOperationAction(ISD::FREM , MVT::f32, Expand);
397 // Special cases we handle for FP constants.
398 addLegalFPImmediate(APFloat(+0.0f)); // xorps
399 addLegalFPImmediate(APFloat(+0.0)); // FLD0
400 addLegalFPImmediate(APFloat(+1.0)); // FLD1
401 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
402 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
404 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
405 // this though and handle it in InstructionSelectPreprocess so that
406 // dagcombine2 can hack on these.
408 setConvertAction(MVT::f32, MVT::f64, Expand);
409 setConvertAction(MVT::f32, MVT::f80, Expand);
410 setConvertAction(MVT::f80, MVT::f32, Expand);
411 setConvertAction(MVT::f64, MVT::f32, Expand);
412 // And x87->x87 truncations also.
413 setConvertAction(MVT::f80, MVT::f64, Expand);
417 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
418 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
421 // f32 and f64 in x87.
422 // Set up the FP register classes.
423 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
424 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
426 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
427 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
428 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
429 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
431 // Floating truncations go through memory. If optimizing, we lie about
432 // this though and handle it in InstructionSelectPreprocess so that
433 // dagcombine2 can hack on these.
435 setConvertAction(MVT::f80, MVT::f32, Expand);
436 setConvertAction(MVT::f64, MVT::f32, Expand);
437 setConvertAction(MVT::f80, MVT::f64, Expand);
441 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
442 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
444 addLegalFPImmediate(APFloat(+0.0)); // FLD0
445 addLegalFPImmediate(APFloat(+1.0)); // FLD1
446 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
447 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
448 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
449 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
450 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
451 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
454 // Long double always uses X87.
455 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
456 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
457 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
459 APFloat TmpFlt(+0.0);
460 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
461 addLegalFPImmediate(TmpFlt); // FLD0
463 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
464 APFloat TmpFlt2(+1.0);
465 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
466 addLegalFPImmediate(TmpFlt2); // FLD1
467 TmpFlt2.changeSign();
468 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
472 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
473 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
476 // Always use a library call for pow.
477 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
478 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
479 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
481 // First set operation action for all vector types to expand. Then we
482 // will selectively turn on ones that can be effectively codegen'd.
483 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
484 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
485 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
486 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
487 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
488 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
489 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
490 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
491 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
492 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
493 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
494 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
495 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
496 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
497 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
498 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
499 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
500 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
501 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
502 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
503 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
504 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
505 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
506 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
507 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
508 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
509 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
510 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
511 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
512 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
513 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
514 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
515 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
516 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
517 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
518 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
519 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
520 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
521 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
524 if (Subtarget->hasMMX()) {
525 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
526 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
527 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
528 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
530 // FIXME: add MMX packed arithmetics
532 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
533 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
534 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
535 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
537 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
538 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
539 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
540 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
542 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
543 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
545 setOperationAction(ISD::AND, MVT::v8i8, Promote);
546 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
547 setOperationAction(ISD::AND, MVT::v4i16, Promote);
548 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
549 setOperationAction(ISD::AND, MVT::v2i32, Promote);
550 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
551 setOperationAction(ISD::AND, MVT::v1i64, Legal);
553 setOperationAction(ISD::OR, MVT::v8i8, Promote);
554 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
555 setOperationAction(ISD::OR, MVT::v4i16, Promote);
556 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
557 setOperationAction(ISD::OR, MVT::v2i32, Promote);
558 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
559 setOperationAction(ISD::OR, MVT::v1i64, Legal);
561 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
562 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
563 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
564 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
565 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
566 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
567 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
569 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
570 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
571 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
572 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
573 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
574 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
575 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
577 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
578 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
579 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
580 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
582 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
583 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
584 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
585 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
587 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
588 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
589 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
592 if (Subtarget->hasSSE1()) {
593 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
595 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
596 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
597 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
598 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
599 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
600 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
601 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
602 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
603 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
604 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
605 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
608 if (Subtarget->hasSSE2()) {
609 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
610 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
611 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
612 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
613 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
615 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
616 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
617 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
618 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
619 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
620 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
621 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
622 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
623 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
624 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
625 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
626 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
627 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
628 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
629 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
631 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
632 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
633 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
634 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
635 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
637 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
638 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
639 // Do not attempt to custom lower non-power-of-2 vectors
640 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
642 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
646 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
648 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
649 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
650 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
651 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
652 if (Subtarget->is64Bit()) {
653 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
654 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
657 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
658 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
659 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
660 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
661 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
662 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
663 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
664 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
665 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
666 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
667 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
668 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
671 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
673 // Custom lower v2i64 and v2f64 selects.
674 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
675 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
676 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
677 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
680 if (Subtarget->hasSSE41()) {
681 // FIXME: Do we need to handle scalar-to-vector here?
682 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
684 // i8 and i16 vectors are custom , because the source register and source
685 // source memory operand types are not the same width. f32 vectors are
686 // custom since the immediate controlling the insert encodes additional
688 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
689 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
690 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
691 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
693 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
694 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
695 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
696 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal);
698 if (Subtarget->is64Bit()) {
699 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
704 // We want to custom lower some of our intrinsics.
705 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
707 // We have target-specific dag combine patterns for the following nodes:
708 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
709 setTargetDAGCombine(ISD::SELECT);
710 setTargetDAGCombine(ISD::STORE);
712 computeRegisterProperties();
714 // FIXME: These should be based on subtarget info. Plus, the values should
715 // be smaller when we are in optimizing for size mode.
716 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
717 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
718 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
719 allowUnalignedMemoryAccesses = true; // x86 supports it!
720 setPrefLoopAlignment(16);
723 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
724 /// the desired ByVal argument alignment.
725 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
728 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
729 if (VTy->getBitWidth() == 128)
731 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
732 unsigned EltAlign = 0;
733 getMaxByValAlign(ATy->getElementType(), EltAlign);
734 if (EltAlign > MaxAlign)
736 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
737 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
738 unsigned EltAlign = 0;
739 getMaxByValAlign(STy->getElementType(i), EltAlign);
740 if (EltAlign > MaxAlign)
749 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
750 /// function arguments in the caller parameter area. For X86, aggregates
751 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
752 /// are at 4-byte boundaries.
753 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
754 if (Subtarget->is64Bit())
755 return getTargetData()->getABITypeAlignment(Ty);
757 if (Subtarget->hasSSE1())
758 getMaxByValAlign(Ty, Align);
762 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
764 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
765 SelectionDAG &DAG) const {
766 if (usesGlobalOffsetTable())
767 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
768 if (!Subtarget->isPICStyleRIPRel())
769 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
773 //===----------------------------------------------------------------------===//
774 // Return Value Calling Convention Implementation
775 //===----------------------------------------------------------------------===//
777 #include "X86GenCallingConv.inc"
779 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
780 /// exists skip possible ISD:TokenFactor.
781 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
782 if (Chain.getOpcode() == X86ISD::TAILCALL) {
784 } else if (Chain.getOpcode() == ISD::TokenFactor) {
785 if (Chain.getNumOperands() &&
786 Chain.getOperand(0).getOpcode() == X86ISD::TAILCALL)
787 return Chain.getOperand(0);
792 /// LowerRET - Lower an ISD::RET node.
793 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
794 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
796 SmallVector<CCValAssign, 16> RVLocs;
797 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
798 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
799 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
800 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
802 // If this is the first return lowered for this function, add the regs to the
803 // liveout set for the function.
804 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
805 for (unsigned i = 0; i != RVLocs.size(); ++i)
806 if (RVLocs[i].isRegLoc())
807 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
809 SDOperand Chain = Op.getOperand(0);
811 // Handle tail call return.
812 Chain = GetPossiblePreceedingTailCall(Chain);
813 if (Chain.getOpcode() == X86ISD::TAILCALL) {
814 SDOperand TailCall = Chain;
815 SDOperand TargetAddress = TailCall.getOperand(1);
816 SDOperand StackAdjustment = TailCall.getOperand(2);
817 assert(((TargetAddress.getOpcode() == ISD::Register &&
818 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
819 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
820 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
821 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
822 "Expecting an global address, external symbol, or register");
823 assert(StackAdjustment.getOpcode() == ISD::Constant &&
824 "Expecting a const value");
826 SmallVector<SDOperand,8> Operands;
827 Operands.push_back(Chain.getOperand(0));
828 Operands.push_back(TargetAddress);
829 Operands.push_back(StackAdjustment);
830 // Copy registers used by the call. Last operand is a flag so it is not
832 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
833 Operands.push_back(Chain.getOperand(i));
835 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
842 // Copy the result values into the output registers.
843 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
844 RVLocs[0].getLocReg() != X86::ST0) {
845 for (unsigned i = 0; i != RVLocs.size(); ++i) {
846 CCValAssign &VA = RVLocs[i];
847 assert(VA.isRegLoc() && "Can only return in registers!");
848 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
850 Flag = Chain.getValue(1);
853 // We need to handle a destination of ST0 specially, because it isn't really
855 SDOperand Value = Op.getOperand(1);
857 // an XMM register onto the fp-stack. Do this with an FP_EXTEND to f80.
858 // This will get legalized into a load/store if it can't get optimized away.
859 if (isScalarFPTypeInSSEReg(RVLocs[0].getValVT()))
860 Value = DAG.getNode(ISD::FP_EXTEND, MVT::f80, Value);
862 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
863 SDOperand Ops[] = { Chain, Value };
864 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
865 Flag = Chain.getValue(1);
868 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
870 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
872 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
876 /// LowerCallResult - Lower the result values of an ISD::CALL into the
877 /// appropriate copies out of appropriate physical registers. This assumes that
878 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
879 /// being lowered. The returns a SDNode with the same number of values as the
881 SDNode *X86TargetLowering::
882 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
883 unsigned CallingConv, SelectionDAG &DAG) {
885 // Assign locations to each value returned by this call.
886 SmallVector<CCValAssign, 16> RVLocs;
887 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
888 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
889 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
891 SmallVector<SDOperand, 8> ResultVals;
893 // Copy all of the result registers out of their specified physreg.
894 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
895 for (unsigned i = 0; i != RVLocs.size(); ++i) {
896 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
897 RVLocs[i].getValVT(), InFlag).getValue(1);
898 InFlag = Chain.getValue(2);
899 ResultVals.push_back(Chain.getValue(0));
902 // Copies from the FP stack are special, as ST0 isn't a valid register
903 // before the fp stackifier runs.
905 // Copy ST0 into an RFP register with FP_GET_RESULT. If this will end up
906 // in an SSE register, copy it out as F80 and do a truncate, otherwise use
907 // the specified value type.
908 MVT::ValueType GetResultTy = RVLocs[0].getValVT();
909 if (isScalarFPTypeInSSEReg(GetResultTy))
910 GetResultTy = MVT::f80;
911 SDVTList Tys = DAG.getVTList(GetResultTy, MVT::Other, MVT::Flag);
913 SDOperand GROps[] = { Chain, InFlag };
914 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
915 Chain = RetVal.getValue(1);
916 InFlag = RetVal.getValue(2);
918 // If we want the result in an SSE register, use an FP_TRUNCATE to get it
920 if (GetResultTy != RVLocs[0].getValVT())
921 RetVal = DAG.getNode(ISD::FP_ROUND, RVLocs[0].getValVT(), RetVal,
922 // This truncation won't change the value.
923 DAG.getIntPtrConstant(1));
925 ResultVals.push_back(RetVal);
928 // Merge everything together with a MERGE_VALUES node.
929 ResultVals.push_back(Chain);
930 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
931 &ResultVals[0], ResultVals.size()).Val;
934 /// LowerCallResultToTwo64BitRegs - Lower the result values of an x86-64
935 /// ISD::CALL where the results are known to be in two 64-bit registers,
936 /// e.g. XMM0 and XMM1. This simplify store the two values back to the
937 /// fixed stack slot allocated for StructRet.
938 SDNode *X86TargetLowering::
939 LowerCallResultToTwo64BitRegs(SDOperand Chain, SDOperand InFlag,
940 SDNode *TheCall, unsigned Reg1, unsigned Reg2,
941 MVT::ValueType VT, SelectionDAG &DAG) {
942 SDOperand RetVal1 = DAG.getCopyFromReg(Chain, Reg1, VT, InFlag);
943 Chain = RetVal1.getValue(1);
944 InFlag = RetVal1.getValue(2);
945 SDOperand RetVal2 = DAG.getCopyFromReg(Chain, Reg2, VT, InFlag);
946 Chain = RetVal2.getValue(1);
947 InFlag = RetVal2.getValue(2);
948 SDOperand FIN = TheCall->getOperand(5);
949 Chain = DAG.getStore(Chain, RetVal1, FIN, NULL, 0);
950 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
951 Chain = DAG.getStore(Chain, RetVal2, FIN, NULL, 0);
955 /// LowerCallResultToTwoX87Regs - Lower the result values of an x86-64 ISD::CALL
956 /// where the results are known to be in ST0 and ST1.
957 SDNode *X86TargetLowering::
958 LowerCallResultToTwoX87Regs(SDOperand Chain, SDOperand InFlag,
959 SDNode *TheCall, SelectionDAG &DAG) {
960 SmallVector<SDOperand, 8> ResultVals;
961 const MVT::ValueType VTs[] = { MVT::f80, MVT::f80, MVT::Other, MVT::Flag };
962 SDVTList Tys = DAG.getVTList(VTs, 4);
963 SDOperand Ops[] = { Chain, InFlag };
964 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT2, Tys, Ops, 2);
965 Chain = RetVal.getValue(2);
966 SDOperand FIN = TheCall->getOperand(5);
967 Chain = DAG.getStore(Chain, RetVal.getValue(1), FIN, NULL, 0);
968 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(16));
969 Chain = DAG.getStore(Chain, RetVal, FIN, NULL, 0);
973 //===----------------------------------------------------------------------===//
974 // C & StdCall & Fast Calling Convention implementation
975 //===----------------------------------------------------------------------===//
976 // StdCall calling convention seems to be standard for many Windows' API
977 // routines and around. It differs from C calling convention just a little:
978 // callee should clean up the stack, not caller. Symbols should be also
979 // decorated in some fancy way :) It doesn't support any vector arguments.
980 // For info on fast calling convention see Fast Calling Convention (tail call)
981 // implementation LowerX86_32FastCCCallTo.
983 /// AddLiveIn - This helper function adds the specified physical register to the
984 /// MachineFunction as a live in value. It also creates a corresponding virtual
986 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
987 const TargetRegisterClass *RC) {
988 assert(RC->contains(PReg) && "Not the correct regclass!");
989 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
990 MF.getRegInfo().addLiveIn(PReg, VReg);
994 /// CallIsStructReturn - Determines whether a CALL node uses struct return
996 static bool CallIsStructReturn(SDOperand Op) {
997 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1001 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(6));
1002 return Flags->getValue() & ISD::ParamFlags::StructReturn;
1005 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1006 /// return semantics.
1007 static bool ArgsAreStructReturn(SDOperand Op) {
1008 unsigned NumArgs = Op.Val->getNumValues() - 1;
1012 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(3));
1013 return Flags->getValue() & ISD::ParamFlags::StructReturn;
1016 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires the
1017 /// callee to pop its own arguments. Callee pop is necessary to support tail
1019 bool X86TargetLowering::IsCalleePop(SDOperand Op) {
1020 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1024 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1027 case CallingConv::X86_StdCall:
1028 return !Subtarget->is64Bit();
1029 case CallingConv::X86_FastCall:
1030 return !Subtarget->is64Bit();
1031 case CallingConv::Fast:
1032 return PerformTailCallOpt;
1036 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1037 /// FORMAL_ARGUMENTS node.
1038 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDOperand Op) const {
1039 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1041 if (Subtarget->is64Bit()) {
1042 if (CC == CallingConv::Fast && PerformTailCallOpt)
1043 return CC_X86_64_TailCall;
1048 if (CC == CallingConv::X86_FastCall)
1049 return CC_X86_32_FastCall;
1050 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1051 return CC_X86_32_TailCall;
1056 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1057 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1059 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDOperand Op) {
1060 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1061 if (CC == CallingConv::X86_FastCall)
1063 else if (CC == CallingConv::X86_StdCall)
1068 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
1069 /// possibly be overwritten when lowering the outgoing arguments in a tail
1070 /// call. Currently the implementation of this call is very conservative and
1071 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
1072 /// virtual registers would be overwritten by direct lowering.
1073 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op,
1074 MachineFrameInfo * MFI) {
1075 RegisterSDNode * OpReg = NULL;
1076 FrameIndexSDNode * FrameIdxNode = NULL;
1078 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
1079 (Op.getOpcode()== ISD::CopyFromReg &&
1080 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
1081 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
1082 (Op.getOpcode() == ISD::LOAD &&
1083 (FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op.getOperand(1))) &&
1084 (MFI->isFixedObjectIndex((FrameIdx = FrameIdxNode->getIndex()))) &&
1085 (MFI->getObjectOffset(FrameIdx) >= 0)))
1090 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1091 /// in a register before calling.
1092 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1093 return !IsTailCall && !Is64Bit &&
1094 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1095 Subtarget->isPICStyleGOT();
1099 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1100 /// address to be loaded in a register.
1102 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1103 return !Is64Bit && IsTailCall &&
1104 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1105 Subtarget->isPICStyleGOT();
1108 /// CopyTailCallClobberedArgumentsToVRegs - Create virtual registers for all
1109 /// arguments to force loading and guarantee that arguments sourcing from
1110 /// incomming parameters are not overwriting each other.
1112 CopyTailCallClobberedArgumentsToVRegs(SDOperand Chain,
1113 SmallVector<std::pair<unsigned, SDOperand>, 8> &TailCallClobberedVRegs,
1115 MachineFunction &MF,
1116 const TargetLowering * TL) {
1119 for (unsigned i = 0, e = TailCallClobberedVRegs.size(); i != e; i++) {
1120 SDOperand Arg = TailCallClobberedVRegs[i].second;
1121 unsigned Idx = TailCallClobberedVRegs[i].first;
1124 createVirtualRegister(TL->getRegClassFor(Arg.getValueType()));
1125 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
1126 InFlag = Chain.getValue(1);
1127 Arg = DAG.getCopyFromReg(Chain, VReg, Arg.getValueType(), InFlag);
1128 TailCallClobberedVRegs[i] = std::make_pair(Idx, Arg);
1129 Chain = Arg.getValue(1);
1130 InFlag = Arg.getValue(2);
1135 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1136 /// by "Src" to address "Dst" with size and alignment information specified by
1137 /// the specific parameter attribute. The copy will be passed as a byval function
1140 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1141 unsigned Flags, SelectionDAG &DAG) {
1142 unsigned Align = 1 <<
1143 ((Flags & ISD::ParamFlags::ByValAlign) >> ISD::ParamFlags::ByValAlignOffs);
1144 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1145 ISD::ParamFlags::ByValSizeOffs;
1146 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1147 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1148 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1149 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1152 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
1153 const CCValAssign &VA,
1154 MachineFrameInfo *MFI,
1156 SDOperand Root, unsigned i) {
1157 // Create the nodes corresponding to a load from this parameter slot.
1158 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
1159 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1160 bool isByVal = Flags & ISD::ParamFlags::ByVal;
1161 bool isImmutable = !AlwaysUseMutable && !isByVal;
1163 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1164 // changed with more analysis.
1165 // In case of tail call optimization mark all arguments mutable. Since they
1166 // could be overwritten by lowering of arguments in case of a tail call.
1167 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1168 VA.getLocMemOffset(), isImmutable);
1169 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1172 return DAG.getLoad(VA.getValVT(), Root, FIN,
1173 PseudoSourceValue::getFixedStack(), FI);
1177 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
1178 MachineFunction &MF = DAG.getMachineFunction();
1179 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1181 const Function* Fn = MF.getFunction();
1182 if (Fn->hasExternalLinkage() &&
1183 Subtarget->isTargetCygMing() &&
1184 Fn->getName() == "main")
1185 FuncInfo->setForceFramePointer(true);
1187 // Decorate the function name.
1188 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1190 MachineFrameInfo *MFI = MF.getFrameInfo();
1191 SDOperand Root = Op.getOperand(0);
1192 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1193 unsigned CC = MF.getFunction()->getCallingConv();
1194 bool Is64Bit = Subtarget->is64Bit();
1196 assert(!(isVarArg && CC == CallingConv::Fast) &&
1197 "Var args not supported with calling convention fastcc");
1199 // Assign locations to all of the incoming arguments.
1200 SmallVector<CCValAssign, 16> ArgLocs;
1201 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1202 CCInfo.AnalyzeFormalArguments(Op.Val, CCAssignFnForNode(Op));
1204 SmallVector<SDOperand, 8> ArgValues;
1205 unsigned LastVal = ~0U;
1206 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1207 CCValAssign &VA = ArgLocs[i];
1208 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1210 assert(VA.getValNo() != LastVal &&
1211 "Don't support value assigned to multiple locs yet");
1212 LastVal = VA.getValNo();
1214 if (VA.isRegLoc()) {
1215 MVT::ValueType RegVT = VA.getLocVT();
1216 TargetRegisterClass *RC;
1217 if (RegVT == MVT::i32)
1218 RC = X86::GR32RegisterClass;
1219 else if (Is64Bit && RegVT == MVT::i64)
1220 RC = X86::GR64RegisterClass;
1221 else if (RegVT == MVT::f32)
1222 RC = X86::FR32RegisterClass;
1223 else if (RegVT == MVT::f64)
1224 RC = X86::FR64RegisterClass;
1226 assert(MVT::isVector(RegVT));
1227 if (Is64Bit && MVT::getSizeInBits(RegVT) == 64) {
1228 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1231 RC = X86::VR128RegisterClass;
1234 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1235 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1237 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1238 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1240 if (VA.getLocInfo() == CCValAssign::SExt)
1241 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1242 DAG.getValueType(VA.getValVT()));
1243 else if (VA.getLocInfo() == CCValAssign::ZExt)
1244 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1245 DAG.getValueType(VA.getValVT()));
1247 if (VA.getLocInfo() != CCValAssign::Full)
1248 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1250 // Handle MMX values passed in GPRs.
1251 if (Is64Bit && RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1252 MVT::getSizeInBits(RegVT) == 64)
1253 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1255 ArgValues.push_back(ArgValue);
1257 assert(VA.isMemLoc());
1258 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1262 unsigned StackSize = CCInfo.getNextStackOffset();
1263 // align stack specially for tail calls
1264 if (CC == CallingConv::Fast)
1265 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1267 // If the function takes variable number of arguments, make a frame index for
1268 // the start of the first vararg value... for expansion of llvm.va_start.
1270 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1271 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1274 static const unsigned GPR64ArgRegs[] = {
1275 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1277 static const unsigned XMMArgRegs[] = {
1278 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1279 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1282 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1283 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1285 // For X86-64, if there are vararg parameters that are passed via
1286 // registers, then we must store them to their spots on the stack so they
1287 // may be loaded by deferencing the result of va_next.
1288 VarArgsGPOffset = NumIntRegs * 8;
1289 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1290 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1292 // Store the integer parameter registers.
1293 SmallVector<SDOperand, 8> MemOps;
1294 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1295 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1296 DAG.getIntPtrConstant(VarArgsGPOffset));
1297 for (; NumIntRegs != 6; ++NumIntRegs) {
1298 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1299 X86::GR64RegisterClass);
1300 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1302 DAG.getStore(Val.getValue(1), Val, FIN,
1303 PseudoSourceValue::getFixedStack(),
1305 MemOps.push_back(Store);
1306 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1307 DAG.getIntPtrConstant(8));
1310 // Now store the XMM (fp + vector) parameter registers.
1311 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1312 DAG.getIntPtrConstant(VarArgsFPOffset));
1313 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1314 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1315 X86::VR128RegisterClass);
1316 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1318 DAG.getStore(Val.getValue(1), Val, FIN,
1319 PseudoSourceValue::getFixedStack(),
1321 MemOps.push_back(Store);
1322 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1323 DAG.getIntPtrConstant(16));
1325 if (!MemOps.empty())
1326 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1327 &MemOps[0], MemOps.size());
1331 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1332 // arguments and the arguments after the retaddr has been pushed are
1334 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1335 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1336 (StackSize & 7) == 0)
1339 ArgValues.push_back(Root);
1341 // Some CCs need callee pop.
1342 if (IsCalleePop(Op)) {
1343 BytesToPopOnReturn = StackSize; // Callee pops everything.
1344 BytesCallerReserves = 0;
1346 BytesToPopOnReturn = 0; // Callee pops nothing.
1347 // If this is an sret function, the return should pop the hidden pointer.
1348 if (!Is64Bit && ArgsAreStructReturn(Op))
1349 BytesToPopOnReturn = 4;
1350 BytesCallerReserves = StackSize;
1354 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1355 if (CC == CallingConv::X86_FastCall)
1356 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1359 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1361 // Return the new list of results.
1362 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1363 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1367 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1368 const SDOperand &StackPtr,
1369 const CCValAssign &VA,
1372 unsigned LocMemOffset = VA.getLocMemOffset();
1373 SDOperand PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1374 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1375 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1376 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1377 if (Flags & ISD::ParamFlags::ByVal) {
1378 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1380 return DAG.getStore(Chain, Arg, PtrOff,
1381 PseudoSourceValue::getStack(), LocMemOffset);
1384 /// ClassifyX86_64SRetCallReturn - Classify how to implement a x86-64
1385 /// struct return call to the specified function. X86-64 ABI specifies
1386 /// some SRet calls are actually returned in registers. Since current
1387 /// LLVM cannot represent multi-value calls, they are represent as
1388 /// calls where the results are passed in a hidden struct provided by
1389 /// the caller. This function examines the type of the struct to
1390 /// determine the correct way to implement the call.
1392 X86TargetLowering::ClassifyX86_64SRetCallReturn(const Function *Fn) {
1393 // FIXME: Disabled for now.
1394 return X86::InMemory;
1396 const PointerType *PTy = cast<PointerType>(Fn->arg_begin()->getType());
1397 const Type *RTy = PTy->getElementType();
1398 unsigned Size = getTargetData()->getABITypeSize(RTy);
1399 if (Size != 16 && Size != 32)
1400 return X86::InMemory;
1403 const StructType *STy = dyn_cast<StructType>(RTy);
1404 if (!STy) return X86::InMemory;
1405 if (STy->getNumElements() == 2 &&
1406 STy->getElementType(0) == Type::X86_FP80Ty &&
1407 STy->getElementType(1) == Type::X86_FP80Ty)
1412 for (Type::subtype_iterator I = RTy->subtype_begin(), E = RTy->subtype_end();
1414 const Type *STy = I->get();
1415 if (!STy->isFPOrFPVector()) {
1423 return X86::InGPR64;
1426 void X86TargetLowering::X86_64AnalyzeSRetCallOperands(SDNode *TheCall,
1429 unsigned NumOps = (TheCall->getNumOperands() - 5) / 2;
1430 for (unsigned i = 1; i != NumOps; ++i) {
1431 MVT::ValueType ArgVT = TheCall->getOperand(5+2*i).getValueType();
1432 SDOperand FlagOp = TheCall->getOperand(5+2*i+1);
1433 unsigned ArgFlags =cast<ConstantSDNode>(FlagOp)->getValue();
1434 if (Fn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo)) {
1435 cerr << "Call operand #" << i << " has unhandled type "
1436 << MVT::getValueTypeString(ArgVT) << "\n";
1442 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 MachineFrameInfo * MFI = MF.getFrameInfo();
1445 SDOperand Chain = Op.getOperand(0);
1446 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1447 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1448 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1449 && CC == CallingConv::Fast && PerformTailCallOpt;
1450 SDOperand Callee = Op.getOperand(4);
1451 bool Is64Bit = Subtarget->is64Bit();
1452 bool IsStructRet = CallIsStructReturn(Op);
1454 assert(!(isVarArg && CC == CallingConv::Fast) &&
1455 "Var args not supported with calling convention fastcc");
1457 // Analyze operands of the call, assigning locations to each operand.
1458 SmallVector<CCValAssign, 16> ArgLocs;
1459 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1460 CCAssignFn *CCFn = CCAssignFnForNode(Op);
1462 X86::X86_64SRet SRetMethod = X86::InMemory;
1463 if (Is64Bit && IsStructRet)
1464 // FIXME: We can't figure out type of the sret structure for indirect
1465 // calls. We need to copy more information from CallSite to the ISD::CALL
1467 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1469 ClassifyX86_64SRetCallReturn(dyn_cast<Function>(G->getGlobal()));
1471 // UGLY HACK! For x86-64, some 128-bit aggregates are returns in a pair of
1472 // registers. Unfortunately, llvm does not support i128 yet so we pretend it's
1474 if (SRetMethod != X86::InMemory)
1475 X86_64AnalyzeSRetCallOperands(Op.Val, CCFn, CCInfo);
1477 CCInfo.AnalyzeCallOperands(Op.Val, CCFn);
1479 // Get a count of how many bytes are to be pushed on the stack.
1480 unsigned NumBytes = CCInfo.getNextStackOffset();
1481 if (CC == CallingConv::Fast)
1482 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1484 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1485 // arguments and the arguments after the retaddr has been pushed are aligned.
1486 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1487 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1488 (NumBytes & 7) == 0)
1493 // Lower arguments at fp - stackoffset + fpdiff.
1494 unsigned NumBytesCallerPushed =
1495 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1496 FPDiff = NumBytesCallerPushed - NumBytes;
1498 // Set the delta of movement of the returnaddr stackslot.
1499 // But only set if delta is greater than previous delta.
1500 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1501 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1504 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1506 SDOperand RetAddrFrIdx;
1508 // Adjust the Return address stack slot.
1510 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1511 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1512 // Load the "old" Return address.
1514 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1515 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1519 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1520 SmallVector<std::pair<unsigned, SDOperand>, 8> TailCallClobberedVRegs;
1521 SmallVector<SDOperand, 8> MemOpChains;
1525 // Walk the register/memloc assignments, inserting copies/loads. For tail
1526 // calls, remember all arguments for later special lowering.
1527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1528 CCValAssign &VA = ArgLocs[i];
1529 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1531 // Promote the value if needed.
1532 switch (VA.getLocInfo()) {
1533 default: assert(0 && "Unknown loc info!");
1534 case CCValAssign::Full: break;
1535 case CCValAssign::SExt:
1536 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1538 case CCValAssign::ZExt:
1539 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1541 case CCValAssign::AExt:
1542 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1546 if (VA.isRegLoc()) {
1547 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1550 assert(VA.isMemLoc());
1551 if (StackPtr.Val == 0)
1552 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1554 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1556 } else if (IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
1557 TailCallClobberedVRegs.push_back(std::make_pair(i,Arg));
1562 if (!MemOpChains.empty())
1563 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1564 &MemOpChains[0], MemOpChains.size());
1566 // Build a sequence of copy-to-reg nodes chained together with token chain
1567 // and flag operands which copy the outgoing args into registers.
1569 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1570 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1572 InFlag = Chain.getValue(1);
1575 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1577 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1578 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1579 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1581 InFlag = Chain.getValue(1);
1583 // If we are tail calling and generating PIC/GOT style code load the address
1584 // of the callee into ecx. The value in ecx is used as target of the tail
1585 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1586 // calls on PIC/GOT architectures. Normally we would just put the address of
1587 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1588 // restored (since ebx is callee saved) before jumping to the target@PLT.
1589 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1590 // Note: The actual moving to ecx is done further down.
1591 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1592 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1593 !G->getGlobal()->hasProtectedVisibility())
1594 Callee = LowerGlobalAddress(Callee, DAG);
1595 else if (isa<ExternalSymbolSDNode>(Callee))
1596 Callee = LowerExternalSymbol(Callee,DAG);
1599 if (Is64Bit && isVarArg) {
1600 // From AMD64 ABI document:
1601 // For calls that may call functions that use varargs or stdargs
1602 // (prototype-less calls or calls to functions containing ellipsis (...) in
1603 // the declaration) %al is used as hidden argument to specify the number
1604 // of SSE registers used. The contents of %al do not need to match exactly
1605 // the number of registers, but must be an ubound on the number of SSE
1606 // registers used and is in the range 0 - 8 inclusive.
1608 // Count the number of XMM registers allocated.
1609 static const unsigned XMMArgRegs[] = {
1610 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1611 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1613 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1615 Chain = DAG.getCopyToReg(Chain, X86::AL,
1616 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1617 InFlag = Chain.getValue(1);
1621 // For tail calls lower the arguments to the 'real' stack slot.
1623 SmallVector<SDOperand, 8> MemOpChains2;
1626 // Do not flag preceeding copytoreg stuff together with the following stuff.
1627 InFlag = SDOperand();
1629 Chain = CopyTailCallClobberedArgumentsToVRegs(Chain, TailCallClobberedVRegs,
1632 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1633 CCValAssign &VA = ArgLocs[i];
1634 if (!VA.isRegLoc()) {
1635 assert(VA.isMemLoc());
1636 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1637 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1638 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1639 // Create frame index.
1640 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1641 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1642 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1643 FIN = DAG.getFrameIndex(FI, MVT::i32);
1645 // Find virtual register for this argument.
1647 for (unsigned idx=0, e= TailCallClobberedVRegs.size(); idx < e; idx++)
1648 if (TailCallClobberedVRegs[idx].first==i) {
1649 Arg = TailCallClobberedVRegs[idx].second;
1653 assert(IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)==false ||
1654 (Found==true && "No corresponding Argument was found"));
1656 if (Flags & ISD::ParamFlags::ByVal) {
1657 // Copy relative to framepointer.
1658 MemOpChains2.push_back(CreateCopyOfByValArgument(Arg, FIN, Chain,
1661 // Store relative to framepointer.
1662 MemOpChains2.push_back(
1663 DAG.getStore(Chain, Arg, FIN,
1664 PseudoSourceValue::getFixedStack(), FI));
1669 if (!MemOpChains2.empty())
1670 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1671 &MemOpChains2[0], MemOpChains2.size());
1673 // Store the return address to the appropriate stack slot.
1675 // Calculate the new stack slot for the return address.
1676 int SlotSize = Is64Bit ? 8 : 4;
1677 int NewReturnAddrFI =
1678 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1679 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1680 SDOperand NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1681 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1682 PseudoSourceValue::getFixedStack(), NewReturnAddrFI);
1686 // If the callee is a GlobalAddress node (quite common, every direct call is)
1687 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1688 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1689 // We should use extra load for direct calls to dllimported functions in
1691 if ((IsTailCall || !Is64Bit ||
1692 getTargetMachine().getCodeModel() != CodeModel::Large)
1693 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1694 getTargetMachine(), true))
1695 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1696 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1697 if (IsTailCall || !Is64Bit ||
1698 getTargetMachine().getCodeModel() != CodeModel::Large)
1699 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1700 } else if (IsTailCall) {
1701 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1703 Chain = DAG.getCopyToReg(Chain,
1704 DAG.getRegister(Opc, getPointerTy()),
1706 Callee = DAG.getRegister(Opc, getPointerTy());
1707 // Add register as live out.
1708 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1711 // Returns a chain & a flag for retval copy to use.
1712 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1713 SmallVector<SDOperand, 8> Ops;
1716 Ops.push_back(Chain);
1717 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1718 Ops.push_back(DAG.getIntPtrConstant(0));
1720 Ops.push_back(InFlag);
1721 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1722 InFlag = Chain.getValue(1);
1724 // Returns a chain & a flag for retval copy to use.
1725 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1729 Ops.push_back(Chain);
1730 Ops.push_back(Callee);
1733 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1735 // Add an implicit use GOT pointer in EBX.
1736 if (!IsTailCall && !Is64Bit &&
1737 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1738 Subtarget->isPICStyleGOT())
1739 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1741 // Add argument registers to the end of the list so that they are known live
1743 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1744 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1745 RegsToPass[i].second.getValueType()));
1748 Ops.push_back(InFlag);
1751 assert(InFlag.Val &&
1752 "Flag must be set. Depend on flag being set in LowerRET");
1753 Chain = DAG.getNode(X86ISD::TAILCALL,
1754 Op.Val->getVTList(), &Ops[0], Ops.size());
1756 return SDOperand(Chain.Val, Op.ResNo);
1759 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1760 InFlag = Chain.getValue(1);
1762 // Create the CALLSEQ_END node.
1763 unsigned NumBytesForCalleeToPush;
1764 if (IsCalleePop(Op))
1765 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1766 else if (!Is64Bit && IsStructRet)
1767 // If this is is a call to a struct-return function, the callee
1768 // pops the hidden struct pointer, so we have to push it back.
1769 // This is common for Darwin/X86, Linux & Mingw32 targets.
1770 NumBytesForCalleeToPush = 4;
1772 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1774 // Returns a flag for retval copy to use.
1775 Chain = DAG.getCALLSEQ_END(Chain,
1776 DAG.getIntPtrConstant(NumBytes),
1777 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1779 InFlag = Chain.getValue(1);
1781 // Handle result values, copying them out of physregs into vregs that we
1783 switch (SRetMethod) {
1785 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1787 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1789 MVT::i64, DAG), Op.ResNo);
1791 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1792 X86::XMM0, X86::XMM1,
1793 MVT::f64, DAG), Op.ResNo);
1795 return SDOperand(LowerCallResultToTwoX87Regs(Chain, InFlag, Op.Val, DAG),
1801 //===----------------------------------------------------------------------===//
1802 // Fast Calling Convention (tail call) implementation
1803 //===----------------------------------------------------------------------===//
1805 // Like std call, callee cleans arguments, convention except that ECX is
1806 // reserved for storing the tail called function address. Only 2 registers are
1807 // free for argument passing (inreg). Tail call optimization is performed
1809 // * tailcallopt is enabled
1810 // * caller/callee are fastcc
1811 // On X86_64 architecture with GOT-style position independent code only local
1812 // (within module) calls are supported at the moment.
1813 // To keep the stack aligned according to platform abi the function
1814 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1815 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1816 // If a tail called function callee has more arguments than the caller the
1817 // caller needs to make sure that there is room to move the RETADDR to. This is
1818 // achieved by reserving an area the size of the argument delta right after the
1819 // original REtADDR, but before the saved framepointer or the spilled registers
1820 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1832 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1833 /// for a 16 byte align requirement.
1834 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1835 SelectionDAG& DAG) {
1836 if (PerformTailCallOpt) {
1837 MachineFunction &MF = DAG.getMachineFunction();
1838 const TargetMachine &TM = MF.getTarget();
1839 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1840 unsigned StackAlignment = TFI.getStackAlignment();
1841 uint64_t AlignMask = StackAlignment - 1;
1842 int64_t Offset = StackSize;
1843 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1844 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1845 // Number smaller than 12 so just add the difference.
1846 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1848 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1849 Offset = ((~AlignMask) & Offset) + StackAlignment +
1850 (StackAlignment-SlotSize);
1857 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1858 /// following the call is a return. A function is eligible if caller/callee
1859 /// calling conventions match, currently only fastcc supports tail calls, and
1860 /// the function CALL is immediatly followed by a RET.
1861 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1863 SelectionDAG& DAG) const {
1864 if (!PerformTailCallOpt)
1867 // Check whether CALL node immediatly preceeds the RET node and whether the
1868 // return uses the result of the node or is a void return.
1869 unsigned NumOps = Ret.getNumOperands();
1871 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1872 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1874 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1875 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1876 MachineFunction &MF = DAG.getMachineFunction();
1877 unsigned CallerCC = MF.getFunction()->getCallingConv();
1878 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1879 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1880 SDOperand Callee = Call.getOperand(4);
1881 // On x86/32Bit PIC/GOT tail calls are supported.
1882 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1883 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1886 // Can only do local tail calls (in same module, hidden or protected) on
1887 // x86_64 PIC/GOT at the moment.
1888 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1889 return G->getGlobal()->hasHiddenVisibility()
1890 || G->getGlobal()->hasProtectedVisibility();
1897 //===----------------------------------------------------------------------===//
1898 // Other Lowering Hooks
1899 //===----------------------------------------------------------------------===//
1902 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1903 MachineFunction &MF = DAG.getMachineFunction();
1904 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1905 int ReturnAddrIndex = FuncInfo->getRAIndex();
1907 if (ReturnAddrIndex == 0) {
1908 // Set up a frame object for the return address.
1909 if (Subtarget->is64Bit())
1910 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1912 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1914 FuncInfo->setRAIndex(ReturnAddrIndex);
1917 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1922 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1923 /// specific condition code. It returns a false if it cannot do a direct
1924 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1926 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1927 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1928 SelectionDAG &DAG) {
1929 X86CC = X86::COND_INVALID;
1931 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1932 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1933 // X > -1 -> X == 0, jump !sign.
1934 RHS = DAG.getConstant(0, RHS.getValueType());
1935 X86CC = X86::COND_NS;
1937 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1938 // X < 0 -> X == 0, jump on sign.
1939 X86CC = X86::COND_S;
1941 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1943 RHS = DAG.getConstant(0, RHS.getValueType());
1944 X86CC = X86::COND_LE;
1949 switch (SetCCOpcode) {
1951 case ISD::SETEQ: X86CC = X86::COND_E; break;
1952 case ISD::SETGT: X86CC = X86::COND_G; break;
1953 case ISD::SETGE: X86CC = X86::COND_GE; break;
1954 case ISD::SETLT: X86CC = X86::COND_L; break;
1955 case ISD::SETLE: X86CC = X86::COND_LE; break;
1956 case ISD::SETNE: X86CC = X86::COND_NE; break;
1957 case ISD::SETULT: X86CC = X86::COND_B; break;
1958 case ISD::SETUGT: X86CC = X86::COND_A; break;
1959 case ISD::SETULE: X86CC = X86::COND_BE; break;
1960 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1963 // On a floating point condition, the flags are set as follows:
1965 // 0 | 0 | 0 | X > Y
1966 // 0 | 0 | 1 | X < Y
1967 // 1 | 0 | 0 | X == Y
1968 // 1 | 1 | 1 | unordered
1970 switch (SetCCOpcode) {
1973 case ISD::SETEQ: X86CC = X86::COND_E; break;
1974 case ISD::SETOLT: Flip = true; // Fallthrough
1976 case ISD::SETGT: X86CC = X86::COND_A; break;
1977 case ISD::SETOLE: Flip = true; // Fallthrough
1979 case ISD::SETGE: X86CC = X86::COND_AE; break;
1980 case ISD::SETUGT: Flip = true; // Fallthrough
1982 case ISD::SETLT: X86CC = X86::COND_B; break;
1983 case ISD::SETUGE: Flip = true; // Fallthrough
1985 case ISD::SETLE: X86CC = X86::COND_BE; break;
1987 case ISD::SETNE: X86CC = X86::COND_NE; break;
1988 case ISD::SETUO: X86CC = X86::COND_P; break;
1989 case ISD::SETO: X86CC = X86::COND_NP; break;
1992 std::swap(LHS, RHS);
1995 return X86CC != X86::COND_INVALID;
1998 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1999 /// code. Current x86 isa includes the following FP cmov instructions:
2000 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2001 static bool hasFPCMov(unsigned X86CC) {
2017 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2018 /// true if Op is undef or if its value falls within the specified range (L, H].
2019 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2020 if (Op.getOpcode() == ISD::UNDEF)
2023 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2024 return (Val >= Low && Val < Hi);
2027 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2028 /// true if Op is undef or if its value equal to the specified value.
2029 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2030 if (Op.getOpcode() == ISD::UNDEF)
2032 return cast<ConstantSDNode>(Op)->getValue() == Val;
2035 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2036 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2037 bool X86::isPSHUFDMask(SDNode *N) {
2038 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2040 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2043 // Check if the value doesn't reference the second vector.
2044 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2045 SDOperand Arg = N->getOperand(i);
2046 if (Arg.getOpcode() == ISD::UNDEF) continue;
2047 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2048 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2055 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2056 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2057 bool X86::isPSHUFHWMask(SDNode *N) {
2058 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2060 if (N->getNumOperands() != 8)
2063 // Lower quadword copied in order.
2064 for (unsigned i = 0; i != 4; ++i) {
2065 SDOperand Arg = N->getOperand(i);
2066 if (Arg.getOpcode() == ISD::UNDEF) continue;
2067 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2068 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2072 // Upper quadword shuffled.
2073 for (unsigned i = 4; i != 8; ++i) {
2074 SDOperand Arg = N->getOperand(i);
2075 if (Arg.getOpcode() == ISD::UNDEF) continue;
2076 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2077 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2078 if (Val < 4 || Val > 7)
2085 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2086 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2087 bool X86::isPSHUFLWMask(SDNode *N) {
2088 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2090 if (N->getNumOperands() != 8)
2093 // Upper quadword copied in order.
2094 for (unsigned i = 4; i != 8; ++i)
2095 if (!isUndefOrEqual(N->getOperand(i), i))
2098 // Lower quadword shuffled.
2099 for (unsigned i = 0; i != 4; ++i)
2100 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2106 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2107 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2108 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2109 if (NumElems != 2 && NumElems != 4) return false;
2111 unsigned Half = NumElems / 2;
2112 for (unsigned i = 0; i < Half; ++i)
2113 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2115 for (unsigned i = Half; i < NumElems; ++i)
2116 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2122 bool X86::isSHUFPMask(SDNode *N) {
2123 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2124 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2127 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2128 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2129 /// half elements to come from vector 1 (which would equal the dest.) and
2130 /// the upper half to come from vector 2.
2131 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2132 if (NumOps != 2 && NumOps != 4) return false;
2134 unsigned Half = NumOps / 2;
2135 for (unsigned i = 0; i < Half; ++i)
2136 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2138 for (unsigned i = Half; i < NumOps; ++i)
2139 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2144 static bool isCommutedSHUFP(SDNode *N) {
2145 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2146 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2149 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2150 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2151 bool X86::isMOVHLPSMask(SDNode *N) {
2152 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2154 if (N->getNumOperands() != 4)
2157 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2158 return isUndefOrEqual(N->getOperand(0), 6) &&
2159 isUndefOrEqual(N->getOperand(1), 7) &&
2160 isUndefOrEqual(N->getOperand(2), 2) &&
2161 isUndefOrEqual(N->getOperand(3), 3);
2164 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2165 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2167 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2168 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2170 if (N->getNumOperands() != 4)
2173 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2174 return isUndefOrEqual(N->getOperand(0), 2) &&
2175 isUndefOrEqual(N->getOperand(1), 3) &&
2176 isUndefOrEqual(N->getOperand(2), 2) &&
2177 isUndefOrEqual(N->getOperand(3), 3);
2180 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2181 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2182 bool X86::isMOVLPMask(SDNode *N) {
2183 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2185 unsigned NumElems = N->getNumOperands();
2186 if (NumElems != 2 && NumElems != 4)
2189 for (unsigned i = 0; i < NumElems/2; ++i)
2190 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2193 for (unsigned i = NumElems/2; i < NumElems; ++i)
2194 if (!isUndefOrEqual(N->getOperand(i), i))
2200 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2201 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2203 bool X86::isMOVHPMask(SDNode *N) {
2204 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2206 unsigned NumElems = N->getNumOperands();
2207 if (NumElems != 2 && NumElems != 4)
2210 for (unsigned i = 0; i < NumElems/2; ++i)
2211 if (!isUndefOrEqual(N->getOperand(i), i))
2214 for (unsigned i = 0; i < NumElems/2; ++i) {
2215 SDOperand Arg = N->getOperand(i + NumElems/2);
2216 if (!isUndefOrEqual(Arg, i + NumElems))
2223 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2224 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2225 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2226 bool V2IsSplat = false) {
2227 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2230 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2231 SDOperand BitI = Elts[i];
2232 SDOperand BitI1 = Elts[i+1];
2233 if (!isUndefOrEqual(BitI, j))
2236 if (isUndefOrEqual(BitI1, NumElts))
2239 if (!isUndefOrEqual(BitI1, j + NumElts))
2247 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2248 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2249 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2252 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2253 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2254 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2255 bool V2IsSplat = false) {
2256 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2259 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2260 SDOperand BitI = Elts[i];
2261 SDOperand BitI1 = Elts[i+1];
2262 if (!isUndefOrEqual(BitI, j + NumElts/2))
2265 if (isUndefOrEqual(BitI1, NumElts))
2268 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2276 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2277 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2278 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2281 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2282 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2284 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2285 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2287 unsigned NumElems = N->getNumOperands();
2288 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2291 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2292 SDOperand BitI = N->getOperand(i);
2293 SDOperand BitI1 = N->getOperand(i+1);
2295 if (!isUndefOrEqual(BitI, j))
2297 if (!isUndefOrEqual(BitI1, j))
2304 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2305 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2307 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2308 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2310 unsigned NumElems = N->getNumOperands();
2311 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2314 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2315 SDOperand BitI = N->getOperand(i);
2316 SDOperand BitI1 = N->getOperand(i + 1);
2318 if (!isUndefOrEqual(BitI, j))
2320 if (!isUndefOrEqual(BitI1, j))
2327 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2328 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2329 /// MOVSD, and MOVD, i.e. setting the lowest element.
2330 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2331 if (NumElts != 2 && NumElts != 4)
2334 if (!isUndefOrEqual(Elts[0], NumElts))
2337 for (unsigned i = 1; i < NumElts; ++i) {
2338 if (!isUndefOrEqual(Elts[i], i))
2345 bool X86::isMOVLMask(SDNode *N) {
2346 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2347 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2350 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2351 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2352 /// element of vector 2 and the other elements to come from vector 1 in order.
2353 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2354 bool V2IsSplat = false,
2355 bool V2IsUndef = false) {
2356 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2359 if (!isUndefOrEqual(Ops[0], 0))
2362 for (unsigned i = 1; i < NumOps; ++i) {
2363 SDOperand Arg = Ops[i];
2364 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2365 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2366 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2373 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2374 bool V2IsUndef = false) {
2375 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2376 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2377 V2IsSplat, V2IsUndef);
2380 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2381 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2382 bool X86::isMOVSHDUPMask(SDNode *N) {
2383 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2385 if (N->getNumOperands() != 4)
2388 // Expect 1, 1, 3, 3
2389 for (unsigned i = 0; i < 2; ++i) {
2390 SDOperand Arg = N->getOperand(i);
2391 if (Arg.getOpcode() == ISD::UNDEF) continue;
2392 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2393 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2394 if (Val != 1) return false;
2398 for (unsigned i = 2; i < 4; ++i) {
2399 SDOperand Arg = N->getOperand(i);
2400 if (Arg.getOpcode() == ISD::UNDEF) continue;
2401 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2402 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2403 if (Val != 3) return false;
2407 // Don't use movshdup if it can be done with a shufps.
2411 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2412 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2413 bool X86::isMOVSLDUPMask(SDNode *N) {
2414 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2416 if (N->getNumOperands() != 4)
2419 // Expect 0, 0, 2, 2
2420 for (unsigned i = 0; i < 2; ++i) {
2421 SDOperand Arg = N->getOperand(i);
2422 if (Arg.getOpcode() == ISD::UNDEF) continue;
2423 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2424 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2425 if (Val != 0) return false;
2429 for (unsigned i = 2; i < 4; ++i) {
2430 SDOperand Arg = N->getOperand(i);
2431 if (Arg.getOpcode() == ISD::UNDEF) continue;
2432 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2433 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2434 if (Val != 2) return false;
2438 // Don't use movshdup if it can be done with a shufps.
2442 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2443 /// specifies a identity operation on the LHS or RHS.
2444 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2445 unsigned NumElems = N->getNumOperands();
2446 for (unsigned i = 0; i < NumElems; ++i)
2447 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2452 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2453 /// a splat of a single element.
2454 static bool isSplatMask(SDNode *N) {
2455 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2457 // This is a splat operation if each element of the permute is the same, and
2458 // if the value doesn't reference the second vector.
2459 unsigned NumElems = N->getNumOperands();
2460 SDOperand ElementBase;
2462 for (; i != NumElems; ++i) {
2463 SDOperand Elt = N->getOperand(i);
2464 if (isa<ConstantSDNode>(Elt)) {
2470 if (!ElementBase.Val)
2473 for (; i != NumElems; ++i) {
2474 SDOperand Arg = N->getOperand(i);
2475 if (Arg.getOpcode() == ISD::UNDEF) continue;
2476 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2477 if (Arg != ElementBase) return false;
2480 // Make sure it is a splat of the first vector operand.
2481 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2484 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2485 /// a splat of a single element and it's a 2 or 4 element mask.
2486 bool X86::isSplatMask(SDNode *N) {
2487 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2489 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2490 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2492 return ::isSplatMask(N);
2495 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2496 /// specifies a splat of zero element.
2497 bool X86::isSplatLoMask(SDNode *N) {
2498 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2500 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2501 if (!isUndefOrEqual(N->getOperand(i), 0))
2506 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2507 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2509 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2510 unsigned NumOperands = N->getNumOperands();
2511 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2513 for (unsigned i = 0; i < NumOperands; ++i) {
2515 SDOperand Arg = N->getOperand(NumOperands-i-1);
2516 if (Arg.getOpcode() != ISD::UNDEF)
2517 Val = cast<ConstantSDNode>(Arg)->getValue();
2518 if (Val >= NumOperands) Val -= NumOperands;
2520 if (i != NumOperands - 1)
2527 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2528 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2530 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2532 // 8 nodes, but we only care about the last 4.
2533 for (unsigned i = 7; i >= 4; --i) {
2535 SDOperand Arg = N->getOperand(i);
2536 if (Arg.getOpcode() != ISD::UNDEF)
2537 Val = cast<ConstantSDNode>(Arg)->getValue();
2546 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2547 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2549 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2551 // 8 nodes, but we only care about the first 4.
2552 for (int i = 3; i >= 0; --i) {
2554 SDOperand Arg = N->getOperand(i);
2555 if (Arg.getOpcode() != ISD::UNDEF)
2556 Val = cast<ConstantSDNode>(Arg)->getValue();
2565 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2566 /// specifies a 8 element shuffle that can be broken into a pair of
2567 /// PSHUFHW and PSHUFLW.
2568 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2569 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2571 if (N->getNumOperands() != 8)
2574 // Lower quadword shuffled.
2575 for (unsigned i = 0; i != 4; ++i) {
2576 SDOperand Arg = N->getOperand(i);
2577 if (Arg.getOpcode() == ISD::UNDEF) continue;
2578 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2579 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2584 // Upper quadword shuffled.
2585 for (unsigned i = 4; i != 8; ++i) {
2586 SDOperand Arg = N->getOperand(i);
2587 if (Arg.getOpcode() == ISD::UNDEF) continue;
2588 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2589 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2590 if (Val < 4 || Val > 7)
2597 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2598 /// values in ther permute mask.
2599 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2600 SDOperand &V2, SDOperand &Mask,
2601 SelectionDAG &DAG) {
2602 MVT::ValueType VT = Op.getValueType();
2603 MVT::ValueType MaskVT = Mask.getValueType();
2604 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2605 unsigned NumElems = Mask.getNumOperands();
2606 SmallVector<SDOperand, 8> MaskVec;
2608 for (unsigned i = 0; i != NumElems; ++i) {
2609 SDOperand Arg = Mask.getOperand(i);
2610 if (Arg.getOpcode() == ISD::UNDEF) {
2611 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2614 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2615 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2617 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2619 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2623 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2624 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2627 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2628 /// the two vector operands have swapped position.
2630 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2631 MVT::ValueType MaskVT = Mask.getValueType();
2632 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2633 unsigned NumElems = Mask.getNumOperands();
2634 SmallVector<SDOperand, 8> MaskVec;
2635 for (unsigned i = 0; i != NumElems; ++i) {
2636 SDOperand Arg = Mask.getOperand(i);
2637 if (Arg.getOpcode() == ISD::UNDEF) {
2638 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2641 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2642 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2644 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2646 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2648 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2652 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2653 /// match movhlps. The lower half elements should come from upper half of
2654 /// V1 (and in order), and the upper half elements should come from the upper
2655 /// half of V2 (and in order).
2656 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2657 unsigned NumElems = Mask->getNumOperands();
2660 for (unsigned i = 0, e = 2; i != e; ++i)
2661 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2663 for (unsigned i = 2; i != 4; ++i)
2664 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2669 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2670 /// is promoted to a vector.
2671 static inline bool isScalarLoadToVector(SDNode *N) {
2672 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2673 N = N->getOperand(0).Val;
2674 return ISD::isNON_EXTLoad(N);
2679 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2680 /// match movlp{s|d}. The lower half elements should come from lower half of
2681 /// V1 (and in order), and the upper half elements should come from the upper
2682 /// half of V2 (and in order). And since V1 will become the source of the
2683 /// MOVLP, it must be either a vector load or a scalar load to vector.
2684 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2685 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2687 // Is V2 is a vector load, don't do this transformation. We will try to use
2688 // load folding shufps op.
2689 if (ISD::isNON_EXTLoad(V2))
2692 unsigned NumElems = Mask->getNumOperands();
2693 if (NumElems != 2 && NumElems != 4)
2695 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2696 if (!isUndefOrEqual(Mask->getOperand(i), i))
2698 for (unsigned i = NumElems/2; i != NumElems; ++i)
2699 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2704 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2706 static bool isSplatVector(SDNode *N) {
2707 if (N->getOpcode() != ISD::BUILD_VECTOR)
2710 SDOperand SplatValue = N->getOperand(0);
2711 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2712 if (N->getOperand(i) != SplatValue)
2717 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2719 static bool isUndefShuffle(SDNode *N) {
2720 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2723 SDOperand V1 = N->getOperand(0);
2724 SDOperand V2 = N->getOperand(1);
2725 SDOperand Mask = N->getOperand(2);
2726 unsigned NumElems = Mask.getNumOperands();
2727 for (unsigned i = 0; i != NumElems; ++i) {
2728 SDOperand Arg = Mask.getOperand(i);
2729 if (Arg.getOpcode() != ISD::UNDEF) {
2730 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2731 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2733 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2740 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2742 static inline bool isZeroNode(SDOperand Elt) {
2743 return ((isa<ConstantSDNode>(Elt) &&
2744 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2745 (isa<ConstantFPSDNode>(Elt) &&
2746 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2749 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2750 /// to an zero vector.
2751 static bool isZeroShuffle(SDNode *N) {
2752 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2755 SDOperand V1 = N->getOperand(0);
2756 SDOperand V2 = N->getOperand(1);
2757 SDOperand Mask = N->getOperand(2);
2758 unsigned NumElems = Mask.getNumOperands();
2759 for (unsigned i = 0; i != NumElems; ++i) {
2760 SDOperand Arg = Mask.getOperand(i);
2761 if (Arg.getOpcode() == ISD::UNDEF)
2764 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2765 if (Idx < NumElems) {
2766 unsigned Opc = V1.Val->getOpcode();
2767 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2769 if (Opc != ISD::BUILD_VECTOR ||
2770 !isZeroNode(V1.Val->getOperand(Idx)))
2772 } else if (Idx >= NumElems) {
2773 unsigned Opc = V2.Val->getOpcode();
2774 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2776 if (Opc != ISD::BUILD_VECTOR ||
2777 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2784 /// getZeroVector - Returns a vector of specified type with all zero elements.
2786 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2787 assert(MVT::isVector(VT) && "Expected a vector type");
2789 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2790 // type. This ensures they get CSE'd.
2791 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2793 if (MVT::getSizeInBits(VT) == 64) // MMX
2794 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2796 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2797 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2800 /// getOnesVector - Returns a vector of specified type with all bits set.
2802 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2803 assert(MVT::isVector(VT) && "Expected a vector type");
2805 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2806 // type. This ensures they get CSE'd.
2807 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2809 if (MVT::getSizeInBits(VT) == 64) // MMX
2810 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2812 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2813 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2817 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2818 /// that point to V2 points to its first element.
2819 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2820 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2822 bool Changed = false;
2823 SmallVector<SDOperand, 8> MaskVec;
2824 unsigned NumElems = Mask.getNumOperands();
2825 for (unsigned i = 0; i != NumElems; ++i) {
2826 SDOperand Arg = Mask.getOperand(i);
2827 if (Arg.getOpcode() != ISD::UNDEF) {
2828 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2829 if (Val > NumElems) {
2830 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2834 MaskVec.push_back(Arg);
2838 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2839 &MaskVec[0], MaskVec.size());
2843 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2844 /// operation of specified width.
2845 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2846 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2847 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2849 SmallVector<SDOperand, 8> MaskVec;
2850 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2851 for (unsigned i = 1; i != NumElems; ++i)
2852 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2853 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2856 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2857 /// of specified width.
2858 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2859 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2860 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2861 SmallVector<SDOperand, 8> MaskVec;
2862 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2863 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2864 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2866 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2869 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2870 /// of specified width.
2871 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2872 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2873 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2874 unsigned Half = NumElems/2;
2875 SmallVector<SDOperand, 8> MaskVec;
2876 for (unsigned i = 0; i != Half; ++i) {
2877 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2878 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2880 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2883 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2885 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2886 SDOperand V1 = Op.getOperand(0);
2887 SDOperand Mask = Op.getOperand(2);
2888 MVT::ValueType VT = Op.getValueType();
2889 unsigned NumElems = Mask.getNumOperands();
2890 Mask = getUnpacklMask(NumElems, DAG);
2891 while (NumElems != 4) {
2892 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2895 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2897 Mask = getZeroVector(MVT::v4i32, DAG);
2898 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2899 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2900 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2903 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2904 /// vector of zero or undef vector. This produces a shuffle where the low
2905 /// element of V2 is swizzled into the zero/undef vector, landing at element
2906 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2907 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2908 unsigned NumElems, unsigned Idx,
2909 bool isZero, SelectionDAG &DAG) {
2910 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2911 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2912 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2913 SmallVector<SDOperand, 16> MaskVec;
2914 for (unsigned i = 0; i != NumElems; ++i)
2915 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2916 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2918 MaskVec.push_back(DAG.getConstant(i, EVT));
2919 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2920 &MaskVec[0], MaskVec.size());
2921 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2924 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2926 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2927 unsigned NumNonZero, unsigned NumZero,
2928 SelectionDAG &DAG, TargetLowering &TLI) {
2934 for (unsigned i = 0; i < 16; ++i) {
2935 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2936 if (ThisIsNonZero && First) {
2938 V = getZeroVector(MVT::v8i16, DAG);
2940 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2945 SDOperand ThisElt(0, 0), LastElt(0, 0);
2946 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2947 if (LastIsNonZero) {
2948 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2950 if (ThisIsNonZero) {
2951 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2952 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2953 ThisElt, DAG.getConstant(8, MVT::i8));
2955 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2960 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2961 DAG.getIntPtrConstant(i/2));
2965 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2968 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2970 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2971 unsigned NumNonZero, unsigned NumZero,
2972 SelectionDAG &DAG, TargetLowering &TLI) {
2978 for (unsigned i = 0; i < 8; ++i) {
2979 bool isNonZero = (NonZeros & (1 << i)) != 0;
2983 V = getZeroVector(MVT::v8i16, DAG);
2985 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2988 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2989 DAG.getIntPtrConstant(i));
2997 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2998 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
2999 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
3000 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3001 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3002 // eliminated on x86-32 hosts.
3003 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3006 if (ISD::isBuildVectorAllOnes(Op.Val))
3007 return getOnesVector(Op.getValueType(), DAG);
3008 return getZeroVector(Op.getValueType(), DAG);
3011 MVT::ValueType VT = Op.getValueType();
3012 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3013 unsigned EVTBits = MVT::getSizeInBits(EVT);
3015 unsigned NumElems = Op.getNumOperands();
3016 unsigned NumZero = 0;
3017 unsigned NumNonZero = 0;
3018 unsigned NonZeros = 0;
3019 bool HasNonImms = false;
3020 SmallSet<SDOperand, 8> Values;
3021 for (unsigned i = 0; i < NumElems; ++i) {
3022 SDOperand Elt = Op.getOperand(i);
3023 if (Elt.getOpcode() == ISD::UNDEF)
3026 if (Elt.getOpcode() != ISD::Constant &&
3027 Elt.getOpcode() != ISD::ConstantFP)
3029 if (isZeroNode(Elt))
3032 NonZeros |= (1 << i);
3037 if (NumNonZero == 0) {
3038 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3039 return DAG.getNode(ISD::UNDEF, VT);
3042 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3043 if (Values.size() == 1)
3046 // Special case for single non-zero element.
3047 if (NumNonZero == 1 && NumElems <= 4) {
3048 unsigned Idx = CountTrailingZeros_32(NonZeros);
3049 SDOperand Item = Op.getOperand(Idx);
3050 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3052 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3053 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3055 else if (!HasNonImms) // Otherwise, it's better to do a constpool load.
3058 if (EVTBits == 32) {
3059 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3060 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3062 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3063 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3064 SmallVector<SDOperand, 8> MaskVec;
3065 for (unsigned i = 0; i < NumElems; i++)
3066 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3067 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3068 &MaskVec[0], MaskVec.size());
3069 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3070 DAG.getNode(ISD::UNDEF, VT), Mask);
3074 // A vector full of immediates; various special cases are already
3075 // handled, so this is best done with a single constant-pool load.
3079 // Let legalizer expand 2-wide build_vectors.
3083 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3084 if (EVTBits == 8 && NumElems == 16) {
3085 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3087 if (V.Val) return V;
3090 if (EVTBits == 16 && NumElems == 8) {
3091 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3093 if (V.Val) return V;
3096 // If element VT is == 32 bits, turn it into a number of shuffles.
3097 SmallVector<SDOperand, 8> V;
3099 if (NumElems == 4 && NumZero > 0) {
3100 for (unsigned i = 0; i < 4; ++i) {
3101 bool isZero = !(NonZeros & (1 << i));
3103 V[i] = getZeroVector(VT, DAG);
3105 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3108 for (unsigned i = 0; i < 2; ++i) {
3109 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3112 V[i] = V[i*2]; // Must be a zero vector.
3115 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3116 getMOVLMask(NumElems, DAG));
3119 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3120 getMOVLMask(NumElems, DAG));
3123 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3124 getUnpacklMask(NumElems, DAG));
3129 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3130 // clears the upper bits.
3131 // FIXME: we can do the same for v4f32 case when we know both parts of
3132 // the lower half come from scalar_to_vector (loadf32). We should do
3133 // that in post legalizer dag combiner with target specific hooks.
3134 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3136 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3137 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3138 SmallVector<SDOperand, 8> MaskVec;
3139 bool Reverse = (NonZeros & 0x3) == 2;
3140 for (unsigned i = 0; i < 2; ++i)
3142 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3144 MaskVec.push_back(DAG.getConstant(i, EVT));
3145 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3146 for (unsigned i = 0; i < 2; ++i)
3148 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3150 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3151 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3152 &MaskVec[0], MaskVec.size());
3153 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3156 if (Values.size() > 2) {
3157 // Expand into a number of unpckl*.
3159 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3160 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3161 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3162 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3163 for (unsigned i = 0; i < NumElems; ++i)
3164 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3166 while (NumElems != 0) {
3167 for (unsigned i = 0; i < NumElems; ++i)
3168 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3179 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3180 SDOperand PermMask, SelectionDAG &DAG,
3181 TargetLowering &TLI) {
3183 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3184 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3185 MVT::ValueType PtrVT = TLI.getPointerTy();
3186 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3187 PermMask.Val->op_end());
3189 // First record which half of which vector the low elements come from.
3190 SmallVector<unsigned, 4> LowQuad(4);
3191 for (unsigned i = 0; i < 4; ++i) {
3192 SDOperand Elt = MaskElts[i];
3193 if (Elt.getOpcode() == ISD::UNDEF)
3195 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3196 int QuadIdx = EltIdx / 4;
3199 int BestLowQuad = -1;
3200 unsigned MaxQuad = 1;
3201 for (unsigned i = 0; i < 4; ++i) {
3202 if (LowQuad[i] > MaxQuad) {
3204 MaxQuad = LowQuad[i];
3208 // Record which half of which vector the high elements come from.
3209 SmallVector<unsigned, 4> HighQuad(4);
3210 for (unsigned i = 4; i < 8; ++i) {
3211 SDOperand Elt = MaskElts[i];
3212 if (Elt.getOpcode() == ISD::UNDEF)
3214 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3215 int QuadIdx = EltIdx / 4;
3216 ++HighQuad[QuadIdx];
3218 int BestHighQuad = -1;
3220 for (unsigned i = 0; i < 4; ++i) {
3221 if (HighQuad[i] > MaxQuad) {
3223 MaxQuad = HighQuad[i];
3227 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3228 if (BestLowQuad != -1 || BestHighQuad != -1) {
3229 // First sort the 4 chunks in order using shufpd.
3230 SmallVector<SDOperand, 8> MaskVec;
3231 if (BestLowQuad != -1)
3232 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3234 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3235 if (BestHighQuad != -1)
3236 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3238 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3239 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3240 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3241 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3242 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3243 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3245 // Now sort high and low parts separately.
3246 BitVector InOrder(8);
3247 if (BestLowQuad != -1) {
3248 // Sort lower half in order using PSHUFLW.
3250 bool AnyOutOrder = false;
3251 for (unsigned i = 0; i != 4; ++i) {
3252 SDOperand Elt = MaskElts[i];
3253 if (Elt.getOpcode() == ISD::UNDEF) {
3254 MaskVec.push_back(Elt);
3257 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3260 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3261 // If this element is in the right place after this shuffle, then
3263 if ((int)(EltIdx / 4) == BestLowQuad)
3268 for (unsigned i = 4; i != 8; ++i)
3269 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3270 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3271 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3275 if (BestHighQuad != -1) {
3276 // Sort high half in order using PSHUFHW if possible.
3278 for (unsigned i = 0; i != 4; ++i)
3279 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3280 bool AnyOutOrder = false;
3281 for (unsigned i = 4; i != 8; ++i) {
3282 SDOperand Elt = MaskElts[i];
3283 if (Elt.getOpcode() == ISD::UNDEF) {
3284 MaskVec.push_back(Elt);
3287 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3290 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3291 // If this element is in the right place after this shuffle, then
3293 if ((int)(EltIdx / 4) == BestHighQuad)
3298 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3299 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3303 // The other elements are put in the right place using pextrw and pinsrw.
3304 for (unsigned i = 0; i != 8; ++i) {
3307 SDOperand Elt = MaskElts[i];
3308 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3311 SDOperand ExtOp = (EltIdx < 8)
3312 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3313 DAG.getConstant(EltIdx, PtrVT))
3314 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3315 DAG.getConstant(EltIdx - 8, PtrVT));
3316 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3317 DAG.getConstant(i, PtrVT));
3322 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3323 ///as few as possible.
3324 // First, let's find out how many elements are already in the right order.
3325 unsigned V1InOrder = 0;
3326 unsigned V1FromV1 = 0;
3327 unsigned V2InOrder = 0;
3328 unsigned V2FromV2 = 0;
3329 SmallVector<SDOperand, 8> V1Elts;
3330 SmallVector<SDOperand, 8> V2Elts;
3331 for (unsigned i = 0; i < 8; ++i) {
3332 SDOperand Elt = MaskElts[i];
3333 if (Elt.getOpcode() == ISD::UNDEF) {
3334 V1Elts.push_back(Elt);
3335 V2Elts.push_back(Elt);
3340 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3342 V1Elts.push_back(Elt);
3343 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3345 } else if (EltIdx == i+8) {
3346 V1Elts.push_back(Elt);
3347 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3349 } else if (EltIdx < 8) {
3350 V1Elts.push_back(Elt);
3353 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3358 if (V2InOrder > V1InOrder) {
3359 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3361 std::swap(V1Elts, V2Elts);
3362 std::swap(V1FromV1, V2FromV2);
3365 if ((V1FromV1 + V1InOrder) != 8) {
3366 // Some elements are from V2.
3368 // If there are elements that are from V1 but out of place,
3369 // then first sort them in place
3370 SmallVector<SDOperand, 8> MaskVec;
3371 for (unsigned i = 0; i < 8; ++i) {
3372 SDOperand Elt = V1Elts[i];
3373 if (Elt.getOpcode() == ISD::UNDEF) {
3374 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3377 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3379 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3381 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3383 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3384 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3388 for (unsigned i = 0; i < 8; ++i) {
3389 SDOperand Elt = V1Elts[i];
3390 if (Elt.getOpcode() == ISD::UNDEF)
3392 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3395 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3396 DAG.getConstant(EltIdx - 8, PtrVT));
3397 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3398 DAG.getConstant(i, PtrVT));
3402 // All elements are from V1.
3404 for (unsigned i = 0; i < 8; ++i) {
3405 SDOperand Elt = V1Elts[i];
3406 if (Elt.getOpcode() == ISD::UNDEF)
3408 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3409 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3410 DAG.getConstant(EltIdx, PtrVT));
3411 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3412 DAG.getConstant(i, PtrVT));
3418 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3419 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3420 /// done when every pair / quad of shuffle mask elements point to elements in
3421 /// the right sequence. e.g.
3422 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3424 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3426 SDOperand PermMask, SelectionDAG &DAG,
3427 TargetLowering &TLI) {
3428 unsigned NumElems = PermMask.getNumOperands();
3429 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3430 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3431 MVT::ValueType NewVT = MaskVT;
3433 case MVT::v4f32: NewVT = MVT::v2f64; break;
3434 case MVT::v4i32: NewVT = MVT::v2i64; break;
3435 case MVT::v8i16: NewVT = MVT::v4i32; break;
3436 case MVT::v16i8: NewVT = MVT::v4i32; break;
3437 default: assert(false && "Unexpected!");
3440 if (NewWidth == 2) {
3441 if (MVT::isInteger(VT))
3446 unsigned Scale = NumElems / NewWidth;
3447 SmallVector<SDOperand, 8> MaskVec;
3448 for (unsigned i = 0; i < NumElems; i += Scale) {
3449 unsigned StartIdx = ~0U;
3450 for (unsigned j = 0; j < Scale; ++j) {
3451 SDOperand Elt = PermMask.getOperand(i+j);
3452 if (Elt.getOpcode() == ISD::UNDEF)
3454 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3455 if (StartIdx == ~0U)
3456 StartIdx = EltIdx - (EltIdx % Scale);
3457 if (EltIdx != StartIdx + j)
3460 if (StartIdx == ~0U)
3461 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3463 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3466 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3467 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3468 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3469 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3470 &MaskVec[0], MaskVec.size()));
3474 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3475 SDOperand V1 = Op.getOperand(0);
3476 SDOperand V2 = Op.getOperand(1);
3477 SDOperand PermMask = Op.getOperand(2);
3478 MVT::ValueType VT = Op.getValueType();
3479 unsigned NumElems = PermMask.getNumOperands();
3480 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3481 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3482 bool V1IsSplat = false;
3483 bool V2IsSplat = false;
3485 if (isUndefShuffle(Op.Val))
3486 return DAG.getNode(ISD::UNDEF, VT);
3488 if (isZeroShuffle(Op.Val))
3489 return getZeroVector(VT, DAG);
3491 if (isIdentityMask(PermMask.Val))
3493 else if (isIdentityMask(PermMask.Val, true))
3496 if (isSplatMask(PermMask.Val)) {
3497 if (NumElems <= 4) return Op;
3498 // Promote it to a v4i32 splat.
3499 return PromoteSplat(Op, DAG);
3502 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3504 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3505 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3507 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3508 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3509 // FIXME: Figure out a cleaner way to do this.
3510 // Try to make use of movq to zero out the top part.
3511 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3512 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3514 SDOperand NewV1 = NewOp.getOperand(0);
3515 SDOperand NewV2 = NewOp.getOperand(1);
3516 SDOperand NewMask = NewOp.getOperand(2);
3517 if (isCommutedMOVL(NewMask.Val, true, false)) {
3518 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3519 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3520 NewV1, NewV2, getMOVLMask(2, DAG));
3521 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3524 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3525 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3526 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3527 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3531 if (X86::isMOVLMask(PermMask.Val))
3532 return (V1IsUndef) ? V2 : Op;
3534 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3535 X86::isMOVSLDUPMask(PermMask.Val) ||
3536 X86::isMOVHLPSMask(PermMask.Val) ||
3537 X86::isMOVHPMask(PermMask.Val) ||
3538 X86::isMOVLPMask(PermMask.Val))
3541 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3542 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3543 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3545 bool Commuted = false;
3546 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3547 // 1,1,1,1 -> v8i16 though.
3548 V1IsSplat = isSplatVector(V1.Val);
3549 V2IsSplat = isSplatVector(V2.Val);
3551 // Canonicalize the splat or undef, if present, to be on the RHS.
3552 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3553 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3554 std::swap(V1IsSplat, V2IsSplat);
3555 std::swap(V1IsUndef, V2IsUndef);
3559 // FIXME: Figure out a cleaner way to do this.
3560 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3561 if (V2IsUndef) return V1;
3562 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3564 // V2 is a splat, so the mask may be malformed. That is, it may point
3565 // to any V2 element. The instruction selectior won't like this. Get
3566 // a corrected mask and commute to form a proper MOVS{S|D}.
3567 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3568 if (NewMask.Val != PermMask.Val)
3569 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3574 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3575 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3576 X86::isUNPCKLMask(PermMask.Val) ||
3577 X86::isUNPCKHMask(PermMask.Val))
3581 // Normalize mask so all entries that point to V2 points to its first
3582 // element then try to match unpck{h|l} again. If match, return a
3583 // new vector_shuffle with the corrected mask.
3584 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3585 if (NewMask.Val != PermMask.Val) {
3586 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3587 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3588 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3589 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3590 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3591 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3596 // Normalize the node to match x86 shuffle ops if needed
3597 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3598 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3601 // Commute is back and try unpck* again.
3602 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3603 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3604 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3605 X86::isUNPCKLMask(PermMask.Val) ||
3606 X86::isUNPCKHMask(PermMask.Val))
3610 // If VT is integer, try PSHUF* first, then SHUFP*.
3611 if (MVT::isInteger(VT)) {
3612 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3613 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3614 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3615 X86::isPSHUFDMask(PermMask.Val)) ||
3616 X86::isPSHUFHWMask(PermMask.Val) ||
3617 X86::isPSHUFLWMask(PermMask.Val)) {
3618 if (V2.getOpcode() != ISD::UNDEF)
3619 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3620 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3624 if (X86::isSHUFPMask(PermMask.Val) &&
3625 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3628 // Floating point cases in the other order.
3629 if (X86::isSHUFPMask(PermMask.Val))
3631 if (X86::isPSHUFDMask(PermMask.Val) ||
3632 X86::isPSHUFHWMask(PermMask.Val) ||
3633 X86::isPSHUFLWMask(PermMask.Val)) {
3634 if (V2.getOpcode() != ISD::UNDEF)
3635 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3636 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3641 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3642 if (VT == MVT::v8i16) {
3643 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3648 // Handle all 4 wide cases with a number of shuffles.
3649 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3650 // Don't do this for MMX.
3651 MVT::ValueType MaskVT = PermMask.getValueType();
3652 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3653 SmallVector<std::pair<int, int>, 8> Locs;
3654 Locs.reserve(NumElems);
3655 SmallVector<SDOperand, 8> Mask1(NumElems,
3656 DAG.getNode(ISD::UNDEF, MaskEVT));
3657 SmallVector<SDOperand, 8> Mask2(NumElems,
3658 DAG.getNode(ISD::UNDEF, MaskEVT));
3661 // If no more than two elements come from either vector. This can be
3662 // implemented with two shuffles. First shuffle gather the elements.
3663 // The second shuffle, which takes the first shuffle as both of its
3664 // vector operands, put the elements into the right order.
3665 for (unsigned i = 0; i != NumElems; ++i) {
3666 SDOperand Elt = PermMask.getOperand(i);
3667 if (Elt.getOpcode() == ISD::UNDEF) {
3668 Locs[i] = std::make_pair(-1, -1);
3670 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3671 if (Val < NumElems) {
3672 Locs[i] = std::make_pair(0, NumLo);
3676 Locs[i] = std::make_pair(1, NumHi);
3677 if (2+NumHi < NumElems)
3678 Mask1[2+NumHi] = Elt;
3683 if (NumLo <= 2 && NumHi <= 2) {
3684 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3685 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3686 &Mask1[0], Mask1.size()));
3687 for (unsigned i = 0; i != NumElems; ++i) {
3688 if (Locs[i].first == -1)
3691 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3692 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3693 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3697 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3698 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3699 &Mask2[0], Mask2.size()));
3702 // Break it into (shuffle shuffle_hi, shuffle_lo).
3704 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3705 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3706 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3707 unsigned MaskIdx = 0;
3709 unsigned HiIdx = NumElems/2;
3710 for (unsigned i = 0; i != NumElems; ++i) {
3711 if (i == NumElems/2) {
3717 SDOperand Elt = PermMask.getOperand(i);
3718 if (Elt.getOpcode() == ISD::UNDEF) {
3719 Locs[i] = std::make_pair(-1, -1);
3720 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3721 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3722 (*MaskPtr)[LoIdx] = Elt;
3725 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3726 (*MaskPtr)[HiIdx] = Elt;
3731 SDOperand LoShuffle =
3732 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3733 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3734 &LoMask[0], LoMask.size()));
3735 SDOperand HiShuffle =
3736 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3737 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3738 &HiMask[0], HiMask.size()));
3739 SmallVector<SDOperand, 8> MaskOps;
3740 for (unsigned i = 0; i != NumElems; ++i) {
3741 if (Locs[i].first == -1) {
3742 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3744 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3745 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3748 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3749 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3750 &MaskOps[0], MaskOps.size()));
3757 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDOperand Op,
3758 SelectionDAG &DAG) {
3759 MVT::ValueType VT = Op.getValueType();
3760 if (MVT::getSizeInBits(VT) == 8) {
3761 SDOperand Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
3762 Op.getOperand(0), Op.getOperand(1));
3763 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3764 DAG.getValueType(VT));
3765 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3766 } else if (MVT::getSizeInBits(VT) == 16) {
3767 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
3768 Op.getOperand(0), Op.getOperand(1));
3769 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3770 DAG.getValueType(VT));
3771 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3778 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3779 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3782 if (Subtarget->hasSSE41())
3783 return LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
3785 MVT::ValueType VT = Op.getValueType();
3786 // TODO: handle v16i8.
3787 if (MVT::getSizeInBits(VT) == 16) {
3788 SDOperand Vec = Op.getOperand(0);
3789 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3791 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3792 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3793 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3795 // Transform it so it match pextrw which produces a 32-bit result.
3796 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3797 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3798 Op.getOperand(0), Op.getOperand(1));
3799 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3800 DAG.getValueType(VT));
3801 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3802 } else if (MVT::getSizeInBits(VT) == 32) {
3803 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3806 // SHUFPS the element to the lowest double word, then movss.
3807 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3808 SmallVector<SDOperand, 8> IdxVec;
3810 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3812 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3814 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3816 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3817 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3818 &IdxVec[0], IdxVec.size());
3819 SDOperand Vec = Op.getOperand(0);
3820 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3821 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3822 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3823 DAG.getIntPtrConstant(0));
3824 } else if (MVT::getSizeInBits(VT) == 64) {
3825 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
3826 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
3827 // to match extract_elt for f64.
3828 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3832 // UNPCKHPD the element to the lowest double word, then movsd.
3833 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3834 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3835 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3836 SmallVector<SDOperand, 8> IdxVec;
3837 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3839 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3840 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3841 &IdxVec[0], IdxVec.size());
3842 SDOperand Vec = Op.getOperand(0);
3843 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3844 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3845 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3846 DAG.getIntPtrConstant(0));
3853 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDOperand Op, SelectionDAG &DAG){
3854 MVT::ValueType VT = Op.getValueType();
3855 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3857 SDOperand N0 = Op.getOperand(0);
3858 SDOperand N1 = Op.getOperand(1);
3859 SDOperand N2 = Op.getOperand(2);
3861 if ((MVT::getSizeInBits(EVT) == 8) || (MVT::getSizeInBits(EVT) == 16)) {
3862 unsigned Opc = (MVT::getSizeInBits(EVT) == 8) ? X86ISD::PINSRB
3864 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
3866 if (N1.getValueType() != MVT::i32)
3867 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3868 if (N2.getValueType() != MVT::i32)
3869 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3870 return DAG.getNode(Opc, VT, N0, N1, N2);
3871 } else if (EVT == MVT::f32) {
3872 // Bits [7:6] of the constant are the source select. This will always be
3873 // zero here. The DAG Combiner may combine an extract_elt index into these
3874 // bits. For example (insert (extract, 3), 2) could be matched by putting
3875 // the '3' into bits [7:6] of X86ISD::INSERTPS.
3876 // Bits [5:4] of the constant are the destination select. This is the
3877 // value of the incoming immediate.
3878 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
3879 // combine either bitwise AND or insert of float 0.0 to set these bits.
3880 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
3881 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
3887 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3888 MVT::ValueType VT = Op.getValueType();
3889 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3891 if (Subtarget->hasSSE41())
3892 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
3897 SDOperand N0 = Op.getOperand(0);
3898 SDOperand N1 = Op.getOperand(1);
3899 SDOperand N2 = Op.getOperand(2);
3901 if (MVT::getSizeInBits(EVT) == 16) {
3902 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3903 // as its second argument.
3904 if (N1.getValueType() != MVT::i32)
3905 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3906 if (N2.getValueType() != MVT::i32)
3907 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3908 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3914 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3915 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3916 MVT::ValueType VT = MVT::v2i32;
3917 switch (Op.getValueType()) {
3924 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
3925 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
3928 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3929 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3930 // one of the above mentioned nodes. It has to be wrapped because otherwise
3931 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3932 // be used to form addressing mode. These wrapped nodes will be selected
3935 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3936 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3937 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3939 CP->getAlignment());
3940 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3941 // With PIC, the address is actually $g + Offset.
3942 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3943 !Subtarget->isPICStyleRIPRel()) {
3944 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3945 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3953 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3954 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3955 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3956 // If it's a debug information descriptor, don't mess with it.
3957 if (DAG.isVerifiedDebugInfoDesc(Op))
3959 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3960 // With PIC, the address is actually $g + Offset.
3961 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3962 !Subtarget->isPICStyleRIPRel()) {
3963 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3964 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3968 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3969 // load the value at address GV, not the value of GV itself. This means that
3970 // the GlobalAddress must be in the base or index register of the address, not
3971 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3972 // The same applies for external symbols during PIC codegen
3973 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3974 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
3975 PseudoSourceValue::getGOT(), 0);
3980 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3982 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3983 const MVT::ValueType PtrVT) {
3985 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3986 DAG.getNode(X86ISD::GlobalBaseReg,
3988 InFlag = Chain.getValue(1);
3990 // emit leal symbol@TLSGD(,%ebx,1), %eax
3991 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3992 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3993 GA->getValueType(0),
3995 SDOperand Ops[] = { Chain, TGA, InFlag };
3996 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3997 InFlag = Result.getValue(2);
3998 Chain = Result.getValue(1);
4000 // call ___tls_get_addr. This function receives its argument in
4001 // the register EAX.
4002 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4003 InFlag = Chain.getValue(1);
4005 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4006 SDOperand Ops1[] = { Chain,
4007 DAG.getTargetExternalSymbol("___tls_get_addr",
4009 DAG.getRegister(X86::EAX, PtrVT),
4010 DAG.getRegister(X86::EBX, PtrVT),
4012 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4013 InFlag = Chain.getValue(1);
4015 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4018 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4019 // "local exec" model.
4021 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4022 const MVT::ValueType PtrVT) {
4023 // Get the Thread Pointer
4024 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4025 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4027 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4028 GA->getValueType(0),
4030 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4032 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4033 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4034 PseudoSourceValue::getGOT(), 0);
4036 // The address of the thread local variable is the add of the thread
4037 // pointer with the offset of the variable.
4038 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4042 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
4043 // TODO: implement the "local dynamic" model
4044 // TODO: implement the "initial exec"model for pic executables
4045 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
4046 "TLS not implemented for non-ELF and 64-bit targets");
4047 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4048 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4049 // otherwise use the "Local Exec"TLS Model
4050 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4051 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
4053 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4057 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
4058 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4059 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4060 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4061 // With PIC, the address is actually $g + Offset.
4062 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4063 !Subtarget->isPICStyleRIPRel()) {
4064 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4065 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4072 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
4073 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4074 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4075 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4076 // With PIC, the address is actually $g + Offset.
4077 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4078 !Subtarget->isPICStyleRIPRel()) {
4079 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4080 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4087 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4088 /// take a 2 x i32 value to shift plus a shift amount.
4089 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
4090 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
4091 "Not an i64 shift!");
4092 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4093 SDOperand ShOpLo = Op.getOperand(0);
4094 SDOperand ShOpHi = Op.getOperand(1);
4095 SDOperand ShAmt = Op.getOperand(2);
4096 SDOperand Tmp1 = isSRA ?
4097 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
4098 DAG.getConstant(0, MVT::i32);
4100 SDOperand Tmp2, Tmp3;
4101 if (Op.getOpcode() == ISD::SHL_PARTS) {
4102 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
4103 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
4105 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
4106 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
4109 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4110 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4111 DAG.getConstant(32, MVT::i8));
4112 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
4113 AndNode, DAG.getConstant(0, MVT::i8));
4116 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4117 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
4118 SmallVector<SDOperand, 4> Ops;
4119 if (Op.getOpcode() == ISD::SHL_PARTS) {
4120 Ops.push_back(Tmp2);
4121 Ops.push_back(Tmp3);
4123 Ops.push_back(Cond);
4124 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4127 Ops.push_back(Tmp3);
4128 Ops.push_back(Tmp1);
4130 Ops.push_back(Cond);
4131 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4133 Ops.push_back(Tmp2);
4134 Ops.push_back(Tmp3);
4136 Ops.push_back(Cond);
4137 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4140 Ops.push_back(Tmp3);
4141 Ops.push_back(Tmp1);
4143 Ops.push_back(Cond);
4144 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4147 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
4151 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
4154 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
4155 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
4156 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
4157 "Unknown SINT_TO_FP to lower!");
4159 // These are really Legal; caller falls through into that case.
4160 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4162 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4163 Subtarget->is64Bit())
4166 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
4167 MachineFunction &MF = DAG.getMachineFunction();
4168 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4169 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4170 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4172 PseudoSourceValue::getFixedStack(),
4177 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4179 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4181 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4182 SmallVector<SDOperand, 8> Ops;
4183 Ops.push_back(Chain);
4184 Ops.push_back(StackSlot);
4185 Ops.push_back(DAG.getValueType(SrcVT));
4186 SDOperand Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4187 Tys, &Ops[0], Ops.size());
4190 Chain = Result.getValue(1);
4191 SDOperand InFlag = Result.getValue(2);
4193 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4194 // shouldn't be necessary except that RFP cannot be live across
4195 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4196 MachineFunction &MF = DAG.getMachineFunction();
4197 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4198 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4199 Tys = DAG.getVTList(MVT::Other);
4200 SmallVector<SDOperand, 8> Ops;
4201 Ops.push_back(Chain);
4202 Ops.push_back(Result);
4203 Ops.push_back(StackSlot);
4204 Ops.push_back(DAG.getValueType(Op.getValueType()));
4205 Ops.push_back(InFlag);
4206 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4207 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4208 PseudoSourceValue::getFixedStack(), SSFI);
4214 std::pair<SDOperand,SDOperand> X86TargetLowering::
4215 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4216 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4217 "Unknown FP_TO_SINT to lower!");
4219 // These are really Legal.
4220 if (Op.getValueType() == MVT::i32 &&
4221 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4222 return std::make_pair(SDOperand(), SDOperand());
4223 if (Subtarget->is64Bit() &&
4224 Op.getValueType() == MVT::i64 &&
4225 Op.getOperand(0).getValueType() != MVT::f80)
4226 return std::make_pair(SDOperand(), SDOperand());
4228 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4230 MachineFunction &MF = DAG.getMachineFunction();
4231 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4232 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4233 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4235 switch (Op.getValueType()) {
4236 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4237 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4238 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4239 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4242 SDOperand Chain = DAG.getEntryNode();
4243 SDOperand Value = Op.getOperand(0);
4244 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4245 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4246 Chain = DAG.getStore(Chain, Value, StackSlot,
4247 PseudoSourceValue::getFixedStack(), SSFI);
4248 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4250 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4252 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4253 Chain = Value.getValue(1);
4254 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4255 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4258 // Build the FP_TO_INT*_IN_MEM
4259 SDOperand Ops[] = { Chain, Value, StackSlot };
4260 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4262 return std::make_pair(FIST, StackSlot);
4265 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4266 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4267 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4268 if (FIST.Val == 0) return SDOperand();
4271 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4274 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4275 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4276 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4277 if (FIST.Val == 0) return 0;
4279 // Return an i64 load from the stack slot.
4280 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4282 // Use a MERGE_VALUES node to drop the chain result value.
4283 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4286 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4287 MVT::ValueType VT = Op.getValueType();
4288 MVT::ValueType EltVT = VT;
4289 if (MVT::isVector(VT))
4290 EltVT = MVT::getVectorElementType(VT);
4291 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4292 std::vector<Constant*> CV;
4293 if (EltVT == MVT::f64) {
4294 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4298 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4304 Constant *C = ConstantVector::get(CV);
4305 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4306 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4307 PseudoSourceValue::getConstantPool(), 0,
4309 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4312 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4313 MVT::ValueType VT = Op.getValueType();
4314 MVT::ValueType EltVT = VT;
4315 unsigned EltNum = 1;
4316 if (MVT::isVector(VT)) {
4317 EltVT = MVT::getVectorElementType(VT);
4318 EltNum = MVT::getVectorNumElements(VT);
4320 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4321 std::vector<Constant*> CV;
4322 if (EltVT == MVT::f64) {
4323 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4327 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4333 Constant *C = ConstantVector::get(CV);
4334 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4335 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4336 PseudoSourceValue::getConstantPool(), 0,
4338 if (MVT::isVector(VT)) {
4339 return DAG.getNode(ISD::BIT_CONVERT, VT,
4340 DAG.getNode(ISD::XOR, MVT::v2i64,
4341 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4342 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4344 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4348 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4349 SDOperand Op0 = Op.getOperand(0);
4350 SDOperand Op1 = Op.getOperand(1);
4351 MVT::ValueType VT = Op.getValueType();
4352 MVT::ValueType SrcVT = Op1.getValueType();
4353 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4355 // If second operand is smaller, extend it first.
4356 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4357 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4359 SrcTy = MVT::getTypeForValueType(SrcVT);
4361 // And if it is bigger, shrink it first.
4362 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4363 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4365 SrcTy = MVT::getTypeForValueType(SrcVT);
4368 // At this point the operands and the result should have the same
4369 // type, and that won't be f80 since that is not custom lowered.
4371 // First get the sign bit of second operand.
4372 std::vector<Constant*> CV;
4373 if (SrcVT == MVT::f64) {
4374 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4375 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4377 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4378 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4379 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4380 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4382 Constant *C = ConstantVector::get(CV);
4383 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4384 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4385 PseudoSourceValue::getConstantPool(), 0,
4387 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4389 // Shift sign bit right or left if the two operands have different types.
4390 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4391 // Op0 is MVT::f32, Op1 is MVT::f64.
4392 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4393 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4394 DAG.getConstant(32, MVT::i32));
4395 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4396 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4397 DAG.getIntPtrConstant(0));
4400 // Clear first operand sign bit.
4402 if (VT == MVT::f64) {
4403 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4404 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4406 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4407 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4408 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4409 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4411 C = ConstantVector::get(CV);
4412 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4413 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4414 PseudoSourceValue::getConstantPool(), 0,
4416 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4418 // Or the value with the sign bit.
4419 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4422 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4423 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4425 SDOperand Op0 = Op.getOperand(0);
4426 SDOperand Op1 = Op.getOperand(1);
4427 SDOperand CC = Op.getOperand(2);
4428 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4429 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4432 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4434 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4435 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4436 DAG.getConstant(X86CC, MVT::i8), Cond);
4439 assert(isFP && "Illegal integer SetCC!");
4441 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4442 switch (SetCCOpcode) {
4443 default: assert(false && "Illegal floating point SetCC!");
4444 case ISD::SETOEQ: { // !PF & ZF
4445 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4446 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4447 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4448 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4449 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4451 case ISD::SETUNE: { // PF | !ZF
4452 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4453 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4454 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4455 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4456 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4462 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4463 bool addTest = true;
4464 SDOperand Cond = Op.getOperand(0);
4467 if (Cond.getOpcode() == ISD::SETCC)
4468 Cond = LowerSETCC(Cond, DAG);
4470 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4471 // setting operand in place of the X86ISD::SETCC.
4472 if (Cond.getOpcode() == X86ISD::SETCC) {
4473 CC = Cond.getOperand(0);
4475 SDOperand Cmp = Cond.getOperand(1);
4476 unsigned Opc = Cmp.getOpcode();
4477 MVT::ValueType VT = Op.getValueType();
4479 bool IllegalFPCMov = false;
4480 if (MVT::isFloatingPoint(VT) && !MVT::isVector(VT) &&
4481 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4482 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4484 if ((Opc == X86ISD::CMP ||
4485 Opc == X86ISD::COMI ||
4486 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4493 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4494 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4497 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4499 SmallVector<SDOperand, 4> Ops;
4500 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4501 // condition is true.
4502 Ops.push_back(Op.getOperand(2));
4503 Ops.push_back(Op.getOperand(1));
4505 Ops.push_back(Cond);
4506 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4509 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4510 bool addTest = true;
4511 SDOperand Chain = Op.getOperand(0);
4512 SDOperand Cond = Op.getOperand(1);
4513 SDOperand Dest = Op.getOperand(2);
4516 if (Cond.getOpcode() == ISD::SETCC)
4517 Cond = LowerSETCC(Cond, DAG);
4519 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4520 // setting operand in place of the X86ISD::SETCC.
4521 if (Cond.getOpcode() == X86ISD::SETCC) {
4522 CC = Cond.getOperand(0);
4524 SDOperand Cmp = Cond.getOperand(1);
4525 unsigned Opc = Cmp.getOpcode();
4526 if (Opc == X86ISD::CMP ||
4527 Opc == X86ISD::COMI ||
4528 Opc == X86ISD::UCOMI) {
4535 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4536 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4538 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4539 Chain, Op.getOperand(2), CC, Cond);
4543 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4544 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4545 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4546 // that the guard pages used by the OS virtual memory manager are allocated in
4547 // correct sequence.
4549 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4550 SelectionDAG &DAG) {
4551 assert(Subtarget->isTargetCygMing() &&
4552 "This should be used only on Cygwin/Mingw targets");
4555 SDOperand Chain = Op.getOperand(0);
4556 SDOperand Size = Op.getOperand(1);
4557 // FIXME: Ensure alignment here
4561 MVT::ValueType IntPtr = getPointerTy();
4562 MVT::ValueType SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
4564 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4565 Flag = Chain.getValue(1);
4567 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4568 SDOperand Ops[] = { Chain,
4569 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4570 DAG.getRegister(X86::EAX, IntPtr),
4572 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4573 Flag = Chain.getValue(1);
4575 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4577 std::vector<MVT::ValueType> Tys;
4578 Tys.push_back(SPTy);
4579 Tys.push_back(MVT::Other);
4580 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4581 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4584 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4585 SDOperand InFlag(0, 0);
4586 SDOperand Chain = Op.getOperand(0);
4588 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4589 if (Align == 0) Align = 1;
4591 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4592 // If not DWORD aligned or size is more than the threshold, call memset.
4593 // The libc version is likely to be faster for these cases. It can use the
4594 // address value and run time information about the CPU.
4595 if ((Align & 3) != 0 ||
4596 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4597 MVT::ValueType IntPtr = getPointerTy();
4598 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4599 TargetLowering::ArgListTy Args;
4600 TargetLowering::ArgListEntry Entry;
4601 Entry.Node = Op.getOperand(1);
4602 Entry.Ty = IntPtrTy;
4603 Args.push_back(Entry);
4604 // Extend the unsigned i8 argument to be an int value for the call.
4605 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4606 Entry.Ty = IntPtrTy;
4607 Args.push_back(Entry);
4608 Entry.Node = Op.getOperand(3);
4609 Args.push_back(Entry);
4610 std::pair<SDOperand,SDOperand> CallResult =
4611 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
4612 false, DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4613 return CallResult.second;
4618 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4619 unsigned BytesLeft = 0;
4620 bool TwoRepStos = false;
4623 uint64_t Val = ValC->getValue() & 255;
4625 // If the value is a constant, then we can potentially use larger sets.
4626 switch (Align & 3) {
4627 case 2: // WORD aligned
4630 Val = (Val << 8) | Val;
4632 case 0: // DWORD aligned
4635 Val = (Val << 8) | Val;
4636 Val = (Val << 16) | Val;
4637 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4640 Val = (Val << 32) | Val;
4643 default: // Byte aligned
4646 Count = Op.getOperand(3);
4650 if (AVT > MVT::i8) {
4652 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4653 Count = DAG.getIntPtrConstant(I->getValue() / UBytes);
4654 BytesLeft = I->getValue() % UBytes;
4656 assert(AVT >= MVT::i32 &&
4657 "Do not use rep;stos if not at least DWORD aligned");
4658 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4659 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4664 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4666 InFlag = Chain.getValue(1);
4669 Count = Op.getOperand(3);
4670 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4671 InFlag = Chain.getValue(1);
4674 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4676 InFlag = Chain.getValue(1);
4677 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4678 Op.getOperand(1), InFlag);
4679 InFlag = Chain.getValue(1);
4681 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4682 SmallVector<SDOperand, 8> Ops;
4683 Ops.push_back(Chain);
4684 Ops.push_back(DAG.getValueType(AVT));
4685 Ops.push_back(InFlag);
4686 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4689 InFlag = Chain.getValue(1);
4690 Count = Op.getOperand(3);
4691 MVT::ValueType CVT = Count.getValueType();
4692 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4693 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4694 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4696 InFlag = Chain.getValue(1);
4697 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4699 Ops.push_back(Chain);
4700 Ops.push_back(DAG.getValueType(MVT::i8));
4701 Ops.push_back(InFlag);
4702 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4703 } else if (BytesLeft) {
4704 // Issue stores for the last 1 - 7 bytes.
4706 unsigned Val = ValC->getValue() & 255;
4707 unsigned Offset = I->getValue() - BytesLeft;
4708 SDOperand DstAddr = Op.getOperand(1);
4709 MVT::ValueType AddrVT = DstAddr.getValueType();
4710 if (BytesLeft >= 4) {
4711 Val = (Val << 8) | Val;
4712 Val = (Val << 16) | Val;
4713 Value = DAG.getConstant(Val, MVT::i32);
4714 Chain = DAG.getStore(Chain, Value,
4715 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4716 DAG.getConstant(Offset, AddrVT)),
4721 if (BytesLeft >= 2) {
4722 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4723 Chain = DAG.getStore(Chain, Value,
4724 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4725 DAG.getConstant(Offset, AddrVT)),
4730 if (BytesLeft == 1) {
4731 Value = DAG.getConstant(Val, MVT::i8);
4732 Chain = DAG.getStore(Chain, Value,
4733 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4734 DAG.getConstant(Offset, AddrVT)),
4742 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4747 SelectionDAG &DAG) {
4749 unsigned BytesLeft = 0;
4750 switch (Align & 3) {
4751 case 2: // WORD aligned
4754 case 0: // DWORD aligned
4756 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4759 default: // Byte aligned
4764 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4765 SDOperand Count = DAG.getIntPtrConstant(Size / UBytes);
4766 BytesLeft = Size % UBytes;
4768 SDOperand InFlag(0, 0);
4769 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4771 InFlag = Chain.getValue(1);
4772 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4774 InFlag = Chain.getValue(1);
4775 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4777 InFlag = Chain.getValue(1);
4779 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4780 SmallVector<SDOperand, 8> Ops;
4781 Ops.push_back(Chain);
4782 Ops.push_back(DAG.getValueType(AVT));
4783 Ops.push_back(InFlag);
4784 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4787 // Issue loads and stores for the last 1 - 7 bytes.
4788 unsigned Offset = Size - BytesLeft;
4789 SDOperand DstAddr = Dest;
4790 MVT::ValueType DstVT = DstAddr.getValueType();
4791 SDOperand SrcAddr = Source;
4792 MVT::ValueType SrcVT = SrcAddr.getValueType();
4794 if (BytesLeft >= 4) {
4795 Value = DAG.getLoad(MVT::i32, Chain,
4796 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4797 DAG.getConstant(Offset, SrcVT)),
4799 Chain = Value.getValue(1);
4800 Chain = DAG.getStore(Chain, Value,
4801 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4802 DAG.getConstant(Offset, DstVT)),
4807 if (BytesLeft >= 2) {
4808 Value = DAG.getLoad(MVT::i16, Chain,
4809 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4810 DAG.getConstant(Offset, SrcVT)),
4812 Chain = Value.getValue(1);
4813 Chain = DAG.getStore(Chain, Value,
4814 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4815 DAG.getConstant(Offset, DstVT)),
4821 if (BytesLeft == 1) {
4822 Value = DAG.getLoad(MVT::i8, Chain,
4823 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4824 DAG.getConstant(Offset, SrcVT)),
4826 Chain = Value.getValue(1);
4827 Chain = DAG.getStore(Chain, Value,
4828 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4829 DAG.getConstant(Offset, DstVT)),
4837 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4838 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4839 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4840 SDOperand TheChain = N->getOperand(0);
4841 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4842 if (Subtarget->is64Bit()) {
4843 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4844 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4845 MVT::i64, rax.getValue(2));
4846 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4847 DAG.getConstant(32, MVT::i8));
4849 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4852 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4853 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4856 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4857 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4858 MVT::i32, eax.getValue(2));
4859 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4860 SDOperand Ops[] = { eax, edx };
4861 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
4863 // Use a MERGE_VALUES to return the value and chain.
4864 Ops[1] = edx.getValue(1);
4865 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4866 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4869 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4870 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4872 if (!Subtarget->is64Bit()) {
4873 // vastart just stores the address of the VarArgsFrameIndex slot into the
4874 // memory location argument.
4875 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4876 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
4880 // gp_offset (0 - 6 * 8)
4881 // fp_offset (48 - 48 + 8 * 16)
4882 // overflow_arg_area (point to parameters coming in memory).
4884 SmallVector<SDOperand, 8> MemOps;
4885 SDOperand FIN = Op.getOperand(1);
4887 SDOperand Store = DAG.getStore(Op.getOperand(0),
4888 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4890 MemOps.push_back(Store);
4893 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4894 Store = DAG.getStore(Op.getOperand(0),
4895 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4897 MemOps.push_back(Store);
4899 // Store ptr to overflow_arg_area
4900 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4901 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4902 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
4903 MemOps.push_back(Store);
4905 // Store ptr to reg_save_area.
4906 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
4907 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4908 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
4909 MemOps.push_back(Store);
4910 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4913 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4914 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4915 SDOperand Chain = Op.getOperand(0);
4916 SDOperand DstPtr = Op.getOperand(1);
4917 SDOperand SrcPtr = Op.getOperand(2);
4918 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4919 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4921 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr, SrcSV, 0);
4922 Chain = SrcPtr.getValue(1);
4923 for (unsigned i = 0; i < 3; ++i) {
4924 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr, SrcSV, 0);
4925 Chain = Val.getValue(1);
4926 Chain = DAG.getStore(Chain, Val, DstPtr, DstSV, 0);
4929 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4930 DAG.getIntPtrConstant(8));
4931 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4932 DAG.getIntPtrConstant(8));
4938 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4939 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4941 default: return SDOperand(); // Don't custom lower most intrinsics.
4942 // Comparison intrinsics.
4943 case Intrinsic::x86_sse_comieq_ss:
4944 case Intrinsic::x86_sse_comilt_ss:
4945 case Intrinsic::x86_sse_comile_ss:
4946 case Intrinsic::x86_sse_comigt_ss:
4947 case Intrinsic::x86_sse_comige_ss:
4948 case Intrinsic::x86_sse_comineq_ss:
4949 case Intrinsic::x86_sse_ucomieq_ss:
4950 case Intrinsic::x86_sse_ucomilt_ss:
4951 case Intrinsic::x86_sse_ucomile_ss:
4952 case Intrinsic::x86_sse_ucomigt_ss:
4953 case Intrinsic::x86_sse_ucomige_ss:
4954 case Intrinsic::x86_sse_ucomineq_ss:
4955 case Intrinsic::x86_sse2_comieq_sd:
4956 case Intrinsic::x86_sse2_comilt_sd:
4957 case Intrinsic::x86_sse2_comile_sd:
4958 case Intrinsic::x86_sse2_comigt_sd:
4959 case Intrinsic::x86_sse2_comige_sd:
4960 case Intrinsic::x86_sse2_comineq_sd:
4961 case Intrinsic::x86_sse2_ucomieq_sd:
4962 case Intrinsic::x86_sse2_ucomilt_sd:
4963 case Intrinsic::x86_sse2_ucomile_sd:
4964 case Intrinsic::x86_sse2_ucomigt_sd:
4965 case Intrinsic::x86_sse2_ucomige_sd:
4966 case Intrinsic::x86_sse2_ucomineq_sd: {
4968 ISD::CondCode CC = ISD::SETCC_INVALID;
4971 case Intrinsic::x86_sse_comieq_ss:
4972 case Intrinsic::x86_sse2_comieq_sd:
4976 case Intrinsic::x86_sse_comilt_ss:
4977 case Intrinsic::x86_sse2_comilt_sd:
4981 case Intrinsic::x86_sse_comile_ss:
4982 case Intrinsic::x86_sse2_comile_sd:
4986 case Intrinsic::x86_sse_comigt_ss:
4987 case Intrinsic::x86_sse2_comigt_sd:
4991 case Intrinsic::x86_sse_comige_ss:
4992 case Intrinsic::x86_sse2_comige_sd:
4996 case Intrinsic::x86_sse_comineq_ss:
4997 case Intrinsic::x86_sse2_comineq_sd:
5001 case Intrinsic::x86_sse_ucomieq_ss:
5002 case Intrinsic::x86_sse2_ucomieq_sd:
5003 Opc = X86ISD::UCOMI;
5006 case Intrinsic::x86_sse_ucomilt_ss:
5007 case Intrinsic::x86_sse2_ucomilt_sd:
5008 Opc = X86ISD::UCOMI;
5011 case Intrinsic::x86_sse_ucomile_ss:
5012 case Intrinsic::x86_sse2_ucomile_sd:
5013 Opc = X86ISD::UCOMI;
5016 case Intrinsic::x86_sse_ucomigt_ss:
5017 case Intrinsic::x86_sse2_ucomigt_sd:
5018 Opc = X86ISD::UCOMI;
5021 case Intrinsic::x86_sse_ucomige_ss:
5022 case Intrinsic::x86_sse2_ucomige_sd:
5023 Opc = X86ISD::UCOMI;
5026 case Intrinsic::x86_sse_ucomineq_ss:
5027 case Intrinsic::x86_sse2_ucomineq_sd:
5028 Opc = X86ISD::UCOMI;
5034 SDOperand LHS = Op.getOperand(1);
5035 SDOperand RHS = Op.getOperand(2);
5036 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5038 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5039 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5040 DAG.getConstant(X86CC, MVT::i8), Cond);
5041 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
5046 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
5047 // Depths > 0 not supported yet!
5048 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5051 // Just load the return address
5052 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5053 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5056 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
5057 // Depths > 0 not supported yet!
5058 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5061 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5062 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5063 DAG.getIntPtrConstant(4));
5066 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
5067 SelectionDAG &DAG) {
5068 // Is not yet supported on x86-64
5069 if (Subtarget->is64Bit())
5072 return DAG.getIntPtrConstant(8);
5075 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
5077 assert(!Subtarget->is64Bit() &&
5078 "Lowering of eh_return builtin is not supported yet on x86-64");
5080 MachineFunction &MF = DAG.getMachineFunction();
5081 SDOperand Chain = Op.getOperand(0);
5082 SDOperand Offset = Op.getOperand(1);
5083 SDOperand Handler = Op.getOperand(2);
5085 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
5088 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5089 DAG.getIntPtrConstant(-4UL));
5090 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5091 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5092 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5093 MF.getRegInfo().addLiveOut(X86::ECX);
5095 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5096 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5099 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
5100 SelectionDAG &DAG) {
5101 SDOperand Root = Op.getOperand(0);
5102 SDOperand Trmp = Op.getOperand(1); // trampoline
5103 SDOperand FPtr = Op.getOperand(2); // nested function
5104 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
5106 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5108 const X86InstrInfo *TII =
5109 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5111 if (Subtarget->is64Bit()) {
5112 SDOperand OutChains[6];
5114 // Large code-model.
5116 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5117 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5119 const unsigned char N86R10 =
5120 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R10);
5121 const unsigned char N86R11 =
5122 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R11);
5124 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5126 // Load the pointer to the nested function into R11.
5127 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5128 SDOperand Addr = Trmp;
5129 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5132 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5133 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5135 // Load the 'nest' parameter value into R10.
5136 // R10 is specified in X86CallingConv.td
5137 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5138 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5139 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5142 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5143 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5145 // Jump to the nested function.
5146 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5147 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5148 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5151 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5152 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5153 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5157 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5158 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5160 const Function *Func =
5161 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5162 unsigned CC = Func->getCallingConv();
5167 assert(0 && "Unsupported calling convention");
5168 case CallingConv::C:
5169 case CallingConv::X86_StdCall: {
5170 // Pass 'nest' parameter in ECX.
5171 // Must be kept in sync with X86CallingConv.td
5174 // Check that ECX wasn't needed by an 'inreg' parameter.
5175 const FunctionType *FTy = Func->getFunctionType();
5176 const ParamAttrsList *Attrs = Func->getParamAttrs();
5178 if (Attrs && !Func->isVarArg()) {
5179 unsigned InRegCount = 0;
5182 for (FunctionType::param_iterator I = FTy->param_begin(),
5183 E = FTy->param_end(); I != E; ++I, ++Idx)
5184 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
5185 // FIXME: should only count parameters that are lowered to integers.
5186 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5188 if (InRegCount > 2) {
5189 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5195 case CallingConv::X86_FastCall:
5196 // Pass 'nest' parameter in EAX.
5197 // Must be kept in sync with X86CallingConv.td
5202 SDOperand OutChains[4];
5203 SDOperand Addr, Disp;
5205 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5206 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5208 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5209 const unsigned char N86Reg =
5210 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5211 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5214 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5215 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5217 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5218 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5219 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5220 TrmpAddr, 5, false, 1);
5222 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5223 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5226 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5227 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5231 SDOperand X86TargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
5233 The rounding mode is in bits 11:10 of FPSR, and has the following
5240 FLT_ROUNDS, on the other hand, expects the following:
5247 To perform the conversion, we do:
5248 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5251 MachineFunction &MF = DAG.getMachineFunction();
5252 const TargetMachine &TM = MF.getTarget();
5253 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5254 unsigned StackAlignment = TFI.getStackAlignment();
5255 MVT::ValueType VT = Op.getValueType();
5257 // Save FP Control Word to stack slot
5258 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5259 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5261 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5262 DAG.getEntryNode(), StackSlot);
5264 // Load FP Control Word from stack slot
5265 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5267 // Transform as necessary
5269 DAG.getNode(ISD::SRL, MVT::i16,
5270 DAG.getNode(ISD::AND, MVT::i16,
5271 CWD, DAG.getConstant(0x800, MVT::i16)),
5272 DAG.getConstant(11, MVT::i8));
5274 DAG.getNode(ISD::SRL, MVT::i16,
5275 DAG.getNode(ISD::AND, MVT::i16,
5276 CWD, DAG.getConstant(0x400, MVT::i16)),
5277 DAG.getConstant(9, MVT::i8));
5280 DAG.getNode(ISD::AND, MVT::i16,
5281 DAG.getNode(ISD::ADD, MVT::i16,
5282 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5283 DAG.getConstant(1, MVT::i16)),
5284 DAG.getConstant(3, MVT::i16));
5287 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5288 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5291 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5292 MVT::ValueType VT = Op.getValueType();
5293 MVT::ValueType OpVT = VT;
5294 unsigned NumBits = MVT::getSizeInBits(VT);
5296 Op = Op.getOperand(0);
5297 if (VT == MVT::i8) {
5298 // Zero extend to i32 since there is not an i8 bsr.
5300 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5303 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5304 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5305 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5307 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5308 SmallVector<SDOperand, 4> Ops;
5310 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5311 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5312 Ops.push_back(Op.getValue(1));
5313 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5315 // Finally xor with NumBits-1.
5316 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5319 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5323 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5324 MVT::ValueType VT = Op.getValueType();
5325 MVT::ValueType OpVT = VT;
5326 unsigned NumBits = MVT::getSizeInBits(VT);
5328 Op = Op.getOperand(0);
5329 if (VT == MVT::i8) {
5331 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5334 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5335 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5336 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5338 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5339 SmallVector<SDOperand, 4> Ops;
5341 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5342 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5343 Ops.push_back(Op.getValue(1));
5344 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5347 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5351 SDOperand X86TargetLowering::LowerCAS(SDOperand Op, SelectionDAG &DAG) {
5352 MVT::ValueType T = cast<AtomicSDNode>(Op.Val)->getVT();
5356 case MVT::i8: Reg = X86::AL; size = 1; break;
5357 case MVT::i16: Reg = X86::AX; size = 2; break;
5358 case MVT::i32: Reg = X86::EAX; size = 4; break;
5360 SDOperand cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5361 Op.getOperand(3), SDOperand());
5362 SDOperand Ops[] = { cpIn.getValue(0),
5365 DAG.getTargetConstant(size, MVT::i8),
5367 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5368 SDOperand Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5370 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5374 /// LowerOperation - Provide custom lowering hooks for some operations.
5376 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5377 switch (Op.getOpcode()) {
5378 default: assert(0 && "Should not custom lower this!");
5379 case ISD::ATOMIC_LCS: return LowerCAS(Op,DAG);
5380 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5381 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5382 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5383 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5384 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5385 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5386 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5387 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5388 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5389 case ISD::SHL_PARTS:
5390 case ISD::SRA_PARTS:
5391 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5392 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5393 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5394 case ISD::FABS: return LowerFABS(Op, DAG);
5395 case ISD::FNEG: return LowerFNEG(Op, DAG);
5396 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5397 case ISD::SETCC: return LowerSETCC(Op, DAG);
5398 case ISD::SELECT: return LowerSELECT(Op, DAG);
5399 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5400 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5401 case ISD::CALL: return LowerCALL(Op, DAG);
5402 case ISD::RET: return LowerRET(Op, DAG);
5403 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5404 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5405 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5406 case ISD::VASTART: return LowerVASTART(Op, DAG);
5407 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5408 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5409 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5410 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5411 case ISD::FRAME_TO_ARGS_OFFSET:
5412 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5413 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5414 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5415 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5416 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5417 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5418 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5420 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5421 case ISD::READCYCLECOUNTER:
5422 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5426 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5427 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5428 switch (N->getOpcode()) {
5429 default: assert(0 && "Should not custom lower this!");
5430 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5431 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5435 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5437 default: return NULL;
5438 case X86ISD::BSF: return "X86ISD::BSF";
5439 case X86ISD::BSR: return "X86ISD::BSR";
5440 case X86ISD::SHLD: return "X86ISD::SHLD";
5441 case X86ISD::SHRD: return "X86ISD::SHRD";
5442 case X86ISD::FAND: return "X86ISD::FAND";
5443 case X86ISD::FOR: return "X86ISD::FOR";
5444 case X86ISD::FXOR: return "X86ISD::FXOR";
5445 case X86ISD::FSRL: return "X86ISD::FSRL";
5446 case X86ISD::FILD: return "X86ISD::FILD";
5447 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5448 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5449 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5450 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5451 case X86ISD::FLD: return "X86ISD::FLD";
5452 case X86ISD::FST: return "X86ISD::FST";
5453 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5454 case X86ISD::FP_GET_RESULT2: return "X86ISD::FP_GET_RESULT2";
5455 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5456 case X86ISD::CALL: return "X86ISD::CALL";
5457 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5458 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5459 case X86ISD::CMP: return "X86ISD::CMP";
5460 case X86ISD::COMI: return "X86ISD::COMI";
5461 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5462 case X86ISD::SETCC: return "X86ISD::SETCC";
5463 case X86ISD::CMOV: return "X86ISD::CMOV";
5464 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5465 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5466 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5467 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5468 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5469 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5470 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
5471 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5472 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
5473 case X86ISD::PINSRB: return "X86ISD::PINSRB";
5474 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5475 case X86ISD::FMAX: return "X86ISD::FMAX";
5476 case X86ISD::FMIN: return "X86ISD::FMIN";
5477 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5478 case X86ISD::FRCP: return "X86ISD::FRCP";
5479 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5480 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5481 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5482 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5483 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5484 case X86ISD::LCMPXCHG_DAG: return "x86ISD::LCMPXCHG_DAG";
5488 // isLegalAddressingMode - Return true if the addressing mode represented
5489 // by AM is legal for this target, for a load/store of the specified type.
5490 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5491 const Type *Ty) const {
5492 // X86 supports extremely general addressing modes.
5494 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5495 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5499 // We can only fold this if we don't need an extra load.
5500 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5503 // X86-64 only supports addr of globals in small code model.
5504 if (Subtarget->is64Bit()) {
5505 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5507 // If lower 4G is not available, then we must use rip-relative addressing.
5508 if (AM.BaseOffs || AM.Scale > 1)
5519 // These scales always work.
5524 // These scales are formed with basereg+scalereg. Only accept if there is
5529 default: // Other stuff never works.
5537 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5538 if (!Ty1->isInteger() || !Ty2->isInteger())
5540 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5541 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5542 if (NumBits1 <= NumBits2)
5544 return Subtarget->is64Bit() || NumBits1 < 64;
5547 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5548 MVT::ValueType VT2) const {
5549 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5551 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5552 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5553 if (NumBits1 <= NumBits2)
5555 return Subtarget->is64Bit() || NumBits1 < 64;
5558 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5559 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5560 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5561 /// are assumed to be legal.
5563 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5564 // Only do shuffles on 128-bit vector types for now.
5565 if (MVT::getSizeInBits(VT) == 64) return false;
5566 return (Mask.Val->getNumOperands() <= 4 ||
5567 isIdentityMask(Mask.Val) ||
5568 isIdentityMask(Mask.Val, true) ||
5569 isSplatMask(Mask.Val) ||
5570 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5571 X86::isUNPCKLMask(Mask.Val) ||
5572 X86::isUNPCKHMask(Mask.Val) ||
5573 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5574 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5577 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5579 SelectionDAG &DAG) const {
5580 unsigned NumElts = BVOps.size();
5581 // Only do shuffles on 128-bit vector types for now.
5582 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5583 if (NumElts == 2) return true;
5585 return (isMOVLMask(&BVOps[0], 4) ||
5586 isCommutedMOVL(&BVOps[0], 4, true) ||
5587 isSHUFPMask(&BVOps[0], 4) ||
5588 isCommutedSHUFP(&BVOps[0], 4));
5593 //===----------------------------------------------------------------------===//
5594 // X86 Scheduler Hooks
5595 //===----------------------------------------------------------------------===//
5598 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5599 MachineBasicBlock *BB) {
5600 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5601 switch (MI->getOpcode()) {
5602 default: assert(false && "Unexpected instr type to insert");
5603 case X86::CMOV_FR32:
5604 case X86::CMOV_FR64:
5605 case X86::CMOV_V4F32:
5606 case X86::CMOV_V2F64:
5607 case X86::CMOV_V2I64: {
5608 // To "insert" a SELECT_CC instruction, we actually have to insert the
5609 // diamond control-flow pattern. The incoming instruction knows the
5610 // destination vreg to set, the condition code register to branch on, the
5611 // true/false values to select between, and a branch opcode to use.
5612 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5613 ilist<MachineBasicBlock>::iterator It = BB;
5619 // cmpTY ccX, r1, r2
5621 // fallthrough --> copy0MBB
5622 MachineBasicBlock *thisMBB = BB;
5623 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5624 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5626 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5627 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5628 MachineFunction *F = BB->getParent();
5629 F->getBasicBlockList().insert(It, copy0MBB);
5630 F->getBasicBlockList().insert(It, sinkMBB);
5631 // Update machine-CFG edges by first adding all successors of the current
5632 // block to the new block which will contain the Phi node for the select.
5633 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5634 e = BB->succ_end(); i != e; ++i)
5635 sinkMBB->addSuccessor(*i);
5636 // Next, remove all successors of the current block, and add the true
5637 // and fallthrough blocks as its successors.
5638 while(!BB->succ_empty())
5639 BB->removeSuccessor(BB->succ_begin());
5640 BB->addSuccessor(copy0MBB);
5641 BB->addSuccessor(sinkMBB);
5644 // %FalseValue = ...
5645 // # fallthrough to sinkMBB
5648 // Update machine-CFG edges
5649 BB->addSuccessor(sinkMBB);
5652 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5655 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5656 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5657 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5659 delete MI; // The pseudo instruction is gone now.
5663 case X86::FP32_TO_INT16_IN_MEM:
5664 case X86::FP32_TO_INT32_IN_MEM:
5665 case X86::FP32_TO_INT64_IN_MEM:
5666 case X86::FP64_TO_INT16_IN_MEM:
5667 case X86::FP64_TO_INT32_IN_MEM:
5668 case X86::FP64_TO_INT64_IN_MEM:
5669 case X86::FP80_TO_INT16_IN_MEM:
5670 case X86::FP80_TO_INT32_IN_MEM:
5671 case X86::FP80_TO_INT64_IN_MEM: {
5672 // Change the floating point control register to use "round towards zero"
5673 // mode when truncating to an integer value.
5674 MachineFunction *F = BB->getParent();
5675 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5676 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5678 // Load the old value of the high byte of the control word...
5680 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
5681 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5683 // Set the high part to be round to zero...
5684 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5687 // Reload the modified control word now...
5688 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5690 // Restore the memory image of control word to original value
5691 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5694 // Get the X86 opcode to use.
5696 switch (MI->getOpcode()) {
5697 default: assert(0 && "illegal opcode!");
5698 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5699 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5700 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5701 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5702 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5703 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5704 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5705 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5706 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5710 MachineOperand &Op = MI->getOperand(0);
5711 if (Op.isRegister()) {
5712 AM.BaseType = X86AddressMode::RegBase;
5713 AM.Base.Reg = Op.getReg();
5715 AM.BaseType = X86AddressMode::FrameIndexBase;
5716 AM.Base.FrameIndex = Op.getIndex();
5718 Op = MI->getOperand(1);
5719 if (Op.isImmediate())
5720 AM.Scale = Op.getImm();
5721 Op = MI->getOperand(2);
5722 if (Op.isImmediate())
5723 AM.IndexReg = Op.getImm();
5724 Op = MI->getOperand(3);
5725 if (Op.isGlobalAddress()) {
5726 AM.GV = Op.getGlobal();
5728 AM.Disp = Op.getImm();
5730 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5731 .addReg(MI->getOperand(4).getReg());
5733 // Reload the original control word now.
5734 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5736 delete MI; // The pseudo instruction is gone now.
5742 //===----------------------------------------------------------------------===//
5743 // X86 Optimization Hooks
5744 //===----------------------------------------------------------------------===//
5746 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5750 const SelectionDAG &DAG,
5751 unsigned Depth) const {
5752 unsigned Opc = Op.getOpcode();
5753 assert((Opc >= ISD::BUILTIN_OP_END ||
5754 Opc == ISD::INTRINSIC_WO_CHAIN ||
5755 Opc == ISD::INTRINSIC_W_CHAIN ||
5756 Opc == ISD::INTRINSIC_VOID) &&
5757 "Should use MaskedValueIsZero if you don't know whether Op"
5758 " is a target node!");
5760 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
5764 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
5765 Mask.getBitWidth() - 1);
5770 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5771 /// element of the result of the vector shuffle.
5772 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5773 MVT::ValueType VT = N->getValueType(0);
5774 SDOperand PermMask = N->getOperand(2);
5775 unsigned NumElems = PermMask.getNumOperands();
5776 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5778 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5780 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5781 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5782 SDOperand Idx = PermMask.getOperand(i);
5783 if (Idx.getOpcode() == ISD::UNDEF)
5784 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5785 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5790 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5791 /// node is a GlobalAddress + an offset.
5792 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5793 unsigned Opc = N->getOpcode();
5794 if (Opc == X86ISD::Wrapper) {
5795 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5796 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5799 } else if (Opc == ISD::ADD) {
5800 SDOperand N1 = N->getOperand(0);
5801 SDOperand N2 = N->getOperand(1);
5802 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5803 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5805 Offset += V->getSignExtended();
5808 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5809 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5811 Offset += V->getSignExtended();
5819 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5821 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5822 MachineFrameInfo *MFI) {
5823 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5826 SDOperand Loc = N->getOperand(1);
5827 SDOperand BaseLoc = Base->getOperand(1);
5828 if (Loc.getOpcode() == ISD::FrameIndex) {
5829 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5831 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5832 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5833 int FS = MFI->getObjectSize(FI);
5834 int BFS = MFI->getObjectSize(BFI);
5835 if (FS != BFS || FS != Size) return false;
5836 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5838 GlobalValue *GV1 = NULL;
5839 GlobalValue *GV2 = NULL;
5840 int64_t Offset1 = 0;
5841 int64_t Offset2 = 0;
5842 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5843 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5844 if (isGA1 && isGA2 && GV1 == GV2)
5845 return Offset1 == (Offset2 + Dist*Size);
5851 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5852 const X86Subtarget *Subtarget) {
5855 if (isGAPlusOffset(Base, GV, Offset))
5856 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5857 // DAG combine handles the stack object case.
5862 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5863 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5864 /// if the load addresses are consecutive, non-overlapping, and in the right
5866 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5867 const X86Subtarget *Subtarget) {
5868 MachineFunction &MF = DAG.getMachineFunction();
5869 MachineFrameInfo *MFI = MF.getFrameInfo();
5870 MVT::ValueType VT = N->getValueType(0);
5871 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5872 SDOperand PermMask = N->getOperand(2);
5873 int NumElems = (int)PermMask.getNumOperands();
5874 SDNode *Base = NULL;
5875 for (int i = 0; i < NumElems; ++i) {
5876 SDOperand Idx = PermMask.getOperand(i);
5877 if (Idx.getOpcode() == ISD::UNDEF) {
5878 if (!Base) return SDOperand();
5881 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5882 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5886 else if (!isConsecutiveLoad(Arg.Val, Base,
5887 i, MVT::getSizeInBits(EVT)/8,MFI))
5892 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5893 LoadSDNode *LD = cast<LoadSDNode>(Base);
5895 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5896 LD->getSrcValueOffset(), LD->isVolatile());
5898 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5899 LD->getSrcValueOffset(), LD->isVolatile(),
5900 LD->getAlignment());
5904 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5905 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5906 const X86Subtarget *Subtarget) {
5907 SDOperand Cond = N->getOperand(0);
5909 // If we have SSE[12] support, try to form min/max nodes.
5910 if (Subtarget->hasSSE2() &&
5911 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5912 if (Cond.getOpcode() == ISD::SETCC) {
5913 // Get the LHS/RHS of the select.
5914 SDOperand LHS = N->getOperand(1);
5915 SDOperand RHS = N->getOperand(2);
5916 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5918 unsigned Opcode = 0;
5919 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5922 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5925 if (!UnsafeFPMath) break;
5927 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5929 Opcode = X86ISD::FMIN;
5932 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5935 if (!UnsafeFPMath) break;
5937 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5939 Opcode = X86ISD::FMAX;
5942 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5945 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5948 if (!UnsafeFPMath) break;
5950 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5952 Opcode = X86ISD::FMIN;
5955 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5958 if (!UnsafeFPMath) break;
5960 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5962 Opcode = X86ISD::FMAX;
5968 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5976 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
5977 static SDOperand PerformSTORECombine(StoreSDNode *St, SelectionDAG &DAG,
5978 const X86Subtarget *Subtarget) {
5979 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
5980 // the FP state in cases where an emms may be missing.
5981 // A preferable solution to the general problem is to figure out the right
5982 // places to insert EMMS. This qualifies as a quick hack.
5983 if (MVT::isVector(St->getValue().getValueType()) &&
5984 MVT::getSizeInBits(St->getValue().getValueType()) == 64 &&
5985 isa<LoadSDNode>(St->getValue()) &&
5986 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
5987 St->getChain().hasOneUse() && !St->isVolatile()) {
5988 SDNode* LdVal = St->getValue().Val;
5990 int TokenFactorIndex = -1;
5991 SmallVector<SDOperand, 8> Ops;
5992 SDNode* ChainVal = St->getChain().Val;
5993 // Must be a store of a load. We currently handle two cases: the load
5994 // is a direct child, and it's under an intervening TokenFactor. It is
5995 // possible to dig deeper under nested TokenFactors.
5996 if (ChainVal == LdVal)
5997 Ld = cast<LoadSDNode>(St->getChain());
5998 else if (St->getValue().hasOneUse() &&
5999 ChainVal->getOpcode() == ISD::TokenFactor) {
6000 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6001 if (ChainVal->getOperand(i).Val == LdVal) {
6002 TokenFactorIndex = i;
6003 Ld = cast<LoadSDNode>(St->getValue());
6005 Ops.push_back(ChainVal->getOperand(i));
6009 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6010 if (Subtarget->is64Bit()) {
6011 SDOperand NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6012 Ld->getBasePtr(), Ld->getSrcValue(),
6013 Ld->getSrcValueOffset(), Ld->isVolatile(),
6014 Ld->getAlignment());
6015 SDOperand NewChain = NewLd.getValue(1);
6016 if (TokenFactorIndex != -1) {
6017 Ops.push_back(NewLd);
6018 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6021 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6022 St->getSrcValue(), St->getSrcValueOffset(),
6023 St->isVolatile(), St->getAlignment());
6026 // Otherwise, lower to two 32-bit copies.
6027 SDOperand LoAddr = Ld->getBasePtr();
6028 SDOperand HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6029 DAG.getConstant(MVT::i32, 4));
6031 SDOperand LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6032 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6033 Ld->isVolatile(), Ld->getAlignment());
6034 SDOperand HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6035 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6037 MinAlign(Ld->getAlignment(), 4));
6039 SDOperand NewChain = LoLd.getValue(1);
6040 if (TokenFactorIndex != -1) {
6041 Ops.push_back(LoLd);
6042 Ops.push_back(HiLd);
6043 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6047 LoAddr = St->getBasePtr();
6048 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6049 DAG.getConstant(MVT::i32, 4));
6051 SDOperand LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6052 St->getSrcValue(), St->getSrcValueOffset(),
6053 St->isVolatile(), St->getAlignment());
6054 SDOperand HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6055 St->getSrcValue(), St->getSrcValueOffset()+4,
6057 MinAlign(St->getAlignment(), 4));
6058 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6064 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6065 /// X86ISD::FXOR nodes.
6066 static SDOperand PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6067 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6068 // F[X]OR(0.0, x) -> x
6069 // F[X]OR(x, 0.0) -> x
6070 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6071 if (C->getValueAPF().isPosZero())
6072 return N->getOperand(1);
6073 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6074 if (C->getValueAPF().isPosZero())
6075 return N->getOperand(0);
6079 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6080 static SDOperand PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6081 // FAND(0.0, x) -> 0.0
6082 // FAND(x, 0.0) -> 0.0
6083 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6084 if (C->getValueAPF().isPosZero())
6085 return N->getOperand(0);
6086 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6087 if (C->getValueAPF().isPosZero())
6088 return N->getOperand(1);
6093 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
6094 DAGCombinerInfo &DCI) const {
6095 SelectionDAG &DAG = DCI.DAG;
6096 switch (N->getOpcode()) {
6098 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, Subtarget);
6099 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
6101 return PerformSTORECombine(cast<StoreSDNode>(N), DAG, Subtarget);
6103 case X86ISD::FOR: return PerformFORCombine(N, DAG);
6104 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
6110 //===----------------------------------------------------------------------===//
6111 // X86 Inline Assembly Support
6112 //===----------------------------------------------------------------------===//
6114 /// getConstraintType - Given a constraint letter, return the type of
6115 /// constraint it is for this target.
6116 X86TargetLowering::ConstraintType
6117 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
6118 if (Constraint.size() == 1) {
6119 switch (Constraint[0]) {
6128 return C_RegisterClass;
6133 return TargetLowering::getConstraintType(Constraint);
6136 /// LowerXConstraint - try to replace an X constraint, which matches anything,
6137 /// with another that has more specific requirements based on the type of the
6138 /// corresponding operand.
6139 void X86TargetLowering::lowerXConstraint(MVT::ValueType ConstraintVT,
6140 std::string& s) const {
6141 if (MVT::isFloatingPoint(ConstraintVT)) {
6142 if (Subtarget->hasSSE2())
6144 else if (Subtarget->hasSSE1())
6149 return TargetLowering::lowerXConstraint(ConstraintVT, s);
6152 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6153 /// vector. If it is invalid, don't add anything to Ops.
6154 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
6156 std::vector<SDOperand>&Ops,
6157 SelectionDAG &DAG) {
6158 SDOperand Result(0, 0);
6160 switch (Constraint) {
6163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6164 if (C->getValue() <= 31) {
6165 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6171 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6172 if (C->getValue() <= 255) {
6173 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6179 // Literal immediates are always ok.
6180 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
6181 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
6185 // If we are in non-pic codegen mode, we allow the address of a global (with
6186 // an optional displacement) to be used with 'i'.
6187 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
6190 // Match either (GA) or (GA+C)
6192 Offset = GA->getOffset();
6193 } else if (Op.getOpcode() == ISD::ADD) {
6194 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6195 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6197 Offset = GA->getOffset()+C->getValue();
6199 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6200 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6202 Offset = GA->getOffset()+C->getValue();
6209 // If addressing this global requires a load (e.g. in PIC mode), we can't
6211 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
6215 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
6221 // Otherwise, not valid for this mode.
6227 Ops.push_back(Result);
6230 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6233 std::vector<unsigned> X86TargetLowering::
6234 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6235 MVT::ValueType VT) const {
6236 if (Constraint.size() == 1) {
6237 // FIXME: not handling fp-stack yet!
6238 switch (Constraint[0]) { // GCC X86 Constraint Letters
6239 default: break; // Unknown constraint letter
6240 case 'A': // EAX/EDX
6241 if (VT == MVT::i32 || VT == MVT::i64)
6242 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
6244 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
6247 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
6248 else if (VT == MVT::i16)
6249 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
6250 else if (VT == MVT::i8)
6251 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
6252 else if (VT == MVT::i64)
6253 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
6258 return std::vector<unsigned>();
6261 std::pair<unsigned, const TargetRegisterClass*>
6262 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6263 MVT::ValueType VT) const {
6264 // First, see if this is a constraint that directly corresponds to an LLVM
6266 if (Constraint.size() == 1) {
6267 // GCC Constraint Letters
6268 switch (Constraint[0]) {
6270 case 'r': // GENERAL_REGS
6271 case 'R': // LEGACY_REGS
6272 case 'l': // INDEX_REGS
6273 if (VT == MVT::i64 && Subtarget->is64Bit())
6274 return std::make_pair(0U, X86::GR64RegisterClass);
6276 return std::make_pair(0U, X86::GR32RegisterClass);
6277 else if (VT == MVT::i16)
6278 return std::make_pair(0U, X86::GR16RegisterClass);
6279 else if (VT == MVT::i8)
6280 return std::make_pair(0U, X86::GR8RegisterClass);
6282 case 'y': // MMX_REGS if MMX allowed.
6283 if (!Subtarget->hasMMX()) break;
6284 return std::make_pair(0U, X86::VR64RegisterClass);
6286 case 'Y': // SSE_REGS if SSE2 allowed
6287 if (!Subtarget->hasSSE2()) break;
6289 case 'x': // SSE_REGS if SSE1 allowed
6290 if (!Subtarget->hasSSE1()) break;
6294 // Scalar SSE types.
6297 return std::make_pair(0U, X86::FR32RegisterClass);
6300 return std::make_pair(0U, X86::FR64RegisterClass);
6308 return std::make_pair(0U, X86::VR128RegisterClass);
6314 // Use the default implementation in TargetLowering to convert the register
6315 // constraint into a member of a register class.
6316 std::pair<unsigned, const TargetRegisterClass*> Res;
6317 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6319 // Not found as a standard register?
6320 if (Res.second == 0) {
6321 // GCC calls "st(0)" just plain "st".
6322 if (StringsEqualNoCase("{st}", Constraint)) {
6323 Res.first = X86::ST0;
6324 Res.second = X86::RFP80RegisterClass;
6330 // Otherwise, check to see if this is a register class of the wrong value
6331 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6332 // turn into {ax},{dx}.
6333 if (Res.second->hasType(VT))
6334 return Res; // Correct type already, nothing to do.
6336 // All of the single-register GCC register classes map their values onto
6337 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6338 // really want an 8-bit or 32-bit register, map to the appropriate register
6339 // class and return the appropriate register.
6340 if (Res.second != X86::GR16RegisterClass)
6343 if (VT == MVT::i8) {
6344 unsigned DestReg = 0;
6345 switch (Res.first) {
6347 case X86::AX: DestReg = X86::AL; break;
6348 case X86::DX: DestReg = X86::DL; break;
6349 case X86::CX: DestReg = X86::CL; break;
6350 case X86::BX: DestReg = X86::BL; break;
6353 Res.first = DestReg;
6354 Res.second = Res.second = X86::GR8RegisterClass;
6356 } else if (VT == MVT::i32) {
6357 unsigned DestReg = 0;
6358 switch (Res.first) {
6360 case X86::AX: DestReg = X86::EAX; break;
6361 case X86::DX: DestReg = X86::EDX; break;
6362 case X86::CX: DestReg = X86::ECX; break;
6363 case X86::BX: DestReg = X86::EBX; break;
6364 case X86::SI: DestReg = X86::ESI; break;
6365 case X86::DI: DestReg = X86::EDI; break;
6366 case X86::BP: DestReg = X86::EBP; break;
6367 case X86::SP: DestReg = X86::ESP; break;
6370 Res.first = DestReg;
6371 Res.second = Res.second = X86::GR32RegisterClass;
6373 } else if (VT == MVT::i64) {
6374 unsigned DestReg = 0;
6375 switch (Res.first) {
6377 case X86::AX: DestReg = X86::RAX; break;
6378 case X86::DX: DestReg = X86::RDX; break;
6379 case X86::CX: DestReg = X86::RCX; break;
6380 case X86::BX: DestReg = X86::RBX; break;
6381 case X86::SI: DestReg = X86::RSI; break;
6382 case X86::DI: DestReg = X86::RDI; break;
6383 case X86::BP: DestReg = X86::RBP; break;
6384 case X86::SP: DestReg = X86::RSP; break;
6387 Res.first = DestReg;
6388 Res.second = Res.second = X86::GR64RegisterClass;