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/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/StringExtras.h"
43 // Forward declarations.
44 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
46 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
47 : TargetLowering(TM) {
48 Subtarget = &TM.getSubtarget<X86Subtarget>();
49 X86ScalarSSEf64 = Subtarget->hasSSE2();
50 X86ScalarSSEf32 = Subtarget->hasSSE1();
51 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
55 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 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
265 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
266 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
267 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
268 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
269 if (Subtarget->is64Bit())
270 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
271 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
272 if (Subtarget->is64Bit()) {
273 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
274 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
275 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
276 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
278 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
279 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
280 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
281 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
282 if (Subtarget->is64Bit()) {
283 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
284 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
285 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
288 if (Subtarget->hasSSE1())
289 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
291 if (!Subtarget->hasSSE2())
292 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
294 // Expand certain atomics
295 setOperationAction(ISD::ATOMIC_CMP_SWAP_8 , MVT::i8, Custom);
296 setOperationAction(ISD::ATOMIC_CMP_SWAP_16, MVT::i16, Custom);
297 setOperationAction(ISD::ATOMIC_CMP_SWAP_32, MVT::i32, Custom);
298 setOperationAction(ISD::ATOMIC_CMP_SWAP_64, MVT::i64, Custom);
300 setOperationAction(ISD::ATOMIC_LOAD_SUB_8, MVT::i8, Expand);
301 setOperationAction(ISD::ATOMIC_LOAD_SUB_16, MVT::i16, Expand);
302 setOperationAction(ISD::ATOMIC_LOAD_SUB_32, MVT::i32, Expand);
303 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Expand);
305 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
306 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
307 // FIXME - use subtarget debug flags
308 if (!Subtarget->isTargetDarwin() &&
309 !Subtarget->isTargetELF() &&
310 !Subtarget->isTargetCygMing()) {
311 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
312 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
315 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
316 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
317 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
318 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
319 if (Subtarget->is64Bit()) {
320 setExceptionPointerRegister(X86::RAX);
321 setExceptionSelectorRegister(X86::RDX);
323 setExceptionPointerRegister(X86::EAX);
324 setExceptionSelectorRegister(X86::EDX);
326 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
327 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
329 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
331 setOperationAction(ISD::TRAP, MVT::Other, Legal);
333 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
334 setOperationAction(ISD::VASTART , MVT::Other, Custom);
335 setOperationAction(ISD::VAEND , MVT::Other, Expand);
336 if (Subtarget->is64Bit()) {
337 setOperationAction(ISD::VAARG , MVT::Other, Custom);
338 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
340 setOperationAction(ISD::VAARG , MVT::Other, Expand);
341 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
344 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
345 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
346 if (Subtarget->is64Bit())
347 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
348 if (Subtarget->isTargetCygMing())
349 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
351 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
353 if (X86ScalarSSEf64) {
354 // f32 and f64 use SSE.
355 // Set up the FP register classes.
356 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
357 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
359 // Use ANDPD to simulate FABS.
360 setOperationAction(ISD::FABS , MVT::f64, Custom);
361 setOperationAction(ISD::FABS , MVT::f32, Custom);
363 // Use XORP to simulate FNEG.
364 setOperationAction(ISD::FNEG , MVT::f64, Custom);
365 setOperationAction(ISD::FNEG , MVT::f32, Custom);
367 // Use ANDPD and ORPD to simulate FCOPYSIGN.
368 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
369 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
371 // We don't support sin/cos/fmod
372 setOperationAction(ISD::FSIN , MVT::f64, Expand);
373 setOperationAction(ISD::FCOS , MVT::f64, Expand);
374 setOperationAction(ISD::FSIN , MVT::f32, Expand);
375 setOperationAction(ISD::FCOS , MVT::f32, Expand);
377 // Expand FP immediates into loads from the stack, except for the special
379 addLegalFPImmediate(APFloat(+0.0)); // xorpd
380 addLegalFPImmediate(APFloat(+0.0f)); // xorps
382 // Floating truncations from f80 and extensions to f80 go through memory.
383 // If optimizing, we lie about this though and handle it in
384 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
386 setConvertAction(MVT::f32, MVT::f80, Expand);
387 setConvertAction(MVT::f64, MVT::f80, Expand);
388 setConvertAction(MVT::f80, MVT::f32, Expand);
389 setConvertAction(MVT::f80, MVT::f64, Expand);
391 } else if (X86ScalarSSEf32) {
392 // Use SSE for f32, x87 for f64.
393 // Set up the FP register classes.
394 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
395 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
397 // Use ANDPS to simulate FABS.
398 setOperationAction(ISD::FABS , MVT::f32, Custom);
400 // Use XORP to simulate FNEG.
401 setOperationAction(ISD::FNEG , MVT::f32, Custom);
403 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
405 // Use ANDPS and ORPS to simulate FCOPYSIGN.
406 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
407 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
409 // We don't support sin/cos/fmod
410 setOperationAction(ISD::FSIN , MVT::f32, Expand);
411 setOperationAction(ISD::FCOS , MVT::f32, Expand);
413 // Special cases we handle for FP constants.
414 addLegalFPImmediate(APFloat(+0.0f)); // xorps
415 addLegalFPImmediate(APFloat(+0.0)); // FLD0
416 addLegalFPImmediate(APFloat(+1.0)); // FLD1
417 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
418 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
420 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
421 // this though and handle it in InstructionSelectPreprocess so that
422 // dagcombine2 can hack on these.
424 setConvertAction(MVT::f32, MVT::f64, Expand);
425 setConvertAction(MVT::f32, MVT::f80, Expand);
426 setConvertAction(MVT::f80, MVT::f32, Expand);
427 setConvertAction(MVT::f64, MVT::f32, Expand);
428 // And x87->x87 truncations also.
429 setConvertAction(MVT::f80, MVT::f64, Expand);
433 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
434 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
437 // f32 and f64 in x87.
438 // Set up the FP register classes.
439 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
440 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
442 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
443 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
447 // Floating truncations go through memory. If optimizing, we lie about
448 // this though and handle it in InstructionSelectPreprocess so that
449 // dagcombine2 can hack on these.
451 setConvertAction(MVT::f80, MVT::f32, Expand);
452 setConvertAction(MVT::f64, MVT::f32, Expand);
453 setConvertAction(MVT::f80, MVT::f64, Expand);
457 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
458 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
460 addLegalFPImmediate(APFloat(+0.0)); // FLD0
461 addLegalFPImmediate(APFloat(+1.0)); // FLD1
462 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
463 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
464 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
465 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
466 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
467 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
470 // Long double always uses X87.
471 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
472 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
475 APFloat TmpFlt(+0.0);
476 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
477 addLegalFPImmediate(TmpFlt); // FLD0
479 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
480 APFloat TmpFlt2(+1.0);
481 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
482 addLegalFPImmediate(TmpFlt2); // FLD1
483 TmpFlt2.changeSign();
484 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
488 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
489 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
492 // Always use a library call for pow.
493 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
494 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
495 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
497 setOperationAction(ISD::FLOG, MVT::f32, Expand);
498 setOperationAction(ISD::FLOG, MVT::f64, Expand);
499 setOperationAction(ISD::FLOG, MVT::f80, Expand);
500 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
501 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
502 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
503 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
504 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
505 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
506 setOperationAction(ISD::FEXP, MVT::f32, Expand);
507 setOperationAction(ISD::FEXP, MVT::f64, Expand);
508 setOperationAction(ISD::FEXP, MVT::f80, Expand);
509 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
510 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
511 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
513 // First set operation action for all vector types to expand. Then we
514 // will selectively turn on ones that can be effectively codegen'd.
515 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
516 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
517 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
532 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
557 if (Subtarget->hasMMX()) {
558 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
559 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
560 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
561 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
562 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
564 // FIXME: add MMX packed arithmetics
566 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
567 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
568 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
569 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
571 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
572 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
573 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
574 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
576 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
577 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
579 setOperationAction(ISD::AND, MVT::v8i8, Promote);
580 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
581 setOperationAction(ISD::AND, MVT::v4i16, Promote);
582 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
583 setOperationAction(ISD::AND, MVT::v2i32, Promote);
584 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
585 setOperationAction(ISD::AND, MVT::v1i64, Legal);
587 setOperationAction(ISD::OR, MVT::v8i8, Promote);
588 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
589 setOperationAction(ISD::OR, MVT::v4i16, Promote);
590 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
591 setOperationAction(ISD::OR, MVT::v2i32, Promote);
592 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
593 setOperationAction(ISD::OR, MVT::v1i64, Legal);
595 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
596 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
597 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
598 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
599 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
600 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
601 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
603 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
604 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
605 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
606 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
607 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
608 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
609 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
610 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
611 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
613 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
614 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
615 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
616 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
617 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
619 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
620 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
621 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
622 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
624 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
625 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
626 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
627 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
629 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
632 if (Subtarget->hasSSE1()) {
633 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
635 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
636 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
637 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
638 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
639 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
640 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
641 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
642 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
645 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
646 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
649 if (Subtarget->hasSSE2()) {
650 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
651 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
652 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
653 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
654 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
656 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
657 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
658 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
659 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
660 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
661 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
662 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
663 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
664 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
665 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
666 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
667 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
668 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
669 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
670 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
672 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
673 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
674 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
675 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
679 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
681 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
683 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
684 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
685 MVT VT = (MVT::SimpleValueType)i;
686 // Do not attempt to custom lower non-power-of-2 vectors
687 if (!isPowerOf2_32(VT.getVectorNumElements()))
689 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
690 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
691 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
693 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
694 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
695 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
696 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
697 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
698 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
699 if (Subtarget->is64Bit()) {
700 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
701 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
704 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
705 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
706 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
707 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
708 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
709 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
710 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
711 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
712 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
713 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
714 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
715 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
718 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
720 // Custom lower v2i64 and v2f64 selects.
721 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
722 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
723 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
724 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
728 if (Subtarget->hasSSE41()) {
729 // FIXME: Do we need to handle scalar-to-vector here?
730 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
731 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
733 // i8 and i16 vectors are custom , because the source register and source
734 // source memory operand types are not the same width. f32 vectors are
735 // custom since the immediate controlling the insert encodes additional
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
740 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
742 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
744 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
745 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
747 if (Subtarget->is64Bit()) {
748 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
749 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
753 if (Subtarget->hasSSE42()) {
754 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
757 // We want to custom lower some of our intrinsics.
758 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
760 // We have target-specific dag combine patterns for the following nodes:
761 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
762 setTargetDAGCombine(ISD::BUILD_VECTOR);
763 setTargetDAGCombine(ISD::SELECT);
764 setTargetDAGCombine(ISD::STORE);
766 computeRegisterProperties();
768 // FIXME: These should be based on subtarget info. Plus, the values should
769 // be smaller when we are in optimizing for size mode.
770 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
771 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
772 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
773 allowUnalignedMemoryAccesses = true; // x86 supports it!
774 setPrefLoopAlignment(16);
778 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
783 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
784 /// the desired ByVal argument alignment.
785 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
788 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
789 if (VTy->getBitWidth() == 128)
791 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
792 unsigned EltAlign = 0;
793 getMaxByValAlign(ATy->getElementType(), EltAlign);
794 if (EltAlign > MaxAlign)
796 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
797 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
798 unsigned EltAlign = 0;
799 getMaxByValAlign(STy->getElementType(i), EltAlign);
800 if (EltAlign > MaxAlign)
809 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
810 /// function arguments in the caller parameter area. For X86, aggregates
811 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
812 /// are at 4-byte boundaries.
813 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
814 if (Subtarget->is64Bit()) {
815 // Max of 8 and alignment of type.
816 unsigned TyAlign = TD->getABITypeAlignment(Ty);
823 if (Subtarget->hasSSE1())
824 getMaxByValAlign(Ty, Align);
828 /// getOptimalMemOpType - Returns the target specific optimal type for load
829 /// and store operations as a result of memset, memcpy, and memmove
830 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
833 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
834 bool isSrcConst, bool isSrcStr) const {
835 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
837 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
839 if (Subtarget->is64Bit() && Size >= 8)
845 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
847 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
848 SelectionDAG &DAG) const {
849 if (usesGlobalOffsetTable())
850 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
851 if (!Subtarget->isPICStyleRIPRel())
852 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
856 //===----------------------------------------------------------------------===//
857 // Return Value Calling Convention Implementation
858 //===----------------------------------------------------------------------===//
860 #include "X86GenCallingConv.inc"
862 /// LowerRET - Lower an ISD::RET node.
863 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
864 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
866 SmallVector<CCValAssign, 16> RVLocs;
867 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
868 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
869 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
870 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
872 // If this is the first return lowered for this function, add the regs to the
873 // liveout set for the function.
874 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
875 for (unsigned i = 0; i != RVLocs.size(); ++i)
876 if (RVLocs[i].isRegLoc())
877 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
879 SDValue Chain = Op.getOperand(0);
881 // Handle tail call return.
882 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
883 if (Chain.getOpcode() == X86ISD::TAILCALL) {
884 SDValue TailCall = Chain;
885 SDValue TargetAddress = TailCall.getOperand(1);
886 SDValue StackAdjustment = TailCall.getOperand(2);
887 assert(((TargetAddress.getOpcode() == ISD::Register &&
888 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
889 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
890 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
891 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
892 "Expecting an global address, external symbol, or register");
893 assert(StackAdjustment.getOpcode() == ISD::Constant &&
894 "Expecting a const value");
896 SmallVector<SDValue,8> Operands;
897 Operands.push_back(Chain.getOperand(0));
898 Operands.push_back(TargetAddress);
899 Operands.push_back(StackAdjustment);
900 // Copy registers used by the call. Last operand is a flag so it is not
902 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
903 Operands.push_back(Chain.getOperand(i));
905 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
912 SmallVector<SDValue, 6> RetOps;
913 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
914 // Operand #1 = Bytes To Pop
915 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
917 // Copy the result values into the output registers.
918 for (unsigned i = 0; i != RVLocs.size(); ++i) {
919 CCValAssign &VA = RVLocs[i];
920 assert(VA.isRegLoc() && "Can only return in registers!");
921 SDValue ValToCopy = Op.getOperand(i*2+1);
923 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
924 // the RET instruction and handled by the FP Stackifier.
925 if (RVLocs[i].getLocReg() == X86::ST0 ||
926 RVLocs[i].getLocReg() == X86::ST1) {
927 // If this is a copy from an xmm register to ST(0), use an FPExtend to
928 // change the value to the FP stack register class.
929 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
930 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
931 RetOps.push_back(ValToCopy);
932 // Don't emit a copytoreg.
936 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
937 Flag = Chain.getValue(1);
940 // The x86-64 ABI for returning structs by value requires that we copy
941 // the sret argument into %rax for the return. We saved the argument into
942 // a virtual register in the entry block, so now we copy the value out
944 if (Subtarget->is64Bit() &&
945 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
946 MachineFunction &MF = DAG.getMachineFunction();
947 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
948 unsigned Reg = FuncInfo->getSRetReturnReg();
950 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
951 FuncInfo->setSRetReturnReg(Reg);
953 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
955 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
956 Flag = Chain.getValue(1);
959 RetOps[0] = Chain; // Update chain.
961 // Add the flag if we have it.
963 RetOps.push_back(Flag);
965 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
969 /// LowerCallResult - Lower the result values of an ISD::CALL into the
970 /// appropriate copies out of appropriate physical registers. This assumes that
971 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
972 /// being lowered. The returns a SDNode with the same number of values as the
974 SDNode *X86TargetLowering::
975 LowerCallResult(SDValue Chain, SDValue InFlag, SDNode *TheCall,
976 unsigned CallingConv, SelectionDAG &DAG) {
978 // Assign locations to each value returned by this call.
979 SmallVector<CCValAssign, 16> RVLocs;
980 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
981 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
982 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
984 SmallVector<SDValue, 8> ResultVals;
986 // Copy all of the result registers out of their specified physreg.
987 for (unsigned i = 0; i != RVLocs.size(); ++i) {
988 MVT CopyVT = RVLocs[i].getValVT();
990 // If this is a call to a function that returns an fp value on the floating
991 // point stack, but where we prefer to use the value in xmm registers, copy
992 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
993 if ((RVLocs[i].getLocReg() == X86::ST0 ||
994 RVLocs[i].getLocReg() == X86::ST1) &&
995 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
999 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1000 CopyVT, InFlag).getValue(1);
1001 SDValue Val = Chain.getValue(0);
1002 InFlag = Chain.getValue(2);
1004 if (CopyVT != RVLocs[i].getValVT()) {
1005 // Round the F80 the right size, which also moves to the appropriate xmm
1007 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1008 // This truncation won't change the value.
1009 DAG.getIntPtrConstant(1));
1012 ResultVals.push_back(Val);
1015 // Merge everything together with a MERGE_VALUES node.
1016 ResultVals.push_back(Chain);
1017 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1018 ResultVals.size()).getNode();
1022 //===----------------------------------------------------------------------===//
1023 // C & StdCall & Fast Calling Convention implementation
1024 //===----------------------------------------------------------------------===//
1025 // StdCall calling convention seems to be standard for many Windows' API
1026 // routines and around. It differs from C calling convention just a little:
1027 // callee should clean up the stack, not caller. Symbols should be also
1028 // decorated in some fancy way :) It doesn't support any vector arguments.
1029 // For info on fast calling convention see Fast Calling Convention (tail call)
1030 // implementation LowerX86_32FastCCCallTo.
1032 /// AddLiveIn - This helper function adds the specified physical register to the
1033 /// MachineFunction as a live in value. It also creates a corresponding virtual
1034 /// register for it.
1035 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1036 const TargetRegisterClass *RC) {
1037 assert(RC->contains(PReg) && "Not the correct regclass!");
1038 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1039 MF.getRegInfo().addLiveIn(PReg, VReg);
1043 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1045 static bool CallIsStructReturn(SDValue Op) {
1046 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1050 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
1053 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1054 /// return semantics.
1055 static bool ArgsAreStructReturn(SDValue Op) {
1056 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1060 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1063 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1064 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1066 bool X86TargetLowering::IsCalleePop(SDValue Op) {
1067 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1071 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1074 case CallingConv::X86_StdCall:
1075 return !Subtarget->is64Bit();
1076 case CallingConv::X86_FastCall:
1077 return !Subtarget->is64Bit();
1078 case CallingConv::Fast:
1079 return PerformTailCallOpt;
1083 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1084 /// FORMAL_ARGUMENTS node.
1085 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
1086 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1088 if (Subtarget->is64Bit()) {
1089 if (Subtarget->isTargetWin64())
1090 return CC_X86_Win64_C;
1091 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1092 return CC_X86_64_TailCall;
1097 if (CC == CallingConv::X86_FastCall)
1098 return CC_X86_32_FastCall;
1099 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1100 return CC_X86_32_TailCall;
1101 else if (CC == CallingConv::Fast)
1102 return CC_X86_32_FastCC;
1107 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1108 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1110 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1111 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1112 if (CC == CallingConv::X86_FastCall)
1114 else if (CC == CallingConv::X86_StdCall)
1120 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1121 /// in a register before calling.
1122 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1123 return !IsTailCall && !Is64Bit &&
1124 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1125 Subtarget->isPICStyleGOT();
1128 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1129 /// address to be loaded in a register.
1131 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1132 return !Is64Bit && IsTailCall &&
1133 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1134 Subtarget->isPICStyleGOT();
1137 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1138 /// by "Src" to address "Dst" with size and alignment information specified by
1139 /// the specific parameter attribute. The copy will be passed as a byval
1140 /// function parameter.
1142 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1143 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1144 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1145 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1146 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1149 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1150 const CCValAssign &VA,
1151 MachineFrameInfo *MFI,
1153 SDValue Root, unsigned i) {
1154 // Create the nodes corresponding to a load from this parameter slot.
1155 ISD::ArgFlagsTy Flags =
1156 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1157 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1158 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1160 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1161 // changed with more analysis.
1162 // In case of tail call optimization mark all arguments mutable. Since they
1163 // could be overwritten by lowering of arguments in case of a tail call.
1164 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1165 VA.getLocMemOffset(), isImmutable);
1166 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1167 if (Flags.isByVal())
1169 return DAG.getLoad(VA.getValVT(), Root, FIN,
1170 PseudoSourceValue::getFixedStack(FI), 0);
1174 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1175 MachineFunction &MF = DAG.getMachineFunction();
1176 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1178 const Function* Fn = MF.getFunction();
1179 if (Fn->hasExternalLinkage() &&
1180 Subtarget->isTargetCygMing() &&
1181 Fn->getName() == "main")
1182 FuncInfo->setForceFramePointer(true);
1184 // Decorate the function name.
1185 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1187 MachineFrameInfo *MFI = MF.getFrameInfo();
1188 SDValue Root = Op.getOperand(0);
1189 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1190 unsigned CC = MF.getFunction()->getCallingConv();
1191 bool Is64Bit = Subtarget->is64Bit();
1192 bool IsWin64 = Subtarget->isTargetWin64();
1194 assert(!(isVarArg && CC == CallingConv::Fast) &&
1195 "Var args not supported with calling convention fastcc");
1197 // Assign locations to all of the incoming arguments.
1198 SmallVector<CCValAssign, 16> ArgLocs;
1199 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1200 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(Op));
1202 SmallVector<SDValue, 8> ArgValues;
1203 unsigned LastVal = ~0U;
1204 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1205 CCValAssign &VA = ArgLocs[i];
1206 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1208 assert(VA.getValNo() != LastVal &&
1209 "Don't support value assigned to multiple locs yet");
1210 LastVal = VA.getValNo();
1212 if (VA.isRegLoc()) {
1213 MVT RegVT = VA.getLocVT();
1214 TargetRegisterClass *RC;
1215 if (RegVT == MVT::i32)
1216 RC = X86::GR32RegisterClass;
1217 else if (Is64Bit && RegVT == MVT::i64)
1218 RC = X86::GR64RegisterClass;
1219 else if (RegVT == MVT::f32)
1220 RC = X86::FR32RegisterClass;
1221 else if (RegVT == MVT::f64)
1222 RC = X86::FR64RegisterClass;
1223 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1224 RC = X86::VR128RegisterClass;
1225 else if (RegVT.isVector()) {
1226 assert(RegVT.getSizeInBits() == 64);
1228 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1230 // Darwin calling convention passes MMX values in either GPRs or
1231 // XMMs in x86-64. Other targets pass them in memory.
1232 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1233 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1236 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1241 assert(0 && "Unknown argument type!");
1244 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1245 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1247 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1248 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1250 if (VA.getLocInfo() == CCValAssign::SExt)
1251 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1252 DAG.getValueType(VA.getValVT()));
1253 else if (VA.getLocInfo() == CCValAssign::ZExt)
1254 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1255 DAG.getValueType(VA.getValVT()));
1257 if (VA.getLocInfo() != CCValAssign::Full)
1258 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1260 // Handle MMX values passed in GPRs.
1261 if (Is64Bit && RegVT != VA.getLocVT()) {
1262 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1263 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1264 else if (RC == X86::VR128RegisterClass) {
1265 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1266 DAG.getConstant(0, MVT::i64));
1267 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1271 ArgValues.push_back(ArgValue);
1273 assert(VA.isMemLoc());
1274 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1278 // The x86-64 ABI for returning structs by value requires that we copy
1279 // the sret argument into %rax for the return. Save the argument into
1280 // a virtual register so that we can access it from the return points.
1281 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1282 MachineFunction &MF = DAG.getMachineFunction();
1283 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1284 unsigned Reg = FuncInfo->getSRetReturnReg();
1286 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1287 FuncInfo->setSRetReturnReg(Reg);
1289 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1290 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1293 unsigned StackSize = CCInfo.getNextStackOffset();
1294 // align stack specially for tail calls
1295 if (PerformTailCallOpt && CC == CallingConv::Fast)
1296 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1298 // If the function takes variable number of arguments, make a frame index for
1299 // the start of the first vararg value... for expansion of llvm.va_start.
1301 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1302 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1305 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1307 // FIXME: We should really autogenerate these arrays
1308 static const unsigned GPR64ArgRegsWin64[] = {
1309 X86::RCX, X86::RDX, X86::R8, X86::R9
1311 static const unsigned XMMArgRegsWin64[] = {
1312 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1314 static const unsigned GPR64ArgRegs64Bit[] = {
1315 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1317 static const unsigned XMMArgRegs64Bit[] = {
1318 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1319 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1321 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1324 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1325 GPR64ArgRegs = GPR64ArgRegsWin64;
1326 XMMArgRegs = XMMArgRegsWin64;
1328 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1329 GPR64ArgRegs = GPR64ArgRegs64Bit;
1330 XMMArgRegs = XMMArgRegs64Bit;
1332 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1334 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1337 // For X86-64, if there are vararg parameters that are passed via
1338 // registers, then we must store them to their spots on the stack so they
1339 // may be loaded by deferencing the result of va_next.
1340 VarArgsGPOffset = NumIntRegs * 8;
1341 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1342 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1343 TotalNumXMMRegs * 16, 16);
1345 // Store the integer parameter registers.
1346 SmallVector<SDValue, 8> MemOps;
1347 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1348 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1349 DAG.getIntPtrConstant(VarArgsGPOffset));
1350 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1351 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1352 X86::GR64RegisterClass);
1353 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1355 DAG.getStore(Val.getValue(1), Val, FIN,
1356 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1357 MemOps.push_back(Store);
1358 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1359 DAG.getIntPtrConstant(8));
1362 // Now store the XMM (fp + vector) parameter registers.
1363 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1364 DAG.getIntPtrConstant(VarArgsFPOffset));
1365 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1366 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1367 X86::VR128RegisterClass);
1368 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1370 DAG.getStore(Val.getValue(1), Val, FIN,
1371 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1372 MemOps.push_back(Store);
1373 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1374 DAG.getIntPtrConstant(16));
1376 if (!MemOps.empty())
1377 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1378 &MemOps[0], MemOps.size());
1382 ArgValues.push_back(Root);
1384 // Some CCs need callee pop.
1385 if (IsCalleePop(Op)) {
1386 BytesToPopOnReturn = StackSize; // Callee pops everything.
1387 BytesCallerReserves = 0;
1389 BytesToPopOnReturn = 0; // Callee pops nothing.
1390 // If this is an sret function, the return should pop the hidden pointer.
1391 if (!Is64Bit && ArgsAreStructReturn(Op))
1392 BytesToPopOnReturn = 4;
1393 BytesCallerReserves = StackSize;
1397 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1398 if (CC == CallingConv::X86_FastCall)
1399 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1402 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1404 // Return the new list of results.
1405 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1406 ArgValues.size()).getValue(Op.getResNo());
1410 X86TargetLowering::LowerMemOpCallTo(SDValue Op, SelectionDAG &DAG,
1411 const SDValue &StackPtr,
1412 const CCValAssign &VA,
1415 unsigned LocMemOffset = VA.getLocMemOffset();
1416 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1417 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1418 ISD::ArgFlagsTy Flags =
1419 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1420 if (Flags.isByVal()) {
1421 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1423 return DAG.getStore(Chain, Arg, PtrOff,
1424 PseudoSourceValue::getStack(), LocMemOffset);
1427 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1428 /// optimization is performed and it is required.
1430 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1431 SDValue &OutRetAddr,
1436 if (!IsTailCall || FPDiff==0) return Chain;
1438 // Adjust the Return address stack slot.
1439 MVT VT = getPointerTy();
1440 OutRetAddr = getReturnAddressFrameIndex(DAG);
1441 // Load the "old" Return address.
1442 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1443 return SDValue(OutRetAddr.getNode(), 1);
1446 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1447 /// optimization is performed and it is required (FPDiff!=0).
1449 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1450 SDValue Chain, SDValue RetAddrFrIdx,
1451 bool Is64Bit, int FPDiff) {
1452 // Store the return address to the appropriate stack slot.
1453 if (!FPDiff) return Chain;
1454 // Calculate the new stack slot for the return address.
1455 int SlotSize = Is64Bit ? 8 : 4;
1456 int NewReturnAddrFI =
1457 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1458 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1459 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1460 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1461 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1465 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1466 MachineFunction &MF = DAG.getMachineFunction();
1467 SDValue Chain = Op.getOperand(0);
1468 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1469 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1470 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1471 && CC == CallingConv::Fast && PerformTailCallOpt;
1472 SDValue Callee = Op.getOperand(4);
1473 bool Is64Bit = Subtarget->is64Bit();
1474 bool IsStructRet = CallIsStructReturn(Op);
1476 assert(!(isVarArg && CC == CallingConv::Fast) &&
1477 "Var args not supported with calling convention fastcc");
1479 // Analyze operands of the call, assigning locations to each operand.
1480 SmallVector<CCValAssign, 16> ArgLocs;
1481 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1482 CCInfo.AnalyzeCallOperands(Op.getNode(), CCAssignFnForNode(Op));
1484 // Get a count of how many bytes are to be pushed on the stack.
1485 unsigned NumBytes = CCInfo.getNextStackOffset();
1487 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1491 // Lower arguments at fp - stackoffset + fpdiff.
1492 unsigned NumBytesCallerPushed =
1493 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1494 FPDiff = NumBytesCallerPushed - NumBytes;
1496 // Set the delta of movement of the returnaddr stackslot.
1497 // But only set if delta is greater than previous delta.
1498 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1499 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1502 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1504 SDValue RetAddrFrIdx;
1505 // Load return adress for tail calls.
1506 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1509 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1510 SmallVector<SDValue, 8> MemOpChains;
1513 // Walk the register/memloc assignments, inserting copies/loads. In the case
1514 // of tail call optimization arguments are handle later.
1515 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1516 CCValAssign &VA = ArgLocs[i];
1517 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1518 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1519 getArgFlags().isByVal();
1521 // Promote the value if needed.
1522 switch (VA.getLocInfo()) {
1523 default: assert(0 && "Unknown loc info!");
1524 case CCValAssign::Full: break;
1525 case CCValAssign::SExt:
1526 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1528 case CCValAssign::ZExt:
1529 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1531 case CCValAssign::AExt:
1532 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1536 if (VA.isRegLoc()) {
1538 MVT RegVT = VA.getLocVT();
1539 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1540 switch (VA.getLocReg()) {
1543 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1545 // Special case: passing MMX values in GPR registers.
1546 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1549 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1550 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1551 // Special case: passing MMX values in XMM registers.
1552 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1553 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1554 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1555 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1556 getMOVLMask(2, DAG));
1561 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1563 if (!IsTailCall || (IsTailCall && isByVal)) {
1564 assert(VA.isMemLoc());
1565 if (StackPtr.getNode() == 0)
1566 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1568 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1574 if (!MemOpChains.empty())
1575 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1576 &MemOpChains[0], MemOpChains.size());
1578 // Build a sequence of copy-to-reg nodes chained together with token chain
1579 // and flag operands which copy the outgoing args into registers.
1581 // Tail call byval lowering might overwrite argument registers so in case of
1582 // tail call optimization the copies to registers are lowered later.
1584 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1585 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1587 InFlag = Chain.getValue(1);
1590 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1592 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1593 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1594 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1596 InFlag = Chain.getValue(1);
1598 // If we are tail calling and generating PIC/GOT style code load the address
1599 // of the callee into ecx. The value in ecx is used as target of the tail
1600 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1601 // calls on PIC/GOT architectures. Normally we would just put the address of
1602 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1603 // restored (since ebx is callee saved) before jumping to the target@PLT.
1604 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1605 // Note: The actual moving to ecx is done further down.
1606 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1607 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1608 !G->getGlobal()->hasProtectedVisibility())
1609 Callee = LowerGlobalAddress(Callee, DAG);
1610 else if (isa<ExternalSymbolSDNode>(Callee))
1611 Callee = LowerExternalSymbol(Callee,DAG);
1614 if (Is64Bit && isVarArg) {
1615 // From AMD64 ABI document:
1616 // For calls that may call functions that use varargs or stdargs
1617 // (prototype-less calls or calls to functions containing ellipsis (...) in
1618 // the declaration) %al is used as hidden argument to specify the number
1619 // of SSE registers used. The contents of %al do not need to match exactly
1620 // the number of registers, but must be an ubound on the number of SSE
1621 // registers used and is in the range 0 - 8 inclusive.
1623 // FIXME: Verify this on Win64
1624 // Count the number of XMM registers allocated.
1625 static const unsigned XMMArgRegs[] = {
1626 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1627 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1629 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1631 Chain = DAG.getCopyToReg(Chain, X86::AL,
1632 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1633 InFlag = Chain.getValue(1);
1637 // For tail calls lower the arguments to the 'real' stack slot.
1639 SmallVector<SDValue, 8> MemOpChains2;
1642 // Do not flag preceeding copytoreg stuff together with the following stuff.
1644 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1645 CCValAssign &VA = ArgLocs[i];
1646 if (!VA.isRegLoc()) {
1647 assert(VA.isMemLoc());
1648 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1649 SDValue FlagsOp = Op.getOperand(6+2*VA.getValNo());
1650 ISD::ArgFlagsTy Flags =
1651 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1652 // Create frame index.
1653 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1654 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1655 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1656 FIN = DAG.getFrameIndex(FI, getPointerTy());
1658 if (Flags.isByVal()) {
1659 // Copy relative to framepointer.
1660 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1661 if (StackPtr.getNode() == 0)
1662 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1663 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1665 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1668 // Store relative to framepointer.
1669 MemOpChains2.push_back(
1670 DAG.getStore(Chain, Arg, FIN,
1671 PseudoSourceValue::getFixedStack(FI), 0));
1676 if (!MemOpChains2.empty())
1677 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1678 &MemOpChains2[0], MemOpChains2.size());
1680 // Copy arguments to their registers.
1681 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1682 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1684 InFlag = Chain.getValue(1);
1688 // Store the return address to the appropriate stack slot.
1689 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1693 // If the callee is a GlobalAddress node (quite common, every direct call is)
1694 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1695 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1696 // We should use extra load for direct calls to dllimported functions in
1698 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1699 getTargetMachine(), true))
1700 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1701 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1702 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1703 } else if (IsTailCall) {
1704 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1706 Chain = DAG.getCopyToReg(Chain,
1707 DAG.getRegister(Opc, getPointerTy()),
1709 Callee = DAG.getRegister(Opc, getPointerTy());
1710 // Add register as live out.
1711 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1714 // Returns a chain & a flag for retval copy to use.
1715 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1716 SmallVector<SDValue, 8> Ops;
1719 Ops.push_back(Chain);
1720 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1721 Ops.push_back(DAG.getIntPtrConstant(0));
1722 if (InFlag.getNode())
1723 Ops.push_back(InFlag);
1724 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1725 InFlag = Chain.getValue(1);
1727 // Returns a chain & a flag for retval copy to use.
1728 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1732 Ops.push_back(Chain);
1733 Ops.push_back(Callee);
1736 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1738 // Add argument registers to the end of the list so that they are known live
1740 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1741 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1742 RegsToPass[i].second.getValueType()));
1744 // Add an implicit use GOT pointer in EBX.
1745 if (!IsTailCall && !Is64Bit &&
1746 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1747 Subtarget->isPICStyleGOT())
1748 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1750 // Add an implicit use of AL for x86 vararg functions.
1751 if (Is64Bit && isVarArg)
1752 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1754 if (InFlag.getNode())
1755 Ops.push_back(InFlag);
1758 assert(InFlag.getNode() &&
1759 "Flag must be set. Depend on flag being set in LowerRET");
1760 Chain = DAG.getNode(X86ISD::TAILCALL,
1761 Op.getNode()->getVTList(), &Ops[0], Ops.size());
1763 return SDValue(Chain.getNode(), Op.getResNo());
1766 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1767 InFlag = Chain.getValue(1);
1769 // Create the CALLSEQ_END node.
1770 unsigned NumBytesForCalleeToPush;
1771 if (IsCalleePop(Op))
1772 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1773 else if (!Is64Bit && IsStructRet)
1774 // If this is is a call to a struct-return function, the callee
1775 // pops the hidden struct pointer, so we have to push it back.
1776 // This is common for Darwin/X86, Linux & Mingw32 targets.
1777 NumBytesForCalleeToPush = 4;
1779 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1781 // Returns a flag for retval copy to use.
1782 Chain = DAG.getCALLSEQ_END(Chain,
1783 DAG.getIntPtrConstant(NumBytes),
1784 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1786 InFlag = Chain.getValue(1);
1788 // Handle result values, copying them out of physregs into vregs that we
1790 return SDValue(LowerCallResult(Chain, InFlag, Op.getNode(), CC, DAG),
1795 //===----------------------------------------------------------------------===//
1796 // Fast Calling Convention (tail call) implementation
1797 //===----------------------------------------------------------------------===//
1799 // Like std call, callee cleans arguments, convention except that ECX is
1800 // reserved for storing the tail called function address. Only 2 registers are
1801 // free for argument passing (inreg). Tail call optimization is performed
1803 // * tailcallopt is enabled
1804 // * caller/callee are fastcc
1805 // On X86_64 architecture with GOT-style position independent code only local
1806 // (within module) calls are supported at the moment.
1807 // To keep the stack aligned according to platform abi the function
1808 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1809 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1810 // If a tail called function callee has more arguments than the caller the
1811 // caller needs to make sure that there is room to move the RETADDR to. This is
1812 // achieved by reserving an area the size of the argument delta right after the
1813 // original REtADDR, but before the saved framepointer or the spilled registers
1814 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1826 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1827 /// for a 16 byte align requirement.
1828 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1829 SelectionDAG& DAG) {
1830 MachineFunction &MF = DAG.getMachineFunction();
1831 const TargetMachine &TM = MF.getTarget();
1832 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1833 unsigned StackAlignment = TFI.getStackAlignment();
1834 uint64_t AlignMask = StackAlignment - 1;
1835 int64_t Offset = StackSize;
1836 uint64_t SlotSize = TD->getPointerSize();
1837 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1838 // Number smaller than 12 so just add the difference.
1839 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1841 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1842 Offset = ((~AlignMask) & Offset) + StackAlignment +
1843 (StackAlignment-SlotSize);
1848 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1849 /// following the call is a return. A function is eligible if caller/callee
1850 /// calling conventions match, currently only fastcc supports tail calls, and
1851 /// the function CALL is immediatly followed by a RET.
1852 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1854 SelectionDAG& DAG) const {
1855 if (!PerformTailCallOpt)
1858 if (CheckTailCallReturnConstraints(Call, Ret)) {
1859 MachineFunction &MF = DAG.getMachineFunction();
1860 unsigned CallerCC = MF.getFunction()->getCallingConv();
1861 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1862 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1863 SDValue Callee = Call.getOperand(4);
1864 // On x86/32Bit PIC/GOT tail calls are supported.
1865 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1866 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1869 // Can only do local tail calls (in same module, hidden or protected) on
1870 // x86_64 PIC/GOT at the moment.
1871 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1872 return G->getGlobal()->hasHiddenVisibility()
1873 || G->getGlobal()->hasProtectedVisibility();
1881 X86TargetLowering::createFastISel(MachineFunction &mf,
1882 DenseMap<const Value *, unsigned> &vm,
1883 DenseMap<const BasicBlock *,
1884 MachineBasicBlock *> &bm) {
1885 return X86::createFastISel(mf, vm, bm);
1889 //===----------------------------------------------------------------------===//
1890 // Other Lowering Hooks
1891 //===----------------------------------------------------------------------===//
1894 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1895 MachineFunction &MF = DAG.getMachineFunction();
1896 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1897 int ReturnAddrIndex = FuncInfo->getRAIndex();
1898 uint64_t SlotSize = TD->getPointerSize();
1900 if (ReturnAddrIndex == 0) {
1901 // Set up a frame object for the return address.
1902 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1903 FuncInfo->setRAIndex(ReturnAddrIndex);
1906 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1910 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1911 /// specific condition code. It returns a false if it cannot do a direct
1912 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1914 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1915 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1916 SelectionDAG &DAG) {
1917 X86CC = X86::COND_INVALID;
1919 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1920 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1921 // X > -1 -> X == 0, jump !sign.
1922 RHS = DAG.getConstant(0, RHS.getValueType());
1923 X86CC = X86::COND_NS;
1925 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1926 // X < 0 -> X == 0, jump on sign.
1927 X86CC = X86::COND_S;
1929 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1931 RHS = DAG.getConstant(0, RHS.getValueType());
1932 X86CC = X86::COND_LE;
1937 switch (SetCCOpcode) {
1939 case ISD::SETEQ: X86CC = X86::COND_E; break;
1940 case ISD::SETGT: X86CC = X86::COND_G; break;
1941 case ISD::SETGE: X86CC = X86::COND_GE; break;
1942 case ISD::SETLT: X86CC = X86::COND_L; break;
1943 case ISD::SETLE: X86CC = X86::COND_LE; break;
1944 case ISD::SETNE: X86CC = X86::COND_NE; break;
1945 case ISD::SETULT: X86CC = X86::COND_B; break;
1946 case ISD::SETUGT: X86CC = X86::COND_A; break;
1947 case ISD::SETULE: X86CC = X86::COND_BE; break;
1948 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1951 // First determine if it requires or is profitable to flip the operands.
1953 switch (SetCCOpcode) {
1963 // If LHS is a foldable load, but RHS is not, flip the condition.
1965 (ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
1966 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
1967 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
1971 std::swap(LHS, RHS);
1973 // On a floating point condition, the flags are set as follows:
1975 // 0 | 0 | 0 | X > Y
1976 // 0 | 0 | 1 | X < Y
1977 // 1 | 0 | 0 | X == Y
1978 // 1 | 1 | 1 | unordered
1979 switch (SetCCOpcode) {
1983 X86CC = X86::COND_E;
1985 case ISD::SETOLT: // flipped
1988 X86CC = X86::COND_A;
1990 case ISD::SETOLE: // flipped
1993 X86CC = X86::COND_AE;
1995 case ISD::SETUGT: // flipped
1998 X86CC = X86::COND_B;
2000 case ISD::SETUGE: // flipped
2003 X86CC = X86::COND_BE;
2007 X86CC = X86::COND_NE;
2010 X86CC = X86::COND_P;
2013 X86CC = X86::COND_NP;
2018 return X86CC != X86::COND_INVALID;
2021 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2022 /// code. Current x86 isa includes the following FP cmov instructions:
2023 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2024 static bool hasFPCMov(unsigned X86CC) {
2040 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2041 /// true if Op is undef or if its value falls within the specified range (L, H].
2042 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2043 if (Op.getOpcode() == ISD::UNDEF)
2046 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2047 return (Val >= Low && Val < Hi);
2050 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2051 /// true if Op is undef or if its value equal to the specified value.
2052 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2053 if (Op.getOpcode() == ISD::UNDEF)
2055 return cast<ConstantSDNode>(Op)->getValue() == Val;
2058 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2059 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2060 bool X86::isPSHUFDMask(SDNode *N) {
2061 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2063 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2066 // Check if the value doesn't reference the second vector.
2067 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2068 SDValue Arg = N->getOperand(i);
2069 if (Arg.getOpcode() == ISD::UNDEF) continue;
2070 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2071 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2078 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2079 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2080 bool X86::isPSHUFHWMask(SDNode *N) {
2081 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2083 if (N->getNumOperands() != 8)
2086 // Lower quadword copied in order.
2087 for (unsigned i = 0; i != 4; ++i) {
2088 SDValue Arg = N->getOperand(i);
2089 if (Arg.getOpcode() == ISD::UNDEF) continue;
2090 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2091 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2095 // Upper quadword shuffled.
2096 for (unsigned i = 4; i != 8; ++i) {
2097 SDValue Arg = N->getOperand(i);
2098 if (Arg.getOpcode() == ISD::UNDEF) continue;
2099 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2100 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2101 if (Val < 4 || Val > 7)
2108 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2109 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2110 bool X86::isPSHUFLWMask(SDNode *N) {
2111 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2113 if (N->getNumOperands() != 8)
2116 // Upper quadword copied in order.
2117 for (unsigned i = 4; i != 8; ++i)
2118 if (!isUndefOrEqual(N->getOperand(i), i))
2121 // Lower quadword shuffled.
2122 for (unsigned i = 0; i != 4; ++i)
2123 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2129 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2130 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2131 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2132 if (NumElems != 2 && NumElems != 4) return false;
2134 unsigned Half = NumElems / 2;
2135 for (unsigned i = 0; i < Half; ++i)
2136 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2138 for (unsigned i = Half; i < NumElems; ++i)
2139 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2145 bool X86::isSHUFPMask(SDNode *N) {
2146 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2147 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2150 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2151 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2152 /// half elements to come from vector 1 (which would equal the dest.) and
2153 /// the upper half to come from vector 2.
2154 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2155 if (NumOps != 2 && NumOps != 4) return false;
2157 unsigned Half = NumOps / 2;
2158 for (unsigned i = 0; i < Half; ++i)
2159 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2161 for (unsigned i = Half; i < NumOps; ++i)
2162 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2167 static bool isCommutedSHUFP(SDNode *N) {
2168 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2169 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2172 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2173 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2174 bool X86::isMOVHLPSMask(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2177 if (N->getNumOperands() != 4)
2180 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2181 return isUndefOrEqual(N->getOperand(0), 6) &&
2182 isUndefOrEqual(N->getOperand(1), 7) &&
2183 isUndefOrEqual(N->getOperand(2), 2) &&
2184 isUndefOrEqual(N->getOperand(3), 3);
2187 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2188 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2190 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2191 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2193 if (N->getNumOperands() != 4)
2196 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2197 return isUndefOrEqual(N->getOperand(0), 2) &&
2198 isUndefOrEqual(N->getOperand(1), 3) &&
2199 isUndefOrEqual(N->getOperand(2), 2) &&
2200 isUndefOrEqual(N->getOperand(3), 3);
2203 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2204 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2205 bool X86::isMOVLPMask(SDNode *N) {
2206 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2208 unsigned NumElems = N->getNumOperands();
2209 if (NumElems != 2 && NumElems != 4)
2212 for (unsigned i = 0; i < NumElems/2; ++i)
2213 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2216 for (unsigned i = NumElems/2; i < NumElems; ++i)
2217 if (!isUndefOrEqual(N->getOperand(i), i))
2223 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2224 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2226 bool X86::isMOVHPMask(SDNode *N) {
2227 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2229 unsigned NumElems = N->getNumOperands();
2230 if (NumElems != 2 && NumElems != 4)
2233 for (unsigned i = 0; i < NumElems/2; ++i)
2234 if (!isUndefOrEqual(N->getOperand(i), i))
2237 for (unsigned i = 0; i < NumElems/2; ++i) {
2238 SDValue Arg = N->getOperand(i + NumElems/2);
2239 if (!isUndefOrEqual(Arg, i + NumElems))
2246 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2247 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2248 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2249 bool V2IsSplat = false) {
2250 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2253 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2254 SDValue BitI = Elts[i];
2255 SDValue BitI1 = Elts[i+1];
2256 if (!isUndefOrEqual(BitI, j))
2259 if (isUndefOrEqual(BitI1, NumElts))
2262 if (!isUndefOrEqual(BitI1, j + NumElts))
2270 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2271 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2272 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2275 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2276 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2277 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2278 bool V2IsSplat = false) {
2279 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2282 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2283 SDValue BitI = Elts[i];
2284 SDValue BitI1 = Elts[i+1];
2285 if (!isUndefOrEqual(BitI, j + NumElts/2))
2288 if (isUndefOrEqual(BitI1, NumElts))
2291 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2299 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2300 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2301 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2304 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2305 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2307 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2308 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2310 unsigned NumElems = N->getNumOperands();
2311 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2314 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2315 SDValue BitI = N->getOperand(i);
2316 SDValue BitI1 = N->getOperand(i+1);
2318 if (!isUndefOrEqual(BitI, j))
2320 if (!isUndefOrEqual(BitI1, j))
2327 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2328 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2330 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2331 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2333 unsigned NumElems = N->getNumOperands();
2334 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2337 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2338 SDValue BitI = N->getOperand(i);
2339 SDValue BitI1 = N->getOperand(i + 1);
2341 if (!isUndefOrEqual(BitI, j))
2343 if (!isUndefOrEqual(BitI1, j))
2350 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2351 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2352 /// MOVSD, and MOVD, i.e. setting the lowest element.
2353 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2354 if (NumElts != 2 && NumElts != 4)
2357 if (!isUndefOrEqual(Elts[0], NumElts))
2360 for (unsigned i = 1; i < NumElts; ++i) {
2361 if (!isUndefOrEqual(Elts[i], i))
2368 bool X86::isMOVLMask(SDNode *N) {
2369 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2370 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2373 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2374 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2375 /// element of vector 2 and the other elements to come from vector 1 in order.
2376 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2377 bool V2IsSplat = false,
2378 bool V2IsUndef = false) {
2379 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2382 if (!isUndefOrEqual(Ops[0], 0))
2385 for (unsigned i = 1; i < NumOps; ++i) {
2386 SDValue Arg = Ops[i];
2387 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2388 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2389 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2396 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2397 bool V2IsUndef = false) {
2398 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2399 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2400 V2IsSplat, V2IsUndef);
2403 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2404 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2405 bool X86::isMOVSHDUPMask(SDNode *N) {
2406 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2408 if (N->getNumOperands() != 4)
2411 // Expect 1, 1, 3, 3
2412 for (unsigned i = 0; i < 2; ++i) {
2413 SDValue Arg = N->getOperand(i);
2414 if (Arg.getOpcode() == ISD::UNDEF) continue;
2415 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2416 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2417 if (Val != 1) return false;
2421 for (unsigned i = 2; i < 4; ++i) {
2422 SDValue Arg = N->getOperand(i);
2423 if (Arg.getOpcode() == ISD::UNDEF) continue;
2424 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2425 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2426 if (Val != 3) return false;
2430 // Don't use movshdup if it can be done with a shufps.
2434 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2435 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2436 bool X86::isMOVSLDUPMask(SDNode *N) {
2437 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2439 if (N->getNumOperands() != 4)
2442 // Expect 0, 0, 2, 2
2443 for (unsigned i = 0; i < 2; ++i) {
2444 SDValue Arg = N->getOperand(i);
2445 if (Arg.getOpcode() == ISD::UNDEF) continue;
2446 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2447 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2448 if (Val != 0) return false;
2452 for (unsigned i = 2; i < 4; ++i) {
2453 SDValue Arg = N->getOperand(i);
2454 if (Arg.getOpcode() == ISD::UNDEF) continue;
2455 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2456 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2457 if (Val != 2) return false;
2461 // Don't use movshdup if it can be done with a shufps.
2465 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2466 /// specifies a identity operation on the LHS or RHS.
2467 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2468 unsigned NumElems = N->getNumOperands();
2469 for (unsigned i = 0; i < NumElems; ++i)
2470 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2475 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2476 /// a splat of a single element.
2477 static bool isSplatMask(SDNode *N) {
2478 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2480 // This is a splat operation if each element of the permute is the same, and
2481 // if the value doesn't reference the second vector.
2482 unsigned NumElems = N->getNumOperands();
2483 SDValue ElementBase;
2485 for (; i != NumElems; ++i) {
2486 SDValue Elt = N->getOperand(i);
2487 if (isa<ConstantSDNode>(Elt)) {
2493 if (!ElementBase.getNode())
2496 for (; i != NumElems; ++i) {
2497 SDValue Arg = N->getOperand(i);
2498 if (Arg.getOpcode() == ISD::UNDEF) continue;
2499 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2500 if (Arg != ElementBase) return false;
2503 // Make sure it is a splat of the first vector operand.
2504 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2507 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2508 /// a splat of a single element and it's a 2 or 4 element mask.
2509 bool X86::isSplatMask(SDNode *N) {
2510 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2512 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2513 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2515 return ::isSplatMask(N);
2518 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2519 /// specifies a splat of zero element.
2520 bool X86::isSplatLoMask(SDNode *N) {
2521 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2523 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2524 if (!isUndefOrEqual(N->getOperand(i), 0))
2529 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2530 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2532 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2533 unsigned NumOperands = N->getNumOperands();
2534 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2536 for (unsigned i = 0; i < NumOperands; ++i) {
2538 SDValue Arg = N->getOperand(NumOperands-i-1);
2539 if (Arg.getOpcode() != ISD::UNDEF)
2540 Val = cast<ConstantSDNode>(Arg)->getValue();
2541 if (Val >= NumOperands) Val -= NumOperands;
2543 if (i != NumOperands - 1)
2550 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2551 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2553 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2555 // 8 nodes, but we only care about the last 4.
2556 for (unsigned i = 7; i >= 4; --i) {
2558 SDValue Arg = N->getOperand(i);
2559 if (Arg.getOpcode() != ISD::UNDEF)
2560 Val = cast<ConstantSDNode>(Arg)->getValue();
2569 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2570 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2572 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2574 // 8 nodes, but we only care about the first 4.
2575 for (int i = 3; i >= 0; --i) {
2577 SDValue Arg = N->getOperand(i);
2578 if (Arg.getOpcode() != ISD::UNDEF)
2579 Val = cast<ConstantSDNode>(Arg)->getValue();
2588 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2589 /// specifies a 8 element shuffle that can be broken into a pair of
2590 /// PSHUFHW and PSHUFLW.
2591 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2592 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2594 if (N->getNumOperands() != 8)
2597 // Lower quadword shuffled.
2598 for (unsigned i = 0; i != 4; ++i) {
2599 SDValue Arg = N->getOperand(i);
2600 if (Arg.getOpcode() == ISD::UNDEF) continue;
2601 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2602 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2607 // Upper quadword shuffled.
2608 for (unsigned i = 4; i != 8; ++i) {
2609 SDValue Arg = N->getOperand(i);
2610 if (Arg.getOpcode() == ISD::UNDEF) continue;
2611 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2612 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2613 if (Val < 4 || Val > 7)
2620 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2621 /// values in ther permute mask.
2622 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2623 SDValue &V2, SDValue &Mask,
2624 SelectionDAG &DAG) {
2625 MVT VT = Op.getValueType();
2626 MVT MaskVT = Mask.getValueType();
2627 MVT EltVT = MaskVT.getVectorElementType();
2628 unsigned NumElems = Mask.getNumOperands();
2629 SmallVector<SDValue, 8> MaskVec;
2631 for (unsigned i = 0; i != NumElems; ++i) {
2632 SDValue Arg = Mask.getOperand(i);
2633 if (Arg.getOpcode() == ISD::UNDEF) {
2634 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2637 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2638 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2640 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2642 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2646 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2647 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2650 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2651 /// the two vector operands have swapped position.
2653 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2654 MVT MaskVT = Mask.getValueType();
2655 MVT EltVT = MaskVT.getVectorElementType();
2656 unsigned NumElems = Mask.getNumOperands();
2657 SmallVector<SDValue, 8> MaskVec;
2658 for (unsigned i = 0; i != NumElems; ++i) {
2659 SDValue Arg = Mask.getOperand(i);
2660 if (Arg.getOpcode() == ISD::UNDEF) {
2661 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2664 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2665 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2667 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2669 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2671 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2675 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2676 /// match movhlps. The lower half elements should come from upper half of
2677 /// V1 (and in order), and the upper half elements should come from the upper
2678 /// half of V2 (and in order).
2679 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2680 unsigned NumElems = Mask->getNumOperands();
2683 for (unsigned i = 0, e = 2; i != e; ++i)
2684 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2686 for (unsigned i = 2; i != 4; ++i)
2687 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2692 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2693 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2695 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2696 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2697 N = N->getOperand(0).getNode();
2698 if (ISD::isNON_EXTLoad(N)) {
2700 *LD = cast<LoadSDNode>(N);
2707 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2708 /// match movlp{s|d}. The lower half elements should come from lower half of
2709 /// V1 (and in order), and the upper half elements should come from the upper
2710 /// half of V2 (and in order). And since V1 will become the source of the
2711 /// MOVLP, it must be either a vector load or a scalar load to vector.
2712 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2713 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2715 // Is V2 is a vector load, don't do this transformation. We will try to use
2716 // load folding shufps op.
2717 if (ISD::isNON_EXTLoad(V2))
2720 unsigned NumElems = Mask->getNumOperands();
2721 if (NumElems != 2 && NumElems != 4)
2723 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2724 if (!isUndefOrEqual(Mask->getOperand(i), i))
2726 for (unsigned i = NumElems/2; i != NumElems; ++i)
2727 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2732 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2734 static bool isSplatVector(SDNode *N) {
2735 if (N->getOpcode() != ISD::BUILD_VECTOR)
2738 SDValue SplatValue = N->getOperand(0);
2739 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2740 if (N->getOperand(i) != SplatValue)
2745 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2747 static bool isUndefShuffle(SDNode *N) {
2748 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2751 SDValue V1 = N->getOperand(0);
2752 SDValue V2 = N->getOperand(1);
2753 SDValue Mask = N->getOperand(2);
2754 unsigned NumElems = Mask.getNumOperands();
2755 for (unsigned i = 0; i != NumElems; ++i) {
2756 SDValue Arg = Mask.getOperand(i);
2757 if (Arg.getOpcode() != ISD::UNDEF) {
2758 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2759 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2761 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2768 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2770 static inline bool isZeroNode(SDValue Elt) {
2771 return ((isa<ConstantSDNode>(Elt) &&
2772 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2773 (isa<ConstantFPSDNode>(Elt) &&
2774 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2777 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2778 /// to an zero vector.
2779 static bool isZeroShuffle(SDNode *N) {
2780 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2783 SDValue V1 = N->getOperand(0);
2784 SDValue V2 = N->getOperand(1);
2785 SDValue Mask = N->getOperand(2);
2786 unsigned NumElems = Mask.getNumOperands();
2787 for (unsigned i = 0; i != NumElems; ++i) {
2788 SDValue Arg = Mask.getOperand(i);
2789 if (Arg.getOpcode() == ISD::UNDEF)
2792 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2793 if (Idx < NumElems) {
2794 unsigned Opc = V1.getNode()->getOpcode();
2795 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2797 if (Opc != ISD::BUILD_VECTOR ||
2798 !isZeroNode(V1.getNode()->getOperand(Idx)))
2800 } else if (Idx >= NumElems) {
2801 unsigned Opc = V2.getNode()->getOpcode();
2802 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2804 if (Opc != ISD::BUILD_VECTOR ||
2805 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2812 /// getZeroVector - Returns a vector of specified type with all zero elements.
2814 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2815 assert(VT.isVector() && "Expected a vector type");
2817 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2818 // type. This ensures they get CSE'd.
2820 if (VT.getSizeInBits() == 64) { // MMX
2821 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2822 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2823 } else if (HasSSE2) { // SSE2
2824 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2825 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2827 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2828 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2830 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2833 /// getOnesVector - Returns a vector of specified type with all bits set.
2835 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2836 assert(VT.isVector() && "Expected a vector type");
2838 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2839 // type. This ensures they get CSE'd.
2840 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2842 if (VT.getSizeInBits() == 64) // MMX
2843 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2845 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2846 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2850 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2851 /// that point to V2 points to its first element.
2852 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2853 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2855 bool Changed = false;
2856 SmallVector<SDValue, 8> MaskVec;
2857 unsigned NumElems = Mask.getNumOperands();
2858 for (unsigned i = 0; i != NumElems; ++i) {
2859 SDValue Arg = Mask.getOperand(i);
2860 if (Arg.getOpcode() != ISD::UNDEF) {
2861 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2862 if (Val > NumElems) {
2863 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2867 MaskVec.push_back(Arg);
2871 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2872 &MaskVec[0], MaskVec.size());
2876 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2877 /// operation of specified width.
2878 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2879 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2880 MVT BaseVT = MaskVT.getVectorElementType();
2882 SmallVector<SDValue, 8> MaskVec;
2883 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2884 for (unsigned i = 1; i != NumElems; ++i)
2885 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2886 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2889 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2890 /// of specified width.
2891 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2892 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2893 MVT BaseVT = MaskVT.getVectorElementType();
2894 SmallVector<SDValue, 8> MaskVec;
2895 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2896 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2897 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2899 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2902 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2903 /// of specified width.
2904 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2905 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2906 MVT BaseVT = MaskVT.getVectorElementType();
2907 unsigned Half = NumElems/2;
2908 SmallVector<SDValue, 8> MaskVec;
2909 for (unsigned i = 0; i != Half; ++i) {
2910 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2911 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2913 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2916 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2917 /// element #0 of a vector with the specified index, leaving the rest of the
2918 /// elements in place.
2919 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2920 SelectionDAG &DAG) {
2921 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2922 MVT BaseVT = MaskVT.getVectorElementType();
2923 SmallVector<SDValue, 8> MaskVec;
2924 // Element #0 of the result gets the elt we are replacing.
2925 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2926 for (unsigned i = 1; i != NumElems; ++i)
2927 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2928 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2931 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2932 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2933 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2934 MVT VT = Op.getValueType();
2937 SDValue V1 = Op.getOperand(0);
2938 SDValue Mask = Op.getOperand(2);
2939 unsigned NumElems = Mask.getNumOperands();
2940 // Special handling of v4f32 -> v4i32.
2941 if (VT != MVT::v4f32) {
2942 Mask = getUnpacklMask(NumElems, DAG);
2943 while (NumElems > 4) {
2944 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2947 Mask = getZeroVector(MVT::v4i32, true, DAG);
2950 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2951 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2952 DAG.getNode(ISD::UNDEF, PVT), Mask);
2953 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2956 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2957 /// vector of zero or undef vector. This produces a shuffle where the low
2958 /// element of V2 is swizzled into the zero/undef vector, landing at element
2959 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2960 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2961 bool isZero, bool HasSSE2,
2962 SelectionDAG &DAG) {
2963 MVT VT = V2.getValueType();
2965 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
2966 unsigned NumElems = V2.getValueType().getVectorNumElements();
2967 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2968 MVT EVT = MaskVT.getVectorElementType();
2969 SmallVector<SDValue, 16> MaskVec;
2970 for (unsigned i = 0; i != NumElems; ++i)
2971 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2972 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2974 MaskVec.push_back(DAG.getConstant(i, EVT));
2975 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2976 &MaskVec[0], MaskVec.size());
2977 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2980 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2981 /// a shuffle that is zero.
2983 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
2984 unsigned NumElems, bool Low,
2985 SelectionDAG &DAG) {
2986 unsigned NumZeros = 0;
2987 for (unsigned i = 0; i < NumElems; ++i) {
2988 unsigned Index = Low ? i : NumElems-i-1;
2989 SDValue Idx = Mask.getOperand(Index);
2990 if (Idx.getOpcode() == ISD::UNDEF) {
2994 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
2995 if (Elt.getNode() && isZeroNode(Elt))
3003 /// isVectorShift - Returns true if the shuffle can be implemented as a
3004 /// logical left or right shift of a vector.
3005 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3006 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3007 unsigned NumElems = Mask.getNumOperands();
3010 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3013 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3018 bool SeenV1 = false;
3019 bool SeenV2 = false;
3020 for (unsigned i = NumZeros; i < NumElems; ++i) {
3021 unsigned Val = isLeft ? (i - NumZeros) : i;
3022 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3023 if (Idx.getOpcode() == ISD::UNDEF)
3025 unsigned Index = cast<ConstantSDNode>(Idx)->getValue();
3026 if (Index < NumElems)
3035 if (SeenV1 && SeenV2)
3038 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3044 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3046 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3047 unsigned NumNonZero, unsigned NumZero,
3048 SelectionDAG &DAG, TargetLowering &TLI) {
3054 for (unsigned i = 0; i < 16; ++i) {
3055 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3056 if (ThisIsNonZero && First) {
3058 V = getZeroVector(MVT::v8i16, true, DAG);
3060 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3065 SDValue ThisElt(0, 0), LastElt(0, 0);
3066 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3067 if (LastIsNonZero) {
3068 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3070 if (ThisIsNonZero) {
3071 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3072 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3073 ThisElt, DAG.getConstant(8, MVT::i8));
3075 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3079 if (ThisElt.getNode())
3080 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3081 DAG.getIntPtrConstant(i/2));
3085 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3088 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3090 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3091 unsigned NumNonZero, unsigned NumZero,
3092 SelectionDAG &DAG, TargetLowering &TLI) {
3098 for (unsigned i = 0; i < 8; ++i) {
3099 bool isNonZero = (NonZeros & (1 << i)) != 0;
3103 V = getZeroVector(MVT::v8i16, true, DAG);
3105 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3108 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3109 DAG.getIntPtrConstant(i));
3116 /// getVShift - Return a vector logical shift node.
3118 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3119 unsigned NumBits, SelectionDAG &DAG,
3120 const TargetLowering &TLI) {
3121 bool isMMX = VT.getSizeInBits() == 64;
3122 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3123 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3124 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3125 return DAG.getNode(ISD::BIT_CONVERT, VT,
3126 DAG.getNode(Opc, ShVT, SrcOp,
3127 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3131 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3132 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3133 if (ISD::isBuildVectorAllZeros(Op.getNode())
3134 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3135 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3136 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3137 // eliminated on x86-32 hosts.
3138 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3141 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3142 return getOnesVector(Op.getValueType(), DAG);
3143 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3146 MVT VT = Op.getValueType();
3147 MVT EVT = VT.getVectorElementType();
3148 unsigned EVTBits = EVT.getSizeInBits();
3150 unsigned NumElems = Op.getNumOperands();
3151 unsigned NumZero = 0;
3152 unsigned NumNonZero = 0;
3153 unsigned NonZeros = 0;
3154 bool IsAllConstants = true;
3155 SmallSet<SDValue, 8> Values;
3156 for (unsigned i = 0; i < NumElems; ++i) {
3157 SDValue Elt = Op.getOperand(i);
3158 if (Elt.getOpcode() == ISD::UNDEF)
3161 if (Elt.getOpcode() != ISD::Constant &&
3162 Elt.getOpcode() != ISD::ConstantFP)
3163 IsAllConstants = false;
3164 if (isZeroNode(Elt))
3167 NonZeros |= (1 << i);
3172 if (NumNonZero == 0) {
3173 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3174 return DAG.getNode(ISD::UNDEF, VT);
3177 // Special case for single non-zero, non-undef, element.
3178 if (NumNonZero == 1 && NumElems <= 4) {
3179 unsigned Idx = CountTrailingZeros_32(NonZeros);
3180 SDValue Item = Op.getOperand(Idx);
3182 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3183 // the value are obviously zero, truncate the value to i32 and do the
3184 // insertion that way. Only do this if the value is non-constant or if the
3185 // value is a constant being inserted into element 0. It is cheaper to do
3186 // a constant pool load than it is to do a movd + shuffle.
3187 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3188 (!IsAllConstants || Idx == 0)) {
3189 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3190 // Handle MMX and SSE both.
3191 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3192 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3194 // Truncate the value (which may itself be a constant) to i32, and
3195 // convert it to a vector with movd (S2V+shuffle to zero extend).
3196 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3197 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3198 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3199 Subtarget->hasSSE2(), DAG);
3201 // Now we have our 32-bit value zero extended in the low element of
3202 // a vector. If Idx != 0, swizzle it into place.
3205 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3206 getSwapEltZeroMask(VecElts, Idx, DAG)
3208 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3210 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3214 // If we have a constant or non-constant insertion into the low element of
3215 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3216 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3217 // depending on what the source datatype is. Because we can only get here
3218 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3220 // Don't do this for i64 values on x86-32.
3221 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3222 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3223 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3224 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3225 Subtarget->hasSSE2(), DAG);
3228 // Is it a vector logical left shift?
3229 if (NumElems == 2 && Idx == 1 &&
3230 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3231 unsigned NumBits = VT.getSizeInBits();
3232 return getVShift(true, VT,
3233 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3234 NumBits/2, DAG, *this);
3237 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3240 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3241 // is a non-constant being inserted into an element other than the low one,
3242 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3243 // movd/movss) to move this into the low element, then shuffle it into
3245 if (EVTBits == 32) {
3246 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3248 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3249 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3250 Subtarget->hasSSE2(), DAG);
3251 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3252 MVT MaskEVT = MaskVT.getVectorElementType();
3253 SmallVector<SDValue, 8> MaskVec;
3254 for (unsigned i = 0; i < NumElems; i++)
3255 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3256 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3257 &MaskVec[0], MaskVec.size());
3258 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3259 DAG.getNode(ISD::UNDEF, VT), Mask);
3263 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3264 if (Values.size() == 1)
3267 // A vector full of immediates; various special cases are already
3268 // handled, so this is best done with a single constant-pool load.
3272 // Let legalizer expand 2-wide build_vectors.
3273 if (EVTBits == 64) {
3274 if (NumNonZero == 1) {
3275 // One half is zero or undef.
3276 unsigned Idx = CountTrailingZeros_32(NonZeros);
3277 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3278 Op.getOperand(Idx));
3279 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3280 Subtarget->hasSSE2(), DAG);
3285 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3286 if (EVTBits == 8 && NumElems == 16) {
3287 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3289 if (V.getNode()) return V;
3292 if (EVTBits == 16 && NumElems == 8) {
3293 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3295 if (V.getNode()) return V;
3298 // If element VT is == 32 bits, turn it into a number of shuffles.
3299 SmallVector<SDValue, 8> V;
3301 if (NumElems == 4 && NumZero > 0) {
3302 for (unsigned i = 0; i < 4; ++i) {
3303 bool isZero = !(NonZeros & (1 << i));
3305 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3307 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3310 for (unsigned i = 0; i < 2; ++i) {
3311 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3314 V[i] = V[i*2]; // Must be a zero vector.
3317 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3318 getMOVLMask(NumElems, DAG));
3321 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3322 getMOVLMask(NumElems, DAG));
3325 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3326 getUnpacklMask(NumElems, DAG));
3331 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3332 MVT EVT = MaskVT.getVectorElementType();
3333 SmallVector<SDValue, 8> MaskVec;
3334 bool Reverse = (NonZeros & 0x3) == 2;
3335 for (unsigned i = 0; i < 2; ++i)
3337 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3339 MaskVec.push_back(DAG.getConstant(i, EVT));
3340 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3341 for (unsigned i = 0; i < 2; ++i)
3343 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3345 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3346 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3347 &MaskVec[0], MaskVec.size());
3348 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3351 if (Values.size() > 2) {
3352 // Expand into a number of unpckl*.
3354 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3355 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3356 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3357 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3358 for (unsigned i = 0; i < NumElems; ++i)
3359 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3361 while (NumElems != 0) {
3362 for (unsigned i = 0; i < NumElems; ++i)
3363 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3374 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3375 SDValue PermMask, SelectionDAG &DAG,
3376 TargetLowering &TLI) {
3378 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3379 MVT MaskEVT = MaskVT.getVectorElementType();
3380 MVT PtrVT = TLI.getPointerTy();
3381 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3382 PermMask.getNode()->op_end());
3384 // First record which half of which vector the low elements come from.
3385 SmallVector<unsigned, 4> LowQuad(4);
3386 for (unsigned i = 0; i < 4; ++i) {
3387 SDValue Elt = MaskElts[i];
3388 if (Elt.getOpcode() == ISD::UNDEF)
3390 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3391 int QuadIdx = EltIdx / 4;
3395 int BestLowQuad = -1;
3396 unsigned MaxQuad = 1;
3397 for (unsigned i = 0; i < 4; ++i) {
3398 if (LowQuad[i] > MaxQuad) {
3400 MaxQuad = LowQuad[i];
3404 // Record which half of which vector the high elements come from.
3405 SmallVector<unsigned, 4> HighQuad(4);
3406 for (unsigned i = 4; i < 8; ++i) {
3407 SDValue Elt = MaskElts[i];
3408 if (Elt.getOpcode() == ISD::UNDEF)
3410 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3411 int QuadIdx = EltIdx / 4;
3412 ++HighQuad[QuadIdx];
3415 int BestHighQuad = -1;
3417 for (unsigned i = 0; i < 4; ++i) {
3418 if (HighQuad[i] > MaxQuad) {
3420 MaxQuad = HighQuad[i];
3424 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3425 if (BestLowQuad != -1 || BestHighQuad != -1) {
3426 // First sort the 4 chunks in order using shufpd.
3427 SmallVector<SDValue, 8> MaskVec;
3429 if (BestLowQuad != -1)
3430 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3432 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3434 if (BestHighQuad != -1)
3435 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3437 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3439 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3440 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3441 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3442 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3443 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3445 // Now sort high and low parts separately.
3446 BitVector InOrder(8);
3447 if (BestLowQuad != -1) {
3448 // Sort lower half in order using PSHUFLW.
3450 bool AnyOutOrder = false;
3452 for (unsigned i = 0; i != 4; ++i) {
3453 SDValue Elt = MaskElts[i];
3454 if (Elt.getOpcode() == ISD::UNDEF) {
3455 MaskVec.push_back(Elt);
3458 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3462 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3464 // If this element is in the right place after this shuffle, then
3466 if ((int)(EltIdx / 4) == BestLowQuad)
3471 for (unsigned i = 4; i != 8; ++i)
3472 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3473 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3474 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3478 if (BestHighQuad != -1) {
3479 // Sort high half in order using PSHUFHW if possible.
3482 for (unsigned i = 0; i != 4; ++i)
3483 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3485 bool AnyOutOrder = false;
3486 for (unsigned i = 4; i != 8; ++i) {
3487 SDValue Elt = MaskElts[i];
3488 if (Elt.getOpcode() == ISD::UNDEF) {
3489 MaskVec.push_back(Elt);
3492 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3496 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3498 // If this element is in the right place after this shuffle, then
3500 if ((int)(EltIdx / 4) == BestHighQuad)
3506 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3507 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3511 // The other elements are put in the right place using pextrw and pinsrw.
3512 for (unsigned i = 0; i != 8; ++i) {
3515 SDValue Elt = MaskElts[i];
3516 if (Elt.getOpcode() == ISD::UNDEF)
3518 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3519 SDValue ExtOp = (EltIdx < 8)
3520 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3521 DAG.getConstant(EltIdx, PtrVT))
3522 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3523 DAG.getConstant(EltIdx - 8, PtrVT));
3524 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3525 DAG.getConstant(i, PtrVT));
3531 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3532 // few as possible. First, let's find out how many elements are already in the
3534 unsigned V1InOrder = 0;
3535 unsigned V1FromV1 = 0;
3536 unsigned V2InOrder = 0;
3537 unsigned V2FromV2 = 0;
3538 SmallVector<SDValue, 8> V1Elts;
3539 SmallVector<SDValue, 8> V2Elts;
3540 for (unsigned i = 0; i < 8; ++i) {
3541 SDValue Elt = MaskElts[i];
3542 if (Elt.getOpcode() == ISD::UNDEF) {
3543 V1Elts.push_back(Elt);
3544 V2Elts.push_back(Elt);
3549 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3551 V1Elts.push_back(Elt);
3552 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3554 } else if (EltIdx == i+8) {
3555 V1Elts.push_back(Elt);
3556 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3558 } else if (EltIdx < 8) {
3559 V1Elts.push_back(Elt);
3562 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3567 if (V2InOrder > V1InOrder) {
3568 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3570 std::swap(V1Elts, V2Elts);
3571 std::swap(V1FromV1, V2FromV2);
3574 if ((V1FromV1 + V1InOrder) != 8) {
3575 // Some elements are from V2.
3577 // If there are elements that are from V1 but out of place,
3578 // then first sort them in place
3579 SmallVector<SDValue, 8> MaskVec;
3580 for (unsigned i = 0; i < 8; ++i) {
3581 SDValue Elt = V1Elts[i];
3582 if (Elt.getOpcode() == ISD::UNDEF) {
3583 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3586 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3588 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3590 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3592 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3593 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3597 for (unsigned i = 0; i < 8; ++i) {
3598 SDValue Elt = V1Elts[i];
3599 if (Elt.getOpcode() == ISD::UNDEF)
3601 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3604 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3605 DAG.getConstant(EltIdx - 8, PtrVT));
3606 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3607 DAG.getConstant(i, PtrVT));
3611 // All elements are from V1.
3613 for (unsigned i = 0; i < 8; ++i) {
3614 SDValue Elt = V1Elts[i];
3615 if (Elt.getOpcode() == ISD::UNDEF)
3617 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3618 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3619 DAG.getConstant(EltIdx, PtrVT));
3620 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3621 DAG.getConstant(i, PtrVT));
3627 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3628 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3629 /// done when every pair / quad of shuffle mask elements point to elements in
3630 /// the right sequence. e.g.
3631 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3633 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3635 SDValue PermMask, SelectionDAG &DAG,
3636 TargetLowering &TLI) {
3637 unsigned NumElems = PermMask.getNumOperands();
3638 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3639 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3640 MVT MaskEltVT = MaskVT.getVectorElementType();
3642 switch (VT.getSimpleVT()) {
3643 default: assert(false && "Unexpected!");
3644 case MVT::v4f32: NewVT = MVT::v2f64; break;
3645 case MVT::v4i32: NewVT = MVT::v2i64; break;
3646 case MVT::v8i16: NewVT = MVT::v4i32; break;
3647 case MVT::v16i8: NewVT = MVT::v4i32; break;
3650 if (NewWidth == 2) {
3656 unsigned Scale = NumElems / NewWidth;
3657 SmallVector<SDValue, 8> MaskVec;
3658 for (unsigned i = 0; i < NumElems; i += Scale) {
3659 unsigned StartIdx = ~0U;
3660 for (unsigned j = 0; j < Scale; ++j) {
3661 SDValue Elt = PermMask.getOperand(i+j);
3662 if (Elt.getOpcode() == ISD::UNDEF)
3664 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3665 if (StartIdx == ~0U)
3666 StartIdx = EltIdx - (EltIdx % Scale);
3667 if (EltIdx != StartIdx + j)
3670 if (StartIdx == ~0U)
3671 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3673 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3676 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3677 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3678 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3679 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3680 &MaskVec[0], MaskVec.size()));
3683 /// getVZextMovL - Return a zero-extending vector move low node.
3685 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3686 SDValue SrcOp, SelectionDAG &DAG,
3687 const X86Subtarget *Subtarget) {
3688 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3689 LoadSDNode *LD = NULL;
3690 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3691 LD = dyn_cast<LoadSDNode>(SrcOp);
3693 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3695 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3696 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3697 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3698 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3699 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3701 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3702 return DAG.getNode(ISD::BIT_CONVERT, VT,
3703 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3704 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3711 return DAG.getNode(ISD::BIT_CONVERT, VT,
3712 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3713 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3716 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3719 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3720 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3721 MVT MaskVT = PermMask.getValueType();
3722 MVT MaskEVT = MaskVT.getVectorElementType();
3723 SmallVector<std::pair<int, int>, 8> Locs;
3725 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3728 for (unsigned i = 0; i != 4; ++i) {
3729 SDValue Elt = PermMask.getOperand(i);
3730 if (Elt.getOpcode() == ISD::UNDEF) {
3731 Locs[i] = std::make_pair(-1, -1);
3733 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3734 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3736 Locs[i] = std::make_pair(0, NumLo);
3740 Locs[i] = std::make_pair(1, NumHi);
3742 Mask1[2+NumHi] = Elt;
3748 if (NumLo <= 2 && NumHi <= 2) {
3749 // If no more than two elements come from either vector. This can be
3750 // implemented with two shuffles. First shuffle gather the elements.
3751 // The second shuffle, which takes the first shuffle as both of its
3752 // vector operands, put the elements into the right order.
3753 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3754 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3755 &Mask1[0], Mask1.size()));
3757 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3758 for (unsigned i = 0; i != 4; ++i) {
3759 if (Locs[i].first == -1)
3762 unsigned Idx = (i < 2) ? 0 : 4;
3763 Idx += Locs[i].first * 2 + Locs[i].second;
3764 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3768 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3769 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3770 &Mask2[0], Mask2.size()));
3771 } else if (NumLo == 3 || NumHi == 3) {
3772 // Otherwise, we must have three elements from one vector, call it X, and
3773 // one element from the other, call it Y. First, use a shufps to build an
3774 // intermediate vector with the one element from Y and the element from X
3775 // that will be in the same half in the final destination (the indexes don't
3776 // matter). Then, use a shufps to build the final vector, taking the half
3777 // containing the element from Y from the intermediate, and the other half
3780 // Normalize it so the 3 elements come from V1.
3781 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3785 // Find the element from V2.
3787 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3788 SDValue Elt = PermMask.getOperand(HiIndex);
3789 if (Elt.getOpcode() == ISD::UNDEF)
3791 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3796 Mask1[0] = PermMask.getOperand(HiIndex);
3797 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3798 Mask1[2] = PermMask.getOperand(HiIndex^1);
3799 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3800 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3801 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3804 Mask1[0] = PermMask.getOperand(0);
3805 Mask1[1] = PermMask.getOperand(1);
3806 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3807 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3808 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3809 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3811 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3812 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3813 Mask1[2] = PermMask.getOperand(2);
3814 Mask1[3] = PermMask.getOperand(3);
3815 if (Mask1[2].getOpcode() != ISD::UNDEF)
3816 Mask1[2] = DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getValue()+4,
3818 if (Mask1[3].getOpcode() != ISD::UNDEF)
3819 Mask1[3] = DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getValue()+4,
3821 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3822 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3826 // Break it into (shuffle shuffle_hi, shuffle_lo).
3828 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3829 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3830 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3831 unsigned MaskIdx = 0;
3834 for (unsigned i = 0; i != 4; ++i) {
3841 SDValue Elt = PermMask.getOperand(i);
3842 if (Elt.getOpcode() == ISD::UNDEF) {
3843 Locs[i] = std::make_pair(-1, -1);
3844 } else if (cast<ConstantSDNode>(Elt)->getValue() < 4) {
3845 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3846 (*MaskPtr)[LoIdx] = Elt;
3849 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3850 (*MaskPtr)[HiIdx] = Elt;
3855 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3856 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3857 &LoMask[0], LoMask.size()));
3858 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3859 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3860 &HiMask[0], HiMask.size()));
3861 SmallVector<SDValue, 8> MaskOps;
3862 for (unsigned i = 0; i != 4; ++i) {
3863 if (Locs[i].first == -1) {
3864 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3866 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3867 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3870 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3871 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3872 &MaskOps[0], MaskOps.size()));
3876 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3877 SDValue V1 = Op.getOperand(0);
3878 SDValue V2 = Op.getOperand(1);
3879 SDValue PermMask = Op.getOperand(2);
3880 MVT VT = Op.getValueType();
3881 unsigned NumElems = PermMask.getNumOperands();
3882 bool isMMX = VT.getSizeInBits() == 64;
3883 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3884 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3885 bool V1IsSplat = false;
3886 bool V2IsSplat = false;
3888 if (isUndefShuffle(Op.getNode()))
3889 return DAG.getNode(ISD::UNDEF, VT);
3891 if (isZeroShuffle(Op.getNode()))
3892 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3894 if (isIdentityMask(PermMask.getNode()))
3896 else if (isIdentityMask(PermMask.getNode(), true))
3899 if (isSplatMask(PermMask.getNode())) {
3900 if (isMMX || NumElems < 4) return Op;
3901 // Promote it to a v4{if}32 splat.
3902 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3905 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3907 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3908 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3909 if (NewOp.getNode())
3910 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3911 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3912 // FIXME: Figure out a cleaner way to do this.
3913 // Try to make use of movq to zero out the top part.
3914 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3915 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3917 if (NewOp.getNode()) {
3918 SDValue NewV1 = NewOp.getOperand(0);
3919 SDValue NewV2 = NewOp.getOperand(1);
3920 SDValue NewMask = NewOp.getOperand(2);
3921 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3922 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3923 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3926 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3927 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3929 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3930 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3935 // Check if this can be converted into a logical shift.
3936 bool isLeft = false;
3939 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
3940 if (isShift && ShVal.hasOneUse()) {
3941 // If the shifted value has multiple uses, it may be cheaper to use
3942 // v_set0 + movlhps or movhlps, etc.
3943 MVT EVT = VT.getVectorElementType();
3944 ShAmt *= EVT.getSizeInBits();
3945 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3948 if (X86::isMOVLMask(PermMask.getNode())) {
3951 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3952 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
3957 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
3958 X86::isMOVSLDUPMask(PermMask.getNode()) ||
3959 X86::isMOVHLPSMask(PermMask.getNode()) ||
3960 X86::isMOVHPMask(PermMask.getNode()) ||
3961 X86::isMOVLPMask(PermMask.getNode())))
3964 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
3965 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
3966 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3969 // No better options. Use a vshl / vsrl.
3970 MVT EVT = VT.getVectorElementType();
3971 ShAmt *= EVT.getSizeInBits();
3972 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3975 bool Commuted = false;
3976 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3977 // 1,1,1,1 -> v8i16 though.
3978 V1IsSplat = isSplatVector(V1.getNode());
3979 V2IsSplat = isSplatVector(V2.getNode());
3981 // Canonicalize the splat or undef, if present, to be on the RHS.
3982 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3983 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3984 std::swap(V1IsSplat, V2IsSplat);
3985 std::swap(V1IsUndef, V2IsUndef);
3989 // FIXME: Figure out a cleaner way to do this.
3990 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
3991 if (V2IsUndef) return V1;
3992 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3994 // V2 is a splat, so the mask may be malformed. That is, it may point
3995 // to any V2 element. The instruction selectior won't like this. Get
3996 // a corrected mask and commute to form a proper MOVS{S|D}.
3997 SDValue NewMask = getMOVLMask(NumElems, DAG);
3998 if (NewMask.getNode() != PermMask.getNode())
3999 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4004 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4005 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4006 X86::isUNPCKLMask(PermMask.getNode()) ||
4007 X86::isUNPCKHMask(PermMask.getNode()))
4011 // Normalize mask so all entries that point to V2 points to its first
4012 // element then try to match unpck{h|l} again. If match, return a
4013 // new vector_shuffle with the corrected mask.
4014 SDValue NewMask = NormalizeMask(PermMask, DAG);
4015 if (NewMask.getNode() != PermMask.getNode()) {
4016 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4017 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4018 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4019 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4020 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4021 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4026 // Normalize the node to match x86 shuffle ops if needed
4027 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4028 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4031 // Commute is back and try unpck* again.
4032 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4033 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4034 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4035 X86::isUNPCKLMask(PermMask.getNode()) ||
4036 X86::isUNPCKHMask(PermMask.getNode()))
4040 // Try PSHUF* first, then SHUFP*.
4041 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4042 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4043 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4044 if (V2.getOpcode() != ISD::UNDEF)
4045 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4046 DAG.getNode(ISD::UNDEF, VT), PermMask);
4051 if (Subtarget->hasSSE2() &&
4052 (X86::isPSHUFDMask(PermMask.getNode()) ||
4053 X86::isPSHUFHWMask(PermMask.getNode()) ||
4054 X86::isPSHUFLWMask(PermMask.getNode()))) {
4056 if (VT == MVT::v4f32) {
4058 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4059 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4060 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4061 } else if (V2.getOpcode() != ISD::UNDEF)
4062 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4063 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4065 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4069 // Binary or unary shufps.
4070 if (X86::isSHUFPMask(PermMask.getNode()) ||
4071 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4075 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4076 if (VT == MVT::v8i16) {
4077 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4078 if (NewOp.getNode())
4082 // Handle all 4 wide cases with a number of shuffles except for MMX.
4083 if (NumElems == 4 && !isMMX)
4084 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4090 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4091 SelectionDAG &DAG) {
4092 MVT VT = Op.getValueType();
4093 if (VT.getSizeInBits() == 8) {
4094 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4095 Op.getOperand(0), Op.getOperand(1));
4096 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4097 DAG.getValueType(VT));
4098 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4099 } else if (VT.getSizeInBits() == 16) {
4100 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4101 Op.getOperand(0), Op.getOperand(1));
4102 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4103 DAG.getValueType(VT));
4104 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4105 } else if (VT == MVT::f32) {
4106 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4107 // the result back to FR32 register. It's only worth matching if the
4108 // result has a single use which is a store or a bitcast to i32.
4109 if (!Op.hasOneUse())
4111 SDNode *User = *Op.getNode()->use_begin();
4112 if (User->getOpcode() != ISD::STORE &&
4113 (User->getOpcode() != ISD::BIT_CONVERT ||
4114 User->getValueType(0) != MVT::i32))
4116 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4117 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4119 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4126 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4127 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4130 if (Subtarget->hasSSE41()) {
4131 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4136 MVT VT = Op.getValueType();
4137 // TODO: handle v16i8.
4138 if (VT.getSizeInBits() == 16) {
4139 SDValue Vec = Op.getOperand(0);
4140 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4142 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4143 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4144 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4146 // Transform it so it match pextrw which produces a 32-bit result.
4147 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4148 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4149 Op.getOperand(0), Op.getOperand(1));
4150 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4151 DAG.getValueType(VT));
4152 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4153 } else if (VT.getSizeInBits() == 32) {
4154 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4157 // SHUFPS the element to the lowest double word, then movss.
4158 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4159 SmallVector<SDValue, 8> IdxVec;
4161 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4163 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4165 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4167 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4168 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4169 &IdxVec[0], IdxVec.size());
4170 SDValue Vec = Op.getOperand(0);
4171 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4172 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4173 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4174 DAG.getIntPtrConstant(0));
4175 } else if (VT.getSizeInBits() == 64) {
4176 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4177 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4178 // to match extract_elt for f64.
4179 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4183 // UNPCKHPD the element to the lowest double word, then movsd.
4184 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4185 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4186 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4187 SmallVector<SDValue, 8> IdxVec;
4188 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4190 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4191 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4192 &IdxVec[0], IdxVec.size());
4193 SDValue Vec = Op.getOperand(0);
4194 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4195 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4196 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4197 DAG.getIntPtrConstant(0));
4204 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4205 MVT VT = Op.getValueType();
4206 MVT EVT = VT.getVectorElementType();
4208 SDValue N0 = Op.getOperand(0);
4209 SDValue N1 = Op.getOperand(1);
4210 SDValue N2 = Op.getOperand(2);
4212 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4213 isa<ConstantSDNode>(N2)) {
4214 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4216 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4218 if (N1.getValueType() != MVT::i32)
4219 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4220 if (N2.getValueType() != MVT::i32)
4221 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4222 return DAG.getNode(Opc, VT, N0, N1, N2);
4223 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4224 // Bits [7:6] of the constant are the source select. This will always be
4225 // zero here. The DAG Combiner may combine an extract_elt index into these
4226 // bits. For example (insert (extract, 3), 2) could be matched by putting
4227 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4228 // Bits [5:4] of the constant are the destination select. This is the
4229 // value of the incoming immediate.
4230 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4231 // combine either bitwise AND or insert of float 0.0 to set these bits.
4232 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
4233 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4239 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4240 MVT VT = Op.getValueType();
4241 MVT EVT = VT.getVectorElementType();
4243 if (Subtarget->hasSSE41())
4244 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4249 SDValue N0 = Op.getOperand(0);
4250 SDValue N1 = Op.getOperand(1);
4251 SDValue N2 = Op.getOperand(2);
4253 if (EVT.getSizeInBits() == 16) {
4254 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4255 // as its second argument.
4256 if (N1.getValueType() != MVT::i32)
4257 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4258 if (N2.getValueType() != MVT::i32)
4259 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4260 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4266 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4267 if (Op.getValueType() == MVT::v2f32)
4268 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4269 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4270 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4271 Op.getOperand(0))));
4273 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4274 MVT VT = MVT::v2i32;
4275 switch (Op.getValueType().getSimpleVT()) {
4282 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4283 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4286 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4287 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4288 // one of the above mentioned nodes. It has to be wrapped because otherwise
4289 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4290 // be used to form addressing mode. These wrapped nodes will be selected
4293 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4294 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4295 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4297 CP->getAlignment());
4298 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4299 // With PIC, the address is actually $g + Offset.
4300 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4301 !Subtarget->isPICStyleRIPRel()) {
4302 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4303 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4311 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4312 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4313 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4314 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4315 // With PIC, the address is actually $g + Offset.
4316 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4317 !Subtarget->isPICStyleRIPRel()) {
4318 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4319 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4323 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4324 // load the value at address GV, not the value of GV itself. This means that
4325 // the GlobalAddress must be in the base or index register of the address, not
4326 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4327 // The same applies for external symbols during PIC codegen
4328 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4329 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4330 PseudoSourceValue::getGOT(), 0);
4335 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4337 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4340 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4341 DAG.getNode(X86ISD::GlobalBaseReg,
4343 InFlag = Chain.getValue(1);
4345 // emit leal symbol@TLSGD(,%ebx,1), %eax
4346 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4347 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4348 GA->getValueType(0),
4350 SDValue Ops[] = { Chain, TGA, InFlag };
4351 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4352 InFlag = Result.getValue(2);
4353 Chain = Result.getValue(1);
4355 // call ___tls_get_addr. This function receives its argument in
4356 // the register EAX.
4357 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4358 InFlag = Chain.getValue(1);
4360 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4361 SDValue Ops1[] = { Chain,
4362 DAG.getTargetExternalSymbol("___tls_get_addr",
4364 DAG.getRegister(X86::EAX, PtrVT),
4365 DAG.getRegister(X86::EBX, PtrVT),
4367 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4368 InFlag = Chain.getValue(1);
4370 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4373 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4375 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4377 SDValue InFlag, Chain;
4379 // emit leaq symbol@TLSGD(%rip), %rdi
4380 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4381 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4382 GA->getValueType(0),
4384 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4385 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4386 Chain = Result.getValue(1);
4387 InFlag = Result.getValue(2);
4389 // call __tls_get_addr. This function receives its argument in
4390 // the register RDI.
4391 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4392 InFlag = Chain.getValue(1);
4394 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4395 SDValue Ops1[] = { Chain,
4396 DAG.getTargetExternalSymbol("__tls_get_addr",
4398 DAG.getRegister(X86::RDI, PtrVT),
4400 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4401 InFlag = Chain.getValue(1);
4403 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4406 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4407 // "local exec" model.
4408 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4410 // Get the Thread Pointer
4411 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4412 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4414 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4415 GA->getValueType(0),
4417 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4419 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4420 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4421 PseudoSourceValue::getGOT(), 0);
4423 // The address of the thread local variable is the add of the thread
4424 // pointer with the offset of the variable.
4425 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4429 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4430 // TODO: implement the "local dynamic" model
4431 // TODO: implement the "initial exec"model for pic executables
4432 assert(Subtarget->isTargetELF() &&
4433 "TLS not implemented for non-ELF targets");
4434 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4435 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4436 // otherwise use the "Local Exec"TLS Model
4437 if (Subtarget->is64Bit()) {
4438 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4440 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4441 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4443 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4448 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4449 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4450 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4451 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4452 // With PIC, the address is actually $g + Offset.
4453 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4454 !Subtarget->isPICStyleRIPRel()) {
4455 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4456 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4463 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4464 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4465 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4466 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4467 // With PIC, the address is actually $g + Offset.
4468 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4469 !Subtarget->isPICStyleRIPRel()) {
4470 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4471 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4478 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4479 /// take a 2 x i32 value to shift plus a shift amount.
4480 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4481 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4482 MVT VT = Op.getValueType();
4483 unsigned VTBits = VT.getSizeInBits();
4484 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4485 SDValue ShOpLo = Op.getOperand(0);
4486 SDValue ShOpHi = Op.getOperand(1);
4487 SDValue ShAmt = Op.getOperand(2);
4488 SDValue Tmp1 = isSRA ?
4489 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4490 DAG.getConstant(0, VT);
4493 if (Op.getOpcode() == ISD::SHL_PARTS) {
4494 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4495 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4497 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4498 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4501 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4502 DAG.getConstant(VTBits, MVT::i8));
4503 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4504 AndNode, DAG.getConstant(0, MVT::i8));
4507 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4508 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4509 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4511 if (Op.getOpcode() == ISD::SHL_PARTS) {
4512 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4513 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4515 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4516 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4519 SDValue Ops[2] = { Lo, Hi };
4520 return DAG.getMergeValues(Ops, 2);
4523 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4524 MVT SrcVT = Op.getOperand(0).getValueType();
4525 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4526 "Unknown SINT_TO_FP to lower!");
4528 // These are really Legal; caller falls through into that case.
4529 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4531 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4532 Subtarget->is64Bit())
4535 unsigned Size = SrcVT.getSizeInBits()/8;
4536 MachineFunction &MF = DAG.getMachineFunction();
4537 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4538 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4539 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4541 PseudoSourceValue::getFixedStack(SSFI), 0);
4545 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4547 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4549 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4550 SmallVector<SDValue, 8> Ops;
4551 Ops.push_back(Chain);
4552 Ops.push_back(StackSlot);
4553 Ops.push_back(DAG.getValueType(SrcVT));
4554 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4555 Tys, &Ops[0], Ops.size());
4558 Chain = Result.getValue(1);
4559 SDValue InFlag = Result.getValue(2);
4561 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4562 // shouldn't be necessary except that RFP cannot be live across
4563 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4564 MachineFunction &MF = DAG.getMachineFunction();
4565 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4566 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4567 Tys = DAG.getVTList(MVT::Other);
4568 SmallVector<SDValue, 8> Ops;
4569 Ops.push_back(Chain);
4570 Ops.push_back(Result);
4571 Ops.push_back(StackSlot);
4572 Ops.push_back(DAG.getValueType(Op.getValueType()));
4573 Ops.push_back(InFlag);
4574 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4575 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4576 PseudoSourceValue::getFixedStack(SSFI), 0);
4582 std::pair<SDValue,SDValue> X86TargetLowering::
4583 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4584 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4585 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4586 "Unknown FP_TO_SINT to lower!");
4588 // These are really Legal.
4589 if (Op.getValueType() == MVT::i32 &&
4590 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4591 return std::make_pair(SDValue(), SDValue());
4592 if (Subtarget->is64Bit() &&
4593 Op.getValueType() == MVT::i64 &&
4594 Op.getOperand(0).getValueType() != MVT::f80)
4595 return std::make_pair(SDValue(), SDValue());
4597 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4599 MachineFunction &MF = DAG.getMachineFunction();
4600 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4601 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4602 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4604 switch (Op.getValueType().getSimpleVT()) {
4605 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4606 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4607 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4608 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4611 SDValue Chain = DAG.getEntryNode();
4612 SDValue Value = Op.getOperand(0);
4613 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4614 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4615 Chain = DAG.getStore(Chain, Value, StackSlot,
4616 PseudoSourceValue::getFixedStack(SSFI), 0);
4617 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4619 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4621 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4622 Chain = Value.getValue(1);
4623 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4624 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4627 // Build the FP_TO_INT*_IN_MEM
4628 SDValue Ops[] = { Chain, Value, StackSlot };
4629 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4631 return std::make_pair(FIST, StackSlot);
4634 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4635 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4636 SDValue FIST = Vals.first, StackSlot = Vals.second;
4637 if (FIST.getNode() == 0) return SDValue();
4640 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4643 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4644 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4645 SDValue FIST = Vals.first, StackSlot = Vals.second;
4646 if (FIST.getNode() == 0) return 0;
4648 MVT VT = N->getValueType(0);
4650 // Return a load from the stack slot.
4651 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4653 // Use MERGE_VALUES to drop the chain result value and get a node with one
4654 // result. This requires turning off getMergeValues simplification, since
4655 // otherwise it will give us Res back.
4656 return DAG.getMergeValues(&Res, 1, false).getNode();
4659 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4660 MVT VT = Op.getValueType();
4663 EltVT = VT.getVectorElementType();
4664 std::vector<Constant*> CV;
4665 if (EltVT == MVT::f64) {
4666 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4670 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4676 Constant *C = ConstantVector::get(CV);
4677 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4678 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4679 PseudoSourceValue::getConstantPool(), 0,
4681 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4684 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4685 MVT VT = Op.getValueType();
4687 unsigned EltNum = 1;
4688 if (VT.isVector()) {
4689 EltVT = VT.getVectorElementType();
4690 EltNum = VT.getVectorNumElements();
4692 std::vector<Constant*> CV;
4693 if (EltVT == MVT::f64) {
4694 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4698 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4704 Constant *C = ConstantVector::get(CV);
4705 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4706 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4707 PseudoSourceValue::getConstantPool(), 0,
4709 if (VT.isVector()) {
4710 return DAG.getNode(ISD::BIT_CONVERT, VT,
4711 DAG.getNode(ISD::XOR, MVT::v2i64,
4712 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4713 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4715 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4719 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4720 SDValue Op0 = Op.getOperand(0);
4721 SDValue Op1 = Op.getOperand(1);
4722 MVT VT = Op.getValueType();
4723 MVT SrcVT = Op1.getValueType();
4725 // If second operand is smaller, extend it first.
4726 if (SrcVT.bitsLT(VT)) {
4727 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4730 // And if it is bigger, shrink it first.
4731 if (SrcVT.bitsGT(VT)) {
4732 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4736 // At this point the operands and the result should have the same
4737 // type, and that won't be f80 since that is not custom lowered.
4739 // First get the sign bit of second operand.
4740 std::vector<Constant*> CV;
4741 if (SrcVT == MVT::f64) {
4742 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4743 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4745 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4746 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4747 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4748 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4750 Constant *C = ConstantVector::get(CV);
4751 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4752 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4753 PseudoSourceValue::getConstantPool(), 0,
4755 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4757 // Shift sign bit right or left if the two operands have different types.
4758 if (SrcVT.bitsGT(VT)) {
4759 // Op0 is MVT::f32, Op1 is MVT::f64.
4760 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4761 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4762 DAG.getConstant(32, MVT::i32));
4763 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4764 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4765 DAG.getIntPtrConstant(0));
4768 // Clear first operand sign bit.
4770 if (VT == MVT::f64) {
4771 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4772 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4774 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4775 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4776 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4777 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4779 C = ConstantVector::get(CV);
4780 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4781 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4782 PseudoSourceValue::getConstantPool(), 0,
4784 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4786 // Or the value with the sign bit.
4787 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4790 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4791 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4793 SDValue Op0 = Op.getOperand(0);
4794 SDValue Op1 = Op.getOperand(1);
4795 SDValue CC = Op.getOperand(2);
4796 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4797 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4800 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4802 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4803 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4804 DAG.getConstant(X86CC, MVT::i8), Cond);
4807 assert(isFP && "Illegal integer SetCC!");
4809 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4810 switch (SetCCOpcode) {
4811 default: assert(false && "Illegal floating point SetCC!");
4812 case ISD::SETOEQ: { // !PF & ZF
4813 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4814 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4815 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4816 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4817 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4819 case ISD::SETUNE: { // PF | !ZF
4820 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4821 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4822 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4823 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4824 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4829 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4831 SDValue Op0 = Op.getOperand(0);
4832 SDValue Op1 = Op.getOperand(1);
4833 SDValue CC = Op.getOperand(2);
4834 MVT VT = Op.getValueType();
4835 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4836 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4840 MVT VT0 = Op0.getValueType();
4841 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4842 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4845 switch (SetCCOpcode) {
4848 case ISD::SETEQ: SSECC = 0; break;
4850 case ISD::SETGT: Swap = true; // Fallthrough
4852 case ISD::SETOLT: SSECC = 1; break;
4854 case ISD::SETGE: Swap = true; // Fallthrough
4856 case ISD::SETOLE: SSECC = 2; break;
4857 case ISD::SETUO: SSECC = 3; break;
4859 case ISD::SETNE: SSECC = 4; break;
4860 case ISD::SETULE: Swap = true;
4861 case ISD::SETUGE: SSECC = 5; break;
4862 case ISD::SETULT: Swap = true;
4863 case ISD::SETUGT: SSECC = 6; break;
4864 case ISD::SETO: SSECC = 7; break;
4867 std::swap(Op0, Op1);
4869 // In the two special cases we can't handle, emit two comparisons.
4871 if (SetCCOpcode == ISD::SETUEQ) {
4873 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4874 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4875 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4877 else if (SetCCOpcode == ISD::SETONE) {
4879 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4880 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4881 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4883 assert(0 && "Illegal FP comparison");
4885 // Handle all other FP comparisons here.
4886 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4889 // We are handling one of the integer comparisons here. Since SSE only has
4890 // GT and EQ comparisons for integer, swapping operands and multiple
4891 // operations may be required for some comparisons.
4892 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4893 bool Swap = false, Invert = false, FlipSigns = false;
4895 switch (VT.getSimpleVT()) {
4897 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4898 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4899 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4900 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4903 switch (SetCCOpcode) {
4905 case ISD::SETNE: Invert = true;
4906 case ISD::SETEQ: Opc = EQOpc; break;
4907 case ISD::SETLT: Swap = true;
4908 case ISD::SETGT: Opc = GTOpc; break;
4909 case ISD::SETGE: Swap = true;
4910 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4911 case ISD::SETULT: Swap = true;
4912 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4913 case ISD::SETUGE: Swap = true;
4914 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4917 std::swap(Op0, Op1);
4919 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4920 // bits of the inputs before performing those operations.
4922 MVT EltVT = VT.getVectorElementType();
4923 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4924 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4925 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4927 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
4928 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
4931 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
4933 // If the logical-not of the result is required, perform that now.
4935 MVT EltVT = VT.getVectorElementType();
4936 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
4937 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
4938 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
4940 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
4945 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
4946 bool addTest = true;
4947 SDValue Cond = Op.getOperand(0);
4950 if (Cond.getOpcode() == ISD::SETCC)
4951 Cond = LowerSETCC(Cond, DAG);
4953 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4954 // setting operand in place of the X86ISD::SETCC.
4955 if (Cond.getOpcode() == X86ISD::SETCC) {
4956 CC = Cond.getOperand(0);
4958 SDValue Cmp = Cond.getOperand(1);
4959 unsigned Opc = Cmp.getOpcode();
4960 MVT VT = Op.getValueType();
4962 bool IllegalFPCMov = false;
4963 if (VT.isFloatingPoint() && !VT.isVector() &&
4964 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4965 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4967 if ((Opc == X86ISD::CMP ||
4968 Opc == X86ISD::COMI ||
4969 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4976 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4977 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4980 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4982 SmallVector<SDValue, 4> Ops;
4983 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4984 // condition is true.
4985 Ops.push_back(Op.getOperand(2));
4986 Ops.push_back(Op.getOperand(1));
4988 Ops.push_back(Cond);
4989 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4992 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
4993 bool addTest = true;
4994 SDValue Chain = Op.getOperand(0);
4995 SDValue Cond = Op.getOperand(1);
4996 SDValue Dest = Op.getOperand(2);
4999 if (Cond.getOpcode() == ISD::SETCC)
5000 Cond = LowerSETCC(Cond, DAG);
5002 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5003 // setting operand in place of the X86ISD::SETCC.
5004 if (Cond.getOpcode() == X86ISD::SETCC) {
5005 CC = Cond.getOperand(0);
5007 SDValue Cmp = Cond.getOperand(1);
5008 unsigned Opc = Cmp.getOpcode();
5009 if (Opc == X86ISD::CMP ||
5010 Opc == X86ISD::COMI ||
5011 Opc == X86ISD::UCOMI) {
5018 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5019 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5021 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5022 Chain, Op.getOperand(2), CC, Cond);
5026 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5027 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5028 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5029 // that the guard pages used by the OS virtual memory manager are allocated in
5030 // correct sequence.
5032 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5033 SelectionDAG &DAG) {
5034 assert(Subtarget->isTargetCygMing() &&
5035 "This should be used only on Cygwin/Mingw targets");
5038 SDValue Chain = Op.getOperand(0);
5039 SDValue Size = Op.getOperand(1);
5040 // FIXME: Ensure alignment here
5044 MVT IntPtr = getPointerTy();
5045 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5047 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0));
5049 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5050 Flag = Chain.getValue(1);
5052 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5053 SDValue Ops[] = { Chain,
5054 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5055 DAG.getRegister(X86::EAX, IntPtr),
5056 DAG.getRegister(X86StackPtr, SPTy),
5058 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5059 Flag = Chain.getValue(1);
5061 Chain = DAG.getCALLSEQ_END(Chain,
5062 DAG.getIntPtrConstant(0),
5063 DAG.getIntPtrConstant(0),
5066 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5068 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5069 return DAG.getMergeValues(Ops1, 2);
5073 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5075 SDValue Dst, SDValue Src,
5076 SDValue Size, unsigned Align,
5077 const Value *DstSV, uint64_t DstSVOff) {
5078 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5080 /// If not DWORD aligned or size is more than the threshold, call the library.
5081 /// The libc version is likely to be faster for these cases. It can use the
5082 /// address value and run time information about the CPU.
5083 if ((Align & 3) != 0 ||
5085 ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
5086 SDValue InFlag(0, 0);
5088 // Check to see if there is a specialized entry-point for memory zeroing.
5089 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5090 if (const char *bzeroEntry =
5091 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5092 MVT IntPtr = getPointerTy();
5093 const Type *IntPtrTy = TD->getIntPtrType();
5094 TargetLowering::ArgListTy Args;
5095 TargetLowering::ArgListEntry Entry;
5097 Entry.Ty = IntPtrTy;
5098 Args.push_back(Entry);
5100 Args.push_back(Entry);
5101 std::pair<SDValue,SDValue> CallResult =
5102 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
5103 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
5105 return CallResult.second;
5108 // Otherwise have the target-independent code call memset.
5112 uint64_t SizeVal = ConstantSize->getValue();
5113 SDValue InFlag(0, 0);
5116 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5117 unsigned BytesLeft = 0;
5118 bool TwoRepStos = false;
5121 uint64_t Val = ValC->getValue() & 255;
5123 // If the value is a constant, then we can potentially use larger sets.
5124 switch (Align & 3) {
5125 case 2: // WORD aligned
5128 Val = (Val << 8) | Val;
5130 case 0: // DWORD aligned
5133 Val = (Val << 8) | Val;
5134 Val = (Val << 16) | Val;
5135 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5138 Val = (Val << 32) | Val;
5141 default: // Byte aligned
5144 Count = DAG.getIntPtrConstant(SizeVal);
5148 if (AVT.bitsGT(MVT::i8)) {
5149 unsigned UBytes = AVT.getSizeInBits() / 8;
5150 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5151 BytesLeft = SizeVal % UBytes;
5154 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5156 InFlag = Chain.getValue(1);
5159 Count = DAG.getIntPtrConstant(SizeVal);
5160 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5161 InFlag = Chain.getValue(1);
5164 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5166 InFlag = Chain.getValue(1);
5167 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5169 InFlag = Chain.getValue(1);
5171 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5172 SmallVector<SDValue, 8> Ops;
5173 Ops.push_back(Chain);
5174 Ops.push_back(DAG.getValueType(AVT));
5175 Ops.push_back(InFlag);
5176 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5179 InFlag = Chain.getValue(1);
5181 MVT CVT = Count.getValueType();
5182 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5183 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5184 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5186 InFlag = Chain.getValue(1);
5187 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5189 Ops.push_back(Chain);
5190 Ops.push_back(DAG.getValueType(MVT::i8));
5191 Ops.push_back(InFlag);
5192 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5193 } else if (BytesLeft) {
5194 // Handle the last 1 - 7 bytes.
5195 unsigned Offset = SizeVal - BytesLeft;
5196 MVT AddrVT = Dst.getValueType();
5197 MVT SizeVT = Size.getValueType();
5199 Chain = DAG.getMemset(Chain,
5200 DAG.getNode(ISD::ADD, AddrVT, Dst,
5201 DAG.getConstant(Offset, AddrVT)),
5203 DAG.getConstant(BytesLeft, SizeVT),
5204 Align, DstSV, DstSVOff + Offset);
5207 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5212 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5213 SDValue Chain, SDValue Dst, SDValue Src,
5214 SDValue Size, unsigned Align,
5216 const Value *DstSV, uint64_t DstSVOff,
5217 const Value *SrcSV, uint64_t SrcSVOff) {
5218 // This requires the copy size to be a constant, preferrably
5219 // within a subtarget-specific limit.
5220 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5223 uint64_t SizeVal = ConstantSize->getValue();
5224 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5227 /// If not DWORD aligned, call the library.
5228 if ((Align & 3) != 0)
5233 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5236 unsigned UBytes = AVT.getSizeInBits() / 8;
5237 unsigned CountVal = SizeVal / UBytes;
5238 SDValue Count = DAG.getIntPtrConstant(CountVal);
5239 unsigned BytesLeft = SizeVal % UBytes;
5241 SDValue InFlag(0, 0);
5242 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5244 InFlag = Chain.getValue(1);
5245 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5247 InFlag = Chain.getValue(1);
5248 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5250 InFlag = Chain.getValue(1);
5252 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5253 SmallVector<SDValue, 8> Ops;
5254 Ops.push_back(Chain);
5255 Ops.push_back(DAG.getValueType(AVT));
5256 Ops.push_back(InFlag);
5257 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5259 SmallVector<SDValue, 4> Results;
5260 Results.push_back(RepMovs);
5262 // Handle the last 1 - 7 bytes.
5263 unsigned Offset = SizeVal - BytesLeft;
5264 MVT DstVT = Dst.getValueType();
5265 MVT SrcVT = Src.getValueType();
5266 MVT SizeVT = Size.getValueType();
5267 Results.push_back(DAG.getMemcpy(Chain,
5268 DAG.getNode(ISD::ADD, DstVT, Dst,
5269 DAG.getConstant(Offset, DstVT)),
5270 DAG.getNode(ISD::ADD, SrcVT, Src,
5271 DAG.getConstant(Offset, SrcVT)),
5272 DAG.getConstant(BytesLeft, SizeVT),
5273 Align, AlwaysInline,
5274 DstSV, DstSVOff + Offset,
5275 SrcSV, SrcSVOff + Offset));
5278 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5281 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5282 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5283 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5284 SDValue TheChain = N->getOperand(0);
5285 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5286 if (Subtarget->is64Bit()) {
5287 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5288 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5289 MVT::i64, rax.getValue(2));
5290 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5291 DAG.getConstant(32, MVT::i8));
5293 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5296 return DAG.getMergeValues(Ops, 2).getNode();
5299 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5300 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5301 MVT::i32, eax.getValue(2));
5302 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5303 SDValue Ops[] = { eax, edx };
5304 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5306 // Use a MERGE_VALUES to return the value and chain.
5307 Ops[1] = edx.getValue(1);
5308 return DAG.getMergeValues(Ops, 2).getNode();
5311 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5312 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5314 if (!Subtarget->is64Bit()) {
5315 // vastart just stores the address of the VarArgsFrameIndex slot into the
5316 // memory location argument.
5317 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5318 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5322 // gp_offset (0 - 6 * 8)
5323 // fp_offset (48 - 48 + 8 * 16)
5324 // overflow_arg_area (point to parameters coming in memory).
5326 SmallVector<SDValue, 8> MemOps;
5327 SDValue FIN = Op.getOperand(1);
5329 SDValue Store = DAG.getStore(Op.getOperand(0),
5330 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5332 MemOps.push_back(Store);
5335 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5336 Store = DAG.getStore(Op.getOperand(0),
5337 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5339 MemOps.push_back(Store);
5341 // Store ptr to overflow_arg_area
5342 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5343 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5344 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5345 MemOps.push_back(Store);
5347 // Store ptr to reg_save_area.
5348 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5349 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5350 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5351 MemOps.push_back(Store);
5352 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5355 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5356 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5357 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5358 SDValue Chain = Op.getOperand(0);
5359 SDValue SrcPtr = Op.getOperand(1);
5360 SDValue SrcSV = Op.getOperand(2);
5362 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5367 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5368 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5369 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5370 SDValue Chain = Op.getOperand(0);
5371 SDValue DstPtr = Op.getOperand(1);
5372 SDValue SrcPtr = Op.getOperand(2);
5373 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5374 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5376 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5377 DAG.getIntPtrConstant(24), 8, false,
5378 DstSV, 0, SrcSV, 0);
5382 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5383 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5385 default: return SDValue(); // Don't custom lower most intrinsics.
5386 // Comparison intrinsics.
5387 case Intrinsic::x86_sse_comieq_ss:
5388 case Intrinsic::x86_sse_comilt_ss:
5389 case Intrinsic::x86_sse_comile_ss:
5390 case Intrinsic::x86_sse_comigt_ss:
5391 case Intrinsic::x86_sse_comige_ss:
5392 case Intrinsic::x86_sse_comineq_ss:
5393 case Intrinsic::x86_sse_ucomieq_ss:
5394 case Intrinsic::x86_sse_ucomilt_ss:
5395 case Intrinsic::x86_sse_ucomile_ss:
5396 case Intrinsic::x86_sse_ucomigt_ss:
5397 case Intrinsic::x86_sse_ucomige_ss:
5398 case Intrinsic::x86_sse_ucomineq_ss:
5399 case Intrinsic::x86_sse2_comieq_sd:
5400 case Intrinsic::x86_sse2_comilt_sd:
5401 case Intrinsic::x86_sse2_comile_sd:
5402 case Intrinsic::x86_sse2_comigt_sd:
5403 case Intrinsic::x86_sse2_comige_sd:
5404 case Intrinsic::x86_sse2_comineq_sd:
5405 case Intrinsic::x86_sse2_ucomieq_sd:
5406 case Intrinsic::x86_sse2_ucomilt_sd:
5407 case Intrinsic::x86_sse2_ucomile_sd:
5408 case Intrinsic::x86_sse2_ucomigt_sd:
5409 case Intrinsic::x86_sse2_ucomige_sd:
5410 case Intrinsic::x86_sse2_ucomineq_sd: {
5412 ISD::CondCode CC = ISD::SETCC_INVALID;
5415 case Intrinsic::x86_sse_comieq_ss:
5416 case Intrinsic::x86_sse2_comieq_sd:
5420 case Intrinsic::x86_sse_comilt_ss:
5421 case Intrinsic::x86_sse2_comilt_sd:
5425 case Intrinsic::x86_sse_comile_ss:
5426 case Intrinsic::x86_sse2_comile_sd:
5430 case Intrinsic::x86_sse_comigt_ss:
5431 case Intrinsic::x86_sse2_comigt_sd:
5435 case Intrinsic::x86_sse_comige_ss:
5436 case Intrinsic::x86_sse2_comige_sd:
5440 case Intrinsic::x86_sse_comineq_ss:
5441 case Intrinsic::x86_sse2_comineq_sd:
5445 case Intrinsic::x86_sse_ucomieq_ss:
5446 case Intrinsic::x86_sse2_ucomieq_sd:
5447 Opc = X86ISD::UCOMI;
5450 case Intrinsic::x86_sse_ucomilt_ss:
5451 case Intrinsic::x86_sse2_ucomilt_sd:
5452 Opc = X86ISD::UCOMI;
5455 case Intrinsic::x86_sse_ucomile_ss:
5456 case Intrinsic::x86_sse2_ucomile_sd:
5457 Opc = X86ISD::UCOMI;
5460 case Intrinsic::x86_sse_ucomigt_ss:
5461 case Intrinsic::x86_sse2_ucomigt_sd:
5462 Opc = X86ISD::UCOMI;
5465 case Intrinsic::x86_sse_ucomige_ss:
5466 case Intrinsic::x86_sse2_ucomige_sd:
5467 Opc = X86ISD::UCOMI;
5470 case Intrinsic::x86_sse_ucomineq_ss:
5471 case Intrinsic::x86_sse2_ucomineq_sd:
5472 Opc = X86ISD::UCOMI;
5478 SDValue LHS = Op.getOperand(1);
5479 SDValue RHS = Op.getOperand(2);
5480 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5482 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5483 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5484 DAG.getConstant(X86CC, MVT::i8), Cond);
5485 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5488 // Fix vector shift instructions where the last operand is a non-immediate
5490 case Intrinsic::x86_sse2_pslli_w:
5491 case Intrinsic::x86_sse2_pslli_d:
5492 case Intrinsic::x86_sse2_pslli_q:
5493 case Intrinsic::x86_sse2_psrli_w:
5494 case Intrinsic::x86_sse2_psrli_d:
5495 case Intrinsic::x86_sse2_psrli_q:
5496 case Intrinsic::x86_sse2_psrai_w:
5497 case Intrinsic::x86_sse2_psrai_d:
5498 case Intrinsic::x86_mmx_pslli_w:
5499 case Intrinsic::x86_mmx_pslli_d:
5500 case Intrinsic::x86_mmx_pslli_q:
5501 case Intrinsic::x86_mmx_psrli_w:
5502 case Intrinsic::x86_mmx_psrli_d:
5503 case Intrinsic::x86_mmx_psrli_q:
5504 case Intrinsic::x86_mmx_psrai_w:
5505 case Intrinsic::x86_mmx_psrai_d: {
5506 SDValue ShAmt = Op.getOperand(2);
5507 if (isa<ConstantSDNode>(ShAmt))
5510 unsigned NewIntNo = 0;
5511 MVT ShAmtVT = MVT::v4i32;
5513 case Intrinsic::x86_sse2_pslli_w:
5514 NewIntNo = Intrinsic::x86_sse2_psll_w;
5516 case Intrinsic::x86_sse2_pslli_d:
5517 NewIntNo = Intrinsic::x86_sse2_psll_d;
5519 case Intrinsic::x86_sse2_pslli_q:
5520 NewIntNo = Intrinsic::x86_sse2_psll_q;
5522 case Intrinsic::x86_sse2_psrli_w:
5523 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5525 case Intrinsic::x86_sse2_psrli_d:
5526 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5528 case Intrinsic::x86_sse2_psrli_q:
5529 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5531 case Intrinsic::x86_sse2_psrai_w:
5532 NewIntNo = Intrinsic::x86_sse2_psra_w;
5534 case Intrinsic::x86_sse2_psrai_d:
5535 NewIntNo = Intrinsic::x86_sse2_psra_d;
5538 ShAmtVT = MVT::v2i32;
5540 case Intrinsic::x86_mmx_pslli_w:
5541 NewIntNo = Intrinsic::x86_mmx_psll_w;
5543 case Intrinsic::x86_mmx_pslli_d:
5544 NewIntNo = Intrinsic::x86_mmx_psll_d;
5546 case Intrinsic::x86_mmx_pslli_q:
5547 NewIntNo = Intrinsic::x86_mmx_psll_q;
5549 case Intrinsic::x86_mmx_psrli_w:
5550 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5552 case Intrinsic::x86_mmx_psrli_d:
5553 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5555 case Intrinsic::x86_mmx_psrli_q:
5556 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5558 case Intrinsic::x86_mmx_psrai_w:
5559 NewIntNo = Intrinsic::x86_mmx_psra_w;
5561 case Intrinsic::x86_mmx_psrai_d:
5562 NewIntNo = Intrinsic::x86_mmx_psra_d;
5564 default: abort(); // Can't reach here.
5569 MVT VT = Op.getValueType();
5570 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5571 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5572 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5573 DAG.getConstant(NewIntNo, MVT::i32),
5574 Op.getOperand(1), ShAmt);
5579 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5580 // Depths > 0 not supported yet!
5581 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5584 // Just load the return address
5585 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5586 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5589 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5590 // Depths > 0 not supported yet!
5591 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5594 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5595 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5596 DAG.getIntPtrConstant(TD->getPointerSize()));
5599 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5600 SelectionDAG &DAG) {
5601 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5604 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5606 MachineFunction &MF = DAG.getMachineFunction();
5607 SDValue Chain = Op.getOperand(0);
5608 SDValue Offset = Op.getOperand(1);
5609 SDValue Handler = Op.getOperand(2);
5611 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5613 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5615 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5616 DAG.getIntPtrConstant(-TD->getPointerSize()));
5617 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5618 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5619 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5620 MF.getRegInfo().addLiveOut(StoreAddrReg);
5622 return DAG.getNode(X86ISD::EH_RETURN,
5624 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5627 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5628 SelectionDAG &DAG) {
5629 SDValue Root = Op.getOperand(0);
5630 SDValue Trmp = Op.getOperand(1); // trampoline
5631 SDValue FPtr = Op.getOperand(2); // nested function
5632 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5634 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5636 const X86InstrInfo *TII =
5637 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5639 if (Subtarget->is64Bit()) {
5640 SDValue OutChains[6];
5642 // Large code-model.
5644 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5645 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5647 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5648 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5650 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5652 // Load the pointer to the nested function into R11.
5653 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5654 SDValue Addr = Trmp;
5655 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5658 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5659 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5661 // Load the 'nest' parameter value into R10.
5662 // R10 is specified in X86CallingConv.td
5663 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5664 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5665 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5668 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5669 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5671 // Jump to the nested function.
5672 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5673 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5674 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5677 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5678 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5679 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5683 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5684 return DAG.getMergeValues(Ops, 2);
5686 const Function *Func =
5687 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5688 unsigned CC = Func->getCallingConv();
5693 assert(0 && "Unsupported calling convention");
5694 case CallingConv::C:
5695 case CallingConv::X86_StdCall: {
5696 // Pass 'nest' parameter in ECX.
5697 // Must be kept in sync with X86CallingConv.td
5700 // Check that ECX wasn't needed by an 'inreg' parameter.
5701 const FunctionType *FTy = Func->getFunctionType();
5702 const PAListPtr &Attrs = Func->getParamAttrs();
5704 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5705 unsigned InRegCount = 0;
5708 for (FunctionType::param_iterator I = FTy->param_begin(),
5709 E = FTy->param_end(); I != E; ++I, ++Idx)
5710 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5711 // FIXME: should only count parameters that are lowered to integers.
5712 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
5714 if (InRegCount > 2) {
5715 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5721 case CallingConv::X86_FastCall:
5722 // Pass 'nest' parameter in EAX.
5723 // Must be kept in sync with X86CallingConv.td
5728 SDValue OutChains[4];
5731 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5732 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5734 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5735 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5736 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5739 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5740 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5742 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5743 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5744 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5745 TrmpAddr, 5, false, 1);
5747 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5748 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5751 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5752 return DAG.getMergeValues(Ops, 2);
5756 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5758 The rounding mode is in bits 11:10 of FPSR, and has the following
5765 FLT_ROUNDS, on the other hand, expects the following:
5772 To perform the conversion, we do:
5773 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5776 MachineFunction &MF = DAG.getMachineFunction();
5777 const TargetMachine &TM = MF.getTarget();
5778 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5779 unsigned StackAlignment = TFI.getStackAlignment();
5780 MVT VT = Op.getValueType();
5782 // Save FP Control Word to stack slot
5783 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5784 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5786 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5787 DAG.getEntryNode(), StackSlot);
5789 // Load FP Control Word from stack slot
5790 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5792 // Transform as necessary
5794 DAG.getNode(ISD::SRL, MVT::i16,
5795 DAG.getNode(ISD::AND, MVT::i16,
5796 CWD, DAG.getConstant(0x800, MVT::i16)),
5797 DAG.getConstant(11, MVT::i8));
5799 DAG.getNode(ISD::SRL, MVT::i16,
5800 DAG.getNode(ISD::AND, MVT::i16,
5801 CWD, DAG.getConstant(0x400, MVT::i16)),
5802 DAG.getConstant(9, MVT::i8));
5805 DAG.getNode(ISD::AND, MVT::i16,
5806 DAG.getNode(ISD::ADD, MVT::i16,
5807 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5808 DAG.getConstant(1, MVT::i16)),
5809 DAG.getConstant(3, MVT::i16));
5812 return DAG.getNode((VT.getSizeInBits() < 16 ?
5813 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5816 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5817 MVT VT = Op.getValueType();
5819 unsigned NumBits = VT.getSizeInBits();
5821 Op = Op.getOperand(0);
5822 if (VT == MVT::i8) {
5823 // Zero extend to i32 since there is not an i8 bsr.
5825 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5828 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5829 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5830 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5832 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5833 SmallVector<SDValue, 4> Ops;
5835 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5836 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5837 Ops.push_back(Op.getValue(1));
5838 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5840 // Finally xor with NumBits-1.
5841 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5844 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5848 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5849 MVT VT = Op.getValueType();
5851 unsigned NumBits = VT.getSizeInBits();
5853 Op = Op.getOperand(0);
5854 if (VT == MVT::i8) {
5856 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5859 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5860 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5861 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5863 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5864 SmallVector<SDValue, 4> Ops;
5866 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5867 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5868 Ops.push_back(Op.getValue(1));
5869 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5872 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5876 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5877 MVT T = Op.getValueType();
5880 switch(T.getSimpleVT()) {
5882 assert(false && "Invalid value type!");
5883 case MVT::i8: Reg = X86::AL; size = 1; break;
5884 case MVT::i16: Reg = X86::AX; size = 2; break;
5885 case MVT::i32: Reg = X86::EAX; size = 4; break;
5887 if (Subtarget->is64Bit()) {
5888 Reg = X86::RAX; size = 8;
5889 } else //Should go away when LowerType stuff lands
5890 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5893 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5894 Op.getOperand(3), SDValue());
5895 SDValue Ops[] = { cpIn.getValue(0),
5898 DAG.getTargetConstant(size, MVT::i8),
5900 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5901 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5903 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5907 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5908 SelectionDAG &DAG) {
5909 MVT T = Op->getValueType(0);
5910 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5911 SDValue cpInL, cpInH;
5912 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5913 DAG.getConstant(0, MVT::i32));
5914 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5915 DAG.getConstant(1, MVT::i32));
5916 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5918 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5919 cpInH, cpInL.getValue(1));
5920 SDValue swapInL, swapInH;
5921 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5922 DAG.getConstant(0, MVT::i32));
5923 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5924 DAG.getConstant(1, MVT::i32));
5925 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5926 swapInL, cpInH.getValue(1));
5927 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5928 swapInH, swapInL.getValue(1));
5929 SDValue Ops[] = { swapInH.getValue(0),
5931 swapInH.getValue(1)};
5932 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5933 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5934 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5935 Result.getValue(1));
5936 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5937 cpOutL.getValue(2));
5938 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5939 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5940 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
5941 return DAG.getMergeValues(Vals, 2).getNode();
5944 SDNode* X86TargetLowering::ExpandATOMIC_LOAD_SUB(SDNode* Op,
5945 SelectionDAG &DAG) {
5946 MVT T = Op->getValueType(0);
5947 SDValue negOp = DAG.getNode(ISD::SUB, T,
5948 DAG.getConstant(0, T), Op->getOperand(2));
5949 return DAG.getAtomic((T==MVT::i8 ? ISD::ATOMIC_LOAD_ADD_8:
5950 T==MVT::i16 ? ISD::ATOMIC_LOAD_ADD_16:
5951 T==MVT::i32 ? ISD::ATOMIC_LOAD_ADD_32:
5952 T==MVT::i64 ? ISD::ATOMIC_LOAD_ADD_64: 0),
5953 Op->getOperand(0), Op->getOperand(1), negOp,
5954 cast<AtomicSDNode>(Op)->getSrcValue(),
5955 cast<AtomicSDNode>(Op)->getAlignment()).getNode();
5958 /// LowerOperation - Provide custom lowering hooks for some operations.
5960 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5961 switch (Op.getOpcode()) {
5962 default: assert(0 && "Should not custom lower this!");
5963 case ISD::ATOMIC_CMP_SWAP_8: return LowerCMP_SWAP(Op,DAG);
5964 case ISD::ATOMIC_CMP_SWAP_16: return LowerCMP_SWAP(Op,DAG);
5965 case ISD::ATOMIC_CMP_SWAP_32: return LowerCMP_SWAP(Op,DAG);
5966 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
5967 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5968 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5969 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5970 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5971 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5972 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5973 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5974 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5975 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5976 case ISD::SHL_PARTS:
5977 case ISD::SRA_PARTS:
5978 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5979 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5980 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5981 case ISD::FABS: return LowerFABS(Op, DAG);
5982 case ISD::FNEG: return LowerFNEG(Op, DAG);
5983 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5984 case ISD::SETCC: return LowerSETCC(Op, DAG);
5985 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
5986 case ISD::SELECT: return LowerSELECT(Op, DAG);
5987 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5988 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5989 case ISD::CALL: return LowerCALL(Op, DAG);
5990 case ISD::RET: return LowerRET(Op, DAG);
5991 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5992 case ISD::VASTART: return LowerVASTART(Op, DAG);
5993 case ISD::VAARG: return LowerVAARG(Op, DAG);
5994 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5995 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5996 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5997 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5998 case ISD::FRAME_TO_ARGS_OFFSET:
5999 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6000 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6001 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6002 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6003 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6004 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6005 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6007 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
6008 case ISD::READCYCLECOUNTER:
6009 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
6013 /// ReplaceNodeResults - Replace a node with an illegal result type
6014 /// with a new node built out of custom code.
6015 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
6016 switch (N->getOpcode()) {
6017 default: assert(0 && "Should not custom lower this!");
6018 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
6019 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
6020 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
6021 case ISD::ATOMIC_LOAD_SUB_8: return ExpandATOMIC_LOAD_SUB(N,DAG);
6022 case ISD::ATOMIC_LOAD_SUB_16: return ExpandATOMIC_LOAD_SUB(N,DAG);
6023 case ISD::ATOMIC_LOAD_SUB_32: return ExpandATOMIC_LOAD_SUB(N,DAG);
6024 case ISD::ATOMIC_LOAD_SUB_64: return ExpandATOMIC_LOAD_SUB(N,DAG);
6028 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6030 default: return NULL;
6031 case X86ISD::BSF: return "X86ISD::BSF";
6032 case X86ISD::BSR: return "X86ISD::BSR";
6033 case X86ISD::SHLD: return "X86ISD::SHLD";
6034 case X86ISD::SHRD: return "X86ISD::SHRD";
6035 case X86ISD::FAND: return "X86ISD::FAND";
6036 case X86ISD::FOR: return "X86ISD::FOR";
6037 case X86ISD::FXOR: return "X86ISD::FXOR";
6038 case X86ISD::FSRL: return "X86ISD::FSRL";
6039 case X86ISD::FILD: return "X86ISD::FILD";
6040 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6041 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6042 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6043 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6044 case X86ISD::FLD: return "X86ISD::FLD";
6045 case X86ISD::FST: return "X86ISD::FST";
6046 case X86ISD::CALL: return "X86ISD::CALL";
6047 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6048 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6049 case X86ISD::CMP: return "X86ISD::CMP";
6050 case X86ISD::COMI: return "X86ISD::COMI";
6051 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6052 case X86ISD::SETCC: return "X86ISD::SETCC";
6053 case X86ISD::CMOV: return "X86ISD::CMOV";
6054 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6055 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6056 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6057 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6058 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6059 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6060 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6061 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6062 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6063 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6064 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6065 case X86ISD::FMAX: return "X86ISD::FMAX";
6066 case X86ISD::FMIN: return "X86ISD::FMIN";
6067 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6068 case X86ISD::FRCP: return "X86ISD::FRCP";
6069 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6070 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6071 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6072 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6073 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6074 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6075 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6076 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6077 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6078 case X86ISD::VSHL: return "X86ISD::VSHL";
6079 case X86ISD::VSRL: return "X86ISD::VSRL";
6080 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6081 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6082 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6083 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6084 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6085 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6086 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6087 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6088 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6089 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6093 // isLegalAddressingMode - Return true if the addressing mode represented
6094 // by AM is legal for this target, for a load/store of the specified type.
6095 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6096 const Type *Ty) const {
6097 // X86 supports extremely general addressing modes.
6099 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6100 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6104 // We can only fold this if we don't need an extra load.
6105 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6108 // X86-64 only supports addr of globals in small code model.
6109 if (Subtarget->is64Bit()) {
6110 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6112 // If lower 4G is not available, then we must use rip-relative addressing.
6113 if (AM.BaseOffs || AM.Scale > 1)
6124 // These scales always work.
6129 // These scales are formed with basereg+scalereg. Only accept if there is
6134 default: // Other stuff never works.
6142 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6143 if (!Ty1->isInteger() || !Ty2->isInteger())
6145 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6146 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6147 if (NumBits1 <= NumBits2)
6149 return Subtarget->is64Bit() || NumBits1 < 64;
6152 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6153 if (!VT1.isInteger() || !VT2.isInteger())
6155 unsigned NumBits1 = VT1.getSizeInBits();
6156 unsigned NumBits2 = VT2.getSizeInBits();
6157 if (NumBits1 <= NumBits2)
6159 return Subtarget->is64Bit() || NumBits1 < 64;
6162 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6163 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6164 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6165 /// are assumed to be legal.
6167 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6168 // Only do shuffles on 128-bit vector types for now.
6169 if (VT.getSizeInBits() == 64) return false;
6170 return (Mask.getNode()->getNumOperands() <= 4 ||
6171 isIdentityMask(Mask.getNode()) ||
6172 isIdentityMask(Mask.getNode(), true) ||
6173 isSplatMask(Mask.getNode()) ||
6174 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6175 X86::isUNPCKLMask(Mask.getNode()) ||
6176 X86::isUNPCKHMask(Mask.getNode()) ||
6177 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6178 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6182 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6183 MVT EVT, SelectionDAG &DAG) const {
6184 unsigned NumElts = BVOps.size();
6185 // Only do shuffles on 128-bit vector types for now.
6186 if (EVT.getSizeInBits() * NumElts == 64) return false;
6187 if (NumElts == 2) return true;
6189 return (isMOVLMask(&BVOps[0], 4) ||
6190 isCommutedMOVL(&BVOps[0], 4, true) ||
6191 isSHUFPMask(&BVOps[0], 4) ||
6192 isCommutedSHUFP(&BVOps[0], 4));
6197 //===----------------------------------------------------------------------===//
6198 // X86 Scheduler Hooks
6199 //===----------------------------------------------------------------------===//
6201 // private utility function
6203 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6204 MachineBasicBlock *MBB,
6212 TargetRegisterClass *RC,
6214 // For the atomic bitwise operator, we generate
6217 // ld t1 = [bitinstr.addr]
6218 // op t2 = t1, [bitinstr.val]
6220 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6222 // fallthrough -->nextMBB
6223 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6224 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6225 MachineFunction::iterator MBBIter = MBB;
6228 /// First build the CFG
6229 MachineFunction *F = MBB->getParent();
6230 MachineBasicBlock *thisMBB = MBB;
6231 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6232 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6233 F->insert(MBBIter, newMBB);
6234 F->insert(MBBIter, nextMBB);
6236 // Move all successors to thisMBB to nextMBB
6237 nextMBB->transferSuccessors(thisMBB);
6239 // Update thisMBB to fall through to newMBB
6240 thisMBB->addSuccessor(newMBB);
6242 // newMBB jumps to itself and fall through to nextMBB
6243 newMBB->addSuccessor(nextMBB);
6244 newMBB->addSuccessor(newMBB);
6246 // Insert instructions into newMBB based on incoming instruction
6247 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6248 MachineOperand& destOper = bInstr->getOperand(0);
6249 MachineOperand* argOpers[6];
6250 int numArgs = bInstr->getNumOperands() - 1;
6251 for (int i=0; i < numArgs; ++i)
6252 argOpers[i] = &bInstr->getOperand(i+1);
6254 // x86 address has 4 operands: base, index, scale, and displacement
6255 int lastAddrIndx = 3; // [0,3]
6258 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6259 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6260 for (int i=0; i <= lastAddrIndx; ++i)
6261 (*MIB).addOperand(*argOpers[i]);
6263 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6265 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6270 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6271 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6272 && "invalid operand");
6273 if (argOpers[valArgIndx]->isReg())
6274 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6276 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6278 (*MIB).addOperand(*argOpers[valArgIndx]);
6280 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6283 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6284 for (int i=0; i <= lastAddrIndx; ++i)
6285 (*MIB).addOperand(*argOpers[i]);
6287 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6288 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6290 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6294 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6296 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6300 // private utility function
6302 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6303 MachineBasicBlock *MBB,
6305 // For the atomic min/max operator, we generate
6308 // ld t1 = [min/max.addr]
6309 // mov t2 = [min/max.val]
6311 // cmov[cond] t2 = t1
6313 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6315 // fallthrough -->nextMBB
6317 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6318 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6319 MachineFunction::iterator MBBIter = MBB;
6322 /// First build the CFG
6323 MachineFunction *F = MBB->getParent();
6324 MachineBasicBlock *thisMBB = MBB;
6325 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6326 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6327 F->insert(MBBIter, newMBB);
6328 F->insert(MBBIter, nextMBB);
6330 // Move all successors to thisMBB to nextMBB
6331 nextMBB->transferSuccessors(thisMBB);
6333 // Update thisMBB to fall through to newMBB
6334 thisMBB->addSuccessor(newMBB);
6336 // newMBB jumps to newMBB and fall through to nextMBB
6337 newMBB->addSuccessor(nextMBB);
6338 newMBB->addSuccessor(newMBB);
6340 // Insert instructions into newMBB based on incoming instruction
6341 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6342 MachineOperand& destOper = mInstr->getOperand(0);
6343 MachineOperand* argOpers[6];
6344 int numArgs = mInstr->getNumOperands() - 1;
6345 for (int i=0; i < numArgs; ++i)
6346 argOpers[i] = &mInstr->getOperand(i+1);
6348 // x86 address has 4 operands: base, index, scale, and displacement
6349 int lastAddrIndx = 3; // [0,3]
6352 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6353 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6354 for (int i=0; i <= lastAddrIndx; ++i)
6355 (*MIB).addOperand(*argOpers[i]);
6357 // We only support register and immediate values
6358 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6359 && "invalid operand");
6361 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6362 if (argOpers[valArgIndx]->isReg())
6363 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6365 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6366 (*MIB).addOperand(*argOpers[valArgIndx]);
6368 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6371 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6376 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6377 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6381 // Cmp and exchange if none has modified the memory location
6382 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6383 for (int i=0; i <= lastAddrIndx; ++i)
6384 (*MIB).addOperand(*argOpers[i]);
6386 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6387 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6389 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6390 MIB.addReg(X86::EAX);
6393 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6395 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6401 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6402 MachineBasicBlock *BB) {
6403 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6404 switch (MI->getOpcode()) {
6405 default: assert(false && "Unexpected instr type to insert");
6406 case X86::CMOV_FR32:
6407 case X86::CMOV_FR64:
6408 case X86::CMOV_V4F32:
6409 case X86::CMOV_V2F64:
6410 case X86::CMOV_V2I64: {
6411 // To "insert" a SELECT_CC instruction, we actually have to insert the
6412 // diamond control-flow pattern. The incoming instruction knows the
6413 // destination vreg to set, the condition code register to branch on, the
6414 // true/false values to select between, and a branch opcode to use.
6415 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6416 MachineFunction::iterator It = BB;
6422 // cmpTY ccX, r1, r2
6424 // fallthrough --> copy0MBB
6425 MachineBasicBlock *thisMBB = BB;
6426 MachineFunction *F = BB->getParent();
6427 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6428 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6430 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6431 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6432 F->insert(It, copy0MBB);
6433 F->insert(It, sinkMBB);
6434 // Update machine-CFG edges by transferring all successors of the current
6435 // block to the new block which will contain the Phi node for the select.
6436 sinkMBB->transferSuccessors(BB);
6438 // Add the true and fallthrough blocks as its successors.
6439 BB->addSuccessor(copy0MBB);
6440 BB->addSuccessor(sinkMBB);
6443 // %FalseValue = ...
6444 // # fallthrough to sinkMBB
6447 // Update machine-CFG edges
6448 BB->addSuccessor(sinkMBB);
6451 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6454 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6455 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6456 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6458 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6462 case X86::FP32_TO_INT16_IN_MEM:
6463 case X86::FP32_TO_INT32_IN_MEM:
6464 case X86::FP32_TO_INT64_IN_MEM:
6465 case X86::FP64_TO_INT16_IN_MEM:
6466 case X86::FP64_TO_INT32_IN_MEM:
6467 case X86::FP64_TO_INT64_IN_MEM:
6468 case X86::FP80_TO_INT16_IN_MEM:
6469 case X86::FP80_TO_INT32_IN_MEM:
6470 case X86::FP80_TO_INT64_IN_MEM: {
6471 // Change the floating point control register to use "round towards zero"
6472 // mode when truncating to an integer value.
6473 MachineFunction *F = BB->getParent();
6474 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6475 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6477 // Load the old value of the high byte of the control word...
6479 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6480 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6482 // Set the high part to be round to zero...
6483 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6486 // Reload the modified control word now...
6487 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6489 // Restore the memory image of control word to original value
6490 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6493 // Get the X86 opcode to use.
6495 switch (MI->getOpcode()) {
6496 default: assert(0 && "illegal opcode!");
6497 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6498 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6499 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6500 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6501 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6502 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6503 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6504 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6505 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6509 MachineOperand &Op = MI->getOperand(0);
6510 if (Op.isRegister()) {
6511 AM.BaseType = X86AddressMode::RegBase;
6512 AM.Base.Reg = Op.getReg();
6514 AM.BaseType = X86AddressMode::FrameIndexBase;
6515 AM.Base.FrameIndex = Op.getIndex();
6517 Op = MI->getOperand(1);
6518 if (Op.isImmediate())
6519 AM.Scale = Op.getImm();
6520 Op = MI->getOperand(2);
6521 if (Op.isImmediate())
6522 AM.IndexReg = Op.getImm();
6523 Op = MI->getOperand(3);
6524 if (Op.isGlobalAddress()) {
6525 AM.GV = Op.getGlobal();
6527 AM.Disp = Op.getImm();
6529 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6530 .addReg(MI->getOperand(4).getReg());
6532 // Reload the original control word now.
6533 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6535 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6538 case X86::ATOMAND32:
6539 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6540 X86::AND32ri, X86::MOV32rm,
6541 X86::LCMPXCHG32, X86::MOV32rr,
6542 X86::NOT32r, X86::EAX,
6543 X86::GR32RegisterClass);
6545 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6546 X86::OR32ri, X86::MOV32rm,
6547 X86::LCMPXCHG32, X86::MOV32rr,
6548 X86::NOT32r, X86::EAX,
6549 X86::GR32RegisterClass);
6550 case X86::ATOMXOR32:
6551 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6552 X86::XOR32ri, X86::MOV32rm,
6553 X86::LCMPXCHG32, X86::MOV32rr,
6554 X86::NOT32r, X86::EAX,
6555 X86::GR32RegisterClass);
6556 case X86::ATOMNAND32:
6557 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6558 X86::AND32ri, X86::MOV32rm,
6559 X86::LCMPXCHG32, X86::MOV32rr,
6560 X86::NOT32r, X86::EAX,
6561 X86::GR32RegisterClass, true);
6562 case X86::ATOMMIN32:
6563 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6564 case X86::ATOMMAX32:
6565 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6566 case X86::ATOMUMIN32:
6567 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6568 case X86::ATOMUMAX32:
6569 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6571 case X86::ATOMAND16:
6572 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6573 X86::AND16ri, X86::MOV16rm,
6574 X86::LCMPXCHG16, X86::MOV16rr,
6575 X86::NOT16r, X86::AX,
6576 X86::GR16RegisterClass);
6578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6579 X86::OR16ri, X86::MOV16rm,
6580 X86::LCMPXCHG16, X86::MOV16rr,
6581 X86::NOT16r, X86::AX,
6582 X86::GR16RegisterClass);
6583 case X86::ATOMXOR16:
6584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6585 X86::XOR16ri, X86::MOV16rm,
6586 X86::LCMPXCHG16, X86::MOV16rr,
6587 X86::NOT16r, X86::AX,
6588 X86::GR16RegisterClass);
6589 case X86::ATOMNAND16:
6590 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6591 X86::AND16ri, X86::MOV16rm,
6592 X86::LCMPXCHG16, X86::MOV16rr,
6593 X86::NOT16r, X86::AX,
6594 X86::GR16RegisterClass, true);
6595 case X86::ATOMMIN16:
6596 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6597 case X86::ATOMMAX16:
6598 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6599 case X86::ATOMUMIN16:
6600 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6601 case X86::ATOMUMAX16:
6602 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6605 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6606 X86::AND8ri, X86::MOV8rm,
6607 X86::LCMPXCHG8, X86::MOV8rr,
6608 X86::NOT8r, X86::AL,
6609 X86::GR8RegisterClass);
6611 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6612 X86::OR8ri, X86::MOV8rm,
6613 X86::LCMPXCHG8, X86::MOV8rr,
6614 X86::NOT8r, X86::AL,
6615 X86::GR8RegisterClass);
6617 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6618 X86::XOR8ri, X86::MOV8rm,
6619 X86::LCMPXCHG8, X86::MOV8rr,
6620 X86::NOT8r, X86::AL,
6621 X86::GR8RegisterClass);
6622 case X86::ATOMNAND8:
6623 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6624 X86::AND8ri, X86::MOV8rm,
6625 X86::LCMPXCHG8, X86::MOV8rr,
6626 X86::NOT8r, X86::AL,
6627 X86::GR8RegisterClass, true);
6628 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6629 case X86::ATOMAND64:
6630 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6631 X86::AND64ri32, X86::MOV64rm,
6632 X86::LCMPXCHG64, X86::MOV64rr,
6633 X86::NOT64r, X86::RAX,
6634 X86::GR64RegisterClass);
6636 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6637 X86::OR64ri32, X86::MOV64rm,
6638 X86::LCMPXCHG64, X86::MOV64rr,
6639 X86::NOT64r, X86::RAX,
6640 X86::GR64RegisterClass);
6641 case X86::ATOMXOR64:
6642 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6643 X86::XOR64ri32, X86::MOV64rm,
6644 X86::LCMPXCHG64, X86::MOV64rr,
6645 X86::NOT64r, X86::RAX,
6646 X86::GR64RegisterClass);
6647 case X86::ATOMNAND64:
6648 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6649 X86::AND64ri32, X86::MOV64rm,
6650 X86::LCMPXCHG64, X86::MOV64rr,
6651 X86::NOT64r, X86::RAX,
6652 X86::GR64RegisterClass, true);
6653 case X86::ATOMMIN64:
6654 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6655 case X86::ATOMMAX64:
6656 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6657 case X86::ATOMUMIN64:
6658 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6659 case X86::ATOMUMAX64:
6660 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6664 //===----------------------------------------------------------------------===//
6665 // X86 Optimization Hooks
6666 //===----------------------------------------------------------------------===//
6668 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6672 const SelectionDAG &DAG,
6673 unsigned Depth) const {
6674 unsigned Opc = Op.getOpcode();
6675 assert((Opc >= ISD::BUILTIN_OP_END ||
6676 Opc == ISD::INTRINSIC_WO_CHAIN ||
6677 Opc == ISD::INTRINSIC_W_CHAIN ||
6678 Opc == ISD::INTRINSIC_VOID) &&
6679 "Should use MaskedValueIsZero if you don't know whether Op"
6680 " is a target node!");
6682 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
6686 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
6687 Mask.getBitWidth() - 1);
6692 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
6693 /// node is a GlobalAddress + offset.
6694 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
6695 GlobalValue* &GA, int64_t &Offset) const{
6696 if (N->getOpcode() == X86ISD::Wrapper) {
6697 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
6698 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
6702 return TargetLowering::isGAPlusOffset(N, GA, Offset);
6705 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
6706 const TargetLowering &TLI) {
6709 if (TLI.isGAPlusOffset(Base, GV, Offset))
6710 return (GV->getAlignment() >= N && (Offset % N) == 0);
6711 // DAG combine handles the stack object case.
6715 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
6716 unsigned NumElems, MVT EVT,
6718 SelectionDAG &DAG, MachineFrameInfo *MFI,
6719 const TargetLowering &TLI) {
6721 for (unsigned i = 0; i < NumElems; ++i) {
6722 SDValue Idx = PermMask.getOperand(i);
6723 if (Idx.getOpcode() == ISD::UNDEF) {
6729 SDValue Elt = DAG.getShuffleScalarElt(N, i);
6730 if (!Elt.getNode() ||
6731 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6734 Base = Elt.getNode();
6735 if (Base->getOpcode() == ISD::UNDEF)
6739 if (Elt.getOpcode() == ISD::UNDEF)
6742 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
6743 EVT.getSizeInBits()/8, i, MFI))
6749 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6750 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6751 /// if the load addresses are consecutive, non-overlapping, and in the right
6753 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6754 const TargetLowering &TLI) {
6755 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6756 MVT VT = N->getValueType(0);
6757 MVT EVT = VT.getVectorElementType();
6758 SDValue PermMask = N->getOperand(2);
6759 unsigned NumElems = PermMask.getNumOperands();
6760 SDNode *Base = NULL;
6761 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
6765 LoadSDNode *LD = cast<LoadSDNode>(Base);
6766 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
6767 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6768 LD->getSrcValueOffset(), LD->isVolatile());
6769 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6770 LD->getSrcValueOffset(), LD->isVolatile(),
6771 LD->getAlignment());
6774 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
6775 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
6776 const X86Subtarget *Subtarget,
6777 const TargetLowering &TLI) {
6778 unsigned NumOps = N->getNumOperands();
6780 // Ignore single operand BUILD_VECTOR.
6784 MVT VT = N->getValueType(0);
6785 MVT EVT = VT.getVectorElementType();
6786 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
6787 // We are looking for load i64 and zero extend. We want to transform
6788 // it before legalizer has a chance to expand it. Also look for i64
6789 // BUILD_PAIR bit casted to f64.
6791 // This must be an insertion into a zero vector.
6792 SDValue HighElt = N->getOperand(1);
6793 if (!isZeroNode(HighElt))
6796 // Value must be a load.
6797 SDNode *Base = N->getOperand(0).getNode();
6798 if (!isa<LoadSDNode>(Base)) {
6799 if (Base->getOpcode() != ISD::BIT_CONVERT)
6801 Base = Base->getOperand(0).getNode();
6802 if (!isa<LoadSDNode>(Base))
6806 // Transform it into VZEXT_LOAD addr.
6807 LoadSDNode *LD = cast<LoadSDNode>(Base);
6809 // Load must not be an extload.
6810 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
6813 return DAG.getNode(X86ISD::VZEXT_LOAD, VT, LD->getChain(), LD->getBasePtr());
6816 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6817 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6818 const X86Subtarget *Subtarget) {
6819 SDValue Cond = N->getOperand(0);
6821 // If we have SSE[12] support, try to form min/max nodes.
6822 if (Subtarget->hasSSE2() &&
6823 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6824 if (Cond.getOpcode() == ISD::SETCC) {
6825 // Get the LHS/RHS of the select.
6826 SDValue LHS = N->getOperand(1);
6827 SDValue RHS = N->getOperand(2);
6828 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6830 unsigned Opcode = 0;
6831 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6834 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6837 if (!UnsafeFPMath) break;
6839 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6841 Opcode = X86ISD::FMIN;
6844 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6847 if (!UnsafeFPMath) break;
6849 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6851 Opcode = X86ISD::FMAX;
6854 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6857 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6860 if (!UnsafeFPMath) break;
6862 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6864 Opcode = X86ISD::FMIN;
6867 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6870 if (!UnsafeFPMath) break;
6872 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6874 Opcode = X86ISD::FMAX;
6880 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6888 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6889 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
6890 const X86Subtarget *Subtarget) {
6891 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6892 // the FP state in cases where an emms may be missing.
6893 // A preferable solution to the general problem is to figure out the right
6894 // places to insert EMMS. This qualifies as a quick hack.
6895 StoreSDNode *St = cast<StoreSDNode>(N);
6896 if (St->getValue().getValueType().isVector() &&
6897 St->getValue().getValueType().getSizeInBits() == 64 &&
6898 isa<LoadSDNode>(St->getValue()) &&
6899 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6900 St->getChain().hasOneUse() && !St->isVolatile()) {
6901 SDNode* LdVal = St->getValue().getNode();
6903 int TokenFactorIndex = -1;
6904 SmallVector<SDValue, 8> Ops;
6905 SDNode* ChainVal = St->getChain().getNode();
6906 // Must be a store of a load. We currently handle two cases: the load
6907 // is a direct child, and it's under an intervening TokenFactor. It is
6908 // possible to dig deeper under nested TokenFactors.
6909 if (ChainVal == LdVal)
6910 Ld = cast<LoadSDNode>(St->getChain());
6911 else if (St->getValue().hasOneUse() &&
6912 ChainVal->getOpcode() == ISD::TokenFactor) {
6913 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6914 if (ChainVal->getOperand(i).getNode() == LdVal) {
6915 TokenFactorIndex = i;
6916 Ld = cast<LoadSDNode>(St->getValue());
6918 Ops.push_back(ChainVal->getOperand(i));
6922 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6923 if (Subtarget->is64Bit()) {
6924 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6925 Ld->getBasePtr(), Ld->getSrcValue(),
6926 Ld->getSrcValueOffset(), Ld->isVolatile(),
6927 Ld->getAlignment());
6928 SDValue NewChain = NewLd.getValue(1);
6929 if (TokenFactorIndex != -1) {
6930 Ops.push_back(NewChain);
6931 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6934 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6935 St->getSrcValue(), St->getSrcValueOffset(),
6936 St->isVolatile(), St->getAlignment());
6939 // Otherwise, lower to two 32-bit copies.
6940 SDValue LoAddr = Ld->getBasePtr();
6941 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6942 DAG.getConstant(4, MVT::i32));
6944 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6945 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6946 Ld->isVolatile(), Ld->getAlignment());
6947 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6948 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6950 MinAlign(Ld->getAlignment(), 4));
6952 SDValue NewChain = LoLd.getValue(1);
6953 if (TokenFactorIndex != -1) {
6954 Ops.push_back(LoLd);
6955 Ops.push_back(HiLd);
6956 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6960 LoAddr = St->getBasePtr();
6961 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6962 DAG.getConstant(4, MVT::i32));
6964 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6965 St->getSrcValue(), St->getSrcValueOffset(),
6966 St->isVolatile(), St->getAlignment());
6967 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6969 St->getSrcValueOffset() + 4,
6971 MinAlign(St->getAlignment(), 4));
6972 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6978 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6979 /// X86ISD::FXOR nodes.
6980 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6981 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6982 // F[X]OR(0.0, x) -> x
6983 // F[X]OR(x, 0.0) -> x
6984 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6985 if (C->getValueAPF().isPosZero())
6986 return N->getOperand(1);
6987 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6988 if (C->getValueAPF().isPosZero())
6989 return N->getOperand(0);
6993 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6994 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6995 // FAND(0.0, x) -> 0.0
6996 // FAND(x, 0.0) -> 0.0
6997 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6998 if (C->getValueAPF().isPosZero())
6999 return N->getOperand(0);
7000 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7001 if (C->getValueAPF().isPosZero())
7002 return N->getOperand(1);
7007 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7008 DAGCombinerInfo &DCI) const {
7009 SelectionDAG &DAG = DCI.DAG;
7010 switch (N->getOpcode()) {
7012 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7013 case ISD::BUILD_VECTOR:
7014 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7015 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7016 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7018 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7019 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7025 //===----------------------------------------------------------------------===//
7026 // X86 Inline Assembly Support
7027 //===----------------------------------------------------------------------===//
7029 /// getConstraintType - Given a constraint letter, return the type of
7030 /// constraint it is for this target.
7031 X86TargetLowering::ConstraintType
7032 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7033 if (Constraint.size() == 1) {
7034 switch (Constraint[0]) {
7045 return C_RegisterClass;
7050 return TargetLowering::getConstraintType(Constraint);
7053 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7054 /// with another that has more specific requirements based on the type of the
7055 /// corresponding operand.
7056 const char *X86TargetLowering::
7057 LowerXConstraint(MVT ConstraintVT) const {
7058 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7059 // 'f' like normal targets.
7060 if (ConstraintVT.isFloatingPoint()) {
7061 if (Subtarget->hasSSE2())
7063 if (Subtarget->hasSSE1())
7067 return TargetLowering::LowerXConstraint(ConstraintVT);
7070 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7071 /// vector. If it is invalid, don't add anything to Ops.
7072 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7074 std::vector<SDValue>&Ops,
7075 SelectionDAG &DAG) const {
7076 SDValue Result(0, 0);
7078 switch (Constraint) {
7081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7082 if (C->getValue() <= 31) {
7083 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7090 if (C->getValue() <= 255) {
7091 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7097 // Literal immediates are always ok.
7098 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7099 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
7103 // If we are in non-pic codegen mode, we allow the address of a global (with
7104 // an optional displacement) to be used with 'i'.
7105 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7108 // Match either (GA) or (GA+C)
7110 Offset = GA->getOffset();
7111 } else if (Op.getOpcode() == ISD::ADD) {
7112 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7113 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7115 Offset = GA->getOffset()+C->getValue();
7117 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7118 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7120 Offset = GA->getOffset()+C->getValue();
7127 // If addressing this global requires a load (e.g. in PIC mode), we can't
7129 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
7133 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7139 // Otherwise, not valid for this mode.
7144 if (Result.getNode()) {
7145 Ops.push_back(Result);
7148 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7151 std::vector<unsigned> X86TargetLowering::
7152 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7154 if (Constraint.size() == 1) {
7155 // FIXME: not handling fp-stack yet!
7156 switch (Constraint[0]) { // GCC X86 Constraint Letters
7157 default: break; // Unknown constraint letter
7158 case 'A': // EAX/EDX
7159 if (VT == MVT::i32 || VT == MVT::i64)
7160 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7162 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7165 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7166 else if (VT == MVT::i16)
7167 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7168 else if (VT == MVT::i8)
7169 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7170 else if (VT == MVT::i64)
7171 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7176 return std::vector<unsigned>();
7179 std::pair<unsigned, const TargetRegisterClass*>
7180 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7182 // First, see if this is a constraint that directly corresponds to an LLVM
7184 if (Constraint.size() == 1) {
7185 // GCC Constraint Letters
7186 switch (Constraint[0]) {
7188 case 'r': // GENERAL_REGS
7189 case 'R': // LEGACY_REGS
7190 case 'l': // INDEX_REGS
7191 if (VT == MVT::i64 && Subtarget->is64Bit())
7192 return std::make_pair(0U, X86::GR64RegisterClass);
7194 return std::make_pair(0U, X86::GR32RegisterClass);
7195 else if (VT == MVT::i16)
7196 return std::make_pair(0U, X86::GR16RegisterClass);
7197 else if (VT == MVT::i8)
7198 return std::make_pair(0U, X86::GR8RegisterClass);
7200 case 'f': // FP Stack registers.
7201 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7202 // value to the correct fpstack register class.
7203 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7204 return std::make_pair(0U, X86::RFP32RegisterClass);
7205 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7206 return std::make_pair(0U, X86::RFP64RegisterClass);
7207 return std::make_pair(0U, X86::RFP80RegisterClass);
7208 case 'y': // MMX_REGS if MMX allowed.
7209 if (!Subtarget->hasMMX()) break;
7210 return std::make_pair(0U, X86::VR64RegisterClass);
7212 case 'Y': // SSE_REGS if SSE2 allowed
7213 if (!Subtarget->hasSSE2()) break;
7215 case 'x': // SSE_REGS if SSE1 allowed
7216 if (!Subtarget->hasSSE1()) break;
7218 switch (VT.getSimpleVT()) {
7220 // Scalar SSE types.
7223 return std::make_pair(0U, X86::FR32RegisterClass);
7226 return std::make_pair(0U, X86::FR64RegisterClass);
7234 return std::make_pair(0U, X86::VR128RegisterClass);
7240 // Use the default implementation in TargetLowering to convert the register
7241 // constraint into a member of a register class.
7242 std::pair<unsigned, const TargetRegisterClass*> Res;
7243 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7245 // Not found as a standard register?
7246 if (Res.second == 0) {
7247 // GCC calls "st(0)" just plain "st".
7248 if (StringsEqualNoCase("{st}", Constraint)) {
7249 Res.first = X86::ST0;
7250 Res.second = X86::RFP80RegisterClass;
7256 // Otherwise, check to see if this is a register class of the wrong value
7257 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7258 // turn into {ax},{dx}.
7259 if (Res.second->hasType(VT))
7260 return Res; // Correct type already, nothing to do.
7262 // All of the single-register GCC register classes map their values onto
7263 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7264 // really want an 8-bit or 32-bit register, map to the appropriate register
7265 // class and return the appropriate register.
7266 if (Res.second == X86::GR16RegisterClass) {
7267 if (VT == MVT::i8) {
7268 unsigned DestReg = 0;
7269 switch (Res.first) {
7271 case X86::AX: DestReg = X86::AL; break;
7272 case X86::DX: DestReg = X86::DL; break;
7273 case X86::CX: DestReg = X86::CL; break;
7274 case X86::BX: DestReg = X86::BL; break;
7277 Res.first = DestReg;
7278 Res.second = Res.second = X86::GR8RegisterClass;
7280 } else if (VT == MVT::i32) {
7281 unsigned DestReg = 0;
7282 switch (Res.first) {
7284 case X86::AX: DestReg = X86::EAX; break;
7285 case X86::DX: DestReg = X86::EDX; break;
7286 case X86::CX: DestReg = X86::ECX; break;
7287 case X86::BX: DestReg = X86::EBX; break;
7288 case X86::SI: DestReg = X86::ESI; break;
7289 case X86::DI: DestReg = X86::EDI; break;
7290 case X86::BP: DestReg = X86::EBP; break;
7291 case X86::SP: DestReg = X86::ESP; break;
7294 Res.first = DestReg;
7295 Res.second = Res.second = X86::GR32RegisterClass;
7297 } else if (VT == MVT::i64) {
7298 unsigned DestReg = 0;
7299 switch (Res.first) {
7301 case X86::AX: DestReg = X86::RAX; break;
7302 case X86::DX: DestReg = X86::RDX; break;
7303 case X86::CX: DestReg = X86::RCX; break;
7304 case X86::BX: DestReg = X86::RBX; break;
7305 case X86::SI: DestReg = X86::RSI; break;
7306 case X86::DI: DestReg = X86::RDI; break;
7307 case X86::BP: DestReg = X86::RBP; break;
7308 case X86::SP: DestReg = X86::RSP; break;
7311 Res.first = DestReg;
7312 Res.second = Res.second = X86::GR64RegisterClass;
7315 } else if (Res.second == X86::FR32RegisterClass ||
7316 Res.second == X86::FR64RegisterClass ||
7317 Res.second == X86::VR128RegisterClass) {
7318 // Handle references to XMM physical registers that got mapped into the
7319 // wrong class. This can happen with constraints like {xmm0} where the
7320 // target independent register mapper will just pick the first match it can
7321 // find, ignoring the required type.
7323 Res.second = X86::FR32RegisterClass;
7324 else if (VT == MVT::f64)
7325 Res.second = X86::FR64RegisterClass;
7326 else if (X86::VR128RegisterClass->hasType(VT))
7327 Res.second = X86::VR128RegisterClass;