1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/SmallSet.h"
41 #include "llvm/ADT/StringExtras.h"
44 // Forward declarations.
45 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
47 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
48 : TargetLowering(TM) {
49 Subtarget = &TM.getSubtarget<X86Subtarget>();
50 X86ScalarSSEf64 = Subtarget->hasSSE2();
51 X86ScalarSSEf32 = Subtarget->hasSSE1();
52 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
56 RegInfo = TM.getRegisterInfo();
58 // Set up the TargetLowering object.
60 // X86 is weird, it always uses i8 for shift amounts and setcc results.
61 setShiftAmountType(MVT::i8);
62 setSetCCResultContents(ZeroOrOneSetCCResult);
63 setSchedulingPreference(SchedulingForRegPressure);
64 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
65 setStackPointerRegisterToSaveRestore(X86StackPtr);
67 if (Subtarget->isTargetDarwin()) {
68 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
69 setUseUnderscoreSetJmp(false);
70 setUseUnderscoreLongJmp(false);
71 } else if (Subtarget->isTargetMingw()) {
72 // MS runtime is weird: it exports _setjmp, but longjmp!
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(false);
76 setUseUnderscoreSetJmp(true);
77 setUseUnderscoreLongJmp(true);
80 // Set up the register classes.
81 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
82 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
83 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
84 if (Subtarget->is64Bit())
85 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
87 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 // We don't accept any truncstore of integer registers.
90 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
92 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
93 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
94 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
95 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
97 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
99 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
101 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
103 if (Subtarget->is64Bit()) {
104 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
105 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
108 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
109 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
111 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
114 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
116 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
117 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
118 // SSE has no i16 to fp conversion, only i32
119 if (X86ScalarSSEf32) {
120 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
121 // f32 and f64 cases are Legal, f80 case is not
122 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
124 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
125 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
128 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
129 // are Legal, f80 is custom lowered.
130 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
131 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
133 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
135 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
136 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
138 if (X86ScalarSSEf32) {
139 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
140 // f32 and f64 cases are Legal, f80 case is not
141 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
143 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
144 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
147 // Handle FP_TO_UINT by promoting the destination to a larger signed
149 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
150 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
151 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
153 if (Subtarget->is64Bit()) {
154 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
155 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
158 // Expand FP_TO_UINT into a select.
159 // FIXME: We would like to use a Custom expander here eventually to do
160 // the optimal thing for SSE vs. the default expansion in the legalizer.
161 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
163 // With SSE3 we can use fisttpll to convert to a signed i64.
164 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
167 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
168 if (!X86ScalarSSEf64) {
169 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
170 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
173 // Scalar integer divide and remainder are lowered to use operations that
174 // produce two results, to match the available instructions. This exposes
175 // the two-result form to trivial CSE, which is able to combine x/y and x%y
176 // into a single instruction.
178 // Scalar integer multiply-high is also lowered to use two-result
179 // operations, to match the available instructions. However, plain multiply
180 // (low) operations are left as Legal, as there are single-result
181 // instructions for this in x86. Using the two-result multiply instructions
182 // when both high and low results are needed must be arranged by dagcombine.
183 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
184 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
185 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
186 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
187 setOperationAction(ISD::SREM , MVT::i8 , Expand);
188 setOperationAction(ISD::UREM , MVT::i8 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
193 setOperationAction(ISD::SREM , MVT::i16 , Expand);
194 setOperationAction(ISD::UREM , MVT::i16 , Expand);
195 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
196 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
197 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
198 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
199 setOperationAction(ISD::SREM , MVT::i32 , Expand);
200 setOperationAction(ISD::UREM , MVT::i32 , Expand);
201 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
202 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
203 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
205 setOperationAction(ISD::SREM , MVT::i64 , Expand);
206 setOperationAction(ISD::UREM , MVT::i64 , Expand);
208 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
209 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
210 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
211 setOperationAction(ISD::SELECT_CC , 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::f32 , Expand);
219 setOperationAction(ISD::FREM , MVT::f64 , Expand);
220 setOperationAction(ISD::FREM , MVT::f80 , Expand);
221 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
223 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
224 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
225 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
226 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
227 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
228 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
229 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
230 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
231 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
232 if (Subtarget->is64Bit()) {
233 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
234 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
235 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
238 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
239 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
241 // These should be promoted to a larger select which is supported.
242 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
243 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
244 // X86 wants to expand cmov itself.
245 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
246 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
248 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
249 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
251 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
252 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
254 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
255 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
258 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
260 // X86 ret instruction may pop stack.
261 setOperationAction(ISD::RET , MVT::Other, Custom);
262 if (!Subtarget->is64Bit())
263 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
266 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
267 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
268 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
269 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
270 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
271 if (Subtarget->is64Bit()) {
272 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
273 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
274 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
275 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
277 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
278 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
279 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
280 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
281 if (Subtarget->is64Bit()) {
282 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
283 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
284 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
287 if (Subtarget->hasSSE1())
288 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
290 if (!Subtarget->hasSSE2())
291 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
293 setOperationAction(ISD::ATOMIC_LCS , MVT::i8, Custom);
294 setOperationAction(ISD::ATOMIC_LCS , MVT::i16, Custom);
295 setOperationAction(ISD::ATOMIC_LCS , MVT::i32, Custom);
296 setOperationAction(ISD::ATOMIC_LCS , MVT::i64, Custom);
298 // Use the default ISD::LOCATION, ISD::DECLARE expansion.
299 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
300 // FIXME - use subtarget debug flags
301 if (!Subtarget->isTargetDarwin() &&
302 !Subtarget->isTargetELF() &&
303 !Subtarget->isTargetCygMing())
304 setOperationAction(ISD::LABEL, MVT::Other, Expand);
306 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
307 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
308 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
309 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
310 if (Subtarget->is64Bit()) {
312 setExceptionPointerRegister(X86::RAX);
313 setExceptionSelectorRegister(X86::RDX);
315 setExceptionPointerRegister(X86::EAX);
316 setExceptionSelectorRegister(X86::EDX);
318 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
320 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
322 setOperationAction(ISD::TRAP, MVT::Other, Legal);
324 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
325 setOperationAction(ISD::VASTART , MVT::Other, Custom);
326 setOperationAction(ISD::VAARG , MVT::Other, Expand);
327 setOperationAction(ISD::VAEND , MVT::Other, Expand);
328 if (Subtarget->is64Bit())
329 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
331 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
333 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
334 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
335 if (Subtarget->is64Bit())
336 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
337 if (Subtarget->isTargetCygMing())
338 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
340 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
342 if (X86ScalarSSEf64) {
343 // f32 and f64 use SSE.
344 // Set up the FP register classes.
345 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
346 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
348 // Use ANDPD to simulate FABS.
349 setOperationAction(ISD::FABS , MVT::f64, Custom);
350 setOperationAction(ISD::FABS , MVT::f32, Custom);
352 // Use XORP to simulate FNEG.
353 setOperationAction(ISD::FNEG , MVT::f64, Custom);
354 setOperationAction(ISD::FNEG , MVT::f32, Custom);
356 // Use ANDPD and ORPD to simulate FCOPYSIGN.
357 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
358 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
360 // We don't support sin/cos/fmod
361 setOperationAction(ISD::FSIN , MVT::f64, Expand);
362 setOperationAction(ISD::FCOS , MVT::f64, Expand);
363 setOperationAction(ISD::FSIN , MVT::f32, Expand);
364 setOperationAction(ISD::FCOS , MVT::f32, Expand);
366 // Expand FP immediates into loads from the stack, except for the special
368 addLegalFPImmediate(APFloat(+0.0)); // xorpd
369 addLegalFPImmediate(APFloat(+0.0f)); // xorps
371 // Floating truncations from f80 and extensions to f80 go through memory.
372 // If optimizing, we lie about this though and handle it in
373 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
375 setConvertAction(MVT::f32, MVT::f80, Expand);
376 setConvertAction(MVT::f64, MVT::f80, Expand);
377 setConvertAction(MVT::f80, MVT::f32, Expand);
378 setConvertAction(MVT::f80, MVT::f64, Expand);
380 } else if (X86ScalarSSEf32) {
381 // Use SSE for f32, x87 for f64.
382 // Set up the FP register classes.
383 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
384 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
386 // Use ANDPS to simulate FABS.
387 setOperationAction(ISD::FABS , MVT::f32, Custom);
389 // Use XORP to simulate FNEG.
390 setOperationAction(ISD::FNEG , MVT::f32, Custom);
392 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
394 // Use ANDPS and ORPS to simulate FCOPYSIGN.
395 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
396 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
398 // We don't support sin/cos/fmod
399 setOperationAction(ISD::FSIN , MVT::f32, Expand);
400 setOperationAction(ISD::FCOS , MVT::f32, Expand);
402 // Special cases we handle for FP constants.
403 addLegalFPImmediate(APFloat(+0.0f)); // xorps
404 addLegalFPImmediate(APFloat(+0.0)); // FLD0
405 addLegalFPImmediate(APFloat(+1.0)); // FLD1
406 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
407 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
409 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
410 // this though and handle it in InstructionSelectPreprocess so that
411 // dagcombine2 can hack on these.
413 setConvertAction(MVT::f32, MVT::f64, Expand);
414 setConvertAction(MVT::f32, MVT::f80, Expand);
415 setConvertAction(MVT::f80, MVT::f32, Expand);
416 setConvertAction(MVT::f64, MVT::f32, Expand);
417 // And x87->x87 truncations also.
418 setConvertAction(MVT::f80, MVT::f64, Expand);
422 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
423 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
426 // f32 and f64 in x87.
427 // Set up the FP register classes.
428 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
429 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
431 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
432 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
433 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
434 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
436 // Floating truncations go through memory. If optimizing, we lie about
437 // this though and handle it in InstructionSelectPreprocess so that
438 // dagcombine2 can hack on these.
440 setConvertAction(MVT::f80, MVT::f32, Expand);
441 setConvertAction(MVT::f64, MVT::f32, Expand);
442 setConvertAction(MVT::f80, MVT::f64, Expand);
446 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
447 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
449 addLegalFPImmediate(APFloat(+0.0)); // FLD0
450 addLegalFPImmediate(APFloat(+1.0)); // FLD1
451 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
452 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
453 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
454 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
455 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
456 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
459 // Long double always uses X87.
460 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
461 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
462 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
464 APFloat TmpFlt(+0.0);
465 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
466 addLegalFPImmediate(TmpFlt); // FLD0
468 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
469 APFloat TmpFlt2(+1.0);
470 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
471 addLegalFPImmediate(TmpFlt2); // FLD1
472 TmpFlt2.changeSign();
473 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
477 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
478 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
481 // Always use a library call for pow.
482 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
483 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
484 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
486 // First set operation action for all vector types to expand. Then we
487 // will selectively turn on ones that can be effectively codegen'd.
488 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
489 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
490 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
491 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
492 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
493 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
494 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
495 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
496 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
497 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
498 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
499 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
500 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
501 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
502 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
503 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
504 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
505 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
506 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
507 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
508 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
509 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
510 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
511 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
512 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
513 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
514 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
515 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
516 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
517 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
518 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
519 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
520 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
521 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
522 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
523 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
524 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
525 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
526 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
529 if (Subtarget->hasMMX()) {
530 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
531 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
532 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
533 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
535 // FIXME: add MMX packed arithmetics
537 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
538 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
539 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
540 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
542 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
543 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
544 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
545 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
547 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
548 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
550 setOperationAction(ISD::AND, MVT::v8i8, Promote);
551 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
552 setOperationAction(ISD::AND, MVT::v4i16, Promote);
553 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
554 setOperationAction(ISD::AND, MVT::v2i32, Promote);
555 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
556 setOperationAction(ISD::AND, MVT::v1i64, Legal);
558 setOperationAction(ISD::OR, MVT::v8i8, Promote);
559 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
560 setOperationAction(ISD::OR, MVT::v4i16, Promote);
561 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
562 setOperationAction(ISD::OR, MVT::v2i32, Promote);
563 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
564 setOperationAction(ISD::OR, MVT::v1i64, Legal);
566 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
567 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
568 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
569 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
570 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
571 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
572 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
574 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
575 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
576 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
577 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
578 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
579 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
580 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
582 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
583 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
584 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
585 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
587 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
588 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
589 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
590 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
592 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
593 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
594 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
597 if (Subtarget->hasSSE1()) {
598 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
600 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
601 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
602 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
603 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
604 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
605 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
606 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
607 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
608 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
609 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
610 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
613 if (Subtarget->hasSSE2()) {
614 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
615 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
616 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
617 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
618 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
620 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
621 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
624 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
625 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
626 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
627 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
628 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
629 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
630 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
631 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
632 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
633 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
634 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
636 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
637 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
638 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
639 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
640 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
642 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
643 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
644 // Do not attempt to custom lower non-power-of-2 vectors
645 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
647 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
648 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
649 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
652 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
655 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
656 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
657 if (Subtarget->is64Bit()) {
658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
659 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
662 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
663 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
664 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
665 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
666 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
667 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
668 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
669 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
670 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
671 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
672 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
673 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
676 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
678 // Custom lower v2i64 and v2f64 selects.
679 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
680 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
681 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
682 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
685 if (Subtarget->hasSSE41()) {
686 // FIXME: Do we need to handle scalar-to-vector here?
687 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
689 // i8 and i16 vectors are custom , because the source register and source
690 // source memory operand types are not the same width. f32 vectors are
691 // custom since the immediate controlling the insert encodes additional
693 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
694 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
696 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
698 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
699 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
701 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
703 if (Subtarget->is64Bit()) {
704 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
705 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
709 // We want to custom lower some of our intrinsics.
710 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
712 // We have target-specific dag combine patterns for the following nodes:
713 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
714 setTargetDAGCombine(ISD::SELECT);
715 setTargetDAGCombine(ISD::STORE);
717 computeRegisterProperties();
719 // FIXME: These should be based on subtarget info. Plus, the values should
720 // be smaller when we are in optimizing for size mode.
721 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
722 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
723 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
724 allowUnalignedMemoryAccesses = true; // x86 supports it!
725 setPrefLoopAlignment(16);
730 X86TargetLowering::getSetCCResultType(const SDOperand &) const {
735 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
736 /// the desired ByVal argument alignment.
737 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
740 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
741 if (VTy->getBitWidth() == 128)
743 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
744 unsigned EltAlign = 0;
745 getMaxByValAlign(ATy->getElementType(), EltAlign);
746 if (EltAlign > MaxAlign)
748 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
749 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
750 unsigned EltAlign = 0;
751 getMaxByValAlign(STy->getElementType(i), EltAlign);
752 if (EltAlign > MaxAlign)
761 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
762 /// function arguments in the caller parameter area. For X86, aggregates
763 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
764 /// are at 4-byte boundaries.
765 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
766 if (Subtarget->is64Bit())
767 return getTargetData()->getABITypeAlignment(Ty);
769 if (Subtarget->hasSSE1())
770 getMaxByValAlign(Ty, Align);
774 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
776 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
777 SelectionDAG &DAG) const {
778 if (usesGlobalOffsetTable())
779 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
780 if (!Subtarget->isPICStyleRIPRel())
781 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
785 //===----------------------------------------------------------------------===//
786 // Return Value Calling Convention Implementation
787 //===----------------------------------------------------------------------===//
789 #include "X86GenCallingConv.inc"
791 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
792 /// exists skip possible ISD:TokenFactor.
793 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
794 if (Chain.getOpcode() == X86ISD::TAILCALL) {
796 } else if (Chain.getOpcode() == ISD::TokenFactor) {
797 if (Chain.getNumOperands() &&
798 Chain.getOperand(0).getOpcode() == X86ISD::TAILCALL)
799 return Chain.getOperand(0);
804 /// LowerRET - Lower an ISD::RET node.
805 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
806 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
808 SmallVector<CCValAssign, 16> RVLocs;
809 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
810 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
811 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
812 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
814 // If this is the first return lowered for this function, add the regs to the
815 // liveout set for the function.
816 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
817 for (unsigned i = 0; i != RVLocs.size(); ++i)
818 if (RVLocs[i].isRegLoc())
819 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
821 SDOperand Chain = Op.getOperand(0);
823 // Handle tail call return.
824 Chain = GetPossiblePreceedingTailCall(Chain);
825 if (Chain.getOpcode() == X86ISD::TAILCALL) {
826 SDOperand TailCall = Chain;
827 SDOperand TargetAddress = TailCall.getOperand(1);
828 SDOperand StackAdjustment = TailCall.getOperand(2);
829 assert(((TargetAddress.getOpcode() == ISD::Register &&
830 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
831 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
832 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
833 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
834 "Expecting an global address, external symbol, or register");
835 assert(StackAdjustment.getOpcode() == ISD::Constant &&
836 "Expecting a const value");
838 SmallVector<SDOperand,8> Operands;
839 Operands.push_back(Chain.getOperand(0));
840 Operands.push_back(TargetAddress);
841 Operands.push_back(StackAdjustment);
842 // Copy registers used by the call. Last operand is a flag so it is not
844 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
845 Operands.push_back(Chain.getOperand(i));
847 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
854 SmallVector<SDOperand, 6> RetOps;
855 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
856 // Operand #1 = Bytes To Pop
857 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
859 // Copy the result values into the output registers.
860 for (unsigned i = 0; i != RVLocs.size(); ++i) {
861 CCValAssign &VA = RVLocs[i];
862 assert(VA.isRegLoc() && "Can only return in registers!");
863 SDOperand ValToCopy = Op.getOperand(i*2+1);
865 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
866 // the RET instruction and handled by the FP Stackifier.
867 if (RVLocs[i].getLocReg() == X86::ST0 ||
868 RVLocs[i].getLocReg() == X86::ST1) {
869 // If this is a copy from an xmm register to ST(0), use an FPExtend to
870 // change the value to the FP stack register class.
871 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
872 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
873 RetOps.push_back(ValToCopy);
874 // Don't emit a copytoreg.
878 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
879 Flag = Chain.getValue(1);
882 // The x86-64 ABI for returning structs by value requires that we copy
883 // the sret argument into %rax for the return. We saved the argument into
884 // a virtual register in the entry block, so now we copy the value out
886 if (Subtarget->is64Bit() &&
887 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
888 MachineFunction &MF = DAG.getMachineFunction();
889 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
890 unsigned Reg = FuncInfo->getSRetReturnReg();
892 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
893 FuncInfo->setSRetReturnReg(Reg);
895 SDOperand Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
897 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
898 Flag = Chain.getValue(1);
901 RetOps[0] = Chain; // Update chain.
903 // Add the flag if we have it.
905 RetOps.push_back(Flag);
907 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
911 /// LowerCallResult - Lower the result values of an ISD::CALL into the
912 /// appropriate copies out of appropriate physical registers. This assumes that
913 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
914 /// being lowered. The returns a SDNode with the same number of values as the
916 SDNode *X86TargetLowering::
917 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
918 unsigned CallingConv, SelectionDAG &DAG) {
920 // Assign locations to each value returned by this call.
921 SmallVector<CCValAssign, 16> RVLocs;
922 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
923 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
924 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
926 SmallVector<SDOperand, 8> ResultVals;
928 // Copy all of the result registers out of their specified physreg.
929 for (unsigned i = 0; i != RVLocs.size(); ++i) {
930 MVT::ValueType CopyVT = RVLocs[i].getValVT();
932 // If this is a call to a function that returns an fp value on the floating
933 // point stack, but where we prefer to use the value in xmm registers, copy
934 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
935 if (RVLocs[i].getLocReg() == X86::ST0 &&
936 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
940 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
941 CopyVT, InFlag).getValue(1);
942 SDOperand Val = Chain.getValue(0);
943 InFlag = Chain.getValue(2);
945 if (CopyVT != RVLocs[i].getValVT()) {
946 // Round the F80 the right size, which also moves to the appropriate xmm
948 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
949 // This truncation won't change the value.
950 DAG.getIntPtrConstant(1));
953 ResultVals.push_back(Val);
956 // Merge everything together with a MERGE_VALUES node.
957 ResultVals.push_back(Chain);
958 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
959 &ResultVals[0], ResultVals.size()).Val;
963 //===----------------------------------------------------------------------===//
964 // C & StdCall & Fast Calling Convention implementation
965 //===----------------------------------------------------------------------===//
966 // StdCall calling convention seems to be standard for many Windows' API
967 // routines and around. It differs from C calling convention just a little:
968 // callee should clean up the stack, not caller. Symbols should be also
969 // decorated in some fancy way :) It doesn't support any vector arguments.
970 // For info on fast calling convention see Fast Calling Convention (tail call)
971 // implementation LowerX86_32FastCCCallTo.
973 /// AddLiveIn - This helper function adds the specified physical register to the
974 /// MachineFunction as a live in value. It also creates a corresponding virtual
976 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
977 const TargetRegisterClass *RC) {
978 assert(RC->contains(PReg) && "Not the correct regclass!");
979 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
980 MF.getRegInfo().addLiveIn(PReg, VReg);
984 /// CallIsStructReturn - Determines whether a CALL node uses struct return
986 static bool CallIsStructReturn(SDOperand Op) {
987 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
991 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
994 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
995 /// return semantics.
996 static bool ArgsAreStructReturn(SDOperand Op) {
997 unsigned NumArgs = Op.Val->getNumValues() - 1;
1001 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1004 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1005 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1007 bool X86TargetLowering::IsCalleePop(SDOperand Op) {
1008 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1012 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1015 case CallingConv::X86_StdCall:
1016 return !Subtarget->is64Bit();
1017 case CallingConv::X86_FastCall:
1018 return !Subtarget->is64Bit();
1019 case CallingConv::Fast:
1020 return PerformTailCallOpt;
1024 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1025 /// FORMAL_ARGUMENTS node.
1026 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDOperand Op) const {
1027 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1029 if (Subtarget->is64Bit()) {
1030 if (Subtarget->isTargetWin64())
1031 return CC_X86_Win64_C;
1033 if (CC == CallingConv::Fast && PerformTailCallOpt)
1034 return CC_X86_64_TailCall;
1040 if (CC == CallingConv::X86_FastCall)
1041 return CC_X86_32_FastCall;
1042 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1043 return CC_X86_32_TailCall;
1048 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1049 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1051 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDOperand Op) {
1052 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1053 if (CC == CallingConv::X86_FastCall)
1055 else if (CC == CallingConv::X86_StdCall)
1060 /// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
1061 /// possibly be overwritten when lowering the outgoing arguments in a tail
1062 /// call. Currently the implementation of this call is very conservative and
1063 /// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
1064 /// virtual registers would be overwritten by direct lowering.
1065 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op,
1066 MachineFrameInfo * MFI) {
1067 RegisterSDNode * OpReg = NULL;
1068 FrameIndexSDNode * FrameIdxNode = NULL;
1070 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
1071 (Op.getOpcode()== ISD::CopyFromReg &&
1072 (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
1073 (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
1074 (Op.getOpcode() == ISD::LOAD &&
1075 (FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op.getOperand(1))) &&
1076 (MFI->isFixedObjectIndex((FrameIdx = FrameIdxNode->getIndex()))) &&
1077 (MFI->getObjectOffset(FrameIdx) >= 0)))
1082 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1083 /// in a register before calling.
1084 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1085 return !IsTailCall && !Is64Bit &&
1086 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1087 Subtarget->isPICStyleGOT();
1091 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1092 /// address to be loaded in a register.
1094 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1095 return !Is64Bit && IsTailCall &&
1096 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1097 Subtarget->isPICStyleGOT();
1100 /// CopyTailCallClobberedArgumentsToVRegs - Create virtual registers for all
1101 /// arguments to force loading and guarantee that arguments sourcing from
1102 /// incomming parameters are not overwriting each other.
1104 CopyTailCallClobberedArgumentsToVRegs(SDOperand Chain,
1105 SmallVector<std::pair<unsigned, SDOperand>, 8> &TailCallClobberedVRegs,
1107 MachineFunction &MF,
1108 const TargetLowering * TL) {
1111 for (unsigned i = 0, e = TailCallClobberedVRegs.size(); i != e; i++) {
1112 SDOperand Arg = TailCallClobberedVRegs[i].second;
1113 unsigned Idx = TailCallClobberedVRegs[i].first;
1116 createVirtualRegister(TL->getRegClassFor(Arg.getValueType()));
1117 Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
1118 InFlag = Chain.getValue(1);
1119 Arg = DAG.getCopyFromReg(Chain, VReg, Arg.getValueType(), InFlag);
1120 TailCallClobberedVRegs[i] = std::make_pair(Idx, Arg);
1121 Chain = Arg.getValue(1);
1122 InFlag = Arg.getValue(2);
1127 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1128 /// by "Src" to address "Dst" with size and alignment information specified by
1129 /// the specific parameter attribute. The copy will be passed as a byval
1130 /// function parameter.
1132 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1133 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1134 SDOperand SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1135 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1136 /*AlwaysInline=*/true,
1140 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
1141 const CCValAssign &VA,
1142 MachineFrameInfo *MFI,
1144 SDOperand Root, unsigned i) {
1145 // Create the nodes corresponding to a load from this parameter slot.
1146 ISD::ArgFlagsTy Flags =
1147 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1148 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1149 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1151 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1152 // changed with more analysis.
1153 // In case of tail call optimization mark all arguments mutable. Since they
1154 // could be overwritten by lowering of arguments in case of a tail call.
1155 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1156 VA.getLocMemOffset(), isImmutable);
1157 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1158 if (Flags.isByVal())
1160 return DAG.getLoad(VA.getValVT(), Root, FIN,
1161 PseudoSourceValue::getFixedStack(), FI);
1165 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
1166 MachineFunction &MF = DAG.getMachineFunction();
1167 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1169 const Function* Fn = MF.getFunction();
1170 if (Fn->hasExternalLinkage() &&
1171 Subtarget->isTargetCygMing() &&
1172 Fn->getName() == "main")
1173 FuncInfo->setForceFramePointer(true);
1175 // Decorate the function name.
1176 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1178 MachineFrameInfo *MFI = MF.getFrameInfo();
1179 SDOperand Root = Op.getOperand(0);
1180 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1181 unsigned CC = MF.getFunction()->getCallingConv();
1182 bool Is64Bit = Subtarget->is64Bit();
1183 bool IsWin64 = Subtarget->isTargetWin64();
1185 assert(!(isVarArg && CC == CallingConv::Fast) &&
1186 "Var args not supported with calling convention fastcc");
1188 // Assign locations to all of the incoming arguments.
1189 SmallVector<CCValAssign, 16> ArgLocs;
1190 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1191 CCInfo.AnalyzeFormalArguments(Op.Val, CCAssignFnForNode(Op));
1193 SmallVector<SDOperand, 8> ArgValues;
1194 unsigned LastVal = ~0U;
1195 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1196 CCValAssign &VA = ArgLocs[i];
1197 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1199 assert(VA.getValNo() != LastVal &&
1200 "Don't support value assigned to multiple locs yet");
1201 LastVal = VA.getValNo();
1203 if (VA.isRegLoc()) {
1204 MVT::ValueType RegVT = VA.getLocVT();
1205 TargetRegisterClass *RC;
1206 if (RegVT == MVT::i32)
1207 RC = X86::GR32RegisterClass;
1208 else if (Is64Bit && RegVT == MVT::i64)
1209 RC = X86::GR64RegisterClass;
1210 else if (RegVT == MVT::f32)
1211 RC = X86::FR32RegisterClass;
1212 else if (RegVT == MVT::f64)
1213 RC = X86::FR64RegisterClass;
1214 else if (MVT::isVector(RegVT) && MVT::getSizeInBits(RegVT) == 128)
1215 RC = X86::VR128RegisterClass;
1216 else if (MVT::isVector(RegVT)) {
1217 assert(MVT::getSizeInBits(RegVT) == 64);
1219 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1221 // Darwin calling convention passes MMX values in either GPRs or
1222 // XMMs in x86-64. Other targets pass them in memory.
1223 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1224 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1227 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1232 assert(0 && "Unknown argument type!");
1235 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1236 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1238 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1239 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1241 if (VA.getLocInfo() == CCValAssign::SExt)
1242 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1243 DAG.getValueType(VA.getValVT()));
1244 else if (VA.getLocInfo() == CCValAssign::ZExt)
1245 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1246 DAG.getValueType(VA.getValVT()));
1248 if (VA.getLocInfo() != CCValAssign::Full)
1249 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1251 // Handle MMX values passed in GPRs.
1252 if (Is64Bit && RegVT != VA.getLocVT()) {
1253 if (MVT::getSizeInBits(RegVT) == 64 && RC == X86::GR64RegisterClass)
1254 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1255 else if (RC == X86::VR128RegisterClass) {
1256 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1257 DAG.getConstant(0, MVT::i64));
1258 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1262 ArgValues.push_back(ArgValue);
1264 assert(VA.isMemLoc());
1265 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1269 // The x86-64 ABI for returning structs by value requires that we copy
1270 // the sret argument into %rax for the return. Save the argument into
1271 // a virtual register so that we can access it from the return points.
1272 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1273 MachineFunction &MF = DAG.getMachineFunction();
1274 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1275 unsigned Reg = FuncInfo->getSRetReturnReg();
1277 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1278 FuncInfo->setSRetReturnReg(Reg);
1280 SDOperand Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1281 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1284 unsigned StackSize = CCInfo.getNextStackOffset();
1285 // align stack specially for tail calls
1286 if (CC == CallingConv::Fast)
1287 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1289 // If the function takes variable number of arguments, make a frame index for
1290 // the start of the first vararg value... for expansion of llvm.va_start.
1292 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1293 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1296 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1298 // FIXME: We should really autogenerate these arrays
1299 static const unsigned GPR64ArgRegsWin64[] = {
1300 X86::RCX, X86::RDX, X86::R8, X86::R9
1302 static const unsigned XMMArgRegsWin64[] = {
1303 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1305 static const unsigned GPR64ArgRegs64Bit[] = {
1306 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1308 static const unsigned XMMArgRegs64Bit[] = {
1309 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1310 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1312 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1315 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1316 GPR64ArgRegs = GPR64ArgRegsWin64;
1317 XMMArgRegs = XMMArgRegsWin64;
1319 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1320 GPR64ArgRegs = GPR64ArgRegs64Bit;
1321 XMMArgRegs = XMMArgRegs64Bit;
1323 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1325 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1328 // For X86-64, if there are vararg parameters that are passed via
1329 // registers, then we must store them to their spots on the stack so they
1330 // may be loaded by deferencing the result of va_next.
1331 VarArgsGPOffset = NumIntRegs * 8;
1332 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1333 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1334 TotalNumXMMRegs * 16, 16);
1336 // Store the integer parameter registers.
1337 SmallVector<SDOperand, 8> MemOps;
1338 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1339 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1340 DAG.getIntPtrConstant(VarArgsGPOffset));
1341 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1342 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1343 X86::GR64RegisterClass);
1344 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1346 DAG.getStore(Val.getValue(1), Val, FIN,
1347 PseudoSourceValue::getFixedStack(),
1349 MemOps.push_back(Store);
1350 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1351 DAG.getIntPtrConstant(8));
1354 // Now store the XMM (fp + vector) parameter registers.
1355 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1356 DAG.getIntPtrConstant(VarArgsFPOffset));
1357 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1358 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1359 X86::VR128RegisterClass);
1360 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1362 DAG.getStore(Val.getValue(1), Val, FIN,
1363 PseudoSourceValue::getFixedStack(),
1365 MemOps.push_back(Store);
1366 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1367 DAG.getIntPtrConstant(16));
1369 if (!MemOps.empty())
1370 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1371 &MemOps[0], MemOps.size());
1375 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1376 // arguments and the arguments after the retaddr has been pushed are
1378 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1379 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1380 (StackSize & 7) == 0)
1383 ArgValues.push_back(Root);
1385 // Some CCs need callee pop.
1386 if (IsCalleePop(Op)) {
1387 BytesToPopOnReturn = StackSize; // Callee pops everything.
1388 BytesCallerReserves = 0;
1390 BytesToPopOnReturn = 0; // Callee pops nothing.
1391 // If this is an sret function, the return should pop the hidden pointer.
1392 if (!Is64Bit && ArgsAreStructReturn(Op))
1393 BytesToPopOnReturn = 4;
1394 BytesCallerReserves = StackSize;
1398 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1399 if (CC == CallingConv::X86_FastCall)
1400 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1403 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1405 // Return the new list of results.
1406 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1407 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1411 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1412 const SDOperand &StackPtr,
1413 const CCValAssign &VA,
1416 unsigned LocMemOffset = VA.getLocMemOffset();
1417 SDOperand PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1419 ISD::ArgFlagsTy Flags =
1420 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1421 if (Flags.isByVal()) {
1422 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1424 return DAG.getStore(Chain, Arg, PtrOff,
1425 PseudoSourceValue::getStack(), LocMemOffset);
1428 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1429 /// optimization is performed and it is required.
1431 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1432 SDOperand &OutRetAddr,
1437 if (!IsTailCall || FPDiff==0) return Chain;
1439 // Adjust the Return address stack slot.
1440 MVT::ValueType VT = getPointerTy();
1441 OutRetAddr = getReturnAddressFrameIndex(DAG);
1442 // Load the "old" Return address.
1443 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1444 return SDOperand(OutRetAddr.Val, 1);
1447 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1448 /// optimization is performed and it is required (FPDiff!=0).
1450 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1451 SDOperand Chain, SDOperand RetAddrFrIdx,
1452 bool Is64Bit, int FPDiff) {
1453 // Store the return address to the appropriate stack slot.
1454 if (!FPDiff) return Chain;
1455 // Calculate the new stack slot for the return address.
1456 int SlotSize = Is64Bit ? 8 : 4;
1457 int NewReturnAddrFI =
1458 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1459 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1460 SDOperand NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1461 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1462 PseudoSourceValue::getFixedStack(), NewReturnAddrFI);
1466 /// CopyTailCallByValClobberedRegToVirtReg - Copy arguments with register target
1467 /// which might be overwritten by later byval tail call lowering to a virtual
1470 X86TargetLowering::CopyTailCallByValClobberedRegToVirtReg(bool containsByValArg,
1471 SmallVector< std::pair<unsigned, unsigned>, 8> &TailCallByValClobberedVRegs,
1472 SmallVector<MVT::ValueType, 8> &TailCallByValClobberedVRegTypes,
1473 std::pair<unsigned, SDOperand> &RegToPass,
1474 SDOperand &OutChain,
1476 MachineFunction &MF,
1477 SelectionDAG & DAG) {
1478 if (!containsByValArg) return false;
1480 std::pair<unsigned, unsigned> ArgRegVReg;
1481 MVT::ValueType VT = RegToPass.second.getValueType();
1483 ArgRegVReg.first = RegToPass.first;
1484 ArgRegVReg.second = MF.getRegInfo().createVirtualRegister(getRegClassFor(VT));
1486 // Copy Argument to virtual register.
1487 OutChain = DAG.getCopyToReg(OutChain, ArgRegVReg.second,
1488 RegToPass.second, OutFlag);
1489 OutFlag = OutChain.getValue(1);
1490 // Remember virtual register and type.
1491 TailCallByValClobberedVRegs.push_back(ArgRegVReg);
1492 TailCallByValClobberedVRegTypes.push_back(VT);
1497 /// RestoreTailCallByValClobberedReg - Restore registers which were saved to
1498 /// virtual registers to prevent tail call byval lowering from overwriting
1499 /// parameter registers.
1501 RestoreTailCallByValClobberedRegs(SelectionDAG & DAG, SDOperand Chain,
1502 SmallVector< std::pair<unsigned, unsigned>, 8> &TailCallByValClobberedVRegs,
1503 SmallVector<MVT::ValueType, 8> &TailCallByValClobberedVRegTypes) {
1504 if (TailCallByValClobberedVRegs.size()==0) return Chain;
1506 SmallVector<SDOperand, 8> RegOpChains;
1507 for (unsigned i = 0, e=TailCallByValClobberedVRegs.size(); i != e; i++) {
1509 unsigned DestReg = TailCallByValClobberedVRegs[i].first;
1510 unsigned VirtReg = TailCallByValClobberedVRegs[i].second;
1511 MVT::ValueType VT = TailCallByValClobberedVRegTypes[i];
1512 SDOperand Tmp = DAG.getCopyFromReg(Chain, VirtReg, VT, InFlag);
1513 Chain = DAG.getCopyToReg(Chain, DestReg, Tmp, InFlag);
1514 RegOpChains.push_back(Chain);
1516 if (!RegOpChains.empty())
1517 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1518 &RegOpChains[0], RegOpChains.size());
1522 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
1523 MachineFunction &MF = DAG.getMachineFunction();
1524 MachineFrameInfo * MFI = MF.getFrameInfo();
1525 SDOperand Chain = Op.getOperand(0);
1526 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1527 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1528 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1529 && CC == CallingConv::Fast && PerformTailCallOpt;
1530 SDOperand Callee = Op.getOperand(4);
1531 bool Is64Bit = Subtarget->is64Bit();
1532 bool IsStructRet = CallIsStructReturn(Op);
1534 assert(!(isVarArg && CC == CallingConv::Fast) &&
1535 "Var args not supported with calling convention fastcc");
1537 // Analyze operands of the call, assigning locations to each operand.
1538 SmallVector<CCValAssign, 16> ArgLocs;
1539 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1540 CCInfo.AnalyzeCallOperands(Op.Val, CCAssignFnForNode(Op));
1542 // Get a count of how many bytes are to be pushed on the stack.
1543 unsigned NumBytes = CCInfo.getNextStackOffset();
1544 if (CC == CallingConv::Fast)
1545 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1547 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1548 // arguments and the arguments after the retaddr has been pushed are aligned.
1549 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1550 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1551 (NumBytes & 7) == 0)
1556 // Lower arguments at fp - stackoffset + fpdiff.
1557 unsigned NumBytesCallerPushed =
1558 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1559 FPDiff = NumBytesCallerPushed - NumBytes;
1561 // Set the delta of movement of the returnaddr stackslot.
1562 // But only set if delta is greater than previous delta.
1563 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1564 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1567 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1569 SDOperand RetAddrFrIdx;
1570 // Load return adress for tail calls.
1571 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1574 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1575 SmallVector<std::pair<unsigned, SDOperand>, 8> TailCallClobberedVRegs;
1577 SmallVector<SDOperand, 8> MemOpChains;
1580 bool containsTailCallByValArg = false;
1581 SmallVector<std::pair<unsigned, unsigned>, 8> TailCallByValClobberedVRegs;
1582 SmallVector<MVT::ValueType, 8> TailCallByValClobberedVRegTypes;
1584 // Walk the register/memloc assignments, inserting copies/loads. For tail
1585 // calls, remember all arguments for later special lowering.
1586 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1587 CCValAssign &VA = ArgLocs[i];
1588 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1589 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1590 getArgFlags().isByVal();
1592 // Promote the value if needed.
1593 switch (VA.getLocInfo()) {
1594 default: assert(0 && "Unknown loc info!");
1595 case CCValAssign::Full: break;
1596 case CCValAssign::SExt:
1597 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1599 case CCValAssign::ZExt:
1600 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1602 case CCValAssign::AExt:
1603 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1607 if (VA.isRegLoc()) {
1609 MVT::ValueType RegVT = VA.getLocVT();
1610 if (MVT::isVector(RegVT) && MVT::getSizeInBits(RegVT) == 64)
1611 switch (VA.getLocReg()) {
1614 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1616 // Special case: passing MMX values in GPR registers.
1617 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1620 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1621 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1622 // Special case: passing MMX values in XMM registers.
1623 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1624 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1625 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1626 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1627 getMOVLMask(2, DAG));
1632 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1634 if (!IsTailCall || (IsTailCall && isByVal)) {
1635 assert(VA.isMemLoc());
1636 if (StackPtr.Val == 0)
1637 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1639 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1641 // Remember fact that this call contains byval arguments.
1642 containsTailCallByValArg |= IsTailCall && isByVal;
1643 } else if (IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
1644 TailCallClobberedVRegs.push_back(std::make_pair(i,Arg));
1649 if (!MemOpChains.empty())
1650 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1651 &MemOpChains[0], MemOpChains.size());
1653 // Build a sequence of copy-to-reg nodes chained together with token chain
1654 // and flag operands which copy the outgoing args into registers.
1656 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1657 // Tail call byval lowering might overwrite argument registers so arguments
1658 // passed to be copied to a virtual register for
1659 // later processing.
1660 if (CopyTailCallByValClobberedRegToVirtReg(containsTailCallByValArg,
1661 TailCallByValClobberedVRegs,
1662 TailCallByValClobberedVRegTypes,
1663 RegsToPass[i], Chain, InFlag, MF,
1667 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1669 InFlag = Chain.getValue(1);
1672 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1674 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1675 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1676 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1678 InFlag = Chain.getValue(1);
1680 // If we are tail calling and generating PIC/GOT style code load the address
1681 // of the callee into ecx. The value in ecx is used as target of the tail
1682 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1683 // calls on PIC/GOT architectures. Normally we would just put the address of
1684 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1685 // restored (since ebx is callee saved) before jumping to the target@PLT.
1686 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1687 // Note: The actual moving to ecx is done further down.
1688 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1689 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1690 !G->getGlobal()->hasProtectedVisibility())
1691 Callee = LowerGlobalAddress(Callee, DAG);
1692 else if (isa<ExternalSymbolSDNode>(Callee))
1693 Callee = LowerExternalSymbol(Callee,DAG);
1696 if (Is64Bit && isVarArg) {
1697 // From AMD64 ABI document:
1698 // For calls that may call functions that use varargs or stdargs
1699 // (prototype-less calls or calls to functions containing ellipsis (...) in
1700 // the declaration) %al is used as hidden argument to specify the number
1701 // of SSE registers used. The contents of %al do not need to match exactly
1702 // the number of registers, but must be an ubound on the number of SSE
1703 // registers used and is in the range 0 - 8 inclusive.
1705 // FIXME: Verify this on Win64
1706 // Count the number of XMM registers allocated.
1707 static const unsigned XMMArgRegs[] = {
1708 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1709 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1711 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1713 Chain = DAG.getCopyToReg(Chain, X86::AL,
1714 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1715 InFlag = Chain.getValue(1);
1719 // For tail calls lower the arguments to the 'real' stack slot.
1721 SmallVector<SDOperand, 8> MemOpChains2;
1724 // Do not flag preceeding copytoreg stuff together with the following stuff.
1725 InFlag = SDOperand();
1727 Chain = CopyTailCallClobberedArgumentsToVRegs(Chain, TailCallClobberedVRegs,
1730 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1731 CCValAssign &VA = ArgLocs[i];
1732 if (!VA.isRegLoc()) {
1733 assert(VA.isMemLoc());
1734 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1735 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1736 ISD::ArgFlagsTy Flags =
1737 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1738 // Create frame index.
1739 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1740 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1741 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1742 FIN = DAG.getFrameIndex(FI, getPointerTy());
1744 // Find virtual register for this argument.
1746 for (unsigned idx=0, e= TailCallClobberedVRegs.size(); idx < e; idx++)
1747 if (TailCallClobberedVRegs[idx].first==i) {
1748 Arg = TailCallClobberedVRegs[idx].second;
1752 assert(IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)==false ||
1753 (Found==true && "No corresponding Argument was found"));
1755 if (Flags.isByVal()) {
1756 // Copy relative to framepointer.
1757 SDOperand Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1758 if (StackPtr.Val == 0)
1759 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1760 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1762 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1765 // Store relative to framepointer.
1766 MemOpChains2.push_back(
1767 DAG.getStore(Chain, Arg, FIN,
1768 PseudoSourceValue::getFixedStack(), FI));
1773 if (!MemOpChains2.empty())
1774 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1775 &MemOpChains2[0], MemOpChains2.size());
1777 // Restore byval lowering clobbered registers.
1778 Chain = RestoreTailCallByValClobberedRegs(DAG, Chain,
1779 TailCallByValClobberedVRegs,
1780 TailCallByValClobberedVRegTypes);
1782 // Store the return address to the appropriate stack slot.
1783 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1787 // If the callee is a GlobalAddress node (quite common, every direct call is)
1788 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1789 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1790 // We should use extra load for direct calls to dllimported functions in
1792 if ((IsTailCall || !Is64Bit ||
1793 getTargetMachine().getCodeModel() != CodeModel::Large)
1794 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1795 getTargetMachine(), true))
1796 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1797 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1798 if (IsTailCall || !Is64Bit ||
1799 getTargetMachine().getCodeModel() != CodeModel::Large)
1800 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1801 } else if (IsTailCall) {
1802 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1804 Chain = DAG.getCopyToReg(Chain,
1805 DAG.getRegister(Opc, getPointerTy()),
1807 Callee = DAG.getRegister(Opc, getPointerTy());
1808 // Add register as live out.
1809 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1812 // Returns a chain & a flag for retval copy to use.
1813 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1814 SmallVector<SDOperand, 8> Ops;
1817 Ops.push_back(Chain);
1818 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1819 Ops.push_back(DAG.getIntPtrConstant(0));
1821 Ops.push_back(InFlag);
1822 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1823 InFlag = Chain.getValue(1);
1825 // Returns a chain & a flag for retval copy to use.
1826 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1830 Ops.push_back(Chain);
1831 Ops.push_back(Callee);
1834 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1836 // Add argument registers to the end of the list so that they are known live
1838 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1839 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1840 RegsToPass[i].second.getValueType()));
1842 // Add an implicit use GOT pointer in EBX.
1843 if (!IsTailCall && !Is64Bit &&
1844 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1845 Subtarget->isPICStyleGOT())
1846 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1848 // Add an implicit use of AL for x86 vararg functions.
1849 if (Is64Bit && isVarArg)
1850 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1853 Ops.push_back(InFlag);
1856 assert(InFlag.Val &&
1857 "Flag must be set. Depend on flag being set in LowerRET");
1858 Chain = DAG.getNode(X86ISD::TAILCALL,
1859 Op.Val->getVTList(), &Ops[0], Ops.size());
1861 return SDOperand(Chain.Val, Op.ResNo);
1864 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1865 InFlag = Chain.getValue(1);
1867 // Create the CALLSEQ_END node.
1868 unsigned NumBytesForCalleeToPush;
1869 if (IsCalleePop(Op))
1870 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1871 else if (!Is64Bit && IsStructRet)
1872 // If this is is a call to a struct-return function, the callee
1873 // pops the hidden struct pointer, so we have to push it back.
1874 // This is common for Darwin/X86, Linux & Mingw32 targets.
1875 NumBytesForCalleeToPush = 4;
1877 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1879 // Returns a flag for retval copy to use.
1880 Chain = DAG.getCALLSEQ_END(Chain,
1881 DAG.getIntPtrConstant(NumBytes),
1882 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1884 InFlag = Chain.getValue(1);
1886 // Handle result values, copying them out of physregs into vregs that we
1888 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1892 //===----------------------------------------------------------------------===//
1893 // Fast Calling Convention (tail call) implementation
1894 //===----------------------------------------------------------------------===//
1896 // Like std call, callee cleans arguments, convention except that ECX is
1897 // reserved for storing the tail called function address. Only 2 registers are
1898 // free for argument passing (inreg). Tail call optimization is performed
1900 // * tailcallopt is enabled
1901 // * caller/callee are fastcc
1902 // On X86_64 architecture with GOT-style position independent code only local
1903 // (within module) calls are supported at the moment.
1904 // To keep the stack aligned according to platform abi the function
1905 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1906 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1907 // If a tail called function callee has more arguments than the caller the
1908 // caller needs to make sure that there is room to move the RETADDR to. This is
1909 // achieved by reserving an area the size of the argument delta right after the
1910 // original REtADDR, but before the saved framepointer or the spilled registers
1911 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1923 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1924 /// for a 16 byte align requirement.
1925 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1926 SelectionDAG& DAG) {
1927 if (PerformTailCallOpt) {
1928 MachineFunction &MF = DAG.getMachineFunction();
1929 const TargetMachine &TM = MF.getTarget();
1930 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1931 unsigned StackAlignment = TFI.getStackAlignment();
1932 uint64_t AlignMask = StackAlignment - 1;
1933 int64_t Offset = StackSize;
1934 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1935 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1936 // Number smaller than 12 so just add the difference.
1937 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1939 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1940 Offset = ((~AlignMask) & Offset) + StackAlignment +
1941 (StackAlignment-SlotSize);
1948 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1949 /// following the call is a return. A function is eligible if caller/callee
1950 /// calling conventions match, currently only fastcc supports tail calls, and
1951 /// the function CALL is immediatly followed by a RET.
1952 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1954 SelectionDAG& DAG) const {
1955 if (!PerformTailCallOpt)
1958 // Check whether CALL node immediatly preceeds the RET node and whether the
1959 // return uses the result of the node or is a void return.
1960 unsigned NumOps = Ret.getNumOperands();
1962 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1963 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1965 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1966 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1967 MachineFunction &MF = DAG.getMachineFunction();
1968 unsigned CallerCC = MF.getFunction()->getCallingConv();
1969 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1970 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1971 SDOperand Callee = Call.getOperand(4);
1972 // On x86/32Bit PIC/GOT tail calls are supported.
1973 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1974 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1977 // Can only do local tail calls (in same module, hidden or protected) on
1978 // x86_64 PIC/GOT at the moment.
1979 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1980 return G->getGlobal()->hasHiddenVisibility()
1981 || G->getGlobal()->hasProtectedVisibility();
1988 //===----------------------------------------------------------------------===//
1989 // Other Lowering Hooks
1990 //===----------------------------------------------------------------------===//
1993 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1994 MachineFunction &MF = DAG.getMachineFunction();
1995 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1996 int ReturnAddrIndex = FuncInfo->getRAIndex();
1998 if (ReturnAddrIndex == 0) {
1999 // Set up a frame object for the return address.
2000 if (Subtarget->is64Bit())
2001 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2003 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2005 FuncInfo->setRAIndex(ReturnAddrIndex);
2008 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2013 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2014 /// specific condition code. It returns a false if it cannot do a direct
2015 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2017 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2018 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2019 SelectionDAG &DAG) {
2020 X86CC = X86::COND_INVALID;
2022 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2023 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2024 // X > -1 -> X == 0, jump !sign.
2025 RHS = DAG.getConstant(0, RHS.getValueType());
2026 X86CC = X86::COND_NS;
2028 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2029 // X < 0 -> X == 0, jump on sign.
2030 X86CC = X86::COND_S;
2032 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2034 RHS = DAG.getConstant(0, RHS.getValueType());
2035 X86CC = X86::COND_LE;
2040 switch (SetCCOpcode) {
2042 case ISD::SETEQ: X86CC = X86::COND_E; break;
2043 case ISD::SETGT: X86CC = X86::COND_G; break;
2044 case ISD::SETGE: X86CC = X86::COND_GE; break;
2045 case ISD::SETLT: X86CC = X86::COND_L; break;
2046 case ISD::SETLE: X86CC = X86::COND_LE; break;
2047 case ISD::SETNE: X86CC = X86::COND_NE; break;
2048 case ISD::SETULT: X86CC = X86::COND_B; break;
2049 case ISD::SETUGT: X86CC = X86::COND_A; break;
2050 case ISD::SETULE: X86CC = X86::COND_BE; break;
2051 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2054 // On a floating point condition, the flags are set as follows:
2056 // 0 | 0 | 0 | X > Y
2057 // 0 | 0 | 1 | X < Y
2058 // 1 | 0 | 0 | X == Y
2059 // 1 | 1 | 1 | unordered
2061 switch (SetCCOpcode) {
2064 case ISD::SETEQ: X86CC = X86::COND_E; break;
2065 case ISD::SETOLT: Flip = true; // Fallthrough
2067 case ISD::SETGT: X86CC = X86::COND_A; break;
2068 case ISD::SETOLE: Flip = true; // Fallthrough
2070 case ISD::SETGE: X86CC = X86::COND_AE; break;
2071 case ISD::SETUGT: Flip = true; // Fallthrough
2073 case ISD::SETLT: X86CC = X86::COND_B; break;
2074 case ISD::SETUGE: Flip = true; // Fallthrough
2076 case ISD::SETLE: X86CC = X86::COND_BE; break;
2078 case ISD::SETNE: X86CC = X86::COND_NE; break;
2079 case ISD::SETUO: X86CC = X86::COND_P; break;
2080 case ISD::SETO: X86CC = X86::COND_NP; break;
2083 std::swap(LHS, RHS);
2086 return X86CC != X86::COND_INVALID;
2089 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2090 /// code. Current x86 isa includes the following FP cmov instructions:
2091 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2092 static bool hasFPCMov(unsigned X86CC) {
2108 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2109 /// true if Op is undef or if its value falls within the specified range (L, H].
2110 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2111 if (Op.getOpcode() == ISD::UNDEF)
2114 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2115 return (Val >= Low && Val < Hi);
2118 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2119 /// true if Op is undef or if its value equal to the specified value.
2120 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2121 if (Op.getOpcode() == ISD::UNDEF)
2123 return cast<ConstantSDNode>(Op)->getValue() == Val;
2126 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2127 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2128 bool X86::isPSHUFDMask(SDNode *N) {
2129 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2131 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2134 // Check if the value doesn't reference the second vector.
2135 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2136 SDOperand Arg = N->getOperand(i);
2137 if (Arg.getOpcode() == ISD::UNDEF) continue;
2138 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2139 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2146 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2147 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2148 bool X86::isPSHUFHWMask(SDNode *N) {
2149 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2151 if (N->getNumOperands() != 8)
2154 // Lower quadword copied in order.
2155 for (unsigned i = 0; i != 4; ++i) {
2156 SDOperand Arg = N->getOperand(i);
2157 if (Arg.getOpcode() == ISD::UNDEF) continue;
2158 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2159 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2163 // Upper quadword shuffled.
2164 for (unsigned i = 4; i != 8; ++i) {
2165 SDOperand Arg = N->getOperand(i);
2166 if (Arg.getOpcode() == ISD::UNDEF) continue;
2167 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2168 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2169 if (Val < 4 || Val > 7)
2176 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2177 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2178 bool X86::isPSHUFLWMask(SDNode *N) {
2179 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2181 if (N->getNumOperands() != 8)
2184 // Upper quadword copied in order.
2185 for (unsigned i = 4; i != 8; ++i)
2186 if (!isUndefOrEqual(N->getOperand(i), i))
2189 // Lower quadword shuffled.
2190 for (unsigned i = 0; i != 4; ++i)
2191 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2197 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2198 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2199 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2200 if (NumElems != 2 && NumElems != 4) return false;
2202 unsigned Half = NumElems / 2;
2203 for (unsigned i = 0; i < Half; ++i)
2204 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2206 for (unsigned i = Half; i < NumElems; ++i)
2207 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2213 bool X86::isSHUFPMask(SDNode *N) {
2214 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2215 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2218 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2219 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2220 /// half elements to come from vector 1 (which would equal the dest.) and
2221 /// the upper half to come from vector 2.
2222 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2223 if (NumOps != 2 && NumOps != 4) return false;
2225 unsigned Half = NumOps / 2;
2226 for (unsigned i = 0; i < Half; ++i)
2227 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2229 for (unsigned i = Half; i < NumOps; ++i)
2230 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2235 static bool isCommutedSHUFP(SDNode *N) {
2236 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2237 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2240 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2241 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2242 bool X86::isMOVHLPSMask(SDNode *N) {
2243 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2245 if (N->getNumOperands() != 4)
2248 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2249 return isUndefOrEqual(N->getOperand(0), 6) &&
2250 isUndefOrEqual(N->getOperand(1), 7) &&
2251 isUndefOrEqual(N->getOperand(2), 2) &&
2252 isUndefOrEqual(N->getOperand(3), 3);
2255 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2256 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2258 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2259 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2261 if (N->getNumOperands() != 4)
2264 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2265 return isUndefOrEqual(N->getOperand(0), 2) &&
2266 isUndefOrEqual(N->getOperand(1), 3) &&
2267 isUndefOrEqual(N->getOperand(2), 2) &&
2268 isUndefOrEqual(N->getOperand(3), 3);
2271 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2272 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2273 bool X86::isMOVLPMask(SDNode *N) {
2274 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2276 unsigned NumElems = N->getNumOperands();
2277 if (NumElems != 2 && NumElems != 4)
2280 for (unsigned i = 0; i < NumElems/2; ++i)
2281 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2284 for (unsigned i = NumElems/2; i < NumElems; ++i)
2285 if (!isUndefOrEqual(N->getOperand(i), i))
2291 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2292 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2294 bool X86::isMOVHPMask(SDNode *N) {
2295 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2297 unsigned NumElems = N->getNumOperands();
2298 if (NumElems != 2 && NumElems != 4)
2301 for (unsigned i = 0; i < NumElems/2; ++i)
2302 if (!isUndefOrEqual(N->getOperand(i), i))
2305 for (unsigned i = 0; i < NumElems/2; ++i) {
2306 SDOperand Arg = N->getOperand(i + NumElems/2);
2307 if (!isUndefOrEqual(Arg, i + NumElems))
2314 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2315 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2316 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2317 bool V2IsSplat = false) {
2318 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2321 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2322 SDOperand BitI = Elts[i];
2323 SDOperand BitI1 = Elts[i+1];
2324 if (!isUndefOrEqual(BitI, j))
2327 if (isUndefOrEqual(BitI1, NumElts))
2330 if (!isUndefOrEqual(BitI1, j + NumElts))
2338 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2339 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2340 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2343 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2344 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2345 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2346 bool V2IsSplat = false) {
2347 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2350 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2351 SDOperand BitI = Elts[i];
2352 SDOperand BitI1 = Elts[i+1];
2353 if (!isUndefOrEqual(BitI, j + NumElts/2))
2356 if (isUndefOrEqual(BitI1, NumElts))
2359 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2367 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2368 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2369 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2372 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2373 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2375 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2376 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2378 unsigned NumElems = N->getNumOperands();
2379 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2382 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2383 SDOperand BitI = N->getOperand(i);
2384 SDOperand BitI1 = N->getOperand(i+1);
2386 if (!isUndefOrEqual(BitI, j))
2388 if (!isUndefOrEqual(BitI1, j))
2395 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2396 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2398 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2399 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2401 unsigned NumElems = N->getNumOperands();
2402 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2405 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2406 SDOperand BitI = N->getOperand(i);
2407 SDOperand BitI1 = N->getOperand(i + 1);
2409 if (!isUndefOrEqual(BitI, j))
2411 if (!isUndefOrEqual(BitI1, j))
2418 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2419 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2420 /// MOVSD, and MOVD, i.e. setting the lowest element.
2421 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2422 if (NumElts != 2 && NumElts != 4)
2425 if (!isUndefOrEqual(Elts[0], NumElts))
2428 for (unsigned i = 1; i < NumElts; ++i) {
2429 if (!isUndefOrEqual(Elts[i], i))
2436 bool X86::isMOVLMask(SDNode *N) {
2437 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2438 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2441 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2442 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2443 /// element of vector 2 and the other elements to come from vector 1 in order.
2444 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2445 bool V2IsSplat = false,
2446 bool V2IsUndef = false) {
2447 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2450 if (!isUndefOrEqual(Ops[0], 0))
2453 for (unsigned i = 1; i < NumOps; ++i) {
2454 SDOperand Arg = Ops[i];
2455 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2456 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2457 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2464 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2465 bool V2IsUndef = false) {
2466 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2467 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2468 V2IsSplat, V2IsUndef);
2471 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2472 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2473 bool X86::isMOVSHDUPMask(SDNode *N) {
2474 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2476 if (N->getNumOperands() != 4)
2479 // Expect 1, 1, 3, 3
2480 for (unsigned i = 0; i < 2; ++i) {
2481 SDOperand Arg = N->getOperand(i);
2482 if (Arg.getOpcode() == ISD::UNDEF) continue;
2483 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2484 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2485 if (Val != 1) return false;
2489 for (unsigned i = 2; i < 4; ++i) {
2490 SDOperand Arg = N->getOperand(i);
2491 if (Arg.getOpcode() == ISD::UNDEF) continue;
2492 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2493 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2494 if (Val != 3) return false;
2498 // Don't use movshdup if it can be done with a shufps.
2502 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2503 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2504 bool X86::isMOVSLDUPMask(SDNode *N) {
2505 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2507 if (N->getNumOperands() != 4)
2510 // Expect 0, 0, 2, 2
2511 for (unsigned i = 0; i < 2; ++i) {
2512 SDOperand Arg = N->getOperand(i);
2513 if (Arg.getOpcode() == ISD::UNDEF) continue;
2514 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2515 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2516 if (Val != 0) return false;
2520 for (unsigned i = 2; i < 4; ++i) {
2521 SDOperand Arg = N->getOperand(i);
2522 if (Arg.getOpcode() == ISD::UNDEF) continue;
2523 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2524 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2525 if (Val != 2) return false;
2529 // Don't use movshdup if it can be done with a shufps.
2533 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2534 /// specifies a identity operation on the LHS or RHS.
2535 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2536 unsigned NumElems = N->getNumOperands();
2537 for (unsigned i = 0; i < NumElems; ++i)
2538 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2543 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2544 /// a splat of a single element.
2545 static bool isSplatMask(SDNode *N) {
2546 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2548 // This is a splat operation if each element of the permute is the same, and
2549 // if the value doesn't reference the second vector.
2550 unsigned NumElems = N->getNumOperands();
2551 SDOperand ElementBase;
2553 for (; i != NumElems; ++i) {
2554 SDOperand Elt = N->getOperand(i);
2555 if (isa<ConstantSDNode>(Elt)) {
2561 if (!ElementBase.Val)
2564 for (; i != NumElems; ++i) {
2565 SDOperand Arg = N->getOperand(i);
2566 if (Arg.getOpcode() == ISD::UNDEF) continue;
2567 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2568 if (Arg != ElementBase) return false;
2571 // Make sure it is a splat of the first vector operand.
2572 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2575 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2576 /// a splat of a single element and it's a 2 or 4 element mask.
2577 bool X86::isSplatMask(SDNode *N) {
2578 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2580 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2581 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2583 return ::isSplatMask(N);
2586 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2587 /// specifies a splat of zero element.
2588 bool X86::isSplatLoMask(SDNode *N) {
2589 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2591 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2592 if (!isUndefOrEqual(N->getOperand(i), 0))
2597 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2598 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2600 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2601 unsigned NumOperands = N->getNumOperands();
2602 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2604 for (unsigned i = 0; i < NumOperands; ++i) {
2606 SDOperand Arg = N->getOperand(NumOperands-i-1);
2607 if (Arg.getOpcode() != ISD::UNDEF)
2608 Val = cast<ConstantSDNode>(Arg)->getValue();
2609 if (Val >= NumOperands) Val -= NumOperands;
2611 if (i != NumOperands - 1)
2618 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2619 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2621 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2623 // 8 nodes, but we only care about the last 4.
2624 for (unsigned i = 7; i >= 4; --i) {
2626 SDOperand Arg = N->getOperand(i);
2627 if (Arg.getOpcode() != ISD::UNDEF)
2628 Val = cast<ConstantSDNode>(Arg)->getValue();
2637 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2638 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2640 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2642 // 8 nodes, but we only care about the first 4.
2643 for (int i = 3; i >= 0; --i) {
2645 SDOperand Arg = N->getOperand(i);
2646 if (Arg.getOpcode() != ISD::UNDEF)
2647 Val = cast<ConstantSDNode>(Arg)->getValue();
2656 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2657 /// specifies a 8 element shuffle that can be broken into a pair of
2658 /// PSHUFHW and PSHUFLW.
2659 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2660 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2662 if (N->getNumOperands() != 8)
2665 // Lower quadword shuffled.
2666 for (unsigned i = 0; i != 4; ++i) {
2667 SDOperand Arg = N->getOperand(i);
2668 if (Arg.getOpcode() == ISD::UNDEF) continue;
2669 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2670 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2675 // Upper quadword shuffled.
2676 for (unsigned i = 4; i != 8; ++i) {
2677 SDOperand Arg = N->getOperand(i);
2678 if (Arg.getOpcode() == ISD::UNDEF) continue;
2679 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2680 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2681 if (Val < 4 || Val > 7)
2688 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2689 /// values in ther permute mask.
2690 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2691 SDOperand &V2, SDOperand &Mask,
2692 SelectionDAG &DAG) {
2693 MVT::ValueType VT = Op.getValueType();
2694 MVT::ValueType MaskVT = Mask.getValueType();
2695 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2696 unsigned NumElems = Mask.getNumOperands();
2697 SmallVector<SDOperand, 8> MaskVec;
2699 for (unsigned i = 0; i != NumElems; ++i) {
2700 SDOperand Arg = Mask.getOperand(i);
2701 if (Arg.getOpcode() == ISD::UNDEF) {
2702 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2705 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2706 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2708 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2710 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2714 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2715 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2718 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2719 /// the two vector operands have swapped position.
2721 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2722 MVT::ValueType MaskVT = Mask.getValueType();
2723 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2724 unsigned NumElems = Mask.getNumOperands();
2725 SmallVector<SDOperand, 8> MaskVec;
2726 for (unsigned i = 0; i != NumElems; ++i) {
2727 SDOperand Arg = Mask.getOperand(i);
2728 if (Arg.getOpcode() == ISD::UNDEF) {
2729 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2732 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2733 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2735 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2737 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2739 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2743 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2744 /// match movhlps. The lower half elements should come from upper half of
2745 /// V1 (and in order), and the upper half elements should come from the upper
2746 /// half of V2 (and in order).
2747 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2748 unsigned NumElems = Mask->getNumOperands();
2751 for (unsigned i = 0, e = 2; i != e; ++i)
2752 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2754 for (unsigned i = 2; i != 4; ++i)
2755 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2760 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2761 /// is promoted to a vector.
2762 static inline bool isScalarLoadToVector(SDNode *N) {
2763 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2764 N = N->getOperand(0).Val;
2765 return ISD::isNON_EXTLoad(N);
2770 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2771 /// match movlp{s|d}. The lower half elements should come from lower half of
2772 /// V1 (and in order), and the upper half elements should come from the upper
2773 /// half of V2 (and in order). And since V1 will become the source of the
2774 /// MOVLP, it must be either a vector load or a scalar load to vector.
2775 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2776 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2778 // Is V2 is a vector load, don't do this transformation. We will try to use
2779 // load folding shufps op.
2780 if (ISD::isNON_EXTLoad(V2))
2783 unsigned NumElems = Mask->getNumOperands();
2784 if (NumElems != 2 && NumElems != 4)
2786 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2787 if (!isUndefOrEqual(Mask->getOperand(i), i))
2789 for (unsigned i = NumElems/2; i != NumElems; ++i)
2790 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2795 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2797 static bool isSplatVector(SDNode *N) {
2798 if (N->getOpcode() != ISD::BUILD_VECTOR)
2801 SDOperand SplatValue = N->getOperand(0);
2802 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2803 if (N->getOperand(i) != SplatValue)
2808 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2810 static bool isUndefShuffle(SDNode *N) {
2811 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2814 SDOperand V1 = N->getOperand(0);
2815 SDOperand V2 = N->getOperand(1);
2816 SDOperand Mask = N->getOperand(2);
2817 unsigned NumElems = Mask.getNumOperands();
2818 for (unsigned i = 0; i != NumElems; ++i) {
2819 SDOperand Arg = Mask.getOperand(i);
2820 if (Arg.getOpcode() != ISD::UNDEF) {
2821 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2822 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2824 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2831 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2833 static inline bool isZeroNode(SDOperand Elt) {
2834 return ((isa<ConstantSDNode>(Elt) &&
2835 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2836 (isa<ConstantFPSDNode>(Elt) &&
2837 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2840 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2841 /// to an zero vector.
2842 static bool isZeroShuffle(SDNode *N) {
2843 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2846 SDOperand V1 = N->getOperand(0);
2847 SDOperand V2 = N->getOperand(1);
2848 SDOperand Mask = N->getOperand(2);
2849 unsigned NumElems = Mask.getNumOperands();
2850 for (unsigned i = 0; i != NumElems; ++i) {
2851 SDOperand Arg = Mask.getOperand(i);
2852 if (Arg.getOpcode() == ISD::UNDEF)
2855 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2856 if (Idx < NumElems) {
2857 unsigned Opc = V1.Val->getOpcode();
2858 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2860 if (Opc != ISD::BUILD_VECTOR ||
2861 !isZeroNode(V1.Val->getOperand(Idx)))
2863 } else if (Idx >= NumElems) {
2864 unsigned Opc = V2.Val->getOpcode();
2865 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2867 if (Opc != ISD::BUILD_VECTOR ||
2868 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2875 /// getZeroVector - Returns a vector of specified type with all zero elements.
2877 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2878 assert(MVT::isVector(VT) && "Expected a vector type");
2880 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2881 // type. This ensures they get CSE'd.
2882 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2884 if (MVT::getSizeInBits(VT) == 64) // MMX
2885 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2887 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2888 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2891 /// getOnesVector - Returns a vector of specified type with all bits set.
2893 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2894 assert(MVT::isVector(VT) && "Expected a vector type");
2896 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2897 // type. This ensures they get CSE'd.
2898 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2900 if (MVT::getSizeInBits(VT) == 64) // MMX
2901 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2903 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2904 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2908 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2909 /// that point to V2 points to its first element.
2910 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2911 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2913 bool Changed = false;
2914 SmallVector<SDOperand, 8> MaskVec;
2915 unsigned NumElems = Mask.getNumOperands();
2916 for (unsigned i = 0; i != NumElems; ++i) {
2917 SDOperand Arg = Mask.getOperand(i);
2918 if (Arg.getOpcode() != ISD::UNDEF) {
2919 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2920 if (Val > NumElems) {
2921 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2925 MaskVec.push_back(Arg);
2929 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2930 &MaskVec[0], MaskVec.size());
2934 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2935 /// operation of specified width.
2936 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2937 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2938 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2940 SmallVector<SDOperand, 8> MaskVec;
2941 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2942 for (unsigned i = 1; i != NumElems; ++i)
2943 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2944 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2947 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2948 /// of specified width.
2949 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2950 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2951 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2952 SmallVector<SDOperand, 8> MaskVec;
2953 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2954 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2955 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2957 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2960 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2961 /// of specified width.
2962 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2963 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2964 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2965 unsigned Half = NumElems/2;
2966 SmallVector<SDOperand, 8> MaskVec;
2967 for (unsigned i = 0; i != Half; ++i) {
2968 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2969 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2971 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2974 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2975 /// element #0 of a vector with the specified index, leaving the rest of the
2976 /// elements in place.
2977 static SDOperand getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2978 SelectionDAG &DAG) {
2979 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2980 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2981 SmallVector<SDOperand, 8> MaskVec;
2982 // Element #0 of the result gets the elt we are replacing.
2983 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2984 for (unsigned i = 1; i != NumElems; ++i)
2985 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2986 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2989 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2990 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG, bool HasSSE2) {
2991 MVT::ValueType PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2992 MVT::ValueType VT = Op.getValueType();
2995 SDOperand V1 = Op.getOperand(0);
2996 SDOperand Mask = Op.getOperand(2);
2997 unsigned NumElems = Mask.getNumOperands();
2998 // Special handling of v4f32 -> v4i32.
2999 if (VT != MVT::v4f32) {
3000 Mask = getUnpacklMask(NumElems, DAG);
3001 while (NumElems > 4) {
3002 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
3005 Mask = getZeroVector(MVT::v4i32, DAG);
3008 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3009 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3010 DAG.getNode(ISD::UNDEF, PVT), Mask);
3011 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3014 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3015 /// vector of zero or undef vector. This produces a shuffle where the low
3016 /// element of V2 is swizzled into the zero/undef vector, landing at element
3017 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3018 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, unsigned Idx,
3019 bool isZero, SelectionDAG &DAG) {
3020 MVT::ValueType VT = V2.getValueType();
3021 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
3022 unsigned NumElems = MVT::getVectorNumElements(V2.getValueType());
3023 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3024 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3025 SmallVector<SDOperand, 16> MaskVec;
3026 for (unsigned i = 0; i != NumElems; ++i)
3027 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3028 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3030 MaskVec.push_back(DAG.getConstant(i, EVT));
3031 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3032 &MaskVec[0], MaskVec.size());
3033 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3036 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3038 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
3039 unsigned NumNonZero, unsigned NumZero,
3040 SelectionDAG &DAG, TargetLowering &TLI) {
3046 for (unsigned i = 0; i < 16; ++i) {
3047 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3048 if (ThisIsNonZero && First) {
3050 V = getZeroVector(MVT::v8i16, DAG);
3052 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3057 SDOperand ThisElt(0, 0), LastElt(0, 0);
3058 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3059 if (LastIsNonZero) {
3060 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3062 if (ThisIsNonZero) {
3063 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3064 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3065 ThisElt, DAG.getConstant(8, MVT::i8));
3067 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3072 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3073 DAG.getIntPtrConstant(i/2));
3077 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3080 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3082 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3083 unsigned NumNonZero, unsigned NumZero,
3084 SelectionDAG &DAG, TargetLowering &TLI) {
3090 for (unsigned i = 0; i < 8; ++i) {
3091 bool isNonZero = (NonZeros & (1 << i)) != 0;
3095 V = getZeroVector(MVT::v8i16, DAG);
3097 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3100 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3101 DAG.getIntPtrConstant(i));
3109 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3110 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3111 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
3112 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3113 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3114 // eliminated on x86-32 hosts.
3115 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3118 if (ISD::isBuildVectorAllOnes(Op.Val))
3119 return getOnesVector(Op.getValueType(), DAG);
3120 return getZeroVector(Op.getValueType(), DAG);
3123 MVT::ValueType VT = Op.getValueType();
3124 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3125 unsigned EVTBits = MVT::getSizeInBits(EVT);
3127 unsigned NumElems = Op.getNumOperands();
3128 unsigned NumZero = 0;
3129 unsigned NumNonZero = 0;
3130 unsigned NonZeros = 0;
3131 bool IsAllConstants = true;
3132 SmallSet<SDOperand, 8> Values;
3133 for (unsigned i = 0; i < NumElems; ++i) {
3134 SDOperand Elt = Op.getOperand(i);
3135 if (Elt.getOpcode() == ISD::UNDEF)
3138 if (Elt.getOpcode() != ISD::Constant &&
3139 Elt.getOpcode() != ISD::ConstantFP)
3140 IsAllConstants = false;
3141 if (isZeroNode(Elt))
3144 NonZeros |= (1 << i);
3149 if (NumNonZero == 0) {
3150 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3151 return DAG.getNode(ISD::UNDEF, VT);
3154 // Special case for single non-zero, non-undef, element.
3155 if (NumNonZero == 1 && NumElems <= 4) {
3156 unsigned Idx = CountTrailingZeros_32(NonZeros);
3157 SDOperand Item = Op.getOperand(Idx);
3159 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3160 // the value are obviously zero, truncate the value to i32 and do the
3161 // insertion that way. Only do this if the value is non-constant or if the
3162 // value is a constant being inserted into element 0. It is cheaper to do
3163 // a constant pool load than it is to do a movd + shuffle.
3164 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3165 (!IsAllConstants || Idx == 0)) {
3166 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3167 // Handle MMX and SSE both.
3168 MVT::ValueType VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3169 MVT::ValueType VecElts = VT == MVT::v2i64 ? 4 : 2;
3171 // Truncate the value (which may itself be a constant) to i32, and
3172 // convert it to a vector with movd (S2V+shuffle to zero extend).
3173 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3174 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3175 Item = getShuffleVectorZeroOrUndef(Item, 0, true, DAG);
3177 // Now we have our 32-bit value zero extended in the low element of
3178 // a vector. If Idx != 0, swizzle it into place.
3181 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3182 getSwapEltZeroMask(VecElts, Idx, DAG)
3184 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3186 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3190 // If we have a constant or non-constant insertion into the low element of
3191 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3192 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3193 // depending on what the source datatype is. Because we can only get here
3194 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3196 // Don't do this for i64 values on x86-32.
3197 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3198 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3199 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3200 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, DAG);
3203 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3206 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3207 // is a non-constant being inserted into an element other than the low one,
3208 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3209 // movd/movss) to move this into the low element, then shuffle it into
3211 if (EVTBits == 32) {
3212 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3214 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3215 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, DAG);
3216 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3217 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3218 SmallVector<SDOperand, 8> MaskVec;
3219 for (unsigned i = 0; i < NumElems; i++)
3220 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3221 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3222 &MaskVec[0], MaskVec.size());
3223 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3224 DAG.getNode(ISD::UNDEF, VT), Mask);
3228 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3229 if (Values.size() == 1)
3232 // A vector full of immediates; various special cases are already
3233 // handled, so this is best done with a single constant-pool load.
3237 // Let legalizer expand 2-wide build_vectors.
3241 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3242 if (EVTBits == 8 && NumElems == 16) {
3243 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3245 if (V.Val) return V;
3248 if (EVTBits == 16 && NumElems == 8) {
3249 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3251 if (V.Val) return V;
3254 // If element VT is == 32 bits, turn it into a number of shuffles.
3255 SmallVector<SDOperand, 8> V;
3257 if (NumElems == 4 && NumZero > 0) {
3258 for (unsigned i = 0; i < 4; ++i) {
3259 bool isZero = !(NonZeros & (1 << i));
3261 V[i] = getZeroVector(VT, DAG);
3263 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3266 for (unsigned i = 0; i < 2; ++i) {
3267 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3270 V[i] = V[i*2]; // Must be a zero vector.
3273 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3274 getMOVLMask(NumElems, DAG));
3277 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3278 getMOVLMask(NumElems, DAG));
3281 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3282 getUnpacklMask(NumElems, DAG));
3287 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3288 // clears the upper bits.
3289 // FIXME: we can do the same for v4f32 case when we know both parts of
3290 // the lower half come from scalar_to_vector (loadf32). We should do
3291 // that in post legalizer dag combiner with target specific hooks.
3292 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3294 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3295 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3296 SmallVector<SDOperand, 8> MaskVec;
3297 bool Reverse = (NonZeros & 0x3) == 2;
3298 for (unsigned i = 0; i < 2; ++i)
3300 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3302 MaskVec.push_back(DAG.getConstant(i, EVT));
3303 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3304 for (unsigned i = 0; i < 2; ++i)
3306 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3308 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3309 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3310 &MaskVec[0], MaskVec.size());
3311 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3314 if (Values.size() > 2) {
3315 // Expand into a number of unpckl*.
3317 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3318 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3319 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3320 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3321 for (unsigned i = 0; i < NumElems; ++i)
3322 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3324 while (NumElems != 0) {
3325 for (unsigned i = 0; i < NumElems; ++i)
3326 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3337 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3338 SDOperand PermMask, SelectionDAG &DAG,
3339 TargetLowering &TLI) {
3341 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3342 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3343 MVT::ValueType PtrVT = TLI.getPointerTy();
3344 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3345 PermMask.Val->op_end());
3347 // First record which half of which vector the low elements come from.
3348 SmallVector<unsigned, 4> LowQuad(4);
3349 for (unsigned i = 0; i < 4; ++i) {
3350 SDOperand Elt = MaskElts[i];
3351 if (Elt.getOpcode() == ISD::UNDEF)
3353 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3354 int QuadIdx = EltIdx / 4;
3357 int BestLowQuad = -1;
3358 unsigned MaxQuad = 1;
3359 for (unsigned i = 0; i < 4; ++i) {
3360 if (LowQuad[i] > MaxQuad) {
3362 MaxQuad = LowQuad[i];
3366 // Record which half of which vector the high elements come from.
3367 SmallVector<unsigned, 4> HighQuad(4);
3368 for (unsigned i = 4; i < 8; ++i) {
3369 SDOperand Elt = MaskElts[i];
3370 if (Elt.getOpcode() == ISD::UNDEF)
3372 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3373 int QuadIdx = EltIdx / 4;
3374 ++HighQuad[QuadIdx];
3376 int BestHighQuad = -1;
3378 for (unsigned i = 0; i < 4; ++i) {
3379 if (HighQuad[i] > MaxQuad) {
3381 MaxQuad = HighQuad[i];
3385 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3386 if (BestLowQuad != -1 || BestHighQuad != -1) {
3387 // First sort the 4 chunks in order using shufpd.
3388 SmallVector<SDOperand, 8> MaskVec;
3389 if (BestLowQuad != -1)
3390 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3392 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3393 if (BestHighQuad != -1)
3394 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3396 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3397 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3398 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3399 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3400 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3401 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3403 // Now sort high and low parts separately.
3404 BitVector InOrder(8);
3405 if (BestLowQuad != -1) {
3406 // Sort lower half in order using PSHUFLW.
3408 bool AnyOutOrder = false;
3409 for (unsigned i = 0; i != 4; ++i) {
3410 SDOperand Elt = MaskElts[i];
3411 if (Elt.getOpcode() == ISD::UNDEF) {
3412 MaskVec.push_back(Elt);
3415 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3418 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3419 // If this element is in the right place after this shuffle, then
3421 if ((int)(EltIdx / 4) == BestLowQuad)
3426 for (unsigned i = 4; i != 8; ++i)
3427 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3428 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3429 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3433 if (BestHighQuad != -1) {
3434 // Sort high half in order using PSHUFHW if possible.
3436 for (unsigned i = 0; i != 4; ++i)
3437 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3438 bool AnyOutOrder = false;
3439 for (unsigned i = 4; i != 8; ++i) {
3440 SDOperand Elt = MaskElts[i];
3441 if (Elt.getOpcode() == ISD::UNDEF) {
3442 MaskVec.push_back(Elt);
3445 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3448 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3449 // If this element is in the right place after this shuffle, then
3451 if ((int)(EltIdx / 4) == BestHighQuad)
3456 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3457 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3461 // The other elements are put in the right place using pextrw and pinsrw.
3462 for (unsigned i = 0; i != 8; ++i) {
3465 SDOperand Elt = MaskElts[i];
3466 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3469 SDOperand ExtOp = (EltIdx < 8)
3470 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3471 DAG.getConstant(EltIdx, PtrVT))
3472 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3473 DAG.getConstant(EltIdx - 8, PtrVT));
3474 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3475 DAG.getConstant(i, PtrVT));
3480 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3481 ///as few as possible.
3482 // First, let's find out how many elements are already in the right order.
3483 unsigned V1InOrder = 0;
3484 unsigned V1FromV1 = 0;
3485 unsigned V2InOrder = 0;
3486 unsigned V2FromV2 = 0;
3487 SmallVector<SDOperand, 8> V1Elts;
3488 SmallVector<SDOperand, 8> V2Elts;
3489 for (unsigned i = 0; i < 8; ++i) {
3490 SDOperand Elt = MaskElts[i];
3491 if (Elt.getOpcode() == ISD::UNDEF) {
3492 V1Elts.push_back(Elt);
3493 V2Elts.push_back(Elt);
3498 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3500 V1Elts.push_back(Elt);
3501 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3503 } else if (EltIdx == i+8) {
3504 V1Elts.push_back(Elt);
3505 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3507 } else if (EltIdx < 8) {
3508 V1Elts.push_back(Elt);
3511 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3516 if (V2InOrder > V1InOrder) {
3517 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3519 std::swap(V1Elts, V2Elts);
3520 std::swap(V1FromV1, V2FromV2);
3523 if ((V1FromV1 + V1InOrder) != 8) {
3524 // Some elements are from V2.
3526 // If there are elements that are from V1 but out of place,
3527 // then first sort them in place
3528 SmallVector<SDOperand, 8> MaskVec;
3529 for (unsigned i = 0; i < 8; ++i) {
3530 SDOperand Elt = V1Elts[i];
3531 if (Elt.getOpcode() == ISD::UNDEF) {
3532 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3535 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3537 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3539 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3541 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3542 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3546 for (unsigned i = 0; i < 8; ++i) {
3547 SDOperand Elt = V1Elts[i];
3548 if (Elt.getOpcode() == ISD::UNDEF)
3550 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3553 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3554 DAG.getConstant(EltIdx - 8, PtrVT));
3555 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3556 DAG.getConstant(i, PtrVT));
3560 // All elements are from V1.
3562 for (unsigned i = 0; i < 8; ++i) {
3563 SDOperand Elt = V1Elts[i];
3564 if (Elt.getOpcode() == ISD::UNDEF)
3566 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3567 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3568 DAG.getConstant(EltIdx, PtrVT));
3569 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3570 DAG.getConstant(i, PtrVT));
3576 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3577 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3578 /// done when every pair / quad of shuffle mask elements point to elements in
3579 /// the right sequence. e.g.
3580 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3582 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3584 SDOperand PermMask, SelectionDAG &DAG,
3585 TargetLowering &TLI) {
3586 unsigned NumElems = PermMask.getNumOperands();
3587 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3588 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3589 MVT::ValueType NewVT = MaskVT;
3591 case MVT::v4f32: NewVT = MVT::v2f64; break;
3592 case MVT::v4i32: NewVT = MVT::v2i64; break;
3593 case MVT::v8i16: NewVT = MVT::v4i32; break;
3594 case MVT::v16i8: NewVT = MVT::v4i32; break;
3595 default: assert(false && "Unexpected!");
3598 if (NewWidth == 2) {
3599 if (MVT::isInteger(VT))
3604 unsigned Scale = NumElems / NewWidth;
3605 SmallVector<SDOperand, 8> MaskVec;
3606 for (unsigned i = 0; i < NumElems; i += Scale) {
3607 unsigned StartIdx = ~0U;
3608 for (unsigned j = 0; j < Scale; ++j) {
3609 SDOperand Elt = PermMask.getOperand(i+j);
3610 if (Elt.getOpcode() == ISD::UNDEF)
3612 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3613 if (StartIdx == ~0U)
3614 StartIdx = EltIdx - (EltIdx % Scale);
3615 if (EltIdx != StartIdx + j)
3618 if (StartIdx == ~0U)
3619 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3621 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3624 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3625 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3626 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3627 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3628 &MaskVec[0], MaskVec.size()));
3632 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3633 SDOperand V1 = Op.getOperand(0);
3634 SDOperand V2 = Op.getOperand(1);
3635 SDOperand PermMask = Op.getOperand(2);
3636 MVT::ValueType VT = Op.getValueType();
3637 unsigned NumElems = PermMask.getNumOperands();
3638 bool isMMX = MVT::getSizeInBits(VT) == 64;
3639 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3640 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3641 bool V1IsSplat = false;
3642 bool V2IsSplat = false;
3644 if (isUndefShuffle(Op.Val))
3645 return DAG.getNode(ISD::UNDEF, VT);
3647 if (isZeroShuffle(Op.Val))
3648 return getZeroVector(VT, DAG);
3650 if (isIdentityMask(PermMask.Val))
3652 else if (isIdentityMask(PermMask.Val, true))
3655 if (isSplatMask(PermMask.Val)) {
3656 if (isMMX || NumElems < 4) return Op;
3657 // Promote it to a v4{if}32 splat.
3658 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3661 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3663 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3664 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3666 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3667 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3668 // FIXME: Figure out a cleaner way to do this.
3669 // Try to make use of movq to zero out the top part.
3670 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3671 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3673 SDOperand NewV1 = NewOp.getOperand(0);
3674 SDOperand NewV2 = NewOp.getOperand(1);
3675 SDOperand NewMask = NewOp.getOperand(2);
3676 if (isCommutedMOVL(NewMask.Val, true, false)) {
3677 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3678 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3679 NewV1, NewV2, getMOVLMask(2, DAG));
3680 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3683 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3684 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3685 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3686 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3690 if (X86::isMOVLMask(PermMask.Val))
3691 return (V1IsUndef) ? V2 : Op;
3693 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3694 X86::isMOVSLDUPMask(PermMask.Val) ||
3695 X86::isMOVHLPSMask(PermMask.Val) ||
3696 X86::isMOVHPMask(PermMask.Val) ||
3697 X86::isMOVLPMask(PermMask.Val))
3700 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3701 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3702 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3704 bool Commuted = false;
3705 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3706 // 1,1,1,1 -> v8i16 though.
3707 V1IsSplat = isSplatVector(V1.Val);
3708 V2IsSplat = isSplatVector(V2.Val);
3710 // Canonicalize the splat or undef, if present, to be on the RHS.
3711 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3712 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3713 std::swap(V1IsSplat, V2IsSplat);
3714 std::swap(V1IsUndef, V2IsUndef);
3718 // FIXME: Figure out a cleaner way to do this.
3719 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3720 if (V2IsUndef) return V1;
3721 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3723 // V2 is a splat, so the mask may be malformed. That is, it may point
3724 // to any V2 element. The instruction selectior won't like this. Get
3725 // a corrected mask and commute to form a proper MOVS{S|D}.
3726 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3727 if (NewMask.Val != PermMask.Val)
3728 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3733 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3734 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3735 X86::isUNPCKLMask(PermMask.Val) ||
3736 X86::isUNPCKHMask(PermMask.Val))
3740 // Normalize mask so all entries that point to V2 points to its first
3741 // element then try to match unpck{h|l} again. If match, return a
3742 // new vector_shuffle with the corrected mask.
3743 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3744 if (NewMask.Val != PermMask.Val) {
3745 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3746 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3747 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3748 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3749 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3750 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3755 // Normalize the node to match x86 shuffle ops if needed
3756 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3757 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3760 // Commute is back and try unpck* again.
3761 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3762 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3763 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3764 X86::isUNPCKLMask(PermMask.Val) ||
3765 X86::isUNPCKHMask(PermMask.Val))
3769 // Try PSHUF* first, then SHUFP*.
3770 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
3771 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3772 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.Val)) {
3773 if (V2.getOpcode() != ISD::UNDEF)
3774 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3775 DAG.getNode(ISD::UNDEF, VT), PermMask);
3780 if (Subtarget->hasSSE2() &&
3781 (X86::isPSHUFDMask(PermMask.Val) ||
3782 X86::isPSHUFHWMask(PermMask.Val) ||
3783 X86::isPSHUFLWMask(PermMask.Val))) {
3784 MVT::ValueType RVT = VT;
3785 if (VT == MVT::v4f32) {
3787 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
3788 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
3789 DAG.getNode(ISD::UNDEF, RVT), PermMask);
3790 } else if (V2.getOpcode() != ISD::UNDEF)
3791 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
3792 DAG.getNode(ISD::UNDEF, RVT), PermMask);
3794 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
3798 // Binary or unary shufps.
3799 if (X86::isSHUFPMask(PermMask.Val) ||
3800 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.Val)))
3804 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3805 if (VT == MVT::v8i16) {
3806 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3811 // Handle all 4 wide cases with a number of shuffles.
3812 if (NumElems == 4 && !isMMX) {
3813 // Don't do this for MMX.
3814 MVT::ValueType MaskVT = PermMask.getValueType();
3815 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3816 SmallVector<std::pair<int, int>, 8> Locs;
3817 Locs.reserve(NumElems);
3818 SmallVector<SDOperand, 8> Mask1(NumElems,
3819 DAG.getNode(ISD::UNDEF, MaskEVT));
3820 SmallVector<SDOperand, 8> Mask2(NumElems,
3821 DAG.getNode(ISD::UNDEF, MaskEVT));
3824 // If no more than two elements come from either vector. This can be
3825 // implemented with two shuffles. First shuffle gather the elements.
3826 // The second shuffle, which takes the first shuffle as both of its
3827 // vector operands, put the elements into the right order.
3828 for (unsigned i = 0; i != NumElems; ++i) {
3829 SDOperand Elt = PermMask.getOperand(i);
3830 if (Elt.getOpcode() == ISD::UNDEF) {
3831 Locs[i] = std::make_pair(-1, -1);
3833 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3834 if (Val < NumElems) {
3835 Locs[i] = std::make_pair(0, NumLo);
3839 Locs[i] = std::make_pair(1, NumHi);
3840 if (2+NumHi < NumElems)
3841 Mask1[2+NumHi] = Elt;
3846 if (NumLo <= 2 && NumHi <= 2) {
3847 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3848 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3849 &Mask1[0], Mask1.size()));
3850 for (unsigned i = 0; i != NumElems; ++i) {
3851 if (Locs[i].first == -1)
3854 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3855 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3856 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3860 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3861 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3862 &Mask2[0], Mask2.size()));
3865 // Break it into (shuffle shuffle_hi, shuffle_lo).
3867 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3868 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3869 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3870 unsigned MaskIdx = 0;
3872 unsigned HiIdx = NumElems/2;
3873 for (unsigned i = 0; i != NumElems; ++i) {
3874 if (i == NumElems/2) {
3880 SDOperand Elt = PermMask.getOperand(i);
3881 if (Elt.getOpcode() == ISD::UNDEF) {
3882 Locs[i] = std::make_pair(-1, -1);
3883 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3884 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3885 (*MaskPtr)[LoIdx] = Elt;
3888 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3889 (*MaskPtr)[HiIdx] = Elt;
3894 SDOperand LoShuffle =
3895 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3896 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3897 &LoMask[0], LoMask.size()));
3898 SDOperand HiShuffle =
3899 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3900 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3901 &HiMask[0], HiMask.size()));
3902 SmallVector<SDOperand, 8> MaskOps;
3903 for (unsigned i = 0; i != NumElems; ++i) {
3904 if (Locs[i].first == -1) {
3905 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3907 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3908 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3911 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3912 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3913 &MaskOps[0], MaskOps.size()));
3920 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDOperand Op,
3921 SelectionDAG &DAG) {
3922 MVT::ValueType VT = Op.getValueType();
3923 if (MVT::getSizeInBits(VT) == 8) {
3924 SDOperand Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
3925 Op.getOperand(0), Op.getOperand(1));
3926 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3927 DAG.getValueType(VT));
3928 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3929 } else if (MVT::getSizeInBits(VT) == 16) {
3930 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
3931 Op.getOperand(0), Op.getOperand(1));
3932 SDOperand Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
3933 DAG.getValueType(VT));
3934 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3935 } else if (VT == MVT::f32) {
3936 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
3937 // the result back to FR32 register. It's only worth matching if the
3938 // result has a single use which is a store or a bitcast to i32.
3939 if (!Op.hasOneUse())
3941 SDNode *User = Op.Val->use_begin()->getUser();
3942 if (User->getOpcode() != ISD::STORE &&
3943 (User->getOpcode() != ISD::BIT_CONVERT ||
3944 User->getValueType(0) != MVT::i32))
3946 SDOperand Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3947 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
3949 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
3956 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3957 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3960 if (Subtarget->hasSSE41()) {
3961 SDOperand Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
3966 MVT::ValueType VT = Op.getValueType();
3967 // TODO: handle v16i8.
3968 if (MVT::getSizeInBits(VT) == 16) {
3969 SDOperand Vec = Op.getOperand(0);
3970 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3972 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3973 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3974 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3976 // Transform it so it match pextrw which produces a 32-bit result.
3977 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3978 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3979 Op.getOperand(0), Op.getOperand(1));
3980 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3981 DAG.getValueType(VT));
3982 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3983 } else if (MVT::getSizeInBits(VT) == 32) {
3984 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3987 // SHUFPS the element to the lowest double word, then movss.
3988 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3989 SmallVector<SDOperand, 8> IdxVec;
3991 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3993 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3995 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3997 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3998 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3999 &IdxVec[0], IdxVec.size());
4000 SDOperand Vec = Op.getOperand(0);
4001 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4002 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4003 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4004 DAG.getIntPtrConstant(0));
4005 } else if (MVT::getSizeInBits(VT) == 64) {
4006 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4007 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4008 // to match extract_elt for f64.
4009 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4013 // UNPCKHPD the element to the lowest double word, then movsd.
4014 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4015 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4016 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
4017 SmallVector<SDOperand, 8> IdxVec;
4018 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
4020 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
4021 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4022 &IdxVec[0], IdxVec.size());
4023 SDOperand Vec = Op.getOperand(0);
4024 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4025 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4026 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4027 DAG.getIntPtrConstant(0));
4034 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDOperand Op, SelectionDAG &DAG){
4035 MVT::ValueType VT = Op.getValueType();
4036 MVT::ValueType EVT = MVT::getVectorElementType(VT);
4038 SDOperand N0 = Op.getOperand(0);
4039 SDOperand N1 = Op.getOperand(1);
4040 SDOperand N2 = Op.getOperand(2);
4042 if ((MVT::getSizeInBits(EVT) == 8) || (MVT::getSizeInBits(EVT) == 16)) {
4043 unsigned Opc = (MVT::getSizeInBits(EVT) == 8) ? X86ISD::PINSRB
4045 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4047 if (N1.getValueType() != MVT::i32)
4048 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4049 if (N2.getValueType() != MVT::i32)
4050 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4051 return DAG.getNode(Opc, VT, N0, N1, N2);
4052 } else if (EVT == MVT::f32) {
4053 // Bits [7:6] of the constant are the source select. This will always be
4054 // zero here. The DAG Combiner may combine an extract_elt index into these
4055 // bits. For example (insert (extract, 3), 2) could be matched by putting
4056 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4057 // Bits [5:4] of the constant are the destination select. This is the
4058 // value of the incoming immediate.
4059 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4060 // combine either bitwise AND or insert of float 0.0 to set these bits.
4061 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
4062 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4068 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
4069 MVT::ValueType VT = Op.getValueType();
4070 MVT::ValueType EVT = MVT::getVectorElementType(VT);
4072 if (Subtarget->hasSSE41())
4073 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4078 SDOperand N0 = Op.getOperand(0);
4079 SDOperand N1 = Op.getOperand(1);
4080 SDOperand N2 = Op.getOperand(2);
4082 if (MVT::getSizeInBits(EVT) == 16) {
4083 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4084 // as its second argument.
4085 if (N1.getValueType() != MVT::i32)
4086 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4087 if (N2.getValueType() != MVT::i32)
4088 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4089 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4095 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
4096 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4097 MVT::ValueType VT = MVT::v2i32;
4098 switch (Op.getValueType()) {
4105 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4106 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4109 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4110 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4111 // one of the above mentioned nodes. It has to be wrapped because otherwise
4112 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4113 // be used to form addressing mode. These wrapped nodes will be selected
4116 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
4117 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4118 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
4120 CP->getAlignment());
4121 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4122 // With PIC, the address is actually $g + Offset.
4123 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4124 !Subtarget->isPICStyleRIPRel()) {
4125 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4126 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4134 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
4135 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4136 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4137 // If it's a debug information descriptor, don't mess with it.
4138 if (DAG.isVerifiedDebugInfoDesc(Op))
4140 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4141 // With PIC, the address is actually $g + Offset.
4142 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4143 !Subtarget->isPICStyleRIPRel()) {
4144 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4145 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4149 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4150 // load the value at address GV, not the value of GV itself. This means that
4151 // the GlobalAddress must be in the base or index register of the address, not
4152 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4153 // The same applies for external symbols during PIC codegen
4154 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4155 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4156 PseudoSourceValue::getGOT(), 0);
4161 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
4163 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4164 const MVT::ValueType PtrVT) {
4166 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4167 DAG.getNode(X86ISD::GlobalBaseReg,
4169 InFlag = Chain.getValue(1);
4171 // emit leal symbol@TLSGD(,%ebx,1), %eax
4172 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4173 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4174 GA->getValueType(0),
4176 SDOperand Ops[] = { Chain, TGA, InFlag };
4177 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4178 InFlag = Result.getValue(2);
4179 Chain = Result.getValue(1);
4181 // call ___tls_get_addr. This function receives its argument in
4182 // the register EAX.
4183 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4184 InFlag = Chain.getValue(1);
4186 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4187 SDOperand Ops1[] = { Chain,
4188 DAG.getTargetExternalSymbol("___tls_get_addr",
4190 DAG.getRegister(X86::EAX, PtrVT),
4191 DAG.getRegister(X86::EBX, PtrVT),
4193 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4194 InFlag = Chain.getValue(1);
4196 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4199 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4200 // "local exec" model.
4202 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4203 const MVT::ValueType PtrVT) {
4204 // Get the Thread Pointer
4205 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4206 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4208 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4209 GA->getValueType(0),
4211 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4213 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4214 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4215 PseudoSourceValue::getGOT(), 0);
4217 // The address of the thread local variable is the add of the thread
4218 // pointer with the offset of the variable.
4219 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4223 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
4224 // TODO: implement the "local dynamic" model
4225 // TODO: implement the "initial exec"model for pic executables
4226 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
4227 "TLS not implemented for non-ELF and 64-bit targets");
4228 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4229 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4230 // otherwise use the "Local Exec"TLS Model
4231 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4232 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
4234 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4238 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
4239 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4240 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4241 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4242 // With PIC, the address is actually $g + Offset.
4243 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4244 !Subtarget->isPICStyleRIPRel()) {
4245 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4246 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4253 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
4254 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4255 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4256 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4257 // With PIC, the address is actually $g + Offset.
4258 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4259 !Subtarget->isPICStyleRIPRel()) {
4260 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4261 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4268 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4269 /// take a 2 x i32 value to shift plus a shift amount.
4270 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
4271 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4272 MVT::ValueType VT = Op.getValueType();
4273 unsigned VTBits = MVT::getSizeInBits(VT);
4274 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4275 SDOperand ShOpLo = Op.getOperand(0);
4276 SDOperand ShOpHi = Op.getOperand(1);
4277 SDOperand ShAmt = Op.getOperand(2);
4278 SDOperand Tmp1 = isSRA ?
4279 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4280 DAG.getConstant(0, VT);
4282 SDOperand Tmp2, Tmp3;
4283 if (Op.getOpcode() == ISD::SHL_PARTS) {
4284 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4285 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4287 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4288 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4291 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4292 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4293 DAG.getConstant(VTBits, MVT::i8));
4294 SDOperand Cond = DAG.getNode(X86ISD::CMP, VT,
4295 AndNode, DAG.getConstant(0, MVT::i8));
4298 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4299 VTs = DAG.getNodeValueTypes(VT, MVT::Flag);
4300 SmallVector<SDOperand, 4> Ops;
4301 if (Op.getOpcode() == ISD::SHL_PARTS) {
4302 Ops.push_back(Tmp2);
4303 Ops.push_back(Tmp3);
4305 Ops.push_back(Cond);
4306 Hi = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4309 Ops.push_back(Tmp3);
4310 Ops.push_back(Tmp1);
4312 Ops.push_back(Cond);
4313 Lo = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4315 Ops.push_back(Tmp2);
4316 Ops.push_back(Tmp3);
4318 Ops.push_back(Cond);
4319 Lo = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4322 Ops.push_back(Tmp3);
4323 Ops.push_back(Tmp1);
4325 Ops.push_back(Cond);
4326 Hi = DAG.getNode(X86ISD::CMOV, VT, &Ops[0], Ops.size());
4329 VTs = DAG.getNodeValueTypes(VT, VT);
4333 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
4336 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
4337 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
4338 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
4339 "Unknown SINT_TO_FP to lower!");
4341 // These are really Legal; caller falls through into that case.
4342 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4344 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4345 Subtarget->is64Bit())
4348 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
4349 MachineFunction &MF = DAG.getMachineFunction();
4350 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4351 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4352 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4354 PseudoSourceValue::getFixedStack(),
4359 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4361 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4363 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4364 SmallVector<SDOperand, 8> Ops;
4365 Ops.push_back(Chain);
4366 Ops.push_back(StackSlot);
4367 Ops.push_back(DAG.getValueType(SrcVT));
4368 SDOperand Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4369 Tys, &Ops[0], Ops.size());
4372 Chain = Result.getValue(1);
4373 SDOperand InFlag = Result.getValue(2);
4375 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4376 // shouldn't be necessary except that RFP cannot be live across
4377 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4378 MachineFunction &MF = DAG.getMachineFunction();
4379 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4380 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4381 Tys = DAG.getVTList(MVT::Other);
4382 SmallVector<SDOperand, 8> Ops;
4383 Ops.push_back(Chain);
4384 Ops.push_back(Result);
4385 Ops.push_back(StackSlot);
4386 Ops.push_back(DAG.getValueType(Op.getValueType()));
4387 Ops.push_back(InFlag);
4388 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4389 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4390 PseudoSourceValue::getFixedStack(), SSFI);
4396 std::pair<SDOperand,SDOperand> X86TargetLowering::
4397 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4398 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4399 "Unknown FP_TO_SINT to lower!");
4401 // These are really Legal.
4402 if (Op.getValueType() == MVT::i32 &&
4403 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4404 return std::make_pair(SDOperand(), SDOperand());
4405 if (Subtarget->is64Bit() &&
4406 Op.getValueType() == MVT::i64 &&
4407 Op.getOperand(0).getValueType() != MVT::f80)
4408 return std::make_pair(SDOperand(), SDOperand());
4410 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4412 MachineFunction &MF = DAG.getMachineFunction();
4413 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4414 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4415 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4417 switch (Op.getValueType()) {
4418 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4419 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4420 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4421 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4424 SDOperand Chain = DAG.getEntryNode();
4425 SDOperand Value = Op.getOperand(0);
4426 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4427 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4428 Chain = DAG.getStore(Chain, Value, StackSlot,
4429 PseudoSourceValue::getFixedStack(), SSFI);
4430 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4432 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4434 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4435 Chain = Value.getValue(1);
4436 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4437 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4440 // Build the FP_TO_INT*_IN_MEM
4441 SDOperand Ops[] = { Chain, Value, StackSlot };
4442 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4444 return std::make_pair(FIST, StackSlot);
4447 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4448 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4449 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4450 if (FIST.Val == 0) return SDOperand();
4453 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4456 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4457 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4458 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4459 if (FIST.Val == 0) return 0;
4461 // Return an i64 load from the stack slot.
4462 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4464 // Use a MERGE_VALUES node to drop the chain result value.
4465 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4468 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4469 MVT::ValueType VT = Op.getValueType();
4470 MVT::ValueType EltVT = VT;
4471 if (MVT::isVector(VT))
4472 EltVT = MVT::getVectorElementType(VT);
4473 std::vector<Constant*> CV;
4474 if (EltVT == MVT::f64) {
4475 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4479 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4485 Constant *C = ConstantVector::get(CV);
4486 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4487 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4488 PseudoSourceValue::getConstantPool(), 0,
4490 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4493 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4494 MVT::ValueType VT = Op.getValueType();
4495 MVT::ValueType EltVT = VT;
4496 unsigned EltNum = 1;
4497 if (MVT::isVector(VT)) {
4498 EltVT = MVT::getVectorElementType(VT);
4499 EltNum = MVT::getVectorNumElements(VT);
4501 std::vector<Constant*> CV;
4502 if (EltVT == MVT::f64) {
4503 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4507 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4513 Constant *C = ConstantVector::get(CV);
4514 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4515 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4516 PseudoSourceValue::getConstantPool(), 0,
4518 if (MVT::isVector(VT)) {
4519 return DAG.getNode(ISD::BIT_CONVERT, VT,
4520 DAG.getNode(ISD::XOR, MVT::v2i64,
4521 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4522 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4524 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4528 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4529 SDOperand Op0 = Op.getOperand(0);
4530 SDOperand Op1 = Op.getOperand(1);
4531 MVT::ValueType VT = Op.getValueType();
4532 MVT::ValueType SrcVT = Op1.getValueType();
4534 // If second operand is smaller, extend it first.
4535 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4536 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4539 // And if it is bigger, shrink it first.
4540 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4541 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4545 // At this point the operands and the result should have the same
4546 // type, and that won't be f80 since that is not custom lowered.
4548 // First get the sign bit of second operand.
4549 std::vector<Constant*> CV;
4550 if (SrcVT == MVT::f64) {
4551 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4552 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4554 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4555 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4556 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4557 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4559 Constant *C = ConstantVector::get(CV);
4560 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4561 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4562 PseudoSourceValue::getConstantPool(), 0,
4564 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4566 // Shift sign bit right or left if the two operands have different types.
4567 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4568 // Op0 is MVT::f32, Op1 is MVT::f64.
4569 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4570 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4571 DAG.getConstant(32, MVT::i32));
4572 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4573 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4574 DAG.getIntPtrConstant(0));
4577 // Clear first operand sign bit.
4579 if (VT == MVT::f64) {
4580 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4581 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4583 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4584 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4585 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4586 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4588 C = ConstantVector::get(CV);
4589 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4590 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4591 PseudoSourceValue::getConstantPool(), 0,
4593 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4595 // Or the value with the sign bit.
4596 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4599 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4600 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4602 SDOperand Op0 = Op.getOperand(0);
4603 SDOperand Op1 = Op.getOperand(1);
4604 SDOperand CC = Op.getOperand(2);
4605 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4606 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4609 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4611 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4612 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4613 DAG.getConstant(X86CC, MVT::i8), Cond);
4616 assert(isFP && "Illegal integer SetCC!");
4618 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4619 switch (SetCCOpcode) {
4620 default: assert(false && "Illegal floating point SetCC!");
4621 case ISD::SETOEQ: { // !PF & ZF
4622 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4623 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4624 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4625 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4626 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4628 case ISD::SETUNE: { // PF | !ZF
4629 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4630 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4631 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4632 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4633 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4639 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4640 bool addTest = true;
4641 SDOperand Cond = Op.getOperand(0);
4644 if (Cond.getOpcode() == ISD::SETCC)
4645 Cond = LowerSETCC(Cond, DAG);
4647 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4648 // setting operand in place of the X86ISD::SETCC.
4649 if (Cond.getOpcode() == X86ISD::SETCC) {
4650 CC = Cond.getOperand(0);
4652 SDOperand Cmp = Cond.getOperand(1);
4653 unsigned Opc = Cmp.getOpcode();
4654 MVT::ValueType VT = Op.getValueType();
4656 bool IllegalFPCMov = false;
4657 if (MVT::isFloatingPoint(VT) && !MVT::isVector(VT) &&
4658 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4659 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4661 if ((Opc == X86ISD::CMP ||
4662 Opc == X86ISD::COMI ||
4663 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4670 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4671 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4674 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4676 SmallVector<SDOperand, 4> Ops;
4677 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4678 // condition is true.
4679 Ops.push_back(Op.getOperand(2));
4680 Ops.push_back(Op.getOperand(1));
4682 Ops.push_back(Cond);
4683 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4686 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4687 bool addTest = true;
4688 SDOperand Chain = Op.getOperand(0);
4689 SDOperand Cond = Op.getOperand(1);
4690 SDOperand Dest = Op.getOperand(2);
4693 if (Cond.getOpcode() == ISD::SETCC)
4694 Cond = LowerSETCC(Cond, DAG);
4696 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4697 // setting operand in place of the X86ISD::SETCC.
4698 if (Cond.getOpcode() == X86ISD::SETCC) {
4699 CC = Cond.getOperand(0);
4701 SDOperand Cmp = Cond.getOperand(1);
4702 unsigned Opc = Cmp.getOpcode();
4703 if (Opc == X86ISD::CMP ||
4704 Opc == X86ISD::COMI ||
4705 Opc == X86ISD::UCOMI) {
4712 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4713 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4715 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4716 Chain, Op.getOperand(2), CC, Cond);
4720 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4721 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4722 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4723 // that the guard pages used by the OS virtual memory manager are allocated in
4724 // correct sequence.
4726 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4727 SelectionDAG &DAG) {
4728 assert(Subtarget->isTargetCygMing() &&
4729 "This should be used only on Cygwin/Mingw targets");
4732 SDOperand Chain = Op.getOperand(0);
4733 SDOperand Size = Op.getOperand(1);
4734 // FIXME: Ensure alignment here
4738 MVT::ValueType IntPtr = getPointerTy();
4739 MVT::ValueType SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
4741 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4742 Flag = Chain.getValue(1);
4744 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4745 SDOperand Ops[] = { Chain,
4746 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4747 DAG.getRegister(X86::EAX, IntPtr),
4749 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4750 Flag = Chain.getValue(1);
4752 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4754 std::vector<MVT::ValueType> Tys;
4755 Tys.push_back(SPTy);
4756 Tys.push_back(MVT::Other);
4757 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4758 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4762 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
4764 SDOperand Dst, SDOperand Src,
4765 SDOperand Size, unsigned Align,
4766 const Value *DstSV, uint64_t DstSVOff) {
4767 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
4769 /// If not DWORD aligned or size is more than the threshold, call the library.
4770 /// The libc version is likely to be faster for these cases. It can use the
4771 /// address value and run time information about the CPU.
4772 if ((Align & 3) == 0 ||
4774 ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
4775 SDOperand InFlag(0, 0);
4777 // Check to see if there is a specialized entry-point for memory zeroing.
4778 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
4779 if (const char *bzeroEntry =
4780 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
4781 MVT::ValueType IntPtr = getPointerTy();
4782 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4783 TargetLowering::ArgListTy Args;
4784 TargetLowering::ArgListEntry Entry;
4786 Entry.Ty = IntPtrTy;
4787 Args.push_back(Entry);
4789 Args.push_back(Entry);
4790 std::pair<SDOperand,SDOperand> CallResult =
4791 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
4792 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
4794 return CallResult.second;
4797 // Otherwise have the target-independent code call memset.
4801 uint64_t SizeVal = ConstantSize->getValue();
4802 SDOperand InFlag(0, 0);
4805 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
4806 unsigned BytesLeft = 0;
4807 bool TwoRepStos = false;
4810 uint64_t Val = ValC->getValue() & 255;
4812 // If the value is a constant, then we can potentially use larger sets.
4813 switch (Align & 3) {
4814 case 2: // WORD aligned
4817 Val = (Val << 8) | Val;
4819 case 0: // DWORD aligned
4822 Val = (Val << 8) | Val;
4823 Val = (Val << 16) | Val;
4824 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
4827 Val = (Val << 32) | Val;
4830 default: // Byte aligned
4833 Count = DAG.getIntPtrConstant(SizeVal);
4837 if (AVT > MVT::i8) {
4838 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4839 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
4840 BytesLeft = SizeVal % UBytes;
4843 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4845 InFlag = Chain.getValue(1);
4848 Count = DAG.getIntPtrConstant(SizeVal);
4849 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
4850 InFlag = Chain.getValue(1);
4853 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4855 InFlag = Chain.getValue(1);
4856 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4858 InFlag = Chain.getValue(1);
4860 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4861 SmallVector<SDOperand, 8> Ops;
4862 Ops.push_back(Chain);
4863 Ops.push_back(DAG.getValueType(AVT));
4864 Ops.push_back(InFlag);
4865 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4868 InFlag = Chain.getValue(1);
4870 MVT::ValueType CVT = Count.getValueType();
4871 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4872 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4873 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4875 InFlag = Chain.getValue(1);
4876 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4878 Ops.push_back(Chain);
4879 Ops.push_back(DAG.getValueType(MVT::i8));
4880 Ops.push_back(InFlag);
4881 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4882 } else if (BytesLeft) {
4883 // Handle the last 1 - 7 bytes.
4884 unsigned Offset = SizeVal - BytesLeft;
4885 MVT::ValueType AddrVT = Dst.getValueType();
4886 MVT::ValueType SizeVT = Size.getValueType();
4888 Chain = DAG.getMemset(Chain,
4889 DAG.getNode(ISD::ADD, AddrVT, Dst,
4890 DAG.getConstant(Offset, AddrVT)),
4892 DAG.getConstant(BytesLeft, SizeVT),
4893 Align, DstSV, DstSVOff + Offset);
4896 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
4901 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
4903 SDOperand Dst, SDOperand Src,
4904 SDOperand Size, unsigned Align,
4906 const Value *DstSV, uint64_t DstSVOff,
4907 const Value *SrcSV, uint64_t SrcSVOff){
4909 // This requires the copy size to be a constant, preferrably
4910 // within a subtarget-specific limit.
4911 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
4914 uint64_t SizeVal = ConstantSize->getValue();
4915 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
4919 unsigned BytesLeft = 0;
4920 if (Align >= 8 && Subtarget->is64Bit())
4922 else if (Align >= 4)
4924 else if (Align >= 2)
4929 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4930 unsigned CountVal = SizeVal / UBytes;
4931 SDOperand Count = DAG.getIntPtrConstant(CountVal);
4932 BytesLeft = SizeVal % UBytes;
4934 SDOperand InFlag(0, 0);
4935 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4937 InFlag = Chain.getValue(1);
4938 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4940 InFlag = Chain.getValue(1);
4941 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4943 InFlag = Chain.getValue(1);
4945 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4946 SmallVector<SDOperand, 8> Ops;
4947 Ops.push_back(Chain);
4948 Ops.push_back(DAG.getValueType(AVT));
4949 Ops.push_back(InFlag);
4950 SDOperand RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4952 SmallVector<SDOperand, 4> Results;
4953 Results.push_back(RepMovs);
4955 // Handle the last 1 - 7 bytes.
4956 unsigned Offset = SizeVal - BytesLeft;
4957 MVT::ValueType DstVT = Dst.getValueType();
4958 MVT::ValueType SrcVT = Src.getValueType();
4959 MVT::ValueType SizeVT = Size.getValueType();
4960 Results.push_back(DAG.getMemcpy(Chain,
4961 DAG.getNode(ISD::ADD, DstVT, Dst,
4962 DAG.getConstant(Offset, DstVT)),
4963 DAG.getNode(ISD::ADD, SrcVT, Src,
4964 DAG.getConstant(Offset, SrcVT)),
4965 DAG.getConstant(BytesLeft, SizeVT),
4966 Align, AlwaysInline,
4967 DstSV, DstSVOff + Offset,
4968 SrcSV, SrcSVOff + Offset));
4971 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
4974 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4975 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4976 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4977 SDOperand TheChain = N->getOperand(0);
4978 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4979 if (Subtarget->is64Bit()) {
4980 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4981 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4982 MVT::i64, rax.getValue(2));
4983 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4984 DAG.getConstant(32, MVT::i8));
4986 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4989 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4990 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4993 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4994 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4995 MVT::i32, eax.getValue(2));
4996 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4997 SDOperand Ops[] = { eax, edx };
4998 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5000 // Use a MERGE_VALUES to return the value and chain.
5001 Ops[1] = edx.getValue(1);
5002 Tys = DAG.getVTList(MVT::i64, MVT::Other);
5003 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
5006 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
5007 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5009 if (!Subtarget->is64Bit()) {
5010 // vastart just stores the address of the VarArgsFrameIndex slot into the
5011 // memory location argument.
5012 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5013 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5017 // gp_offset (0 - 6 * 8)
5018 // fp_offset (48 - 48 + 8 * 16)
5019 // overflow_arg_area (point to parameters coming in memory).
5021 SmallVector<SDOperand, 8> MemOps;
5022 SDOperand FIN = Op.getOperand(1);
5024 SDOperand Store = DAG.getStore(Op.getOperand(0),
5025 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5027 MemOps.push_back(Store);
5030 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5031 Store = DAG.getStore(Op.getOperand(0),
5032 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5034 MemOps.push_back(Store);
5036 // Store ptr to overflow_arg_area
5037 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5038 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5039 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5040 MemOps.push_back(Store);
5042 // Store ptr to reg_save_area.
5043 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5044 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5045 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5046 MemOps.push_back(Store);
5047 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5050 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
5051 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5052 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5053 SDOperand Chain = Op.getOperand(0);
5054 SDOperand DstPtr = Op.getOperand(1);
5055 SDOperand SrcPtr = Op.getOperand(2);
5056 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5057 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5059 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5060 DAG.getIntPtrConstant(24), 8, false,
5061 DstSV, 0, SrcSV, 0);
5065 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
5066 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5068 default: return SDOperand(); // Don't custom lower most intrinsics.
5069 // Comparison intrinsics.
5070 case Intrinsic::x86_sse_comieq_ss:
5071 case Intrinsic::x86_sse_comilt_ss:
5072 case Intrinsic::x86_sse_comile_ss:
5073 case Intrinsic::x86_sse_comigt_ss:
5074 case Intrinsic::x86_sse_comige_ss:
5075 case Intrinsic::x86_sse_comineq_ss:
5076 case Intrinsic::x86_sse_ucomieq_ss:
5077 case Intrinsic::x86_sse_ucomilt_ss:
5078 case Intrinsic::x86_sse_ucomile_ss:
5079 case Intrinsic::x86_sse_ucomigt_ss:
5080 case Intrinsic::x86_sse_ucomige_ss:
5081 case Intrinsic::x86_sse_ucomineq_ss:
5082 case Intrinsic::x86_sse2_comieq_sd:
5083 case Intrinsic::x86_sse2_comilt_sd:
5084 case Intrinsic::x86_sse2_comile_sd:
5085 case Intrinsic::x86_sse2_comigt_sd:
5086 case Intrinsic::x86_sse2_comige_sd:
5087 case Intrinsic::x86_sse2_comineq_sd:
5088 case Intrinsic::x86_sse2_ucomieq_sd:
5089 case Intrinsic::x86_sse2_ucomilt_sd:
5090 case Intrinsic::x86_sse2_ucomile_sd:
5091 case Intrinsic::x86_sse2_ucomigt_sd:
5092 case Intrinsic::x86_sse2_ucomige_sd:
5093 case Intrinsic::x86_sse2_ucomineq_sd: {
5095 ISD::CondCode CC = ISD::SETCC_INVALID;
5098 case Intrinsic::x86_sse_comieq_ss:
5099 case Intrinsic::x86_sse2_comieq_sd:
5103 case Intrinsic::x86_sse_comilt_ss:
5104 case Intrinsic::x86_sse2_comilt_sd:
5108 case Intrinsic::x86_sse_comile_ss:
5109 case Intrinsic::x86_sse2_comile_sd:
5113 case Intrinsic::x86_sse_comigt_ss:
5114 case Intrinsic::x86_sse2_comigt_sd:
5118 case Intrinsic::x86_sse_comige_ss:
5119 case Intrinsic::x86_sse2_comige_sd:
5123 case Intrinsic::x86_sse_comineq_ss:
5124 case Intrinsic::x86_sse2_comineq_sd:
5128 case Intrinsic::x86_sse_ucomieq_ss:
5129 case Intrinsic::x86_sse2_ucomieq_sd:
5130 Opc = X86ISD::UCOMI;
5133 case Intrinsic::x86_sse_ucomilt_ss:
5134 case Intrinsic::x86_sse2_ucomilt_sd:
5135 Opc = X86ISD::UCOMI;
5138 case Intrinsic::x86_sse_ucomile_ss:
5139 case Intrinsic::x86_sse2_ucomile_sd:
5140 Opc = X86ISD::UCOMI;
5143 case Intrinsic::x86_sse_ucomigt_ss:
5144 case Intrinsic::x86_sse2_ucomigt_sd:
5145 Opc = X86ISD::UCOMI;
5148 case Intrinsic::x86_sse_ucomige_ss:
5149 case Intrinsic::x86_sse2_ucomige_sd:
5150 Opc = X86ISD::UCOMI;
5153 case Intrinsic::x86_sse_ucomineq_ss:
5154 case Intrinsic::x86_sse2_ucomineq_sd:
5155 Opc = X86ISD::UCOMI;
5161 SDOperand LHS = Op.getOperand(1);
5162 SDOperand RHS = Op.getOperand(2);
5163 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5165 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5166 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5167 DAG.getConstant(X86CC, MVT::i8), Cond);
5168 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
5173 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
5174 // Depths > 0 not supported yet!
5175 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5178 // Just load the return address
5179 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5180 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5183 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
5184 // Depths > 0 not supported yet!
5185 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5188 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5189 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5190 DAG.getIntPtrConstant(4));
5193 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
5194 SelectionDAG &DAG) {
5195 // Is not yet supported on x86-64
5196 if (Subtarget->is64Bit())
5199 return DAG.getIntPtrConstant(8);
5202 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
5204 assert(!Subtarget->is64Bit() &&
5205 "Lowering of eh_return builtin is not supported yet on x86-64");
5207 MachineFunction &MF = DAG.getMachineFunction();
5208 SDOperand Chain = Op.getOperand(0);
5209 SDOperand Offset = Op.getOperand(1);
5210 SDOperand Handler = Op.getOperand(2);
5212 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
5215 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5216 DAG.getIntPtrConstant(-4UL));
5217 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5218 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5219 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5220 MF.getRegInfo().addLiveOut(X86::ECX);
5222 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5223 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5226 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
5227 SelectionDAG &DAG) {
5228 SDOperand Root = Op.getOperand(0);
5229 SDOperand Trmp = Op.getOperand(1); // trampoline
5230 SDOperand FPtr = Op.getOperand(2); // nested function
5231 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
5233 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5235 const X86InstrInfo *TII =
5236 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5238 if (Subtarget->is64Bit()) {
5239 SDOperand OutChains[6];
5241 // Large code-model.
5243 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5244 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5246 const unsigned char N86R10 =
5247 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R10);
5248 const unsigned char N86R11 =
5249 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R11);
5251 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5253 // Load the pointer to the nested function into R11.
5254 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5255 SDOperand Addr = Trmp;
5256 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5259 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5260 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5262 // Load the 'nest' parameter value into R10.
5263 // R10 is specified in X86CallingConv.td
5264 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5265 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5266 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5269 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5270 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5272 // Jump to the nested function.
5273 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5274 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5275 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5278 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5279 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5280 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5284 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5285 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5287 const Function *Func =
5288 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5289 unsigned CC = Func->getCallingConv();
5294 assert(0 && "Unsupported calling convention");
5295 case CallingConv::C:
5296 case CallingConv::X86_StdCall: {
5297 // Pass 'nest' parameter in ECX.
5298 // Must be kept in sync with X86CallingConv.td
5301 // Check that ECX wasn't needed by an 'inreg' parameter.
5302 const FunctionType *FTy = Func->getFunctionType();
5303 const PAListPtr &Attrs = Func->getParamAttrs();
5305 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5306 unsigned InRegCount = 0;
5309 for (FunctionType::param_iterator I = FTy->param_begin(),
5310 E = FTy->param_end(); I != E; ++I, ++Idx)
5311 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5312 // FIXME: should only count parameters that are lowered to integers.
5313 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5315 if (InRegCount > 2) {
5316 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5322 case CallingConv::X86_FastCall:
5323 // Pass 'nest' parameter in EAX.
5324 // Must be kept in sync with X86CallingConv.td
5329 SDOperand OutChains[4];
5330 SDOperand Addr, Disp;
5332 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5333 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5335 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5336 const unsigned char N86Reg =
5337 ((const X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5338 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5341 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5342 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5344 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5345 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5346 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5347 TrmpAddr, 5, false, 1);
5349 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5350 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5353 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5354 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5358 SDOperand X86TargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
5360 The rounding mode is in bits 11:10 of FPSR, and has the following
5367 FLT_ROUNDS, on the other hand, expects the following:
5374 To perform the conversion, we do:
5375 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5378 MachineFunction &MF = DAG.getMachineFunction();
5379 const TargetMachine &TM = MF.getTarget();
5380 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5381 unsigned StackAlignment = TFI.getStackAlignment();
5382 MVT::ValueType VT = Op.getValueType();
5384 // Save FP Control Word to stack slot
5385 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5386 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5388 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5389 DAG.getEntryNode(), StackSlot);
5391 // Load FP Control Word from stack slot
5392 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5394 // Transform as necessary
5396 DAG.getNode(ISD::SRL, MVT::i16,
5397 DAG.getNode(ISD::AND, MVT::i16,
5398 CWD, DAG.getConstant(0x800, MVT::i16)),
5399 DAG.getConstant(11, MVT::i8));
5401 DAG.getNode(ISD::SRL, MVT::i16,
5402 DAG.getNode(ISD::AND, MVT::i16,
5403 CWD, DAG.getConstant(0x400, MVT::i16)),
5404 DAG.getConstant(9, MVT::i8));
5407 DAG.getNode(ISD::AND, MVT::i16,
5408 DAG.getNode(ISD::ADD, MVT::i16,
5409 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5410 DAG.getConstant(1, MVT::i16)),
5411 DAG.getConstant(3, MVT::i16));
5414 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5415 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5418 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5419 MVT::ValueType VT = Op.getValueType();
5420 MVT::ValueType OpVT = VT;
5421 unsigned NumBits = MVT::getSizeInBits(VT);
5423 Op = Op.getOperand(0);
5424 if (VT == MVT::i8) {
5425 // Zero extend to i32 since there is not an i8 bsr.
5427 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5430 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5431 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5432 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5434 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5435 SmallVector<SDOperand, 4> Ops;
5437 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5438 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5439 Ops.push_back(Op.getValue(1));
5440 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5442 // Finally xor with NumBits-1.
5443 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5446 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5450 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5451 MVT::ValueType VT = Op.getValueType();
5452 MVT::ValueType OpVT = VT;
5453 unsigned NumBits = MVT::getSizeInBits(VT);
5455 Op = Op.getOperand(0);
5456 if (VT == MVT::i8) {
5458 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5461 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5462 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5463 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5465 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5466 SmallVector<SDOperand, 4> Ops;
5468 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5469 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5470 Ops.push_back(Op.getValue(1));
5471 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5474 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5478 SDOperand X86TargetLowering::LowerLCS(SDOperand Op, SelectionDAG &DAG) {
5479 MVT::ValueType T = cast<AtomicSDNode>(Op.Val)->getVT();
5483 case MVT::i8: Reg = X86::AL; size = 1; break;
5484 case MVT::i16: Reg = X86::AX; size = 2; break;
5485 case MVT::i32: Reg = X86::EAX; size = 4; break;
5487 if (Subtarget->is64Bit()) {
5488 Reg = X86::RAX; size = 8;
5489 } else //Should go away when LowerType stuff lands
5490 return SDOperand(ExpandATOMIC_LCS(Op.Val, DAG), 0);
5493 SDOperand cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5494 Op.getOperand(3), SDOperand());
5495 SDOperand Ops[] = { cpIn.getValue(0),
5498 DAG.getTargetConstant(size, MVT::i8),
5500 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5501 SDOperand Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5503 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5507 SDNode* X86TargetLowering::ExpandATOMIC_LCS(SDNode* Op, SelectionDAG &DAG) {
5508 MVT::ValueType T = cast<AtomicSDNode>(Op)->getVT();
5509 assert (T == MVT::i64 && "Only know how to expand i64 CAS");
5510 SDOperand cpInL, cpInH;
5511 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5512 DAG.getConstant(0, MVT::i32));
5513 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5514 DAG.getConstant(1, MVT::i32));
5515 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5516 cpInL, SDOperand());
5517 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5518 cpInH, cpInL.getValue(1));
5519 SDOperand swapInL, swapInH;
5520 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5521 DAG.getConstant(0, MVT::i32));
5522 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5523 DAG.getConstant(1, MVT::i32));
5524 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5525 swapInL, cpInH.getValue(1));
5526 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5527 swapInH, swapInL.getValue(1));
5528 SDOperand Ops[] = { swapInH.getValue(0),
5530 swapInH.getValue(1)};
5531 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5532 SDOperand Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5533 SDOperand cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5534 Result.getValue(1));
5535 SDOperand cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5536 cpOutL.getValue(2));
5537 SDOperand OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5538 SDOperand ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5539 Tys = DAG.getVTList(MVT::i64, MVT::Other);
5540 return DAG.getNode(ISD::MERGE_VALUES, Tys, ResultVal, cpOutH.getValue(1)).Val;
5543 /// LowerOperation - Provide custom lowering hooks for some operations.
5545 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5546 switch (Op.getOpcode()) {
5547 default: assert(0 && "Should not custom lower this!");
5548 case ISD::ATOMIC_LCS: return LowerLCS(Op,DAG);
5549 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5550 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5551 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5552 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5553 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5554 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5555 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5556 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5557 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5558 case ISD::SHL_PARTS:
5559 case ISD::SRA_PARTS:
5560 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5561 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5562 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5563 case ISD::FABS: return LowerFABS(Op, DAG);
5564 case ISD::FNEG: return LowerFNEG(Op, DAG);
5565 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5566 case ISD::SETCC: return LowerSETCC(Op, DAG);
5567 case ISD::SELECT: return LowerSELECT(Op, DAG);
5568 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5569 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5570 case ISD::CALL: return LowerCALL(Op, DAG);
5571 case ISD::RET: return LowerRET(Op, DAG);
5572 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5573 case ISD::VASTART: return LowerVASTART(Op, DAG);
5574 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5575 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5576 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5577 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5578 case ISD::FRAME_TO_ARGS_OFFSET:
5579 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5580 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5581 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5582 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5583 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5584 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5585 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5587 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5588 case ISD::READCYCLECOUNTER:
5589 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5593 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5594 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5595 switch (N->getOpcode()) {
5596 default: assert(0 && "Should not custom lower this!");
5597 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5598 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5599 case ISD::ATOMIC_LCS: return ExpandATOMIC_LCS(N, DAG);
5603 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5605 default: return NULL;
5606 case X86ISD::BSF: return "X86ISD::BSF";
5607 case X86ISD::BSR: return "X86ISD::BSR";
5608 case X86ISD::SHLD: return "X86ISD::SHLD";
5609 case X86ISD::SHRD: return "X86ISD::SHRD";
5610 case X86ISD::FAND: return "X86ISD::FAND";
5611 case X86ISD::FOR: return "X86ISD::FOR";
5612 case X86ISD::FXOR: return "X86ISD::FXOR";
5613 case X86ISD::FSRL: return "X86ISD::FSRL";
5614 case X86ISD::FILD: return "X86ISD::FILD";
5615 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5616 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5617 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5618 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5619 case X86ISD::FLD: return "X86ISD::FLD";
5620 case X86ISD::FST: return "X86ISD::FST";
5621 case X86ISD::CALL: return "X86ISD::CALL";
5622 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5623 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5624 case X86ISD::CMP: return "X86ISD::CMP";
5625 case X86ISD::COMI: return "X86ISD::COMI";
5626 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5627 case X86ISD::SETCC: return "X86ISD::SETCC";
5628 case X86ISD::CMOV: return "X86ISD::CMOV";
5629 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5630 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5631 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5632 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5633 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5634 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5635 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
5636 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5637 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
5638 case X86ISD::PINSRB: return "X86ISD::PINSRB";
5639 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5640 case X86ISD::FMAX: return "X86ISD::FMAX";
5641 case X86ISD::FMIN: return "X86ISD::FMIN";
5642 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5643 case X86ISD::FRCP: return "X86ISD::FRCP";
5644 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5645 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5646 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5647 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5648 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5649 case X86ISD::LCMPXCHG_DAG: return "x86ISD::LCMPXCHG_DAG";
5650 case X86ISD::LCMPXCHG8_DAG: return "x86ISD::LCMPXCHG8_DAG";
5654 // isLegalAddressingMode - Return true if the addressing mode represented
5655 // by AM is legal for this target, for a load/store of the specified type.
5656 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5657 const Type *Ty) const {
5658 // X86 supports extremely general addressing modes.
5660 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5661 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5665 // We can only fold this if we don't need an extra load.
5666 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5669 // X86-64 only supports addr of globals in small code model.
5670 if (Subtarget->is64Bit()) {
5671 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5673 // If lower 4G is not available, then we must use rip-relative addressing.
5674 if (AM.BaseOffs || AM.Scale > 1)
5685 // These scales always work.
5690 // These scales are formed with basereg+scalereg. Only accept if there is
5695 default: // Other stuff never works.
5703 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5704 if (!Ty1->isInteger() || !Ty2->isInteger())
5706 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5707 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5708 if (NumBits1 <= NumBits2)
5710 return Subtarget->is64Bit() || NumBits1 < 64;
5713 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5714 MVT::ValueType VT2) const {
5715 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5717 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5718 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5719 if (NumBits1 <= NumBits2)
5721 return Subtarget->is64Bit() || NumBits1 < 64;
5724 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5725 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5726 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5727 /// are assumed to be legal.
5729 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5730 // Only do shuffles on 128-bit vector types for now.
5731 if (MVT::getSizeInBits(VT) == 64) return false;
5732 return (Mask.Val->getNumOperands() <= 4 ||
5733 isIdentityMask(Mask.Val) ||
5734 isIdentityMask(Mask.Val, true) ||
5735 isSplatMask(Mask.Val) ||
5736 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5737 X86::isUNPCKLMask(Mask.Val) ||
5738 X86::isUNPCKHMask(Mask.Val) ||
5739 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5740 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5744 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDOperand> &BVOps,
5746 SelectionDAG &DAG) const {
5747 unsigned NumElts = BVOps.size();
5748 // Only do shuffles on 128-bit vector types for now.
5749 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5750 if (NumElts == 2) return true;
5752 return (isMOVLMask(&BVOps[0], 4) ||
5753 isCommutedMOVL(&BVOps[0], 4, true) ||
5754 isSHUFPMask(&BVOps[0], 4) ||
5755 isCommutedSHUFP(&BVOps[0], 4));
5760 //===----------------------------------------------------------------------===//
5761 // X86 Scheduler Hooks
5762 //===----------------------------------------------------------------------===//
5765 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5766 MachineBasicBlock *BB) {
5767 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5768 switch (MI->getOpcode()) {
5769 default: assert(false && "Unexpected instr type to insert");
5770 case X86::CMOV_FR32:
5771 case X86::CMOV_FR64:
5772 case X86::CMOV_V4F32:
5773 case X86::CMOV_V2F64:
5774 case X86::CMOV_V2I64: {
5775 // To "insert" a SELECT_CC instruction, we actually have to insert the
5776 // diamond control-flow pattern. The incoming instruction knows the
5777 // destination vreg to set, the condition code register to branch on, the
5778 // true/false values to select between, and a branch opcode to use.
5779 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5780 ilist<MachineBasicBlock>::iterator It = BB;
5786 // cmpTY ccX, r1, r2
5788 // fallthrough --> copy0MBB
5789 MachineBasicBlock *thisMBB = BB;
5790 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5791 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5793 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5794 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5795 MachineFunction *F = BB->getParent();
5796 F->getBasicBlockList().insert(It, copy0MBB);
5797 F->getBasicBlockList().insert(It, sinkMBB);
5798 // Update machine-CFG edges by first adding all successors of the current
5799 // block to the new block which will contain the Phi node for the select.
5800 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5801 e = BB->succ_end(); i != e; ++i)
5802 sinkMBB->addSuccessor(*i);
5803 // Next, remove all successors of the current block, and add the true
5804 // and fallthrough blocks as its successors.
5805 while(!BB->succ_empty())
5806 BB->removeSuccessor(BB->succ_begin());
5807 BB->addSuccessor(copy0MBB);
5808 BB->addSuccessor(sinkMBB);
5811 // %FalseValue = ...
5812 // # fallthrough to sinkMBB
5815 // Update machine-CFG edges
5816 BB->addSuccessor(sinkMBB);
5819 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5822 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5823 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5824 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5826 delete MI; // The pseudo instruction is gone now.
5830 case X86::FP32_TO_INT16_IN_MEM:
5831 case X86::FP32_TO_INT32_IN_MEM:
5832 case X86::FP32_TO_INT64_IN_MEM:
5833 case X86::FP64_TO_INT16_IN_MEM:
5834 case X86::FP64_TO_INT32_IN_MEM:
5835 case X86::FP64_TO_INT64_IN_MEM:
5836 case X86::FP80_TO_INT16_IN_MEM:
5837 case X86::FP80_TO_INT32_IN_MEM:
5838 case X86::FP80_TO_INT64_IN_MEM: {
5839 // Change the floating point control register to use "round towards zero"
5840 // mode when truncating to an integer value.
5841 MachineFunction *F = BB->getParent();
5842 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5843 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5845 // Load the old value of the high byte of the control word...
5847 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
5848 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5850 // Set the high part to be round to zero...
5851 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5854 // Reload the modified control word now...
5855 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5857 // Restore the memory image of control word to original value
5858 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5861 // Get the X86 opcode to use.
5863 switch (MI->getOpcode()) {
5864 default: assert(0 && "illegal opcode!");
5865 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5866 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5867 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5868 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5869 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5870 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5871 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5872 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5873 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5877 MachineOperand &Op = MI->getOperand(0);
5878 if (Op.isRegister()) {
5879 AM.BaseType = X86AddressMode::RegBase;
5880 AM.Base.Reg = Op.getReg();
5882 AM.BaseType = X86AddressMode::FrameIndexBase;
5883 AM.Base.FrameIndex = Op.getIndex();
5885 Op = MI->getOperand(1);
5886 if (Op.isImmediate())
5887 AM.Scale = Op.getImm();
5888 Op = MI->getOperand(2);
5889 if (Op.isImmediate())
5890 AM.IndexReg = Op.getImm();
5891 Op = MI->getOperand(3);
5892 if (Op.isGlobalAddress()) {
5893 AM.GV = Op.getGlobal();
5895 AM.Disp = Op.getImm();
5897 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5898 .addReg(MI->getOperand(4).getReg());
5900 // Reload the original control word now.
5901 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5903 delete MI; // The pseudo instruction is gone now.
5909 //===----------------------------------------------------------------------===//
5910 // X86 Optimization Hooks
5911 //===----------------------------------------------------------------------===//
5913 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5917 const SelectionDAG &DAG,
5918 unsigned Depth) const {
5919 unsigned Opc = Op.getOpcode();
5920 assert((Opc >= ISD::BUILTIN_OP_END ||
5921 Opc == ISD::INTRINSIC_WO_CHAIN ||
5922 Opc == ISD::INTRINSIC_W_CHAIN ||
5923 Opc == ISD::INTRINSIC_VOID) &&
5924 "Should use MaskedValueIsZero if you don't know whether Op"
5925 " is a target node!");
5927 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
5931 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
5932 Mask.getBitWidth() - 1);
5937 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5938 /// element of the result of the vector shuffle.
5939 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5940 MVT::ValueType VT = N->getValueType(0);
5941 SDOperand PermMask = N->getOperand(2);
5942 unsigned NumElems = PermMask.getNumOperands();
5943 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5945 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5947 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5948 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5949 SDOperand Idx = PermMask.getOperand(i);
5950 if (Idx.getOpcode() == ISD::UNDEF)
5951 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5952 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5957 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5958 /// node is a GlobalAddress + an offset.
5959 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5960 unsigned Opc = N->getOpcode();
5961 if (Opc == X86ISD::Wrapper) {
5962 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5963 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5966 } else if (Opc == ISD::ADD) {
5967 SDOperand N1 = N->getOperand(0);
5968 SDOperand N2 = N->getOperand(1);
5969 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5970 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5972 Offset += V->getSignExtended();
5975 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5976 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5978 Offset += V->getSignExtended();
5986 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5988 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5989 MachineFrameInfo *MFI) {
5990 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5993 SDOperand Loc = N->getOperand(1);
5994 SDOperand BaseLoc = Base->getOperand(1);
5995 if (Loc.getOpcode() == ISD::FrameIndex) {
5996 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5998 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5999 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
6000 int FS = MFI->getObjectSize(FI);
6001 int BFS = MFI->getObjectSize(BFI);
6002 if (FS != BFS || FS != Size) return false;
6003 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
6005 GlobalValue *GV1 = NULL;
6006 GlobalValue *GV2 = NULL;
6007 int64_t Offset1 = 0;
6008 int64_t Offset2 = 0;
6009 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
6010 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
6011 if (isGA1 && isGA2 && GV1 == GV2)
6012 return Offset1 == (Offset2 + Dist*Size);
6018 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
6019 const X86Subtarget *Subtarget) {
6022 if (isGAPlusOffset(Base, GV, Offset))
6023 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
6024 // DAG combine handles the stack object case.
6029 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6030 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6031 /// if the load addresses are consecutive, non-overlapping, and in the right
6033 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6034 const X86Subtarget *Subtarget) {
6035 MachineFunction &MF = DAG.getMachineFunction();
6036 MachineFrameInfo *MFI = MF.getFrameInfo();
6037 MVT::ValueType VT = N->getValueType(0);
6038 MVT::ValueType EVT = MVT::getVectorElementType(VT);
6039 SDOperand PermMask = N->getOperand(2);
6040 int NumElems = (int)PermMask.getNumOperands();
6041 SDNode *Base = NULL;
6042 for (int i = 0; i < NumElems; ++i) {
6043 SDOperand Idx = PermMask.getOperand(i);
6044 if (Idx.getOpcode() == ISD::UNDEF) {
6045 if (!Base) return SDOperand();
6048 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
6049 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
6053 else if (!isConsecutiveLoad(Arg.Val, Base,
6054 i, MVT::getSizeInBits(EVT)/8,MFI))
6059 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
6060 LoadSDNode *LD = cast<LoadSDNode>(Base);
6062 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6063 LD->getSrcValueOffset(), LD->isVolatile());
6065 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6066 LD->getSrcValueOffset(), LD->isVolatile(),
6067 LD->getAlignment());
6071 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6072 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6073 const X86Subtarget *Subtarget) {
6074 SDOperand Cond = N->getOperand(0);
6076 // If we have SSE[12] support, try to form min/max nodes.
6077 if (Subtarget->hasSSE2() &&
6078 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6079 if (Cond.getOpcode() == ISD::SETCC) {
6080 // Get the LHS/RHS of the select.
6081 SDOperand LHS = N->getOperand(1);
6082 SDOperand RHS = N->getOperand(2);
6083 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6085 unsigned Opcode = 0;
6086 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6089 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6092 if (!UnsafeFPMath) break;
6094 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6096 Opcode = X86ISD::FMIN;
6099 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6102 if (!UnsafeFPMath) break;
6104 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6106 Opcode = X86ISD::FMAX;
6109 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6112 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6115 if (!UnsafeFPMath) break;
6117 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6119 Opcode = X86ISD::FMIN;
6122 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6125 if (!UnsafeFPMath) break;
6127 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6129 Opcode = X86ISD::FMAX;
6135 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6143 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6144 static SDOperand PerformSTORECombine(StoreSDNode *St, SelectionDAG &DAG,
6145 const X86Subtarget *Subtarget) {
6146 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6147 // the FP state in cases where an emms may be missing.
6148 // A preferable solution to the general problem is to figure out the right
6149 // places to insert EMMS. This qualifies as a quick hack.
6150 if (MVT::isVector(St->getValue().getValueType()) &&
6151 MVT::getSizeInBits(St->getValue().getValueType()) == 64 &&
6152 isa<LoadSDNode>(St->getValue()) &&
6153 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6154 St->getChain().hasOneUse() && !St->isVolatile()) {
6155 SDNode* LdVal = St->getValue().Val;
6157 int TokenFactorIndex = -1;
6158 SmallVector<SDOperand, 8> Ops;
6159 SDNode* ChainVal = St->getChain().Val;
6160 // Must be a store of a load. We currently handle two cases: the load
6161 // is a direct child, and it's under an intervening TokenFactor. It is
6162 // possible to dig deeper under nested TokenFactors.
6163 if (ChainVal == LdVal)
6164 Ld = cast<LoadSDNode>(St->getChain());
6165 else if (St->getValue().hasOneUse() &&
6166 ChainVal->getOpcode() == ISD::TokenFactor) {
6167 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6168 if (ChainVal->getOperand(i).Val == LdVal) {
6169 TokenFactorIndex = i;
6170 Ld = cast<LoadSDNode>(St->getValue());
6172 Ops.push_back(ChainVal->getOperand(i));
6176 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6177 if (Subtarget->is64Bit()) {
6178 SDOperand NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6179 Ld->getBasePtr(), Ld->getSrcValue(),
6180 Ld->getSrcValueOffset(), Ld->isVolatile(),
6181 Ld->getAlignment());
6182 SDOperand NewChain = NewLd.getValue(1);
6183 if (TokenFactorIndex != -1) {
6184 Ops.push_back(NewChain);
6185 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6188 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6189 St->getSrcValue(), St->getSrcValueOffset(),
6190 St->isVolatile(), St->getAlignment());
6193 // Otherwise, lower to two 32-bit copies.
6194 SDOperand LoAddr = Ld->getBasePtr();
6195 SDOperand HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6196 DAG.getConstant(MVT::i32, 4));
6198 SDOperand LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6199 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6200 Ld->isVolatile(), Ld->getAlignment());
6201 SDOperand HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6202 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6204 MinAlign(Ld->getAlignment(), 4));
6206 SDOperand NewChain = LoLd.getValue(1);
6207 if (TokenFactorIndex != -1) {
6208 Ops.push_back(LoLd);
6209 Ops.push_back(HiLd);
6210 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6214 LoAddr = St->getBasePtr();
6215 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6216 DAG.getConstant(MVT::i32, 4));
6218 SDOperand LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6219 St->getSrcValue(), St->getSrcValueOffset(),
6220 St->isVolatile(), St->getAlignment());
6221 SDOperand HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6222 St->getSrcValue(), St->getSrcValueOffset()+4,
6224 MinAlign(St->getAlignment(), 4));
6225 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6231 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6232 /// X86ISD::FXOR nodes.
6233 static SDOperand PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6234 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6235 // F[X]OR(0.0, x) -> x
6236 // F[X]OR(x, 0.0) -> x
6237 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6238 if (C->getValueAPF().isPosZero())
6239 return N->getOperand(1);
6240 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6241 if (C->getValueAPF().isPosZero())
6242 return N->getOperand(0);
6246 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6247 static SDOperand PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6248 // FAND(0.0, x) -> 0.0
6249 // FAND(x, 0.0) -> 0.0
6250 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6251 if (C->getValueAPF().isPosZero())
6252 return N->getOperand(0);
6253 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6254 if (C->getValueAPF().isPosZero())
6255 return N->getOperand(1);
6260 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
6261 DAGCombinerInfo &DCI) const {
6262 SelectionDAG &DAG = DCI.DAG;
6263 switch (N->getOpcode()) {
6265 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, Subtarget);
6266 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
6268 return PerformSTORECombine(cast<StoreSDNode>(N), DAG, Subtarget);
6270 case X86ISD::FOR: return PerformFORCombine(N, DAG);
6271 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
6277 //===----------------------------------------------------------------------===//
6278 // X86 Inline Assembly Support
6279 //===----------------------------------------------------------------------===//
6281 /// getConstraintType - Given a constraint letter, return the type of
6282 /// constraint it is for this target.
6283 X86TargetLowering::ConstraintType
6284 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
6285 if (Constraint.size() == 1) {
6286 switch (Constraint[0]) {
6297 return C_RegisterClass;
6302 return TargetLowering::getConstraintType(Constraint);
6305 /// LowerXConstraint - try to replace an X constraint, which matches anything,
6306 /// with another that has more specific requirements based on the type of the
6307 /// corresponding operand.
6308 const char *X86TargetLowering::
6309 LowerXConstraint(MVT::ValueType ConstraintVT) const {
6310 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
6311 // 'f' like normal targets.
6312 if (MVT::isFloatingPoint(ConstraintVT)) {
6313 if (Subtarget->hasSSE2())
6315 if (Subtarget->hasSSE1())
6319 return TargetLowering::LowerXConstraint(ConstraintVT);
6322 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6323 /// vector. If it is invalid, don't add anything to Ops.
6324 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
6326 std::vector<SDOperand>&Ops,
6327 SelectionDAG &DAG) const {
6328 SDOperand Result(0, 0);
6330 switch (Constraint) {
6333 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6334 if (C->getValue() <= 31) {
6335 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6341 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6342 if (C->getValue() <= 255) {
6343 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6349 // Literal immediates are always ok.
6350 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
6351 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
6355 // If we are in non-pic codegen mode, we allow the address of a global (with
6356 // an optional displacement) to be used with 'i'.
6357 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
6360 // Match either (GA) or (GA+C)
6362 Offset = GA->getOffset();
6363 } else if (Op.getOpcode() == ISD::ADD) {
6364 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6365 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6367 Offset = GA->getOffset()+C->getValue();
6369 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6370 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6372 Offset = GA->getOffset()+C->getValue();
6379 // If addressing this global requires a load (e.g. in PIC mode), we can't
6381 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
6385 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
6391 // Otherwise, not valid for this mode.
6397 Ops.push_back(Result);
6400 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6403 std::vector<unsigned> X86TargetLowering::
6404 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6405 MVT::ValueType VT) const {
6406 if (Constraint.size() == 1) {
6407 // FIXME: not handling fp-stack yet!
6408 switch (Constraint[0]) { // GCC X86 Constraint Letters
6409 default: break; // Unknown constraint letter
6410 case 'A': // EAX/EDX
6411 if (VT == MVT::i32 || VT == MVT::i64)
6412 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
6414 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
6417 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
6418 else if (VT == MVT::i16)
6419 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
6420 else if (VT == MVT::i8)
6421 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
6422 else if (VT == MVT::i64)
6423 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
6428 return std::vector<unsigned>();
6431 std::pair<unsigned, const TargetRegisterClass*>
6432 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6433 MVT::ValueType VT) const {
6434 // First, see if this is a constraint that directly corresponds to an LLVM
6436 if (Constraint.size() == 1) {
6437 // GCC Constraint Letters
6438 switch (Constraint[0]) {
6440 case 'r': // GENERAL_REGS
6441 case 'R': // LEGACY_REGS
6442 case 'l': // INDEX_REGS
6443 if (VT == MVT::i64 && Subtarget->is64Bit())
6444 return std::make_pair(0U, X86::GR64RegisterClass);
6446 return std::make_pair(0U, X86::GR32RegisterClass);
6447 else if (VT == MVT::i16)
6448 return std::make_pair(0U, X86::GR16RegisterClass);
6449 else if (VT == MVT::i8)
6450 return std::make_pair(0U, X86::GR8RegisterClass);
6452 case 'f': // FP Stack registers.
6453 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
6454 // value to the correct fpstack register class.
6455 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
6456 return std::make_pair(0U, X86::RFP32RegisterClass);
6457 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
6458 return std::make_pair(0U, X86::RFP64RegisterClass);
6459 return std::make_pair(0U, X86::RFP80RegisterClass);
6460 case 'y': // MMX_REGS if MMX allowed.
6461 if (!Subtarget->hasMMX()) break;
6462 return std::make_pair(0U, X86::VR64RegisterClass);
6464 case 'Y': // SSE_REGS if SSE2 allowed
6465 if (!Subtarget->hasSSE2()) break;
6467 case 'x': // SSE_REGS if SSE1 allowed
6468 if (!Subtarget->hasSSE1()) break;
6472 // Scalar SSE types.
6475 return std::make_pair(0U, X86::FR32RegisterClass);
6478 return std::make_pair(0U, X86::FR64RegisterClass);
6486 return std::make_pair(0U, X86::VR128RegisterClass);
6492 // Use the default implementation in TargetLowering to convert the register
6493 // constraint into a member of a register class.
6494 std::pair<unsigned, const TargetRegisterClass*> Res;
6495 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6497 // Not found as a standard register?
6498 if (Res.second == 0) {
6499 // GCC calls "st(0)" just plain "st".
6500 if (StringsEqualNoCase("{st}", Constraint)) {
6501 Res.first = X86::ST0;
6502 Res.second = X86::RFP80RegisterClass;
6508 // Otherwise, check to see if this is a register class of the wrong value
6509 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6510 // turn into {ax},{dx}.
6511 if (Res.second->hasType(VT))
6512 return Res; // Correct type already, nothing to do.
6514 // All of the single-register GCC register classes map their values onto
6515 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6516 // really want an 8-bit or 32-bit register, map to the appropriate register
6517 // class and return the appropriate register.
6518 if (Res.second != X86::GR16RegisterClass)
6521 if (VT == MVT::i8) {
6522 unsigned DestReg = 0;
6523 switch (Res.first) {
6525 case X86::AX: DestReg = X86::AL; break;
6526 case X86::DX: DestReg = X86::DL; break;
6527 case X86::CX: DestReg = X86::CL; break;
6528 case X86::BX: DestReg = X86::BL; break;
6531 Res.first = DestReg;
6532 Res.second = Res.second = X86::GR8RegisterClass;
6534 } else if (VT == MVT::i32) {
6535 unsigned DestReg = 0;
6536 switch (Res.first) {
6538 case X86::AX: DestReg = X86::EAX; break;
6539 case X86::DX: DestReg = X86::EDX; break;
6540 case X86::CX: DestReg = X86::ECX; break;
6541 case X86::BX: DestReg = X86::EBX; break;
6542 case X86::SI: DestReg = X86::ESI; break;
6543 case X86::DI: DestReg = X86::EDI; break;
6544 case X86::BP: DestReg = X86::EBP; break;
6545 case X86::SP: DestReg = X86::ESP; break;
6548 Res.first = DestReg;
6549 Res.second = Res.second = X86::GR32RegisterClass;
6551 } else if (VT == MVT::i64) {
6552 unsigned DestReg = 0;
6553 switch (Res.first) {
6555 case X86::AX: DestReg = X86::RAX; break;
6556 case X86::DX: DestReg = X86::RDX; break;
6557 case X86::CX: DestReg = X86::RCX; break;
6558 case X86::BX: DestReg = X86::RBX; break;
6559 case X86::SI: DestReg = X86::RSI; break;
6560 case X86::DI: DestReg = X86::RDI; break;
6561 case X86::BP: DestReg = X86::RBP; break;
6562 case X86::SP: DestReg = X86::RSP; break;
6565 Res.first = DestReg;
6566 Res.second = Res.second = X86::GR64RegisterClass;