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"
41 #include "llvm/Support/CommandLine.h"
45 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
47 // Forward declarations.
48 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
50 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
51 : TargetLowering(TM) {
52 Subtarget = &TM.getSubtarget<X86Subtarget>();
53 X86ScalarSSEf64 = Subtarget->hasSSE2();
54 X86ScalarSSEf32 = Subtarget->hasSSE1();
55 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
59 RegInfo = TM.getRegisterInfo();
62 // Set up the TargetLowering object.
64 // X86 is weird, it always uses i8 for shift amounts and setcc results.
65 setShiftAmountType(MVT::i8);
66 setBooleanContents(ZeroOrOneBooleanContent);
67 setSchedulingPreference(SchedulingForRegPressure);
68 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
69 setStackPointerRegisterToSaveRestore(X86StackPtr);
71 if (Subtarget->isTargetDarwin()) {
72 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(false);
74 setUseUnderscoreLongJmp(false);
75 } else if (Subtarget->isTargetMingw()) {
76 // MS runtime is weird: it exports _setjmp, but longjmp!
77 setUseUnderscoreSetJmp(true);
78 setUseUnderscoreLongJmp(false);
80 setUseUnderscoreSetJmp(true);
81 setUseUnderscoreLongJmp(true);
84 // Set up the register classes.
85 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
86 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
87 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
88 if (Subtarget->is64Bit())
89 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
91 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
93 // We don't accept any truncstore of integer registers.
94 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
95 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
96 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
97 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
98 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
99 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
101 // SETOEQ and SETUNE require checking two conditions.
102 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
103 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
104 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
105 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
106 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
107 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
109 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
111 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
112 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
113 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
115 if (Subtarget->is64Bit()) {
116 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
117 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
119 if (X86ScalarSSEf64) {
120 // We have an impenetrably clever algorithm for ui64->double only.
121 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
122 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
123 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
125 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
128 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
130 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
131 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
132 // SSE has no i16 to fp conversion, only i32
133 if (X86ScalarSSEf32) {
134 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
135 // f32 and f64 cases are Legal, f80 case is not
136 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
138 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
139 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
142 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
143 // are Legal, f80 is custom lowered.
144 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
145 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
147 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
149 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
150 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
152 if (X86ScalarSSEf32) {
153 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
154 // f32 and f64 cases are Legal, f80 case is not
155 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
157 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
158 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
161 // Handle FP_TO_UINT by promoting the destination to a larger signed
163 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
164 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
165 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
167 if (Subtarget->is64Bit()) {
168 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
169 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
171 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
172 // Expand FP_TO_UINT into a select.
173 // FIXME: We would like to use a Custom expander here eventually to do
174 // the optimal thing for SSE vs. the default expansion in the legalizer.
175 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
177 // With SSE3 we can use fisttpll to convert to a signed i64.
178 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
181 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
182 if (!X86ScalarSSEf64) {
183 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
184 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
187 // Scalar integer divide and remainder are lowered to use operations that
188 // produce two results, to match the available instructions. This exposes
189 // the two-result form to trivial CSE, which is able to combine x/y and x%y
190 // into a single instruction.
192 // Scalar integer multiply-high is also lowered to use two-result
193 // operations, to match the available instructions. However, plain multiply
194 // (low) operations are left as Legal, as there are single-result
195 // instructions for this in x86. Using the two-result multiply instructions
196 // when both high and low results are needed must be arranged by dagcombine.
197 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
198 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
199 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
200 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
201 setOperationAction(ISD::SREM , MVT::i8 , Expand);
202 setOperationAction(ISD::UREM , MVT::i8 , Expand);
203 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
204 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
205 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
206 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
207 setOperationAction(ISD::SREM , MVT::i16 , Expand);
208 setOperationAction(ISD::UREM , MVT::i16 , Expand);
209 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
210 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
211 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
212 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
213 setOperationAction(ISD::SREM , MVT::i32 , Expand);
214 setOperationAction(ISD::UREM , MVT::i32 , Expand);
215 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
216 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
217 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
218 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
219 setOperationAction(ISD::SREM , MVT::i64 , Expand);
220 setOperationAction(ISD::UREM , MVT::i64 , Expand);
222 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
223 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
224 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
225 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
226 if (Subtarget->is64Bit())
227 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
228 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
229 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
230 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
231 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
232 setOperationAction(ISD::FREM , MVT::f32 , Expand);
233 setOperationAction(ISD::FREM , MVT::f64 , Expand);
234 setOperationAction(ISD::FREM , MVT::f80 , Expand);
235 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
237 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
238 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
239 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
240 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
241 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
242 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
243 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
244 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
245 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
246 if (Subtarget->is64Bit()) {
247 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
248 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
249 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
252 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
253 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
255 // These should be promoted to a larger select which is supported.
256 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
257 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
258 // X86 wants to expand cmov itself.
259 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
260 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
261 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
262 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
263 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
264 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
265 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
266 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
267 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
268 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
269 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
270 if (Subtarget->is64Bit()) {
271 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
272 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
274 // X86 ret instruction may pop stack.
275 setOperationAction(ISD::RET , MVT::Other, Custom);
276 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
279 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
280 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
281 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
282 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
283 if (Subtarget->is64Bit())
284 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
285 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
286 if (Subtarget->is64Bit()) {
287 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
288 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
289 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
290 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
292 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
293 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
294 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
295 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
298 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
299 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
302 if (Subtarget->hasSSE1())
303 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
305 if (!Subtarget->hasSSE2())
306 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
308 // Expand certain atomics
309 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
310 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
311 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
312 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
314 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
315 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
316 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
317 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
319 if (!Subtarget->is64Bit()) {
320 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
321 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
326 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
329 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
330 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
331 // FIXME - use subtarget debug flags
332 if (!Subtarget->isTargetDarwin() &&
333 !Subtarget->isTargetELF() &&
334 !Subtarget->isTargetCygMing()) {
335 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
336 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
339 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
340 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
341 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
342 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
343 if (Subtarget->is64Bit()) {
344 setExceptionPointerRegister(X86::RAX);
345 setExceptionSelectorRegister(X86::RDX);
347 setExceptionPointerRegister(X86::EAX);
348 setExceptionSelectorRegister(X86::EDX);
350 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
351 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
353 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
355 setOperationAction(ISD::TRAP, MVT::Other, Legal);
357 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
358 setOperationAction(ISD::VASTART , MVT::Other, Custom);
359 setOperationAction(ISD::VAEND , MVT::Other, Expand);
360 if (Subtarget->is64Bit()) {
361 setOperationAction(ISD::VAARG , MVT::Other, Custom);
362 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
364 setOperationAction(ISD::VAARG , MVT::Other, Expand);
365 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
368 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
369 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
370 if (Subtarget->is64Bit())
371 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
372 if (Subtarget->isTargetCygMing())
373 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
375 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
377 if (X86ScalarSSEf64) {
378 // f32 and f64 use SSE.
379 // Set up the FP register classes.
380 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
381 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
383 // Use ANDPD to simulate FABS.
384 setOperationAction(ISD::FABS , MVT::f64, Custom);
385 setOperationAction(ISD::FABS , MVT::f32, Custom);
387 // Use XORP to simulate FNEG.
388 setOperationAction(ISD::FNEG , MVT::f64, Custom);
389 setOperationAction(ISD::FNEG , MVT::f32, Custom);
391 // Use ANDPD and ORPD to simulate FCOPYSIGN.
392 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
393 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
395 // We don't support sin/cos/fmod
396 setOperationAction(ISD::FSIN , MVT::f64, Expand);
397 setOperationAction(ISD::FCOS , MVT::f64, Expand);
398 setOperationAction(ISD::FSIN , MVT::f32, Expand);
399 setOperationAction(ISD::FCOS , MVT::f32, Expand);
401 // Expand FP immediates into loads from the stack, except for the special
403 addLegalFPImmediate(APFloat(+0.0)); // xorpd
404 addLegalFPImmediate(APFloat(+0.0f)); // xorps
406 // Floating truncations from f80 and extensions to f80 go through memory.
407 // If optimizing, we lie about this though and handle it in
408 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
410 setConvertAction(MVT::f32, MVT::f80, Expand);
411 setConvertAction(MVT::f64, MVT::f80, Expand);
412 setConvertAction(MVT::f80, MVT::f32, Expand);
413 setConvertAction(MVT::f80, MVT::f64, Expand);
415 } else if (X86ScalarSSEf32) {
416 // Use SSE for f32, x87 for f64.
417 // Set up the FP register classes.
418 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
419 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
421 // Use ANDPS to simulate FABS.
422 setOperationAction(ISD::FABS , MVT::f32, Custom);
424 // Use XORP to simulate FNEG.
425 setOperationAction(ISD::FNEG , MVT::f32, Custom);
427 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
429 // Use ANDPS and ORPS to simulate FCOPYSIGN.
430 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
431 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
433 // We don't support sin/cos/fmod
434 setOperationAction(ISD::FSIN , MVT::f32, Expand);
435 setOperationAction(ISD::FCOS , MVT::f32, Expand);
437 // Special cases we handle for FP constants.
438 addLegalFPImmediate(APFloat(+0.0f)); // xorps
439 addLegalFPImmediate(APFloat(+0.0)); // FLD0
440 addLegalFPImmediate(APFloat(+1.0)); // FLD1
441 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
442 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
444 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
445 // this though and handle it in InstructionSelectPreprocess so that
446 // dagcombine2 can hack on these.
448 setConvertAction(MVT::f32, MVT::f64, Expand);
449 setConvertAction(MVT::f32, MVT::f80, Expand);
450 setConvertAction(MVT::f80, MVT::f32, Expand);
451 setConvertAction(MVT::f64, MVT::f32, Expand);
452 // And x87->x87 truncations also.
453 setConvertAction(MVT::f80, MVT::f64, Expand);
457 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
458 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
461 // f32 and f64 in x87.
462 // Set up the FP register classes.
463 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
464 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
466 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
467 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
468 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
471 // Floating truncations go through memory. If optimizing, we lie about
472 // this though and handle it in InstructionSelectPreprocess so that
473 // dagcombine2 can hack on these.
475 setConvertAction(MVT::f80, MVT::f32, Expand);
476 setConvertAction(MVT::f64, MVT::f32, Expand);
477 setConvertAction(MVT::f80, MVT::f64, Expand);
481 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
482 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
484 addLegalFPImmediate(APFloat(+0.0)); // FLD0
485 addLegalFPImmediate(APFloat(+1.0)); // FLD1
486 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
487 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
488 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
489 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
490 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
491 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
494 // Long double always uses X87.
495 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
496 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
497 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
500 APFloat TmpFlt(+0.0);
501 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
503 addLegalFPImmediate(TmpFlt); // FLD0
505 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
506 APFloat TmpFlt2(+1.0);
507 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
509 addLegalFPImmediate(TmpFlt2); // FLD1
510 TmpFlt2.changeSign();
511 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
515 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
516 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
519 // Always use a library call for pow.
520 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
521 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
522 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
524 setOperationAction(ISD::FLOG, MVT::f80, Expand);
525 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
526 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
527 setOperationAction(ISD::FEXP, MVT::f80, Expand);
528 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
530 // First set operation action for all vector types to either promote
531 // (for widening) or expand (for scalarization). Then we will selectively
532 // turn on ones that can be effectively codegen'd.
533 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
534 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
535 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
550 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
580 if (!DisableMMX && Subtarget->hasMMX()) {
581 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
582 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
583 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
584 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
585 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
587 // FIXME: add MMX packed arithmetics
589 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
590 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
591 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
592 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
594 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
595 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
596 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
597 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
599 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
600 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
602 setOperationAction(ISD::AND, MVT::v8i8, Promote);
603 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
604 setOperationAction(ISD::AND, MVT::v4i16, Promote);
605 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
606 setOperationAction(ISD::AND, MVT::v2i32, Promote);
607 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
608 setOperationAction(ISD::AND, MVT::v1i64, Legal);
610 setOperationAction(ISD::OR, MVT::v8i8, Promote);
611 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
612 setOperationAction(ISD::OR, MVT::v4i16, Promote);
613 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
614 setOperationAction(ISD::OR, MVT::v2i32, Promote);
615 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
616 setOperationAction(ISD::OR, MVT::v1i64, Legal);
618 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
619 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
620 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
621 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
622 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
623 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
624 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
626 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
627 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
628 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
629 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
630 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
631 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
632 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
633 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
634 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
636 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
637 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
638 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
639 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
640 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
645 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
647 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
648 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
649 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
650 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
652 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
654 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
655 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
656 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
657 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
658 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
659 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
662 if (Subtarget->hasSSE1()) {
663 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
665 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
666 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
667 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
668 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
669 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
670 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
671 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
674 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
675 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
676 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
679 if (Subtarget->hasSSE2()) {
680 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
681 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
682 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
683 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
684 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
686 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
687 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
688 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
689 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
690 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
691 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
692 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
693 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
694 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
695 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
696 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
697 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
698 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
699 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
700 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
701 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
703 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
704 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
705 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
706 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
708 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
709 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
710 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
711 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
712 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
714 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
715 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
716 MVT VT = (MVT::SimpleValueType)i;
717 // Do not attempt to custom lower non-power-of-2 vectors
718 if (!isPowerOf2_32(VT.getVectorNumElements()))
720 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
721 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
722 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
724 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
725 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
726 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
727 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
728 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
729 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
730 if (Subtarget->is64Bit()) {
731 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
732 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
735 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
736 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
737 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
738 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
739 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
740 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
741 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
742 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
743 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
744 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
745 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
746 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
749 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
751 // Custom lower v2i64 and v2f64 selects.
752 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
753 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
754 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
755 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
759 if (Subtarget->hasSSE41()) {
760 // FIXME: Do we need to handle scalar-to-vector here?
761 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
763 // i8 and i16 vectors are custom , because the source register and source
764 // source memory operand types are not the same width. f32 vectors are
765 // custom since the immediate controlling the insert encodes additional
767 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
770 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
772 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
774 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
775 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
777 if (Subtarget->is64Bit()) {
778 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
779 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
783 if (Subtarget->hasSSE42()) {
784 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
787 // We want to custom lower some of our intrinsics.
788 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
790 // Add/Sub/Mul with overflow operations are custom lowered.
791 setOperationAction(ISD::SADDO, MVT::i32, Custom);
792 setOperationAction(ISD::SADDO, MVT::i64, Custom);
793 setOperationAction(ISD::UADDO, MVT::i32, Custom);
794 setOperationAction(ISD::UADDO, MVT::i64, Custom);
795 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
796 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
797 setOperationAction(ISD::USUBO, MVT::i32, Custom);
798 setOperationAction(ISD::USUBO, MVT::i64, Custom);
799 setOperationAction(ISD::SMULO, MVT::i32, Custom);
800 setOperationAction(ISD::SMULO, MVT::i64, Custom);
801 setOperationAction(ISD::UMULO, MVT::i32, Custom);
802 setOperationAction(ISD::UMULO, MVT::i64, Custom);
804 // We have target-specific dag combine patterns for the following nodes:
805 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
806 setTargetDAGCombine(ISD::BUILD_VECTOR);
807 setTargetDAGCombine(ISD::SELECT);
808 setTargetDAGCombine(ISD::STORE);
810 computeRegisterProperties();
812 // FIXME: These should be based on subtarget info. Plus, the values should
813 // be smaller when we are in optimizing for size mode.
814 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
815 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
816 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
817 allowUnalignedMemoryAccesses = true; // x86 supports it!
818 setPrefLoopAlignment(16);
822 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
827 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
828 /// the desired ByVal argument alignment.
829 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
832 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
833 if (VTy->getBitWidth() == 128)
835 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
836 unsigned EltAlign = 0;
837 getMaxByValAlign(ATy->getElementType(), EltAlign);
838 if (EltAlign > MaxAlign)
840 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
841 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
842 unsigned EltAlign = 0;
843 getMaxByValAlign(STy->getElementType(i), EltAlign);
844 if (EltAlign > MaxAlign)
853 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
854 /// function arguments in the caller parameter area. For X86, aggregates
855 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
856 /// are at 4-byte boundaries.
857 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
858 if (Subtarget->is64Bit()) {
859 // Max of 8 and alignment of type.
860 unsigned TyAlign = TD->getABITypeAlignment(Ty);
867 if (Subtarget->hasSSE1())
868 getMaxByValAlign(Ty, Align);
872 /// getOptimalMemOpType - Returns the target specific optimal type for load
873 /// and store operations as a result of memset, memcpy, and memmove
874 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
877 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
878 bool isSrcConst, bool isSrcStr) const {
879 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
880 // linux. This is because the stack realignment code can't handle certain
881 // cases like PR2962. This should be removed when PR2962 is fixed.
882 if (Subtarget->getStackAlignment() >= 16) {
883 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
885 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
888 if (Subtarget->is64Bit() && Size >= 8)
894 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
896 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
897 SelectionDAG &DAG) const {
898 if (usesGlobalOffsetTable())
899 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
900 if (!Subtarget->isPICStyleRIPRel())
901 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
905 //===----------------------------------------------------------------------===//
906 // Return Value Calling Convention Implementation
907 //===----------------------------------------------------------------------===//
909 #include "X86GenCallingConv.inc"
911 /// LowerRET - Lower an ISD::RET node.
912 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
913 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
915 SmallVector<CCValAssign, 16> RVLocs;
916 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
917 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
918 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
919 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
921 // If this is the first return lowered for this function, add the regs to the
922 // liveout set for the function.
923 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
924 for (unsigned i = 0; i != RVLocs.size(); ++i)
925 if (RVLocs[i].isRegLoc())
926 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
928 SDValue Chain = Op.getOperand(0);
930 // Handle tail call return.
931 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
932 if (Chain.getOpcode() == X86ISD::TAILCALL) {
933 SDValue TailCall = Chain;
934 SDValue TargetAddress = TailCall.getOperand(1);
935 SDValue StackAdjustment = TailCall.getOperand(2);
936 assert(((TargetAddress.getOpcode() == ISD::Register &&
937 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
938 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
939 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
940 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
941 "Expecting an global address, external symbol, or register");
942 assert(StackAdjustment.getOpcode() == ISD::Constant &&
943 "Expecting a const value");
945 SmallVector<SDValue,8> Operands;
946 Operands.push_back(Chain.getOperand(0));
947 Operands.push_back(TargetAddress);
948 Operands.push_back(StackAdjustment);
949 // Copy registers used by the call. Last operand is a flag so it is not
951 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
952 Operands.push_back(Chain.getOperand(i));
954 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
961 SmallVector<SDValue, 6> RetOps;
962 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
963 // Operand #1 = Bytes To Pop
964 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
966 // Copy the result values into the output registers.
967 for (unsigned i = 0; i != RVLocs.size(); ++i) {
968 CCValAssign &VA = RVLocs[i];
969 assert(VA.isRegLoc() && "Can only return in registers!");
970 SDValue ValToCopy = Op.getOperand(i*2+1);
972 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
973 // the RET instruction and handled by the FP Stackifier.
974 if (RVLocs[i].getLocReg() == X86::ST0 ||
975 RVLocs[i].getLocReg() == X86::ST1) {
976 // If this is a copy from an xmm register to ST(0), use an FPExtend to
977 // change the value to the FP stack register class.
978 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
979 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
980 RetOps.push_back(ValToCopy);
981 // Don't emit a copytoreg.
985 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
986 Flag = Chain.getValue(1);
989 // The x86-64 ABI for returning structs by value requires that we copy
990 // the sret argument into %rax for the return. We saved the argument into
991 // a virtual register in the entry block, so now we copy the value out
993 if (Subtarget->is64Bit() &&
994 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
995 MachineFunction &MF = DAG.getMachineFunction();
996 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
997 unsigned Reg = FuncInfo->getSRetReturnReg();
999 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1000 FuncInfo->setSRetReturnReg(Reg);
1002 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
1004 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
1005 Flag = Chain.getValue(1);
1008 RetOps[0] = Chain; // Update chain.
1010 // Add the flag if we have it.
1012 RetOps.push_back(Flag);
1014 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
1018 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1019 /// appropriate copies out of appropriate physical registers. This assumes that
1020 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1021 /// being lowered. The returns a SDNode with the same number of values as the
1023 SDNode *X86TargetLowering::
1024 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1025 unsigned CallingConv, SelectionDAG &DAG) {
1027 // Assign locations to each value returned by this call.
1028 SmallVector<CCValAssign, 16> RVLocs;
1029 bool isVarArg = TheCall->isVarArg();
1030 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1031 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1033 SmallVector<SDValue, 8> ResultVals;
1035 // Copy all of the result registers out of their specified physreg.
1036 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1037 MVT CopyVT = RVLocs[i].getValVT();
1039 // If this is a call to a function that returns an fp value on the floating
1040 // point stack, but where we prefer to use the value in xmm registers, copy
1041 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1042 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1043 RVLocs[i].getLocReg() == X86::ST1) &&
1044 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
1048 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1049 CopyVT, InFlag).getValue(1);
1050 SDValue Val = Chain.getValue(0);
1051 InFlag = Chain.getValue(2);
1053 if (CopyVT != RVLocs[i].getValVT()) {
1054 // Round the F80 the right size, which also moves to the appropriate xmm
1056 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1057 // This truncation won't change the value.
1058 DAG.getIntPtrConstant(1));
1061 ResultVals.push_back(Val);
1064 // Merge everything together with a MERGE_VALUES node.
1065 ResultVals.push_back(Chain);
1066 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(), &ResultVals[0],
1067 ResultVals.size()).getNode();
1071 //===----------------------------------------------------------------------===//
1072 // C & StdCall & Fast Calling Convention implementation
1073 //===----------------------------------------------------------------------===//
1074 // StdCall calling convention seems to be standard for many Windows' API
1075 // routines and around. It differs from C calling convention just a little:
1076 // callee should clean up the stack, not caller. Symbols should be also
1077 // decorated in some fancy way :) It doesn't support any vector arguments.
1078 // For info on fast calling convention see Fast Calling Convention (tail call)
1079 // implementation LowerX86_32FastCCCallTo.
1081 /// AddLiveIn - This helper function adds the specified physical register to the
1082 /// MachineFunction as a live in value. It also creates a corresponding virtual
1083 /// register for it.
1084 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1085 const TargetRegisterClass *RC) {
1086 assert(RC->contains(PReg) && "Not the correct regclass!");
1087 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1088 MF.getRegInfo().addLiveIn(PReg, VReg);
1092 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1094 static bool CallIsStructReturn(CallSDNode *TheCall) {
1095 unsigned NumOps = TheCall->getNumArgs();
1099 return TheCall->getArgFlags(0).isSRet();
1102 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1103 /// return semantics.
1104 static bool ArgsAreStructReturn(SDValue Op) {
1105 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1109 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1112 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1113 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1115 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1119 switch (CallingConv) {
1122 case CallingConv::X86_StdCall:
1123 return !Subtarget->is64Bit();
1124 case CallingConv::X86_FastCall:
1125 return !Subtarget->is64Bit();
1126 case CallingConv::Fast:
1127 return PerformTailCallOpt;
1131 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1132 /// given CallingConvention value.
1133 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1134 if (Subtarget->is64Bit()) {
1135 if (Subtarget->isTargetWin64())
1136 return CC_X86_Win64_C;
1137 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1138 return CC_X86_64_TailCall;
1143 if (CC == CallingConv::X86_FastCall)
1144 return CC_X86_32_FastCall;
1145 else if (CC == CallingConv::Fast)
1146 return CC_X86_32_FastCC;
1151 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1152 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1154 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1155 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1156 if (CC == CallingConv::X86_FastCall)
1158 else if (CC == CallingConv::X86_StdCall)
1164 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1165 /// in a register before calling.
1166 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1167 return !IsTailCall && !Is64Bit &&
1168 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1169 Subtarget->isPICStyleGOT();
1172 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1173 /// address to be loaded in a register.
1175 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1176 return !Is64Bit && IsTailCall &&
1177 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1178 Subtarget->isPICStyleGOT();
1181 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1182 /// by "Src" to address "Dst" with size and alignment information specified by
1183 /// the specific parameter attribute. The copy will be passed as a byval
1184 /// function parameter.
1186 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1187 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1188 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1189 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1190 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1193 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1194 const CCValAssign &VA,
1195 MachineFrameInfo *MFI,
1197 SDValue Root, unsigned i) {
1198 // Create the nodes corresponding to a load from this parameter slot.
1199 ISD::ArgFlagsTy Flags =
1200 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1201 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1202 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1204 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1205 // changed with more analysis.
1206 // In case of tail call optimization mark all arguments mutable. Since they
1207 // could be overwritten by lowering of arguments in case of a tail call.
1208 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1209 VA.getLocMemOffset(), isImmutable);
1210 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1211 if (Flags.isByVal())
1213 return DAG.getLoad(VA.getValVT(), Root, FIN,
1214 PseudoSourceValue::getFixedStack(FI), 0);
1218 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1219 MachineFunction &MF = DAG.getMachineFunction();
1220 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1222 const Function* Fn = MF.getFunction();
1223 if (Fn->hasExternalLinkage() &&
1224 Subtarget->isTargetCygMing() &&
1225 Fn->getName() == "main")
1226 FuncInfo->setForceFramePointer(true);
1228 // Decorate the function name.
1229 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1231 MachineFrameInfo *MFI = MF.getFrameInfo();
1232 SDValue Root = Op.getOperand(0);
1233 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1234 unsigned CC = MF.getFunction()->getCallingConv();
1235 bool Is64Bit = Subtarget->is64Bit();
1236 bool IsWin64 = Subtarget->isTargetWin64();
1238 assert(!(isVarArg && CC == CallingConv::Fast) &&
1239 "Var args not supported with calling convention fastcc");
1241 // Assign locations to all of the incoming arguments.
1242 SmallVector<CCValAssign, 16> ArgLocs;
1243 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1244 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1246 SmallVector<SDValue, 8> ArgValues;
1247 unsigned LastVal = ~0U;
1248 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1249 CCValAssign &VA = ArgLocs[i];
1250 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1252 assert(VA.getValNo() != LastVal &&
1253 "Don't support value assigned to multiple locs yet");
1254 LastVal = VA.getValNo();
1256 if (VA.isRegLoc()) {
1257 MVT RegVT = VA.getLocVT();
1258 TargetRegisterClass *RC = NULL;
1259 if (RegVT == MVT::i32)
1260 RC = X86::GR32RegisterClass;
1261 else if (Is64Bit && RegVT == MVT::i64)
1262 RC = X86::GR64RegisterClass;
1263 else if (RegVT == MVT::f32)
1264 RC = X86::FR32RegisterClass;
1265 else if (RegVT == MVT::f64)
1266 RC = X86::FR64RegisterClass;
1267 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1268 RC = X86::VR128RegisterClass;
1269 else if (RegVT.isVector()) {
1270 assert(RegVT.getSizeInBits() == 64);
1272 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1274 // Darwin calling convention passes MMX values in either GPRs or
1275 // XMMs in x86-64. Other targets pass them in memory.
1276 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1277 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1280 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1285 assert(0 && "Unknown argument type!");
1288 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1289 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1291 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1292 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1294 if (VA.getLocInfo() == CCValAssign::SExt)
1295 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1296 DAG.getValueType(VA.getValVT()));
1297 else if (VA.getLocInfo() == CCValAssign::ZExt)
1298 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1299 DAG.getValueType(VA.getValVT()));
1301 if (VA.getLocInfo() != CCValAssign::Full)
1302 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1304 // Handle MMX values passed in GPRs.
1305 if (Is64Bit && RegVT != VA.getLocVT()) {
1306 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1307 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1308 else if (RC == X86::VR128RegisterClass) {
1309 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1310 DAG.getConstant(0, MVT::i64));
1311 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1315 ArgValues.push_back(ArgValue);
1317 assert(VA.isMemLoc());
1318 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1322 // The x86-64 ABI for returning structs by value requires that we copy
1323 // the sret argument into %rax for the return. Save the argument into
1324 // a virtual register so that we can access it from the return points.
1325 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1326 MachineFunction &MF = DAG.getMachineFunction();
1327 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1328 unsigned Reg = FuncInfo->getSRetReturnReg();
1330 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1331 FuncInfo->setSRetReturnReg(Reg);
1333 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1334 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1337 unsigned StackSize = CCInfo.getNextStackOffset();
1338 // align stack specially for tail calls
1339 if (PerformTailCallOpt && CC == CallingConv::Fast)
1340 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1342 // If the function takes variable number of arguments, make a frame index for
1343 // the start of the first vararg value... for expansion of llvm.va_start.
1345 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1346 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1349 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1351 // FIXME: We should really autogenerate these arrays
1352 static const unsigned GPR64ArgRegsWin64[] = {
1353 X86::RCX, X86::RDX, X86::R8, X86::R9
1355 static const unsigned XMMArgRegsWin64[] = {
1356 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1358 static const unsigned GPR64ArgRegs64Bit[] = {
1359 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1361 static const unsigned XMMArgRegs64Bit[] = {
1362 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1363 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1365 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1368 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1369 GPR64ArgRegs = GPR64ArgRegsWin64;
1370 XMMArgRegs = XMMArgRegsWin64;
1372 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1373 GPR64ArgRegs = GPR64ArgRegs64Bit;
1374 XMMArgRegs = XMMArgRegs64Bit;
1376 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1378 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1381 // For X86-64, if there are vararg parameters that are passed via
1382 // registers, then we must store them to their spots on the stack so they
1383 // may be loaded by deferencing the result of va_next.
1384 VarArgsGPOffset = NumIntRegs * 8;
1385 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1386 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1387 TotalNumXMMRegs * 16, 16);
1389 // Store the integer parameter registers.
1390 SmallVector<SDValue, 8> MemOps;
1391 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1392 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1393 DAG.getIntPtrConstant(VarArgsGPOffset));
1394 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1395 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1396 X86::GR64RegisterClass);
1397 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1399 DAG.getStore(Val.getValue(1), Val, FIN,
1400 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1401 MemOps.push_back(Store);
1402 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1403 DAG.getIntPtrConstant(8));
1406 // Now store the XMM (fp + vector) parameter registers.
1407 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1408 DAG.getIntPtrConstant(VarArgsFPOffset));
1409 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1410 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1411 X86::VR128RegisterClass);
1412 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1414 DAG.getStore(Val.getValue(1), Val, FIN,
1415 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1416 MemOps.push_back(Store);
1417 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1418 DAG.getIntPtrConstant(16));
1420 if (!MemOps.empty())
1421 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1422 &MemOps[0], MemOps.size());
1426 ArgValues.push_back(Root);
1428 // Some CCs need callee pop.
1429 if (IsCalleePop(isVarArg, CC)) {
1430 BytesToPopOnReturn = StackSize; // Callee pops everything.
1431 BytesCallerReserves = 0;
1433 BytesToPopOnReturn = 0; // Callee pops nothing.
1434 // If this is an sret function, the return should pop the hidden pointer.
1435 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1436 BytesToPopOnReturn = 4;
1437 BytesCallerReserves = StackSize;
1441 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1442 if (CC == CallingConv::X86_FastCall)
1443 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1446 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1448 // Return the new list of results.
1449 return DAG.getNode(ISD::MERGE_VALUES, Op.getNode()->getVTList(),
1450 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1454 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1455 const SDValue &StackPtr,
1456 const CCValAssign &VA,
1458 SDValue Arg, ISD::ArgFlagsTy Flags) {
1459 unsigned LocMemOffset = VA.getLocMemOffset();
1460 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1461 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1462 if (Flags.isByVal()) {
1463 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1465 return DAG.getStore(Chain, Arg, PtrOff,
1466 PseudoSourceValue::getStack(), LocMemOffset);
1469 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1470 /// optimization is performed and it is required.
1472 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1473 SDValue &OutRetAddr,
1478 if (!IsTailCall || FPDiff==0) return Chain;
1480 // Adjust the Return address stack slot.
1481 MVT VT = getPointerTy();
1482 OutRetAddr = getReturnAddressFrameIndex(DAG);
1483 // Load the "old" Return address.
1484 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1485 return SDValue(OutRetAddr.getNode(), 1);
1488 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1489 /// optimization is performed and it is required (FPDiff!=0).
1491 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1492 SDValue Chain, SDValue RetAddrFrIdx,
1493 bool Is64Bit, int FPDiff) {
1494 // Store the return address to the appropriate stack slot.
1495 if (!FPDiff) return Chain;
1496 // Calculate the new stack slot for the return address.
1497 int SlotSize = Is64Bit ? 8 : 4;
1498 int NewReturnAddrFI =
1499 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1500 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1501 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1502 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1503 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1507 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1508 MachineFunction &MF = DAG.getMachineFunction();
1509 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1510 SDValue Chain = TheCall->getChain();
1511 unsigned CC = TheCall->getCallingConv();
1512 bool isVarArg = TheCall->isVarArg();
1513 bool IsTailCall = TheCall->isTailCall() &&
1514 CC == CallingConv::Fast && PerformTailCallOpt;
1515 SDValue Callee = TheCall->getCallee();
1516 bool Is64Bit = Subtarget->is64Bit();
1517 bool IsStructRet = CallIsStructReturn(TheCall);
1519 assert(!(isVarArg && CC == CallingConv::Fast) &&
1520 "Var args not supported with calling convention fastcc");
1522 // Analyze operands of the call, assigning locations to each operand.
1523 SmallVector<CCValAssign, 16> ArgLocs;
1524 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1525 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1527 // Get a count of how many bytes are to be pushed on the stack.
1528 unsigned NumBytes = CCInfo.getNextStackOffset();
1529 if (PerformTailCallOpt && CC == CallingConv::Fast)
1530 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1534 // Lower arguments at fp - stackoffset + fpdiff.
1535 unsigned NumBytesCallerPushed =
1536 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1537 FPDiff = NumBytesCallerPushed - NumBytes;
1539 // Set the delta of movement of the returnaddr stackslot.
1540 // But only set if delta is greater than previous delta.
1541 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1542 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1545 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1547 SDValue RetAddrFrIdx;
1548 // Load return adress for tail calls.
1549 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1552 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1553 SmallVector<SDValue, 8> MemOpChains;
1556 // Walk the register/memloc assignments, inserting copies/loads. In the case
1557 // of tail call optimization arguments are handle later.
1558 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1559 CCValAssign &VA = ArgLocs[i];
1560 SDValue Arg = TheCall->getArg(i);
1561 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1562 bool isByVal = Flags.isByVal();
1564 // Promote the value if needed.
1565 switch (VA.getLocInfo()) {
1566 default: assert(0 && "Unknown loc info!");
1567 case CCValAssign::Full: break;
1568 case CCValAssign::SExt:
1569 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1571 case CCValAssign::ZExt:
1572 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1574 case CCValAssign::AExt:
1575 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1579 if (VA.isRegLoc()) {
1581 MVT RegVT = VA.getLocVT();
1582 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1583 switch (VA.getLocReg()) {
1586 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1588 // Special case: passing MMX values in GPR registers.
1589 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1592 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1593 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1594 // Special case: passing MMX values in XMM registers.
1595 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1596 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1597 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1598 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1599 getMOVLMask(2, DAG));
1604 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1606 if (!IsTailCall || (IsTailCall && isByVal)) {
1607 assert(VA.isMemLoc());
1608 if (StackPtr.getNode() == 0)
1609 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1611 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1612 Chain, Arg, Flags));
1617 if (!MemOpChains.empty())
1618 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1619 &MemOpChains[0], MemOpChains.size());
1621 // Build a sequence of copy-to-reg nodes chained together with token chain
1622 // and flag operands which copy the outgoing args into registers.
1624 // Tail call byval lowering might overwrite argument registers so in case of
1625 // tail call optimization the copies to registers are lowered later.
1627 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1628 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1630 InFlag = Chain.getValue(1);
1633 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1635 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1636 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1637 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1639 InFlag = Chain.getValue(1);
1641 // If we are tail calling and generating PIC/GOT style code load the address
1642 // of the callee into ecx. The value in ecx is used as target of the tail
1643 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1644 // calls on PIC/GOT architectures. Normally we would just put the address of
1645 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1646 // restored (since ebx is callee saved) before jumping to the target@PLT.
1647 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1648 // Note: The actual moving to ecx is done further down.
1649 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1650 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1651 !G->getGlobal()->hasProtectedVisibility())
1652 Callee = LowerGlobalAddress(Callee, DAG);
1653 else if (isa<ExternalSymbolSDNode>(Callee))
1654 Callee = LowerExternalSymbol(Callee,DAG);
1657 if (Is64Bit && isVarArg) {
1658 // From AMD64 ABI document:
1659 // For calls that may call functions that use varargs or stdargs
1660 // (prototype-less calls or calls to functions containing ellipsis (...) in
1661 // the declaration) %al is used as hidden argument to specify the number
1662 // of SSE registers used. The contents of %al do not need to match exactly
1663 // the number of registers, but must be an ubound on the number of SSE
1664 // registers used and is in the range 0 - 8 inclusive.
1666 // FIXME: Verify this on Win64
1667 // Count the number of XMM registers allocated.
1668 static const unsigned XMMArgRegs[] = {
1669 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1670 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1672 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1674 Chain = DAG.getCopyToReg(Chain, X86::AL,
1675 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1676 InFlag = Chain.getValue(1);
1680 // For tail calls lower the arguments to the 'real' stack slot.
1682 SmallVector<SDValue, 8> MemOpChains2;
1685 // Do not flag preceeding copytoreg stuff together with the following stuff.
1687 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1688 CCValAssign &VA = ArgLocs[i];
1689 if (!VA.isRegLoc()) {
1690 assert(VA.isMemLoc());
1691 SDValue Arg = TheCall->getArg(i);
1692 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1693 // Create frame index.
1694 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1695 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1696 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1697 FIN = DAG.getFrameIndex(FI, getPointerTy());
1699 if (Flags.isByVal()) {
1700 // Copy relative to framepointer.
1701 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1702 if (StackPtr.getNode() == 0)
1703 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1704 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1706 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1709 // Store relative to framepointer.
1710 MemOpChains2.push_back(
1711 DAG.getStore(Chain, Arg, FIN,
1712 PseudoSourceValue::getFixedStack(FI), 0));
1717 if (!MemOpChains2.empty())
1718 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1719 &MemOpChains2[0], MemOpChains2.size());
1721 // Copy arguments to their registers.
1722 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1723 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1725 InFlag = Chain.getValue(1);
1729 // Store the return address to the appropriate stack slot.
1730 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1734 // If the callee is a GlobalAddress node (quite common, every direct call is)
1735 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1736 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1737 // We should use extra load for direct calls to dllimported functions in
1739 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1740 getTargetMachine(), true))
1741 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1743 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1744 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1745 } else if (IsTailCall) {
1746 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1748 Chain = DAG.getCopyToReg(Chain,
1749 DAG.getRegister(Opc, getPointerTy()),
1751 Callee = DAG.getRegister(Opc, getPointerTy());
1752 // Add register as live out.
1753 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1756 // Returns a chain & a flag for retval copy to use.
1757 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1758 SmallVector<SDValue, 8> Ops;
1761 Ops.push_back(Chain);
1762 Ops.push_back(DAG.getIntPtrConstant(NumBytes, true));
1763 Ops.push_back(DAG.getIntPtrConstant(0, true));
1764 if (InFlag.getNode())
1765 Ops.push_back(InFlag);
1766 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1767 InFlag = Chain.getValue(1);
1769 // Returns a chain & a flag for retval copy to use.
1770 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1774 Ops.push_back(Chain);
1775 Ops.push_back(Callee);
1778 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1780 // Add argument registers to the end of the list so that they are known live
1782 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1783 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1784 RegsToPass[i].second.getValueType()));
1786 // Add an implicit use GOT pointer in EBX.
1787 if (!IsTailCall && !Is64Bit &&
1788 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1789 Subtarget->isPICStyleGOT())
1790 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1792 // Add an implicit use of AL for x86 vararg functions.
1793 if (Is64Bit && isVarArg)
1794 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1796 if (InFlag.getNode())
1797 Ops.push_back(InFlag);
1800 assert(InFlag.getNode() &&
1801 "Flag must be set. Depend on flag being set in LowerRET");
1802 Chain = DAG.getNode(X86ISD::TAILCALL,
1803 TheCall->getVTList(), &Ops[0], Ops.size());
1805 return SDValue(Chain.getNode(), Op.getResNo());
1808 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1809 InFlag = Chain.getValue(1);
1811 // Create the CALLSEQ_END node.
1812 unsigned NumBytesForCalleeToPush;
1813 if (IsCalleePop(isVarArg, CC))
1814 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1815 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1816 // If this is is a call to a struct-return function, the callee
1817 // pops the hidden struct pointer, so we have to push it back.
1818 // This is common for Darwin/X86, Linux & Mingw32 targets.
1819 NumBytesForCalleeToPush = 4;
1821 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1823 // Returns a flag for retval copy to use.
1824 Chain = DAG.getCALLSEQ_END(Chain,
1825 DAG.getIntPtrConstant(NumBytes, true),
1826 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1829 InFlag = Chain.getValue(1);
1831 // Handle result values, copying them out of physregs into vregs that we
1833 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1838 //===----------------------------------------------------------------------===//
1839 // Fast Calling Convention (tail call) implementation
1840 //===----------------------------------------------------------------------===//
1842 // Like std call, callee cleans arguments, convention except that ECX is
1843 // reserved for storing the tail called function address. Only 2 registers are
1844 // free for argument passing (inreg). Tail call optimization is performed
1846 // * tailcallopt is enabled
1847 // * caller/callee are fastcc
1848 // On X86_64 architecture with GOT-style position independent code only local
1849 // (within module) calls are supported at the moment.
1850 // To keep the stack aligned according to platform abi the function
1851 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1852 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1853 // If a tail called function callee has more arguments than the caller the
1854 // caller needs to make sure that there is room to move the RETADDR to. This is
1855 // achieved by reserving an area the size of the argument delta right after the
1856 // original REtADDR, but before the saved framepointer or the spilled registers
1857 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1869 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1870 /// for a 16 byte align requirement.
1871 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1872 SelectionDAG& DAG) {
1873 MachineFunction &MF = DAG.getMachineFunction();
1874 const TargetMachine &TM = MF.getTarget();
1875 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1876 unsigned StackAlignment = TFI.getStackAlignment();
1877 uint64_t AlignMask = StackAlignment - 1;
1878 int64_t Offset = StackSize;
1879 uint64_t SlotSize = TD->getPointerSize();
1880 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1881 // Number smaller than 12 so just add the difference.
1882 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1884 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1885 Offset = ((~AlignMask) & Offset) + StackAlignment +
1886 (StackAlignment-SlotSize);
1891 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1892 /// following the call is a return. A function is eligible if caller/callee
1893 /// calling conventions match, currently only fastcc supports tail calls, and
1894 /// the function CALL is immediatly followed by a RET.
1895 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1897 SelectionDAG& DAG) const {
1898 if (!PerformTailCallOpt)
1901 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 unsigned CallerCC = MF.getFunction()->getCallingConv();
1904 unsigned CalleeCC= TheCall->getCallingConv();
1905 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1906 SDValue Callee = TheCall->getCallee();
1907 // On x86/32Bit PIC/GOT tail calls are supported.
1908 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1909 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1912 // Can only do local tail calls (in same module, hidden or protected) on
1913 // x86_64 PIC/GOT at the moment.
1914 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1915 return G->getGlobal()->hasHiddenVisibility()
1916 || G->getGlobal()->hasProtectedVisibility();
1924 X86TargetLowering::createFastISel(MachineFunction &mf,
1925 MachineModuleInfo *mmo,
1926 DenseMap<const Value *, unsigned> &vm,
1927 DenseMap<const BasicBlock *,
1928 MachineBasicBlock *> &bm,
1929 DenseMap<const AllocaInst *, int> &am
1931 , SmallSet<Instruction*, 8> &cil
1934 return X86::createFastISel(mf, mmo, vm, bm, am
1942 //===----------------------------------------------------------------------===//
1943 // Other Lowering Hooks
1944 //===----------------------------------------------------------------------===//
1947 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1948 MachineFunction &MF = DAG.getMachineFunction();
1949 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1950 int ReturnAddrIndex = FuncInfo->getRAIndex();
1951 uint64_t SlotSize = TD->getPointerSize();
1953 if (ReturnAddrIndex == 0) {
1954 // Set up a frame object for the return address.
1955 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1956 FuncInfo->setRAIndex(ReturnAddrIndex);
1959 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1963 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
1964 /// specific condition code, returning the condition code and the LHS/RHS of the
1965 /// comparison to make.
1966 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1967 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
1969 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1970 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1971 // X > -1 -> X == 0, jump !sign.
1972 RHS = DAG.getConstant(0, RHS.getValueType());
1973 return X86::COND_NS;
1974 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1975 // X < 0 -> X == 0, jump on sign.
1977 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
1979 RHS = DAG.getConstant(0, RHS.getValueType());
1980 return X86::COND_LE;
1984 switch (SetCCOpcode) {
1985 default: assert(0 && "Invalid integer condition!");
1986 case ISD::SETEQ: return X86::COND_E;
1987 case ISD::SETGT: return X86::COND_G;
1988 case ISD::SETGE: return X86::COND_GE;
1989 case ISD::SETLT: return X86::COND_L;
1990 case ISD::SETLE: return X86::COND_LE;
1991 case ISD::SETNE: return X86::COND_NE;
1992 case ISD::SETULT: return X86::COND_B;
1993 case ISD::SETUGT: return X86::COND_A;
1994 case ISD::SETULE: return X86::COND_BE;
1995 case ISD::SETUGE: return X86::COND_AE;
1999 // First determine if it is required or is profitable to flip the operands.
2001 // If LHS is a foldable load, but RHS is not, flip the condition.
2002 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2003 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2004 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2005 std::swap(LHS, RHS);
2008 switch (SetCCOpcode) {
2014 std::swap(LHS, RHS);
2018 // On a floating point condition, the flags are set as follows:
2020 // 0 | 0 | 0 | X > Y
2021 // 0 | 0 | 1 | X < Y
2022 // 1 | 0 | 0 | X == Y
2023 // 1 | 1 | 1 | unordered
2024 switch (SetCCOpcode) {
2025 default: assert(0 && "Condcode should be pre-legalized away");
2027 case ISD::SETEQ: return X86::COND_E;
2028 case ISD::SETOLT: // flipped
2030 case ISD::SETGT: return X86::COND_A;
2031 case ISD::SETOLE: // flipped
2033 case ISD::SETGE: return X86::COND_AE;
2034 case ISD::SETUGT: // flipped
2036 case ISD::SETLT: return X86::COND_B;
2037 case ISD::SETUGE: // flipped
2039 case ISD::SETLE: return X86::COND_BE;
2041 case ISD::SETNE: return X86::COND_NE;
2042 case ISD::SETUO: return X86::COND_P;
2043 case ISD::SETO: return X86::COND_NP;
2047 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2048 /// code. Current x86 isa includes the following FP cmov instructions:
2049 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2050 static bool hasFPCMov(unsigned X86CC) {
2066 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2067 /// true if Op is undef or if its value falls within the specified range (L, H].
2068 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2069 if (Op.getOpcode() == ISD::UNDEF)
2072 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2073 return (Val >= Low && Val < Hi);
2076 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2077 /// true if Op is undef or if its value equal to the specified value.
2078 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2079 if (Op.getOpcode() == ISD::UNDEF)
2081 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2084 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2085 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2086 bool X86::isPSHUFDMask(SDNode *N) {
2087 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2089 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2092 // Check if the value doesn't reference the second vector.
2093 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2094 SDValue Arg = N->getOperand(i);
2095 if (Arg.getOpcode() == ISD::UNDEF) continue;
2096 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2097 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2104 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2105 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2106 bool X86::isPSHUFHWMask(SDNode *N) {
2107 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2109 if (N->getNumOperands() != 8)
2112 // Lower quadword copied in order.
2113 for (unsigned i = 0; i != 4; ++i) {
2114 SDValue Arg = N->getOperand(i);
2115 if (Arg.getOpcode() == ISD::UNDEF) continue;
2116 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2117 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2121 // Upper quadword shuffled.
2122 for (unsigned i = 4; i != 8; ++i) {
2123 SDValue Arg = N->getOperand(i);
2124 if (Arg.getOpcode() == ISD::UNDEF) continue;
2125 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2126 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2127 if (Val < 4 || Val > 7)
2134 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2135 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2136 bool X86::isPSHUFLWMask(SDNode *N) {
2137 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2139 if (N->getNumOperands() != 8)
2142 // Upper quadword copied in order.
2143 for (unsigned i = 4; i != 8; ++i)
2144 if (!isUndefOrEqual(N->getOperand(i), i))
2147 // Lower quadword shuffled.
2148 for (unsigned i = 0; i != 4; ++i)
2149 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2155 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2156 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2157 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2158 if (NumElems != 2 && NumElems != 4) return false;
2160 unsigned Half = NumElems / 2;
2161 for (unsigned i = 0; i < Half; ++i)
2162 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2164 for (unsigned i = Half; i < NumElems; ++i)
2165 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2171 bool X86::isSHUFPMask(SDNode *N) {
2172 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2173 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2176 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2177 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2178 /// half elements to come from vector 1 (which would equal the dest.) and
2179 /// the upper half to come from vector 2.
2180 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2181 if (NumOps != 2 && NumOps != 4) return false;
2183 unsigned Half = NumOps / 2;
2184 for (unsigned i = 0; i < Half; ++i)
2185 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2187 for (unsigned i = Half; i < NumOps; ++i)
2188 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2193 static bool isCommutedSHUFP(SDNode *N) {
2194 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2195 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2198 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2199 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2200 bool X86::isMOVHLPSMask(SDNode *N) {
2201 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2203 if (N->getNumOperands() != 4)
2206 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2207 return isUndefOrEqual(N->getOperand(0), 6) &&
2208 isUndefOrEqual(N->getOperand(1), 7) &&
2209 isUndefOrEqual(N->getOperand(2), 2) &&
2210 isUndefOrEqual(N->getOperand(3), 3);
2213 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2214 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2216 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2217 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2219 if (N->getNumOperands() != 4)
2222 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2223 return isUndefOrEqual(N->getOperand(0), 2) &&
2224 isUndefOrEqual(N->getOperand(1), 3) &&
2225 isUndefOrEqual(N->getOperand(2), 2) &&
2226 isUndefOrEqual(N->getOperand(3), 3);
2229 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2230 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2231 bool X86::isMOVLPMask(SDNode *N) {
2232 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2234 unsigned NumElems = N->getNumOperands();
2235 if (NumElems != 2 && NumElems != 4)
2238 for (unsigned i = 0; i < NumElems/2; ++i)
2239 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2242 for (unsigned i = NumElems/2; i < NumElems; ++i)
2243 if (!isUndefOrEqual(N->getOperand(i), i))
2249 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2250 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2252 bool X86::isMOVHPMask(SDNode *N) {
2253 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2255 unsigned NumElems = N->getNumOperands();
2256 if (NumElems != 2 && NumElems != 4)
2259 for (unsigned i = 0; i < NumElems/2; ++i)
2260 if (!isUndefOrEqual(N->getOperand(i), i))
2263 for (unsigned i = 0; i < NumElems/2; ++i) {
2264 SDValue Arg = N->getOperand(i + NumElems/2);
2265 if (!isUndefOrEqual(Arg, i + NumElems))
2272 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2273 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2274 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2275 bool V2IsSplat = false) {
2276 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2279 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2280 SDValue BitI = Elts[i];
2281 SDValue BitI1 = Elts[i+1];
2282 if (!isUndefOrEqual(BitI, j))
2285 if (isUndefOrEqual(BitI1, NumElts))
2288 if (!isUndefOrEqual(BitI1, j + NumElts))
2296 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2297 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2298 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2301 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2302 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2303 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2304 bool V2IsSplat = false) {
2305 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2308 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2309 SDValue BitI = Elts[i];
2310 SDValue BitI1 = Elts[i+1];
2311 if (!isUndefOrEqual(BitI, j + NumElts/2))
2314 if (isUndefOrEqual(BitI1, NumElts))
2317 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2325 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2326 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2327 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2330 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2331 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2333 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2334 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2336 unsigned NumElems = N->getNumOperands();
2337 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2340 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2341 SDValue BitI = N->getOperand(i);
2342 SDValue BitI1 = N->getOperand(i+1);
2344 if (!isUndefOrEqual(BitI, j))
2346 if (!isUndefOrEqual(BitI1, j))
2353 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2354 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2356 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2357 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2359 unsigned NumElems = N->getNumOperands();
2360 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2363 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2364 SDValue BitI = N->getOperand(i);
2365 SDValue BitI1 = N->getOperand(i + 1);
2367 if (!isUndefOrEqual(BitI, j))
2369 if (!isUndefOrEqual(BitI1, j))
2376 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2377 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2378 /// MOVSD, and MOVD, i.e. setting the lowest element.
2379 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2380 if (NumElts != 2 && NumElts != 4)
2383 if (!isUndefOrEqual(Elts[0], NumElts))
2386 for (unsigned i = 1; i < NumElts; ++i) {
2387 if (!isUndefOrEqual(Elts[i], i))
2394 bool X86::isMOVLMask(SDNode *N) {
2395 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2396 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2399 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2400 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2401 /// element of vector 2 and the other elements to come from vector 1 in order.
2402 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2403 bool V2IsSplat = false,
2404 bool V2IsUndef = false) {
2405 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2408 if (!isUndefOrEqual(Ops[0], 0))
2411 for (unsigned i = 1; i < NumOps; ++i) {
2412 SDValue Arg = Ops[i];
2413 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2414 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2415 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2422 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2423 bool V2IsUndef = false) {
2424 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2425 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2426 V2IsSplat, V2IsUndef);
2429 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2430 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2431 bool X86::isMOVSHDUPMask(SDNode *N) {
2432 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2434 if (N->getNumOperands() != 4)
2437 // Expect 1, 1, 3, 3
2438 for (unsigned i = 0; i < 2; ++i) {
2439 SDValue Arg = N->getOperand(i);
2440 if (Arg.getOpcode() == ISD::UNDEF) continue;
2441 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2442 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2443 if (Val != 1) return false;
2447 for (unsigned i = 2; i < 4; ++i) {
2448 SDValue Arg = N->getOperand(i);
2449 if (Arg.getOpcode() == ISD::UNDEF) continue;
2450 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2451 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2452 if (Val != 3) return false;
2456 // Don't use movshdup if it can be done with a shufps.
2460 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2461 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2462 bool X86::isMOVSLDUPMask(SDNode *N) {
2463 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2465 if (N->getNumOperands() != 4)
2468 // Expect 0, 0, 2, 2
2469 for (unsigned i = 0; i < 2; ++i) {
2470 SDValue Arg = N->getOperand(i);
2471 if (Arg.getOpcode() == ISD::UNDEF) continue;
2472 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2473 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2474 if (Val != 0) return false;
2478 for (unsigned i = 2; i < 4; ++i) {
2479 SDValue Arg = N->getOperand(i);
2480 if (Arg.getOpcode() == ISD::UNDEF) continue;
2481 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2482 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2483 if (Val != 2) return false;
2487 // Don't use movshdup if it can be done with a shufps.
2491 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2492 /// specifies a identity operation on the LHS or RHS.
2493 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2494 unsigned NumElems = N->getNumOperands();
2495 for (unsigned i = 0; i < NumElems; ++i)
2496 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2501 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2502 /// a splat of a single element.
2503 static bool isSplatMask(SDNode *N) {
2504 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2506 // This is a splat operation if each element of the permute is the same, and
2507 // if the value doesn't reference the second vector.
2508 unsigned NumElems = N->getNumOperands();
2509 SDValue ElementBase;
2511 for (; i != NumElems; ++i) {
2512 SDValue Elt = N->getOperand(i);
2513 if (isa<ConstantSDNode>(Elt)) {
2519 if (!ElementBase.getNode())
2522 for (; i != NumElems; ++i) {
2523 SDValue Arg = N->getOperand(i);
2524 if (Arg.getOpcode() == ISD::UNDEF) continue;
2525 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2526 if (Arg != ElementBase) return false;
2529 // Make sure it is a splat of the first vector operand.
2530 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2533 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2534 /// we want to splat.
2535 static SDValue getSplatMaskEltNo(SDNode *N) {
2536 assert(isSplatMask(N) && "Not a splat mask");
2537 unsigned NumElems = N->getNumOperands();
2538 SDValue ElementBase;
2540 for (; i != NumElems; ++i) {
2541 SDValue Elt = N->getOperand(i);
2542 if (isa<ConstantSDNode>(Elt))
2545 assert(0 && " No splat value found!");
2550 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2551 /// a splat of a single element and it's a 2 or 4 element mask.
2552 bool X86::isSplatMask(SDNode *N) {
2553 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2555 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2556 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2558 return ::isSplatMask(N);
2561 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2562 /// specifies a splat of zero element.
2563 bool X86::isSplatLoMask(SDNode *N) {
2564 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2566 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2567 if (!isUndefOrEqual(N->getOperand(i), 0))
2572 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2573 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2574 bool X86::isMOVDDUPMask(SDNode *N) {
2575 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2577 unsigned e = N->getNumOperands() / 2;
2578 for (unsigned i = 0; i < e; ++i)
2579 if (!isUndefOrEqual(N->getOperand(i), i))
2581 for (unsigned i = 0; i < e; ++i)
2582 if (!isUndefOrEqual(N->getOperand(e+i), i))
2587 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2588 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2590 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2591 unsigned NumOperands = N->getNumOperands();
2592 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2594 for (unsigned i = 0; i < NumOperands; ++i) {
2596 SDValue Arg = N->getOperand(NumOperands-i-1);
2597 if (Arg.getOpcode() != ISD::UNDEF)
2598 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2599 if (Val >= NumOperands) Val -= NumOperands;
2601 if (i != NumOperands - 1)
2608 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2609 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2611 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2613 // 8 nodes, but we only care about the last 4.
2614 for (unsigned i = 7; i >= 4; --i) {
2616 SDValue Arg = N->getOperand(i);
2617 if (Arg.getOpcode() != ISD::UNDEF)
2618 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2627 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2628 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2630 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2632 // 8 nodes, but we only care about the first 4.
2633 for (int i = 3; i >= 0; --i) {
2635 SDValue Arg = N->getOperand(i);
2636 if (Arg.getOpcode() != ISD::UNDEF)
2637 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2646 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2647 /// specifies a 8 element shuffle that can be broken into a pair of
2648 /// PSHUFHW and PSHUFLW.
2649 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2650 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2652 if (N->getNumOperands() != 8)
2655 // Lower quadword shuffled.
2656 for (unsigned i = 0; i != 4; ++i) {
2657 SDValue Arg = N->getOperand(i);
2658 if (Arg.getOpcode() == ISD::UNDEF) continue;
2659 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2660 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2665 // Upper quadword shuffled.
2666 for (unsigned i = 4; i != 8; ++i) {
2667 SDValue Arg = N->getOperand(i);
2668 if (Arg.getOpcode() == ISD::UNDEF) continue;
2669 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2670 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2671 if (Val < 4 || Val > 7)
2678 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2679 /// values in ther permute mask.
2680 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2681 SDValue &V2, SDValue &Mask,
2682 SelectionDAG &DAG) {
2683 MVT VT = Op.getValueType();
2684 MVT MaskVT = Mask.getValueType();
2685 MVT EltVT = MaskVT.getVectorElementType();
2686 unsigned NumElems = Mask.getNumOperands();
2687 SmallVector<SDValue, 8> MaskVec;
2689 for (unsigned i = 0; i != NumElems; ++i) {
2690 SDValue Arg = Mask.getOperand(i);
2691 if (Arg.getOpcode() == ISD::UNDEF) {
2692 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2695 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2696 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2698 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2700 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2704 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2705 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2708 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2709 /// the two vector operands have swapped position.
2711 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2712 MVT MaskVT = Mask.getValueType();
2713 MVT EltVT = MaskVT.getVectorElementType();
2714 unsigned NumElems = Mask.getNumOperands();
2715 SmallVector<SDValue, 8> MaskVec;
2716 for (unsigned i = 0; i != NumElems; ++i) {
2717 SDValue Arg = Mask.getOperand(i);
2718 if (Arg.getOpcode() == ISD::UNDEF) {
2719 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2722 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2723 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2725 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2727 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2729 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2733 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2734 /// match movhlps. The lower half elements should come from upper half of
2735 /// V1 (and in order), and the upper half elements should come from the upper
2736 /// half of V2 (and in order).
2737 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2738 unsigned NumElems = Mask->getNumOperands();
2741 for (unsigned i = 0, e = 2; i != e; ++i)
2742 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2744 for (unsigned i = 2; i != 4; ++i)
2745 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2750 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2751 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2753 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2754 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2756 N = N->getOperand(0).getNode();
2757 if (!ISD::isNON_EXTLoad(N))
2760 *LD = cast<LoadSDNode>(N);
2764 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2765 /// match movlp{s|d}. The lower half elements should come from lower half of
2766 /// V1 (and in order), and the upper half elements should come from the upper
2767 /// half of V2 (and in order). And since V1 will become the source of the
2768 /// MOVLP, it must be either a vector load or a scalar load to vector.
2769 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2770 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2772 // Is V2 is a vector load, don't do this transformation. We will try to use
2773 // load folding shufps op.
2774 if (ISD::isNON_EXTLoad(V2))
2777 unsigned NumElems = Mask->getNumOperands();
2778 if (NumElems != 2 && NumElems != 4)
2780 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2781 if (!isUndefOrEqual(Mask->getOperand(i), i))
2783 for (unsigned i = NumElems/2; i != NumElems; ++i)
2784 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2789 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2791 static bool isSplatVector(SDNode *N) {
2792 if (N->getOpcode() != ISD::BUILD_VECTOR)
2795 SDValue SplatValue = N->getOperand(0);
2796 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2797 if (N->getOperand(i) != SplatValue)
2802 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2804 static bool isUndefShuffle(SDNode *N) {
2805 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2808 SDValue V1 = N->getOperand(0);
2809 SDValue V2 = N->getOperand(1);
2810 SDValue Mask = N->getOperand(2);
2811 unsigned NumElems = Mask.getNumOperands();
2812 for (unsigned i = 0; i != NumElems; ++i) {
2813 SDValue Arg = Mask.getOperand(i);
2814 if (Arg.getOpcode() != ISD::UNDEF) {
2815 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2816 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2818 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2825 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2827 static inline bool isZeroNode(SDValue Elt) {
2828 return ((isa<ConstantSDNode>(Elt) &&
2829 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2830 (isa<ConstantFPSDNode>(Elt) &&
2831 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2834 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2835 /// to an zero vector.
2836 static bool isZeroShuffle(SDNode *N) {
2837 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2840 SDValue V1 = N->getOperand(0);
2841 SDValue V2 = N->getOperand(1);
2842 SDValue Mask = N->getOperand(2);
2843 unsigned NumElems = Mask.getNumOperands();
2844 for (unsigned i = 0; i != NumElems; ++i) {
2845 SDValue Arg = Mask.getOperand(i);
2846 if (Arg.getOpcode() == ISD::UNDEF)
2849 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2850 if (Idx < NumElems) {
2851 unsigned Opc = V1.getNode()->getOpcode();
2852 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2854 if (Opc != ISD::BUILD_VECTOR ||
2855 !isZeroNode(V1.getNode()->getOperand(Idx)))
2857 } else if (Idx >= NumElems) {
2858 unsigned Opc = V2.getNode()->getOpcode();
2859 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2861 if (Opc != ISD::BUILD_VECTOR ||
2862 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2869 /// getZeroVector - Returns a vector of specified type with all zero elements.
2871 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2872 assert(VT.isVector() && "Expected a vector type");
2874 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2875 // type. This ensures they get CSE'd.
2877 if (VT.getSizeInBits() == 64) { // MMX
2878 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2879 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2880 } else if (HasSSE2) { // SSE2
2881 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2882 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2884 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2885 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2887 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2890 /// getOnesVector - Returns a vector of specified type with all bits set.
2892 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2893 assert(VT.isVector() && "Expected a vector type");
2895 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2896 // type. This ensures they get CSE'd.
2897 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2899 if (VT.getSizeInBits() == 64) // MMX
2900 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2902 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2903 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2907 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2908 /// that point to V2 points to its first element.
2909 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2910 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2912 bool Changed = false;
2913 SmallVector<SDValue, 8> MaskVec;
2914 unsigned NumElems = Mask.getNumOperands();
2915 for (unsigned i = 0; i != NumElems; ++i) {
2916 SDValue Arg = Mask.getOperand(i);
2917 if (Arg.getOpcode() != ISD::UNDEF) {
2918 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2919 if (Val > NumElems) {
2920 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2924 MaskVec.push_back(Arg);
2928 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2929 &MaskVec[0], MaskVec.size());
2933 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2934 /// operation of specified width.
2935 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2936 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2937 MVT BaseVT = MaskVT.getVectorElementType();
2939 SmallVector<SDValue, 8> MaskVec;
2940 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2941 for (unsigned i = 1; i != NumElems; ++i)
2942 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2943 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2946 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2947 /// of specified width.
2948 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2949 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2950 MVT BaseVT = MaskVT.getVectorElementType();
2951 SmallVector<SDValue, 8> MaskVec;
2952 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2953 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2954 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2956 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2959 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2960 /// of specified width.
2961 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2962 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2963 MVT BaseVT = MaskVT.getVectorElementType();
2964 unsigned Half = NumElems/2;
2965 SmallVector<SDValue, 8> MaskVec;
2966 for (unsigned i = 0; i != Half; ++i) {
2967 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2968 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2970 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2973 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2974 /// element #0 of a vector with the specified index, leaving the rest of the
2975 /// elements in place.
2976 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2977 SelectionDAG &DAG) {
2978 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2979 MVT BaseVT = MaskVT.getVectorElementType();
2980 SmallVector<SDValue, 8> MaskVec;
2981 // Element #0 of the result gets the elt we are replacing.
2982 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2983 for (unsigned i = 1; i != NumElems; ++i)
2984 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2985 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2988 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2989 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2990 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2991 MVT VT = Op.getValueType();
2994 SDValue V1 = Op.getOperand(0);
2995 SDValue Mask = Op.getOperand(2);
2996 unsigned MaskNumElems = Mask.getNumOperands();
2997 unsigned NumElems = MaskNumElems;
2998 // Special handling of v4f32 -> v4i32.
2999 if (VT != MVT::v4f32) {
3000 // Find which element we want to splat.
3001 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3002 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3003 // unpack elements to the correct location
3004 while (NumElems > 4) {
3005 if (EltNo < NumElems/2) {
3006 Mask = getUnpacklMask(MaskNumElems, DAG);
3008 Mask = getUnpackhMask(MaskNumElems, DAG);
3009 EltNo -= NumElems/2;
3011 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
3014 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3015 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
3018 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3019 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3020 DAG.getNode(ISD::UNDEF, PVT), Mask);
3021 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3024 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3025 /// load that's promoted to vector, or a load bitcasted.
3026 static bool isVectorLoad(SDValue Op) {
3027 assert(Op.getValueType().isVector() && "Expected a vector type");
3028 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3029 Op.getOpcode() == ISD::BIT_CONVERT) {
3030 return isa<LoadSDNode>(Op.getOperand(0));
3032 return isa<LoadSDNode>(Op);
3036 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3038 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3039 SelectionDAG &DAG, bool HasSSE3) {
3040 // If we have sse3 and shuffle has more than one use or input is a load, then
3041 // use movddup. Otherwise, use movlhps.
3042 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3043 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3044 MVT VT = Op.getValueType();
3047 unsigned NumElems = PVT.getVectorNumElements();
3048 if (NumElems == 2) {
3049 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3050 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
3052 assert(NumElems == 4);
3053 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3054 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3055 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst0, Cst1, Cst0, Cst1);
3058 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3059 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3060 DAG.getNode(ISD::UNDEF, PVT), Mask);
3061 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3064 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3065 /// vector of zero or undef vector. This produces a shuffle where the low
3066 /// element of V2 is swizzled into the zero/undef vector, landing at element
3067 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3068 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3069 bool isZero, bool HasSSE2,
3070 SelectionDAG &DAG) {
3071 MVT VT = V2.getValueType();
3073 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
3074 unsigned NumElems = V2.getValueType().getVectorNumElements();
3075 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3076 MVT EVT = MaskVT.getVectorElementType();
3077 SmallVector<SDValue, 16> MaskVec;
3078 for (unsigned i = 0; i != NumElems; ++i)
3079 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3080 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3082 MaskVec.push_back(DAG.getConstant(i, EVT));
3083 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3084 &MaskVec[0], MaskVec.size());
3085 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3088 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3089 /// a shuffle that is zero.
3091 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3092 unsigned NumElems, bool Low,
3093 SelectionDAG &DAG) {
3094 unsigned NumZeros = 0;
3095 for (unsigned i = 0; i < NumElems; ++i) {
3096 unsigned Index = Low ? i : NumElems-i-1;
3097 SDValue Idx = Mask.getOperand(Index);
3098 if (Idx.getOpcode() == ISD::UNDEF) {
3102 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3103 if (Elt.getNode() && isZeroNode(Elt))
3111 /// isVectorShift - Returns true if the shuffle can be implemented as a
3112 /// logical left or right shift of a vector.
3113 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3114 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3115 unsigned NumElems = Mask.getNumOperands();
3118 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3121 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3126 bool SeenV1 = false;
3127 bool SeenV2 = false;
3128 for (unsigned i = NumZeros; i < NumElems; ++i) {
3129 unsigned Val = isLeft ? (i - NumZeros) : i;
3130 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3131 if (Idx.getOpcode() == ISD::UNDEF)
3133 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3134 if (Index < NumElems)
3143 if (SeenV1 && SeenV2)
3146 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3152 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3154 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3155 unsigned NumNonZero, unsigned NumZero,
3156 SelectionDAG &DAG, TargetLowering &TLI) {
3162 for (unsigned i = 0; i < 16; ++i) {
3163 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3164 if (ThisIsNonZero && First) {
3166 V = getZeroVector(MVT::v8i16, true, DAG);
3168 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3173 SDValue ThisElt(0, 0), LastElt(0, 0);
3174 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3175 if (LastIsNonZero) {
3176 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3178 if (ThisIsNonZero) {
3179 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3180 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3181 ThisElt, DAG.getConstant(8, MVT::i8));
3183 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3187 if (ThisElt.getNode())
3188 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3189 DAG.getIntPtrConstant(i/2));
3193 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3196 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3198 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3199 unsigned NumNonZero, unsigned NumZero,
3200 SelectionDAG &DAG, TargetLowering &TLI) {
3206 for (unsigned i = 0; i < 8; ++i) {
3207 bool isNonZero = (NonZeros & (1 << i)) != 0;
3211 V = getZeroVector(MVT::v8i16, true, DAG);
3213 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3216 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3217 DAG.getIntPtrConstant(i));
3224 /// getVShift - Return a vector logical shift node.
3226 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3227 unsigned NumBits, SelectionDAG &DAG,
3228 const TargetLowering &TLI) {
3229 bool isMMX = VT.getSizeInBits() == 64;
3230 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3231 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3232 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3233 return DAG.getNode(ISD::BIT_CONVERT, VT,
3234 DAG.getNode(Opc, ShVT, SrcOp,
3235 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3239 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3240 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3241 if (ISD::isBuildVectorAllZeros(Op.getNode())
3242 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3243 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3244 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3245 // eliminated on x86-32 hosts.
3246 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3249 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3250 return getOnesVector(Op.getValueType(), DAG);
3251 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3254 MVT VT = Op.getValueType();
3255 MVT EVT = VT.getVectorElementType();
3256 unsigned EVTBits = EVT.getSizeInBits();
3258 unsigned NumElems = Op.getNumOperands();
3259 unsigned NumZero = 0;
3260 unsigned NumNonZero = 0;
3261 unsigned NonZeros = 0;
3262 bool IsAllConstants = true;
3263 SmallSet<SDValue, 8> Values;
3264 for (unsigned i = 0; i < NumElems; ++i) {
3265 SDValue Elt = Op.getOperand(i);
3266 if (Elt.getOpcode() == ISD::UNDEF)
3269 if (Elt.getOpcode() != ISD::Constant &&
3270 Elt.getOpcode() != ISD::ConstantFP)
3271 IsAllConstants = false;
3272 if (isZeroNode(Elt))
3275 NonZeros |= (1 << i);
3280 if (NumNonZero == 0) {
3281 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3282 return DAG.getNode(ISD::UNDEF, VT);
3285 // Special case for single non-zero, non-undef, element.
3286 if (NumNonZero == 1 && NumElems <= 4) {
3287 unsigned Idx = CountTrailingZeros_32(NonZeros);
3288 SDValue Item = Op.getOperand(Idx);
3290 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3291 // the value are obviously zero, truncate the value to i32 and do the
3292 // insertion that way. Only do this if the value is non-constant or if the
3293 // value is a constant being inserted into element 0. It is cheaper to do
3294 // a constant pool load than it is to do a movd + shuffle.
3295 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3296 (!IsAllConstants || Idx == 0)) {
3297 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3298 // Handle MMX and SSE both.
3299 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3300 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3302 // Truncate the value (which may itself be a constant) to i32, and
3303 // convert it to a vector with movd (S2V+shuffle to zero extend).
3304 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3305 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3306 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3307 Subtarget->hasSSE2(), DAG);
3309 // Now we have our 32-bit value zero extended in the low element of
3310 // a vector. If Idx != 0, swizzle it into place.
3313 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3314 getSwapEltZeroMask(VecElts, Idx, DAG)
3316 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3318 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3322 // If we have a constant or non-constant insertion into the low element of
3323 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3324 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3325 // depending on what the source datatype is. Because we can only get here
3326 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3328 // Don't do this for i64 values on x86-32.
3329 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3330 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3331 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3332 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3333 Subtarget->hasSSE2(), DAG);
3336 // Is it a vector logical left shift?
3337 if (NumElems == 2 && Idx == 1 &&
3338 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3339 unsigned NumBits = VT.getSizeInBits();
3340 return getVShift(true, VT,
3341 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3342 NumBits/2, DAG, *this);
3345 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3348 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3349 // is a non-constant being inserted into an element other than the low one,
3350 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3351 // movd/movss) to move this into the low element, then shuffle it into
3353 if (EVTBits == 32) {
3354 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3356 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3357 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3358 Subtarget->hasSSE2(), DAG);
3359 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3360 MVT MaskEVT = MaskVT.getVectorElementType();
3361 SmallVector<SDValue, 8> MaskVec;
3362 for (unsigned i = 0; i < NumElems; i++)
3363 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3364 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3365 &MaskVec[0], MaskVec.size());
3366 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3367 DAG.getNode(ISD::UNDEF, VT), Mask);
3371 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3372 if (Values.size() == 1)
3375 // A vector full of immediates; various special cases are already
3376 // handled, so this is best done with a single constant-pool load.
3380 // Let legalizer expand 2-wide build_vectors.
3381 if (EVTBits == 64) {
3382 if (NumNonZero == 1) {
3383 // One half is zero or undef.
3384 unsigned Idx = CountTrailingZeros_32(NonZeros);
3385 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3386 Op.getOperand(Idx));
3387 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3388 Subtarget->hasSSE2(), DAG);
3393 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3394 if (EVTBits == 8 && NumElems == 16) {
3395 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3397 if (V.getNode()) return V;
3400 if (EVTBits == 16 && NumElems == 8) {
3401 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3403 if (V.getNode()) return V;
3406 // If element VT is == 32 bits, turn it into a number of shuffles.
3407 SmallVector<SDValue, 8> V;
3409 if (NumElems == 4 && NumZero > 0) {
3410 for (unsigned i = 0; i < 4; ++i) {
3411 bool isZero = !(NonZeros & (1 << i));
3413 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3415 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3418 for (unsigned i = 0; i < 2; ++i) {
3419 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3422 V[i] = V[i*2]; // Must be a zero vector.
3425 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3426 getMOVLMask(NumElems, DAG));
3429 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3430 getMOVLMask(NumElems, DAG));
3433 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3434 getUnpacklMask(NumElems, DAG));
3439 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3440 MVT EVT = MaskVT.getVectorElementType();
3441 SmallVector<SDValue, 8> MaskVec;
3442 bool Reverse = (NonZeros & 0x3) == 2;
3443 for (unsigned i = 0; i < 2; ++i)
3445 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3447 MaskVec.push_back(DAG.getConstant(i, EVT));
3448 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3449 for (unsigned i = 0; i < 2; ++i)
3451 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3453 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3454 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3455 &MaskVec[0], MaskVec.size());
3456 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3459 if (Values.size() > 2) {
3460 // Expand into a number of unpckl*.
3462 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3463 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3464 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3465 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3466 for (unsigned i = 0; i < NumElems; ++i)
3467 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3469 while (NumElems != 0) {
3470 for (unsigned i = 0; i < NumElems; ++i)
3471 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3482 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3483 SDValue PermMask, SelectionDAG &DAG,
3484 TargetLowering &TLI) {
3486 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3487 MVT MaskEVT = MaskVT.getVectorElementType();
3488 MVT PtrVT = TLI.getPointerTy();
3489 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3490 PermMask.getNode()->op_end());
3492 // First record which half of which vector the low elements come from.
3493 SmallVector<unsigned, 4> LowQuad(4);
3494 for (unsigned i = 0; i < 4; ++i) {
3495 SDValue Elt = MaskElts[i];
3496 if (Elt.getOpcode() == ISD::UNDEF)
3498 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3499 int QuadIdx = EltIdx / 4;
3503 int BestLowQuad = -1;
3504 unsigned MaxQuad = 1;
3505 for (unsigned i = 0; i < 4; ++i) {
3506 if (LowQuad[i] > MaxQuad) {
3508 MaxQuad = LowQuad[i];
3512 // Record which half of which vector the high elements come from.
3513 SmallVector<unsigned, 4> HighQuad(4);
3514 for (unsigned i = 4; i < 8; ++i) {
3515 SDValue Elt = MaskElts[i];
3516 if (Elt.getOpcode() == ISD::UNDEF)
3518 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3519 int QuadIdx = EltIdx / 4;
3520 ++HighQuad[QuadIdx];
3523 int BestHighQuad = -1;
3525 for (unsigned i = 0; i < 4; ++i) {
3526 if (HighQuad[i] > MaxQuad) {
3528 MaxQuad = HighQuad[i];
3532 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3533 if (BestLowQuad != -1 || BestHighQuad != -1) {
3534 // First sort the 4 chunks in order using shufpd.
3535 SmallVector<SDValue, 8> MaskVec;
3537 if (BestLowQuad != -1)
3538 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3540 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3542 if (BestHighQuad != -1)
3543 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3545 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3547 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3548 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3549 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3550 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3551 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3553 // Now sort high and low parts separately.
3554 BitVector InOrder(8);
3555 if (BestLowQuad != -1) {
3556 // Sort lower half in order using PSHUFLW.
3558 bool AnyOutOrder = false;
3560 for (unsigned i = 0; i != 4; ++i) {
3561 SDValue Elt = MaskElts[i];
3562 if (Elt.getOpcode() == ISD::UNDEF) {
3563 MaskVec.push_back(Elt);
3566 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3570 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3572 // If this element is in the right place after this shuffle, then
3574 if ((int)(EltIdx / 4) == BestLowQuad)
3579 for (unsigned i = 4; i != 8; ++i)
3580 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3581 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3582 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3586 if (BestHighQuad != -1) {
3587 // Sort high half in order using PSHUFHW if possible.
3590 for (unsigned i = 0; i != 4; ++i)
3591 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3593 bool AnyOutOrder = false;
3594 for (unsigned i = 4; i != 8; ++i) {
3595 SDValue Elt = MaskElts[i];
3596 if (Elt.getOpcode() == ISD::UNDEF) {
3597 MaskVec.push_back(Elt);
3600 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3604 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3606 // If this element is in the right place after this shuffle, then
3608 if ((int)(EltIdx / 4) == BestHighQuad)
3614 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3615 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3619 // The other elements are put in the right place using pextrw and pinsrw.
3620 for (unsigned i = 0; i != 8; ++i) {
3623 SDValue Elt = MaskElts[i];
3624 if (Elt.getOpcode() == ISD::UNDEF)
3626 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3627 SDValue ExtOp = (EltIdx < 8)
3628 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3629 DAG.getConstant(EltIdx, PtrVT))
3630 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3631 DAG.getConstant(EltIdx - 8, PtrVT));
3632 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3633 DAG.getConstant(i, PtrVT));
3639 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3640 // few as possible. First, let's find out how many elements are already in the
3642 unsigned V1InOrder = 0;
3643 unsigned V1FromV1 = 0;
3644 unsigned V2InOrder = 0;
3645 unsigned V2FromV2 = 0;
3646 SmallVector<SDValue, 8> V1Elts;
3647 SmallVector<SDValue, 8> V2Elts;
3648 for (unsigned i = 0; i < 8; ++i) {
3649 SDValue Elt = MaskElts[i];
3650 if (Elt.getOpcode() == ISD::UNDEF) {
3651 V1Elts.push_back(Elt);
3652 V2Elts.push_back(Elt);
3657 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3659 V1Elts.push_back(Elt);
3660 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3662 } else if (EltIdx == i+8) {
3663 V1Elts.push_back(Elt);
3664 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3666 } else if (EltIdx < 8) {
3667 V1Elts.push_back(Elt);
3668 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3671 V1Elts.push_back(Elt);
3672 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3677 if (V2InOrder > V1InOrder) {
3678 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3680 std::swap(V1Elts, V2Elts);
3681 std::swap(V1FromV1, V2FromV2);
3684 if ((V1FromV1 + V1InOrder) != 8) {
3685 // Some elements are from V2.
3687 // If there are elements that are from V1 but out of place,
3688 // then first sort them in place
3689 SmallVector<SDValue, 8> MaskVec;
3690 for (unsigned i = 0; i < 8; ++i) {
3691 SDValue Elt = V1Elts[i];
3692 if (Elt.getOpcode() == ISD::UNDEF) {
3693 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3696 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3698 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3700 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3702 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3703 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3707 for (unsigned i = 0; i < 8; ++i) {
3708 SDValue Elt = V1Elts[i];
3709 if (Elt.getOpcode() == ISD::UNDEF)
3711 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3714 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3715 DAG.getConstant(EltIdx - 8, PtrVT));
3716 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3717 DAG.getConstant(i, PtrVT));
3721 // All elements are from V1.
3723 for (unsigned i = 0; i < 8; ++i) {
3724 SDValue Elt = V1Elts[i];
3725 if (Elt.getOpcode() == ISD::UNDEF)
3727 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3728 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3729 DAG.getConstant(EltIdx, PtrVT));
3730 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3731 DAG.getConstant(i, PtrVT));
3737 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3738 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3739 /// done when every pair / quad of shuffle mask elements point to elements in
3740 /// the right sequence. e.g.
3741 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3743 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3745 SDValue PermMask, SelectionDAG &DAG,
3746 TargetLowering &TLI) {
3747 unsigned NumElems = PermMask.getNumOperands();
3748 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3749 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3750 MVT MaskEltVT = MaskVT.getVectorElementType();
3752 switch (VT.getSimpleVT()) {
3753 default: assert(false && "Unexpected!");
3754 case MVT::v4f32: NewVT = MVT::v2f64; break;
3755 case MVT::v4i32: NewVT = MVT::v2i64; break;
3756 case MVT::v8i16: NewVT = MVT::v4i32; break;
3757 case MVT::v16i8: NewVT = MVT::v4i32; break;
3760 if (NewWidth == 2) {
3766 unsigned Scale = NumElems / NewWidth;
3767 SmallVector<SDValue, 8> MaskVec;
3768 for (unsigned i = 0; i < NumElems; i += Scale) {
3769 unsigned StartIdx = ~0U;
3770 for (unsigned j = 0; j < Scale; ++j) {
3771 SDValue Elt = PermMask.getOperand(i+j);
3772 if (Elt.getOpcode() == ISD::UNDEF)
3774 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3775 if (StartIdx == ~0U)
3776 StartIdx = EltIdx - (EltIdx % Scale);
3777 if (EltIdx != StartIdx + j)
3780 if (StartIdx == ~0U)
3781 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3783 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3786 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3787 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3788 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3789 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3790 &MaskVec[0], MaskVec.size()));
3793 /// getVZextMovL - Return a zero-extending vector move low node.
3795 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3796 SDValue SrcOp, SelectionDAG &DAG,
3797 const X86Subtarget *Subtarget) {
3798 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3799 LoadSDNode *LD = NULL;
3800 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3801 LD = dyn_cast<LoadSDNode>(SrcOp);
3803 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3805 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3806 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3807 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3808 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3809 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3811 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3812 return DAG.getNode(ISD::BIT_CONVERT, VT,
3813 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3814 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3821 return DAG.getNode(ISD::BIT_CONVERT, VT,
3822 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3823 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3826 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3829 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3830 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3831 MVT MaskVT = PermMask.getValueType();
3832 MVT MaskEVT = MaskVT.getVectorElementType();
3833 SmallVector<std::pair<int, int>, 8> Locs;
3835 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3838 for (unsigned i = 0; i != 4; ++i) {
3839 SDValue Elt = PermMask.getOperand(i);
3840 if (Elt.getOpcode() == ISD::UNDEF) {
3841 Locs[i] = std::make_pair(-1, -1);
3843 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3844 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3846 Locs[i] = std::make_pair(0, NumLo);
3850 Locs[i] = std::make_pair(1, NumHi);
3852 Mask1[2+NumHi] = Elt;
3858 if (NumLo <= 2 && NumHi <= 2) {
3859 // If no more than two elements come from either vector. This can be
3860 // implemented with two shuffles. First shuffle gather the elements.
3861 // The second shuffle, which takes the first shuffle as both of its
3862 // vector operands, put the elements into the right order.
3863 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3864 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3865 &Mask1[0], Mask1.size()));
3867 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3868 for (unsigned i = 0; i != 4; ++i) {
3869 if (Locs[i].first == -1)
3872 unsigned Idx = (i < 2) ? 0 : 4;
3873 Idx += Locs[i].first * 2 + Locs[i].second;
3874 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3878 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3879 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3880 &Mask2[0], Mask2.size()));
3881 } else if (NumLo == 3 || NumHi == 3) {
3882 // Otherwise, we must have three elements from one vector, call it X, and
3883 // one element from the other, call it Y. First, use a shufps to build an
3884 // intermediate vector with the one element from Y and the element from X
3885 // that will be in the same half in the final destination (the indexes don't
3886 // matter). Then, use a shufps to build the final vector, taking the half
3887 // containing the element from Y from the intermediate, and the other half
3890 // Normalize it so the 3 elements come from V1.
3891 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3895 // Find the element from V2.
3897 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3898 SDValue Elt = PermMask.getOperand(HiIndex);
3899 if (Elt.getOpcode() == ISD::UNDEF)
3901 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3906 Mask1[0] = PermMask.getOperand(HiIndex);
3907 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3908 Mask1[2] = PermMask.getOperand(HiIndex^1);
3909 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3910 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3911 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3914 Mask1[0] = PermMask.getOperand(0);
3915 Mask1[1] = PermMask.getOperand(1);
3916 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3917 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3918 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3919 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3921 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3922 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3923 Mask1[2] = PermMask.getOperand(2);
3924 Mask1[3] = PermMask.getOperand(3);
3925 if (Mask1[2].getOpcode() != ISD::UNDEF)
3927 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
3929 if (Mask1[3].getOpcode() != ISD::UNDEF)
3931 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
3933 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3934 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3938 // Break it into (shuffle shuffle_hi, shuffle_lo).
3940 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3941 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3942 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3943 unsigned MaskIdx = 0;
3946 for (unsigned i = 0; i != 4; ++i) {
3953 SDValue Elt = PermMask.getOperand(i);
3954 if (Elt.getOpcode() == ISD::UNDEF) {
3955 Locs[i] = std::make_pair(-1, -1);
3956 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
3957 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3958 (*MaskPtr)[LoIdx] = Elt;
3961 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3962 (*MaskPtr)[HiIdx] = Elt;
3967 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3968 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3969 &LoMask[0], LoMask.size()));
3970 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3971 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3972 &HiMask[0], HiMask.size()));
3973 SmallVector<SDValue, 8> MaskOps;
3974 for (unsigned i = 0; i != 4; ++i) {
3975 if (Locs[i].first == -1) {
3976 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3978 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3979 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3982 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3983 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3984 &MaskOps[0], MaskOps.size()));
3988 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3989 SDValue V1 = Op.getOperand(0);
3990 SDValue V2 = Op.getOperand(1);
3991 SDValue PermMask = Op.getOperand(2);
3992 MVT VT = Op.getValueType();
3993 unsigned NumElems = PermMask.getNumOperands();
3994 bool isMMX = VT.getSizeInBits() == 64;
3995 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3996 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3997 bool V1IsSplat = false;
3998 bool V2IsSplat = false;
4000 if (isUndefShuffle(Op.getNode()))
4001 return DAG.getNode(ISD::UNDEF, VT);
4003 if (isZeroShuffle(Op.getNode()))
4004 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
4006 if (isIdentityMask(PermMask.getNode()))
4008 else if (isIdentityMask(PermMask.getNode(), true))
4011 // Canonicalize movddup shuffles.
4012 if (V2IsUndef && Subtarget->hasSSE2() &&
4013 VT.getSizeInBits() == 128 &&
4014 X86::isMOVDDUPMask(PermMask.getNode()))
4015 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4017 if (isSplatMask(PermMask.getNode())) {
4018 if (isMMX || NumElems < 4) return Op;
4019 // Promote it to a v4{if}32 splat.
4020 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4023 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4025 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4026 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
4027 if (NewOp.getNode())
4028 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
4029 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4030 // FIXME: Figure out a cleaner way to do this.
4031 // Try to make use of movq to zero out the top part.
4032 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4033 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4035 if (NewOp.getNode()) {
4036 SDValue NewV1 = NewOp.getOperand(0);
4037 SDValue NewV2 = NewOp.getOperand(1);
4038 SDValue NewMask = NewOp.getOperand(2);
4039 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4040 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4041 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
4044 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4045 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4047 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4048 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4053 // Check if this can be converted into a logical shift.
4054 bool isLeft = false;
4057 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4058 if (isShift && ShVal.hasOneUse()) {
4059 // If the shifted value has multiple uses, it may be cheaper to use
4060 // v_set0 + movlhps or movhlps, etc.
4061 MVT EVT = VT.getVectorElementType();
4062 ShAmt *= EVT.getSizeInBits();
4063 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4066 if (X86::isMOVLMask(PermMask.getNode())) {
4069 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4070 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
4075 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4076 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4077 X86::isMOVHLPSMask(PermMask.getNode()) ||
4078 X86::isMOVHPMask(PermMask.getNode()) ||
4079 X86::isMOVLPMask(PermMask.getNode())))
4082 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4083 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4084 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4087 // No better options. Use a vshl / vsrl.
4088 MVT EVT = VT.getVectorElementType();
4089 ShAmt *= EVT.getSizeInBits();
4090 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4093 bool Commuted = false;
4094 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4095 // 1,1,1,1 -> v8i16 though.
4096 V1IsSplat = isSplatVector(V1.getNode());
4097 V2IsSplat = isSplatVector(V2.getNode());
4099 // Canonicalize the splat or undef, if present, to be on the RHS.
4100 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4101 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4102 std::swap(V1IsSplat, V2IsSplat);
4103 std::swap(V1IsUndef, V2IsUndef);
4107 // FIXME: Figure out a cleaner way to do this.
4108 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4109 if (V2IsUndef) return V1;
4110 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4112 // V2 is a splat, so the mask may be malformed. That is, it may point
4113 // to any V2 element. The instruction selectior won't like this. Get
4114 // a corrected mask and commute to form a proper MOVS{S|D}.
4115 SDValue NewMask = getMOVLMask(NumElems, DAG);
4116 if (NewMask.getNode() != PermMask.getNode())
4117 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4122 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4123 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4124 X86::isUNPCKLMask(PermMask.getNode()) ||
4125 X86::isUNPCKHMask(PermMask.getNode()))
4129 // Normalize mask so all entries that point to V2 points to its first
4130 // element then try to match unpck{h|l} again. If match, return a
4131 // new vector_shuffle with the corrected mask.
4132 SDValue NewMask = NormalizeMask(PermMask, DAG);
4133 if (NewMask.getNode() != PermMask.getNode()) {
4134 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4135 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4136 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4137 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4138 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4139 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4144 // Normalize the node to match x86 shuffle ops if needed
4145 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4146 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4149 // Commute is back and try unpck* again.
4150 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4151 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4152 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4153 X86::isUNPCKLMask(PermMask.getNode()) ||
4154 X86::isUNPCKHMask(PermMask.getNode()))
4158 // Try PSHUF* first, then SHUFP*.
4159 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4160 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4161 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4162 if (V2.getOpcode() != ISD::UNDEF)
4163 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4164 DAG.getNode(ISD::UNDEF, VT), PermMask);
4169 if (Subtarget->hasSSE2() &&
4170 (X86::isPSHUFDMask(PermMask.getNode()) ||
4171 X86::isPSHUFHWMask(PermMask.getNode()) ||
4172 X86::isPSHUFLWMask(PermMask.getNode()))) {
4174 if (VT == MVT::v4f32) {
4176 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4177 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4178 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4179 } else if (V2.getOpcode() != ISD::UNDEF)
4180 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4181 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4183 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4187 // Binary or unary shufps.
4188 if (X86::isSHUFPMask(PermMask.getNode()) ||
4189 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4193 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4194 if (VT == MVT::v8i16) {
4195 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4196 if (NewOp.getNode())
4200 // Handle all 4 wide cases with a number of shuffles except for MMX.
4201 if (NumElems == 4 && !isMMX)
4202 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4208 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4209 SelectionDAG &DAG) {
4210 MVT VT = Op.getValueType();
4211 if (VT.getSizeInBits() == 8) {
4212 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4213 Op.getOperand(0), Op.getOperand(1));
4214 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4215 DAG.getValueType(VT));
4216 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4217 } else if (VT.getSizeInBits() == 16) {
4218 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4219 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4221 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4222 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4223 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
4226 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4227 Op.getOperand(0), Op.getOperand(1));
4228 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4229 DAG.getValueType(VT));
4230 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4231 } else if (VT == MVT::f32) {
4232 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4233 // the result back to FR32 register. It's only worth matching if the
4234 // result has a single use which is a store or a bitcast to i32. And in
4235 // the case of a store, it's not worth it if the index is a constant 0,
4236 // because a MOVSSmr can be used instead, which is smaller and faster.
4237 if (!Op.hasOneUse())
4239 SDNode *User = *Op.getNode()->use_begin();
4240 if ((User->getOpcode() != ISD::STORE ||
4241 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4242 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4243 (User->getOpcode() != ISD::BIT_CONVERT ||
4244 User->getValueType(0) != MVT::i32))
4246 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4247 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4249 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4256 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4257 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4260 if (Subtarget->hasSSE41()) {
4261 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4266 MVT VT = Op.getValueType();
4267 // TODO: handle v16i8.
4268 if (VT.getSizeInBits() == 16) {
4269 SDValue Vec = Op.getOperand(0);
4270 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4272 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4273 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4274 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4276 // Transform it so it match pextrw which produces a 32-bit result.
4277 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4278 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4279 Op.getOperand(0), Op.getOperand(1));
4280 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4281 DAG.getValueType(VT));
4282 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4283 } else if (VT.getSizeInBits() == 32) {
4284 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4287 // SHUFPS the element to the lowest double word, then movss.
4288 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4289 SmallVector<SDValue, 8> IdxVec;
4291 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4293 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4295 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4297 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4298 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4299 &IdxVec[0], IdxVec.size());
4300 SDValue Vec = Op.getOperand(0);
4301 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4302 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4303 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4304 DAG.getIntPtrConstant(0));
4305 } else if (VT.getSizeInBits() == 64) {
4306 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4307 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4308 // to match extract_elt for f64.
4309 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4313 // UNPCKHPD the element to the lowest double word, then movsd.
4314 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4315 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4316 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4317 SmallVector<SDValue, 8> IdxVec;
4318 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4320 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4321 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4322 &IdxVec[0], IdxVec.size());
4323 SDValue Vec = Op.getOperand(0);
4324 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4325 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4326 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4327 DAG.getIntPtrConstant(0));
4334 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4335 MVT VT = Op.getValueType();
4336 MVT EVT = VT.getVectorElementType();
4338 SDValue N0 = Op.getOperand(0);
4339 SDValue N1 = Op.getOperand(1);
4340 SDValue N2 = Op.getOperand(2);
4342 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4343 isa<ConstantSDNode>(N2)) {
4344 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4346 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4348 if (N1.getValueType() != MVT::i32)
4349 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4350 if (N2.getValueType() != MVT::i32)
4351 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4352 return DAG.getNode(Opc, VT, N0, N1, N2);
4353 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4354 // Bits [7:6] of the constant are the source select. This will always be
4355 // zero here. The DAG Combiner may combine an extract_elt index into these
4356 // bits. For example (insert (extract, 3), 2) could be matched by putting
4357 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4358 // Bits [5:4] of the constant are the destination select. This is the
4359 // value of the incoming immediate.
4360 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4361 // combine either bitwise AND or insert of float 0.0 to set these bits.
4362 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4363 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4369 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4370 MVT VT = Op.getValueType();
4371 MVT EVT = VT.getVectorElementType();
4373 if (Subtarget->hasSSE41())
4374 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4379 SDValue N0 = Op.getOperand(0);
4380 SDValue N1 = Op.getOperand(1);
4381 SDValue N2 = Op.getOperand(2);
4383 if (EVT.getSizeInBits() == 16) {
4384 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4385 // as its second argument.
4386 if (N1.getValueType() != MVT::i32)
4387 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4388 if (N2.getValueType() != MVT::i32)
4389 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4390 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4396 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4397 if (Op.getValueType() == MVT::v2f32)
4398 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4399 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4400 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4401 Op.getOperand(0))));
4403 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4404 MVT VT = MVT::v2i32;
4405 switch (Op.getValueType().getSimpleVT()) {
4412 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4413 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4416 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4417 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4418 // one of the above mentioned nodes. It has to be wrapped because otherwise
4419 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4420 // be used to form addressing mode. These wrapped nodes will be selected
4423 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4424 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4425 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4427 CP->getAlignment());
4428 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4429 // With PIC, the address is actually $g + Offset.
4430 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4431 !Subtarget->isPICStyleRIPRel()) {
4432 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4433 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4441 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV,
4443 SelectionDAG &DAG) const {
4444 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4445 bool ExtraLoadRequired =
4446 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4448 // Create the TargetGlobalAddress node, folding in the constant
4449 // offset if it is legal.
4451 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4452 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4455 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4456 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4458 // With PIC, the address is actually $g + Offset.
4459 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4460 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4461 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4465 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4466 // load the value at address GV, not the value of GV itself. This means that
4467 // the GlobalAddress must be in the base or index register of the address, not
4468 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4469 // The same applies for external symbols during PIC codegen
4470 if (ExtraLoadRequired)
4471 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4472 PseudoSourceValue::getGOT(), 0);
4474 // If there was a non-zero offset that we didn't fold, create an explicit
4477 Result = DAG.getNode(ISD::ADD, getPointerTy(), Result,
4478 DAG.getConstant(Offset, getPointerTy()));
4484 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4485 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4486 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4487 return LowerGlobalAddress(GV, Offset, DAG);
4490 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4492 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4495 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4496 DAG.getNode(X86ISD::GlobalBaseReg,
4498 InFlag = Chain.getValue(1);
4500 // emit leal symbol@TLSGD(,%ebx,1), %eax
4501 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4502 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4503 GA->getValueType(0),
4505 SDValue Ops[] = { Chain, TGA, InFlag };
4506 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4507 InFlag = Result.getValue(2);
4508 Chain = Result.getValue(1);
4510 // call ___tls_get_addr. This function receives its argument in
4511 // the register EAX.
4512 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4513 InFlag = Chain.getValue(1);
4515 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4516 SDValue Ops1[] = { Chain,
4517 DAG.getTargetExternalSymbol("___tls_get_addr",
4519 DAG.getRegister(X86::EAX, PtrVT),
4520 DAG.getRegister(X86::EBX, PtrVT),
4522 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4523 InFlag = Chain.getValue(1);
4525 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4528 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4530 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4532 SDValue InFlag, Chain;
4534 // emit leaq symbol@TLSGD(%rip), %rdi
4535 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4536 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4537 GA->getValueType(0),
4539 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4540 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4541 Chain = Result.getValue(1);
4542 InFlag = Result.getValue(2);
4544 // call __tls_get_addr. This function receives its argument in
4545 // the register RDI.
4546 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4547 InFlag = Chain.getValue(1);
4549 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4550 SDValue Ops1[] = { Chain,
4551 DAG.getTargetExternalSymbol("__tls_get_addr",
4553 DAG.getRegister(X86::RDI, PtrVT),
4555 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4556 InFlag = Chain.getValue(1);
4558 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4561 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4562 // "local exec" model.
4563 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4565 // Get the Thread Pointer
4566 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4567 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4569 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4570 GA->getValueType(0),
4572 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4574 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4575 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4576 PseudoSourceValue::getGOT(), 0);
4578 // The address of the thread local variable is the add of the thread
4579 // pointer with the offset of the variable.
4580 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4584 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4585 // TODO: implement the "local dynamic" model
4586 // TODO: implement the "initial exec"model for pic executables
4587 assert(Subtarget->isTargetELF() &&
4588 "TLS not implemented for non-ELF targets");
4589 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4590 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4591 // otherwise use the "Local Exec"TLS Model
4592 if (Subtarget->is64Bit()) {
4593 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4595 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4596 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4598 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4603 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4604 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4605 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4606 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4607 // With PIC, the address is actually $g + Offset.
4608 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4609 !Subtarget->isPICStyleRIPRel()) {
4610 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4611 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4618 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4619 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4620 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4621 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4622 // With PIC, the address is actually $g + Offset.
4623 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4624 !Subtarget->isPICStyleRIPRel()) {
4625 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4626 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4633 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4634 /// take a 2 x i32 value to shift plus a shift amount.
4635 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4636 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4637 MVT VT = Op.getValueType();
4638 unsigned VTBits = VT.getSizeInBits();
4639 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4640 SDValue ShOpLo = Op.getOperand(0);
4641 SDValue ShOpHi = Op.getOperand(1);
4642 SDValue ShAmt = Op.getOperand(2);
4643 SDValue Tmp1 = isSRA ?
4644 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4645 DAG.getConstant(0, VT);
4648 if (Op.getOpcode() == ISD::SHL_PARTS) {
4649 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4650 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4652 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4653 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4656 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4657 DAG.getConstant(VTBits, MVT::i8));
4658 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4659 AndNode, DAG.getConstant(0, MVT::i8));
4662 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4663 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4664 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4666 if (Op.getOpcode() == ISD::SHL_PARTS) {
4667 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4668 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4670 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4671 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4674 SDValue Ops[2] = { Lo, Hi };
4675 return DAG.getMergeValues(Ops, 2);
4678 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4679 MVT SrcVT = Op.getOperand(0).getValueType();
4680 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4681 "Unknown SINT_TO_FP to lower!");
4683 // These are really Legal; caller falls through into that case.
4684 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4686 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4687 Subtarget->is64Bit())
4690 unsigned Size = SrcVT.getSizeInBits()/8;
4691 MachineFunction &MF = DAG.getMachineFunction();
4692 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4693 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4694 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4696 PseudoSourceValue::getFixedStack(SSFI), 0);
4700 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4702 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4704 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4705 SmallVector<SDValue, 8> Ops;
4706 Ops.push_back(Chain);
4707 Ops.push_back(StackSlot);
4708 Ops.push_back(DAG.getValueType(SrcVT));
4709 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4710 Tys, &Ops[0], Ops.size());
4713 Chain = Result.getValue(1);
4714 SDValue InFlag = Result.getValue(2);
4716 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4717 // shouldn't be necessary except that RFP cannot be live across
4718 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4719 MachineFunction &MF = DAG.getMachineFunction();
4720 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4721 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4722 Tys = DAG.getVTList(MVT::Other);
4723 SmallVector<SDValue, 8> Ops;
4724 Ops.push_back(Chain);
4725 Ops.push_back(Result);
4726 Ops.push_back(StackSlot);
4727 Ops.push_back(DAG.getValueType(Op.getValueType()));
4728 Ops.push_back(InFlag);
4729 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4730 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4731 PseudoSourceValue::getFixedStack(SSFI), 0);
4737 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4738 MVT SrcVT = Op.getOperand(0).getValueType();
4739 assert(SrcVT.getSimpleVT() == MVT::i64 && "Unknown UINT_TO_FP to lower!");
4741 // We only handle SSE2 f64 target here; caller can handle the rest.
4742 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4745 // This algorithm is not obvious. Here it is in C code, more or less:
4747 double uint64_to_double( uint32_t hi, uint32_t lo )
4749 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4750 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4752 // copy ints to xmm registers
4753 __m128i xh = _mm_cvtsi32_si128( hi );
4754 __m128i xl = _mm_cvtsi32_si128( lo );
4756 // combine into low half of a single xmm register
4757 __m128i x = _mm_unpacklo_epi32( xh, xl );
4761 // merge in appropriate exponents to give the integer bits the
4763 x = _mm_unpacklo_epi32( x, exp );
4765 // subtract away the biases to deal with the IEEE-754 double precision
4767 d = _mm_sub_pd( (__m128d) x, bias );
4769 // All conversions up to here are exact. The correctly rounded result is
4770 // calculated using the
4771 // current rounding mode using the following horizontal add.
4772 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4773 _mm_store_sd( &sd, d ); //since we are returning doubles in XMM, this
4774 // store doesn't really need to be here (except maybe to zero the other
4780 // Build some magic constants.
4781 std::vector<Constant*>CV0;
4782 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4783 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4784 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4785 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4786 Constant *C0 = ConstantVector::get(CV0);
4787 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 4);
4789 std::vector<Constant*>CV1;
4790 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4791 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4792 Constant *C1 = ConstantVector::get(CV1);
4793 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 4);
4795 SmallVector<SDValue, 4> MaskVec;
4796 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
4797 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
4798 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
4799 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
4800 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, &MaskVec[0],
4802 SmallVector<SDValue, 4> MaskVec2;
4803 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
4804 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
4805 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec2[0],
4808 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4i32,
4809 DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
4811 DAG.getIntPtrConstant(1)));
4812 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4i32,
4813 DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
4815 DAG.getIntPtrConstant(0)));
4816 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32,
4817 XR1, XR2, UnpcklMask);
4818 SDValue CLod0 = DAG.getLoad(MVT::v4i32, DAG.getEntryNode(), CPIdx0,
4819 PseudoSourceValue::getConstantPool(), 0, false, 16);
4820 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32,
4821 Unpck1, CLod0, UnpcklMask);
4822 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64, Unpck2);
4823 SDValue CLod1 = DAG.getLoad(MVT::v2f64, CLod0.getValue(1), CPIdx1,
4824 PseudoSourceValue::getConstantPool(), 0, false, 16);
4825 SDValue Sub = DAG.getNode(ISD::FSUB, MVT::v2f64, XR2F, CLod1);
4826 // Add the halves; easiest way is to swap them into another reg first.
4827 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2f64,
4828 Sub, Sub, ShufMask);
4829 SDValue Add = DAG.getNode(ISD::FADD, MVT::v2f64, Shuf, Sub);
4830 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f64, Add,
4831 DAG.getIntPtrConstant(0));
4834 std::pair<SDValue,SDValue> X86TargetLowering::
4835 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4836 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4837 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4838 "Unknown FP_TO_SINT to lower!");
4840 // These are really Legal.
4841 if (Op.getValueType() == MVT::i32 &&
4842 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4843 return std::make_pair(SDValue(), SDValue());
4844 if (Subtarget->is64Bit() &&
4845 Op.getValueType() == MVT::i64 &&
4846 Op.getOperand(0).getValueType() != MVT::f80)
4847 return std::make_pair(SDValue(), SDValue());
4849 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4851 MachineFunction &MF = DAG.getMachineFunction();
4852 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4853 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4854 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4856 switch (Op.getValueType().getSimpleVT()) {
4857 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4858 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4859 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4860 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4863 SDValue Chain = DAG.getEntryNode();
4864 SDValue Value = Op.getOperand(0);
4865 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4866 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4867 Chain = DAG.getStore(Chain, Value, StackSlot,
4868 PseudoSourceValue::getFixedStack(SSFI), 0);
4869 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4871 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4873 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4874 Chain = Value.getValue(1);
4875 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4876 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4879 // Build the FP_TO_INT*_IN_MEM
4880 SDValue Ops[] = { Chain, Value, StackSlot };
4881 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4883 return std::make_pair(FIST, StackSlot);
4886 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4887 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4888 SDValue FIST = Vals.first, StackSlot = Vals.second;
4889 if (FIST.getNode() == 0) return SDValue();
4892 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4895 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4896 MVT VT = Op.getValueType();
4899 EltVT = VT.getVectorElementType();
4900 std::vector<Constant*> CV;
4901 if (EltVT == MVT::f64) {
4902 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4906 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4912 Constant *C = ConstantVector::get(CV);
4913 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4914 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4915 PseudoSourceValue::getConstantPool(), 0,
4917 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4920 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4921 MVT VT = Op.getValueType();
4923 unsigned EltNum = 1;
4924 if (VT.isVector()) {
4925 EltVT = VT.getVectorElementType();
4926 EltNum = VT.getVectorNumElements();
4928 std::vector<Constant*> CV;
4929 if (EltVT == MVT::f64) {
4930 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4934 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4940 Constant *C = ConstantVector::get(CV);
4941 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4942 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4943 PseudoSourceValue::getConstantPool(), 0,
4945 if (VT.isVector()) {
4946 return DAG.getNode(ISD::BIT_CONVERT, VT,
4947 DAG.getNode(ISD::XOR, MVT::v2i64,
4948 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4949 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4951 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4955 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4956 SDValue Op0 = Op.getOperand(0);
4957 SDValue Op1 = Op.getOperand(1);
4958 MVT VT = Op.getValueType();
4959 MVT SrcVT = Op1.getValueType();
4961 // If second operand is smaller, extend it first.
4962 if (SrcVT.bitsLT(VT)) {
4963 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4966 // And if it is bigger, shrink it first.
4967 if (SrcVT.bitsGT(VT)) {
4968 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4972 // At this point the operands and the result should have the same
4973 // type, and that won't be f80 since that is not custom lowered.
4975 // First get the sign bit of second operand.
4976 std::vector<Constant*> CV;
4977 if (SrcVT == MVT::f64) {
4978 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4979 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4981 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4982 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4983 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4984 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4986 Constant *C = ConstantVector::get(CV);
4987 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4988 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4989 PseudoSourceValue::getConstantPool(), 0,
4991 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4993 // Shift sign bit right or left if the two operands have different types.
4994 if (SrcVT.bitsGT(VT)) {
4995 // Op0 is MVT::f32, Op1 is MVT::f64.
4996 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4997 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4998 DAG.getConstant(32, MVT::i32));
4999 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
5000 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
5001 DAG.getIntPtrConstant(0));
5004 // Clear first operand sign bit.
5006 if (VT == MVT::f64) {
5007 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5008 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5010 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5011 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5012 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5013 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5015 C = ConstantVector::get(CV);
5016 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5017 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
5018 PseudoSourceValue::getConstantPool(), 0,
5020 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
5022 // Or the value with the sign bit.
5023 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
5026 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5027 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5028 SDValue Op0 = Op.getOperand(0);
5029 SDValue Op1 = Op.getOperand(1);
5030 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5032 // Lower (X & (1 << N)) == 0 to BT.
5033 // Lower ((X >>u N) & 1) != 0 to BT.
5034 // Lower ((X >>s N) & 1) != 0 to BT.
5035 // FIXME: Is i386 or later or available only on some chips?
5036 if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::Constant &&
5037 Op0.getOperand(1).getOpcode() == ISD::Constant &&
5038 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5039 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5040 ConstantSDNode *CmpRHS = cast<ConstantSDNode>(Op1);
5041 SDValue AndLHS = Op0.getOperand(0);
5042 if (CmpRHS->getZExtValue() == 0 && AndRHS->getZExtValue() == 1 &&
5043 AndLHS.getOpcode() == ISD::SRL) {
5044 SDValue LHS = AndLHS.getOperand(0);
5045 SDValue RHS = AndLHS.getOperand(1);
5047 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5048 // instruction. Since the shift amount is in-range-or-undefined, we know
5049 // that doing a bittest on the i16 value is ok. We extend to i32 because
5050 // the encoding for the i16 version is larger than the i32 version.
5051 if (LHS.getValueType() == MVT::i8)
5052 LHS = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, LHS);
5054 // If the operand types disagree, extend the shift amount to match. Since
5055 // BT ignores high bits (like shifts) we can use anyextend.
5056 if (LHS.getValueType() != RHS.getValueType())
5057 RHS = DAG.getNode(ISD::ANY_EXTEND, LHS.getValueType(), RHS);
5059 SDValue BT = DAG.getNode(X86ISD::BT, MVT::i32, LHS, RHS);
5060 unsigned Cond = CC == ISD::SETEQ ? X86::COND_NC : X86::COND_C;
5061 return DAG.getNode(X86ISD::SETCC, MVT::i8,
5062 DAG.getConstant(Cond, MVT::i8), BT);
5066 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5067 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5069 SDValue Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
5070 return DAG.getNode(X86ISD::SETCC, MVT::i8,
5071 DAG.getConstant(X86CC, MVT::i8), Cond);
5074 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5076 SDValue Op0 = Op.getOperand(0);
5077 SDValue Op1 = Op.getOperand(1);
5078 SDValue CC = Op.getOperand(2);
5079 MVT VT = Op.getValueType();
5080 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5081 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5085 MVT VT0 = Op0.getValueType();
5086 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5087 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5090 switch (SetCCOpcode) {
5093 case ISD::SETEQ: SSECC = 0; break;
5095 case ISD::SETGT: Swap = true; // Fallthrough
5097 case ISD::SETOLT: SSECC = 1; break;
5099 case ISD::SETGE: Swap = true; // Fallthrough
5101 case ISD::SETOLE: SSECC = 2; break;
5102 case ISD::SETUO: SSECC = 3; break;
5104 case ISD::SETNE: SSECC = 4; break;
5105 case ISD::SETULE: Swap = true;
5106 case ISD::SETUGE: SSECC = 5; break;
5107 case ISD::SETULT: Swap = true;
5108 case ISD::SETUGT: SSECC = 6; break;
5109 case ISD::SETO: SSECC = 7; break;
5112 std::swap(Op0, Op1);
5114 // In the two special cases we can't handle, emit two comparisons.
5116 if (SetCCOpcode == ISD::SETUEQ) {
5118 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5119 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5120 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
5122 else if (SetCCOpcode == ISD::SETONE) {
5124 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5125 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5126 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
5128 assert(0 && "Illegal FP comparison");
5130 // Handle all other FP comparisons here.
5131 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5134 // We are handling one of the integer comparisons here. Since SSE only has
5135 // GT and EQ comparisons for integer, swapping operands and multiple
5136 // operations may be required for some comparisons.
5137 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5138 bool Swap = false, Invert = false, FlipSigns = false;
5140 switch (VT.getSimpleVT()) {
5142 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5143 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5144 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5145 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5148 switch (SetCCOpcode) {
5150 case ISD::SETNE: Invert = true;
5151 case ISD::SETEQ: Opc = EQOpc; break;
5152 case ISD::SETLT: Swap = true;
5153 case ISD::SETGT: Opc = GTOpc; break;
5154 case ISD::SETGE: Swap = true;
5155 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5156 case ISD::SETULT: Swap = true;
5157 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5158 case ISD::SETUGE: Swap = true;
5159 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5162 std::swap(Op0, Op1);
5164 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5165 // bits of the inputs before performing those operations.
5167 MVT EltVT = VT.getVectorElementType();
5168 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
5169 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5170 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
5172 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
5173 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
5176 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
5178 // If the logical-not of the result is required, perform that now.
5180 MVT EltVT = VT.getVectorElementType();
5181 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
5182 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
5183 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
5185 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
5190 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5191 static bool isX86LogicalCmp(unsigned Opc) {
5192 return Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI;
5195 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5196 bool addTest = true;
5197 SDValue Cond = Op.getOperand(0);
5200 if (Cond.getOpcode() == ISD::SETCC)
5201 Cond = LowerSETCC(Cond, DAG);
5203 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5204 // setting operand in place of the X86ISD::SETCC.
5205 if (Cond.getOpcode() == X86ISD::SETCC) {
5206 CC = Cond.getOperand(0);
5208 SDValue Cmp = Cond.getOperand(1);
5209 unsigned Opc = Cmp.getOpcode();
5210 MVT VT = Op.getValueType();
5212 bool IllegalFPCMov = false;
5213 if (VT.isFloatingPoint() && !VT.isVector() &&
5214 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5215 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5217 if (isX86LogicalCmp(Opc) && !IllegalFPCMov) {
5224 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5225 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5228 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5230 SmallVector<SDValue, 4> Ops;
5231 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5232 // condition is true.
5233 Ops.push_back(Op.getOperand(2));
5234 Ops.push_back(Op.getOperand(1));
5236 Ops.push_back(Cond);
5237 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
5240 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5241 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5242 // from the AND / OR.
5243 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5244 Opc = Op.getOpcode();
5245 if (Opc != ISD::OR && Opc != ISD::AND)
5247 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5248 Op.getOperand(0).hasOneUse() &&
5249 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5250 Op.getOperand(1).hasOneUse());
5253 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5254 bool addTest = true;
5255 SDValue Chain = Op.getOperand(0);
5256 SDValue Cond = Op.getOperand(1);
5257 SDValue Dest = Op.getOperand(2);
5260 if (Cond.getOpcode() == ISD::SETCC)
5261 Cond = LowerSETCC(Cond, DAG);
5263 // FIXME: LowerXALUO doesn't handle these!!
5264 else if (Cond.getOpcode() == X86ISD::ADD ||
5265 Cond.getOpcode() == X86ISD::SUB ||
5266 Cond.getOpcode() == X86ISD::SMUL ||
5267 Cond.getOpcode() == X86ISD::UMUL)
5268 Cond = LowerXALUO(Cond, DAG);
5271 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5272 // setting operand in place of the X86ISD::SETCC.
5273 if (Cond.getOpcode() == X86ISD::SETCC) {
5274 CC = Cond.getOperand(0);
5276 SDValue Cmp = Cond.getOperand(1);
5277 unsigned Opc = Cmp.getOpcode();
5278 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5279 if (isX86LogicalCmp(Opc) || Opc == X86ISD::BT) {
5283 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5287 // These can only come from an arithmetic instruction with overflow,
5288 // e.g. SADDO, UADDO.
5289 Cond = Cond.getNode()->getOperand(1);
5296 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5297 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5298 unsigned Opc = Cmp.getOpcode();
5299 if (CondOpc == ISD::OR) {
5300 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5301 // two branches instead of an explicit OR instruction with a
5303 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5304 isX86LogicalCmp(Opc)) {
5305 CC = Cond.getOperand(0).getOperand(0);
5306 Chain = DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5307 Chain, Dest, CC, Cmp);
5308 CC = Cond.getOperand(1).getOperand(0);
5312 } else { // ISD::AND
5313 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5314 // two branches instead of an explicit AND instruction with a
5315 // separate test. However, we only do this if this block doesn't
5316 // have a fall-through edge, because this requires an explicit
5317 // jmp when the condition is false.
5318 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5319 isX86LogicalCmp(Opc) &&
5320 Op.getNode()->hasOneUse()) {
5321 X86::CondCode CCode =
5322 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5323 CCode = X86::GetOppositeBranchCondition(CCode);
5324 CC = DAG.getConstant(CCode, MVT::i8);
5325 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5326 // Look for an unconditional branch following this conditional branch.
5327 // We need this because we need to reverse the successors in order
5328 // to implement FCMP_OEQ.
5329 if (User.getOpcode() == ISD::BR) {
5330 SDValue FalseBB = User.getOperand(1);
5332 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5333 assert(NewBR == User);
5336 Chain = DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5337 Chain, Dest, CC, Cmp);
5338 X86::CondCode CCode =
5339 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5340 CCode = X86::GetOppositeBranchCondition(CCode);
5341 CC = DAG.getConstant(CCode, MVT::i8);
5351 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5352 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5354 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5355 Chain, Dest, CC, Cond);
5359 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5360 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5361 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5362 // that the guard pages used by the OS virtual memory manager are allocated in
5363 // correct sequence.
5365 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5366 SelectionDAG &DAG) {
5367 assert(Subtarget->isTargetCygMing() &&
5368 "This should be used only on Cygwin/Mingw targets");
5371 SDValue Chain = Op.getOperand(0);
5372 SDValue Size = Op.getOperand(1);
5373 // FIXME: Ensure alignment here
5377 MVT IntPtr = getPointerTy();
5378 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5380 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5382 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5383 Flag = Chain.getValue(1);
5385 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5386 SDValue Ops[] = { Chain,
5387 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5388 DAG.getRegister(X86::EAX, IntPtr),
5389 DAG.getRegister(X86StackPtr, SPTy),
5391 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5392 Flag = Chain.getValue(1);
5394 Chain = DAG.getCALLSEQ_END(Chain,
5395 DAG.getIntPtrConstant(0, true),
5396 DAG.getIntPtrConstant(0, true),
5399 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5401 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5402 return DAG.getMergeValues(Ops1, 2);
5406 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5408 SDValue Dst, SDValue Src,
5409 SDValue Size, unsigned Align,
5411 uint64_t DstSVOff) {
5412 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5414 // If not DWORD aligned or size is more than the threshold, call the library.
5415 // The libc version is likely to be faster for these cases. It can use the
5416 // address value and run time information about the CPU.
5417 if ((Align & 3) != 0 ||
5419 ConstantSize->getZExtValue() >
5420 getSubtarget()->getMaxInlineSizeThreshold()) {
5421 SDValue InFlag(0, 0);
5423 // Check to see if there is a specialized entry-point for memory zeroing.
5424 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5426 if (const char *bzeroEntry = V &&
5427 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5428 MVT IntPtr = getPointerTy();
5429 const Type *IntPtrTy = TD->getIntPtrType();
5430 TargetLowering::ArgListTy Args;
5431 TargetLowering::ArgListEntry Entry;
5433 Entry.Ty = IntPtrTy;
5434 Args.push_back(Entry);
5436 Args.push_back(Entry);
5437 std::pair<SDValue,SDValue> CallResult =
5438 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5439 CallingConv::C, false,
5440 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG);
5441 return CallResult.second;
5444 // Otherwise have the target-independent code call memset.
5448 uint64_t SizeVal = ConstantSize->getZExtValue();
5449 SDValue InFlag(0, 0);
5452 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5453 unsigned BytesLeft = 0;
5454 bool TwoRepStos = false;
5457 uint64_t Val = ValC->getZExtValue() & 255;
5459 // If the value is a constant, then we can potentially use larger sets.
5460 switch (Align & 3) {
5461 case 2: // WORD aligned
5464 Val = (Val << 8) | Val;
5466 case 0: // DWORD aligned
5469 Val = (Val << 8) | Val;
5470 Val = (Val << 16) | Val;
5471 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5474 Val = (Val << 32) | Val;
5477 default: // Byte aligned
5480 Count = DAG.getIntPtrConstant(SizeVal);
5484 if (AVT.bitsGT(MVT::i8)) {
5485 unsigned UBytes = AVT.getSizeInBits() / 8;
5486 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5487 BytesLeft = SizeVal % UBytes;
5490 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5492 InFlag = Chain.getValue(1);
5495 Count = DAG.getIntPtrConstant(SizeVal);
5496 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5497 InFlag = Chain.getValue(1);
5500 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5502 InFlag = Chain.getValue(1);
5503 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5505 InFlag = Chain.getValue(1);
5507 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5508 SmallVector<SDValue, 8> Ops;
5509 Ops.push_back(Chain);
5510 Ops.push_back(DAG.getValueType(AVT));
5511 Ops.push_back(InFlag);
5512 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5515 InFlag = Chain.getValue(1);
5517 MVT CVT = Count.getValueType();
5518 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5519 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5520 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5522 InFlag = Chain.getValue(1);
5523 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5525 Ops.push_back(Chain);
5526 Ops.push_back(DAG.getValueType(MVT::i8));
5527 Ops.push_back(InFlag);
5528 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5529 } else if (BytesLeft) {
5530 // Handle the last 1 - 7 bytes.
5531 unsigned Offset = SizeVal - BytesLeft;
5532 MVT AddrVT = Dst.getValueType();
5533 MVT SizeVT = Size.getValueType();
5535 Chain = DAG.getMemset(Chain,
5536 DAG.getNode(ISD::ADD, AddrVT, Dst,
5537 DAG.getConstant(Offset, AddrVT)),
5539 DAG.getConstant(BytesLeft, SizeVT),
5540 Align, DstSV, DstSVOff + Offset);
5543 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5548 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5549 SDValue Chain, SDValue Dst, SDValue Src,
5550 SDValue Size, unsigned Align,
5552 const Value *DstSV, uint64_t DstSVOff,
5553 const Value *SrcSV, uint64_t SrcSVOff) {
5554 // This requires the copy size to be a constant, preferrably
5555 // within a subtarget-specific limit.
5556 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5559 uint64_t SizeVal = ConstantSize->getZExtValue();
5560 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5563 /// If not DWORD aligned, call the library.
5564 if ((Align & 3) != 0)
5569 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5572 unsigned UBytes = AVT.getSizeInBits() / 8;
5573 unsigned CountVal = SizeVal / UBytes;
5574 SDValue Count = DAG.getIntPtrConstant(CountVal);
5575 unsigned BytesLeft = SizeVal % UBytes;
5577 SDValue InFlag(0, 0);
5578 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5580 InFlag = Chain.getValue(1);
5581 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5583 InFlag = Chain.getValue(1);
5584 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5586 InFlag = Chain.getValue(1);
5588 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5589 SmallVector<SDValue, 8> Ops;
5590 Ops.push_back(Chain);
5591 Ops.push_back(DAG.getValueType(AVT));
5592 Ops.push_back(InFlag);
5593 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5595 SmallVector<SDValue, 4> Results;
5596 Results.push_back(RepMovs);
5598 // Handle the last 1 - 7 bytes.
5599 unsigned Offset = SizeVal - BytesLeft;
5600 MVT DstVT = Dst.getValueType();
5601 MVT SrcVT = Src.getValueType();
5602 MVT SizeVT = Size.getValueType();
5603 Results.push_back(DAG.getMemcpy(Chain,
5604 DAG.getNode(ISD::ADD, DstVT, Dst,
5605 DAG.getConstant(Offset, DstVT)),
5606 DAG.getNode(ISD::ADD, SrcVT, Src,
5607 DAG.getConstant(Offset, SrcVT)),
5608 DAG.getConstant(BytesLeft, SizeVT),
5609 Align, AlwaysInline,
5610 DstSV, DstSVOff + Offset,
5611 SrcSV, SrcSVOff + Offset));
5614 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5617 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5618 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5620 if (!Subtarget->is64Bit()) {
5621 // vastart just stores the address of the VarArgsFrameIndex slot into the
5622 // memory location argument.
5623 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5624 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5628 // gp_offset (0 - 6 * 8)
5629 // fp_offset (48 - 48 + 8 * 16)
5630 // overflow_arg_area (point to parameters coming in memory).
5632 SmallVector<SDValue, 8> MemOps;
5633 SDValue FIN = Op.getOperand(1);
5635 SDValue Store = DAG.getStore(Op.getOperand(0),
5636 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5638 MemOps.push_back(Store);
5641 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5642 Store = DAG.getStore(Op.getOperand(0),
5643 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5645 MemOps.push_back(Store);
5647 // Store ptr to overflow_arg_area
5648 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5649 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5650 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5651 MemOps.push_back(Store);
5653 // Store ptr to reg_save_area.
5654 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5655 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5656 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5657 MemOps.push_back(Store);
5658 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5661 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5662 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5663 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5664 SDValue Chain = Op.getOperand(0);
5665 SDValue SrcPtr = Op.getOperand(1);
5666 SDValue SrcSV = Op.getOperand(2);
5668 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5673 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5674 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5675 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5676 SDValue Chain = Op.getOperand(0);
5677 SDValue DstPtr = Op.getOperand(1);
5678 SDValue SrcPtr = Op.getOperand(2);
5679 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5680 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5682 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5683 DAG.getIntPtrConstant(24), 8, false,
5684 DstSV, 0, SrcSV, 0);
5688 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5689 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5691 default: return SDValue(); // Don't custom lower most intrinsics.
5692 // Comparison intrinsics.
5693 case Intrinsic::x86_sse_comieq_ss:
5694 case Intrinsic::x86_sse_comilt_ss:
5695 case Intrinsic::x86_sse_comile_ss:
5696 case Intrinsic::x86_sse_comigt_ss:
5697 case Intrinsic::x86_sse_comige_ss:
5698 case Intrinsic::x86_sse_comineq_ss:
5699 case Intrinsic::x86_sse_ucomieq_ss:
5700 case Intrinsic::x86_sse_ucomilt_ss:
5701 case Intrinsic::x86_sse_ucomile_ss:
5702 case Intrinsic::x86_sse_ucomigt_ss:
5703 case Intrinsic::x86_sse_ucomige_ss:
5704 case Intrinsic::x86_sse_ucomineq_ss:
5705 case Intrinsic::x86_sse2_comieq_sd:
5706 case Intrinsic::x86_sse2_comilt_sd:
5707 case Intrinsic::x86_sse2_comile_sd:
5708 case Intrinsic::x86_sse2_comigt_sd:
5709 case Intrinsic::x86_sse2_comige_sd:
5710 case Intrinsic::x86_sse2_comineq_sd:
5711 case Intrinsic::x86_sse2_ucomieq_sd:
5712 case Intrinsic::x86_sse2_ucomilt_sd:
5713 case Intrinsic::x86_sse2_ucomile_sd:
5714 case Intrinsic::x86_sse2_ucomigt_sd:
5715 case Intrinsic::x86_sse2_ucomige_sd:
5716 case Intrinsic::x86_sse2_ucomineq_sd: {
5718 ISD::CondCode CC = ISD::SETCC_INVALID;
5721 case Intrinsic::x86_sse_comieq_ss:
5722 case Intrinsic::x86_sse2_comieq_sd:
5726 case Intrinsic::x86_sse_comilt_ss:
5727 case Intrinsic::x86_sse2_comilt_sd:
5731 case Intrinsic::x86_sse_comile_ss:
5732 case Intrinsic::x86_sse2_comile_sd:
5736 case Intrinsic::x86_sse_comigt_ss:
5737 case Intrinsic::x86_sse2_comigt_sd:
5741 case Intrinsic::x86_sse_comige_ss:
5742 case Intrinsic::x86_sse2_comige_sd:
5746 case Intrinsic::x86_sse_comineq_ss:
5747 case Intrinsic::x86_sse2_comineq_sd:
5751 case Intrinsic::x86_sse_ucomieq_ss:
5752 case Intrinsic::x86_sse2_ucomieq_sd:
5753 Opc = X86ISD::UCOMI;
5756 case Intrinsic::x86_sse_ucomilt_ss:
5757 case Intrinsic::x86_sse2_ucomilt_sd:
5758 Opc = X86ISD::UCOMI;
5761 case Intrinsic::x86_sse_ucomile_ss:
5762 case Intrinsic::x86_sse2_ucomile_sd:
5763 Opc = X86ISD::UCOMI;
5766 case Intrinsic::x86_sse_ucomigt_ss:
5767 case Intrinsic::x86_sse2_ucomigt_sd:
5768 Opc = X86ISD::UCOMI;
5771 case Intrinsic::x86_sse_ucomige_ss:
5772 case Intrinsic::x86_sse2_ucomige_sd:
5773 Opc = X86ISD::UCOMI;
5776 case Intrinsic::x86_sse_ucomineq_ss:
5777 case Intrinsic::x86_sse2_ucomineq_sd:
5778 Opc = X86ISD::UCOMI;
5783 SDValue LHS = Op.getOperand(1);
5784 SDValue RHS = Op.getOperand(2);
5785 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
5786 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5787 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5788 DAG.getConstant(X86CC, MVT::i8), Cond);
5789 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5792 // Fix vector shift instructions where the last operand is a non-immediate
5794 case Intrinsic::x86_sse2_pslli_w:
5795 case Intrinsic::x86_sse2_pslli_d:
5796 case Intrinsic::x86_sse2_pslli_q:
5797 case Intrinsic::x86_sse2_psrli_w:
5798 case Intrinsic::x86_sse2_psrli_d:
5799 case Intrinsic::x86_sse2_psrli_q:
5800 case Intrinsic::x86_sse2_psrai_w:
5801 case Intrinsic::x86_sse2_psrai_d:
5802 case Intrinsic::x86_mmx_pslli_w:
5803 case Intrinsic::x86_mmx_pslli_d:
5804 case Intrinsic::x86_mmx_pslli_q:
5805 case Intrinsic::x86_mmx_psrli_w:
5806 case Intrinsic::x86_mmx_psrli_d:
5807 case Intrinsic::x86_mmx_psrli_q:
5808 case Intrinsic::x86_mmx_psrai_w:
5809 case Intrinsic::x86_mmx_psrai_d: {
5810 SDValue ShAmt = Op.getOperand(2);
5811 if (isa<ConstantSDNode>(ShAmt))
5814 unsigned NewIntNo = 0;
5815 MVT ShAmtVT = MVT::v4i32;
5817 case Intrinsic::x86_sse2_pslli_w:
5818 NewIntNo = Intrinsic::x86_sse2_psll_w;
5820 case Intrinsic::x86_sse2_pslli_d:
5821 NewIntNo = Intrinsic::x86_sse2_psll_d;
5823 case Intrinsic::x86_sse2_pslli_q:
5824 NewIntNo = Intrinsic::x86_sse2_psll_q;
5826 case Intrinsic::x86_sse2_psrli_w:
5827 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5829 case Intrinsic::x86_sse2_psrli_d:
5830 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5832 case Intrinsic::x86_sse2_psrli_q:
5833 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5835 case Intrinsic::x86_sse2_psrai_w:
5836 NewIntNo = Intrinsic::x86_sse2_psra_w;
5838 case Intrinsic::x86_sse2_psrai_d:
5839 NewIntNo = Intrinsic::x86_sse2_psra_d;
5842 ShAmtVT = MVT::v2i32;
5844 case Intrinsic::x86_mmx_pslli_w:
5845 NewIntNo = Intrinsic::x86_mmx_psll_w;
5847 case Intrinsic::x86_mmx_pslli_d:
5848 NewIntNo = Intrinsic::x86_mmx_psll_d;
5850 case Intrinsic::x86_mmx_pslli_q:
5851 NewIntNo = Intrinsic::x86_mmx_psll_q;
5853 case Intrinsic::x86_mmx_psrli_w:
5854 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5856 case Intrinsic::x86_mmx_psrli_d:
5857 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5859 case Intrinsic::x86_mmx_psrli_q:
5860 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5862 case Intrinsic::x86_mmx_psrai_w:
5863 NewIntNo = Intrinsic::x86_mmx_psra_w;
5865 case Intrinsic::x86_mmx_psrai_d:
5866 NewIntNo = Intrinsic::x86_mmx_psra_d;
5868 default: abort(); // Can't reach here.
5873 MVT VT = Op.getValueType();
5874 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5875 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5876 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5877 DAG.getConstant(NewIntNo, MVT::i32),
5878 Op.getOperand(1), ShAmt);
5883 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5884 // Depths > 0 not supported yet!
5885 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5888 // Just load the return address
5889 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5890 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5893 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5894 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5895 MFI->setFrameAddressIsTaken(true);
5896 MVT VT = Op.getValueType();
5897 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5898 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
5899 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), FrameReg, VT);
5901 FrameAddr = DAG.getLoad(VT, DAG.getEntryNode(), FrameAddr, NULL, 0);
5905 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5906 SelectionDAG &DAG) {
5907 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5910 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5912 MachineFunction &MF = DAG.getMachineFunction();
5913 SDValue Chain = Op.getOperand(0);
5914 SDValue Offset = Op.getOperand(1);
5915 SDValue Handler = Op.getOperand(2);
5917 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5919 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5921 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5922 DAG.getIntPtrConstant(-TD->getPointerSize()));
5923 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5924 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5925 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5926 MF.getRegInfo().addLiveOut(StoreAddrReg);
5928 return DAG.getNode(X86ISD::EH_RETURN,
5930 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5933 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5934 SelectionDAG &DAG) {
5935 SDValue Root = Op.getOperand(0);
5936 SDValue Trmp = Op.getOperand(1); // trampoline
5937 SDValue FPtr = Op.getOperand(2); // nested function
5938 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5940 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5942 const X86InstrInfo *TII =
5943 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5945 if (Subtarget->is64Bit()) {
5946 SDValue OutChains[6];
5948 // Large code-model.
5950 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5951 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5953 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5954 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5956 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5958 // Load the pointer to the nested function into R11.
5959 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5960 SDValue Addr = Trmp;
5961 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5964 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5965 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5967 // Load the 'nest' parameter value into R10.
5968 // R10 is specified in X86CallingConv.td
5969 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5970 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5971 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5974 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5975 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5977 // Jump to the nested function.
5978 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5979 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5980 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5983 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5984 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5985 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5989 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5990 return DAG.getMergeValues(Ops, 2);
5992 const Function *Func =
5993 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5994 unsigned CC = Func->getCallingConv();
5999 assert(0 && "Unsupported calling convention");
6000 case CallingConv::C:
6001 case CallingConv::X86_StdCall: {
6002 // Pass 'nest' parameter in ECX.
6003 // Must be kept in sync with X86CallingConv.td
6006 // Check that ECX wasn't needed by an 'inreg' parameter.
6007 const FunctionType *FTy = Func->getFunctionType();
6008 const AttrListPtr &Attrs = Func->getAttributes();
6010 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6011 unsigned InRegCount = 0;
6014 for (FunctionType::param_iterator I = FTy->param_begin(),
6015 E = FTy->param_end(); I != E; ++I, ++Idx)
6016 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6017 // FIXME: should only count parameters that are lowered to integers.
6018 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6020 if (InRegCount > 2) {
6021 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6027 case CallingConv::X86_FastCall:
6028 case CallingConv::Fast:
6029 // Pass 'nest' parameter in EAX.
6030 // Must be kept in sync with X86CallingConv.td
6035 SDValue OutChains[4];
6038 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
6039 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
6041 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6042 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6043 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6046 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
6047 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
6049 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6050 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
6051 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
6052 TrmpAddr, 5, false, 1);
6054 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
6055 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
6058 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
6059 return DAG.getMergeValues(Ops, 2);
6063 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6065 The rounding mode is in bits 11:10 of FPSR, and has the following
6072 FLT_ROUNDS, on the other hand, expects the following:
6079 To perform the conversion, we do:
6080 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6083 MachineFunction &MF = DAG.getMachineFunction();
6084 const TargetMachine &TM = MF.getTarget();
6085 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6086 unsigned StackAlignment = TFI.getStackAlignment();
6087 MVT VT = Op.getValueType();
6089 // Save FP Control Word to stack slot
6090 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6091 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6093 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
6094 DAG.getEntryNode(), StackSlot);
6096 // Load FP Control Word from stack slot
6097 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
6099 // Transform as necessary
6101 DAG.getNode(ISD::SRL, MVT::i16,
6102 DAG.getNode(ISD::AND, MVT::i16,
6103 CWD, DAG.getConstant(0x800, MVT::i16)),
6104 DAG.getConstant(11, MVT::i8));
6106 DAG.getNode(ISD::SRL, MVT::i16,
6107 DAG.getNode(ISD::AND, MVT::i16,
6108 CWD, DAG.getConstant(0x400, MVT::i16)),
6109 DAG.getConstant(9, MVT::i8));
6112 DAG.getNode(ISD::AND, MVT::i16,
6113 DAG.getNode(ISD::ADD, MVT::i16,
6114 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
6115 DAG.getConstant(1, MVT::i16)),
6116 DAG.getConstant(3, MVT::i16));
6119 return DAG.getNode((VT.getSizeInBits() < 16 ?
6120 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
6123 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6124 MVT VT = Op.getValueType();
6126 unsigned NumBits = VT.getSizeInBits();
6128 Op = Op.getOperand(0);
6129 if (VT == MVT::i8) {
6130 // Zero extend to i32 since there is not an i8 bsr.
6132 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
6135 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6136 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6137 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
6139 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6140 SmallVector<SDValue, 4> Ops;
6142 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6143 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6144 Ops.push_back(Op.getValue(1));
6145 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
6147 // Finally xor with NumBits-1.
6148 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6151 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
6155 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6156 MVT VT = Op.getValueType();
6158 unsigned NumBits = VT.getSizeInBits();
6160 Op = Op.getOperand(0);
6161 if (VT == MVT::i8) {
6163 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
6166 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6167 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6168 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
6170 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6171 SmallVector<SDValue, 4> Ops;
6173 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6174 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6175 Ops.push_back(Op.getValue(1));
6176 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
6179 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
6183 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6184 MVT VT = Op.getValueType();
6185 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6187 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6188 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6189 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6190 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6191 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6193 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6194 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6195 // return AloBlo + AloBhi + AhiBlo;
6197 SDValue A = Op.getOperand(0);
6198 SDValue B = Op.getOperand(1);
6200 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6201 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6202 A, DAG.getConstant(32, MVT::i32));
6203 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6204 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6205 B, DAG.getConstant(32, MVT::i32));
6206 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6207 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6209 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6210 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6212 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6213 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6215 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6216 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6217 AloBhi, DAG.getConstant(32, MVT::i32));
6218 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6219 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6220 AhiBlo, DAG.getConstant(32, MVT::i32));
6221 SDValue Res = DAG.getNode(ISD::ADD, VT, AloBlo, AloBhi);
6222 Res = DAG.getNode(ISD::ADD, VT, Res, AhiBlo);
6227 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6228 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6229 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6230 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6231 // has only one use.
6232 SDNode *N = Op.getNode();
6233 SDValue LHS = N->getOperand(0);
6234 SDValue RHS = N->getOperand(1);
6235 unsigned BaseOp = 0;
6238 switch (Op.getOpcode()) {
6239 default: assert(0 && "Unknown ovf instruction!");
6241 BaseOp = X86ISD::ADD;
6245 BaseOp = X86ISD::ADD;
6249 BaseOp = X86ISD::SUB;
6253 BaseOp = X86ISD::SUB;
6257 BaseOp = X86ISD::SMUL;
6261 BaseOp = X86ISD::UMUL;
6266 // Also sets EFLAGS.
6267 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6268 SDValue Sum = DAG.getNode(BaseOp, VTs, LHS, RHS);
6271 DAG.getNode(X86ISD::SETCC, N->getValueType(1),
6272 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6274 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6278 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6279 MVT T = Op.getValueType();
6282 switch(T.getSimpleVT()) {
6284 assert(false && "Invalid value type!");
6285 case MVT::i8: Reg = X86::AL; size = 1; break;
6286 case MVT::i16: Reg = X86::AX; size = 2; break;
6287 case MVT::i32: Reg = X86::EAX; size = 4; break;
6289 assert(Subtarget->is64Bit() && "Node not type legal!");
6290 Reg = X86::RAX; size = 8;
6293 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
6294 Op.getOperand(2), SDValue());
6295 SDValue Ops[] = { cpIn.getValue(0),
6298 DAG.getTargetConstant(size, MVT::i8),
6300 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6301 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
6303 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
6307 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6308 SelectionDAG &DAG) {
6309 assert(Subtarget->is64Bit() && "Result not type legalized?");
6310 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6311 SDValue TheChain = Op.getOperand(0);
6312 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
6313 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
6314 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX, MVT::i64,
6316 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
6317 DAG.getConstant(32, MVT::i8));
6319 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp),
6322 return DAG.getMergeValues(Ops, 2);
6325 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6326 SDNode *Node = Op.getNode();
6327 MVT T = Node->getValueType(0);
6328 SDValue negOp = DAG.getNode(ISD::SUB, T,
6329 DAG.getConstant(0, T), Node->getOperand(2));
6330 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD,
6331 cast<AtomicSDNode>(Node)->getMemoryVT(),
6332 Node->getOperand(0),
6333 Node->getOperand(1), negOp,
6334 cast<AtomicSDNode>(Node)->getSrcValue(),
6335 cast<AtomicSDNode>(Node)->getAlignment());
6338 /// LowerOperation - Provide custom lowering hooks for some operations.
6340 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6341 switch (Op.getOpcode()) {
6342 default: assert(0 && "Should not custom lower this!");
6343 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6344 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6345 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6346 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6347 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6348 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6349 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6350 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6351 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6352 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6353 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6354 case ISD::SHL_PARTS:
6355 case ISD::SRA_PARTS:
6356 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6357 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6358 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6359 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6360 case ISD::FABS: return LowerFABS(Op, DAG);
6361 case ISD::FNEG: return LowerFNEG(Op, DAG);
6362 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6363 case ISD::SETCC: return LowerSETCC(Op, DAG);
6364 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6365 case ISD::SELECT: return LowerSELECT(Op, DAG);
6366 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6367 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6368 case ISD::CALL: return LowerCALL(Op, DAG);
6369 case ISD::RET: return LowerRET(Op, DAG);
6370 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6371 case ISD::VASTART: return LowerVASTART(Op, DAG);
6372 case ISD::VAARG: return LowerVAARG(Op, DAG);
6373 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6374 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6375 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6376 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6377 case ISD::FRAME_TO_ARGS_OFFSET:
6378 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6379 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6380 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6381 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6382 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6383 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6384 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6385 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6391 case ISD::UMULO: return LowerXALUO(Op, DAG);
6392 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6396 void X86TargetLowering::
6397 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6398 SelectionDAG &DAG, unsigned NewOp) {
6399 MVT T = Node->getValueType(0);
6400 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6402 SDValue Chain = Node->getOperand(0);
6403 SDValue In1 = Node->getOperand(1);
6404 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
6405 Node->getOperand(2), DAG.getIntPtrConstant(0));
6406 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
6407 Node->getOperand(2), DAG.getIntPtrConstant(1));
6408 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6409 // have a MemOperand. Pass the info through as a normal operand.
6410 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6411 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6412 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6413 SDValue Result = DAG.getNode(NewOp, Tys, Ops, 5);
6414 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6415 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2));
6416 Results.push_back(Result.getValue(2));
6419 /// ReplaceNodeResults - Replace a node with an illegal result type
6420 /// with a new node built out of custom code.
6421 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6422 SmallVectorImpl<SDValue>&Results,
6423 SelectionDAG &DAG) {
6424 switch (N->getOpcode()) {
6426 assert(false && "Do not know how to custom type legalize this operation!");
6428 case ISD::FP_TO_SINT: {
6429 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
6430 SDValue FIST = Vals.first, StackSlot = Vals.second;
6431 if (FIST.getNode() != 0) {
6432 MVT VT = N->getValueType(0);
6433 // Return a load from the stack slot.
6434 Results.push_back(DAG.getLoad(VT, FIST, StackSlot, NULL, 0));
6438 case ISD::READCYCLECOUNTER: {
6439 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6440 SDValue TheChain = N->getOperand(0);
6441 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
6442 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
6443 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX, MVT::i32,
6445 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6446 SDValue Ops[] = { eax, edx };
6447 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2));
6448 Results.push_back(edx.getValue(1));
6451 case ISD::ATOMIC_CMP_SWAP: {
6452 MVT T = N->getValueType(0);
6453 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6454 SDValue cpInL, cpInH;
6455 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(2),
6456 DAG.getConstant(0, MVT::i32));
6457 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(2),
6458 DAG.getConstant(1, MVT::i32));
6459 cpInL = DAG.getCopyToReg(N->getOperand(0), X86::EAX, cpInL, SDValue());
6460 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX, cpInH,
6462 SDValue swapInL, swapInH;
6463 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(3),
6464 DAG.getConstant(0, MVT::i32));
6465 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(3),
6466 DAG.getConstant(1, MVT::i32));
6467 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX, swapInL,
6469 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX, swapInH,
6470 swapInL.getValue(1));
6471 SDValue Ops[] = { swapInH.getValue(0),
6473 swapInH.getValue(1) };
6474 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6475 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
6476 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
6477 Result.getValue(1));
6478 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
6479 cpOutL.getValue(2));
6480 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6481 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2));
6482 Results.push_back(cpOutH.getValue(1));
6485 case ISD::ATOMIC_LOAD_ADD:
6486 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6488 case ISD::ATOMIC_LOAD_AND:
6489 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6491 case ISD::ATOMIC_LOAD_NAND:
6492 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6494 case ISD::ATOMIC_LOAD_OR:
6495 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6497 case ISD::ATOMIC_LOAD_SUB:
6498 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6500 case ISD::ATOMIC_LOAD_XOR:
6501 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6503 case ISD::ATOMIC_SWAP:
6504 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6509 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6511 default: return NULL;
6512 case X86ISD::BSF: return "X86ISD::BSF";
6513 case X86ISD::BSR: return "X86ISD::BSR";
6514 case X86ISD::SHLD: return "X86ISD::SHLD";
6515 case X86ISD::SHRD: return "X86ISD::SHRD";
6516 case X86ISD::FAND: return "X86ISD::FAND";
6517 case X86ISD::FOR: return "X86ISD::FOR";
6518 case X86ISD::FXOR: return "X86ISD::FXOR";
6519 case X86ISD::FSRL: return "X86ISD::FSRL";
6520 case X86ISD::FILD: return "X86ISD::FILD";
6521 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6522 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6523 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6524 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6525 case X86ISD::FLD: return "X86ISD::FLD";
6526 case X86ISD::FST: return "X86ISD::FST";
6527 case X86ISD::CALL: return "X86ISD::CALL";
6528 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6529 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6530 case X86ISD::BT: return "X86ISD::BT";
6531 case X86ISD::CMP: return "X86ISD::CMP";
6532 case X86ISD::COMI: return "X86ISD::COMI";
6533 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6534 case X86ISD::SETCC: return "X86ISD::SETCC";
6535 case X86ISD::CMOV: return "X86ISD::CMOV";
6536 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6537 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6538 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6539 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6540 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6541 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6542 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6543 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6544 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6545 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6546 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6547 case X86ISD::FMAX: return "X86ISD::FMAX";
6548 case X86ISD::FMIN: return "X86ISD::FMIN";
6549 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6550 case X86ISD::FRCP: return "X86ISD::FRCP";
6551 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6552 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6553 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6554 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6555 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6556 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6557 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6558 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6559 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6560 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6561 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6562 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6563 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6564 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6565 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6566 case X86ISD::VSHL: return "X86ISD::VSHL";
6567 case X86ISD::VSRL: return "X86ISD::VSRL";
6568 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6569 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6570 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6571 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6572 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6573 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6574 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6575 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6576 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6577 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6578 case X86ISD::ADD: return "X86ISD::ADD";
6579 case X86ISD::SUB: return "X86ISD::SUB";
6580 case X86ISD::SMUL: return "X86ISD::SMUL";
6581 case X86ISD::UMUL: return "X86ISD::UMUL";
6585 // isLegalAddressingMode - Return true if the addressing mode represented
6586 // by AM is legal for this target, for a load/store of the specified type.
6587 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6588 const Type *Ty) const {
6589 // X86 supports extremely general addressing modes.
6591 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6592 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6596 // We can only fold this if we don't need an extra load.
6597 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6599 // If BaseGV requires a register, we cannot also have a BaseReg.
6600 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6604 // X86-64 only supports addr of globals in small code model.
6605 if (Subtarget->is64Bit()) {
6606 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6608 // If lower 4G is not available, then we must use rip-relative addressing.
6609 if (AM.BaseOffs || AM.Scale > 1)
6620 // These scales always work.
6625 // These scales are formed with basereg+scalereg. Only accept if there is
6630 default: // Other stuff never works.
6638 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6639 if (!Ty1->isInteger() || !Ty2->isInteger())
6641 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6642 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6643 if (NumBits1 <= NumBits2)
6645 return Subtarget->is64Bit() || NumBits1 < 64;
6648 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6649 if (!VT1.isInteger() || !VT2.isInteger())
6651 unsigned NumBits1 = VT1.getSizeInBits();
6652 unsigned NumBits2 = VT2.getSizeInBits();
6653 if (NumBits1 <= NumBits2)
6655 return Subtarget->is64Bit() || NumBits1 < 64;
6658 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6659 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6660 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6661 /// are assumed to be legal.
6663 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6664 // Only do shuffles on 128-bit vector types for now.
6665 if (VT.getSizeInBits() == 64) return false;
6666 return (Mask.getNode()->getNumOperands() <= 4 ||
6667 isIdentityMask(Mask.getNode()) ||
6668 isIdentityMask(Mask.getNode(), true) ||
6669 isSplatMask(Mask.getNode()) ||
6670 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6671 X86::isUNPCKLMask(Mask.getNode()) ||
6672 X86::isUNPCKHMask(Mask.getNode()) ||
6673 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6674 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6678 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6679 MVT EVT, SelectionDAG &DAG) const {
6680 unsigned NumElts = BVOps.size();
6681 // Only do shuffles on 128-bit vector types for now.
6682 if (EVT.getSizeInBits() * NumElts == 64) return false;
6683 if (NumElts == 2) return true;
6685 return (isMOVLMask(&BVOps[0], 4) ||
6686 isCommutedMOVL(&BVOps[0], 4, true) ||
6687 isSHUFPMask(&BVOps[0], 4) ||
6688 isCommutedSHUFP(&BVOps[0], 4));
6693 //===----------------------------------------------------------------------===//
6694 // X86 Scheduler Hooks
6695 //===----------------------------------------------------------------------===//
6697 // private utility function
6699 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6700 MachineBasicBlock *MBB,
6708 TargetRegisterClass *RC,
6710 // For the atomic bitwise operator, we generate
6713 // ld t1 = [bitinstr.addr]
6714 // op t2 = t1, [bitinstr.val]
6716 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6718 // fallthrough -->nextMBB
6719 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6720 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6721 MachineFunction::iterator MBBIter = MBB;
6724 /// First build the CFG
6725 MachineFunction *F = MBB->getParent();
6726 MachineBasicBlock *thisMBB = MBB;
6727 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6728 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6729 F->insert(MBBIter, newMBB);
6730 F->insert(MBBIter, nextMBB);
6732 // Move all successors to thisMBB to nextMBB
6733 nextMBB->transferSuccessors(thisMBB);
6735 // Update thisMBB to fall through to newMBB
6736 thisMBB->addSuccessor(newMBB);
6738 // newMBB jumps to itself and fall through to nextMBB
6739 newMBB->addSuccessor(nextMBB);
6740 newMBB->addSuccessor(newMBB);
6742 // Insert instructions into newMBB based on incoming instruction
6743 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6744 MachineOperand& destOper = bInstr->getOperand(0);
6745 MachineOperand* argOpers[6];
6746 int numArgs = bInstr->getNumOperands() - 1;
6747 for (int i=0; i < numArgs; ++i)
6748 argOpers[i] = &bInstr->getOperand(i+1);
6750 // x86 address has 4 operands: base, index, scale, and displacement
6751 int lastAddrIndx = 3; // [0,3]
6754 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6755 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6756 for (int i=0; i <= lastAddrIndx; ++i)
6757 (*MIB).addOperand(*argOpers[i]);
6759 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6761 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6766 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6767 assert((argOpers[valArgIndx]->isReg() ||
6768 argOpers[valArgIndx]->isImm()) &&
6770 if (argOpers[valArgIndx]->isReg())
6771 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6773 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6775 (*MIB).addOperand(*argOpers[valArgIndx]);
6777 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6780 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6781 for (int i=0; i <= lastAddrIndx; ++i)
6782 (*MIB).addOperand(*argOpers[i]);
6784 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6785 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6787 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6791 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6793 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6797 // private utility function: 64 bit atomics on 32 bit host.
6799 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
6800 MachineBasicBlock *MBB,
6806 // For the atomic bitwise operator, we generate
6807 // thisMBB (instructions are in pairs, except cmpxchg8b)
6808 // ld t1,t2 = [bitinstr.addr]
6810 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
6811 // op t5, t6 <- out1, out2, [bitinstr.val]
6812 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
6813 // mov ECX, EBX <- t5, t6
6814 // mov EAX, EDX <- t1, t2
6815 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
6816 // mov t3, t4 <- EAX, EDX
6818 // result in out1, out2
6819 // fallthrough -->nextMBB
6821 const TargetRegisterClass *RC = X86::GR32RegisterClass;
6822 const unsigned LoadOpc = X86::MOV32rm;
6823 const unsigned copyOpc = X86::MOV32rr;
6824 const unsigned NotOpc = X86::NOT32r;
6825 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6826 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6827 MachineFunction::iterator MBBIter = MBB;
6830 /// First build the CFG
6831 MachineFunction *F = MBB->getParent();
6832 MachineBasicBlock *thisMBB = MBB;
6833 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6834 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6835 F->insert(MBBIter, newMBB);
6836 F->insert(MBBIter, nextMBB);
6838 // Move all successors to thisMBB to nextMBB
6839 nextMBB->transferSuccessors(thisMBB);
6841 // Update thisMBB to fall through to newMBB
6842 thisMBB->addSuccessor(newMBB);
6844 // newMBB jumps to itself and fall through to nextMBB
6845 newMBB->addSuccessor(nextMBB);
6846 newMBB->addSuccessor(newMBB);
6848 // Insert instructions into newMBB based on incoming instruction
6849 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
6850 assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
6851 MachineOperand& dest1Oper = bInstr->getOperand(0);
6852 MachineOperand& dest2Oper = bInstr->getOperand(1);
6853 MachineOperand* argOpers[6];
6854 for (int i=0; i < 6; ++i)
6855 argOpers[i] = &bInstr->getOperand(i+2);
6857 // x86 address has 4 operands: base, index, scale, and displacement
6858 int lastAddrIndx = 3; // [0,3]
6860 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6861 MachineInstrBuilder MIB = BuildMI(thisMBB, TII->get(LoadOpc), t1);
6862 for (int i=0; i <= lastAddrIndx; ++i)
6863 (*MIB).addOperand(*argOpers[i]);
6864 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6865 MIB = BuildMI(thisMBB, TII->get(LoadOpc), t2);
6866 // add 4 to displacement.
6867 for (int i=0; i <= lastAddrIndx-1; ++i)
6868 (*MIB).addOperand(*argOpers[i]);
6869 MachineOperand newOp3 = *(argOpers[3]);
6871 newOp3.setImm(newOp3.getImm()+4);
6873 newOp3.setOffset(newOp3.getOffset()+4);
6874 (*MIB).addOperand(newOp3);
6876 // t3/4 are defined later, at the bottom of the loop
6877 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
6878 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
6879 BuildMI(newMBB, TII->get(X86::PHI), dest1Oper.getReg())
6880 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
6881 BuildMI(newMBB, TII->get(X86::PHI), dest2Oper.getReg())
6882 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
6884 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
6885 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
6887 MIB = BuildMI(newMBB, TII->get(NotOpc), tt1).addReg(t1);
6888 MIB = BuildMI(newMBB, TII->get(NotOpc), tt2).addReg(t2);
6894 assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
6896 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
6897 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
6898 if (argOpers[4]->isReg())
6899 MIB = BuildMI(newMBB, TII->get(regOpcL), t5);
6901 MIB = BuildMI(newMBB, TII->get(immOpcL), t5);
6902 if (regOpcL != X86::MOV32rr)
6904 (*MIB).addOperand(*argOpers[4]);
6905 assert(argOpers[5]->isReg() == argOpers[4]->isReg());
6906 assert(argOpers[5]->isImm() == argOpers[4]->isImm());
6907 if (argOpers[5]->isReg())
6908 MIB = BuildMI(newMBB, TII->get(regOpcH), t6);
6910 MIB = BuildMI(newMBB, TII->get(immOpcH), t6);
6911 if (regOpcH != X86::MOV32rr)
6913 (*MIB).addOperand(*argOpers[5]);
6915 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EAX);
6917 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EDX);
6920 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EBX);
6922 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::ECX);
6925 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG8B));
6926 for (int i=0; i <= lastAddrIndx; ++i)
6927 (*MIB).addOperand(*argOpers[i]);
6929 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6930 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6932 MIB = BuildMI(newMBB, TII->get(copyOpc), t3);
6933 MIB.addReg(X86::EAX);
6934 MIB = BuildMI(newMBB, TII->get(copyOpc), t4);
6935 MIB.addReg(X86::EDX);
6938 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6940 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6944 // private utility function
6946 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6947 MachineBasicBlock *MBB,
6949 // For the atomic min/max operator, we generate
6952 // ld t1 = [min/max.addr]
6953 // mov t2 = [min/max.val]
6955 // cmov[cond] t2 = t1
6957 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6959 // fallthrough -->nextMBB
6961 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6962 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6963 MachineFunction::iterator MBBIter = MBB;
6966 /// First build the CFG
6967 MachineFunction *F = MBB->getParent();
6968 MachineBasicBlock *thisMBB = MBB;
6969 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6970 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6971 F->insert(MBBIter, newMBB);
6972 F->insert(MBBIter, nextMBB);
6974 // Move all successors to thisMBB to nextMBB
6975 nextMBB->transferSuccessors(thisMBB);
6977 // Update thisMBB to fall through to newMBB
6978 thisMBB->addSuccessor(newMBB);
6980 // newMBB jumps to newMBB and fall through to nextMBB
6981 newMBB->addSuccessor(nextMBB);
6982 newMBB->addSuccessor(newMBB);
6984 // Insert instructions into newMBB based on incoming instruction
6985 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6986 MachineOperand& destOper = mInstr->getOperand(0);
6987 MachineOperand* argOpers[6];
6988 int numArgs = mInstr->getNumOperands() - 1;
6989 for (int i=0; i < numArgs; ++i)
6990 argOpers[i] = &mInstr->getOperand(i+1);
6992 // x86 address has 4 operands: base, index, scale, and displacement
6993 int lastAddrIndx = 3; // [0,3]
6996 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6997 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6998 for (int i=0; i <= lastAddrIndx; ++i)
6999 (*MIB).addOperand(*argOpers[i]);
7001 // We only support register and immediate values
7002 assert((argOpers[valArgIndx]->isReg() ||
7003 argOpers[valArgIndx]->isImm()) &&
7006 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7007 if (argOpers[valArgIndx]->isReg())
7008 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
7010 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
7011 (*MIB).addOperand(*argOpers[valArgIndx]);
7013 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
7016 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
7021 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7022 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
7026 // Cmp and exchange if none has modified the memory location
7027 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
7028 for (int i=0; i <= lastAddrIndx; ++i)
7029 (*MIB).addOperand(*argOpers[i]);
7031 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7032 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7034 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
7035 MIB.addReg(X86::EAX);
7038 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
7040 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7046 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7047 MachineBasicBlock *BB) {
7048 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7049 switch (MI->getOpcode()) {
7050 default: assert(false && "Unexpected instr type to insert");
7051 case X86::CMOV_V1I64:
7052 case X86::CMOV_FR32:
7053 case X86::CMOV_FR64:
7054 case X86::CMOV_V4F32:
7055 case X86::CMOV_V2F64:
7056 case X86::CMOV_V2I64: {
7057 // To "insert" a SELECT_CC instruction, we actually have to insert the
7058 // diamond control-flow pattern. The incoming instruction knows the
7059 // destination vreg to set, the condition code register to branch on, the
7060 // true/false values to select between, and a branch opcode to use.
7061 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7062 MachineFunction::iterator It = BB;
7068 // cmpTY ccX, r1, r2
7070 // fallthrough --> copy0MBB
7071 MachineBasicBlock *thisMBB = BB;
7072 MachineFunction *F = BB->getParent();
7073 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7074 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7076 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7077 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
7078 F->insert(It, copy0MBB);
7079 F->insert(It, sinkMBB);
7080 // Update machine-CFG edges by transferring all successors of the current
7081 // block to the new block which will contain the Phi node for the select.
7082 sinkMBB->transferSuccessors(BB);
7084 // Add the true and fallthrough blocks as its successors.
7085 BB->addSuccessor(copy0MBB);
7086 BB->addSuccessor(sinkMBB);
7089 // %FalseValue = ...
7090 // # fallthrough to sinkMBB
7093 // Update machine-CFG edges
7094 BB->addSuccessor(sinkMBB);
7097 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7100 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
7101 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7102 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7104 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7108 case X86::FP32_TO_INT16_IN_MEM:
7109 case X86::FP32_TO_INT32_IN_MEM:
7110 case X86::FP32_TO_INT64_IN_MEM:
7111 case X86::FP64_TO_INT16_IN_MEM:
7112 case X86::FP64_TO_INT32_IN_MEM:
7113 case X86::FP64_TO_INT64_IN_MEM:
7114 case X86::FP80_TO_INT16_IN_MEM:
7115 case X86::FP80_TO_INT32_IN_MEM:
7116 case X86::FP80_TO_INT64_IN_MEM: {
7117 // Change the floating point control register to use "round towards zero"
7118 // mode when truncating to an integer value.
7119 MachineFunction *F = BB->getParent();
7120 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7121 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7123 // Load the old value of the high byte of the control word...
7125 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7126 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
7128 // Set the high part to be round to zero...
7129 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
7132 // Reload the modified control word now...
7133 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
7135 // Restore the memory image of control word to original value
7136 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
7139 // Get the X86 opcode to use.
7141 switch (MI->getOpcode()) {
7142 default: assert(0 && "illegal opcode!");
7143 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7144 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7145 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7146 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7147 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7148 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7149 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7150 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7151 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7155 MachineOperand &Op = MI->getOperand(0);
7157 AM.BaseType = X86AddressMode::RegBase;
7158 AM.Base.Reg = Op.getReg();
7160 AM.BaseType = X86AddressMode::FrameIndexBase;
7161 AM.Base.FrameIndex = Op.getIndex();
7163 Op = MI->getOperand(1);
7165 AM.Scale = Op.getImm();
7166 Op = MI->getOperand(2);
7168 AM.IndexReg = Op.getImm();
7169 Op = MI->getOperand(3);
7170 if (Op.isGlobal()) {
7171 AM.GV = Op.getGlobal();
7173 AM.Disp = Op.getImm();
7175 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
7176 .addReg(MI->getOperand(4).getReg());
7178 // Reload the original control word now.
7179 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
7181 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7184 case X86::ATOMAND32:
7185 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7186 X86::AND32ri, X86::MOV32rm,
7187 X86::LCMPXCHG32, X86::MOV32rr,
7188 X86::NOT32r, X86::EAX,
7189 X86::GR32RegisterClass);
7191 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7192 X86::OR32ri, X86::MOV32rm,
7193 X86::LCMPXCHG32, X86::MOV32rr,
7194 X86::NOT32r, X86::EAX,
7195 X86::GR32RegisterClass);
7196 case X86::ATOMXOR32:
7197 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7198 X86::XOR32ri, X86::MOV32rm,
7199 X86::LCMPXCHG32, X86::MOV32rr,
7200 X86::NOT32r, X86::EAX,
7201 X86::GR32RegisterClass);
7202 case X86::ATOMNAND32:
7203 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7204 X86::AND32ri, X86::MOV32rm,
7205 X86::LCMPXCHG32, X86::MOV32rr,
7206 X86::NOT32r, X86::EAX,
7207 X86::GR32RegisterClass, true);
7208 case X86::ATOMMIN32:
7209 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7210 case X86::ATOMMAX32:
7211 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7212 case X86::ATOMUMIN32:
7213 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7214 case X86::ATOMUMAX32:
7215 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7217 case X86::ATOMAND16:
7218 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7219 X86::AND16ri, X86::MOV16rm,
7220 X86::LCMPXCHG16, X86::MOV16rr,
7221 X86::NOT16r, X86::AX,
7222 X86::GR16RegisterClass);
7224 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7225 X86::OR16ri, X86::MOV16rm,
7226 X86::LCMPXCHG16, X86::MOV16rr,
7227 X86::NOT16r, X86::AX,
7228 X86::GR16RegisterClass);
7229 case X86::ATOMXOR16:
7230 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7231 X86::XOR16ri, X86::MOV16rm,
7232 X86::LCMPXCHG16, X86::MOV16rr,
7233 X86::NOT16r, X86::AX,
7234 X86::GR16RegisterClass);
7235 case X86::ATOMNAND16:
7236 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7237 X86::AND16ri, X86::MOV16rm,
7238 X86::LCMPXCHG16, X86::MOV16rr,
7239 X86::NOT16r, X86::AX,
7240 X86::GR16RegisterClass, true);
7241 case X86::ATOMMIN16:
7242 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7243 case X86::ATOMMAX16:
7244 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7245 case X86::ATOMUMIN16:
7246 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7247 case X86::ATOMUMAX16:
7248 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7251 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7252 X86::AND8ri, X86::MOV8rm,
7253 X86::LCMPXCHG8, X86::MOV8rr,
7254 X86::NOT8r, X86::AL,
7255 X86::GR8RegisterClass);
7257 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7258 X86::OR8ri, X86::MOV8rm,
7259 X86::LCMPXCHG8, X86::MOV8rr,
7260 X86::NOT8r, X86::AL,
7261 X86::GR8RegisterClass);
7263 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7264 X86::XOR8ri, X86::MOV8rm,
7265 X86::LCMPXCHG8, X86::MOV8rr,
7266 X86::NOT8r, X86::AL,
7267 X86::GR8RegisterClass);
7268 case X86::ATOMNAND8:
7269 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7270 X86::AND8ri, X86::MOV8rm,
7271 X86::LCMPXCHG8, X86::MOV8rr,
7272 X86::NOT8r, X86::AL,
7273 X86::GR8RegisterClass, true);
7274 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7275 // This group is for 64-bit host.
7276 case X86::ATOMAND64:
7277 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7278 X86::AND64ri32, X86::MOV64rm,
7279 X86::LCMPXCHG64, X86::MOV64rr,
7280 X86::NOT64r, X86::RAX,
7281 X86::GR64RegisterClass);
7283 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7284 X86::OR64ri32, X86::MOV64rm,
7285 X86::LCMPXCHG64, X86::MOV64rr,
7286 X86::NOT64r, X86::RAX,
7287 X86::GR64RegisterClass);
7288 case X86::ATOMXOR64:
7289 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7290 X86::XOR64ri32, X86::MOV64rm,
7291 X86::LCMPXCHG64, X86::MOV64rr,
7292 X86::NOT64r, X86::RAX,
7293 X86::GR64RegisterClass);
7294 case X86::ATOMNAND64:
7295 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7296 X86::AND64ri32, X86::MOV64rm,
7297 X86::LCMPXCHG64, X86::MOV64rr,
7298 X86::NOT64r, X86::RAX,
7299 X86::GR64RegisterClass, true);
7300 case X86::ATOMMIN64:
7301 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7302 case X86::ATOMMAX64:
7303 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7304 case X86::ATOMUMIN64:
7305 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7306 case X86::ATOMUMAX64:
7307 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7309 // This group does 64-bit operations on a 32-bit host.
7310 case X86::ATOMAND6432:
7311 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7312 X86::AND32rr, X86::AND32rr,
7313 X86::AND32ri, X86::AND32ri,
7315 case X86::ATOMOR6432:
7316 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7317 X86::OR32rr, X86::OR32rr,
7318 X86::OR32ri, X86::OR32ri,
7320 case X86::ATOMXOR6432:
7321 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7322 X86::XOR32rr, X86::XOR32rr,
7323 X86::XOR32ri, X86::XOR32ri,
7325 case X86::ATOMNAND6432:
7326 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7327 X86::AND32rr, X86::AND32rr,
7328 X86::AND32ri, X86::AND32ri,
7330 case X86::ATOMADD6432:
7331 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7332 X86::ADD32rr, X86::ADC32rr,
7333 X86::ADD32ri, X86::ADC32ri,
7335 case X86::ATOMSUB6432:
7336 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7337 X86::SUB32rr, X86::SBB32rr,
7338 X86::SUB32ri, X86::SBB32ri,
7340 case X86::ATOMSWAP6432:
7341 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7342 X86::MOV32rr, X86::MOV32rr,
7343 X86::MOV32ri, X86::MOV32ri,
7348 //===----------------------------------------------------------------------===//
7349 // X86 Optimization Hooks
7350 //===----------------------------------------------------------------------===//
7352 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7356 const SelectionDAG &DAG,
7357 unsigned Depth) const {
7358 unsigned Opc = Op.getOpcode();
7359 assert((Opc >= ISD::BUILTIN_OP_END ||
7360 Opc == ISD::INTRINSIC_WO_CHAIN ||
7361 Opc == ISD::INTRINSIC_W_CHAIN ||
7362 Opc == ISD::INTRINSIC_VOID) &&
7363 "Should use MaskedValueIsZero if you don't know whether Op"
7364 " is a target node!");
7366 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7370 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7371 Mask.getBitWidth() - 1);
7376 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7377 /// node is a GlobalAddress + offset.
7378 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7379 GlobalValue* &GA, int64_t &Offset) const{
7380 if (N->getOpcode() == X86ISD::Wrapper) {
7381 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7382 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7383 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7387 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7390 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7391 const TargetLowering &TLI) {
7394 if (TLI.isGAPlusOffset(Base, GV, Offset))
7395 return (GV->getAlignment() >= N && (Offset % N) == 0);
7396 // DAG combine handles the stack object case.
7400 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
7401 unsigned NumElems, MVT EVT,
7403 SelectionDAG &DAG, MachineFrameInfo *MFI,
7404 const TargetLowering &TLI) {
7406 for (unsigned i = 0; i < NumElems; ++i) {
7407 SDValue Idx = PermMask.getOperand(i);
7408 if (Idx.getOpcode() == ISD::UNDEF) {
7414 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7415 if (!Elt.getNode() ||
7416 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7419 Base = Elt.getNode();
7420 if (Base->getOpcode() == ISD::UNDEF)
7424 if (Elt.getOpcode() == ISD::UNDEF)
7427 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7428 EVT.getSizeInBits()/8, i, MFI))
7434 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7435 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7436 /// if the load addresses are consecutive, non-overlapping, and in the right
7438 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7439 const TargetLowering &TLI) {
7440 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7441 MVT VT = N->getValueType(0);
7442 MVT EVT = VT.getVectorElementType();
7443 SDValue PermMask = N->getOperand(2);
7444 unsigned NumElems = PermMask.getNumOperands();
7445 SDNode *Base = NULL;
7446 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
7450 LoadSDNode *LD = cast<LoadSDNode>(Base);
7451 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7452 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7453 LD->getSrcValueOffset(), LD->isVolatile());
7454 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7455 LD->getSrcValueOffset(), LD->isVolatile(),
7456 LD->getAlignment());
7459 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7460 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7461 const X86Subtarget *Subtarget,
7462 const TargetLowering &TLI) {
7463 unsigned NumOps = N->getNumOperands();
7465 // Ignore single operand BUILD_VECTOR.
7469 MVT VT = N->getValueType(0);
7470 MVT EVT = VT.getVectorElementType();
7471 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7472 // We are looking for load i64 and zero extend. We want to transform
7473 // it before legalizer has a chance to expand it. Also look for i64
7474 // BUILD_PAIR bit casted to f64.
7476 // This must be an insertion into a zero vector.
7477 SDValue HighElt = N->getOperand(1);
7478 if (!isZeroNode(HighElt))
7481 // Value must be a load.
7482 SDNode *Base = N->getOperand(0).getNode();
7483 if (!isa<LoadSDNode>(Base)) {
7484 if (Base->getOpcode() != ISD::BIT_CONVERT)
7486 Base = Base->getOperand(0).getNode();
7487 if (!isa<LoadSDNode>(Base))
7491 // Transform it into VZEXT_LOAD addr.
7492 LoadSDNode *LD = cast<LoadSDNode>(Base);
7494 // Load must not be an extload.
7495 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7498 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7499 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7500 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, Tys, Ops, 2);
7501 DAG.ReplaceAllUsesOfValueWith(SDValue(Base, 1), ResNode.getValue(1));
7505 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7506 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7507 const X86Subtarget *Subtarget) {
7508 SDValue Cond = N->getOperand(0);
7510 // If we have SSE[12] support, try to form min/max nodes.
7511 if (Subtarget->hasSSE2() &&
7512 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
7513 if (Cond.getOpcode() == ISD::SETCC) {
7514 // Get the LHS/RHS of the select.
7515 SDValue LHS = N->getOperand(1);
7516 SDValue RHS = N->getOperand(2);
7517 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7519 unsigned Opcode = 0;
7520 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7523 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7526 if (!UnsafeFPMath) break;
7528 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7530 Opcode = X86ISD::FMIN;
7533 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7536 if (!UnsafeFPMath) break;
7538 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7540 Opcode = X86ISD::FMAX;
7543 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7546 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7549 if (!UnsafeFPMath) break;
7551 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7553 Opcode = X86ISD::FMIN;
7556 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7559 if (!UnsafeFPMath) break;
7561 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7563 Opcode = X86ISD::FMAX;
7569 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
7577 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
7578 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
7579 const X86Subtarget *Subtarget) {
7580 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
7581 // the FP state in cases where an emms may be missing.
7582 // A preferable solution to the general problem is to figure out the right
7583 // places to insert EMMS. This qualifies as a quick hack.
7584 StoreSDNode *St = cast<StoreSDNode>(N);
7585 if (St->getValue().getValueType().isVector() &&
7586 St->getValue().getValueType().getSizeInBits() == 64 &&
7587 isa<LoadSDNode>(St->getValue()) &&
7588 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
7589 St->getChain().hasOneUse() && !St->isVolatile()) {
7590 SDNode* LdVal = St->getValue().getNode();
7592 int TokenFactorIndex = -1;
7593 SmallVector<SDValue, 8> Ops;
7594 SDNode* ChainVal = St->getChain().getNode();
7595 // Must be a store of a load. We currently handle two cases: the load
7596 // is a direct child, and it's under an intervening TokenFactor. It is
7597 // possible to dig deeper under nested TokenFactors.
7598 if (ChainVal == LdVal)
7599 Ld = cast<LoadSDNode>(St->getChain());
7600 else if (St->getValue().hasOneUse() &&
7601 ChainVal->getOpcode() == ISD::TokenFactor) {
7602 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
7603 if (ChainVal->getOperand(i).getNode() == LdVal) {
7604 TokenFactorIndex = i;
7605 Ld = cast<LoadSDNode>(St->getValue());
7607 Ops.push_back(ChainVal->getOperand(i));
7611 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
7612 if (Subtarget->is64Bit()) {
7613 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
7614 Ld->getBasePtr(), Ld->getSrcValue(),
7615 Ld->getSrcValueOffset(), Ld->isVolatile(),
7616 Ld->getAlignment());
7617 SDValue NewChain = NewLd.getValue(1);
7618 if (TokenFactorIndex != -1) {
7619 Ops.push_back(NewChain);
7620 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7623 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
7624 St->getSrcValue(), St->getSrcValueOffset(),
7625 St->isVolatile(), St->getAlignment());
7628 // Otherwise, lower to two 32-bit copies.
7629 SDValue LoAddr = Ld->getBasePtr();
7630 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7631 DAG.getConstant(4, MVT::i32));
7633 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
7634 Ld->getSrcValue(), Ld->getSrcValueOffset(),
7635 Ld->isVolatile(), Ld->getAlignment());
7636 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
7637 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
7639 MinAlign(Ld->getAlignment(), 4));
7641 SDValue NewChain = LoLd.getValue(1);
7642 if (TokenFactorIndex != -1) {
7643 Ops.push_back(LoLd);
7644 Ops.push_back(HiLd);
7645 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7649 LoAddr = St->getBasePtr();
7650 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7651 DAG.getConstant(4, MVT::i32));
7653 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
7654 St->getSrcValue(), St->getSrcValueOffset(),
7655 St->isVolatile(), St->getAlignment());
7656 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
7658 St->getSrcValueOffset() + 4,
7660 MinAlign(St->getAlignment(), 4));
7661 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
7667 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
7668 /// X86ISD::FXOR nodes.
7669 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
7670 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
7671 // F[X]OR(0.0, x) -> x
7672 // F[X]OR(x, 0.0) -> x
7673 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7674 if (C->getValueAPF().isPosZero())
7675 return N->getOperand(1);
7676 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7677 if (C->getValueAPF().isPosZero())
7678 return N->getOperand(0);
7682 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
7683 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
7684 // FAND(0.0, x) -> 0.0
7685 // FAND(x, 0.0) -> 0.0
7686 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7687 if (C->getValueAPF().isPosZero())
7688 return N->getOperand(0);
7689 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7690 if (C->getValueAPF().isPosZero())
7691 return N->getOperand(1);
7696 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7697 DAGCombinerInfo &DCI) const {
7698 SelectionDAG &DAG = DCI.DAG;
7699 switch (N->getOpcode()) {
7701 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7702 case ISD::BUILD_VECTOR:
7703 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7704 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7705 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7707 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7708 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7714 //===----------------------------------------------------------------------===//
7715 // X86 Inline Assembly Support
7716 //===----------------------------------------------------------------------===//
7718 /// getConstraintType - Given a constraint letter, return the type of
7719 /// constraint it is for this target.
7720 X86TargetLowering::ConstraintType
7721 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7722 if (Constraint.size() == 1) {
7723 switch (Constraint[0]) {
7735 return C_RegisterClass;
7740 return TargetLowering::getConstraintType(Constraint);
7743 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7744 /// with another that has more specific requirements based on the type of the
7745 /// corresponding operand.
7746 const char *X86TargetLowering::
7747 LowerXConstraint(MVT ConstraintVT) const {
7748 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7749 // 'f' like normal targets.
7750 if (ConstraintVT.isFloatingPoint()) {
7751 if (Subtarget->hasSSE2())
7753 if (Subtarget->hasSSE1())
7757 return TargetLowering::LowerXConstraint(ConstraintVT);
7760 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7761 /// vector. If it is invalid, don't add anything to Ops.
7762 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7765 std::vector<SDValue>&Ops,
7766 SelectionDAG &DAG) const {
7767 SDValue Result(0, 0);
7769 switch (Constraint) {
7772 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7773 if (C->getZExtValue() <= 31) {
7774 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7780 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7781 if (C->getZExtValue() <= 63) {
7782 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7788 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7789 if (C->getZExtValue() <= 255) {
7790 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7796 // Literal immediates are always ok.
7797 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7798 Result = DAG.getTargetConstant(CST->getZExtValue(), Op.getValueType());
7802 // If we are in non-pic codegen mode, we allow the address of a global (with
7803 // an optional displacement) to be used with 'i'.
7804 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7807 // Match either (GA) or (GA+C)
7809 Offset = GA->getOffset();
7810 } else if (Op.getOpcode() == ISD::ADD) {
7811 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7812 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7814 Offset = GA->getOffset()+C->getZExtValue();
7816 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7817 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7819 Offset = GA->getOffset()+C->getZExtValue();
7827 Op = LowerGlobalAddress(GA->getGlobal(), Offset, DAG);
7829 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7835 // Otherwise, not valid for this mode.
7840 if (Result.getNode()) {
7841 Ops.push_back(Result);
7844 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
7848 std::vector<unsigned> X86TargetLowering::
7849 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7851 if (Constraint.size() == 1) {
7852 // FIXME: not handling fp-stack yet!
7853 switch (Constraint[0]) { // GCC X86 Constraint Letters
7854 default: break; // Unknown constraint letter
7855 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7858 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7859 else if (VT == MVT::i16)
7860 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7861 else if (VT == MVT::i8)
7862 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7863 else if (VT == MVT::i64)
7864 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7869 return std::vector<unsigned>();
7872 std::pair<unsigned, const TargetRegisterClass*>
7873 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7875 // First, see if this is a constraint that directly corresponds to an LLVM
7877 if (Constraint.size() == 1) {
7878 // GCC Constraint Letters
7879 switch (Constraint[0]) {
7881 case 'r': // GENERAL_REGS
7882 case 'R': // LEGACY_REGS
7883 case 'l': // INDEX_REGS
7885 return std::make_pair(0U, X86::GR8RegisterClass);
7887 return std::make_pair(0U, X86::GR16RegisterClass);
7888 if (VT == MVT::i32 || !Subtarget->is64Bit())
7889 return std::make_pair(0U, X86::GR32RegisterClass);
7890 return std::make_pair(0U, X86::GR64RegisterClass);
7891 case 'f': // FP Stack registers.
7892 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7893 // value to the correct fpstack register class.
7894 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7895 return std::make_pair(0U, X86::RFP32RegisterClass);
7896 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7897 return std::make_pair(0U, X86::RFP64RegisterClass);
7898 return std::make_pair(0U, X86::RFP80RegisterClass);
7899 case 'y': // MMX_REGS if MMX allowed.
7900 if (!Subtarget->hasMMX()) break;
7901 return std::make_pair(0U, X86::VR64RegisterClass);
7902 case 'Y': // SSE_REGS if SSE2 allowed
7903 if (!Subtarget->hasSSE2()) break;
7905 case 'x': // SSE_REGS if SSE1 allowed
7906 if (!Subtarget->hasSSE1()) break;
7908 switch (VT.getSimpleVT()) {
7910 // Scalar SSE types.
7913 return std::make_pair(0U, X86::FR32RegisterClass);
7916 return std::make_pair(0U, X86::FR64RegisterClass);
7924 return std::make_pair(0U, X86::VR128RegisterClass);
7930 // Use the default implementation in TargetLowering to convert the register
7931 // constraint into a member of a register class.
7932 std::pair<unsigned, const TargetRegisterClass*> Res;
7933 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7935 // Not found as a standard register?
7936 if (Res.second == 0) {
7937 // GCC calls "st(0)" just plain "st".
7938 if (StringsEqualNoCase("{st}", Constraint)) {
7939 Res.first = X86::ST0;
7940 Res.second = X86::RFP80RegisterClass;
7942 // 'A' means EAX + EDX.
7943 if (Constraint == "A") {
7944 Res.first = X86::EAX;
7945 Res.second = X86::GRADRegisterClass;
7950 // Otherwise, check to see if this is a register class of the wrong value
7951 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7952 // turn into {ax},{dx}.
7953 if (Res.second->hasType(VT))
7954 return Res; // Correct type already, nothing to do.
7956 // All of the single-register GCC register classes map their values onto
7957 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7958 // really want an 8-bit or 32-bit register, map to the appropriate register
7959 // class and return the appropriate register.
7960 if (Res.second == X86::GR16RegisterClass) {
7961 if (VT == MVT::i8) {
7962 unsigned DestReg = 0;
7963 switch (Res.first) {
7965 case X86::AX: DestReg = X86::AL; break;
7966 case X86::DX: DestReg = X86::DL; break;
7967 case X86::CX: DestReg = X86::CL; break;
7968 case X86::BX: DestReg = X86::BL; break;
7971 Res.first = DestReg;
7972 Res.second = Res.second = X86::GR8RegisterClass;
7974 } else if (VT == MVT::i32) {
7975 unsigned DestReg = 0;
7976 switch (Res.first) {
7978 case X86::AX: DestReg = X86::EAX; break;
7979 case X86::DX: DestReg = X86::EDX; break;
7980 case X86::CX: DestReg = X86::ECX; break;
7981 case X86::BX: DestReg = X86::EBX; break;
7982 case X86::SI: DestReg = X86::ESI; break;
7983 case X86::DI: DestReg = X86::EDI; break;
7984 case X86::BP: DestReg = X86::EBP; break;
7985 case X86::SP: DestReg = X86::ESP; break;
7988 Res.first = DestReg;
7989 Res.second = Res.second = X86::GR32RegisterClass;
7991 } else if (VT == MVT::i64) {
7992 unsigned DestReg = 0;
7993 switch (Res.first) {
7995 case X86::AX: DestReg = X86::RAX; break;
7996 case X86::DX: DestReg = X86::RDX; break;
7997 case X86::CX: DestReg = X86::RCX; break;
7998 case X86::BX: DestReg = X86::RBX; break;
7999 case X86::SI: DestReg = X86::RSI; break;
8000 case X86::DI: DestReg = X86::RDI; break;
8001 case X86::BP: DestReg = X86::RBP; break;
8002 case X86::SP: DestReg = X86::RSP; break;
8005 Res.first = DestReg;
8006 Res.second = Res.second = X86::GR64RegisterClass;
8009 } else if (Res.second == X86::FR32RegisterClass ||
8010 Res.second == X86::FR64RegisterClass ||
8011 Res.second == X86::VR128RegisterClass) {
8012 // Handle references to XMM physical registers that got mapped into the
8013 // wrong class. This can happen with constraints like {xmm0} where the
8014 // target independent register mapper will just pick the first match it can
8015 // find, ignoring the required type.
8017 Res.second = X86::FR32RegisterClass;
8018 else if (VT == MVT::f64)
8019 Res.second = X86::FR64RegisterClass;
8020 else if (X86::VR128RegisterClass->hasType(VT))
8021 Res.second = X86::VR128RegisterClass;
8027 //===----------------------------------------------------------------------===//
8028 // X86 Widen vector type
8029 //===----------------------------------------------------------------------===//
8031 /// getWidenVectorType: given a vector type, returns the type to widen
8032 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8033 /// If there is no vector type that we want to widen to, returns MVT::Other
8034 /// When and where to widen is target dependent based on the cost of
8035 /// scalarizing vs using the wider vector type.
8037 MVT X86TargetLowering::getWidenVectorType(MVT VT) {
8038 assert(VT.isVector());
8039 if (isTypeLegal(VT))
8042 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8043 // type based on element type. This would speed up our search (though
8044 // it may not be worth it since the size of the list is relatively
8046 MVT EltVT = VT.getVectorElementType();
8047 unsigned NElts = VT.getVectorNumElements();
8049 // On X86, it make sense to widen any vector wider than 1
8053 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8054 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8055 MVT SVT = (MVT::SimpleValueType)nVT;
8057 if (isTypeLegal(SVT) &&
8058 SVT.getVectorElementType() == EltVT &&
8059 SVT.getVectorNumElements() > NElts)