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);
555 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
562 if (Subtarget->hasMMX()) {
563 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
564 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
565 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
566 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
567 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
569 // FIXME: add MMX packed arithmetics
571 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
572 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
573 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
574 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
576 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
577 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
578 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
579 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
581 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
582 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
584 setOperationAction(ISD::AND, MVT::v8i8, Promote);
585 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
586 setOperationAction(ISD::AND, MVT::v4i16, Promote);
587 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
588 setOperationAction(ISD::AND, MVT::v2i32, Promote);
589 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
590 setOperationAction(ISD::AND, MVT::v1i64, Legal);
592 setOperationAction(ISD::OR, MVT::v8i8, Promote);
593 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
594 setOperationAction(ISD::OR, MVT::v4i16, Promote);
595 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
596 setOperationAction(ISD::OR, MVT::v2i32, Promote);
597 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
598 setOperationAction(ISD::OR, MVT::v1i64, Legal);
600 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
601 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
602 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
603 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
604 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
605 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
606 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
608 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
609 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
610 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
611 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
612 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
613 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
614 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
615 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
616 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
619 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
620 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
621 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
622 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
624 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
625 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
626 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
629 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
630 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
631 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
632 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
634 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
637 if (Subtarget->hasSSE1()) {
638 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
640 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
641 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
642 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
643 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
644 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
645 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
646 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
648 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
649 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
650 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
651 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
654 if (Subtarget->hasSSE2()) {
655 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
656 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
657 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
658 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
659 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
661 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
662 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
663 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
664 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
665 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
666 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
667 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
668 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
669 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
670 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
671 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
672 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
673 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
674 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
675 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
677 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
678 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
679 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
680 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
683 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
685 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
686 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
688 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
689 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
690 MVT VT = (MVT::SimpleValueType)i;
691 // Do not attempt to custom lower non-power-of-2 vectors
692 if (!isPowerOf2_32(VT.getVectorNumElements()))
694 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
695 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
696 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
698 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
699 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
700 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
701 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
702 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
703 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
704 if (Subtarget->is64Bit()) {
705 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
706 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
709 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
710 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
711 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
712 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
713 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
714 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
715 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
716 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
717 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
718 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
719 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
720 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
723 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
725 // Custom lower v2i64 and v2f64 selects.
726 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
727 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
728 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
729 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
733 if (Subtarget->hasSSE41()) {
734 // FIXME: Do we need to handle scalar-to-vector here?
735 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
736 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
738 // i8 and i16 vectors are custom , because the source register and source
739 // source memory operand types are not the same width. f32 vectors are
740 // custom since the immediate controlling the insert encodes additional
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
743 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
744 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
745 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
747 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
748 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
749 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
750 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
752 if (Subtarget->is64Bit()) {
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
754 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
758 if (Subtarget->hasSSE42()) {
759 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
762 // We want to custom lower some of our intrinsics.
763 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
765 // We have target-specific dag combine patterns for the following nodes:
766 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
767 setTargetDAGCombine(ISD::BUILD_VECTOR);
768 setTargetDAGCombine(ISD::SELECT);
769 setTargetDAGCombine(ISD::STORE);
771 computeRegisterProperties();
773 // FIXME: These should be based on subtarget info. Plus, the values should
774 // be smaller when we are in optimizing for size mode.
775 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
776 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
777 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
778 allowUnalignedMemoryAccesses = true; // x86 supports it!
779 setPrefLoopAlignment(16);
783 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
788 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
789 /// the desired ByVal argument alignment.
790 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
793 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
794 if (VTy->getBitWidth() == 128)
796 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
797 unsigned EltAlign = 0;
798 getMaxByValAlign(ATy->getElementType(), EltAlign);
799 if (EltAlign > MaxAlign)
801 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
802 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
803 unsigned EltAlign = 0;
804 getMaxByValAlign(STy->getElementType(i), EltAlign);
805 if (EltAlign > MaxAlign)
814 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
815 /// function arguments in the caller parameter area. For X86, aggregates
816 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
817 /// are at 4-byte boundaries.
818 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
819 if (Subtarget->is64Bit()) {
820 // Max of 8 and alignment of type.
821 unsigned TyAlign = TD->getABITypeAlignment(Ty);
828 if (Subtarget->hasSSE1())
829 getMaxByValAlign(Ty, Align);
833 /// getOptimalMemOpType - Returns the target specific optimal type for load
834 /// and store operations as a result of memset, memcpy, and memmove
835 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
838 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
839 bool isSrcConst, bool isSrcStr) const {
840 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
842 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
844 if (Subtarget->is64Bit() && Size >= 8)
850 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
852 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
853 SelectionDAG &DAG) const {
854 if (usesGlobalOffsetTable())
855 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
856 if (!Subtarget->isPICStyleRIPRel())
857 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
861 //===----------------------------------------------------------------------===//
862 // Return Value Calling Convention Implementation
863 //===----------------------------------------------------------------------===//
865 #include "X86GenCallingConv.inc"
867 /// LowerRET - Lower an ISD::RET node.
868 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
869 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
871 SmallVector<CCValAssign, 16> RVLocs;
872 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
873 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
874 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
875 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
877 // If this is the first return lowered for this function, add the regs to the
878 // liveout set for the function.
879 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
880 for (unsigned i = 0; i != RVLocs.size(); ++i)
881 if (RVLocs[i].isRegLoc())
882 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
884 SDValue Chain = Op.getOperand(0);
886 // Handle tail call return.
887 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
888 if (Chain.getOpcode() == X86ISD::TAILCALL) {
889 SDValue TailCall = Chain;
890 SDValue TargetAddress = TailCall.getOperand(1);
891 SDValue StackAdjustment = TailCall.getOperand(2);
892 assert(((TargetAddress.getOpcode() == ISD::Register &&
893 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
894 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
895 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
896 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
897 "Expecting an global address, external symbol, or register");
898 assert(StackAdjustment.getOpcode() == ISD::Constant &&
899 "Expecting a const value");
901 SmallVector<SDValue,8> Operands;
902 Operands.push_back(Chain.getOperand(0));
903 Operands.push_back(TargetAddress);
904 Operands.push_back(StackAdjustment);
905 // Copy registers used by the call. Last operand is a flag so it is not
907 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
908 Operands.push_back(Chain.getOperand(i));
910 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
917 SmallVector<SDValue, 6> RetOps;
918 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
919 // Operand #1 = Bytes To Pop
920 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
922 // Copy the result values into the output registers.
923 for (unsigned i = 0; i != RVLocs.size(); ++i) {
924 CCValAssign &VA = RVLocs[i];
925 assert(VA.isRegLoc() && "Can only return in registers!");
926 SDValue ValToCopy = Op.getOperand(i*2+1);
928 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
929 // the RET instruction and handled by the FP Stackifier.
930 if (RVLocs[i].getLocReg() == X86::ST0 ||
931 RVLocs[i].getLocReg() == X86::ST1) {
932 // If this is a copy from an xmm register to ST(0), use an FPExtend to
933 // change the value to the FP stack register class.
934 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
935 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
936 RetOps.push_back(ValToCopy);
937 // Don't emit a copytoreg.
941 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
942 Flag = Chain.getValue(1);
945 // The x86-64 ABI for returning structs by value requires that we copy
946 // the sret argument into %rax for the return. We saved the argument into
947 // a virtual register in the entry block, so now we copy the value out
949 if (Subtarget->is64Bit() &&
950 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
951 MachineFunction &MF = DAG.getMachineFunction();
952 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
953 unsigned Reg = FuncInfo->getSRetReturnReg();
955 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
956 FuncInfo->setSRetReturnReg(Reg);
958 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
960 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
961 Flag = Chain.getValue(1);
964 RetOps[0] = Chain; // Update chain.
966 // Add the flag if we have it.
968 RetOps.push_back(Flag);
970 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
974 /// LowerCallResult - Lower the result values of an ISD::CALL into the
975 /// appropriate copies out of appropriate physical registers. This assumes that
976 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
977 /// being lowered. The returns a SDNode with the same number of values as the
979 SDNode *X86TargetLowering::
980 LowerCallResult(SDValue Chain, SDValue InFlag, SDNode *TheCall,
981 unsigned CallingConv, SelectionDAG &DAG) {
983 // Assign locations to each value returned by this call.
984 SmallVector<CCValAssign, 16> RVLocs;
986 cast<ConstantSDNode>(TheCall->getOperand(2))->getZExtValue() != 0;
987 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
988 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
990 SmallVector<SDValue, 8> ResultVals;
992 // Copy all of the result registers out of their specified physreg.
993 for (unsigned i = 0; i != RVLocs.size(); ++i) {
994 MVT CopyVT = RVLocs[i].getValVT();
996 // If this is a call to a function that returns an fp value on the floating
997 // point stack, but where we prefer to use the value in xmm registers, copy
998 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
999 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1000 RVLocs[i].getLocReg() == X86::ST1) &&
1001 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
1005 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1006 CopyVT, InFlag).getValue(1);
1007 SDValue Val = Chain.getValue(0);
1008 InFlag = Chain.getValue(2);
1010 if (CopyVT != RVLocs[i].getValVT()) {
1011 // Round the F80 the right size, which also moves to the appropriate xmm
1013 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1014 // This truncation won't change the value.
1015 DAG.getIntPtrConstant(1));
1018 ResultVals.push_back(Val);
1021 // Merge everything together with a MERGE_VALUES node.
1022 ResultVals.push_back(Chain);
1023 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1024 ResultVals.size()).getNode();
1028 //===----------------------------------------------------------------------===//
1029 // C & StdCall & Fast Calling Convention implementation
1030 //===----------------------------------------------------------------------===//
1031 // StdCall calling convention seems to be standard for many Windows' API
1032 // routines and around. It differs from C calling convention just a little:
1033 // callee should clean up the stack, not caller. Symbols should be also
1034 // decorated in some fancy way :) It doesn't support any vector arguments.
1035 // For info on fast calling convention see Fast Calling Convention (tail call)
1036 // implementation LowerX86_32FastCCCallTo.
1038 /// AddLiveIn - This helper function adds the specified physical register to the
1039 /// MachineFunction as a live in value. It also creates a corresponding virtual
1040 /// register for it.
1041 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1042 const TargetRegisterClass *RC) {
1043 assert(RC->contains(PReg) && "Not the correct regclass!");
1044 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1045 MF.getRegInfo().addLiveIn(PReg, VReg);
1049 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1051 static bool CallIsStructReturn(SDValue Op) {
1052 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1056 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
1059 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1060 /// return semantics.
1061 static bool ArgsAreStructReturn(SDValue Op) {
1062 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1066 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1069 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1070 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1072 bool X86TargetLowering::IsCalleePop(SDValue Op) {
1073 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1077 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
1080 case CallingConv::X86_StdCall:
1081 return !Subtarget->is64Bit();
1082 case CallingConv::X86_FastCall:
1083 return !Subtarget->is64Bit();
1084 case CallingConv::Fast:
1085 return PerformTailCallOpt;
1089 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1090 /// FORMAL_ARGUMENTS node.
1091 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
1092 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1094 if (Subtarget->is64Bit()) {
1095 if (Subtarget->isTargetWin64())
1096 return CC_X86_Win64_C;
1097 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1098 return CC_X86_64_TailCall;
1103 if (CC == CallingConv::X86_FastCall)
1104 return CC_X86_32_FastCall;
1105 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1106 return CC_X86_32_TailCall;
1107 else if (CC == CallingConv::Fast)
1108 return CC_X86_32_FastCC;
1113 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1114 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1116 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1117 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1118 if (CC == CallingConv::X86_FastCall)
1120 else if (CC == CallingConv::X86_StdCall)
1126 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1127 /// in a register before calling.
1128 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1129 return !IsTailCall && !Is64Bit &&
1130 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1131 Subtarget->isPICStyleGOT();
1134 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1135 /// address to be loaded in a register.
1137 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1138 return !Is64Bit && IsTailCall &&
1139 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1140 Subtarget->isPICStyleGOT();
1143 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1144 /// by "Src" to address "Dst" with size and alignment information specified by
1145 /// the specific parameter attribute. The copy will be passed as a byval
1146 /// function parameter.
1148 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1149 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1150 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1151 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1152 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1155 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1156 const CCValAssign &VA,
1157 MachineFrameInfo *MFI,
1159 SDValue Root, unsigned i) {
1160 // Create the nodes corresponding to a load from this parameter slot.
1161 ISD::ArgFlagsTy Flags =
1162 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1163 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1164 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1166 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1167 // changed with more analysis.
1168 // In case of tail call optimization mark all arguments mutable. Since they
1169 // could be overwritten by lowering of arguments in case of a tail call.
1170 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1171 VA.getLocMemOffset(), isImmutable);
1172 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1173 if (Flags.isByVal())
1175 return DAG.getLoad(VA.getValVT(), Root, FIN,
1176 PseudoSourceValue::getFixedStack(FI), 0);
1180 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1181 MachineFunction &MF = DAG.getMachineFunction();
1182 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1184 const Function* Fn = MF.getFunction();
1185 if (Fn->hasExternalLinkage() &&
1186 Subtarget->isTargetCygMing() &&
1187 Fn->getName() == "main")
1188 FuncInfo->setForceFramePointer(true);
1190 // Decorate the function name.
1191 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1193 MachineFrameInfo *MFI = MF.getFrameInfo();
1194 SDValue Root = Op.getOperand(0);
1195 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1196 unsigned CC = MF.getFunction()->getCallingConv();
1197 bool Is64Bit = Subtarget->is64Bit();
1198 bool IsWin64 = Subtarget->isTargetWin64();
1200 assert(!(isVarArg && CC == CallingConv::Fast) &&
1201 "Var args not supported with calling convention fastcc");
1203 // Assign locations to all of the incoming arguments.
1204 SmallVector<CCValAssign, 16> ArgLocs;
1205 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1206 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(Op));
1208 SmallVector<SDValue, 8> ArgValues;
1209 unsigned LastVal = ~0U;
1210 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1211 CCValAssign &VA = ArgLocs[i];
1212 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1214 assert(VA.getValNo() != LastVal &&
1215 "Don't support value assigned to multiple locs yet");
1216 LastVal = VA.getValNo();
1218 if (VA.isRegLoc()) {
1219 MVT RegVT = VA.getLocVT();
1220 TargetRegisterClass *RC;
1221 if (RegVT == MVT::i32)
1222 RC = X86::GR32RegisterClass;
1223 else if (Is64Bit && RegVT == MVT::i64)
1224 RC = X86::GR64RegisterClass;
1225 else if (RegVT == MVT::f32)
1226 RC = X86::FR32RegisterClass;
1227 else if (RegVT == MVT::f64)
1228 RC = X86::FR64RegisterClass;
1229 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1230 RC = X86::VR128RegisterClass;
1231 else if (RegVT.isVector()) {
1232 assert(RegVT.getSizeInBits() == 64);
1234 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1236 // Darwin calling convention passes MMX values in either GPRs or
1237 // XMMs in x86-64. Other targets pass them in memory.
1238 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1239 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1242 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1247 assert(0 && "Unknown argument type!");
1250 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1251 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1253 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1254 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1256 if (VA.getLocInfo() == CCValAssign::SExt)
1257 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1258 DAG.getValueType(VA.getValVT()));
1259 else if (VA.getLocInfo() == CCValAssign::ZExt)
1260 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1261 DAG.getValueType(VA.getValVT()));
1263 if (VA.getLocInfo() != CCValAssign::Full)
1264 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1266 // Handle MMX values passed in GPRs.
1267 if (Is64Bit && RegVT != VA.getLocVT()) {
1268 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1269 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1270 else if (RC == X86::VR128RegisterClass) {
1271 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1272 DAG.getConstant(0, MVT::i64));
1273 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1277 ArgValues.push_back(ArgValue);
1279 assert(VA.isMemLoc());
1280 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1284 // The x86-64 ABI for returning structs by value requires that we copy
1285 // the sret argument into %rax for the return. Save the argument into
1286 // a virtual register so that we can access it from the return points.
1287 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1288 MachineFunction &MF = DAG.getMachineFunction();
1289 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1290 unsigned Reg = FuncInfo->getSRetReturnReg();
1292 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1293 FuncInfo->setSRetReturnReg(Reg);
1295 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1296 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1299 unsigned StackSize = CCInfo.getNextStackOffset();
1300 // align stack specially for tail calls
1301 if (PerformTailCallOpt && CC == CallingConv::Fast)
1302 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1304 // If the function takes variable number of arguments, make a frame index for
1305 // the start of the first vararg value... for expansion of llvm.va_start.
1307 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1308 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1311 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1313 // FIXME: We should really autogenerate these arrays
1314 static const unsigned GPR64ArgRegsWin64[] = {
1315 X86::RCX, X86::RDX, X86::R8, X86::R9
1317 static const unsigned XMMArgRegsWin64[] = {
1318 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1320 static const unsigned GPR64ArgRegs64Bit[] = {
1321 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1323 static const unsigned XMMArgRegs64Bit[] = {
1324 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1325 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1327 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1330 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1331 GPR64ArgRegs = GPR64ArgRegsWin64;
1332 XMMArgRegs = XMMArgRegsWin64;
1334 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1335 GPR64ArgRegs = GPR64ArgRegs64Bit;
1336 XMMArgRegs = XMMArgRegs64Bit;
1338 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1340 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1343 // For X86-64, if there are vararg parameters that are passed via
1344 // registers, then we must store them to their spots on the stack so they
1345 // may be loaded by deferencing the result of va_next.
1346 VarArgsGPOffset = NumIntRegs * 8;
1347 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1348 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1349 TotalNumXMMRegs * 16, 16);
1351 // Store the integer parameter registers.
1352 SmallVector<SDValue, 8> MemOps;
1353 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1354 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1355 DAG.getIntPtrConstant(VarArgsGPOffset));
1356 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1357 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1358 X86::GR64RegisterClass);
1359 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1361 DAG.getStore(Val.getValue(1), Val, FIN,
1362 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1363 MemOps.push_back(Store);
1364 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1365 DAG.getIntPtrConstant(8));
1368 // Now store the XMM (fp + vector) parameter registers.
1369 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1370 DAG.getIntPtrConstant(VarArgsFPOffset));
1371 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1372 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1373 X86::VR128RegisterClass);
1374 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1376 DAG.getStore(Val.getValue(1), Val, FIN,
1377 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1378 MemOps.push_back(Store);
1379 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1380 DAG.getIntPtrConstant(16));
1382 if (!MemOps.empty())
1383 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1384 &MemOps[0], MemOps.size());
1388 ArgValues.push_back(Root);
1390 // Some CCs need callee pop.
1391 if (IsCalleePop(Op)) {
1392 BytesToPopOnReturn = StackSize; // Callee pops everything.
1393 BytesCallerReserves = 0;
1395 BytesToPopOnReturn = 0; // Callee pops nothing.
1396 // If this is an sret function, the return should pop the hidden pointer.
1397 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1398 BytesToPopOnReturn = 4;
1399 BytesCallerReserves = StackSize;
1403 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1404 if (CC == CallingConv::X86_FastCall)
1405 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1408 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1410 // Return the new list of results.
1411 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1412 ArgValues.size()).getValue(Op.getResNo());
1416 X86TargetLowering::LowerMemOpCallTo(SDValue Op, SelectionDAG &DAG,
1417 const SDValue &StackPtr,
1418 const CCValAssign &VA,
1421 unsigned LocMemOffset = VA.getLocMemOffset();
1422 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1423 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1424 ISD::ArgFlagsTy Flags =
1425 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1426 if (Flags.isByVal()) {
1427 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1429 return DAG.getStore(Chain, Arg, PtrOff,
1430 PseudoSourceValue::getStack(), LocMemOffset);
1433 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1434 /// optimization is performed and it is required.
1436 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1437 SDValue &OutRetAddr,
1442 if (!IsTailCall || FPDiff==0) return Chain;
1444 // Adjust the Return address stack slot.
1445 MVT VT = getPointerTy();
1446 OutRetAddr = getReturnAddressFrameIndex(DAG);
1447 // Load the "old" Return address.
1448 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1449 return SDValue(OutRetAddr.getNode(), 1);
1452 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1453 /// optimization is performed and it is required (FPDiff!=0).
1455 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1456 SDValue Chain, SDValue RetAddrFrIdx,
1457 bool Is64Bit, int FPDiff) {
1458 // Store the return address to the appropriate stack slot.
1459 if (!FPDiff) return Chain;
1460 // Calculate the new stack slot for the return address.
1461 int SlotSize = Is64Bit ? 8 : 4;
1462 int NewReturnAddrFI =
1463 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1464 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1465 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1466 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1467 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1471 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1472 MachineFunction &MF = DAG.getMachineFunction();
1473 SDValue Chain = Op.getOperand(0);
1474 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1475 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1476 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue() != 0
1477 && CC == CallingConv::Fast && PerformTailCallOpt;
1478 SDValue Callee = Op.getOperand(4);
1479 bool Is64Bit = Subtarget->is64Bit();
1480 bool IsStructRet = CallIsStructReturn(Op);
1482 assert(!(isVarArg && CC == CallingConv::Fast) &&
1483 "Var args not supported with calling convention fastcc");
1485 // Analyze operands of the call, assigning locations to each operand.
1486 SmallVector<CCValAssign, 16> ArgLocs;
1487 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1488 CCInfo.AnalyzeCallOperands(Op.getNode(), CCAssignFnForNode(Op));
1490 // Get a count of how many bytes are to be pushed on the stack.
1491 unsigned NumBytes = CCInfo.getNextStackOffset();
1492 if (PerformTailCallOpt && CC == CallingConv::Fast)
1493 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1497 // Lower arguments at fp - stackoffset + fpdiff.
1498 unsigned NumBytesCallerPushed =
1499 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1500 FPDiff = NumBytesCallerPushed - NumBytes;
1502 // Set the delta of movement of the returnaddr stackslot.
1503 // But only set if delta is greater than previous delta.
1504 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1505 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1508 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1510 SDValue RetAddrFrIdx;
1511 // Load return adress for tail calls.
1512 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1515 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1516 SmallVector<SDValue, 8> MemOpChains;
1519 // Walk the register/memloc assignments, inserting copies/loads. In the case
1520 // of tail call optimization arguments are handle later.
1521 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1522 CCValAssign &VA = ArgLocs[i];
1523 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1524 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1525 getArgFlags().isByVal();
1527 // Promote the value if needed.
1528 switch (VA.getLocInfo()) {
1529 default: assert(0 && "Unknown loc info!");
1530 case CCValAssign::Full: break;
1531 case CCValAssign::SExt:
1532 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1534 case CCValAssign::ZExt:
1535 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1537 case CCValAssign::AExt:
1538 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1542 if (VA.isRegLoc()) {
1544 MVT RegVT = VA.getLocVT();
1545 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1546 switch (VA.getLocReg()) {
1549 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1551 // Special case: passing MMX values in GPR registers.
1552 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1555 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1556 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1557 // Special case: passing MMX values in XMM registers.
1558 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1559 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1560 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1561 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1562 getMOVLMask(2, DAG));
1567 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1569 if (!IsTailCall || (IsTailCall && isByVal)) {
1570 assert(VA.isMemLoc());
1571 if (StackPtr.getNode() == 0)
1572 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1574 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1580 if (!MemOpChains.empty())
1581 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1582 &MemOpChains[0], MemOpChains.size());
1584 // Build a sequence of copy-to-reg nodes chained together with token chain
1585 // and flag operands which copy the outgoing args into registers.
1587 // Tail call byval lowering might overwrite argument registers so in case of
1588 // tail call optimization the copies to registers are lowered later.
1590 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1591 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1593 InFlag = Chain.getValue(1);
1596 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1598 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1599 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1600 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1602 InFlag = Chain.getValue(1);
1604 // If we are tail calling and generating PIC/GOT style code load the address
1605 // of the callee into ecx. The value in ecx is used as target of the tail
1606 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1607 // calls on PIC/GOT architectures. Normally we would just put the address of
1608 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1609 // restored (since ebx is callee saved) before jumping to the target@PLT.
1610 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1611 // Note: The actual moving to ecx is done further down.
1612 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1613 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1614 !G->getGlobal()->hasProtectedVisibility())
1615 Callee = LowerGlobalAddress(Callee, DAG);
1616 else if (isa<ExternalSymbolSDNode>(Callee))
1617 Callee = LowerExternalSymbol(Callee,DAG);
1620 if (Is64Bit && isVarArg) {
1621 // From AMD64 ABI document:
1622 // For calls that may call functions that use varargs or stdargs
1623 // (prototype-less calls or calls to functions containing ellipsis (...) in
1624 // the declaration) %al is used as hidden argument to specify the number
1625 // of SSE registers used. The contents of %al do not need to match exactly
1626 // the number of registers, but must be an ubound on the number of SSE
1627 // registers used and is in the range 0 - 8 inclusive.
1629 // FIXME: Verify this on Win64
1630 // Count the number of XMM registers allocated.
1631 static const unsigned XMMArgRegs[] = {
1632 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1633 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1635 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1637 Chain = DAG.getCopyToReg(Chain, X86::AL,
1638 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1639 InFlag = Chain.getValue(1);
1643 // For tail calls lower the arguments to the 'real' stack slot.
1645 SmallVector<SDValue, 8> MemOpChains2;
1648 // Do not flag preceeding copytoreg stuff together with the following stuff.
1650 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1651 CCValAssign &VA = ArgLocs[i];
1652 if (!VA.isRegLoc()) {
1653 assert(VA.isMemLoc());
1654 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1655 SDValue FlagsOp = Op.getOperand(6+2*VA.getValNo());
1656 ISD::ArgFlagsTy Flags =
1657 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1658 // Create frame index.
1659 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1660 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1661 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1662 FIN = DAG.getFrameIndex(FI, getPointerTy());
1664 if (Flags.isByVal()) {
1665 // Copy relative to framepointer.
1666 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1667 if (StackPtr.getNode() == 0)
1668 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1669 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1671 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1674 // Store relative to framepointer.
1675 MemOpChains2.push_back(
1676 DAG.getStore(Chain, Arg, FIN,
1677 PseudoSourceValue::getFixedStack(FI), 0));
1682 if (!MemOpChains2.empty())
1683 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1684 &MemOpChains2[0], MemOpChains2.size());
1686 // Copy arguments to their registers.
1687 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1688 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1690 InFlag = Chain.getValue(1);
1694 // Store the return address to the appropriate stack slot.
1695 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1699 // If the callee is a GlobalAddress node (quite common, every direct call is)
1700 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1701 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1702 // We should use extra load for direct calls to dllimported functions in
1704 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1705 getTargetMachine(), true))
1706 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1707 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1708 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1709 } else if (IsTailCall) {
1710 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1712 Chain = DAG.getCopyToReg(Chain,
1713 DAG.getRegister(Opc, getPointerTy()),
1715 Callee = DAG.getRegister(Opc, getPointerTy());
1716 // Add register as live out.
1717 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1720 // Returns a chain & a flag for retval copy to use.
1721 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1722 SmallVector<SDValue, 8> Ops;
1725 Ops.push_back(Chain);
1726 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1727 Ops.push_back(DAG.getIntPtrConstant(0));
1728 if (InFlag.getNode())
1729 Ops.push_back(InFlag);
1730 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1731 InFlag = Chain.getValue(1);
1733 // Returns a chain & a flag for retval copy to use.
1734 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1738 Ops.push_back(Chain);
1739 Ops.push_back(Callee);
1742 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1744 // Add argument registers to the end of the list so that they are known live
1746 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1747 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1748 RegsToPass[i].second.getValueType()));
1750 // Add an implicit use GOT pointer in EBX.
1751 if (!IsTailCall && !Is64Bit &&
1752 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1753 Subtarget->isPICStyleGOT())
1754 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1756 // Add an implicit use of AL for x86 vararg functions.
1757 if (Is64Bit && isVarArg)
1758 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1760 if (InFlag.getNode())
1761 Ops.push_back(InFlag);
1764 assert(InFlag.getNode() &&
1765 "Flag must be set. Depend on flag being set in LowerRET");
1766 Chain = DAG.getNode(X86ISD::TAILCALL,
1767 Op.getNode()->getVTList(), &Ops[0], Ops.size());
1769 return SDValue(Chain.getNode(), Op.getResNo());
1772 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1773 InFlag = Chain.getValue(1);
1775 // Create the CALLSEQ_END node.
1776 unsigned NumBytesForCalleeToPush;
1777 if (IsCalleePop(Op))
1778 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1779 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1780 // If this is is a call to a struct-return function, the callee
1781 // pops the hidden struct pointer, so we have to push it back.
1782 // This is common for Darwin/X86, Linux & Mingw32 targets.
1783 NumBytesForCalleeToPush = 4;
1785 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1787 // Returns a flag for retval copy to use.
1788 Chain = DAG.getCALLSEQ_END(Chain,
1789 DAG.getIntPtrConstant(NumBytes),
1790 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1792 InFlag = Chain.getValue(1);
1794 // Handle result values, copying them out of physregs into vregs that we
1796 return SDValue(LowerCallResult(Chain, InFlag, Op.getNode(), CC, DAG),
1801 //===----------------------------------------------------------------------===//
1802 // Fast Calling Convention (tail call) implementation
1803 //===----------------------------------------------------------------------===//
1805 // Like std call, callee cleans arguments, convention except that ECX is
1806 // reserved for storing the tail called function address. Only 2 registers are
1807 // free for argument passing (inreg). Tail call optimization is performed
1809 // * tailcallopt is enabled
1810 // * caller/callee are fastcc
1811 // On X86_64 architecture with GOT-style position independent code only local
1812 // (within module) calls are supported at the moment.
1813 // To keep the stack aligned according to platform abi the function
1814 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1815 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1816 // If a tail called function callee has more arguments than the caller the
1817 // caller needs to make sure that there is room to move the RETADDR to. This is
1818 // achieved by reserving an area the size of the argument delta right after the
1819 // original REtADDR, but before the saved framepointer or the spilled registers
1820 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1832 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1833 /// for a 16 byte align requirement.
1834 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1835 SelectionDAG& DAG) {
1836 MachineFunction &MF = DAG.getMachineFunction();
1837 const TargetMachine &TM = MF.getTarget();
1838 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1839 unsigned StackAlignment = TFI.getStackAlignment();
1840 uint64_t AlignMask = StackAlignment - 1;
1841 int64_t Offset = StackSize;
1842 uint64_t SlotSize = TD->getPointerSize();
1843 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1844 // Number smaller than 12 so just add the difference.
1845 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1847 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1848 Offset = ((~AlignMask) & Offset) + StackAlignment +
1849 (StackAlignment-SlotSize);
1854 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1855 /// following the call is a return. A function is eligible if caller/callee
1856 /// calling conventions match, currently only fastcc supports tail calls, and
1857 /// the function CALL is immediatly followed by a RET.
1858 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1860 SelectionDAG& DAG) const {
1861 if (!PerformTailCallOpt)
1864 if (CheckTailCallReturnConstraints(Call, Ret)) {
1865 MachineFunction &MF = DAG.getMachineFunction();
1866 unsigned CallerCC = MF.getFunction()->getCallingConv();
1867 unsigned CalleeCC= cast<ConstantSDNode>(Call.getOperand(1))->getZExtValue();
1868 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1869 SDValue Callee = Call.getOperand(4);
1870 // On x86/32Bit PIC/GOT tail calls are supported.
1871 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1872 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1875 // Can only do local tail calls (in same module, hidden or protected) on
1876 // x86_64 PIC/GOT at the moment.
1877 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1878 return G->getGlobal()->hasHiddenVisibility()
1879 || G->getGlobal()->hasProtectedVisibility();
1887 X86TargetLowering::createFastISel(MachineFunction &mf,
1888 DenseMap<const Value *, unsigned> &vm,
1889 DenseMap<const BasicBlock *,
1890 MachineBasicBlock *> &bm,
1891 DenseMap<const AllocaInst *, int> &am) {
1893 return X86::createFastISel(mf, vm, bm, am);
1897 //===----------------------------------------------------------------------===//
1898 // Other Lowering Hooks
1899 //===----------------------------------------------------------------------===//
1902 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1903 MachineFunction &MF = DAG.getMachineFunction();
1904 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1905 int ReturnAddrIndex = FuncInfo->getRAIndex();
1906 uint64_t SlotSize = TD->getPointerSize();
1908 if (ReturnAddrIndex == 0) {
1909 // Set up a frame object for the return address.
1910 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1911 FuncInfo->setRAIndex(ReturnAddrIndex);
1914 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1918 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1919 /// specific condition code. It returns a false if it cannot do a direct
1920 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1922 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1923 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1924 SelectionDAG &DAG) {
1925 X86CC = X86::COND_INVALID;
1927 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1928 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1929 // X > -1 -> X == 0, jump !sign.
1930 RHS = DAG.getConstant(0, RHS.getValueType());
1931 X86CC = X86::COND_NS;
1933 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1934 // X < 0 -> X == 0, jump on sign.
1935 X86CC = X86::COND_S;
1937 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
1939 RHS = DAG.getConstant(0, RHS.getValueType());
1940 X86CC = X86::COND_LE;
1945 switch (SetCCOpcode) {
1947 case ISD::SETEQ: X86CC = X86::COND_E; break;
1948 case ISD::SETGT: X86CC = X86::COND_G; break;
1949 case ISD::SETGE: X86CC = X86::COND_GE; break;
1950 case ISD::SETLT: X86CC = X86::COND_L; break;
1951 case ISD::SETLE: X86CC = X86::COND_LE; break;
1952 case ISD::SETNE: X86CC = X86::COND_NE; break;
1953 case ISD::SETULT: X86CC = X86::COND_B; break;
1954 case ISD::SETUGT: X86CC = X86::COND_A; break;
1955 case ISD::SETULE: X86CC = X86::COND_BE; break;
1956 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1959 // First determine if it requires or is profitable to flip the operands.
1961 switch (SetCCOpcode) {
1971 // If LHS is a foldable load, but RHS is not, flip the condition.
1973 (ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
1974 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
1975 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
1979 std::swap(LHS, RHS);
1981 // On a floating point condition, the flags are set as follows:
1983 // 0 | 0 | 0 | X > Y
1984 // 0 | 0 | 1 | X < Y
1985 // 1 | 0 | 0 | X == Y
1986 // 1 | 1 | 1 | unordered
1987 switch (SetCCOpcode) {
1991 X86CC = X86::COND_E;
1993 case ISD::SETOLT: // flipped
1996 X86CC = X86::COND_A;
1998 case ISD::SETOLE: // flipped
2001 X86CC = X86::COND_AE;
2003 case ISD::SETUGT: // flipped
2006 X86CC = X86::COND_B;
2008 case ISD::SETUGE: // flipped
2011 X86CC = X86::COND_BE;
2015 X86CC = X86::COND_NE;
2018 X86CC = X86::COND_P;
2021 X86CC = X86::COND_NP;
2026 return X86CC != X86::COND_INVALID;
2029 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2030 /// code. Current x86 isa includes the following FP cmov instructions:
2031 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2032 static bool hasFPCMov(unsigned X86CC) {
2048 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2049 /// true if Op is undef or if its value falls within the specified range (L, H].
2050 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2051 if (Op.getOpcode() == ISD::UNDEF)
2054 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2055 return (Val >= Low && Val < Hi);
2058 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2059 /// true if Op is undef or if its value equal to the specified value.
2060 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2061 if (Op.getOpcode() == ISD::UNDEF)
2063 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2066 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2067 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2068 bool X86::isPSHUFDMask(SDNode *N) {
2069 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2071 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2074 // Check if the value doesn't reference the second vector.
2075 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2076 SDValue Arg = N->getOperand(i);
2077 if (Arg.getOpcode() == ISD::UNDEF) continue;
2078 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2079 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2086 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2087 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2088 bool X86::isPSHUFHWMask(SDNode *N) {
2089 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2091 if (N->getNumOperands() != 8)
2094 // Lower quadword copied in order.
2095 for (unsigned i = 0; i != 4; ++i) {
2096 SDValue Arg = N->getOperand(i);
2097 if (Arg.getOpcode() == ISD::UNDEF) continue;
2098 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2099 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2103 // Upper quadword shuffled.
2104 for (unsigned i = 4; i != 8; ++i) {
2105 SDValue Arg = N->getOperand(i);
2106 if (Arg.getOpcode() == ISD::UNDEF) continue;
2107 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2108 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2109 if (Val < 4 || Val > 7)
2116 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2117 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2118 bool X86::isPSHUFLWMask(SDNode *N) {
2119 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2121 if (N->getNumOperands() != 8)
2124 // Upper quadword copied in order.
2125 for (unsigned i = 4; i != 8; ++i)
2126 if (!isUndefOrEqual(N->getOperand(i), i))
2129 // Lower quadword shuffled.
2130 for (unsigned i = 0; i != 4; ++i)
2131 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2137 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2138 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2139 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2140 if (NumElems != 2 && NumElems != 4) return false;
2142 unsigned Half = NumElems / 2;
2143 for (unsigned i = 0; i < Half; ++i)
2144 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2146 for (unsigned i = Half; i < NumElems; ++i)
2147 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2153 bool X86::isSHUFPMask(SDNode *N) {
2154 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2155 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2158 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2159 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2160 /// half elements to come from vector 1 (which would equal the dest.) and
2161 /// the upper half to come from vector 2.
2162 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2163 if (NumOps != 2 && NumOps != 4) return false;
2165 unsigned Half = NumOps / 2;
2166 for (unsigned i = 0; i < Half; ++i)
2167 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2169 for (unsigned i = Half; i < NumOps; ++i)
2170 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2175 static bool isCommutedSHUFP(SDNode *N) {
2176 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2177 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2180 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2181 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2182 bool X86::isMOVHLPSMask(SDNode *N) {
2183 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2185 if (N->getNumOperands() != 4)
2188 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2189 return isUndefOrEqual(N->getOperand(0), 6) &&
2190 isUndefOrEqual(N->getOperand(1), 7) &&
2191 isUndefOrEqual(N->getOperand(2), 2) &&
2192 isUndefOrEqual(N->getOperand(3), 3);
2195 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2196 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2198 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2199 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2201 if (N->getNumOperands() != 4)
2204 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2205 return isUndefOrEqual(N->getOperand(0), 2) &&
2206 isUndefOrEqual(N->getOperand(1), 3) &&
2207 isUndefOrEqual(N->getOperand(2), 2) &&
2208 isUndefOrEqual(N->getOperand(3), 3);
2211 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2212 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2213 bool X86::isMOVLPMask(SDNode *N) {
2214 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2216 unsigned NumElems = N->getNumOperands();
2217 if (NumElems != 2 && NumElems != 4)
2220 for (unsigned i = 0; i < NumElems/2; ++i)
2221 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2224 for (unsigned i = NumElems/2; i < NumElems; ++i)
2225 if (!isUndefOrEqual(N->getOperand(i), i))
2231 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2232 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2234 bool X86::isMOVHPMask(SDNode *N) {
2235 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2237 unsigned NumElems = N->getNumOperands();
2238 if (NumElems != 2 && NumElems != 4)
2241 for (unsigned i = 0; i < NumElems/2; ++i)
2242 if (!isUndefOrEqual(N->getOperand(i), i))
2245 for (unsigned i = 0; i < NumElems/2; ++i) {
2246 SDValue Arg = N->getOperand(i + NumElems/2);
2247 if (!isUndefOrEqual(Arg, i + NumElems))
2254 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2255 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2256 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2257 bool V2IsSplat = false) {
2258 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2261 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2262 SDValue BitI = Elts[i];
2263 SDValue BitI1 = Elts[i+1];
2264 if (!isUndefOrEqual(BitI, j))
2267 if (isUndefOrEqual(BitI1, NumElts))
2270 if (!isUndefOrEqual(BitI1, j + NumElts))
2278 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2279 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2280 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2283 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2284 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2285 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2286 bool V2IsSplat = false) {
2287 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2290 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2291 SDValue BitI = Elts[i];
2292 SDValue BitI1 = Elts[i+1];
2293 if (!isUndefOrEqual(BitI, j + NumElts/2))
2296 if (isUndefOrEqual(BitI1, NumElts))
2299 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2307 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2308 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2309 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2312 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2313 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2315 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2316 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2318 unsigned NumElems = N->getNumOperands();
2319 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2322 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2323 SDValue BitI = N->getOperand(i);
2324 SDValue BitI1 = N->getOperand(i+1);
2326 if (!isUndefOrEqual(BitI, j))
2328 if (!isUndefOrEqual(BitI1, j))
2335 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2336 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2338 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2339 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2341 unsigned NumElems = N->getNumOperands();
2342 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2345 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2346 SDValue BitI = N->getOperand(i);
2347 SDValue BitI1 = N->getOperand(i + 1);
2349 if (!isUndefOrEqual(BitI, j))
2351 if (!isUndefOrEqual(BitI1, j))
2358 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2359 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2360 /// MOVSD, and MOVD, i.e. setting the lowest element.
2361 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2362 if (NumElts != 2 && NumElts != 4)
2365 if (!isUndefOrEqual(Elts[0], NumElts))
2368 for (unsigned i = 1; i < NumElts; ++i) {
2369 if (!isUndefOrEqual(Elts[i], i))
2376 bool X86::isMOVLMask(SDNode *N) {
2377 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2378 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2381 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2382 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2383 /// element of vector 2 and the other elements to come from vector 1 in order.
2384 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2385 bool V2IsSplat = false,
2386 bool V2IsUndef = false) {
2387 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2390 if (!isUndefOrEqual(Ops[0], 0))
2393 for (unsigned i = 1; i < NumOps; ++i) {
2394 SDValue Arg = Ops[i];
2395 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2396 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2397 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2404 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2405 bool V2IsUndef = false) {
2406 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2407 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2408 V2IsSplat, V2IsUndef);
2411 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2412 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2413 bool X86::isMOVSHDUPMask(SDNode *N) {
2414 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2416 if (N->getNumOperands() != 4)
2419 // Expect 1, 1, 3, 3
2420 for (unsigned i = 0; i < 2; ++i) {
2421 SDValue Arg = N->getOperand(i);
2422 if (Arg.getOpcode() == ISD::UNDEF) continue;
2423 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2424 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2425 if (Val != 1) return false;
2429 for (unsigned i = 2; i < 4; ++i) {
2430 SDValue Arg = N->getOperand(i);
2431 if (Arg.getOpcode() == ISD::UNDEF) continue;
2432 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2433 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2434 if (Val != 3) return false;
2438 // Don't use movshdup if it can be done with a shufps.
2442 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2443 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2444 bool X86::isMOVSLDUPMask(SDNode *N) {
2445 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2447 if (N->getNumOperands() != 4)
2450 // Expect 0, 0, 2, 2
2451 for (unsigned i = 0; i < 2; ++i) {
2452 SDValue Arg = N->getOperand(i);
2453 if (Arg.getOpcode() == ISD::UNDEF) continue;
2454 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2455 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2456 if (Val != 0) return false;
2460 for (unsigned i = 2; i < 4; ++i) {
2461 SDValue Arg = N->getOperand(i);
2462 if (Arg.getOpcode() == ISD::UNDEF) continue;
2463 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2464 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2465 if (Val != 2) return false;
2469 // Don't use movshdup if it can be done with a shufps.
2473 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2474 /// specifies a identity operation on the LHS or RHS.
2475 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2476 unsigned NumElems = N->getNumOperands();
2477 for (unsigned i = 0; i < NumElems; ++i)
2478 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2483 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2484 /// a splat of a single element.
2485 static bool isSplatMask(SDNode *N) {
2486 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2488 // This is a splat operation if each element of the permute is the same, and
2489 // if the value doesn't reference the second vector.
2490 unsigned NumElems = N->getNumOperands();
2491 SDValue ElementBase;
2493 for (; i != NumElems; ++i) {
2494 SDValue Elt = N->getOperand(i);
2495 if (isa<ConstantSDNode>(Elt)) {
2501 if (!ElementBase.getNode())
2504 for (; i != NumElems; ++i) {
2505 SDValue Arg = N->getOperand(i);
2506 if (Arg.getOpcode() == ISD::UNDEF) continue;
2507 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2508 if (Arg != ElementBase) return false;
2511 // Make sure it is a splat of the first vector operand.
2512 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2515 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2516 /// a splat of a single element and it's a 2 or 4 element mask.
2517 bool X86::isSplatMask(SDNode *N) {
2518 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2520 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2521 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2523 return ::isSplatMask(N);
2526 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2527 /// specifies a splat of zero element.
2528 bool X86::isSplatLoMask(SDNode *N) {
2529 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2531 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2532 if (!isUndefOrEqual(N->getOperand(i), 0))
2537 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2538 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2540 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2541 unsigned NumOperands = N->getNumOperands();
2542 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2544 for (unsigned i = 0; i < NumOperands; ++i) {
2546 SDValue Arg = N->getOperand(NumOperands-i-1);
2547 if (Arg.getOpcode() != ISD::UNDEF)
2548 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2549 if (Val >= NumOperands) Val -= NumOperands;
2551 if (i != NumOperands - 1)
2558 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2559 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2561 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2563 // 8 nodes, but we only care about the last 4.
2564 for (unsigned i = 7; i >= 4; --i) {
2566 SDValue Arg = N->getOperand(i);
2567 if (Arg.getOpcode() != ISD::UNDEF)
2568 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2577 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2578 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2580 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2582 // 8 nodes, but we only care about the first 4.
2583 for (int i = 3; i >= 0; --i) {
2585 SDValue Arg = N->getOperand(i);
2586 if (Arg.getOpcode() != ISD::UNDEF)
2587 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2596 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2597 /// specifies a 8 element shuffle that can be broken into a pair of
2598 /// PSHUFHW and PSHUFLW.
2599 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2600 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2602 if (N->getNumOperands() != 8)
2605 // Lower quadword shuffled.
2606 for (unsigned i = 0; i != 4; ++i) {
2607 SDValue Arg = N->getOperand(i);
2608 if (Arg.getOpcode() == ISD::UNDEF) continue;
2609 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2610 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2615 // Upper quadword shuffled.
2616 for (unsigned i = 4; i != 8; ++i) {
2617 SDValue Arg = N->getOperand(i);
2618 if (Arg.getOpcode() == ISD::UNDEF) continue;
2619 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2620 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2621 if (Val < 4 || Val > 7)
2628 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2629 /// values in ther permute mask.
2630 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2631 SDValue &V2, SDValue &Mask,
2632 SelectionDAG &DAG) {
2633 MVT VT = Op.getValueType();
2634 MVT MaskVT = Mask.getValueType();
2635 MVT EltVT = MaskVT.getVectorElementType();
2636 unsigned NumElems = Mask.getNumOperands();
2637 SmallVector<SDValue, 8> MaskVec;
2639 for (unsigned i = 0; i != NumElems; ++i) {
2640 SDValue Arg = Mask.getOperand(i);
2641 if (Arg.getOpcode() == ISD::UNDEF) {
2642 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2645 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2646 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2648 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2650 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2654 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2655 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2658 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2659 /// the two vector operands have swapped position.
2661 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2662 MVT MaskVT = Mask.getValueType();
2663 MVT EltVT = MaskVT.getVectorElementType();
2664 unsigned NumElems = Mask.getNumOperands();
2665 SmallVector<SDValue, 8> MaskVec;
2666 for (unsigned i = 0; i != NumElems; ++i) {
2667 SDValue Arg = Mask.getOperand(i);
2668 if (Arg.getOpcode() == ISD::UNDEF) {
2669 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2672 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2673 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2675 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2677 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2679 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2683 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2684 /// match movhlps. The lower half elements should come from upper half of
2685 /// V1 (and in order), and the upper half elements should come from the upper
2686 /// half of V2 (and in order).
2687 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2688 unsigned NumElems = Mask->getNumOperands();
2691 for (unsigned i = 0, e = 2; i != e; ++i)
2692 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2694 for (unsigned i = 2; i != 4; ++i)
2695 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2700 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2701 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2703 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2704 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2705 N = N->getOperand(0).getNode();
2706 if (ISD::isNON_EXTLoad(N)) {
2708 *LD = cast<LoadSDNode>(N);
2715 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2716 /// match movlp{s|d}. The lower half elements should come from lower half of
2717 /// V1 (and in order), and the upper half elements should come from the upper
2718 /// half of V2 (and in order). And since V1 will become the source of the
2719 /// MOVLP, it must be either a vector load or a scalar load to vector.
2720 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2721 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2723 // Is V2 is a vector load, don't do this transformation. We will try to use
2724 // load folding shufps op.
2725 if (ISD::isNON_EXTLoad(V2))
2728 unsigned NumElems = Mask->getNumOperands();
2729 if (NumElems != 2 && NumElems != 4)
2731 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2732 if (!isUndefOrEqual(Mask->getOperand(i), i))
2734 for (unsigned i = NumElems/2; i != NumElems; ++i)
2735 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2740 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2742 static bool isSplatVector(SDNode *N) {
2743 if (N->getOpcode() != ISD::BUILD_VECTOR)
2746 SDValue SplatValue = N->getOperand(0);
2747 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2748 if (N->getOperand(i) != SplatValue)
2753 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2755 static bool isUndefShuffle(SDNode *N) {
2756 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2759 SDValue V1 = N->getOperand(0);
2760 SDValue V2 = N->getOperand(1);
2761 SDValue Mask = N->getOperand(2);
2762 unsigned NumElems = Mask.getNumOperands();
2763 for (unsigned i = 0; i != NumElems; ++i) {
2764 SDValue Arg = Mask.getOperand(i);
2765 if (Arg.getOpcode() != ISD::UNDEF) {
2766 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2767 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2769 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2776 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2778 static inline bool isZeroNode(SDValue Elt) {
2779 return ((isa<ConstantSDNode>(Elt) &&
2780 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2781 (isa<ConstantFPSDNode>(Elt) &&
2782 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2785 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2786 /// to an zero vector.
2787 static bool isZeroShuffle(SDNode *N) {
2788 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2791 SDValue V1 = N->getOperand(0);
2792 SDValue V2 = N->getOperand(1);
2793 SDValue Mask = N->getOperand(2);
2794 unsigned NumElems = Mask.getNumOperands();
2795 for (unsigned i = 0; i != NumElems; ++i) {
2796 SDValue Arg = Mask.getOperand(i);
2797 if (Arg.getOpcode() == ISD::UNDEF)
2800 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2801 if (Idx < NumElems) {
2802 unsigned Opc = V1.getNode()->getOpcode();
2803 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2805 if (Opc != ISD::BUILD_VECTOR ||
2806 !isZeroNode(V1.getNode()->getOperand(Idx)))
2808 } else if (Idx >= NumElems) {
2809 unsigned Opc = V2.getNode()->getOpcode();
2810 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2812 if (Opc != ISD::BUILD_VECTOR ||
2813 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2820 /// getZeroVector - Returns a vector of specified type with all zero elements.
2822 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2823 assert(VT.isVector() && "Expected a vector type");
2825 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2826 // type. This ensures they get CSE'd.
2828 if (VT.getSizeInBits() == 64) { // MMX
2829 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2830 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2831 } else if (HasSSE2) { // SSE2
2832 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2833 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2835 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2836 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2838 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2841 /// getOnesVector - Returns a vector of specified type with all bits set.
2843 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2844 assert(VT.isVector() && "Expected a vector type");
2846 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2847 // type. This ensures they get CSE'd.
2848 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2850 if (VT.getSizeInBits() == 64) // MMX
2851 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2853 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2854 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2858 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2859 /// that point to V2 points to its first element.
2860 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2861 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2863 bool Changed = false;
2864 SmallVector<SDValue, 8> MaskVec;
2865 unsigned NumElems = Mask.getNumOperands();
2866 for (unsigned i = 0; i != NumElems; ++i) {
2867 SDValue Arg = Mask.getOperand(i);
2868 if (Arg.getOpcode() != ISD::UNDEF) {
2869 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2870 if (Val > NumElems) {
2871 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2875 MaskVec.push_back(Arg);
2879 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2880 &MaskVec[0], MaskVec.size());
2884 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2885 /// operation of specified width.
2886 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2887 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2888 MVT BaseVT = MaskVT.getVectorElementType();
2890 SmallVector<SDValue, 8> MaskVec;
2891 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2892 for (unsigned i = 1; i != NumElems; ++i)
2893 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2894 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2897 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2898 /// of specified width.
2899 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2900 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2901 MVT BaseVT = MaskVT.getVectorElementType();
2902 SmallVector<SDValue, 8> MaskVec;
2903 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2904 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2905 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2907 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2910 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2911 /// of specified width.
2912 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2913 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2914 MVT BaseVT = MaskVT.getVectorElementType();
2915 unsigned Half = NumElems/2;
2916 SmallVector<SDValue, 8> MaskVec;
2917 for (unsigned i = 0; i != Half; ++i) {
2918 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2919 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2921 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2924 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2925 /// element #0 of a vector with the specified index, leaving the rest of the
2926 /// elements in place.
2927 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2928 SelectionDAG &DAG) {
2929 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2930 MVT BaseVT = MaskVT.getVectorElementType();
2931 SmallVector<SDValue, 8> MaskVec;
2932 // Element #0 of the result gets the elt we are replacing.
2933 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2934 for (unsigned i = 1; i != NumElems; ++i)
2935 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2936 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2939 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2940 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2941 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2942 MVT VT = Op.getValueType();
2945 SDValue V1 = Op.getOperand(0);
2946 SDValue Mask = Op.getOperand(2);
2947 unsigned NumElems = Mask.getNumOperands();
2948 // Special handling of v4f32 -> v4i32.
2949 if (VT != MVT::v4f32) {
2950 Mask = getUnpacklMask(NumElems, DAG);
2951 while (NumElems > 4) {
2952 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2955 Mask = getZeroVector(MVT::v4i32, true, DAG);
2958 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2959 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2960 DAG.getNode(ISD::UNDEF, PVT), Mask);
2961 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2964 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2965 /// vector of zero or undef vector. This produces a shuffle where the low
2966 /// element of V2 is swizzled into the zero/undef vector, landing at element
2967 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2968 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2969 bool isZero, bool HasSSE2,
2970 SelectionDAG &DAG) {
2971 MVT VT = V2.getValueType();
2973 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
2974 unsigned NumElems = V2.getValueType().getVectorNumElements();
2975 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2976 MVT EVT = MaskVT.getVectorElementType();
2977 SmallVector<SDValue, 16> MaskVec;
2978 for (unsigned i = 0; i != NumElems; ++i)
2979 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2980 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2982 MaskVec.push_back(DAG.getConstant(i, EVT));
2983 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2984 &MaskVec[0], MaskVec.size());
2985 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2988 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2989 /// a shuffle that is zero.
2991 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
2992 unsigned NumElems, bool Low,
2993 SelectionDAG &DAG) {
2994 unsigned NumZeros = 0;
2995 for (unsigned i = 0; i < NumElems; ++i) {
2996 unsigned Index = Low ? i : NumElems-i-1;
2997 SDValue Idx = Mask.getOperand(Index);
2998 if (Idx.getOpcode() == ISD::UNDEF) {
3002 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3003 if (Elt.getNode() && isZeroNode(Elt))
3011 /// isVectorShift - Returns true if the shuffle can be implemented as a
3012 /// logical left or right shift of a vector.
3013 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3014 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3015 unsigned NumElems = Mask.getNumOperands();
3018 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3021 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3026 bool SeenV1 = false;
3027 bool SeenV2 = false;
3028 for (unsigned i = NumZeros; i < NumElems; ++i) {
3029 unsigned Val = isLeft ? (i - NumZeros) : i;
3030 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3031 if (Idx.getOpcode() == ISD::UNDEF)
3033 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3034 if (Index < NumElems)
3043 if (SeenV1 && SeenV2)
3046 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3052 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3054 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3055 unsigned NumNonZero, unsigned NumZero,
3056 SelectionDAG &DAG, TargetLowering &TLI) {
3062 for (unsigned i = 0; i < 16; ++i) {
3063 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3064 if (ThisIsNonZero && First) {
3066 V = getZeroVector(MVT::v8i16, true, DAG);
3068 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3073 SDValue ThisElt(0, 0), LastElt(0, 0);
3074 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3075 if (LastIsNonZero) {
3076 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3078 if (ThisIsNonZero) {
3079 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3080 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3081 ThisElt, DAG.getConstant(8, MVT::i8));
3083 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3087 if (ThisElt.getNode())
3088 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3089 DAG.getIntPtrConstant(i/2));
3093 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3096 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3098 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3099 unsigned NumNonZero, unsigned NumZero,
3100 SelectionDAG &DAG, TargetLowering &TLI) {
3106 for (unsigned i = 0; i < 8; ++i) {
3107 bool isNonZero = (NonZeros & (1 << i)) != 0;
3111 V = getZeroVector(MVT::v8i16, true, DAG);
3113 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3116 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3117 DAG.getIntPtrConstant(i));
3124 /// getVShift - Return a vector logical shift node.
3126 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3127 unsigned NumBits, SelectionDAG &DAG,
3128 const TargetLowering &TLI) {
3129 bool isMMX = VT.getSizeInBits() == 64;
3130 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3131 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3132 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3133 return DAG.getNode(ISD::BIT_CONVERT, VT,
3134 DAG.getNode(Opc, ShVT, SrcOp,
3135 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3139 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3140 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3141 if (ISD::isBuildVectorAllZeros(Op.getNode())
3142 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3143 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3144 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3145 // eliminated on x86-32 hosts.
3146 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3149 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3150 return getOnesVector(Op.getValueType(), DAG);
3151 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3154 MVT VT = Op.getValueType();
3155 MVT EVT = VT.getVectorElementType();
3156 unsigned EVTBits = EVT.getSizeInBits();
3158 unsigned NumElems = Op.getNumOperands();
3159 unsigned NumZero = 0;
3160 unsigned NumNonZero = 0;
3161 unsigned NonZeros = 0;
3162 bool IsAllConstants = true;
3163 SmallSet<SDValue, 8> Values;
3164 for (unsigned i = 0; i < NumElems; ++i) {
3165 SDValue Elt = Op.getOperand(i);
3166 if (Elt.getOpcode() == ISD::UNDEF)
3169 if (Elt.getOpcode() != ISD::Constant &&
3170 Elt.getOpcode() != ISD::ConstantFP)
3171 IsAllConstants = false;
3172 if (isZeroNode(Elt))
3175 NonZeros |= (1 << i);
3180 if (NumNonZero == 0) {
3181 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3182 return DAG.getNode(ISD::UNDEF, VT);
3185 // Special case for single non-zero, non-undef, element.
3186 if (NumNonZero == 1 && NumElems <= 4) {
3187 unsigned Idx = CountTrailingZeros_32(NonZeros);
3188 SDValue Item = Op.getOperand(Idx);
3190 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3191 // the value are obviously zero, truncate the value to i32 and do the
3192 // insertion that way. Only do this if the value is non-constant or if the
3193 // value is a constant being inserted into element 0. It is cheaper to do
3194 // a constant pool load than it is to do a movd + shuffle.
3195 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3196 (!IsAllConstants || Idx == 0)) {
3197 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3198 // Handle MMX and SSE both.
3199 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3200 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3202 // Truncate the value (which may itself be a constant) to i32, and
3203 // convert it to a vector with movd (S2V+shuffle to zero extend).
3204 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3205 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3206 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3207 Subtarget->hasSSE2(), DAG);
3209 // Now we have our 32-bit value zero extended in the low element of
3210 // a vector. If Idx != 0, swizzle it into place.
3213 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3214 getSwapEltZeroMask(VecElts, Idx, DAG)
3216 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3218 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3222 // If we have a constant or non-constant insertion into the low element of
3223 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3224 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3225 // depending on what the source datatype is. Because we can only get here
3226 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3228 // Don't do this for i64 values on x86-32.
3229 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3230 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3231 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3232 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3233 Subtarget->hasSSE2(), DAG);
3236 // Is it a vector logical left shift?
3237 if (NumElems == 2 && Idx == 1 &&
3238 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3239 unsigned NumBits = VT.getSizeInBits();
3240 return getVShift(true, VT,
3241 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3242 NumBits/2, DAG, *this);
3245 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3248 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3249 // is a non-constant being inserted into an element other than the low one,
3250 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3251 // movd/movss) to move this into the low element, then shuffle it into
3253 if (EVTBits == 32) {
3254 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3256 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3257 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3258 Subtarget->hasSSE2(), DAG);
3259 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3260 MVT MaskEVT = MaskVT.getVectorElementType();
3261 SmallVector<SDValue, 8> MaskVec;
3262 for (unsigned i = 0; i < NumElems; i++)
3263 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3264 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3265 &MaskVec[0], MaskVec.size());
3266 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3267 DAG.getNode(ISD::UNDEF, VT), Mask);
3271 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3272 if (Values.size() == 1)
3275 // A vector full of immediates; various special cases are already
3276 // handled, so this is best done with a single constant-pool load.
3280 // Let legalizer expand 2-wide build_vectors.
3281 if (EVTBits == 64) {
3282 if (NumNonZero == 1) {
3283 // One half is zero or undef.
3284 unsigned Idx = CountTrailingZeros_32(NonZeros);
3285 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3286 Op.getOperand(Idx));
3287 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3288 Subtarget->hasSSE2(), DAG);
3293 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3294 if (EVTBits == 8 && NumElems == 16) {
3295 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3297 if (V.getNode()) return V;
3300 if (EVTBits == 16 && NumElems == 8) {
3301 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3303 if (V.getNode()) return V;
3306 // If element VT is == 32 bits, turn it into a number of shuffles.
3307 SmallVector<SDValue, 8> V;
3309 if (NumElems == 4 && NumZero > 0) {
3310 for (unsigned i = 0; i < 4; ++i) {
3311 bool isZero = !(NonZeros & (1 << i));
3313 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3315 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3318 for (unsigned i = 0; i < 2; ++i) {
3319 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3322 V[i] = V[i*2]; // Must be a zero vector.
3325 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3326 getMOVLMask(NumElems, DAG));
3329 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3330 getMOVLMask(NumElems, DAG));
3333 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3334 getUnpacklMask(NumElems, DAG));
3339 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3340 MVT EVT = MaskVT.getVectorElementType();
3341 SmallVector<SDValue, 8> MaskVec;
3342 bool Reverse = (NonZeros & 0x3) == 2;
3343 for (unsigned i = 0; i < 2; ++i)
3345 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3347 MaskVec.push_back(DAG.getConstant(i, EVT));
3348 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3349 for (unsigned i = 0; i < 2; ++i)
3351 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3353 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3354 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3355 &MaskVec[0], MaskVec.size());
3356 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3359 if (Values.size() > 2) {
3360 // Expand into a number of unpckl*.
3362 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3363 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3364 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3365 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3366 for (unsigned i = 0; i < NumElems; ++i)
3367 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3369 while (NumElems != 0) {
3370 for (unsigned i = 0; i < NumElems; ++i)
3371 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3382 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3383 SDValue PermMask, SelectionDAG &DAG,
3384 TargetLowering &TLI) {
3386 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3387 MVT MaskEVT = MaskVT.getVectorElementType();
3388 MVT PtrVT = TLI.getPointerTy();
3389 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3390 PermMask.getNode()->op_end());
3392 // First record which half of which vector the low elements come from.
3393 SmallVector<unsigned, 4> LowQuad(4);
3394 for (unsigned i = 0; i < 4; ++i) {
3395 SDValue Elt = MaskElts[i];
3396 if (Elt.getOpcode() == ISD::UNDEF)
3398 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3399 int QuadIdx = EltIdx / 4;
3403 int BestLowQuad = -1;
3404 unsigned MaxQuad = 1;
3405 for (unsigned i = 0; i < 4; ++i) {
3406 if (LowQuad[i] > MaxQuad) {
3408 MaxQuad = LowQuad[i];
3412 // Record which half of which vector the high elements come from.
3413 SmallVector<unsigned, 4> HighQuad(4);
3414 for (unsigned i = 4; i < 8; ++i) {
3415 SDValue Elt = MaskElts[i];
3416 if (Elt.getOpcode() == ISD::UNDEF)
3418 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3419 int QuadIdx = EltIdx / 4;
3420 ++HighQuad[QuadIdx];
3423 int BestHighQuad = -1;
3425 for (unsigned i = 0; i < 4; ++i) {
3426 if (HighQuad[i] > MaxQuad) {
3428 MaxQuad = HighQuad[i];
3432 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3433 if (BestLowQuad != -1 || BestHighQuad != -1) {
3434 // First sort the 4 chunks in order using shufpd.
3435 SmallVector<SDValue, 8> MaskVec;
3437 if (BestLowQuad != -1)
3438 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3440 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3442 if (BestHighQuad != -1)
3443 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3445 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3447 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3448 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3449 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3450 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3451 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3453 // Now sort high and low parts separately.
3454 BitVector InOrder(8);
3455 if (BestLowQuad != -1) {
3456 // Sort lower half in order using PSHUFLW.
3458 bool AnyOutOrder = false;
3460 for (unsigned i = 0; i != 4; ++i) {
3461 SDValue Elt = MaskElts[i];
3462 if (Elt.getOpcode() == ISD::UNDEF) {
3463 MaskVec.push_back(Elt);
3466 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3470 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3472 // If this element is in the right place after this shuffle, then
3474 if ((int)(EltIdx / 4) == BestLowQuad)
3479 for (unsigned i = 4; i != 8; ++i)
3480 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3481 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3482 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3486 if (BestHighQuad != -1) {
3487 // Sort high half in order using PSHUFHW if possible.
3490 for (unsigned i = 0; i != 4; ++i)
3491 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3493 bool AnyOutOrder = false;
3494 for (unsigned i = 4; i != 8; ++i) {
3495 SDValue Elt = MaskElts[i];
3496 if (Elt.getOpcode() == ISD::UNDEF) {
3497 MaskVec.push_back(Elt);
3500 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3504 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3506 // If this element is in the right place after this shuffle, then
3508 if ((int)(EltIdx / 4) == BestHighQuad)
3514 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3515 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3519 // The other elements are put in the right place using pextrw and pinsrw.
3520 for (unsigned i = 0; i != 8; ++i) {
3523 SDValue Elt = MaskElts[i];
3524 if (Elt.getOpcode() == ISD::UNDEF)
3526 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3527 SDValue ExtOp = (EltIdx < 8)
3528 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3529 DAG.getConstant(EltIdx, PtrVT))
3530 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3531 DAG.getConstant(EltIdx - 8, PtrVT));
3532 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3533 DAG.getConstant(i, PtrVT));
3539 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3540 // few as possible. First, let's find out how many elements are already in the
3542 unsigned V1InOrder = 0;
3543 unsigned V1FromV1 = 0;
3544 unsigned V2InOrder = 0;
3545 unsigned V2FromV2 = 0;
3546 SmallVector<SDValue, 8> V1Elts;
3547 SmallVector<SDValue, 8> V2Elts;
3548 for (unsigned i = 0; i < 8; ++i) {
3549 SDValue Elt = MaskElts[i];
3550 if (Elt.getOpcode() == ISD::UNDEF) {
3551 V1Elts.push_back(Elt);
3552 V2Elts.push_back(Elt);
3557 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3559 V1Elts.push_back(Elt);
3560 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3562 } else if (EltIdx == i+8) {
3563 V1Elts.push_back(Elt);
3564 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3566 } else if (EltIdx < 8) {
3567 V1Elts.push_back(Elt);
3570 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3575 if (V2InOrder > V1InOrder) {
3576 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3578 std::swap(V1Elts, V2Elts);
3579 std::swap(V1FromV1, V2FromV2);
3582 if ((V1FromV1 + V1InOrder) != 8) {
3583 // Some elements are from V2.
3585 // If there are elements that are from V1 but out of place,
3586 // then first sort them in place
3587 SmallVector<SDValue, 8> MaskVec;
3588 for (unsigned i = 0; i < 8; ++i) {
3589 SDValue Elt = V1Elts[i];
3590 if (Elt.getOpcode() == ISD::UNDEF) {
3591 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3594 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3596 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3598 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3600 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3601 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3605 for (unsigned i = 0; i < 8; ++i) {
3606 SDValue Elt = V1Elts[i];
3607 if (Elt.getOpcode() == ISD::UNDEF)
3609 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3612 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3613 DAG.getConstant(EltIdx - 8, PtrVT));
3614 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3615 DAG.getConstant(i, PtrVT));
3619 // All elements are from V1.
3621 for (unsigned i = 0; i < 8; ++i) {
3622 SDValue Elt = V1Elts[i];
3623 if (Elt.getOpcode() == ISD::UNDEF)
3625 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3626 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3627 DAG.getConstant(EltIdx, PtrVT));
3628 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3629 DAG.getConstant(i, PtrVT));
3635 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3636 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3637 /// done when every pair / quad of shuffle mask elements point to elements in
3638 /// the right sequence. e.g.
3639 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3641 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3643 SDValue PermMask, SelectionDAG &DAG,
3644 TargetLowering &TLI) {
3645 unsigned NumElems = PermMask.getNumOperands();
3646 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3647 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3648 MVT MaskEltVT = MaskVT.getVectorElementType();
3650 switch (VT.getSimpleVT()) {
3651 default: assert(false && "Unexpected!");
3652 case MVT::v4f32: NewVT = MVT::v2f64; break;
3653 case MVT::v4i32: NewVT = MVT::v2i64; break;
3654 case MVT::v8i16: NewVT = MVT::v4i32; break;
3655 case MVT::v16i8: NewVT = MVT::v4i32; break;
3658 if (NewWidth == 2) {
3664 unsigned Scale = NumElems / NewWidth;
3665 SmallVector<SDValue, 8> MaskVec;
3666 for (unsigned i = 0; i < NumElems; i += Scale) {
3667 unsigned StartIdx = ~0U;
3668 for (unsigned j = 0; j < Scale; ++j) {
3669 SDValue Elt = PermMask.getOperand(i+j);
3670 if (Elt.getOpcode() == ISD::UNDEF)
3672 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3673 if (StartIdx == ~0U)
3674 StartIdx = EltIdx - (EltIdx % Scale);
3675 if (EltIdx != StartIdx + j)
3678 if (StartIdx == ~0U)
3679 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3681 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3684 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3685 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3686 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3687 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3688 &MaskVec[0], MaskVec.size()));
3691 /// getVZextMovL - Return a zero-extending vector move low node.
3693 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3694 SDValue SrcOp, SelectionDAG &DAG,
3695 const X86Subtarget *Subtarget) {
3696 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3697 LoadSDNode *LD = NULL;
3698 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3699 LD = dyn_cast<LoadSDNode>(SrcOp);
3701 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3703 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3704 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3705 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3706 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3707 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3709 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3710 return DAG.getNode(ISD::BIT_CONVERT, VT,
3711 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3712 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3719 return DAG.getNode(ISD::BIT_CONVERT, VT,
3720 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3721 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3724 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3727 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3728 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3729 MVT MaskVT = PermMask.getValueType();
3730 MVT MaskEVT = MaskVT.getVectorElementType();
3731 SmallVector<std::pair<int, int>, 8> Locs;
3733 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3736 for (unsigned i = 0; i != 4; ++i) {
3737 SDValue Elt = PermMask.getOperand(i);
3738 if (Elt.getOpcode() == ISD::UNDEF) {
3739 Locs[i] = std::make_pair(-1, -1);
3741 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3742 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3744 Locs[i] = std::make_pair(0, NumLo);
3748 Locs[i] = std::make_pair(1, NumHi);
3750 Mask1[2+NumHi] = Elt;
3756 if (NumLo <= 2 && NumHi <= 2) {
3757 // If no more than two elements come from either vector. This can be
3758 // implemented with two shuffles. First shuffle gather the elements.
3759 // The second shuffle, which takes the first shuffle as both of its
3760 // vector operands, put the elements into the right order.
3761 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3762 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3763 &Mask1[0], Mask1.size()));
3765 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3766 for (unsigned i = 0; i != 4; ++i) {
3767 if (Locs[i].first == -1)
3770 unsigned Idx = (i < 2) ? 0 : 4;
3771 Idx += Locs[i].first * 2 + Locs[i].second;
3772 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3776 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3777 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3778 &Mask2[0], Mask2.size()));
3779 } else if (NumLo == 3 || NumHi == 3) {
3780 // Otherwise, we must have three elements from one vector, call it X, and
3781 // one element from the other, call it Y. First, use a shufps to build an
3782 // intermediate vector with the one element from Y and the element from X
3783 // that will be in the same half in the final destination (the indexes don't
3784 // matter). Then, use a shufps to build the final vector, taking the half
3785 // containing the element from Y from the intermediate, and the other half
3788 // Normalize it so the 3 elements come from V1.
3789 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3793 // Find the element from V2.
3795 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3796 SDValue Elt = PermMask.getOperand(HiIndex);
3797 if (Elt.getOpcode() == ISD::UNDEF)
3799 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3804 Mask1[0] = PermMask.getOperand(HiIndex);
3805 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3806 Mask1[2] = PermMask.getOperand(HiIndex^1);
3807 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3808 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3809 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3812 Mask1[0] = PermMask.getOperand(0);
3813 Mask1[1] = PermMask.getOperand(1);
3814 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3815 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3816 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3817 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3819 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3820 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3821 Mask1[2] = PermMask.getOperand(2);
3822 Mask1[3] = PermMask.getOperand(3);
3823 if (Mask1[2].getOpcode() != ISD::UNDEF)
3825 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
3827 if (Mask1[3].getOpcode() != ISD::UNDEF)
3829 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
3831 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3832 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3836 // Break it into (shuffle shuffle_hi, shuffle_lo).
3838 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3839 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3840 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3841 unsigned MaskIdx = 0;
3844 for (unsigned i = 0; i != 4; ++i) {
3851 SDValue Elt = PermMask.getOperand(i);
3852 if (Elt.getOpcode() == ISD::UNDEF) {
3853 Locs[i] = std::make_pair(-1, -1);
3854 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
3855 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3856 (*MaskPtr)[LoIdx] = Elt;
3859 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3860 (*MaskPtr)[HiIdx] = Elt;
3865 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3866 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3867 &LoMask[0], LoMask.size()));
3868 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3869 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3870 &HiMask[0], HiMask.size()));
3871 SmallVector<SDValue, 8> MaskOps;
3872 for (unsigned i = 0; i != 4; ++i) {
3873 if (Locs[i].first == -1) {
3874 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3876 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3877 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3880 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3881 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3882 &MaskOps[0], MaskOps.size()));
3886 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3887 SDValue V1 = Op.getOperand(0);
3888 SDValue V2 = Op.getOperand(1);
3889 SDValue PermMask = Op.getOperand(2);
3890 MVT VT = Op.getValueType();
3891 unsigned NumElems = PermMask.getNumOperands();
3892 bool isMMX = VT.getSizeInBits() == 64;
3893 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3894 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3895 bool V1IsSplat = false;
3896 bool V2IsSplat = false;
3898 if (isUndefShuffle(Op.getNode()))
3899 return DAG.getNode(ISD::UNDEF, VT);
3901 if (isZeroShuffle(Op.getNode()))
3902 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3904 if (isIdentityMask(PermMask.getNode()))
3906 else if (isIdentityMask(PermMask.getNode(), true))
3909 if (isSplatMask(PermMask.getNode())) {
3910 if (isMMX || NumElems < 4) return Op;
3911 // Promote it to a v4{if}32 splat.
3912 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3915 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3917 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3918 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3919 if (NewOp.getNode())
3920 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3921 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3922 // FIXME: Figure out a cleaner way to do this.
3923 // Try to make use of movq to zero out the top part.
3924 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3925 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3927 if (NewOp.getNode()) {
3928 SDValue NewV1 = NewOp.getOperand(0);
3929 SDValue NewV2 = NewOp.getOperand(1);
3930 SDValue NewMask = NewOp.getOperand(2);
3931 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3932 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3933 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3936 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3937 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3939 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3940 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3945 // Check if this can be converted into a logical shift.
3946 bool isLeft = false;
3949 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
3950 if (isShift && ShVal.hasOneUse()) {
3951 // If the shifted value has multiple uses, it may be cheaper to use
3952 // v_set0 + movlhps or movhlps, etc.
3953 MVT EVT = VT.getVectorElementType();
3954 ShAmt *= EVT.getSizeInBits();
3955 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3958 if (X86::isMOVLMask(PermMask.getNode())) {
3961 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3962 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
3967 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
3968 X86::isMOVSLDUPMask(PermMask.getNode()) ||
3969 X86::isMOVHLPSMask(PermMask.getNode()) ||
3970 X86::isMOVHPMask(PermMask.getNode()) ||
3971 X86::isMOVLPMask(PermMask.getNode())))
3974 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
3975 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
3976 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3979 // No better options. Use a vshl / vsrl.
3980 MVT EVT = VT.getVectorElementType();
3981 ShAmt *= EVT.getSizeInBits();
3982 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3985 bool Commuted = false;
3986 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3987 // 1,1,1,1 -> v8i16 though.
3988 V1IsSplat = isSplatVector(V1.getNode());
3989 V2IsSplat = isSplatVector(V2.getNode());
3991 // Canonicalize the splat or undef, if present, to be on the RHS.
3992 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3993 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3994 std::swap(V1IsSplat, V2IsSplat);
3995 std::swap(V1IsUndef, V2IsUndef);
3999 // FIXME: Figure out a cleaner way to do this.
4000 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4001 if (V2IsUndef) return V1;
4002 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4004 // V2 is a splat, so the mask may be malformed. That is, it may point
4005 // to any V2 element. The instruction selectior won't like this. Get
4006 // a corrected mask and commute to form a proper MOVS{S|D}.
4007 SDValue NewMask = getMOVLMask(NumElems, DAG);
4008 if (NewMask.getNode() != PermMask.getNode())
4009 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4014 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4015 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4016 X86::isUNPCKLMask(PermMask.getNode()) ||
4017 X86::isUNPCKHMask(PermMask.getNode()))
4021 // Normalize mask so all entries that point to V2 points to its first
4022 // element then try to match unpck{h|l} again. If match, return a
4023 // new vector_shuffle with the corrected mask.
4024 SDValue NewMask = NormalizeMask(PermMask, DAG);
4025 if (NewMask.getNode() != PermMask.getNode()) {
4026 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4027 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4028 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4029 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4030 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4031 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4036 // Normalize the node to match x86 shuffle ops if needed
4037 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4038 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4041 // Commute is back and try unpck* again.
4042 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4043 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4044 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4045 X86::isUNPCKLMask(PermMask.getNode()) ||
4046 X86::isUNPCKHMask(PermMask.getNode()))
4050 // Try PSHUF* first, then SHUFP*.
4051 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4052 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4053 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4054 if (V2.getOpcode() != ISD::UNDEF)
4055 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4056 DAG.getNode(ISD::UNDEF, VT), PermMask);
4061 if (Subtarget->hasSSE2() &&
4062 (X86::isPSHUFDMask(PermMask.getNode()) ||
4063 X86::isPSHUFHWMask(PermMask.getNode()) ||
4064 X86::isPSHUFLWMask(PermMask.getNode()))) {
4066 if (VT == MVT::v4f32) {
4068 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4069 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4070 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4071 } else if (V2.getOpcode() != ISD::UNDEF)
4072 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4073 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4075 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4079 // Binary or unary shufps.
4080 if (X86::isSHUFPMask(PermMask.getNode()) ||
4081 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4085 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4086 if (VT == MVT::v8i16) {
4087 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4088 if (NewOp.getNode())
4092 // Handle all 4 wide cases with a number of shuffles except for MMX.
4093 if (NumElems == 4 && !isMMX)
4094 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4100 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4101 SelectionDAG &DAG) {
4102 MVT VT = Op.getValueType();
4103 if (VT.getSizeInBits() == 8) {
4104 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4105 Op.getOperand(0), Op.getOperand(1));
4106 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4107 DAG.getValueType(VT));
4108 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4109 } else if (VT.getSizeInBits() == 16) {
4110 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4111 Op.getOperand(0), Op.getOperand(1));
4112 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4113 DAG.getValueType(VT));
4114 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4115 } else if (VT == MVT::f32) {
4116 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4117 // the result back to FR32 register. It's only worth matching if the
4118 // result has a single use which is a store or a bitcast to i32.
4119 if (!Op.hasOneUse())
4121 SDNode *User = *Op.getNode()->use_begin();
4122 if (User->getOpcode() != ISD::STORE &&
4123 (User->getOpcode() != ISD::BIT_CONVERT ||
4124 User->getValueType(0) != MVT::i32))
4126 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4127 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4129 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4136 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4137 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4140 if (Subtarget->hasSSE41()) {
4141 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4146 MVT VT = Op.getValueType();
4147 // TODO: handle v16i8.
4148 if (VT.getSizeInBits() == 16) {
4149 SDValue Vec = Op.getOperand(0);
4150 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4152 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4153 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4154 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4156 // Transform it so it match pextrw which produces a 32-bit result.
4157 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4158 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4159 Op.getOperand(0), Op.getOperand(1));
4160 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4161 DAG.getValueType(VT));
4162 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4163 } else if (VT.getSizeInBits() == 32) {
4164 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4167 // SHUFPS the element to the lowest double word, then movss.
4168 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4169 SmallVector<SDValue, 8> IdxVec;
4171 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4173 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4175 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4177 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4178 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4179 &IdxVec[0], IdxVec.size());
4180 SDValue Vec = Op.getOperand(0);
4181 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4182 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4183 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4184 DAG.getIntPtrConstant(0));
4185 } else if (VT.getSizeInBits() == 64) {
4186 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4187 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4188 // to match extract_elt for f64.
4189 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4193 // UNPCKHPD the element to the lowest double word, then movsd.
4194 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4195 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4196 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4197 SmallVector<SDValue, 8> IdxVec;
4198 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4200 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4201 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4202 &IdxVec[0], IdxVec.size());
4203 SDValue Vec = Op.getOperand(0);
4204 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4205 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4206 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4207 DAG.getIntPtrConstant(0));
4214 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4215 MVT VT = Op.getValueType();
4216 MVT EVT = VT.getVectorElementType();
4218 SDValue N0 = Op.getOperand(0);
4219 SDValue N1 = Op.getOperand(1);
4220 SDValue N2 = Op.getOperand(2);
4222 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4223 isa<ConstantSDNode>(N2)) {
4224 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4226 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4228 if (N1.getValueType() != MVT::i32)
4229 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4230 if (N2.getValueType() != MVT::i32)
4231 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4232 return DAG.getNode(Opc, VT, N0, N1, N2);
4233 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4234 // Bits [7:6] of the constant are the source select. This will always be
4235 // zero here. The DAG Combiner may combine an extract_elt index into these
4236 // bits. For example (insert (extract, 3), 2) could be matched by putting
4237 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4238 // Bits [5:4] of the constant are the destination select. This is the
4239 // value of the incoming immediate.
4240 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4241 // combine either bitwise AND or insert of float 0.0 to set these bits.
4242 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4243 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4249 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4250 MVT VT = Op.getValueType();
4251 MVT EVT = VT.getVectorElementType();
4253 if (Subtarget->hasSSE41())
4254 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4259 SDValue N0 = Op.getOperand(0);
4260 SDValue N1 = Op.getOperand(1);
4261 SDValue N2 = Op.getOperand(2);
4263 if (EVT.getSizeInBits() == 16) {
4264 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4265 // as its second argument.
4266 if (N1.getValueType() != MVT::i32)
4267 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4268 if (N2.getValueType() != MVT::i32)
4269 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4270 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4276 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4277 if (Op.getValueType() == MVT::v2f32)
4278 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4279 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4280 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4281 Op.getOperand(0))));
4283 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4284 MVT VT = MVT::v2i32;
4285 switch (Op.getValueType().getSimpleVT()) {
4292 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4293 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4296 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4297 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4298 // one of the above mentioned nodes. It has to be wrapped because otherwise
4299 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4300 // be used to form addressing mode. These wrapped nodes will be selected
4303 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4304 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4305 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4307 CP->getAlignment());
4308 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4309 // With PIC, the address is actually $g + Offset.
4310 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4311 !Subtarget->isPICStyleRIPRel()) {
4312 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4313 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4321 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4322 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4323 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4324 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4325 // With PIC, the address is actually $g + Offset.
4326 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4327 !Subtarget->isPICStyleRIPRel()) {
4328 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4329 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4333 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4334 // load the value at address GV, not the value of GV itself. This means that
4335 // the GlobalAddress must be in the base or index register of the address, not
4336 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4337 // The same applies for external symbols during PIC codegen
4338 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4339 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4340 PseudoSourceValue::getGOT(), 0);
4345 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4347 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4350 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4351 DAG.getNode(X86ISD::GlobalBaseReg,
4353 InFlag = Chain.getValue(1);
4355 // emit leal symbol@TLSGD(,%ebx,1), %eax
4356 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4357 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4358 GA->getValueType(0),
4360 SDValue Ops[] = { Chain, TGA, InFlag };
4361 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4362 InFlag = Result.getValue(2);
4363 Chain = Result.getValue(1);
4365 // call ___tls_get_addr. This function receives its argument in
4366 // the register EAX.
4367 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4368 InFlag = Chain.getValue(1);
4370 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4371 SDValue Ops1[] = { Chain,
4372 DAG.getTargetExternalSymbol("___tls_get_addr",
4374 DAG.getRegister(X86::EAX, PtrVT),
4375 DAG.getRegister(X86::EBX, PtrVT),
4377 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4378 InFlag = Chain.getValue(1);
4380 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4383 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4385 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4387 SDValue InFlag, Chain;
4389 // emit leaq symbol@TLSGD(%rip), %rdi
4390 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4391 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4392 GA->getValueType(0),
4394 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4395 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4396 Chain = Result.getValue(1);
4397 InFlag = Result.getValue(2);
4399 // call __tls_get_addr. This function receives its argument in
4400 // the register RDI.
4401 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4402 InFlag = Chain.getValue(1);
4404 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4405 SDValue Ops1[] = { Chain,
4406 DAG.getTargetExternalSymbol("__tls_get_addr",
4408 DAG.getRegister(X86::RDI, PtrVT),
4410 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4411 InFlag = Chain.getValue(1);
4413 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4416 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4417 // "local exec" model.
4418 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4420 // Get the Thread Pointer
4421 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4422 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4424 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4425 GA->getValueType(0),
4427 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4429 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4430 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4431 PseudoSourceValue::getGOT(), 0);
4433 // The address of the thread local variable is the add of the thread
4434 // pointer with the offset of the variable.
4435 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4439 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4440 // TODO: implement the "local dynamic" model
4441 // TODO: implement the "initial exec"model for pic executables
4442 assert(Subtarget->isTargetELF() &&
4443 "TLS not implemented for non-ELF targets");
4444 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4445 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4446 // otherwise use the "Local Exec"TLS Model
4447 if (Subtarget->is64Bit()) {
4448 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4450 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4451 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4453 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4458 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4459 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4460 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4461 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4462 // With PIC, the address is actually $g + Offset.
4463 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4464 !Subtarget->isPICStyleRIPRel()) {
4465 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4466 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4473 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4474 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4475 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4476 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4477 // With PIC, the address is actually $g + Offset.
4478 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4479 !Subtarget->isPICStyleRIPRel()) {
4480 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4481 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4488 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4489 /// take a 2 x i32 value to shift plus a shift amount.
4490 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4491 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4492 MVT VT = Op.getValueType();
4493 unsigned VTBits = VT.getSizeInBits();
4494 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4495 SDValue ShOpLo = Op.getOperand(0);
4496 SDValue ShOpHi = Op.getOperand(1);
4497 SDValue ShAmt = Op.getOperand(2);
4498 SDValue Tmp1 = isSRA ?
4499 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4500 DAG.getConstant(0, VT);
4503 if (Op.getOpcode() == ISD::SHL_PARTS) {
4504 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4505 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4507 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4508 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4511 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4512 DAG.getConstant(VTBits, MVT::i8));
4513 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4514 AndNode, DAG.getConstant(0, MVT::i8));
4517 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4518 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4519 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4521 if (Op.getOpcode() == ISD::SHL_PARTS) {
4522 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4523 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4525 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4526 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4529 SDValue Ops[2] = { Lo, Hi };
4530 return DAG.getMergeValues(Ops, 2);
4533 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4534 MVT SrcVT = Op.getOperand(0).getValueType();
4535 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4536 "Unknown SINT_TO_FP to lower!");
4538 // These are really Legal; caller falls through into that case.
4539 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4541 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4542 Subtarget->is64Bit())
4545 unsigned Size = SrcVT.getSizeInBits()/8;
4546 MachineFunction &MF = DAG.getMachineFunction();
4547 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4548 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4549 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4551 PseudoSourceValue::getFixedStack(SSFI), 0);
4555 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4557 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4559 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4560 SmallVector<SDValue, 8> Ops;
4561 Ops.push_back(Chain);
4562 Ops.push_back(StackSlot);
4563 Ops.push_back(DAG.getValueType(SrcVT));
4564 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4565 Tys, &Ops[0], Ops.size());
4568 Chain = Result.getValue(1);
4569 SDValue InFlag = Result.getValue(2);
4571 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4572 // shouldn't be necessary except that RFP cannot be live across
4573 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4574 MachineFunction &MF = DAG.getMachineFunction();
4575 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4576 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4577 Tys = DAG.getVTList(MVT::Other);
4578 SmallVector<SDValue, 8> Ops;
4579 Ops.push_back(Chain);
4580 Ops.push_back(Result);
4581 Ops.push_back(StackSlot);
4582 Ops.push_back(DAG.getValueType(Op.getValueType()));
4583 Ops.push_back(InFlag);
4584 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4585 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4586 PseudoSourceValue::getFixedStack(SSFI), 0);
4592 std::pair<SDValue,SDValue> X86TargetLowering::
4593 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4594 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4595 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4596 "Unknown FP_TO_SINT to lower!");
4598 // These are really Legal.
4599 if (Op.getValueType() == MVT::i32 &&
4600 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4601 return std::make_pair(SDValue(), SDValue());
4602 if (Subtarget->is64Bit() &&
4603 Op.getValueType() == MVT::i64 &&
4604 Op.getOperand(0).getValueType() != MVT::f80)
4605 return std::make_pair(SDValue(), SDValue());
4607 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4609 MachineFunction &MF = DAG.getMachineFunction();
4610 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4611 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4612 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4614 switch (Op.getValueType().getSimpleVT()) {
4615 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4616 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4617 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4618 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4621 SDValue Chain = DAG.getEntryNode();
4622 SDValue Value = Op.getOperand(0);
4623 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4624 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4625 Chain = DAG.getStore(Chain, Value, StackSlot,
4626 PseudoSourceValue::getFixedStack(SSFI), 0);
4627 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4629 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4631 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4632 Chain = Value.getValue(1);
4633 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4634 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4637 // Build the FP_TO_INT*_IN_MEM
4638 SDValue Ops[] = { Chain, Value, StackSlot };
4639 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4641 return std::make_pair(FIST, StackSlot);
4644 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4645 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4646 SDValue FIST = Vals.first, StackSlot = Vals.second;
4647 if (FIST.getNode() == 0) return SDValue();
4650 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4653 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4654 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4655 SDValue FIST = Vals.first, StackSlot = Vals.second;
4656 if (FIST.getNode() == 0) return 0;
4658 MVT VT = N->getValueType(0);
4660 // Return a load from the stack slot.
4661 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4663 // Use MERGE_VALUES to drop the chain result value and get a node with one
4664 // result. This requires turning off getMergeValues simplification, since
4665 // otherwise it will give us Res back.
4666 return DAG.getMergeValues(&Res, 1, false).getNode();
4669 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4670 MVT VT = Op.getValueType();
4673 EltVT = VT.getVectorElementType();
4674 std::vector<Constant*> CV;
4675 if (EltVT == MVT::f64) {
4676 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4680 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4686 Constant *C = ConstantVector::get(CV);
4687 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4688 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4689 PseudoSourceValue::getConstantPool(), 0,
4691 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4694 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4695 MVT VT = Op.getValueType();
4697 unsigned EltNum = 1;
4698 if (VT.isVector()) {
4699 EltVT = VT.getVectorElementType();
4700 EltNum = VT.getVectorNumElements();
4702 std::vector<Constant*> CV;
4703 if (EltVT == MVT::f64) {
4704 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4708 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4714 Constant *C = ConstantVector::get(CV);
4715 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4716 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4717 PseudoSourceValue::getConstantPool(), 0,
4719 if (VT.isVector()) {
4720 return DAG.getNode(ISD::BIT_CONVERT, VT,
4721 DAG.getNode(ISD::XOR, MVT::v2i64,
4722 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4723 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4725 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4729 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4730 SDValue Op0 = Op.getOperand(0);
4731 SDValue Op1 = Op.getOperand(1);
4732 MVT VT = Op.getValueType();
4733 MVT SrcVT = Op1.getValueType();
4735 // If second operand is smaller, extend it first.
4736 if (SrcVT.bitsLT(VT)) {
4737 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4740 // And if it is bigger, shrink it first.
4741 if (SrcVT.bitsGT(VT)) {
4742 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4746 // At this point the operands and the result should have the same
4747 // type, and that won't be f80 since that is not custom lowered.
4749 // First get the sign bit of second operand.
4750 std::vector<Constant*> CV;
4751 if (SrcVT == MVT::f64) {
4752 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4753 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4755 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4756 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4757 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4758 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4760 Constant *C = ConstantVector::get(CV);
4761 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4762 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4763 PseudoSourceValue::getConstantPool(), 0,
4765 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4767 // Shift sign bit right or left if the two operands have different types.
4768 if (SrcVT.bitsGT(VT)) {
4769 // Op0 is MVT::f32, Op1 is MVT::f64.
4770 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4771 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4772 DAG.getConstant(32, MVT::i32));
4773 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4774 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4775 DAG.getIntPtrConstant(0));
4778 // Clear first operand sign bit.
4780 if (VT == MVT::f64) {
4781 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4782 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4784 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4785 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4786 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4787 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4789 C = ConstantVector::get(CV);
4790 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4791 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4792 PseudoSourceValue::getConstantPool(), 0,
4794 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4796 // Or the value with the sign bit.
4797 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4800 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4801 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4803 SDValue Op0 = Op.getOperand(0);
4804 SDValue Op1 = Op.getOperand(1);
4805 SDValue CC = Op.getOperand(2);
4806 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4807 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4810 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4812 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4813 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4814 DAG.getConstant(X86CC, MVT::i8), Cond);
4817 assert(isFP && "Illegal integer SetCC!");
4819 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4820 switch (SetCCOpcode) {
4821 default: assert(false && "Illegal floating point SetCC!");
4822 case ISD::SETOEQ: { // !PF & ZF
4823 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4824 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4825 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4826 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4827 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4829 case ISD::SETUNE: { // PF | !ZF
4830 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4831 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4832 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4833 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4834 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4839 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4841 SDValue Op0 = Op.getOperand(0);
4842 SDValue Op1 = Op.getOperand(1);
4843 SDValue CC = Op.getOperand(2);
4844 MVT VT = Op.getValueType();
4845 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4846 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4850 MVT VT0 = Op0.getValueType();
4851 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4852 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4855 switch (SetCCOpcode) {
4858 case ISD::SETEQ: SSECC = 0; break;
4860 case ISD::SETGT: Swap = true; // Fallthrough
4862 case ISD::SETOLT: SSECC = 1; break;
4864 case ISD::SETGE: Swap = true; // Fallthrough
4866 case ISD::SETOLE: SSECC = 2; break;
4867 case ISD::SETUO: SSECC = 3; break;
4869 case ISD::SETNE: SSECC = 4; break;
4870 case ISD::SETULE: Swap = true;
4871 case ISD::SETUGE: SSECC = 5; break;
4872 case ISD::SETULT: Swap = true;
4873 case ISD::SETUGT: SSECC = 6; break;
4874 case ISD::SETO: SSECC = 7; break;
4877 std::swap(Op0, Op1);
4879 // In the two special cases we can't handle, emit two comparisons.
4881 if (SetCCOpcode == ISD::SETUEQ) {
4883 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4884 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4885 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4887 else if (SetCCOpcode == ISD::SETONE) {
4889 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4890 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4891 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4893 assert(0 && "Illegal FP comparison");
4895 // Handle all other FP comparisons here.
4896 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4899 // We are handling one of the integer comparisons here. Since SSE only has
4900 // GT and EQ comparisons for integer, swapping operands and multiple
4901 // operations may be required for some comparisons.
4902 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4903 bool Swap = false, Invert = false, FlipSigns = false;
4905 switch (VT.getSimpleVT()) {
4907 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4908 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4909 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4910 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4913 switch (SetCCOpcode) {
4915 case ISD::SETNE: Invert = true;
4916 case ISD::SETEQ: Opc = EQOpc; break;
4917 case ISD::SETLT: Swap = true;
4918 case ISD::SETGT: Opc = GTOpc; break;
4919 case ISD::SETGE: Swap = true;
4920 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4921 case ISD::SETULT: Swap = true;
4922 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4923 case ISD::SETUGE: Swap = true;
4924 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4927 std::swap(Op0, Op1);
4929 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4930 // bits of the inputs before performing those operations.
4932 MVT EltVT = VT.getVectorElementType();
4933 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4934 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4935 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4937 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
4938 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
4941 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
4943 // If the logical-not of the result is required, perform that now.
4945 MVT EltVT = VT.getVectorElementType();
4946 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
4947 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
4948 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
4950 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
4955 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
4956 bool addTest = true;
4957 SDValue Cond = Op.getOperand(0);
4960 if (Cond.getOpcode() == ISD::SETCC)
4961 Cond = LowerSETCC(Cond, DAG);
4963 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4964 // setting operand in place of the X86ISD::SETCC.
4965 if (Cond.getOpcode() == X86ISD::SETCC) {
4966 CC = Cond.getOperand(0);
4968 SDValue Cmp = Cond.getOperand(1);
4969 unsigned Opc = Cmp.getOpcode();
4970 MVT VT = Op.getValueType();
4972 bool IllegalFPCMov = false;
4973 if (VT.isFloatingPoint() && !VT.isVector() &&
4974 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4975 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4977 if ((Opc == X86ISD::CMP ||
4978 Opc == X86ISD::COMI ||
4979 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4986 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4987 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4990 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4992 SmallVector<SDValue, 4> Ops;
4993 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4994 // condition is true.
4995 Ops.push_back(Op.getOperand(2));
4996 Ops.push_back(Op.getOperand(1));
4998 Ops.push_back(Cond);
4999 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
5002 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5003 bool addTest = true;
5004 SDValue Chain = Op.getOperand(0);
5005 SDValue Cond = Op.getOperand(1);
5006 SDValue Dest = Op.getOperand(2);
5009 if (Cond.getOpcode() == ISD::SETCC)
5010 Cond = LowerSETCC(Cond, DAG);
5012 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5013 // setting operand in place of the X86ISD::SETCC.
5014 if (Cond.getOpcode() == X86ISD::SETCC) {
5015 CC = Cond.getOperand(0);
5017 SDValue Cmp = Cond.getOperand(1);
5018 unsigned Opc = Cmp.getOpcode();
5019 if (Opc == X86ISD::CMP ||
5020 Opc == X86ISD::COMI ||
5021 Opc == X86ISD::UCOMI) {
5028 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5029 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5031 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5032 Chain, Op.getOperand(2), CC, Cond);
5036 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5037 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5038 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5039 // that the guard pages used by the OS virtual memory manager are allocated in
5040 // correct sequence.
5042 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5043 SelectionDAG &DAG) {
5044 assert(Subtarget->isTargetCygMing() &&
5045 "This should be used only on Cygwin/Mingw targets");
5048 SDValue Chain = Op.getOperand(0);
5049 SDValue Size = Op.getOperand(1);
5050 // FIXME: Ensure alignment here
5054 MVT IntPtr = getPointerTy();
5055 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5057 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0));
5059 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5060 Flag = Chain.getValue(1);
5062 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5063 SDValue Ops[] = { Chain,
5064 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5065 DAG.getRegister(X86::EAX, IntPtr),
5066 DAG.getRegister(X86StackPtr, SPTy),
5068 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5069 Flag = Chain.getValue(1);
5071 Chain = DAG.getCALLSEQ_END(Chain,
5072 DAG.getIntPtrConstant(0),
5073 DAG.getIntPtrConstant(0),
5076 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5078 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5079 return DAG.getMergeValues(Ops1, 2);
5083 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5085 SDValue Dst, SDValue Src,
5086 SDValue Size, unsigned Align,
5087 const Value *DstSV, uint64_t DstSVOff) {
5088 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5090 /// If not DWORD aligned or size is more than the threshold, call the library.
5091 /// The libc version is likely to be faster for these cases. It can use the
5092 /// address value and run time information about the CPU.
5093 if ((Align & 3) != 0 ||
5095 ConstantSize->getZExtValue() >
5096 getSubtarget()->getMaxInlineSizeThreshold()) {
5097 SDValue InFlag(0, 0);
5099 // Check to see if there is a specialized entry-point for memory zeroing.
5100 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5101 if (const char *bzeroEntry =
5102 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5103 MVT IntPtr = getPointerTy();
5104 const Type *IntPtrTy = TD->getIntPtrType();
5105 TargetLowering::ArgListTy Args;
5106 TargetLowering::ArgListEntry Entry;
5108 Entry.Ty = IntPtrTy;
5109 Args.push_back(Entry);
5111 Args.push_back(Entry);
5112 std::pair<SDValue,SDValue> CallResult =
5113 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
5114 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
5116 return CallResult.second;
5119 // Otherwise have the target-independent code call memset.
5123 uint64_t SizeVal = ConstantSize->getZExtValue();
5124 SDValue InFlag(0, 0);
5127 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5128 unsigned BytesLeft = 0;
5129 bool TwoRepStos = false;
5132 uint64_t Val = ValC->getZExtValue() & 255;
5134 // If the value is a constant, then we can potentially use larger sets.
5135 switch (Align & 3) {
5136 case 2: // WORD aligned
5139 Val = (Val << 8) | Val;
5141 case 0: // DWORD aligned
5144 Val = (Val << 8) | Val;
5145 Val = (Val << 16) | Val;
5146 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5149 Val = (Val << 32) | Val;
5152 default: // Byte aligned
5155 Count = DAG.getIntPtrConstant(SizeVal);
5159 if (AVT.bitsGT(MVT::i8)) {
5160 unsigned UBytes = AVT.getSizeInBits() / 8;
5161 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5162 BytesLeft = SizeVal % UBytes;
5165 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5167 InFlag = Chain.getValue(1);
5170 Count = DAG.getIntPtrConstant(SizeVal);
5171 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5172 InFlag = Chain.getValue(1);
5175 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5177 InFlag = Chain.getValue(1);
5178 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5180 InFlag = Chain.getValue(1);
5182 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5183 SmallVector<SDValue, 8> Ops;
5184 Ops.push_back(Chain);
5185 Ops.push_back(DAG.getValueType(AVT));
5186 Ops.push_back(InFlag);
5187 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5190 InFlag = Chain.getValue(1);
5192 MVT CVT = Count.getValueType();
5193 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5194 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5195 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5197 InFlag = Chain.getValue(1);
5198 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5200 Ops.push_back(Chain);
5201 Ops.push_back(DAG.getValueType(MVT::i8));
5202 Ops.push_back(InFlag);
5203 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5204 } else if (BytesLeft) {
5205 // Handle the last 1 - 7 bytes.
5206 unsigned Offset = SizeVal - BytesLeft;
5207 MVT AddrVT = Dst.getValueType();
5208 MVT SizeVT = Size.getValueType();
5210 Chain = DAG.getMemset(Chain,
5211 DAG.getNode(ISD::ADD, AddrVT, Dst,
5212 DAG.getConstant(Offset, AddrVT)),
5214 DAG.getConstant(BytesLeft, SizeVT),
5215 Align, DstSV, DstSVOff + Offset);
5218 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5223 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5224 SDValue Chain, SDValue Dst, SDValue Src,
5225 SDValue Size, unsigned Align,
5227 const Value *DstSV, uint64_t DstSVOff,
5228 const Value *SrcSV, uint64_t SrcSVOff) {
5229 // This requires the copy size to be a constant, preferrably
5230 // within a subtarget-specific limit.
5231 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5234 uint64_t SizeVal = ConstantSize->getZExtValue();
5235 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5238 /// If not DWORD aligned, call the library.
5239 if ((Align & 3) != 0)
5244 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5247 unsigned UBytes = AVT.getSizeInBits() / 8;
5248 unsigned CountVal = SizeVal / UBytes;
5249 SDValue Count = DAG.getIntPtrConstant(CountVal);
5250 unsigned BytesLeft = SizeVal % UBytes;
5252 SDValue InFlag(0, 0);
5253 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5255 InFlag = Chain.getValue(1);
5256 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5258 InFlag = Chain.getValue(1);
5259 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5261 InFlag = Chain.getValue(1);
5263 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5264 SmallVector<SDValue, 8> Ops;
5265 Ops.push_back(Chain);
5266 Ops.push_back(DAG.getValueType(AVT));
5267 Ops.push_back(InFlag);
5268 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5270 SmallVector<SDValue, 4> Results;
5271 Results.push_back(RepMovs);
5273 // Handle the last 1 - 7 bytes.
5274 unsigned Offset = SizeVal - BytesLeft;
5275 MVT DstVT = Dst.getValueType();
5276 MVT SrcVT = Src.getValueType();
5277 MVT SizeVT = Size.getValueType();
5278 Results.push_back(DAG.getMemcpy(Chain,
5279 DAG.getNode(ISD::ADD, DstVT, Dst,
5280 DAG.getConstant(Offset, DstVT)),
5281 DAG.getNode(ISD::ADD, SrcVT, Src,
5282 DAG.getConstant(Offset, SrcVT)),
5283 DAG.getConstant(BytesLeft, SizeVT),
5284 Align, AlwaysInline,
5285 DstSV, DstSVOff + Offset,
5286 SrcSV, SrcSVOff + Offset));
5289 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5292 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5293 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5294 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5295 SDValue TheChain = N->getOperand(0);
5296 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5297 if (Subtarget->is64Bit()) {
5298 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5299 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5300 MVT::i64, rax.getValue(2));
5301 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5302 DAG.getConstant(32, MVT::i8));
5304 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5307 return DAG.getMergeValues(Ops, 2).getNode();
5310 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5311 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5312 MVT::i32, eax.getValue(2));
5313 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5314 SDValue Ops[] = { eax, edx };
5315 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5317 // Use a MERGE_VALUES to return the value and chain.
5318 Ops[1] = edx.getValue(1);
5319 return DAG.getMergeValues(Ops, 2).getNode();
5322 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5323 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5325 if (!Subtarget->is64Bit()) {
5326 // vastart just stores the address of the VarArgsFrameIndex slot into the
5327 // memory location argument.
5328 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5329 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5333 // gp_offset (0 - 6 * 8)
5334 // fp_offset (48 - 48 + 8 * 16)
5335 // overflow_arg_area (point to parameters coming in memory).
5337 SmallVector<SDValue, 8> MemOps;
5338 SDValue FIN = Op.getOperand(1);
5340 SDValue Store = DAG.getStore(Op.getOperand(0),
5341 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5343 MemOps.push_back(Store);
5346 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5347 Store = DAG.getStore(Op.getOperand(0),
5348 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5350 MemOps.push_back(Store);
5352 // Store ptr to overflow_arg_area
5353 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5354 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5355 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5356 MemOps.push_back(Store);
5358 // Store ptr to reg_save_area.
5359 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5360 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5361 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5362 MemOps.push_back(Store);
5363 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5366 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5367 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5368 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5369 SDValue Chain = Op.getOperand(0);
5370 SDValue SrcPtr = Op.getOperand(1);
5371 SDValue SrcSV = Op.getOperand(2);
5373 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5378 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5379 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5380 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5381 SDValue Chain = Op.getOperand(0);
5382 SDValue DstPtr = Op.getOperand(1);
5383 SDValue SrcPtr = Op.getOperand(2);
5384 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5385 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5387 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5388 DAG.getIntPtrConstant(24), 8, false,
5389 DstSV, 0, SrcSV, 0);
5393 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5394 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5396 default: return SDValue(); // Don't custom lower most intrinsics.
5397 // Comparison intrinsics.
5398 case Intrinsic::x86_sse_comieq_ss:
5399 case Intrinsic::x86_sse_comilt_ss:
5400 case Intrinsic::x86_sse_comile_ss:
5401 case Intrinsic::x86_sse_comigt_ss:
5402 case Intrinsic::x86_sse_comige_ss:
5403 case Intrinsic::x86_sse_comineq_ss:
5404 case Intrinsic::x86_sse_ucomieq_ss:
5405 case Intrinsic::x86_sse_ucomilt_ss:
5406 case Intrinsic::x86_sse_ucomile_ss:
5407 case Intrinsic::x86_sse_ucomigt_ss:
5408 case Intrinsic::x86_sse_ucomige_ss:
5409 case Intrinsic::x86_sse_ucomineq_ss:
5410 case Intrinsic::x86_sse2_comieq_sd:
5411 case Intrinsic::x86_sse2_comilt_sd:
5412 case Intrinsic::x86_sse2_comile_sd:
5413 case Intrinsic::x86_sse2_comigt_sd:
5414 case Intrinsic::x86_sse2_comige_sd:
5415 case Intrinsic::x86_sse2_comineq_sd:
5416 case Intrinsic::x86_sse2_ucomieq_sd:
5417 case Intrinsic::x86_sse2_ucomilt_sd:
5418 case Intrinsic::x86_sse2_ucomile_sd:
5419 case Intrinsic::x86_sse2_ucomigt_sd:
5420 case Intrinsic::x86_sse2_ucomige_sd:
5421 case Intrinsic::x86_sse2_ucomineq_sd: {
5423 ISD::CondCode CC = ISD::SETCC_INVALID;
5426 case Intrinsic::x86_sse_comieq_ss:
5427 case Intrinsic::x86_sse2_comieq_sd:
5431 case Intrinsic::x86_sse_comilt_ss:
5432 case Intrinsic::x86_sse2_comilt_sd:
5436 case Intrinsic::x86_sse_comile_ss:
5437 case Intrinsic::x86_sse2_comile_sd:
5441 case Intrinsic::x86_sse_comigt_ss:
5442 case Intrinsic::x86_sse2_comigt_sd:
5446 case Intrinsic::x86_sse_comige_ss:
5447 case Intrinsic::x86_sse2_comige_sd:
5451 case Intrinsic::x86_sse_comineq_ss:
5452 case Intrinsic::x86_sse2_comineq_sd:
5456 case Intrinsic::x86_sse_ucomieq_ss:
5457 case Intrinsic::x86_sse2_ucomieq_sd:
5458 Opc = X86ISD::UCOMI;
5461 case Intrinsic::x86_sse_ucomilt_ss:
5462 case Intrinsic::x86_sse2_ucomilt_sd:
5463 Opc = X86ISD::UCOMI;
5466 case Intrinsic::x86_sse_ucomile_ss:
5467 case Intrinsic::x86_sse2_ucomile_sd:
5468 Opc = X86ISD::UCOMI;
5471 case Intrinsic::x86_sse_ucomigt_ss:
5472 case Intrinsic::x86_sse2_ucomigt_sd:
5473 Opc = X86ISD::UCOMI;
5476 case Intrinsic::x86_sse_ucomige_ss:
5477 case Intrinsic::x86_sse2_ucomige_sd:
5478 Opc = X86ISD::UCOMI;
5481 case Intrinsic::x86_sse_ucomineq_ss:
5482 case Intrinsic::x86_sse2_ucomineq_sd:
5483 Opc = X86ISD::UCOMI;
5489 SDValue LHS = Op.getOperand(1);
5490 SDValue RHS = Op.getOperand(2);
5491 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5493 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5494 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5495 DAG.getConstant(X86CC, MVT::i8), Cond);
5496 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5499 // Fix vector shift instructions where the last operand is a non-immediate
5501 case Intrinsic::x86_sse2_pslli_w:
5502 case Intrinsic::x86_sse2_pslli_d:
5503 case Intrinsic::x86_sse2_pslli_q:
5504 case Intrinsic::x86_sse2_psrli_w:
5505 case Intrinsic::x86_sse2_psrli_d:
5506 case Intrinsic::x86_sse2_psrli_q:
5507 case Intrinsic::x86_sse2_psrai_w:
5508 case Intrinsic::x86_sse2_psrai_d:
5509 case Intrinsic::x86_mmx_pslli_w:
5510 case Intrinsic::x86_mmx_pslli_d:
5511 case Intrinsic::x86_mmx_pslli_q:
5512 case Intrinsic::x86_mmx_psrli_w:
5513 case Intrinsic::x86_mmx_psrli_d:
5514 case Intrinsic::x86_mmx_psrli_q:
5515 case Intrinsic::x86_mmx_psrai_w:
5516 case Intrinsic::x86_mmx_psrai_d: {
5517 SDValue ShAmt = Op.getOperand(2);
5518 if (isa<ConstantSDNode>(ShAmt))
5521 unsigned NewIntNo = 0;
5522 MVT ShAmtVT = MVT::v4i32;
5524 case Intrinsic::x86_sse2_pslli_w:
5525 NewIntNo = Intrinsic::x86_sse2_psll_w;
5527 case Intrinsic::x86_sse2_pslli_d:
5528 NewIntNo = Intrinsic::x86_sse2_psll_d;
5530 case Intrinsic::x86_sse2_pslli_q:
5531 NewIntNo = Intrinsic::x86_sse2_psll_q;
5533 case Intrinsic::x86_sse2_psrli_w:
5534 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5536 case Intrinsic::x86_sse2_psrli_d:
5537 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5539 case Intrinsic::x86_sse2_psrli_q:
5540 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5542 case Intrinsic::x86_sse2_psrai_w:
5543 NewIntNo = Intrinsic::x86_sse2_psra_w;
5545 case Intrinsic::x86_sse2_psrai_d:
5546 NewIntNo = Intrinsic::x86_sse2_psra_d;
5549 ShAmtVT = MVT::v2i32;
5551 case Intrinsic::x86_mmx_pslli_w:
5552 NewIntNo = Intrinsic::x86_mmx_psll_w;
5554 case Intrinsic::x86_mmx_pslli_d:
5555 NewIntNo = Intrinsic::x86_mmx_psll_d;
5557 case Intrinsic::x86_mmx_pslli_q:
5558 NewIntNo = Intrinsic::x86_mmx_psll_q;
5560 case Intrinsic::x86_mmx_psrli_w:
5561 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5563 case Intrinsic::x86_mmx_psrli_d:
5564 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5566 case Intrinsic::x86_mmx_psrli_q:
5567 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5569 case Intrinsic::x86_mmx_psrai_w:
5570 NewIntNo = Intrinsic::x86_mmx_psra_w;
5572 case Intrinsic::x86_mmx_psrai_d:
5573 NewIntNo = Intrinsic::x86_mmx_psra_d;
5575 default: abort(); // Can't reach here.
5580 MVT VT = Op.getValueType();
5581 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5582 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5583 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5584 DAG.getConstant(NewIntNo, MVT::i32),
5585 Op.getOperand(1), ShAmt);
5590 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5591 // Depths > 0 not supported yet!
5592 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5595 // Just load the return address
5596 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5597 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5600 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5601 // Depths > 0 not supported yet!
5602 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5605 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5606 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5607 DAG.getIntPtrConstant(TD->getPointerSize()));
5610 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5611 SelectionDAG &DAG) {
5612 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5615 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5617 MachineFunction &MF = DAG.getMachineFunction();
5618 SDValue Chain = Op.getOperand(0);
5619 SDValue Offset = Op.getOperand(1);
5620 SDValue Handler = Op.getOperand(2);
5622 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5624 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5626 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5627 DAG.getIntPtrConstant(-TD->getPointerSize()));
5628 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5629 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5630 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5631 MF.getRegInfo().addLiveOut(StoreAddrReg);
5633 return DAG.getNode(X86ISD::EH_RETURN,
5635 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5638 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5639 SelectionDAG &DAG) {
5640 SDValue Root = Op.getOperand(0);
5641 SDValue Trmp = Op.getOperand(1); // trampoline
5642 SDValue FPtr = Op.getOperand(2); // nested function
5643 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5645 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5647 const X86InstrInfo *TII =
5648 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5650 if (Subtarget->is64Bit()) {
5651 SDValue OutChains[6];
5653 // Large code-model.
5655 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5656 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5658 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5659 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5661 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5663 // Load the pointer to the nested function into R11.
5664 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5665 SDValue Addr = Trmp;
5666 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5669 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5670 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5672 // Load the 'nest' parameter value into R10.
5673 // R10 is specified in X86CallingConv.td
5674 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5675 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5676 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5679 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5680 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5682 // Jump to the nested function.
5683 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5684 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5685 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5688 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5689 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5690 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5694 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5695 return DAG.getMergeValues(Ops, 2);
5697 const Function *Func =
5698 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5699 unsigned CC = Func->getCallingConv();
5704 assert(0 && "Unsupported calling convention");
5705 case CallingConv::C:
5706 case CallingConv::X86_StdCall: {
5707 // Pass 'nest' parameter in ECX.
5708 // Must be kept in sync with X86CallingConv.td
5711 // Check that ECX wasn't needed by an 'inreg' parameter.
5712 const FunctionType *FTy = Func->getFunctionType();
5713 const PAListPtr &Attrs = Func->getParamAttrs();
5715 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5716 unsigned InRegCount = 0;
5719 for (FunctionType::param_iterator I = FTy->param_begin(),
5720 E = FTy->param_end(); I != E; ++I, ++Idx)
5721 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5722 // FIXME: should only count parameters that are lowered to integers.
5723 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
5725 if (InRegCount > 2) {
5726 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5732 case CallingConv::X86_FastCall:
5733 case CallingConv::Fast:
5734 // Pass 'nest' parameter in EAX.
5735 // Must be kept in sync with X86CallingConv.td
5740 SDValue OutChains[4];
5743 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5744 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5746 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5747 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5748 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5751 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5752 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5754 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5755 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5756 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5757 TrmpAddr, 5, false, 1);
5759 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5760 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5763 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5764 return DAG.getMergeValues(Ops, 2);
5768 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5770 The rounding mode is in bits 11:10 of FPSR, and has the following
5777 FLT_ROUNDS, on the other hand, expects the following:
5784 To perform the conversion, we do:
5785 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5788 MachineFunction &MF = DAG.getMachineFunction();
5789 const TargetMachine &TM = MF.getTarget();
5790 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5791 unsigned StackAlignment = TFI.getStackAlignment();
5792 MVT VT = Op.getValueType();
5794 // Save FP Control Word to stack slot
5795 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5796 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5798 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5799 DAG.getEntryNode(), StackSlot);
5801 // Load FP Control Word from stack slot
5802 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5804 // Transform as necessary
5806 DAG.getNode(ISD::SRL, MVT::i16,
5807 DAG.getNode(ISD::AND, MVT::i16,
5808 CWD, DAG.getConstant(0x800, MVT::i16)),
5809 DAG.getConstant(11, MVT::i8));
5811 DAG.getNode(ISD::SRL, MVT::i16,
5812 DAG.getNode(ISD::AND, MVT::i16,
5813 CWD, DAG.getConstant(0x400, MVT::i16)),
5814 DAG.getConstant(9, MVT::i8));
5817 DAG.getNode(ISD::AND, MVT::i16,
5818 DAG.getNode(ISD::ADD, MVT::i16,
5819 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5820 DAG.getConstant(1, MVT::i16)),
5821 DAG.getConstant(3, MVT::i16));
5824 return DAG.getNode((VT.getSizeInBits() < 16 ?
5825 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5828 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5829 MVT VT = Op.getValueType();
5831 unsigned NumBits = VT.getSizeInBits();
5833 Op = Op.getOperand(0);
5834 if (VT == MVT::i8) {
5835 // Zero extend to i32 since there is not an i8 bsr.
5837 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5840 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5841 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5842 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5844 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5845 SmallVector<SDValue, 4> Ops;
5847 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5848 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5849 Ops.push_back(Op.getValue(1));
5850 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5852 // Finally xor with NumBits-1.
5853 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5856 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5860 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5861 MVT VT = Op.getValueType();
5863 unsigned NumBits = VT.getSizeInBits();
5865 Op = Op.getOperand(0);
5866 if (VT == MVT::i8) {
5868 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5871 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5872 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5873 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5875 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5876 SmallVector<SDValue, 4> Ops;
5878 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5879 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5880 Ops.push_back(Op.getValue(1));
5881 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5884 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5888 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5889 MVT T = Op.getValueType();
5892 switch(T.getSimpleVT()) {
5894 assert(false && "Invalid value type!");
5895 case MVT::i8: Reg = X86::AL; size = 1; break;
5896 case MVT::i16: Reg = X86::AX; size = 2; break;
5897 case MVT::i32: Reg = X86::EAX; size = 4; break;
5899 if (Subtarget->is64Bit()) {
5900 Reg = X86::RAX; size = 8;
5901 } else //Should go away when LowerType stuff lands
5902 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5905 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5906 Op.getOperand(2), SDValue());
5907 SDValue Ops[] = { cpIn.getValue(0),
5910 DAG.getTargetConstant(size, MVT::i8),
5912 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5913 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5915 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5919 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5920 SelectionDAG &DAG) {
5921 MVT T = Op->getValueType(0);
5922 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5923 SDValue cpInL, cpInH;
5924 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5925 DAG.getConstant(0, MVT::i32));
5926 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5927 DAG.getConstant(1, MVT::i32));
5928 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5930 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5931 cpInH, cpInL.getValue(1));
5932 SDValue swapInL, swapInH;
5933 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5934 DAG.getConstant(0, MVT::i32));
5935 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5936 DAG.getConstant(1, MVT::i32));
5937 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5938 swapInL, cpInH.getValue(1));
5939 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5940 swapInH, swapInL.getValue(1));
5941 SDValue Ops[] = { swapInH.getValue(0),
5943 swapInH.getValue(1)};
5944 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5945 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5946 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5947 Result.getValue(1));
5948 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5949 cpOutL.getValue(2));
5950 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5951 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5952 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
5953 return DAG.getMergeValues(Vals, 2).getNode();
5956 SDNode* X86TargetLowering::ExpandATOMIC_LOAD_SUB(SDNode* Op,
5957 SelectionDAG &DAG) {
5958 MVT T = Op->getValueType(0);
5959 SDValue negOp = DAG.getNode(ISD::SUB, T,
5960 DAG.getConstant(0, T), Op->getOperand(2));
5961 return DAG.getAtomic((T==MVT::i8 ? ISD::ATOMIC_LOAD_ADD_8:
5962 T==MVT::i16 ? ISD::ATOMIC_LOAD_ADD_16:
5963 T==MVT::i32 ? ISD::ATOMIC_LOAD_ADD_32:
5964 T==MVT::i64 ? ISD::ATOMIC_LOAD_ADD_64: 0),
5965 Op->getOperand(0), Op->getOperand(1), negOp,
5966 cast<AtomicSDNode>(Op)->getSrcValue(),
5967 cast<AtomicSDNode>(Op)->getAlignment()).getNode();
5970 /// LowerOperation - Provide custom lowering hooks for some operations.
5972 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5973 switch (Op.getOpcode()) {
5974 default: assert(0 && "Should not custom lower this!");
5975 case ISD::ATOMIC_CMP_SWAP_8: return LowerCMP_SWAP(Op,DAG);
5976 case ISD::ATOMIC_CMP_SWAP_16: return LowerCMP_SWAP(Op,DAG);
5977 case ISD::ATOMIC_CMP_SWAP_32: return LowerCMP_SWAP(Op,DAG);
5978 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
5979 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5980 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5981 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5982 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5983 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5984 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5985 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5986 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5987 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5988 case ISD::SHL_PARTS:
5989 case ISD::SRA_PARTS:
5990 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5991 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5992 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5993 case ISD::FABS: return LowerFABS(Op, DAG);
5994 case ISD::FNEG: return LowerFNEG(Op, DAG);
5995 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5996 case ISD::SETCC: return LowerSETCC(Op, DAG);
5997 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
5998 case ISD::SELECT: return LowerSELECT(Op, DAG);
5999 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6000 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6001 case ISD::CALL: return LowerCALL(Op, DAG);
6002 case ISD::RET: return LowerRET(Op, DAG);
6003 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6004 case ISD::VASTART: return LowerVASTART(Op, DAG);
6005 case ISD::VAARG: return LowerVAARG(Op, DAG);
6006 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6007 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6008 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6009 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6010 case ISD::FRAME_TO_ARGS_OFFSET:
6011 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6012 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6013 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6014 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6015 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6016 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6017 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6019 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
6020 case ISD::READCYCLECOUNTER:
6021 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
6025 /// ReplaceNodeResults - Replace a node with an illegal result type
6026 /// with a new node built out of custom code.
6027 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
6028 switch (N->getOpcode()) {
6029 default: assert(0 && "Should not custom lower this!");
6030 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
6031 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
6032 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
6033 case ISD::ATOMIC_LOAD_SUB_8: return ExpandATOMIC_LOAD_SUB(N,DAG);
6034 case ISD::ATOMIC_LOAD_SUB_16: return ExpandATOMIC_LOAD_SUB(N,DAG);
6035 case ISD::ATOMIC_LOAD_SUB_32: return ExpandATOMIC_LOAD_SUB(N,DAG);
6036 case ISD::ATOMIC_LOAD_SUB_64: return ExpandATOMIC_LOAD_SUB(N,DAG);
6040 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6042 default: return NULL;
6043 case X86ISD::BSF: return "X86ISD::BSF";
6044 case X86ISD::BSR: return "X86ISD::BSR";
6045 case X86ISD::SHLD: return "X86ISD::SHLD";
6046 case X86ISD::SHRD: return "X86ISD::SHRD";
6047 case X86ISD::FAND: return "X86ISD::FAND";
6048 case X86ISD::FOR: return "X86ISD::FOR";
6049 case X86ISD::FXOR: return "X86ISD::FXOR";
6050 case X86ISD::FSRL: return "X86ISD::FSRL";
6051 case X86ISD::FILD: return "X86ISD::FILD";
6052 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6053 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6054 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6055 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6056 case X86ISD::FLD: return "X86ISD::FLD";
6057 case X86ISD::FST: return "X86ISD::FST";
6058 case X86ISD::CALL: return "X86ISD::CALL";
6059 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6060 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6061 case X86ISD::CMP: return "X86ISD::CMP";
6062 case X86ISD::COMI: return "X86ISD::COMI";
6063 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6064 case X86ISD::SETCC: return "X86ISD::SETCC";
6065 case X86ISD::CMOV: return "X86ISD::CMOV";
6066 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6067 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6068 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6069 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6070 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6071 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6072 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6073 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6074 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6075 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6076 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6077 case X86ISD::FMAX: return "X86ISD::FMAX";
6078 case X86ISD::FMIN: return "X86ISD::FMIN";
6079 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6080 case X86ISD::FRCP: return "X86ISD::FRCP";
6081 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6082 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6083 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6084 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6085 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6086 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6087 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6088 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6089 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6090 case X86ISD::VSHL: return "X86ISD::VSHL";
6091 case X86ISD::VSRL: return "X86ISD::VSRL";
6092 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6093 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6094 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6095 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6096 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6097 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6098 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6099 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6100 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6101 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6105 // isLegalAddressingMode - Return true if the addressing mode represented
6106 // by AM is legal for this target, for a load/store of the specified type.
6107 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6108 const Type *Ty) const {
6109 // X86 supports extremely general addressing modes.
6111 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6112 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6116 // We can only fold this if we don't need an extra load.
6117 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6120 // X86-64 only supports addr of globals in small code model.
6121 if (Subtarget->is64Bit()) {
6122 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6124 // If lower 4G is not available, then we must use rip-relative addressing.
6125 if (AM.BaseOffs || AM.Scale > 1)
6136 // These scales always work.
6141 // These scales are formed with basereg+scalereg. Only accept if there is
6146 default: // Other stuff never works.
6154 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6155 if (!Ty1->isInteger() || !Ty2->isInteger())
6157 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6158 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6159 if (NumBits1 <= NumBits2)
6161 return Subtarget->is64Bit() || NumBits1 < 64;
6164 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6165 if (!VT1.isInteger() || !VT2.isInteger())
6167 unsigned NumBits1 = VT1.getSizeInBits();
6168 unsigned NumBits2 = VT2.getSizeInBits();
6169 if (NumBits1 <= NumBits2)
6171 return Subtarget->is64Bit() || NumBits1 < 64;
6174 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6175 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6176 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6177 /// are assumed to be legal.
6179 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6180 // Only do shuffles on 128-bit vector types for now.
6181 if (VT.getSizeInBits() == 64) return false;
6182 return (Mask.getNode()->getNumOperands() <= 4 ||
6183 isIdentityMask(Mask.getNode()) ||
6184 isIdentityMask(Mask.getNode(), true) ||
6185 isSplatMask(Mask.getNode()) ||
6186 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6187 X86::isUNPCKLMask(Mask.getNode()) ||
6188 X86::isUNPCKHMask(Mask.getNode()) ||
6189 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6190 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6194 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6195 MVT EVT, SelectionDAG &DAG) const {
6196 unsigned NumElts = BVOps.size();
6197 // Only do shuffles on 128-bit vector types for now.
6198 if (EVT.getSizeInBits() * NumElts == 64) return false;
6199 if (NumElts == 2) return true;
6201 return (isMOVLMask(&BVOps[0], 4) ||
6202 isCommutedMOVL(&BVOps[0], 4, true) ||
6203 isSHUFPMask(&BVOps[0], 4) ||
6204 isCommutedSHUFP(&BVOps[0], 4));
6209 //===----------------------------------------------------------------------===//
6210 // X86 Scheduler Hooks
6211 //===----------------------------------------------------------------------===//
6213 // private utility function
6215 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6216 MachineBasicBlock *MBB,
6224 TargetRegisterClass *RC,
6226 // For the atomic bitwise operator, we generate
6229 // ld t1 = [bitinstr.addr]
6230 // op t2 = t1, [bitinstr.val]
6232 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6234 // fallthrough -->nextMBB
6235 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6236 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6237 MachineFunction::iterator MBBIter = MBB;
6240 /// First build the CFG
6241 MachineFunction *F = MBB->getParent();
6242 MachineBasicBlock *thisMBB = MBB;
6243 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6244 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6245 F->insert(MBBIter, newMBB);
6246 F->insert(MBBIter, nextMBB);
6248 // Move all successors to thisMBB to nextMBB
6249 nextMBB->transferSuccessors(thisMBB);
6251 // Update thisMBB to fall through to newMBB
6252 thisMBB->addSuccessor(newMBB);
6254 // newMBB jumps to itself and fall through to nextMBB
6255 newMBB->addSuccessor(nextMBB);
6256 newMBB->addSuccessor(newMBB);
6258 // Insert instructions into newMBB based on incoming instruction
6259 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6260 MachineOperand& destOper = bInstr->getOperand(0);
6261 MachineOperand* argOpers[6];
6262 int numArgs = bInstr->getNumOperands() - 1;
6263 for (int i=0; i < numArgs; ++i)
6264 argOpers[i] = &bInstr->getOperand(i+1);
6266 // x86 address has 4 operands: base, index, scale, and displacement
6267 int lastAddrIndx = 3; // [0,3]
6270 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6271 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6272 for (int i=0; i <= lastAddrIndx; ++i)
6273 (*MIB).addOperand(*argOpers[i]);
6275 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6277 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6282 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6283 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6284 && "invalid operand");
6285 if (argOpers[valArgIndx]->isReg())
6286 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6288 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6290 (*MIB).addOperand(*argOpers[valArgIndx]);
6292 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6295 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6296 for (int i=0; i <= lastAddrIndx; ++i)
6297 (*MIB).addOperand(*argOpers[i]);
6299 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6300 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6302 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6306 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6308 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6312 // private utility function
6314 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6315 MachineBasicBlock *MBB,
6317 // For the atomic min/max operator, we generate
6320 // ld t1 = [min/max.addr]
6321 // mov t2 = [min/max.val]
6323 // cmov[cond] t2 = t1
6325 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6327 // fallthrough -->nextMBB
6329 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6330 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6331 MachineFunction::iterator MBBIter = MBB;
6334 /// First build the CFG
6335 MachineFunction *F = MBB->getParent();
6336 MachineBasicBlock *thisMBB = MBB;
6337 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6338 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6339 F->insert(MBBIter, newMBB);
6340 F->insert(MBBIter, nextMBB);
6342 // Move all successors to thisMBB to nextMBB
6343 nextMBB->transferSuccessors(thisMBB);
6345 // Update thisMBB to fall through to newMBB
6346 thisMBB->addSuccessor(newMBB);
6348 // newMBB jumps to newMBB and fall through to nextMBB
6349 newMBB->addSuccessor(nextMBB);
6350 newMBB->addSuccessor(newMBB);
6352 // Insert instructions into newMBB based on incoming instruction
6353 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6354 MachineOperand& destOper = mInstr->getOperand(0);
6355 MachineOperand* argOpers[6];
6356 int numArgs = mInstr->getNumOperands() - 1;
6357 for (int i=0; i < numArgs; ++i)
6358 argOpers[i] = &mInstr->getOperand(i+1);
6360 // x86 address has 4 operands: base, index, scale, and displacement
6361 int lastAddrIndx = 3; // [0,3]
6364 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6365 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6366 for (int i=0; i <= lastAddrIndx; ++i)
6367 (*MIB).addOperand(*argOpers[i]);
6369 // We only support register and immediate values
6370 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6371 && "invalid operand");
6373 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6374 if (argOpers[valArgIndx]->isReg())
6375 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6377 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6378 (*MIB).addOperand(*argOpers[valArgIndx]);
6380 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6383 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6388 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6389 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6393 // Cmp and exchange if none has modified the memory location
6394 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6395 for (int i=0; i <= lastAddrIndx; ++i)
6396 (*MIB).addOperand(*argOpers[i]);
6398 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6399 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6401 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6402 MIB.addReg(X86::EAX);
6405 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6407 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6413 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6414 MachineBasicBlock *BB) {
6415 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6416 switch (MI->getOpcode()) {
6417 default: assert(false && "Unexpected instr type to insert");
6418 case X86::CMOV_FR32:
6419 case X86::CMOV_FR64:
6420 case X86::CMOV_V4F32:
6421 case X86::CMOV_V2F64:
6422 case X86::CMOV_V2I64: {
6423 // To "insert" a SELECT_CC instruction, we actually have to insert the
6424 // diamond control-flow pattern. The incoming instruction knows the
6425 // destination vreg to set, the condition code register to branch on, the
6426 // true/false values to select between, and a branch opcode to use.
6427 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6428 MachineFunction::iterator It = BB;
6434 // cmpTY ccX, r1, r2
6436 // fallthrough --> copy0MBB
6437 MachineBasicBlock *thisMBB = BB;
6438 MachineFunction *F = BB->getParent();
6439 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6440 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6442 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6443 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6444 F->insert(It, copy0MBB);
6445 F->insert(It, sinkMBB);
6446 // Update machine-CFG edges by transferring all successors of the current
6447 // block to the new block which will contain the Phi node for the select.
6448 sinkMBB->transferSuccessors(BB);
6450 // Add the true and fallthrough blocks as its successors.
6451 BB->addSuccessor(copy0MBB);
6452 BB->addSuccessor(sinkMBB);
6455 // %FalseValue = ...
6456 // # fallthrough to sinkMBB
6459 // Update machine-CFG edges
6460 BB->addSuccessor(sinkMBB);
6463 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6466 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6467 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6468 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6470 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6474 case X86::FP32_TO_INT16_IN_MEM:
6475 case X86::FP32_TO_INT32_IN_MEM:
6476 case X86::FP32_TO_INT64_IN_MEM:
6477 case X86::FP64_TO_INT16_IN_MEM:
6478 case X86::FP64_TO_INT32_IN_MEM:
6479 case X86::FP64_TO_INT64_IN_MEM:
6480 case X86::FP80_TO_INT16_IN_MEM:
6481 case X86::FP80_TO_INT32_IN_MEM:
6482 case X86::FP80_TO_INT64_IN_MEM: {
6483 // Change the floating point control register to use "round towards zero"
6484 // mode when truncating to an integer value.
6485 MachineFunction *F = BB->getParent();
6486 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6487 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6489 // Load the old value of the high byte of the control word...
6491 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6492 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6494 // Set the high part to be round to zero...
6495 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6498 // Reload the modified control word now...
6499 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6501 // Restore the memory image of control word to original value
6502 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6505 // Get the X86 opcode to use.
6507 switch (MI->getOpcode()) {
6508 default: assert(0 && "illegal opcode!");
6509 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6510 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6511 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6512 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6513 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6514 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6515 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6516 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6517 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6521 MachineOperand &Op = MI->getOperand(0);
6522 if (Op.isRegister()) {
6523 AM.BaseType = X86AddressMode::RegBase;
6524 AM.Base.Reg = Op.getReg();
6526 AM.BaseType = X86AddressMode::FrameIndexBase;
6527 AM.Base.FrameIndex = Op.getIndex();
6529 Op = MI->getOperand(1);
6530 if (Op.isImmediate())
6531 AM.Scale = Op.getImm();
6532 Op = MI->getOperand(2);
6533 if (Op.isImmediate())
6534 AM.IndexReg = Op.getImm();
6535 Op = MI->getOperand(3);
6536 if (Op.isGlobalAddress()) {
6537 AM.GV = Op.getGlobal();
6539 AM.Disp = Op.getImm();
6541 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6542 .addReg(MI->getOperand(4).getReg());
6544 // Reload the original control word now.
6545 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6547 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6550 case X86::ATOMAND32:
6551 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6552 X86::AND32ri, X86::MOV32rm,
6553 X86::LCMPXCHG32, X86::MOV32rr,
6554 X86::NOT32r, X86::EAX,
6555 X86::GR32RegisterClass);
6557 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6558 X86::OR32ri, X86::MOV32rm,
6559 X86::LCMPXCHG32, X86::MOV32rr,
6560 X86::NOT32r, X86::EAX,
6561 X86::GR32RegisterClass);
6562 case X86::ATOMXOR32:
6563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6564 X86::XOR32ri, X86::MOV32rm,
6565 X86::LCMPXCHG32, X86::MOV32rr,
6566 X86::NOT32r, X86::EAX,
6567 X86::GR32RegisterClass);
6568 case X86::ATOMNAND32:
6569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6570 X86::AND32ri, X86::MOV32rm,
6571 X86::LCMPXCHG32, X86::MOV32rr,
6572 X86::NOT32r, X86::EAX,
6573 X86::GR32RegisterClass, true);
6574 case X86::ATOMMIN32:
6575 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6576 case X86::ATOMMAX32:
6577 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6578 case X86::ATOMUMIN32:
6579 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6580 case X86::ATOMUMAX32:
6581 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6583 case X86::ATOMAND16:
6584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6585 X86::AND16ri, X86::MOV16rm,
6586 X86::LCMPXCHG16, X86::MOV16rr,
6587 X86::NOT16r, X86::AX,
6588 X86::GR16RegisterClass);
6590 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6591 X86::OR16ri, X86::MOV16rm,
6592 X86::LCMPXCHG16, X86::MOV16rr,
6593 X86::NOT16r, X86::AX,
6594 X86::GR16RegisterClass);
6595 case X86::ATOMXOR16:
6596 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6597 X86::XOR16ri, X86::MOV16rm,
6598 X86::LCMPXCHG16, X86::MOV16rr,
6599 X86::NOT16r, X86::AX,
6600 X86::GR16RegisterClass);
6601 case X86::ATOMNAND16:
6602 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6603 X86::AND16ri, X86::MOV16rm,
6604 X86::LCMPXCHG16, X86::MOV16rr,
6605 X86::NOT16r, X86::AX,
6606 X86::GR16RegisterClass, true);
6607 case X86::ATOMMIN16:
6608 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6609 case X86::ATOMMAX16:
6610 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6611 case X86::ATOMUMIN16:
6612 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6613 case X86::ATOMUMAX16:
6614 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6617 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6618 X86::AND8ri, X86::MOV8rm,
6619 X86::LCMPXCHG8, X86::MOV8rr,
6620 X86::NOT8r, X86::AL,
6621 X86::GR8RegisterClass);
6623 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6624 X86::OR8ri, X86::MOV8rm,
6625 X86::LCMPXCHG8, X86::MOV8rr,
6626 X86::NOT8r, X86::AL,
6627 X86::GR8RegisterClass);
6629 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6630 X86::XOR8ri, X86::MOV8rm,
6631 X86::LCMPXCHG8, X86::MOV8rr,
6632 X86::NOT8r, X86::AL,
6633 X86::GR8RegisterClass);
6634 case X86::ATOMNAND8:
6635 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6636 X86::AND8ri, X86::MOV8rm,
6637 X86::LCMPXCHG8, X86::MOV8rr,
6638 X86::NOT8r, X86::AL,
6639 X86::GR8RegisterClass, true);
6640 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6641 case X86::ATOMAND64:
6642 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6643 X86::AND64ri32, X86::MOV64rm,
6644 X86::LCMPXCHG64, X86::MOV64rr,
6645 X86::NOT64r, X86::RAX,
6646 X86::GR64RegisterClass);
6648 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6649 X86::OR64ri32, X86::MOV64rm,
6650 X86::LCMPXCHG64, X86::MOV64rr,
6651 X86::NOT64r, X86::RAX,
6652 X86::GR64RegisterClass);
6653 case X86::ATOMXOR64:
6654 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6655 X86::XOR64ri32, X86::MOV64rm,
6656 X86::LCMPXCHG64, X86::MOV64rr,
6657 X86::NOT64r, X86::RAX,
6658 X86::GR64RegisterClass);
6659 case X86::ATOMNAND64:
6660 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6661 X86::AND64ri32, X86::MOV64rm,
6662 X86::LCMPXCHG64, X86::MOV64rr,
6663 X86::NOT64r, X86::RAX,
6664 X86::GR64RegisterClass, true);
6665 case X86::ATOMMIN64:
6666 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6667 case X86::ATOMMAX64:
6668 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6669 case X86::ATOMUMIN64:
6670 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6671 case X86::ATOMUMAX64:
6672 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6676 //===----------------------------------------------------------------------===//
6677 // X86 Optimization Hooks
6678 //===----------------------------------------------------------------------===//
6680 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6684 const SelectionDAG &DAG,
6685 unsigned Depth) const {
6686 unsigned Opc = Op.getOpcode();
6687 assert((Opc >= ISD::BUILTIN_OP_END ||
6688 Opc == ISD::INTRINSIC_WO_CHAIN ||
6689 Opc == ISD::INTRINSIC_W_CHAIN ||
6690 Opc == ISD::INTRINSIC_VOID) &&
6691 "Should use MaskedValueIsZero if you don't know whether Op"
6692 " is a target node!");
6694 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
6698 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
6699 Mask.getBitWidth() - 1);
6704 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
6705 /// node is a GlobalAddress + offset.
6706 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
6707 GlobalValue* &GA, int64_t &Offset) const{
6708 if (N->getOpcode() == X86ISD::Wrapper) {
6709 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
6710 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
6714 return TargetLowering::isGAPlusOffset(N, GA, Offset);
6717 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
6718 const TargetLowering &TLI) {
6721 if (TLI.isGAPlusOffset(Base, GV, Offset))
6722 return (GV->getAlignment() >= N && (Offset % N) == 0);
6723 // DAG combine handles the stack object case.
6727 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
6728 unsigned NumElems, MVT EVT,
6730 SelectionDAG &DAG, MachineFrameInfo *MFI,
6731 const TargetLowering &TLI) {
6733 for (unsigned i = 0; i < NumElems; ++i) {
6734 SDValue Idx = PermMask.getOperand(i);
6735 if (Idx.getOpcode() == ISD::UNDEF) {
6741 SDValue Elt = DAG.getShuffleScalarElt(N, i);
6742 if (!Elt.getNode() ||
6743 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6746 Base = Elt.getNode();
6747 if (Base->getOpcode() == ISD::UNDEF)
6751 if (Elt.getOpcode() == ISD::UNDEF)
6754 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
6755 EVT.getSizeInBits()/8, i, MFI))
6761 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6762 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6763 /// if the load addresses are consecutive, non-overlapping, and in the right
6765 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6766 const TargetLowering &TLI) {
6767 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6768 MVT VT = N->getValueType(0);
6769 MVT EVT = VT.getVectorElementType();
6770 SDValue PermMask = N->getOperand(2);
6771 unsigned NumElems = PermMask.getNumOperands();
6772 SDNode *Base = NULL;
6773 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
6777 LoadSDNode *LD = cast<LoadSDNode>(Base);
6778 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
6779 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6780 LD->getSrcValueOffset(), LD->isVolatile());
6781 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6782 LD->getSrcValueOffset(), LD->isVolatile(),
6783 LD->getAlignment());
6786 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
6787 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
6788 const X86Subtarget *Subtarget,
6789 const TargetLowering &TLI) {
6790 unsigned NumOps = N->getNumOperands();
6792 // Ignore single operand BUILD_VECTOR.
6796 MVT VT = N->getValueType(0);
6797 MVT EVT = VT.getVectorElementType();
6798 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
6799 // We are looking for load i64 and zero extend. We want to transform
6800 // it before legalizer has a chance to expand it. Also look for i64
6801 // BUILD_PAIR bit casted to f64.
6803 // This must be an insertion into a zero vector.
6804 SDValue HighElt = N->getOperand(1);
6805 if (!isZeroNode(HighElt))
6808 // Value must be a load.
6809 SDNode *Base = N->getOperand(0).getNode();
6810 if (!isa<LoadSDNode>(Base)) {
6811 if (Base->getOpcode() != ISD::BIT_CONVERT)
6813 Base = Base->getOperand(0).getNode();
6814 if (!isa<LoadSDNode>(Base))
6818 // Transform it into VZEXT_LOAD addr.
6819 LoadSDNode *LD = cast<LoadSDNode>(Base);
6821 // Load must not be an extload.
6822 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
6825 return DAG.getNode(X86ISD::VZEXT_LOAD, VT, LD->getChain(), LD->getBasePtr());
6828 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6829 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6830 const X86Subtarget *Subtarget) {
6831 SDValue Cond = N->getOperand(0);
6833 // If we have SSE[12] support, try to form min/max nodes.
6834 if (Subtarget->hasSSE2() &&
6835 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6836 if (Cond.getOpcode() == ISD::SETCC) {
6837 // Get the LHS/RHS of the select.
6838 SDValue LHS = N->getOperand(1);
6839 SDValue RHS = N->getOperand(2);
6840 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6842 unsigned Opcode = 0;
6843 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6846 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6849 if (!UnsafeFPMath) break;
6851 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6853 Opcode = X86ISD::FMIN;
6856 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6859 if (!UnsafeFPMath) break;
6861 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6863 Opcode = X86ISD::FMAX;
6866 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6869 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6872 if (!UnsafeFPMath) break;
6874 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6876 Opcode = X86ISD::FMIN;
6879 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6882 if (!UnsafeFPMath) break;
6884 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6886 Opcode = X86ISD::FMAX;
6892 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6900 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6901 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
6902 const X86Subtarget *Subtarget) {
6903 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6904 // the FP state in cases where an emms may be missing.
6905 // A preferable solution to the general problem is to figure out the right
6906 // places to insert EMMS. This qualifies as a quick hack.
6907 StoreSDNode *St = cast<StoreSDNode>(N);
6908 if (St->getValue().getValueType().isVector() &&
6909 St->getValue().getValueType().getSizeInBits() == 64 &&
6910 isa<LoadSDNode>(St->getValue()) &&
6911 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6912 St->getChain().hasOneUse() && !St->isVolatile()) {
6913 SDNode* LdVal = St->getValue().getNode();
6915 int TokenFactorIndex = -1;
6916 SmallVector<SDValue, 8> Ops;
6917 SDNode* ChainVal = St->getChain().getNode();
6918 // Must be a store of a load. We currently handle two cases: the load
6919 // is a direct child, and it's under an intervening TokenFactor. It is
6920 // possible to dig deeper under nested TokenFactors.
6921 if (ChainVal == LdVal)
6922 Ld = cast<LoadSDNode>(St->getChain());
6923 else if (St->getValue().hasOneUse() &&
6924 ChainVal->getOpcode() == ISD::TokenFactor) {
6925 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6926 if (ChainVal->getOperand(i).getNode() == LdVal) {
6927 TokenFactorIndex = i;
6928 Ld = cast<LoadSDNode>(St->getValue());
6930 Ops.push_back(ChainVal->getOperand(i));
6934 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6935 if (Subtarget->is64Bit()) {
6936 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6937 Ld->getBasePtr(), Ld->getSrcValue(),
6938 Ld->getSrcValueOffset(), Ld->isVolatile(),
6939 Ld->getAlignment());
6940 SDValue NewChain = NewLd.getValue(1);
6941 if (TokenFactorIndex != -1) {
6942 Ops.push_back(NewChain);
6943 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6946 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6947 St->getSrcValue(), St->getSrcValueOffset(),
6948 St->isVolatile(), St->getAlignment());
6951 // Otherwise, lower to two 32-bit copies.
6952 SDValue LoAddr = Ld->getBasePtr();
6953 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6954 DAG.getConstant(4, MVT::i32));
6956 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6957 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6958 Ld->isVolatile(), Ld->getAlignment());
6959 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6960 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6962 MinAlign(Ld->getAlignment(), 4));
6964 SDValue NewChain = LoLd.getValue(1);
6965 if (TokenFactorIndex != -1) {
6966 Ops.push_back(LoLd);
6967 Ops.push_back(HiLd);
6968 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6972 LoAddr = St->getBasePtr();
6973 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6974 DAG.getConstant(4, MVT::i32));
6976 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6977 St->getSrcValue(), St->getSrcValueOffset(),
6978 St->isVolatile(), St->getAlignment());
6979 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6981 St->getSrcValueOffset() + 4,
6983 MinAlign(St->getAlignment(), 4));
6984 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6990 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6991 /// X86ISD::FXOR nodes.
6992 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6993 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6994 // F[X]OR(0.0, x) -> x
6995 // F[X]OR(x, 0.0) -> x
6996 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6997 if (C->getValueAPF().isPosZero())
6998 return N->getOperand(1);
6999 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7000 if (C->getValueAPF().isPosZero())
7001 return N->getOperand(0);
7005 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
7006 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
7007 // FAND(0.0, x) -> 0.0
7008 // FAND(x, 0.0) -> 0.0
7009 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7010 if (C->getValueAPF().isPosZero())
7011 return N->getOperand(0);
7012 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7013 if (C->getValueAPF().isPosZero())
7014 return N->getOperand(1);
7019 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7020 DAGCombinerInfo &DCI) const {
7021 SelectionDAG &DAG = DCI.DAG;
7022 switch (N->getOpcode()) {
7024 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7025 case ISD::BUILD_VECTOR:
7026 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7027 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7028 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7030 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7031 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7037 //===----------------------------------------------------------------------===//
7038 // X86 Inline Assembly Support
7039 //===----------------------------------------------------------------------===//
7041 /// getConstraintType - Given a constraint letter, return the type of
7042 /// constraint it is for this target.
7043 X86TargetLowering::ConstraintType
7044 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7045 if (Constraint.size() == 1) {
7046 switch (Constraint[0]) {
7057 return C_RegisterClass;
7062 return TargetLowering::getConstraintType(Constraint);
7065 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7066 /// with another that has more specific requirements based on the type of the
7067 /// corresponding operand.
7068 const char *X86TargetLowering::
7069 LowerXConstraint(MVT ConstraintVT) const {
7070 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7071 // 'f' like normal targets.
7072 if (ConstraintVT.isFloatingPoint()) {
7073 if (Subtarget->hasSSE2())
7075 if (Subtarget->hasSSE1())
7079 return TargetLowering::LowerXConstraint(ConstraintVT);
7082 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7083 /// vector. If it is invalid, don't add anything to Ops.
7084 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7086 std::vector<SDValue>&Ops,
7087 SelectionDAG &DAG) const {
7088 SDValue Result(0, 0);
7090 switch (Constraint) {
7093 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7094 if (C->getZExtValue() <= 31) {
7095 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7101 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7102 if (C->getZExtValue() <= 255) {
7103 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7109 // Literal immediates are always ok.
7110 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7111 Result = DAG.getTargetConstant(CST->getZExtValue(), Op.getValueType());
7115 // If we are in non-pic codegen mode, we allow the address of a global (with
7116 // an optional displacement) to be used with 'i'.
7117 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7120 // Match either (GA) or (GA+C)
7122 Offset = GA->getOffset();
7123 } else if (Op.getOpcode() == ISD::ADD) {
7124 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7125 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7127 Offset = GA->getOffset()+C->getZExtValue();
7129 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7130 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7132 Offset = GA->getOffset()+C->getZExtValue();
7139 // If addressing this global requires a load (e.g. in PIC mode), we can't
7141 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
7145 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7151 // Otherwise, not valid for this mode.
7156 if (Result.getNode()) {
7157 Ops.push_back(Result);
7160 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7163 std::vector<unsigned> X86TargetLowering::
7164 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7166 if (Constraint.size() == 1) {
7167 // FIXME: not handling fp-stack yet!
7168 switch (Constraint[0]) { // GCC X86 Constraint Letters
7169 default: break; // Unknown constraint letter
7170 case 'A': // EAX/EDX
7171 if (VT == MVT::i32 || VT == MVT::i64)
7172 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7174 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7177 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7178 else if (VT == MVT::i16)
7179 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7180 else if (VT == MVT::i8)
7181 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7182 else if (VT == MVT::i64)
7183 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7188 return std::vector<unsigned>();
7191 std::pair<unsigned, const TargetRegisterClass*>
7192 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7194 // First, see if this is a constraint that directly corresponds to an LLVM
7196 if (Constraint.size() == 1) {
7197 // GCC Constraint Letters
7198 switch (Constraint[0]) {
7200 case 'r': // GENERAL_REGS
7201 case 'R': // LEGACY_REGS
7202 case 'l': // INDEX_REGS
7203 if (VT == MVT::i64 && Subtarget->is64Bit())
7204 return std::make_pair(0U, X86::GR64RegisterClass);
7206 return std::make_pair(0U, X86::GR32RegisterClass);
7207 else if (VT == MVT::i16)
7208 return std::make_pair(0U, X86::GR16RegisterClass);
7209 else if (VT == MVT::i8)
7210 return std::make_pair(0U, X86::GR8RegisterClass);
7212 case 'f': // FP Stack registers.
7213 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7214 // value to the correct fpstack register class.
7215 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7216 return std::make_pair(0U, X86::RFP32RegisterClass);
7217 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7218 return std::make_pair(0U, X86::RFP64RegisterClass);
7219 return std::make_pair(0U, X86::RFP80RegisterClass);
7220 case 'y': // MMX_REGS if MMX allowed.
7221 if (!Subtarget->hasMMX()) break;
7222 return std::make_pair(0U, X86::VR64RegisterClass);
7224 case 'Y': // SSE_REGS if SSE2 allowed
7225 if (!Subtarget->hasSSE2()) break;
7227 case 'x': // SSE_REGS if SSE1 allowed
7228 if (!Subtarget->hasSSE1()) break;
7230 switch (VT.getSimpleVT()) {
7232 // Scalar SSE types.
7235 return std::make_pair(0U, X86::FR32RegisterClass);
7238 return std::make_pair(0U, X86::FR64RegisterClass);
7246 return std::make_pair(0U, X86::VR128RegisterClass);
7252 // Use the default implementation in TargetLowering to convert the register
7253 // constraint into a member of a register class.
7254 std::pair<unsigned, const TargetRegisterClass*> Res;
7255 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7257 // Not found as a standard register?
7258 if (Res.second == 0) {
7259 // GCC calls "st(0)" just plain "st".
7260 if (StringsEqualNoCase("{st}", Constraint)) {
7261 Res.first = X86::ST0;
7262 Res.second = X86::RFP80RegisterClass;
7268 // Otherwise, check to see if this is a register class of the wrong value
7269 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7270 // turn into {ax},{dx}.
7271 if (Res.second->hasType(VT))
7272 return Res; // Correct type already, nothing to do.
7274 // All of the single-register GCC register classes map their values onto
7275 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7276 // really want an 8-bit or 32-bit register, map to the appropriate register
7277 // class and return the appropriate register.
7278 if (Res.second == X86::GR16RegisterClass) {
7279 if (VT == MVT::i8) {
7280 unsigned DestReg = 0;
7281 switch (Res.first) {
7283 case X86::AX: DestReg = X86::AL; break;
7284 case X86::DX: DestReg = X86::DL; break;
7285 case X86::CX: DestReg = X86::CL; break;
7286 case X86::BX: DestReg = X86::BL; break;
7289 Res.first = DestReg;
7290 Res.second = Res.second = X86::GR8RegisterClass;
7292 } else if (VT == MVT::i32) {
7293 unsigned DestReg = 0;
7294 switch (Res.first) {
7296 case X86::AX: DestReg = X86::EAX; break;
7297 case X86::DX: DestReg = X86::EDX; break;
7298 case X86::CX: DestReg = X86::ECX; break;
7299 case X86::BX: DestReg = X86::EBX; break;
7300 case X86::SI: DestReg = X86::ESI; break;
7301 case X86::DI: DestReg = X86::EDI; break;
7302 case X86::BP: DestReg = X86::EBP; break;
7303 case X86::SP: DestReg = X86::ESP; break;
7306 Res.first = DestReg;
7307 Res.second = Res.second = X86::GR32RegisterClass;
7309 } else if (VT == MVT::i64) {
7310 unsigned DestReg = 0;
7311 switch (Res.first) {
7313 case X86::AX: DestReg = X86::RAX; break;
7314 case X86::DX: DestReg = X86::RDX; break;
7315 case X86::CX: DestReg = X86::RCX; break;
7316 case X86::BX: DestReg = X86::RBX; break;
7317 case X86::SI: DestReg = X86::RSI; break;
7318 case X86::DI: DestReg = X86::RDI; break;
7319 case X86::BP: DestReg = X86::RBP; break;
7320 case X86::SP: DestReg = X86::RSP; break;
7323 Res.first = DestReg;
7324 Res.second = Res.second = X86::GR64RegisterClass;
7327 } else if (Res.second == X86::FR32RegisterClass ||
7328 Res.second == X86::FR64RegisterClass ||
7329 Res.second == X86::VR128RegisterClass) {
7330 // Handle references to XMM physical registers that got mapped into the
7331 // wrong class. This can happen with constraints like {xmm0} where the
7332 // target independent register mapper will just pick the first match it can
7333 // find, ignoring the required type.
7335 Res.second = X86::FR32RegisterClass;
7336 else if (VT == MVT::f64)
7337 Res.second = X86::FR64RegisterClass;
7338 else if (X86::VR128RegisterClass->hasType(VT))
7339 Res.second = X86::VR128RegisterClass;