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, DebugLoc dl);
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::i32 , Promote);
117 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
119 if (!UseSoftFloat && !NoImplicitFloat && X86ScalarSSEf64) {
120 // We have an impenetrably clever algorithm for ui64->double only.
121 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
123 // We have faster algorithm for ui32->single only.
124 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
126 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
130 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
132 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
133 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
135 if (!UseSoftFloat && !NoImplicitFloat) {
136 // SSE has no i16 to fp conversion, only i32
137 if (X86ScalarSSEf32) {
138 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
139 // f32 and f64 cases are Legal, f80 case is not
140 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
142 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
143 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
146 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
150 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
151 // are Legal, f80 is custom lowered.
152 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
153 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
155 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
157 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
158 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
160 if (X86ScalarSSEf32) {
161 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
162 // f32 and f64 cases are Legal, f80 case is not
163 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
165 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
166 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
169 // Handle FP_TO_UINT by promoting the destination to a larger signed
171 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
172 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
173 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
175 if (Subtarget->is64Bit()) {
176 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
177 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
179 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
180 // Expand FP_TO_UINT into a select.
181 // FIXME: We would like to use a Custom expander here eventually to do
182 // the optimal thing for SSE vs. the default expansion in the legalizer.
183 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
185 // With SSE3 we can use fisttpll to convert to a signed i64.
186 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
189 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
190 if (!X86ScalarSSEf64) {
191 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
192 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
195 // Scalar integer divide and remainder are lowered to use operations that
196 // produce two results, to match the available instructions. This exposes
197 // the two-result form to trivial CSE, which is able to combine x/y and x%y
198 // into a single instruction.
200 // Scalar integer multiply-high is also lowered to use two-result
201 // operations, to match the available instructions. However, plain multiply
202 // (low) operations are left as Legal, as there are single-result
203 // instructions for this in x86. Using the two-result multiply instructions
204 // when both high and low results are needed must be arranged by dagcombine.
205 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
206 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
207 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
208 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
209 setOperationAction(ISD::SREM , MVT::i8 , Expand);
210 setOperationAction(ISD::UREM , MVT::i8 , Expand);
211 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
212 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
213 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
214 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
215 setOperationAction(ISD::SREM , MVT::i16 , Expand);
216 setOperationAction(ISD::UREM , MVT::i16 , Expand);
217 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
218 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
219 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
220 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
221 setOperationAction(ISD::SREM , MVT::i32 , Expand);
222 setOperationAction(ISD::UREM , MVT::i32 , Expand);
223 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
224 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
225 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
226 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
227 setOperationAction(ISD::SREM , MVT::i64 , Expand);
228 setOperationAction(ISD::UREM , MVT::i64 , Expand);
230 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
231 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
232 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
233 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
234 if (Subtarget->is64Bit())
235 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
236 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
237 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
238 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
239 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
240 setOperationAction(ISD::FREM , MVT::f32 , Expand);
241 setOperationAction(ISD::FREM , MVT::f64 , Expand);
242 setOperationAction(ISD::FREM , MVT::f80 , Expand);
243 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
245 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
246 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
247 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
248 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
249 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
250 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
251 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
252 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
253 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
256 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
257 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
260 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
261 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
263 // These should be promoted to a larger select which is supported.
264 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
265 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
266 // X86 wants to expand cmov itself.
267 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
268 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
269 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
270 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
271 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
272 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
274 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
275 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
276 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
277 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
278 if (Subtarget->is64Bit()) {
279 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
280 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
282 // X86 ret instruction may pop stack.
283 setOperationAction(ISD::RET , MVT::Other, Custom);
284 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
287 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
288 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
289 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
290 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
291 if (Subtarget->is64Bit())
292 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
293 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
294 if (Subtarget->is64Bit()) {
295 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
296 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
297 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
298 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
300 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
301 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
302 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
303 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
304 if (Subtarget->is64Bit()) {
305 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
306 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
307 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
310 if (Subtarget->hasSSE1())
311 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
313 if (!Subtarget->hasSSE2())
314 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
316 // Expand certain atomics
317 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
318 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
319 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
320 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
327 if (!Subtarget->is64Bit()) {
328 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
329 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
333 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
334 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
337 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
338 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
339 // FIXME - use subtarget debug flags
340 if (!Subtarget->isTargetDarwin() &&
341 !Subtarget->isTargetELF() &&
342 !Subtarget->isTargetCygMing()) {
343 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
344 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
347 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
348 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
349 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
350 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
351 if (Subtarget->is64Bit()) {
352 setExceptionPointerRegister(X86::RAX);
353 setExceptionSelectorRegister(X86::RDX);
355 setExceptionPointerRegister(X86::EAX);
356 setExceptionSelectorRegister(X86::EDX);
358 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
359 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
361 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
363 setOperationAction(ISD::TRAP, MVT::Other, Legal);
365 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
366 setOperationAction(ISD::VASTART , MVT::Other, Custom);
367 setOperationAction(ISD::VAEND , MVT::Other, Expand);
368 if (Subtarget->is64Bit()) {
369 setOperationAction(ISD::VAARG , MVT::Other, Custom);
370 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
372 setOperationAction(ISD::VAARG , MVT::Other, Expand);
373 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
376 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
377 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
378 if (Subtarget->is64Bit())
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
380 if (Subtarget->isTargetCygMing())
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
383 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
385 if (!UseSoftFloat && X86ScalarSSEf64) {
386 // f32 and f64 use SSE.
387 // Set up the FP register classes.
388 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
389 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
391 // Use ANDPD to simulate FABS.
392 setOperationAction(ISD::FABS , MVT::f64, Custom);
393 setOperationAction(ISD::FABS , MVT::f32, Custom);
395 // Use XORP to simulate FNEG.
396 setOperationAction(ISD::FNEG , MVT::f64, Custom);
397 setOperationAction(ISD::FNEG , MVT::f32, Custom);
399 // Use ANDPD and ORPD to simulate FCOPYSIGN.
400 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
401 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
403 // We don't support sin/cos/fmod
404 setOperationAction(ISD::FSIN , MVT::f64, Expand);
405 setOperationAction(ISD::FCOS , MVT::f64, Expand);
406 setOperationAction(ISD::FSIN , MVT::f32, Expand);
407 setOperationAction(ISD::FCOS , MVT::f32, Expand);
409 // Expand FP immediates into loads from the stack, except for the special
411 addLegalFPImmediate(APFloat(+0.0)); // xorpd
412 addLegalFPImmediate(APFloat(+0.0f)); // xorps
414 // Floating truncations from f80 and extensions to f80 go through memory.
415 // If optimizing, we lie about this though and handle it in
416 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
418 setConvertAction(MVT::f32, MVT::f80, Expand);
419 setConvertAction(MVT::f64, MVT::f80, Expand);
420 setConvertAction(MVT::f80, MVT::f32, Expand);
421 setConvertAction(MVT::f80, MVT::f64, Expand);
423 } else if (!UseSoftFloat && X86ScalarSSEf32) {
424 // Use SSE for f32, x87 for f64.
425 // Set up the FP register classes.
426 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
427 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
429 // Use ANDPS to simulate FABS.
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
437 // Use ANDPS and ORPS to simulate FCOPYSIGN.
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 // We don't support sin/cos/fmod
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Special cases we handle for FP constants.
446 addLegalFPImmediate(APFloat(+0.0f)); // xorps
447 addLegalFPImmediate(APFloat(+0.0)); // FLD0
448 addLegalFPImmediate(APFloat(+1.0)); // FLD1
449 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
450 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
452 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
453 // this though and handle it in InstructionSelectPreprocess so that
454 // dagcombine2 can hack on these.
456 setConvertAction(MVT::f32, MVT::f64, Expand);
457 setConvertAction(MVT::f32, MVT::f80, Expand);
458 setConvertAction(MVT::f80, MVT::f32, Expand);
459 setConvertAction(MVT::f64, MVT::f32, Expand);
460 // And x87->x87 truncations also.
461 setConvertAction(MVT::f80, MVT::f64, Expand);
465 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
466 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
468 } else if (!UseSoftFloat) {
469 // f32 and f64 in x87.
470 // Set up the FP register classes.
471 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
472 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
474 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
475 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
476 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
477 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
479 // Floating truncations go through memory. If optimizing, we lie about
480 // this though and handle it in InstructionSelectPreprocess so that
481 // dagcombine2 can hack on these.
483 setConvertAction(MVT::f80, MVT::f32, Expand);
484 setConvertAction(MVT::f64, MVT::f32, Expand);
485 setConvertAction(MVT::f80, MVT::f64, Expand);
489 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
490 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
492 addLegalFPImmediate(APFloat(+0.0)); // FLD0
493 addLegalFPImmediate(APFloat(+1.0)); // FLD1
494 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
495 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
496 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
497 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
498 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
499 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
502 // Long double always uses X87.
503 if (!UseSoftFloat && !NoImplicitFloat) {
504 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
505 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
506 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
509 APFloat TmpFlt(+0.0);
510 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
512 addLegalFPImmediate(TmpFlt); // FLD0
514 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
515 APFloat TmpFlt2(+1.0);
516 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt2); // FLD1
519 TmpFlt2.changeSign();
520 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
524 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
525 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
529 // Always use a library call for pow.
530 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
531 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
534 setOperationAction(ISD::FLOG, MVT::f80, Expand);
535 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
537 setOperationAction(ISD::FEXP, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
540 // First set operation action for all vector types to either promote
541 // (for widening) or expand (for scalarization). Then we will selectively
542 // turn on ones that can be effectively codegen'd.
543 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
544 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
545 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
560 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
590 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
591 // with -msoft-float, disable use of MMX as well.
592 if (!UseSoftFloat && !NoImplicitFloat && !DisableMMX && Subtarget->hasMMX()) {
593 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
594 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
595 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
596 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
597 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
599 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
600 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
601 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
602 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
604 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
605 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
606 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
607 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
609 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
610 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
612 setOperationAction(ISD::AND, MVT::v8i8, Promote);
613 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
614 setOperationAction(ISD::AND, MVT::v4i16, Promote);
615 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
616 setOperationAction(ISD::AND, MVT::v2i32, Promote);
617 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
618 setOperationAction(ISD::AND, MVT::v1i64, Legal);
620 setOperationAction(ISD::OR, MVT::v8i8, Promote);
621 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
622 setOperationAction(ISD::OR, MVT::v4i16, Promote);
623 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
624 setOperationAction(ISD::OR, MVT::v2i32, Promote);
625 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
626 setOperationAction(ISD::OR, MVT::v1i64, Legal);
628 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
629 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
630 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
631 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
632 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
633 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
634 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
636 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
637 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
638 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
639 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
640 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
641 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
642 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
643 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
644 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
646 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
648 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
657 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
658 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
659 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
660 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
662 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
664 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
665 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
666 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
667 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
668 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
669 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
672 if (!UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE1()) {
673 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
675 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
676 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
677 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
678 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
679 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
680 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
681 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
684 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
685 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
689 if (!UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2()) {
690 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
692 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
693 // registers cannot be used even for integer operations.
694 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
695 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
696 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
697 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
699 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
700 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
701 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
702 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
703 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
704 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
705 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
706 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
707 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
708 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
709 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
710 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
711 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
712 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
713 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
714 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
716 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
721 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
722 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
723 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
724 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
727 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
728 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
729 MVT VT = (MVT::SimpleValueType)i;
730 // Do not attempt to custom lower non-power-of-2 vectors
731 if (!isPowerOf2_32(VT.getVectorNumElements()))
733 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
734 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
735 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
738 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
739 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
740 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
741 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
745 if (Subtarget->is64Bit()) {
746 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
747 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
750 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
751 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
752 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
753 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
754 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
755 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
756 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
757 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
758 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
759 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
760 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
761 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
764 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
766 // Custom lower v2i64 and v2f64 selects.
767 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
768 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
769 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
770 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
774 if (Subtarget->hasSSE41()) {
775 // FIXME: Do we need to handle scalar-to-vector here?
776 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
778 // i8 and i16 vectors are custom , because the source register and source
779 // source memory operand types are not the same width. f32 vectors are
780 // custom since the immediate controlling the insert encodes additional
782 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
783 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
792 if (Subtarget->is64Bit()) {
793 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
794 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
798 if (Subtarget->hasSSE42()) {
799 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
802 // We want to custom lower some of our intrinsics.
803 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
805 // Add/Sub/Mul with overflow operations are custom lowered.
806 setOperationAction(ISD::SADDO, MVT::i32, Custom);
807 setOperationAction(ISD::SADDO, MVT::i64, Custom);
808 setOperationAction(ISD::UADDO, MVT::i32, Custom);
809 setOperationAction(ISD::UADDO, MVT::i64, Custom);
810 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
811 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
812 setOperationAction(ISD::USUBO, MVT::i32, Custom);
813 setOperationAction(ISD::USUBO, MVT::i64, Custom);
814 setOperationAction(ISD::SMULO, MVT::i32, Custom);
815 setOperationAction(ISD::SMULO, MVT::i64, Custom);
816 setOperationAction(ISD::UMULO, MVT::i32, Custom);
817 setOperationAction(ISD::UMULO, MVT::i64, Custom);
819 // We have target-specific dag combine patterns for the following nodes:
820 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
821 setTargetDAGCombine(ISD::BUILD_VECTOR);
822 setTargetDAGCombine(ISD::SELECT);
823 setTargetDAGCombine(ISD::SHL);
824 setTargetDAGCombine(ISD::SRA);
825 setTargetDAGCombine(ISD::SRL);
826 setTargetDAGCombine(ISD::STORE);
828 computeRegisterProperties();
830 // FIXME: These should be based on subtarget info. Plus, the values should
831 // be smaller when we are in optimizing for size mode.
832 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
833 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
834 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
835 allowUnalignedMemoryAccesses = true; // x86 supports it!
836 setPrefLoopAlignment(16);
840 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
845 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
846 /// the desired ByVal argument alignment.
847 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
850 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
851 if (VTy->getBitWidth() == 128)
853 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
854 unsigned EltAlign = 0;
855 getMaxByValAlign(ATy->getElementType(), EltAlign);
856 if (EltAlign > MaxAlign)
858 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
859 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
860 unsigned EltAlign = 0;
861 getMaxByValAlign(STy->getElementType(i), EltAlign);
862 if (EltAlign > MaxAlign)
871 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
872 /// function arguments in the caller parameter area. For X86, aggregates
873 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
874 /// are at 4-byte boundaries.
875 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
876 if (Subtarget->is64Bit()) {
877 // Max of 8 and alignment of type.
878 unsigned TyAlign = TD->getABITypeAlignment(Ty);
885 if (Subtarget->hasSSE1())
886 getMaxByValAlign(Ty, Align);
890 /// getOptimalMemOpType - Returns the target specific optimal type for load
891 /// and store operations as a result of memset, memcpy, and memmove
892 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
895 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
896 bool isSrcConst, bool isSrcStr) const {
897 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
898 // linux. This is because the stack realignment code can't handle certain
899 // cases like PR2962. This should be removed when PR2962 is fixed.
900 if (!NoImplicitFloat && Subtarget->getStackAlignment() >= 16) {
901 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
903 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
906 if (Subtarget->is64Bit() && Size >= 8)
911 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
913 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
914 SelectionDAG &DAG) const {
915 if (usesGlobalOffsetTable())
916 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
917 if (!Subtarget->isPICStyleRIPRel())
918 // This doesn't have DebugLoc associated with it, but is not really the
919 // same as a Register.
920 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
925 //===----------------------------------------------------------------------===//
926 // Return Value Calling Convention Implementation
927 //===----------------------------------------------------------------------===//
929 #include "X86GenCallingConv.inc"
931 /// LowerRET - Lower an ISD::RET node.
932 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
933 DebugLoc dl = Op.getDebugLoc();
934 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
936 SmallVector<CCValAssign, 16> RVLocs;
937 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
938 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
939 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
940 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
942 // If this is the first return lowered for this function, add the regs to the
943 // liveout set for the function.
944 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
945 for (unsigned i = 0; i != RVLocs.size(); ++i)
946 if (RVLocs[i].isRegLoc())
947 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
949 SDValue Chain = Op.getOperand(0);
951 // Handle tail call return.
952 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
953 if (Chain.getOpcode() == X86ISD::TAILCALL) {
954 SDValue TailCall = Chain;
955 SDValue TargetAddress = TailCall.getOperand(1);
956 SDValue StackAdjustment = TailCall.getOperand(2);
957 assert(((TargetAddress.getOpcode() == ISD::Register &&
958 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
959 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
960 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
961 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
962 "Expecting an global address, external symbol, or register");
963 assert(StackAdjustment.getOpcode() == ISD::Constant &&
964 "Expecting a const value");
966 SmallVector<SDValue,8> Operands;
967 Operands.push_back(Chain.getOperand(0));
968 Operands.push_back(TargetAddress);
969 Operands.push_back(StackAdjustment);
970 // Copy registers used by the call. Last operand is a flag so it is not
972 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
973 Operands.push_back(Chain.getOperand(i));
975 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
982 SmallVector<SDValue, 6> RetOps;
983 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
984 // Operand #1 = Bytes To Pop
985 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
987 // Copy the result values into the output registers.
988 for (unsigned i = 0; i != RVLocs.size(); ++i) {
989 CCValAssign &VA = RVLocs[i];
990 assert(VA.isRegLoc() && "Can only return in registers!");
991 SDValue ValToCopy = Op.getOperand(i*2+1);
993 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
994 // the RET instruction and handled by the FP Stackifier.
995 if (VA.getLocReg() == X86::ST0 ||
996 VA.getLocReg() == X86::ST1) {
997 // If this is a copy from an xmm register to ST(0), use an FPExtend to
998 // change the value to the FP stack register class.
999 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1000 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1001 RetOps.push_back(ValToCopy);
1002 // Don't emit a copytoreg.
1006 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1007 // which is returned in RAX / RDX.
1008 if (Subtarget->is64Bit()) {
1009 MVT ValVT = ValToCopy.getValueType();
1010 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1011 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1012 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1013 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1017 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1018 Flag = Chain.getValue(1);
1021 // The x86-64 ABI for returning structs by value requires that we copy
1022 // the sret argument into %rax for the return. We saved the argument into
1023 // a virtual register in the entry block, so now we copy the value out
1025 if (Subtarget->is64Bit() &&
1026 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1027 MachineFunction &MF = DAG.getMachineFunction();
1028 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1029 unsigned Reg = FuncInfo->getSRetReturnReg();
1031 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1032 FuncInfo->setSRetReturnReg(Reg);
1034 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1036 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1037 Flag = Chain.getValue(1);
1040 RetOps[0] = Chain; // Update chain.
1042 // Add the flag if we have it.
1044 RetOps.push_back(Flag);
1046 return DAG.getNode(X86ISD::RET_FLAG, dl,
1047 MVT::Other, &RetOps[0], RetOps.size());
1051 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1052 /// appropriate copies out of appropriate physical registers. This assumes that
1053 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1054 /// being lowered. The returns a SDNode with the same number of values as the
1056 SDNode *X86TargetLowering::
1057 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1058 unsigned CallingConv, SelectionDAG &DAG) {
1060 DebugLoc dl = TheCall->getDebugLoc();
1061 // Assign locations to each value returned by this call.
1062 SmallVector<CCValAssign, 16> RVLocs;
1063 bool isVarArg = TheCall->isVarArg();
1064 bool Is64Bit = Subtarget->is64Bit();
1065 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1066 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1068 SmallVector<SDValue, 8> ResultVals;
1070 // Copy all of the result registers out of their specified physreg.
1071 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1072 CCValAssign &VA = RVLocs[i];
1073 MVT CopyVT = VA.getValVT();
1075 // If this is x86-64, and we disabled SSE, we can't return FP values
1076 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1077 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1078 cerr << "SSE register return with SSE disabled\n";
1082 // If this is a call to a function that returns an fp value on the floating
1083 // point stack, but where we prefer to use the value in xmm registers, copy
1084 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1085 if ((VA.getLocReg() == X86::ST0 ||
1086 VA.getLocReg() == X86::ST1) &&
1087 isScalarFPTypeInSSEReg(VA.getValVT())) {
1092 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1093 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1094 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1095 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1096 MVT::v2i64, InFlag).getValue(1);
1097 Val = Chain.getValue(0);
1098 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1099 Val, DAG.getConstant(0, MVT::i64));
1101 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1102 MVT::i64, InFlag).getValue(1);
1103 Val = Chain.getValue(0);
1105 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1107 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1108 CopyVT, InFlag).getValue(1);
1109 Val = Chain.getValue(0);
1111 InFlag = Chain.getValue(2);
1113 if (CopyVT != VA.getValVT()) {
1114 // Round the F80 the right size, which also moves to the appropriate xmm
1116 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1117 // This truncation won't change the value.
1118 DAG.getIntPtrConstant(1));
1121 ResultVals.push_back(Val);
1124 // Merge everything together with a MERGE_VALUES node.
1125 ResultVals.push_back(Chain);
1126 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1127 &ResultVals[0], ResultVals.size()).getNode();
1131 //===----------------------------------------------------------------------===//
1132 // C & StdCall & Fast Calling Convention implementation
1133 //===----------------------------------------------------------------------===//
1134 // StdCall calling convention seems to be standard for many Windows' API
1135 // routines and around. It differs from C calling convention just a little:
1136 // callee should clean up the stack, not caller. Symbols should be also
1137 // decorated in some fancy way :) It doesn't support any vector arguments.
1138 // For info on fast calling convention see Fast Calling Convention (tail call)
1139 // implementation LowerX86_32FastCCCallTo.
1141 /// AddLiveIn - This helper function adds the specified physical register to the
1142 /// MachineFunction as a live in value. It also creates a corresponding virtual
1143 /// register for it.
1144 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1145 const TargetRegisterClass *RC) {
1146 assert(RC->contains(PReg) && "Not the correct regclass!");
1147 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1148 MF.getRegInfo().addLiveIn(PReg, VReg);
1152 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1154 static bool CallIsStructReturn(CallSDNode *TheCall) {
1155 unsigned NumOps = TheCall->getNumArgs();
1159 return TheCall->getArgFlags(0).isSRet();
1162 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1163 /// return semantics.
1164 static bool ArgsAreStructReturn(SDValue Op) {
1165 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1169 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1172 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1173 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1175 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1179 switch (CallingConv) {
1182 case CallingConv::X86_StdCall:
1183 return !Subtarget->is64Bit();
1184 case CallingConv::X86_FastCall:
1185 return !Subtarget->is64Bit();
1186 case CallingConv::Fast:
1187 return PerformTailCallOpt;
1191 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1192 /// given CallingConvention value.
1193 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1194 if (Subtarget->is64Bit()) {
1195 if (Subtarget->isTargetWin64())
1196 return CC_X86_Win64_C;
1197 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1198 return CC_X86_64_TailCall;
1203 if (CC == CallingConv::X86_FastCall)
1204 return CC_X86_32_FastCall;
1205 else if (CC == CallingConv::Fast)
1206 return CC_X86_32_FastCC;
1211 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1212 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1214 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1215 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1216 if (CC == CallingConv::X86_FastCall)
1218 else if (CC == CallingConv::X86_StdCall)
1224 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1225 /// in a register before calling.
1226 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1227 return !IsTailCall && !Is64Bit &&
1228 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1229 Subtarget->isPICStyleGOT();
1232 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1233 /// address to be loaded in a register.
1235 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1236 return !Is64Bit && IsTailCall &&
1237 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1238 Subtarget->isPICStyleGOT();
1241 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1242 /// by "Src" to address "Dst" with size and alignment information specified by
1243 /// the specific parameter attribute. The copy will be passed as a byval
1244 /// function parameter.
1246 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1247 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1249 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1250 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1251 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1254 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1255 const CCValAssign &VA,
1256 MachineFrameInfo *MFI,
1258 SDValue Root, unsigned i) {
1259 // Create the nodes corresponding to a load from this parameter slot.
1260 ISD::ArgFlagsTy Flags =
1261 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1262 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1263 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1265 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1266 // changed with more analysis.
1267 // In case of tail call optimization mark all arguments mutable. Since they
1268 // could be overwritten by lowering of arguments in case of a tail call.
1269 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1270 VA.getLocMemOffset(), isImmutable);
1271 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1272 if (Flags.isByVal())
1274 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1275 PseudoSourceValue::getFixedStack(FI), 0);
1279 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1280 MachineFunction &MF = DAG.getMachineFunction();
1281 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1282 DebugLoc dl = Op.getDebugLoc();
1284 const Function* Fn = MF.getFunction();
1285 if (Fn->hasExternalLinkage() &&
1286 Subtarget->isTargetCygMing() &&
1287 Fn->getName() == "main")
1288 FuncInfo->setForceFramePointer(true);
1290 // Decorate the function name.
1291 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1293 MachineFrameInfo *MFI = MF.getFrameInfo();
1294 SDValue Root = Op.getOperand(0);
1295 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1296 unsigned CC = MF.getFunction()->getCallingConv();
1297 bool Is64Bit = Subtarget->is64Bit();
1298 bool IsWin64 = Subtarget->isTargetWin64();
1300 assert(!(isVarArg && CC == CallingConv::Fast) &&
1301 "Var args not supported with calling convention fastcc");
1303 // Assign locations to all of the incoming arguments.
1304 SmallVector<CCValAssign, 16> ArgLocs;
1305 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1306 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1308 SmallVector<SDValue, 8> ArgValues;
1309 unsigned LastVal = ~0U;
1310 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1311 CCValAssign &VA = ArgLocs[i];
1312 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1314 assert(VA.getValNo() != LastVal &&
1315 "Don't support value assigned to multiple locs yet");
1316 LastVal = VA.getValNo();
1318 if (VA.isRegLoc()) {
1319 MVT RegVT = VA.getLocVT();
1320 TargetRegisterClass *RC = NULL;
1321 if (RegVT == MVT::i32)
1322 RC = X86::GR32RegisterClass;
1323 else if (Is64Bit && RegVT == MVT::i64)
1324 RC = X86::GR64RegisterClass;
1325 else if (RegVT == MVT::f32)
1326 RC = X86::FR32RegisterClass;
1327 else if (RegVT == MVT::f64)
1328 RC = X86::FR64RegisterClass;
1329 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1330 RC = X86::VR128RegisterClass;
1331 else if (RegVT.isVector()) {
1332 assert(RegVT.getSizeInBits() == 64);
1334 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1336 // Darwin calling convention passes MMX values in either GPRs or
1337 // XMMs in x86-64. Other targets pass them in memory.
1338 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1339 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1342 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1347 assert(0 && "Unknown argument type!");
1350 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1351 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1353 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1354 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1356 if (VA.getLocInfo() == CCValAssign::SExt)
1357 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1358 DAG.getValueType(VA.getValVT()));
1359 else if (VA.getLocInfo() == CCValAssign::ZExt)
1360 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1361 DAG.getValueType(VA.getValVT()));
1363 if (VA.getLocInfo() != CCValAssign::Full)
1364 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1366 // Handle MMX values passed in GPRs.
1367 if (Is64Bit && RegVT != VA.getLocVT()) {
1368 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1369 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1370 else if (RC == X86::VR128RegisterClass) {
1371 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1372 ArgValue, DAG.getConstant(0, MVT::i64));
1373 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1377 ArgValues.push_back(ArgValue);
1379 assert(VA.isMemLoc());
1380 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1384 // The x86-64 ABI for returning structs by value requires that we copy
1385 // the sret argument into %rax for the return. Save the argument into
1386 // a virtual register so that we can access it from the return points.
1387 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1388 MachineFunction &MF = DAG.getMachineFunction();
1389 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1390 unsigned Reg = FuncInfo->getSRetReturnReg();
1392 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1393 FuncInfo->setSRetReturnReg(Reg);
1395 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1396 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1399 unsigned StackSize = CCInfo.getNextStackOffset();
1400 // align stack specially for tail calls
1401 if (PerformTailCallOpt && CC == CallingConv::Fast)
1402 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1404 // If the function takes variable number of arguments, make a frame index for
1405 // the start of the first vararg value... for expansion of llvm.va_start.
1407 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1408 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1411 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1413 // FIXME: We should really autogenerate these arrays
1414 static const unsigned GPR64ArgRegsWin64[] = {
1415 X86::RCX, X86::RDX, X86::R8, X86::R9
1417 static const unsigned XMMArgRegsWin64[] = {
1418 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1420 static const unsigned GPR64ArgRegs64Bit[] = {
1421 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1423 static const unsigned XMMArgRegs64Bit[] = {
1424 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1425 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1427 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1430 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1431 GPR64ArgRegs = GPR64ArgRegsWin64;
1432 XMMArgRegs = XMMArgRegsWin64;
1434 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1435 GPR64ArgRegs = GPR64ArgRegs64Bit;
1436 XMMArgRegs = XMMArgRegs64Bit;
1438 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1440 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1443 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1444 "SSE register cannot be used when SSE is disabled!");
1445 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloat) &&
1446 "SSE register cannot be used when SSE is disabled!");
1447 if (UseSoftFloat || NoImplicitFloat || !Subtarget->hasSSE1())
1448 // Kernel mode asks for SSE to be disabled, so don't push them
1450 TotalNumXMMRegs = 0;
1452 // For X86-64, if there are vararg parameters that are passed via
1453 // registers, then we must store them to their spots on the stack so they
1454 // may be loaded by deferencing the result of va_next.
1455 VarArgsGPOffset = NumIntRegs * 8;
1456 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1457 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1458 TotalNumXMMRegs * 16, 16);
1460 // Store the integer parameter registers.
1461 SmallVector<SDValue, 8> MemOps;
1462 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1463 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1464 DAG.getIntPtrConstant(VarArgsGPOffset));
1465 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1466 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1467 X86::GR64RegisterClass);
1468 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1470 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1471 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1472 MemOps.push_back(Store);
1473 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1474 DAG.getIntPtrConstant(8));
1477 // Now store the XMM (fp + vector) parameter registers.
1478 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1479 DAG.getIntPtrConstant(VarArgsFPOffset));
1480 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1481 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1482 X86::VR128RegisterClass);
1483 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1485 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1486 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1487 MemOps.push_back(Store);
1488 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1489 DAG.getIntPtrConstant(16));
1491 if (!MemOps.empty())
1492 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1493 &MemOps[0], MemOps.size());
1497 ArgValues.push_back(Root);
1499 // Some CCs need callee pop.
1500 if (IsCalleePop(isVarArg, CC)) {
1501 BytesToPopOnReturn = StackSize; // Callee pops everything.
1502 BytesCallerReserves = 0;
1504 BytesToPopOnReturn = 0; // Callee pops nothing.
1505 // If this is an sret function, the return should pop the hidden pointer.
1506 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1507 BytesToPopOnReturn = 4;
1508 BytesCallerReserves = StackSize;
1512 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1513 if (CC == CallingConv::X86_FastCall)
1514 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1517 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1519 // Return the new list of results.
1520 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1521 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1525 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1526 const SDValue &StackPtr,
1527 const CCValAssign &VA,
1529 SDValue Arg, ISD::ArgFlagsTy Flags) {
1530 DebugLoc dl = TheCall->getDebugLoc();
1531 unsigned LocMemOffset = VA.getLocMemOffset();
1532 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1533 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1534 if (Flags.isByVal()) {
1535 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1537 return DAG.getStore(Chain, dl, Arg, PtrOff,
1538 PseudoSourceValue::getStack(), LocMemOffset);
1541 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1542 /// optimization is performed and it is required.
1544 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1545 SDValue &OutRetAddr,
1551 if (!IsTailCall || FPDiff==0) return Chain;
1553 // Adjust the Return address stack slot.
1554 MVT VT = getPointerTy();
1555 OutRetAddr = getReturnAddressFrameIndex(DAG);
1557 // Load the "old" Return address.
1558 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1559 return SDValue(OutRetAddr.getNode(), 1);
1562 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1563 /// optimization is performed and it is required (FPDiff!=0).
1565 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1566 SDValue Chain, SDValue RetAddrFrIdx,
1567 bool Is64Bit, int FPDiff, DebugLoc dl) {
1568 // Store the return address to the appropriate stack slot.
1569 if (!FPDiff) return Chain;
1570 // Calculate the new stack slot for the return address.
1571 int SlotSize = Is64Bit ? 8 : 4;
1572 int NewReturnAddrFI =
1573 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1574 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1575 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1576 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1577 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1581 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1582 MachineFunction &MF = DAG.getMachineFunction();
1583 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1584 SDValue Chain = TheCall->getChain();
1585 unsigned CC = TheCall->getCallingConv();
1586 bool isVarArg = TheCall->isVarArg();
1587 bool IsTailCall = TheCall->isTailCall() &&
1588 CC == CallingConv::Fast && PerformTailCallOpt;
1589 SDValue Callee = TheCall->getCallee();
1590 bool Is64Bit = Subtarget->is64Bit();
1591 bool IsStructRet = CallIsStructReturn(TheCall);
1592 DebugLoc dl = TheCall->getDebugLoc();
1594 assert(!(isVarArg && CC == CallingConv::Fast) &&
1595 "Var args not supported with calling convention fastcc");
1597 // Analyze operands of the call, assigning locations to each operand.
1598 SmallVector<CCValAssign, 16> ArgLocs;
1599 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1600 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1602 // Get a count of how many bytes are to be pushed on the stack.
1603 unsigned NumBytes = CCInfo.getNextStackOffset();
1604 if (PerformTailCallOpt && CC == CallingConv::Fast)
1605 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1609 // Lower arguments at fp - stackoffset + fpdiff.
1610 unsigned NumBytesCallerPushed =
1611 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1612 FPDiff = NumBytesCallerPushed - NumBytes;
1614 // Set the delta of movement of the returnaddr stackslot.
1615 // But only set if delta is greater than previous delta.
1616 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1617 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1620 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1622 SDValue RetAddrFrIdx;
1623 // Load return adress for tail calls.
1624 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1627 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1628 SmallVector<SDValue, 8> MemOpChains;
1631 // Walk the register/memloc assignments, inserting copies/loads. In the case
1632 // of tail call optimization arguments are handle later.
1633 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1634 CCValAssign &VA = ArgLocs[i];
1635 SDValue Arg = TheCall->getArg(i);
1636 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1637 bool isByVal = Flags.isByVal();
1639 // Promote the value if needed.
1640 switch (VA.getLocInfo()) {
1641 default: assert(0 && "Unknown loc info!");
1642 case CCValAssign::Full: break;
1643 case CCValAssign::SExt:
1644 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1646 case CCValAssign::ZExt:
1647 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1649 case CCValAssign::AExt:
1650 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1654 if (VA.isRegLoc()) {
1656 MVT RegVT = VA.getLocVT();
1657 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1658 switch (VA.getLocReg()) {
1661 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1663 // Special case: passing MMX values in GPR registers.
1664 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1667 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1668 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1669 // Special case: passing MMX values in XMM registers.
1670 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1671 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1672 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
1673 DAG.getUNDEF(MVT::v2i64), Arg,
1674 getMOVLMask(2, DAG, dl));
1679 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1681 if (!IsTailCall || (IsTailCall && isByVal)) {
1682 assert(VA.isMemLoc());
1683 if (StackPtr.getNode() == 0)
1684 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1686 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1687 Chain, Arg, Flags));
1692 if (!MemOpChains.empty())
1693 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1694 &MemOpChains[0], MemOpChains.size());
1696 // Build a sequence of copy-to-reg nodes chained together with token chain
1697 // and flag operands which copy the outgoing args into registers.
1699 // Tail call byval lowering might overwrite argument registers so in case of
1700 // tail call optimization the copies to registers are lowered later.
1702 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1703 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1704 RegsToPass[i].second, InFlag);
1705 InFlag = Chain.getValue(1);
1708 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1710 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1711 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1712 DAG.getNode(X86ISD::GlobalBaseReg,
1713 DebugLoc::getUnknownLoc(),
1716 InFlag = Chain.getValue(1);
1718 // If we are tail calling and generating PIC/GOT style code load the address
1719 // of the callee into ecx. The value in ecx is used as target of the tail
1720 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1721 // calls on PIC/GOT architectures. Normally we would just put the address of
1722 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1723 // restored (since ebx is callee saved) before jumping to the target@PLT.
1724 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1725 // Note: The actual moving to ecx is done further down.
1726 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1727 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1728 !G->getGlobal()->hasProtectedVisibility())
1729 Callee = LowerGlobalAddress(Callee, DAG);
1730 else if (isa<ExternalSymbolSDNode>(Callee))
1731 Callee = LowerExternalSymbol(Callee,DAG);
1734 if (Is64Bit && isVarArg) {
1735 // From AMD64 ABI document:
1736 // For calls that may call functions that use varargs or stdargs
1737 // (prototype-less calls or calls to functions containing ellipsis (...) in
1738 // the declaration) %al is used as hidden argument to specify the number
1739 // of SSE registers used. The contents of %al do not need to match exactly
1740 // the number of registers, but must be an ubound on the number of SSE
1741 // registers used and is in the range 0 - 8 inclusive.
1743 // FIXME: Verify this on Win64
1744 // Count the number of XMM registers allocated.
1745 static const unsigned XMMArgRegs[] = {
1746 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1747 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1749 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1750 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1751 && "SSE registers cannot be used when SSE is disabled");
1753 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1754 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1755 InFlag = Chain.getValue(1);
1759 // For tail calls lower the arguments to the 'real' stack slot.
1761 SmallVector<SDValue, 8> MemOpChains2;
1764 // Do not flag preceeding copytoreg stuff together with the following stuff.
1766 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1767 CCValAssign &VA = ArgLocs[i];
1768 if (!VA.isRegLoc()) {
1769 assert(VA.isMemLoc());
1770 SDValue Arg = TheCall->getArg(i);
1771 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1772 // Create frame index.
1773 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1774 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1775 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1776 FIN = DAG.getFrameIndex(FI, getPointerTy());
1778 if (Flags.isByVal()) {
1779 // Copy relative to framepointer.
1780 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1781 if (StackPtr.getNode() == 0)
1782 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1784 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1786 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1789 // Store relative to framepointer.
1790 MemOpChains2.push_back(
1791 DAG.getStore(Chain, dl, Arg, FIN,
1792 PseudoSourceValue::getFixedStack(FI), 0));
1797 if (!MemOpChains2.empty())
1798 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1799 &MemOpChains2[0], MemOpChains2.size());
1801 // Copy arguments to their registers.
1802 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1803 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1804 RegsToPass[i].second, InFlag);
1805 InFlag = Chain.getValue(1);
1809 // Store the return address to the appropriate stack slot.
1810 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1814 // If the callee is a GlobalAddress node (quite common, every direct call is)
1815 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1816 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1817 // We should use extra load for direct calls to dllimported functions in
1819 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1820 getTargetMachine(), true))
1821 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1823 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1824 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1825 } else if (IsTailCall) {
1826 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1828 Chain = DAG.getCopyToReg(Chain, dl,
1829 DAG.getRegister(Opc, getPointerTy()),
1831 Callee = DAG.getRegister(Opc, getPointerTy());
1832 // Add register as live out.
1833 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1836 // Returns a chain & a flag for retval copy to use.
1837 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1838 SmallVector<SDValue, 8> Ops;
1841 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1842 DAG.getIntPtrConstant(0, true), InFlag);
1843 InFlag = Chain.getValue(1);
1845 // Returns a chain & a flag for retval copy to use.
1846 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1850 Ops.push_back(Chain);
1851 Ops.push_back(Callee);
1854 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1856 // Add argument registers to the end of the list so that they are known live
1858 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1859 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1860 RegsToPass[i].second.getValueType()));
1862 // Add an implicit use GOT pointer in EBX.
1863 if (!IsTailCall && !Is64Bit &&
1864 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1865 Subtarget->isPICStyleGOT())
1866 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1868 // Add an implicit use of AL for x86 vararg functions.
1869 if (Is64Bit && isVarArg)
1870 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1872 if (InFlag.getNode())
1873 Ops.push_back(InFlag);
1876 assert(InFlag.getNode() &&
1877 "Flag must be set. Depend on flag being set in LowerRET");
1878 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1879 TheCall->getVTList(), &Ops[0], Ops.size());
1881 return SDValue(Chain.getNode(), Op.getResNo());
1884 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1885 InFlag = Chain.getValue(1);
1887 // Create the CALLSEQ_END node.
1888 unsigned NumBytesForCalleeToPush;
1889 if (IsCalleePop(isVarArg, CC))
1890 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1891 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1892 // If this is is a call to a struct-return function, the callee
1893 // pops the hidden struct pointer, so we have to push it back.
1894 // This is common for Darwin/X86, Linux & Mingw32 targets.
1895 NumBytesForCalleeToPush = 4;
1897 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1899 // Returns a flag for retval copy to use.
1900 Chain = DAG.getCALLSEQ_END(Chain,
1901 DAG.getIntPtrConstant(NumBytes, true),
1902 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1905 InFlag = Chain.getValue(1);
1907 // Handle result values, copying them out of physregs into vregs that we
1909 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1914 //===----------------------------------------------------------------------===//
1915 // Fast Calling Convention (tail call) implementation
1916 //===----------------------------------------------------------------------===//
1918 // Like std call, callee cleans arguments, convention except that ECX is
1919 // reserved for storing the tail called function address. Only 2 registers are
1920 // free for argument passing (inreg). Tail call optimization is performed
1922 // * tailcallopt is enabled
1923 // * caller/callee are fastcc
1924 // On X86_64 architecture with GOT-style position independent code only local
1925 // (within module) calls are supported at the moment.
1926 // To keep the stack aligned according to platform abi the function
1927 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1928 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1929 // If a tail called function callee has more arguments than the caller the
1930 // caller needs to make sure that there is room to move the RETADDR to. This is
1931 // achieved by reserving an area the size of the argument delta right after the
1932 // original REtADDR, but before the saved framepointer or the spilled registers
1933 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1945 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1946 /// for a 16 byte align requirement.
1947 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1948 SelectionDAG& DAG) {
1949 MachineFunction &MF = DAG.getMachineFunction();
1950 const TargetMachine &TM = MF.getTarget();
1951 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1952 unsigned StackAlignment = TFI.getStackAlignment();
1953 uint64_t AlignMask = StackAlignment - 1;
1954 int64_t Offset = StackSize;
1955 uint64_t SlotSize = TD->getPointerSize();
1956 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1957 // Number smaller than 12 so just add the difference.
1958 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1960 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1961 Offset = ((~AlignMask) & Offset) + StackAlignment +
1962 (StackAlignment-SlotSize);
1967 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1968 /// following the call is a return. A function is eligible if caller/callee
1969 /// calling conventions match, currently only fastcc supports tail calls, and
1970 /// the function CALL is immediatly followed by a RET.
1971 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1973 SelectionDAG& DAG) const {
1974 if (!PerformTailCallOpt)
1977 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1978 MachineFunction &MF = DAG.getMachineFunction();
1979 unsigned CallerCC = MF.getFunction()->getCallingConv();
1980 unsigned CalleeCC= TheCall->getCallingConv();
1981 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1982 SDValue Callee = TheCall->getCallee();
1983 // On x86/32Bit PIC/GOT tail calls are supported.
1984 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1985 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1988 // Can only do local tail calls (in same module, hidden or protected) on
1989 // x86_64 PIC/GOT at the moment.
1990 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1991 return G->getGlobal()->hasHiddenVisibility()
1992 || G->getGlobal()->hasProtectedVisibility();
2000 X86TargetLowering::createFastISel(MachineFunction &mf,
2001 MachineModuleInfo *mmo,
2003 DenseMap<const Value *, unsigned> &vm,
2004 DenseMap<const BasicBlock *,
2005 MachineBasicBlock *> &bm,
2006 DenseMap<const AllocaInst *, int> &am
2008 , SmallSet<Instruction*, 8> &cil
2011 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2019 //===----------------------------------------------------------------------===//
2020 // Other Lowering Hooks
2021 //===----------------------------------------------------------------------===//
2024 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2025 MachineFunction &MF = DAG.getMachineFunction();
2026 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2027 int ReturnAddrIndex = FuncInfo->getRAIndex();
2029 if (ReturnAddrIndex == 0) {
2030 // Set up a frame object for the return address.
2031 uint64_t SlotSize = TD->getPointerSize();
2032 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2033 FuncInfo->setRAIndex(ReturnAddrIndex);
2036 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2040 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2041 /// specific condition code, returning the condition code and the LHS/RHS of the
2042 /// comparison to make.
2043 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2044 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2046 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2047 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2048 // X > -1 -> X == 0, jump !sign.
2049 RHS = DAG.getConstant(0, RHS.getValueType());
2050 return X86::COND_NS;
2051 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2052 // X < 0 -> X == 0, jump on sign.
2054 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2056 RHS = DAG.getConstant(0, RHS.getValueType());
2057 return X86::COND_LE;
2061 switch (SetCCOpcode) {
2062 default: assert(0 && "Invalid integer condition!");
2063 case ISD::SETEQ: return X86::COND_E;
2064 case ISD::SETGT: return X86::COND_G;
2065 case ISD::SETGE: return X86::COND_GE;
2066 case ISD::SETLT: return X86::COND_L;
2067 case ISD::SETLE: return X86::COND_LE;
2068 case ISD::SETNE: return X86::COND_NE;
2069 case ISD::SETULT: return X86::COND_B;
2070 case ISD::SETUGT: return X86::COND_A;
2071 case ISD::SETULE: return X86::COND_BE;
2072 case ISD::SETUGE: return X86::COND_AE;
2076 // First determine if it is required or is profitable to flip the operands.
2078 // If LHS is a foldable load, but RHS is not, flip the condition.
2079 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2080 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2081 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2082 std::swap(LHS, RHS);
2085 switch (SetCCOpcode) {
2091 std::swap(LHS, RHS);
2095 // On a floating point condition, the flags are set as follows:
2097 // 0 | 0 | 0 | X > Y
2098 // 0 | 0 | 1 | X < Y
2099 // 1 | 0 | 0 | X == Y
2100 // 1 | 1 | 1 | unordered
2101 switch (SetCCOpcode) {
2102 default: assert(0 && "Condcode should be pre-legalized away");
2104 case ISD::SETEQ: return X86::COND_E;
2105 case ISD::SETOLT: // flipped
2107 case ISD::SETGT: return X86::COND_A;
2108 case ISD::SETOLE: // flipped
2110 case ISD::SETGE: return X86::COND_AE;
2111 case ISD::SETUGT: // flipped
2113 case ISD::SETLT: return X86::COND_B;
2114 case ISD::SETUGE: // flipped
2116 case ISD::SETLE: return X86::COND_BE;
2118 case ISD::SETNE: return X86::COND_NE;
2119 case ISD::SETUO: return X86::COND_P;
2120 case ISD::SETO: return X86::COND_NP;
2124 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2125 /// code. Current x86 isa includes the following FP cmov instructions:
2126 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2127 static bool hasFPCMov(unsigned X86CC) {
2143 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2144 /// true if Op is undef or if its value falls within the specified range (L, H].
2145 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2146 if (Op.getOpcode() == ISD::UNDEF)
2149 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2150 return (Val >= Low && Val < Hi);
2153 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2154 /// true if Op is undef or if its value equal to the specified value.
2155 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2156 if (Op.getOpcode() == ISD::UNDEF)
2158 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2161 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2162 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2163 bool X86::isPSHUFDMask(SDNode *N) {
2164 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2166 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2169 // Check if the value doesn't reference the second vector.
2170 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2171 SDValue Arg = N->getOperand(i);
2172 if (Arg.getOpcode() == ISD::UNDEF) continue;
2173 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2174 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2181 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2182 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2183 bool X86::isPSHUFHWMask(SDNode *N) {
2184 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2186 if (N->getNumOperands() != 8)
2189 // Lower quadword copied in order.
2190 for (unsigned i = 0; i != 4; ++i) {
2191 SDValue Arg = N->getOperand(i);
2192 if (Arg.getOpcode() == ISD::UNDEF) continue;
2193 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2194 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2198 // Upper quadword shuffled.
2199 for (unsigned i = 4; i != 8; ++i) {
2200 SDValue Arg = N->getOperand(i);
2201 if (Arg.getOpcode() == ISD::UNDEF) continue;
2202 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2203 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2204 if (Val < 4 || Val > 7)
2211 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2212 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2213 bool X86::isPSHUFLWMask(SDNode *N) {
2214 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2216 if (N->getNumOperands() != 8)
2219 // Upper quadword copied in order.
2220 for (unsigned i = 4; i != 8; ++i)
2221 if (!isUndefOrEqual(N->getOperand(i), i))
2224 // Lower quadword shuffled.
2225 for (unsigned i = 0; i != 4; ++i)
2226 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2232 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2233 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2234 template<class SDOperand>
2235 static bool isSHUFPMask(SDOperand *Elems, unsigned NumElems) {
2236 if (NumElems != 2 && NumElems != 4) return false;
2238 unsigned Half = NumElems / 2;
2239 for (unsigned i = 0; i < Half; ++i)
2240 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2242 for (unsigned i = Half; i < NumElems; ++i)
2243 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2249 bool X86::isSHUFPMask(SDNode *N) {
2250 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2251 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2254 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2255 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2256 /// half elements to come from vector 1 (which would equal the dest.) and
2257 /// the upper half to come from vector 2.
2258 template<class SDOperand>
2259 static bool isCommutedSHUFP(SDOperand *Ops, unsigned NumOps) {
2260 if (NumOps != 2 && NumOps != 4) return false;
2262 unsigned Half = NumOps / 2;
2263 for (unsigned i = 0; i < Half; ++i)
2264 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2266 for (unsigned i = Half; i < NumOps; ++i)
2267 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2272 static bool isCommutedSHUFP(SDNode *N) {
2273 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2274 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2277 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2278 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2279 bool X86::isMOVHLPSMask(SDNode *N) {
2280 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2282 if (N->getNumOperands() != 4)
2285 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2286 return isUndefOrEqual(N->getOperand(0), 6) &&
2287 isUndefOrEqual(N->getOperand(1), 7) &&
2288 isUndefOrEqual(N->getOperand(2), 2) &&
2289 isUndefOrEqual(N->getOperand(3), 3);
2292 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2293 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2295 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2296 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2298 if (N->getNumOperands() != 4)
2301 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2302 return isUndefOrEqual(N->getOperand(0), 2) &&
2303 isUndefOrEqual(N->getOperand(1), 3) &&
2304 isUndefOrEqual(N->getOperand(2), 2) &&
2305 isUndefOrEqual(N->getOperand(3), 3);
2308 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2309 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2310 bool X86::isMOVLPMask(SDNode *N) {
2311 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2313 unsigned NumElems = N->getNumOperands();
2314 if (NumElems != 2 && NumElems != 4)
2317 for (unsigned i = 0; i < NumElems/2; ++i)
2318 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2321 for (unsigned i = NumElems/2; i < NumElems; ++i)
2322 if (!isUndefOrEqual(N->getOperand(i), i))
2328 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2329 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2331 bool X86::isMOVHPMask(SDNode *N) {
2332 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2334 unsigned NumElems = N->getNumOperands();
2335 if (NumElems != 2 && NumElems != 4)
2338 for (unsigned i = 0; i < NumElems/2; ++i)
2339 if (!isUndefOrEqual(N->getOperand(i), i))
2342 for (unsigned i = 0; i < NumElems/2; ++i) {
2343 SDValue Arg = N->getOperand(i + NumElems/2);
2344 if (!isUndefOrEqual(Arg, i + NumElems))
2351 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2352 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2353 template<class SDOperand>
2354 bool static isUNPCKLMask(SDOperand *Elts, unsigned NumElts,
2355 bool V2IsSplat = false) {
2356 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2359 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2360 SDValue BitI = Elts[i];
2361 SDValue BitI1 = Elts[i+1];
2362 if (!isUndefOrEqual(BitI, j))
2365 if (!isUndefOrEqual(BitI1, NumElts))
2368 if (!isUndefOrEqual(BitI1, j + NumElts))
2376 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2377 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2378 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2381 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2382 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2383 template<class SDOperand>
2384 bool static isUNPCKHMask(SDOperand *Elts, unsigned NumElts,
2385 bool V2IsSplat = false) {
2386 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2389 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2390 SDValue BitI = Elts[i];
2391 SDValue BitI1 = Elts[i+1];
2392 if (!isUndefOrEqual(BitI, j + NumElts/2))
2395 if (isUndefOrEqual(BitI1, NumElts))
2398 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2406 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2407 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2408 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2411 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2412 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2414 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2415 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2417 unsigned NumElems = N->getNumOperands();
2418 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2421 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2422 SDValue BitI = N->getOperand(i);
2423 SDValue BitI1 = N->getOperand(i+1);
2425 if (!isUndefOrEqual(BitI, j))
2427 if (!isUndefOrEqual(BitI1, j))
2434 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2435 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2437 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2438 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2440 unsigned NumElems = N->getNumOperands();
2441 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2444 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2445 SDValue BitI = N->getOperand(i);
2446 SDValue BitI1 = N->getOperand(i + 1);
2448 if (!isUndefOrEqual(BitI, j))
2450 if (!isUndefOrEqual(BitI1, j))
2457 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2458 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2459 /// MOVSD, and MOVD, i.e. setting the lowest element.
2460 template<class SDOperand>
2461 static bool isMOVLMask(SDOperand *Elts, unsigned NumElts) {
2462 if (NumElts != 2 && NumElts != 4)
2465 if (!isUndefOrEqual(Elts[0], NumElts))
2468 for (unsigned i = 1; i < NumElts; ++i) {
2469 if (!isUndefOrEqual(Elts[i], i))
2476 bool X86::isMOVLMask(SDNode *N) {
2477 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2478 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2481 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2482 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2483 /// element of vector 2 and the other elements to come from vector 1 in order.
2484 template<class SDOperand>
2485 static bool isCommutedMOVL(SDOperand *Ops, unsigned NumOps,
2486 bool V2IsSplat = false,
2487 bool V2IsUndef = false) {
2488 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2491 if (!isUndefOrEqual(Ops[0], 0))
2494 for (unsigned i = 1; i < NumOps; ++i) {
2495 SDValue Arg = Ops[i];
2496 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2497 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2498 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2505 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2506 bool V2IsUndef = false) {
2507 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2508 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2509 V2IsSplat, V2IsUndef);
2512 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2513 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2514 bool X86::isMOVSHDUPMask(SDNode *N) {
2515 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2517 if (N->getNumOperands() != 4)
2520 // Expect 1, 1, 3, 3
2521 for (unsigned i = 0; i < 2; ++i) {
2522 SDValue Arg = N->getOperand(i);
2523 if (Arg.getOpcode() == ISD::UNDEF) continue;
2524 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2525 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2526 if (Val != 1) return false;
2530 for (unsigned i = 2; i < 4; ++i) {
2531 SDValue Arg = N->getOperand(i);
2532 if (Arg.getOpcode() == ISD::UNDEF) continue;
2533 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2534 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2535 if (Val != 3) return false;
2539 // Don't use movshdup if it can be done with a shufps.
2543 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2544 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2545 bool X86::isMOVSLDUPMask(SDNode *N) {
2546 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2548 if (N->getNumOperands() != 4)
2551 // Expect 0, 0, 2, 2
2552 for (unsigned i = 0; i < 2; ++i) {
2553 SDValue Arg = N->getOperand(i);
2554 if (Arg.getOpcode() == ISD::UNDEF) continue;
2555 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2556 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2557 if (Val != 0) return false;
2561 for (unsigned i = 2; i < 4; ++i) {
2562 SDValue Arg = N->getOperand(i);
2563 if (Arg.getOpcode() == ISD::UNDEF) continue;
2564 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2565 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2566 if (Val != 2) return false;
2570 // Don't use movshdup if it can be done with a shufps.
2574 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2575 /// specifies a identity operation on the LHS or RHS.
2576 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2577 unsigned NumElems = N->getNumOperands();
2578 for (unsigned i = 0; i < NumElems; ++i)
2579 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2584 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2585 /// a splat of a single element.
2586 static bool isSplatMask(SDNode *N) {
2587 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2589 // This is a splat operation if each element of the permute is the same, and
2590 // if the value doesn't reference the second vector.
2591 unsigned NumElems = N->getNumOperands();
2592 SDValue ElementBase;
2594 for (; i != NumElems; ++i) {
2595 SDValue Elt = N->getOperand(i);
2596 if (isa<ConstantSDNode>(Elt)) {
2602 if (!ElementBase.getNode())
2605 for (; i != NumElems; ++i) {
2606 SDValue Arg = N->getOperand(i);
2607 if (Arg.getOpcode() == ISD::UNDEF) continue;
2608 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2609 if (Arg != ElementBase) return false;
2612 // Make sure it is a splat of the first vector operand.
2613 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2616 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2617 /// we want to splat.
2618 static SDValue getSplatMaskEltNo(SDNode *N) {
2619 assert(isSplatMask(N) && "Not a splat mask");
2620 unsigned NumElems = N->getNumOperands();
2621 SDValue ElementBase;
2623 for (; i != NumElems; ++i) {
2624 SDValue Elt = N->getOperand(i);
2625 if (isa<ConstantSDNode>(Elt))
2628 assert(0 && " No splat value found!");
2633 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2634 /// a splat of a single element and it's a 2 or 4 element mask.
2635 bool X86::isSplatMask(SDNode *N) {
2636 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2638 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2639 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2641 return ::isSplatMask(N);
2644 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2645 /// specifies a splat of zero element.
2646 bool X86::isSplatLoMask(SDNode *N) {
2647 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2649 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2650 if (!isUndefOrEqual(N->getOperand(i), 0))
2655 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2656 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2657 bool X86::isMOVDDUPMask(SDNode *N) {
2658 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2660 unsigned e = N->getNumOperands() / 2;
2661 for (unsigned i = 0; i < e; ++i)
2662 if (!isUndefOrEqual(N->getOperand(i), i))
2664 for (unsigned i = 0; i < e; ++i)
2665 if (!isUndefOrEqual(N->getOperand(e+i), i))
2670 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2671 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2673 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2674 unsigned NumOperands = N->getNumOperands();
2675 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2677 for (unsigned i = 0; i < NumOperands; ++i) {
2679 SDValue Arg = N->getOperand(NumOperands-i-1);
2680 if (Arg.getOpcode() != ISD::UNDEF)
2681 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2682 if (Val >= NumOperands) Val -= NumOperands;
2684 if (i != NumOperands - 1)
2691 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2692 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2694 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2696 // 8 nodes, but we only care about the last 4.
2697 for (unsigned i = 7; i >= 4; --i) {
2699 SDValue Arg = N->getOperand(i);
2700 if (Arg.getOpcode() != ISD::UNDEF) {
2701 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2711 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2712 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2714 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2716 // 8 nodes, but we only care about the first 4.
2717 for (int i = 3; i >= 0; --i) {
2719 SDValue Arg = N->getOperand(i);
2720 if (Arg.getOpcode() != ISD::UNDEF)
2721 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2730 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2731 /// values in ther permute mask.
2732 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2733 SDValue &V2, SDValue &Mask,
2734 SelectionDAG &DAG) {
2735 MVT VT = Op.getValueType();
2736 MVT MaskVT = Mask.getValueType();
2737 MVT EltVT = MaskVT.getVectorElementType();
2738 unsigned NumElems = Mask.getNumOperands();
2739 SmallVector<SDValue, 8> MaskVec;
2740 DebugLoc dl = Op.getDebugLoc();
2742 for (unsigned i = 0; i != NumElems; ++i) {
2743 SDValue Arg = Mask.getOperand(i);
2744 if (Arg.getOpcode() == ISD::UNDEF) {
2745 MaskVec.push_back(DAG.getUNDEF(EltVT));
2748 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2749 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2751 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2753 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2757 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2758 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
2761 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2762 /// the two vector operands have swapped position.
2764 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG, DebugLoc dl) {
2765 MVT MaskVT = Mask.getValueType();
2766 MVT EltVT = MaskVT.getVectorElementType();
2767 unsigned NumElems = Mask.getNumOperands();
2768 SmallVector<SDValue, 8> MaskVec;
2769 for (unsigned i = 0; i != NumElems; ++i) {
2770 SDValue Arg = Mask.getOperand(i);
2771 if (Arg.getOpcode() == ISD::UNDEF) {
2772 MaskVec.push_back(DAG.getUNDEF(EltVT));
2775 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2776 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2778 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2780 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2782 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2786 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2787 /// match movhlps. The lower half elements should come from upper half of
2788 /// V1 (and in order), and the upper half elements should come from the upper
2789 /// half of V2 (and in order).
2790 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2791 unsigned NumElems = Mask->getNumOperands();
2794 for (unsigned i = 0, e = 2; i != e; ++i)
2795 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2797 for (unsigned i = 2; i != 4; ++i)
2798 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2803 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2804 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2806 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2807 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2809 N = N->getOperand(0).getNode();
2810 if (!ISD::isNON_EXTLoad(N))
2813 *LD = cast<LoadSDNode>(N);
2817 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2818 /// match movlp{s|d}. The lower half elements should come from lower half of
2819 /// V1 (and in order), and the upper half elements should come from the upper
2820 /// half of V2 (and in order). And since V1 will become the source of the
2821 /// MOVLP, it must be either a vector load or a scalar load to vector.
2822 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2823 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2825 // Is V2 is a vector load, don't do this transformation. We will try to use
2826 // load folding shufps op.
2827 if (ISD::isNON_EXTLoad(V2))
2830 unsigned NumElems = Mask->getNumOperands();
2831 if (NumElems != 2 && NumElems != 4)
2833 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2834 if (!isUndefOrEqual(Mask->getOperand(i), i))
2836 for (unsigned i = NumElems/2; i != NumElems; ++i)
2837 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2842 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2844 static bool isSplatVector(SDNode *N) {
2845 if (N->getOpcode() != ISD::BUILD_VECTOR)
2848 SDValue SplatValue = N->getOperand(0);
2849 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2850 if (N->getOperand(i) != SplatValue)
2855 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2857 static bool isUndefShuffle(SDNode *N) {
2858 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2861 SDValue V1 = N->getOperand(0);
2862 SDValue V2 = N->getOperand(1);
2863 SDValue Mask = N->getOperand(2);
2864 unsigned NumElems = Mask.getNumOperands();
2865 for (unsigned i = 0; i != NumElems; ++i) {
2866 SDValue Arg = Mask.getOperand(i);
2867 if (Arg.getOpcode() != ISD::UNDEF) {
2868 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2869 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2871 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2878 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2880 static inline bool isZeroNode(SDValue Elt) {
2881 return ((isa<ConstantSDNode>(Elt) &&
2882 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2883 (isa<ConstantFPSDNode>(Elt) &&
2884 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2887 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2888 /// to an zero vector.
2889 static bool isZeroShuffle(SDNode *N) {
2890 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2893 SDValue V1 = N->getOperand(0);
2894 SDValue V2 = N->getOperand(1);
2895 SDValue Mask = N->getOperand(2);
2896 unsigned NumElems = Mask.getNumOperands();
2897 for (unsigned i = 0; i != NumElems; ++i) {
2898 SDValue Arg = Mask.getOperand(i);
2899 if (Arg.getOpcode() == ISD::UNDEF)
2902 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2903 if (Idx < NumElems) {
2904 unsigned Opc = V1.getNode()->getOpcode();
2905 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2907 if (Opc != ISD::BUILD_VECTOR ||
2908 !isZeroNode(V1.getNode()->getOperand(Idx)))
2910 } else if (Idx >= NumElems) {
2911 unsigned Opc = V2.getNode()->getOpcode();
2912 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2914 if (Opc != ISD::BUILD_VECTOR ||
2915 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2922 /// getZeroVector - Returns a vector of specified type with all zero elements.
2924 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2926 assert(VT.isVector() && "Expected a vector type");
2928 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2929 // type. This ensures they get CSE'd.
2931 if (VT.getSizeInBits() == 64) { // MMX
2932 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2933 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2934 } else if (HasSSE2) { // SSE2
2935 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2936 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2938 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2939 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2941 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2944 /// getOnesVector - Returns a vector of specified type with all bits set.
2946 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2947 assert(VT.isVector() && "Expected a vector type");
2949 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2950 // type. This ensures they get CSE'd.
2951 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2953 if (VT.getSizeInBits() == 64) // MMX
2954 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2956 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2957 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2961 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2962 /// that point to V2 points to its first element.
2963 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2964 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2966 bool Changed = false;
2967 SmallVector<SDValue, 8> MaskVec;
2968 unsigned NumElems = Mask.getNumOperands();
2969 for (unsigned i = 0; i != NumElems; ++i) {
2970 SDValue Arg = Mask.getOperand(i);
2971 if (Arg.getOpcode() != ISD::UNDEF) {
2972 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2973 if (Val > NumElems) {
2974 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2978 MaskVec.push_back(Arg);
2982 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getDebugLoc(),
2983 Mask.getValueType(),
2984 &MaskVec[0], MaskVec.size());
2988 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2989 /// operation of specified width.
2990 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl) {
2991 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2992 MVT BaseVT = MaskVT.getVectorElementType();
2994 SmallVector<SDValue, 8> MaskVec;
2995 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2996 for (unsigned i = 1; i != NumElems; ++i)
2997 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2998 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
2999 &MaskVec[0], MaskVec.size());
3002 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
3003 /// of specified width.
3004 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG,
3006 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3007 MVT BaseVT = MaskVT.getVectorElementType();
3008 SmallVector<SDValue, 8> MaskVec;
3009 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3010 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3011 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3013 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3014 &MaskVec[0], MaskVec.size());
3017 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3018 /// of specified width.
3019 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG,
3021 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3022 MVT BaseVT = MaskVT.getVectorElementType();
3023 unsigned Half = NumElems/2;
3024 SmallVector<SDValue, 8> MaskVec;
3025 for (unsigned i = 0; i != Half; ++i) {
3026 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3027 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3029 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3030 &MaskVec[0], MaskVec.size());
3033 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
3034 /// element #0 of a vector with the specified index, leaving the rest of the
3035 /// elements in place.
3036 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
3037 SelectionDAG &DAG, DebugLoc dl) {
3038 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3039 MVT BaseVT = MaskVT.getVectorElementType();
3040 SmallVector<SDValue, 8> MaskVec;
3041 // Element #0 of the result gets the elt we are replacing.
3042 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
3043 for (unsigned i = 1; i != NumElems; ++i)
3044 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
3045 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3046 &MaskVec[0], MaskVec.size());
3049 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3050 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
3051 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
3052 MVT VT = Op.getValueType();
3055 SDValue V1 = Op.getOperand(0);
3056 SDValue Mask = Op.getOperand(2);
3057 unsigned MaskNumElems = Mask.getNumOperands();
3058 unsigned NumElems = MaskNumElems;
3059 DebugLoc dl = Op.getDebugLoc();
3060 // Special handling of v4f32 -> v4i32.
3061 if (VT != MVT::v4f32) {
3062 // Find which element we want to splat.
3063 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3064 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3065 // unpack elements to the correct location
3066 while (NumElems > 4) {
3067 if (EltNo < NumElems/2) {
3068 Mask = getUnpacklMask(MaskNumElems, DAG, dl);
3070 Mask = getUnpackhMask(MaskNumElems, DAG, dl);
3071 EltNo -= NumElems/2;
3073 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1, Mask);
3076 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3077 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3080 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3081 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3082 DAG.getUNDEF(PVT), Mask);
3083 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3086 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3087 /// load that's promoted to vector, or a load bitcasted.
3088 static bool isVectorLoad(SDValue Op) {
3089 assert(Op.getValueType().isVector() && "Expected a vector type");
3090 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3091 Op.getOpcode() == ISD::BIT_CONVERT) {
3092 return isa<LoadSDNode>(Op.getOperand(0));
3094 return isa<LoadSDNode>(Op);
3098 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3100 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3101 SelectionDAG &DAG, bool HasSSE3) {
3102 // If we have sse3 and shuffle has more than one use or input is a load, then
3103 // use movddup. Otherwise, use movlhps.
3104 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3105 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3106 MVT VT = Op.getValueType();
3109 DebugLoc dl = Op.getDebugLoc();
3110 unsigned NumElems = PVT.getVectorNumElements();
3111 if (NumElems == 2) {
3112 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3113 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3115 assert(NumElems == 4);
3116 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3117 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3118 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3119 Cst0, Cst1, Cst0, Cst1);
3122 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3123 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3124 DAG.getUNDEF(PVT), Mask);
3125 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3128 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3129 /// vector of zero or undef vector. This produces a shuffle where the low
3130 /// element of V2 is swizzled into the zero/undef vector, landing at element
3131 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3132 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3133 bool isZero, bool HasSSE2,
3134 SelectionDAG &DAG) {
3135 DebugLoc dl = V2.getDebugLoc();
3136 MVT VT = V2.getValueType();
3138 ? getZeroVector(VT, HasSSE2, DAG, dl) : DAG.getUNDEF(VT);
3139 unsigned NumElems = V2.getValueType().getVectorNumElements();
3140 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3141 MVT EVT = MaskVT.getVectorElementType();
3142 SmallVector<SDValue, 16> MaskVec;
3143 for (unsigned i = 0; i != NumElems; ++i)
3144 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3145 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3147 MaskVec.push_back(DAG.getConstant(i, EVT));
3148 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3149 &MaskVec[0], MaskVec.size());
3150 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
3153 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3154 /// a shuffle that is zero.
3156 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3157 unsigned NumElems, bool Low,
3158 SelectionDAG &DAG) {
3159 unsigned NumZeros = 0;
3160 for (unsigned i = 0; i < NumElems; ++i) {
3161 unsigned Index = Low ? i : NumElems-i-1;
3162 SDValue Idx = Mask.getOperand(Index);
3163 if (Idx.getOpcode() == ISD::UNDEF) {
3167 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3168 if (Elt.getNode() && isZeroNode(Elt))
3176 /// isVectorShift - Returns true if the shuffle can be implemented as a
3177 /// logical left or right shift of a vector.
3178 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3179 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3180 unsigned NumElems = Mask.getNumOperands();
3183 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3186 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3191 bool SeenV1 = false;
3192 bool SeenV2 = false;
3193 for (unsigned i = NumZeros; i < NumElems; ++i) {
3194 unsigned Val = isLeft ? (i - NumZeros) : i;
3195 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3196 if (Idx.getOpcode() == ISD::UNDEF)
3198 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3199 if (Index < NumElems)
3208 if (SeenV1 && SeenV2)
3211 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3217 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3219 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3220 unsigned NumNonZero, unsigned NumZero,
3221 SelectionDAG &DAG, TargetLowering &TLI) {
3225 DebugLoc dl = Op.getDebugLoc();
3228 for (unsigned i = 0; i < 16; ++i) {
3229 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3230 if (ThisIsNonZero && First) {
3232 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3234 V = DAG.getUNDEF(MVT::v8i16);
3239 SDValue ThisElt(0, 0), LastElt(0, 0);
3240 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3241 if (LastIsNonZero) {
3242 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3243 MVT::i16, Op.getOperand(i-1));
3245 if (ThisIsNonZero) {
3246 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3247 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3248 ThisElt, DAG.getConstant(8, MVT::i8));
3250 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3254 if (ThisElt.getNode())
3255 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3256 DAG.getIntPtrConstant(i/2));
3260 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3263 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3265 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3266 unsigned NumNonZero, unsigned NumZero,
3267 SelectionDAG &DAG, TargetLowering &TLI) {
3271 DebugLoc dl = Op.getDebugLoc();
3274 for (unsigned i = 0; i < 8; ++i) {
3275 bool isNonZero = (NonZeros & (1 << i)) != 0;
3279 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3281 V = DAG.getUNDEF(MVT::v8i16);
3284 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3285 MVT::v8i16, V, Op.getOperand(i),
3286 DAG.getIntPtrConstant(i));
3293 /// getVShift - Return a vector logical shift node.
3295 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3296 unsigned NumBits, SelectionDAG &DAG,
3297 const TargetLowering &TLI, DebugLoc dl) {
3298 bool isMMX = VT.getSizeInBits() == 64;
3299 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3300 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3301 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3302 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3303 DAG.getNode(Opc, dl, ShVT, SrcOp,
3304 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3308 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3309 DebugLoc dl = Op.getDebugLoc();
3310 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3311 if (ISD::isBuildVectorAllZeros(Op.getNode())
3312 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3313 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3314 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3315 // eliminated on x86-32 hosts.
3316 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3319 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3320 return getOnesVector(Op.getValueType(), DAG, dl);
3321 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3324 MVT VT = Op.getValueType();
3325 MVT EVT = VT.getVectorElementType();
3326 unsigned EVTBits = EVT.getSizeInBits();
3328 unsigned NumElems = Op.getNumOperands();
3329 unsigned NumZero = 0;
3330 unsigned NumNonZero = 0;
3331 unsigned NonZeros = 0;
3332 bool IsAllConstants = true;
3333 SmallSet<SDValue, 8> Values;
3334 for (unsigned i = 0; i < NumElems; ++i) {
3335 SDValue Elt = Op.getOperand(i);
3336 if (Elt.getOpcode() == ISD::UNDEF)
3339 if (Elt.getOpcode() != ISD::Constant &&
3340 Elt.getOpcode() != ISD::ConstantFP)
3341 IsAllConstants = false;
3342 if (isZeroNode(Elt))
3345 NonZeros |= (1 << i);
3350 if (NumNonZero == 0) {
3351 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3352 return DAG.getUNDEF(VT);
3355 // Special case for single non-zero, non-undef, element.
3356 if (NumNonZero == 1 && NumElems <= 4) {
3357 unsigned Idx = CountTrailingZeros_32(NonZeros);
3358 SDValue Item = Op.getOperand(Idx);
3360 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3361 // the value are obviously zero, truncate the value to i32 and do the
3362 // insertion that way. Only do this if the value is non-constant or if the
3363 // value is a constant being inserted into element 0. It is cheaper to do
3364 // a constant pool load than it is to do a movd + shuffle.
3365 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3366 (!IsAllConstants || Idx == 0)) {
3367 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3368 // Handle MMX and SSE both.
3369 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3370 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3372 // Truncate the value (which may itself be a constant) to i32, and
3373 // convert it to a vector with movd (S2V+shuffle to zero extend).
3374 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3375 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3376 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3377 Subtarget->hasSSE2(), DAG);
3379 // Now we have our 32-bit value zero extended in the low element of
3380 // a vector. If Idx != 0, swizzle it into place.
3383 Item, DAG.getUNDEF(Item.getValueType()),
3384 getSwapEltZeroMask(VecElts, Idx, DAG, dl)
3386 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VecVT, Ops, 3);
3388 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3392 // If we have a constant or non-constant insertion into the low element of
3393 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3394 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3395 // depending on what the source datatype is. Because we can only get here
3396 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3398 // Don't do this for i64 values on x86-32.
3399 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3400 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3401 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3402 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3403 Subtarget->hasSSE2(), DAG);
3406 // Is it a vector logical left shift?
3407 if (NumElems == 2 && Idx == 1 &&
3408 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3409 unsigned NumBits = VT.getSizeInBits();
3410 return getVShift(true, VT,
3411 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3412 VT, Op.getOperand(1)),
3413 NumBits/2, DAG, *this, dl);
3416 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3419 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3420 // is a non-constant being inserted into an element other than the low one,
3421 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3422 // movd/movss) to move this into the low element, then shuffle it into
3424 if (EVTBits == 32) {
3425 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3427 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3428 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3429 Subtarget->hasSSE2(), DAG);
3430 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3431 MVT MaskEVT = MaskVT.getVectorElementType();
3432 SmallVector<SDValue, 8> MaskVec;
3433 for (unsigned i = 0; i < NumElems; i++)
3434 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3435 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3436 &MaskVec[0], MaskVec.size());
3437 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, Item,
3438 DAG.getUNDEF(VT), Mask);
3442 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3443 if (Values.size() == 1)
3446 // A vector full of immediates; various special cases are already
3447 // handled, so this is best done with a single constant-pool load.
3451 // Let legalizer expand 2-wide build_vectors.
3452 if (EVTBits == 64) {
3453 if (NumNonZero == 1) {
3454 // One half is zero or undef.
3455 unsigned Idx = CountTrailingZeros_32(NonZeros);
3456 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3457 Op.getOperand(Idx));
3458 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3459 Subtarget->hasSSE2(), DAG);
3464 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3465 if (EVTBits == 8 && NumElems == 16) {
3466 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3468 if (V.getNode()) return V;
3471 if (EVTBits == 16 && NumElems == 8) {
3472 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3474 if (V.getNode()) return V;
3477 // If element VT is == 32 bits, turn it into a number of shuffles.
3478 SmallVector<SDValue, 8> V;
3480 if (NumElems == 4 && NumZero > 0) {
3481 for (unsigned i = 0; i < 4; ++i) {
3482 bool isZero = !(NonZeros & (1 << i));
3484 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3486 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3489 for (unsigned i = 0; i < 2; ++i) {
3490 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3493 V[i] = V[i*2]; // Must be a zero vector.
3496 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2+1], V[i*2],
3497 getMOVLMask(NumElems, DAG, dl));
3500 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3501 getMOVLMask(NumElems, DAG, dl));
3504 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3505 getUnpacklMask(NumElems, DAG, dl));
3510 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3511 MVT EVT = MaskVT.getVectorElementType();
3512 SmallVector<SDValue, 8> MaskVec;
3513 bool Reverse = (NonZeros & 0x3) == 2;
3514 for (unsigned i = 0; i < 2; ++i)
3516 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3518 MaskVec.push_back(DAG.getConstant(i, EVT));
3519 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3520 for (unsigned i = 0; i < 2; ++i)
3522 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3524 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3525 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3526 &MaskVec[0], MaskVec.size());
3527 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[0], V[1], ShufMask);
3530 if (Values.size() > 2) {
3531 // Expand into a number of unpckl*.
3533 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3534 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3535 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3536 SDValue UnpckMask = getUnpacklMask(NumElems, DAG, dl);
3537 for (unsigned i = 0; i < NumElems; ++i)
3538 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3540 while (NumElems != 0) {
3541 for (unsigned i = 0; i < NumElems; ++i)
3542 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i], V[i + NumElems],
3552 // v8i16 shuffles - Prefer shuffles in the following order:
3553 // 1. [all] pshuflw, pshufhw, optional move
3554 // 2. [ssse3] 1 x pshufb
3555 // 3. [ssse3] 2 x pshufb + 1 x por
3556 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3558 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3559 SDValue PermMask, SelectionDAG &DAG,
3560 X86TargetLowering &TLI, DebugLoc dl) {
3561 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3562 PermMask.getNode()->op_end());
3563 SmallVector<int, 8> MaskVals;
3565 // Determine if more than 1 of the words in each of the low and high quadwords
3566 // of the result come from the same quadword of one of the two inputs. Undef
3567 // mask values count as coming from any quadword, for better codegen.
3568 SmallVector<unsigned, 4> LoQuad(4);
3569 SmallVector<unsigned, 4> HiQuad(4);
3570 BitVector InputQuads(4);
3571 for (unsigned i = 0; i < 8; ++i) {
3572 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3573 SDValue Elt = MaskElts[i];
3574 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3575 cast<ConstantSDNode>(Elt)->getZExtValue();
3576 MaskVals.push_back(EltIdx);
3585 InputQuads.set(EltIdx / 4);
3588 int BestLoQuad = -1;
3589 unsigned MaxQuad = 1;
3590 for (unsigned i = 0; i < 4; ++i) {
3591 if (LoQuad[i] > MaxQuad) {
3593 MaxQuad = LoQuad[i];
3597 int BestHiQuad = -1;
3599 for (unsigned i = 0; i < 4; ++i) {
3600 if (HiQuad[i] > MaxQuad) {
3602 MaxQuad = HiQuad[i];
3606 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3607 // of the two input vectors, shuffle them into one input vector so only a
3608 // single pshufb instruction is necessary. If There are more than 2 input
3609 // quads, disable the next transformation since it does not help SSSE3.
3610 bool V1Used = InputQuads[0] || InputQuads[1];
3611 bool V2Used = InputQuads[2] || InputQuads[3];
3612 if (TLI.getSubtarget()->hasSSSE3()) {
3613 if (InputQuads.count() == 2 && V1Used && V2Used) {
3614 BestLoQuad = InputQuads.find_first();
3615 BestHiQuad = InputQuads.find_next(BestLoQuad);
3617 if (InputQuads.count() > 2) {
3623 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3624 // the shuffle mask. If a quad is scored as -1, that means that it contains
3625 // words from all 4 input quadwords.
3627 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3628 SmallVector<SDValue,8> MaskV;
3629 MaskV.push_back(DAG.getConstant(BestLoQuad < 0 ? 0 : BestLoQuad, MVT::i64));
3630 MaskV.push_back(DAG.getConstant(BestHiQuad < 0 ? 1 : BestHiQuad, MVT::i64));
3631 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, &MaskV[0], 2);
3633 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
3634 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3635 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), Mask);
3636 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3638 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3639 // source words for the shuffle, to aid later transformations.
3640 bool AllWordsInNewV = true;
3641 bool InOrder[2] = { true, true };
3642 for (unsigned i = 0; i != 8; ++i) {
3643 int idx = MaskVals[i];
3645 InOrder[i/4] = false;
3646 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3648 AllWordsInNewV = false;
3652 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3653 if (AllWordsInNewV) {
3654 for (int i = 0; i != 8; ++i) {
3655 int idx = MaskVals[i];
3658 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3659 if ((idx != i) && idx < 4)
3661 if ((idx != i) && idx > 3)
3670 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3671 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3672 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3674 for (unsigned i = 0; i != 8; ++i)
3675 MaskV.push_back((MaskVals[i] < 0) ? DAG.getUNDEF(MVT::i16)
3676 : DAG.getConstant(MaskVals[i],
3678 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3679 DAG.getUNDEF(MVT::v8i16),
3680 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i16,
3685 // If we have SSSE3, and all words of the result are from 1 input vector,
3686 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3687 // is present, fall back to case 4.
3688 if (TLI.getSubtarget()->hasSSSE3()) {
3689 SmallVector<SDValue,16> pshufbMask;
3691 // If we have elements from both input vectors, set the high bit of the
3692 // shuffle mask element to zero out elements that come from V2 in the V1
3693 // mask, and elements that come from V1 in the V2 mask, so that the two
3694 // results can be OR'd together.
3695 bool TwoInputs = V1Used && V2Used;
3696 for (unsigned i = 0; i != 8; ++i) {
3697 int EltIdx = MaskVals[i] * 2;
3698 if (TwoInputs && (EltIdx >= 16)) {
3699 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3700 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3703 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3704 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3706 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3707 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3708 DAG.getNode(ISD::BUILD_VECTOR, dl,
3709 MVT::v16i8, &pshufbMask[0], 16));
3711 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3713 // Calculate the shuffle mask for the second input, shuffle it, and
3714 // OR it with the first shuffled input.
3716 for (unsigned i = 0; i != 8; ++i) {
3717 int EltIdx = MaskVals[i] * 2;
3719 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3720 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3723 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3724 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3726 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3727 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3728 DAG.getNode(ISD::BUILD_VECTOR, dl,
3729 MVT::v16i8, &pshufbMask[0], 16));
3730 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3731 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3734 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3735 // and update MaskVals with new element order.
3736 BitVector InOrder(8);
3737 if (BestLoQuad >= 0) {
3738 SmallVector<SDValue, 8> MaskV;
3739 for (int i = 0; i != 4; ++i) {
3740 int idx = MaskVals[i];
3742 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3744 } else if ((idx / 4) == BestLoQuad) {
3745 MaskV.push_back(DAG.getConstant(idx & 3, MVT::i16));
3748 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3751 for (unsigned i = 4; i != 8; ++i)
3752 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3753 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3754 DAG.getUNDEF(MVT::v8i16),
3755 DAG.getNode(ISD::BUILD_VECTOR, dl,
3756 MVT::v8i16, &MaskV[0], 8));
3759 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3760 // and update MaskVals with the new element order.
3761 if (BestHiQuad >= 0) {
3762 SmallVector<SDValue, 8> MaskV;
3763 for (unsigned i = 0; i != 4; ++i)
3764 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3765 for (unsigned i = 4; i != 8; ++i) {
3766 int idx = MaskVals[i];
3768 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3770 } else if ((idx / 4) == BestHiQuad) {
3771 MaskV.push_back(DAG.getConstant((idx & 3) + 4, MVT::i16));
3774 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3777 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3778 DAG.getUNDEF(MVT::v8i16),
3779 DAG.getNode(ISD::BUILD_VECTOR, dl,
3780 MVT::v8i16, &MaskV[0], 8));
3783 // In case BestHi & BestLo were both -1, which means each quadword has a word
3784 // from each of the four input quadwords, calculate the InOrder bitvector now
3785 // before falling through to the insert/extract cleanup.
3786 if (BestLoQuad == -1 && BestHiQuad == -1) {
3788 for (int i = 0; i != 8; ++i)
3789 if (MaskVals[i] < 0 || MaskVals[i] == i)
3793 // The other elements are put in the right place using pextrw and pinsrw.
3794 for (unsigned i = 0; i != 8; ++i) {
3797 int EltIdx = MaskVals[i];
3800 SDValue ExtOp = (EltIdx < 8)
3801 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3802 DAG.getIntPtrConstant(EltIdx))
3803 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3804 DAG.getIntPtrConstant(EltIdx - 8));
3805 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3806 DAG.getIntPtrConstant(i));
3811 // v16i8 shuffles - Prefer shuffles in the following order:
3812 // 1. [ssse3] 1 x pshufb
3813 // 2. [ssse3] 2 x pshufb + 1 x por
3814 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3816 SDValue LowerVECTOR_SHUFFLEv16i8(SDValue V1, SDValue V2,
3817 SDValue PermMask, SelectionDAG &DAG,
3818 X86TargetLowering &TLI, DebugLoc dl) {
3819 SmallVector<SDValue, 16> MaskElts(PermMask.getNode()->op_begin(),
3820 PermMask.getNode()->op_end());
3821 SmallVector<int, 16> MaskVals;
3823 // If we have SSSE3, case 1 is generated when all result bytes come from
3824 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3825 // present, fall back to case 3.
3826 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3829 for (unsigned i = 0; i < 16; ++i) {
3830 SDValue Elt = MaskElts[i];
3831 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3832 cast<ConstantSDNode>(Elt)->getZExtValue();
3833 MaskVals.push_back(EltIdx);
3842 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3843 if (TLI.getSubtarget()->hasSSSE3()) {
3844 SmallVector<SDValue,16> pshufbMask;
3846 // If all result elements are from one input vector, then only translate
3847 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3849 // Otherwise, we have elements from both input vectors, and must zero out
3850 // elements that come from V2 in the first mask, and V1 in the second mask
3851 // so that we can OR them together.
3852 bool TwoInputs = !(V1Only || V2Only);
3853 for (unsigned i = 0; i != 16; ++i) {
3854 int EltIdx = MaskVals[i];
3855 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3856 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3859 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3861 // If all the elements are from V2, assign it to V1 and return after
3862 // building the first pshufb.
3865 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3866 DAG.getNode(ISD::BUILD_VECTOR, dl,
3867 MVT::v16i8, &pshufbMask[0], 16));
3871 // Calculate the shuffle mask for the second input, shuffle it, and
3872 // OR it with the first shuffled input.
3874 for (unsigned i = 0; i != 16; ++i) {
3875 int EltIdx = MaskVals[i];
3877 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3880 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3882 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3883 DAG.getNode(ISD::BUILD_VECTOR, dl,
3884 MVT::v16i8, &pshufbMask[0], 16));
3885 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3888 // No SSSE3 - Calculate in place words and then fix all out of place words
3889 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3890 // the 16 different words that comprise the two doublequadword input vectors.
3891 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3892 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3893 SDValue NewV = V2Only ? V2 : V1;
3894 for (int i = 0; i != 8; ++i) {
3895 int Elt0 = MaskVals[i*2];
3896 int Elt1 = MaskVals[i*2+1];
3898 // This word of the result is all undef, skip it.
3899 if (Elt0 < 0 && Elt1 < 0)
3902 // This word of the result is already in the correct place, skip it.
3903 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3905 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3908 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3909 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3912 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3913 // using a single extract together, load it and store it.
3914 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3915 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3916 DAG.getIntPtrConstant(Elt1 / 2));
3917 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3918 DAG.getIntPtrConstant(i));
3922 // If Elt1 is defined, extract it from the appropriate source. If the
3923 // source byte is not also odd, shift the extracted word left 8 bits
3924 // otherwise clear the bottom 8 bits if we need to do an or.
3926 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3927 DAG.getIntPtrConstant(Elt1 / 2));
3928 if ((Elt1 & 1) == 0)
3929 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3930 DAG.getConstant(8, TLI.getShiftAmountTy()));
3932 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3933 DAG.getConstant(0xFF00, MVT::i16));
3935 // If Elt0 is defined, extract it from the appropriate source. If the
3936 // source byte is not also even, shift the extracted word right 8 bits. If
3937 // Elt1 was also defined, OR the extracted values together before
3938 // inserting them in the result.
3940 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3941 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3942 if ((Elt0 & 1) != 0)
3943 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3944 DAG.getConstant(8, TLI.getShiftAmountTy()));
3946 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3947 DAG.getConstant(0x00FF, MVT::i16));
3948 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3951 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3952 DAG.getIntPtrConstant(i));
3954 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3957 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3958 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3959 /// done when every pair / quad of shuffle mask elements point to elements in
3960 /// the right sequence. e.g.
3961 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3963 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3965 SDValue PermMask, SelectionDAG &DAG,
3966 TargetLowering &TLI, DebugLoc dl) {
3967 unsigned NumElems = PermMask.getNumOperands();
3968 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3969 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3970 MVT MaskEltVT = MaskVT.getVectorElementType();
3972 switch (VT.getSimpleVT()) {
3973 default: assert(false && "Unexpected!");
3974 case MVT::v4f32: NewVT = MVT::v2f64; break;
3975 case MVT::v4i32: NewVT = MVT::v2i64; break;
3976 case MVT::v8i16: NewVT = MVT::v4i32; break;
3977 case MVT::v16i8: NewVT = MVT::v4i32; break;
3980 if (NewWidth == 2) {
3986 unsigned Scale = NumElems / NewWidth;
3987 SmallVector<SDValue, 8> MaskVec;
3988 for (unsigned i = 0; i < NumElems; i += Scale) {
3989 unsigned StartIdx = ~0U;
3990 for (unsigned j = 0; j < Scale; ++j) {
3991 SDValue Elt = PermMask.getOperand(i+j);
3992 if (Elt.getOpcode() == ISD::UNDEF)
3994 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3995 if (StartIdx == ~0U)
3996 StartIdx = EltIdx - (EltIdx % Scale);
3997 if (EltIdx != StartIdx + j)
4000 if (StartIdx == ~0U)
4001 MaskVec.push_back(DAG.getUNDEF(MaskEltVT));
4003 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
4006 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4007 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4008 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, NewVT, V1, V2,
4009 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4010 &MaskVec[0], MaskVec.size()));
4013 /// getVZextMovL - Return a zero-extending vector move low node.
4015 static SDValue getVZextMovL(MVT VT, MVT OpVT,
4016 SDValue SrcOp, SelectionDAG &DAG,
4017 const X86Subtarget *Subtarget, DebugLoc dl) {
4018 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4019 LoadSDNode *LD = NULL;
4020 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4021 LD = dyn_cast<LoadSDNode>(SrcOp);
4023 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4025 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4026 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
4027 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4028 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4029 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
4031 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4032 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4033 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4034 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4042 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4043 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4044 DAG.getNode(ISD::BIT_CONVERT, dl,
4048 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4051 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
4052 SDValue PermMask, MVT VT, SelectionDAG &DAG,
4054 MVT MaskVT = PermMask.getValueType();
4055 MVT MaskEVT = MaskVT.getVectorElementType();
4056 SmallVector<std::pair<int, int>, 8> Locs;
4058 SmallVector<SDValue, 8> Mask1(4, DAG.getUNDEF(MaskEVT));
4061 for (unsigned i = 0; i != 4; ++i) {
4062 SDValue Elt = PermMask.getOperand(i);
4063 if (Elt.getOpcode() == ISD::UNDEF) {
4064 Locs[i] = std::make_pair(-1, -1);
4066 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4067 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
4069 Locs[i] = std::make_pair(0, NumLo);
4073 Locs[i] = std::make_pair(1, NumHi);
4075 Mask1[2+NumHi] = Elt;
4081 if (NumLo <= 2 && NumHi <= 2) {
4082 // If no more than two elements come from either vector. This can be
4083 // implemented with two shuffles. First shuffle gather the elements.
4084 // The second shuffle, which takes the first shuffle as both of its
4085 // vector operands, put the elements into the right order.
4086 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4087 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4088 &Mask1[0], Mask1.size()));
4090 SmallVector<SDValue, 8> Mask2(4, DAG.getUNDEF(MaskEVT));
4091 for (unsigned i = 0; i != 4; ++i) {
4092 if (Locs[i].first == -1)
4095 unsigned Idx = (i < 2) ? 0 : 4;
4096 Idx += Locs[i].first * 2 + Locs[i].second;
4097 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
4101 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1,
4102 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4103 &Mask2[0], Mask2.size()));
4104 } else if (NumLo == 3 || NumHi == 3) {
4105 // Otherwise, we must have three elements from one vector, call it X, and
4106 // one element from the other, call it Y. First, use a shufps to build an
4107 // intermediate vector with the one element from Y and the element from X
4108 // that will be in the same half in the final destination (the indexes don't
4109 // matter). Then, use a shufps to build the final vector, taking the half
4110 // containing the element from Y from the intermediate, and the other half
4113 // Normalize it so the 3 elements come from V1.
4114 PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
4118 // Find the element from V2.
4120 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4121 SDValue Elt = PermMask.getOperand(HiIndex);
4122 if (Elt.getOpcode() == ISD::UNDEF)
4124 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4129 Mask1[0] = PermMask.getOperand(HiIndex);
4130 Mask1[1] = DAG.getUNDEF(MaskEVT);
4131 Mask1[2] = PermMask.getOperand(HiIndex^1);
4132 Mask1[3] = DAG.getUNDEF(MaskEVT);
4133 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4134 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &Mask1[0], 4));
4137 Mask1[0] = PermMask.getOperand(0);
4138 Mask1[1] = PermMask.getOperand(1);
4139 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
4140 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
4141 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4142 DAG.getNode(ISD::BUILD_VECTOR, dl,
4143 MaskVT, &Mask1[0], 4));
4145 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
4146 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
4147 Mask1[2] = PermMask.getOperand(2);
4148 Mask1[3] = PermMask.getOperand(3);
4149 if (Mask1[2].getOpcode() != ISD::UNDEF)
4151 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
4153 if (Mask1[3].getOpcode() != ISD::UNDEF)
4155 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
4157 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V2, V1,
4158 DAG.getNode(ISD::BUILD_VECTOR, dl,
4159 MaskVT, &Mask1[0], 4));
4163 // Break it into (shuffle shuffle_hi, shuffle_lo).
4165 SmallVector<SDValue,8> LoMask(4, DAG.getUNDEF(MaskEVT));
4166 SmallVector<SDValue,8> HiMask(4, DAG.getUNDEF(MaskEVT));
4167 SmallVector<SDValue,8> *MaskPtr = &LoMask;
4168 unsigned MaskIdx = 0;
4171 for (unsigned i = 0; i != 4; ++i) {
4178 SDValue Elt = PermMask.getOperand(i);
4179 if (Elt.getOpcode() == ISD::UNDEF) {
4180 Locs[i] = std::make_pair(-1, -1);
4181 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
4182 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4183 (*MaskPtr)[LoIdx] = Elt;
4186 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4187 (*MaskPtr)[HiIdx] = Elt;
4192 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4193 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4194 &LoMask[0], LoMask.size()));
4195 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4196 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4197 &HiMask[0], HiMask.size()));
4198 SmallVector<SDValue, 8> MaskOps;
4199 for (unsigned i = 0; i != 4; ++i) {
4200 if (Locs[i].first == -1) {
4201 MaskOps.push_back(DAG.getUNDEF(MaskEVT));
4203 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4204 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
4207 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, LoShuffle, HiShuffle,
4208 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4209 &MaskOps[0], MaskOps.size()));
4213 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4214 SDValue V1 = Op.getOperand(0);
4215 SDValue V2 = Op.getOperand(1);
4216 SDValue PermMask = Op.getOperand(2);
4217 MVT VT = Op.getValueType();
4218 DebugLoc dl = Op.getDebugLoc();
4219 unsigned NumElems = PermMask.getNumOperands();
4220 bool isMMX = VT.getSizeInBits() == 64;
4221 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4222 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4223 bool V1IsSplat = false;
4224 bool V2IsSplat = false;
4226 // FIXME: Check for legal shuffle and return?
4228 if (isUndefShuffle(Op.getNode()))
4229 return DAG.getUNDEF(VT);
4231 if (isZeroShuffle(Op.getNode()))
4232 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4234 if (isIdentityMask(PermMask.getNode()))
4236 else if (isIdentityMask(PermMask.getNode(), true))
4239 // Canonicalize movddup shuffles.
4240 if (V2IsUndef && Subtarget->hasSSE2() &&
4241 VT.getSizeInBits() == 128 &&
4242 X86::isMOVDDUPMask(PermMask.getNode()))
4243 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4245 if (isSplatMask(PermMask.getNode())) {
4246 if (isMMX || NumElems < 4) return Op;
4247 // Promote it to a v4{if}32 splat.
4248 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4251 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4253 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4254 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG,
4256 if (NewOp.getNode())
4257 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4258 LowerVECTOR_SHUFFLE(NewOp, DAG));
4259 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4260 // FIXME: Figure out a cleaner way to do this.
4261 // Try to make use of movq to zero out the top part.
4262 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4263 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4265 if (NewOp.getNode()) {
4266 SDValue NewV1 = NewOp.getOperand(0);
4267 SDValue NewV2 = NewOp.getOperand(1);
4268 SDValue NewMask = NewOp.getOperand(2);
4269 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4270 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4271 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget,
4275 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4276 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4278 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4279 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4280 DAG, Subtarget, dl);
4284 // Check if this can be converted into a logical shift.
4285 bool isLeft = false;
4288 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4289 if (isShift && ShVal.hasOneUse()) {
4290 // If the shifted value has multiple uses, it may be cheaper to use
4291 // v_set0 + movlhps or movhlps, etc.
4292 MVT EVT = VT.getVectorElementType();
4293 ShAmt *= EVT.getSizeInBits();
4294 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4297 if (X86::isMOVLMask(PermMask.getNode())) {
4300 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4301 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4306 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4307 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4308 X86::isMOVHLPSMask(PermMask.getNode()) ||
4309 X86::isMOVHPMask(PermMask.getNode()) ||
4310 X86::isMOVLPMask(PermMask.getNode())))
4313 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4314 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4315 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4318 // No better options. Use a vshl / vsrl.
4319 MVT EVT = VT.getVectorElementType();
4320 ShAmt *= EVT.getSizeInBits();
4321 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4324 bool Commuted = false;
4325 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4326 // 1,1,1,1 -> v8i16 though.
4327 V1IsSplat = isSplatVector(V1.getNode());
4328 V2IsSplat = isSplatVector(V2.getNode());
4330 // Canonicalize the splat or undef, if present, to be on the RHS.
4331 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4332 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4333 std::swap(V1IsSplat, V2IsSplat);
4334 std::swap(V1IsUndef, V2IsUndef);
4338 // FIXME: Figure out a cleaner way to do this.
4339 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4340 if (V2IsUndef) return V1;
4341 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4343 // V2 is a splat, so the mask may be malformed. That is, it may point
4344 // to any V2 element. The instruction selectior won't like this. Get
4345 // a corrected mask and commute to form a proper MOVS{S|D}.
4346 SDValue NewMask = getMOVLMask(NumElems, DAG, dl);
4347 if (NewMask.getNode() != PermMask.getNode())
4348 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4353 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4354 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4355 X86::isUNPCKLMask(PermMask.getNode()) ||
4356 X86::isUNPCKHMask(PermMask.getNode()))
4360 // Normalize mask so all entries that point to V2 points to its first
4361 // element then try to match unpck{h|l} again. If match, return a
4362 // new vector_shuffle with the corrected mask.
4363 SDValue NewMask = NormalizeMask(PermMask, DAG);
4364 if (NewMask.getNode() != PermMask.getNode()) {
4365 if (X86::isUNPCKLMask(NewMask.getNode(), true)) {
4366 SDValue NewMask = getUnpacklMask(NumElems, DAG, dl);
4367 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4368 } else if (X86::isUNPCKHMask(NewMask.getNode(), true)) {
4369 SDValue NewMask = getUnpackhMask(NumElems, DAG, dl);
4370 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4375 // Normalize the node to match x86 shuffle ops if needed
4376 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4377 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4380 // Commute is back and try unpck* again.
4381 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4382 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4383 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4384 X86::isUNPCKLMask(PermMask.getNode()) ||
4385 X86::isUNPCKHMask(PermMask.getNode()))
4389 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4390 // Try PSHUF* first, then SHUFP*.
4391 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4392 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4393 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4394 if (V2.getOpcode() != ISD::UNDEF)
4395 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1,
4396 DAG.getUNDEF(VT), PermMask);
4401 if (Subtarget->hasSSE2() &&
4402 (X86::isPSHUFDMask(PermMask.getNode()) ||
4403 X86::isPSHUFHWMask(PermMask.getNode()) ||
4404 X86::isPSHUFLWMask(PermMask.getNode()))) {
4406 if (VT == MVT::v4f32) {
4408 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT,
4409 DAG.getNode(ISD::BIT_CONVERT, dl, RVT, V1),
4410 DAG.getUNDEF(RVT), PermMask);
4411 } else if (V2.getOpcode() != ISD::UNDEF)
4412 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT, V1,
4413 DAG.getUNDEF(RVT), PermMask);
4415 Op = DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
4419 // Binary or unary shufps.
4420 if (X86::isSHUFPMask(PermMask.getNode()) ||
4421 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4425 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4426 if (VT == MVT::v8i16) {
4427 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this, dl);
4428 if (NewOp.getNode())
4432 if (VT == MVT::v16i8) {
4433 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(V1, V2, PermMask, DAG, *this, dl);
4434 if (NewOp.getNode())
4438 // Handle all 4 wide cases with a number of shuffles except for MMX.
4439 if (NumElems == 4 && !isMMX)
4440 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG, dl);
4446 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4447 SelectionDAG &DAG) {
4448 MVT VT = Op.getValueType();
4449 DebugLoc dl = Op.getDebugLoc();
4450 if (VT.getSizeInBits() == 8) {
4451 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4452 Op.getOperand(0), Op.getOperand(1));
4453 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4454 DAG.getValueType(VT));
4455 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4456 } else if (VT.getSizeInBits() == 16) {
4457 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4458 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4460 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4461 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4462 DAG.getNode(ISD::BIT_CONVERT, dl,
4466 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4467 Op.getOperand(0), Op.getOperand(1));
4468 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4469 DAG.getValueType(VT));
4470 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4471 } else if (VT == MVT::f32) {
4472 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4473 // the result back to FR32 register. It's only worth matching if the
4474 // result has a single use which is a store or a bitcast to i32. And in
4475 // the case of a store, it's not worth it if the index is a constant 0,
4476 // because a MOVSSmr can be used instead, which is smaller and faster.
4477 if (!Op.hasOneUse())
4479 SDNode *User = *Op.getNode()->use_begin();
4480 if ((User->getOpcode() != ISD::STORE ||
4481 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4482 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4483 (User->getOpcode() != ISD::BIT_CONVERT ||
4484 User->getValueType(0) != MVT::i32))
4486 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4487 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4490 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4491 } else if (VT == MVT::i32) {
4492 // ExtractPS works with constant index.
4493 if (isa<ConstantSDNode>(Op.getOperand(1)))
4501 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4502 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4505 if (Subtarget->hasSSE41()) {
4506 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4511 MVT VT = Op.getValueType();
4512 DebugLoc dl = Op.getDebugLoc();
4513 // TODO: handle v16i8.
4514 if (VT.getSizeInBits() == 16) {
4515 SDValue Vec = Op.getOperand(0);
4516 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4518 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4519 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4520 DAG.getNode(ISD::BIT_CONVERT, dl,
4523 // Transform it so it match pextrw which produces a 32-bit result.
4524 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4525 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4526 Op.getOperand(0), Op.getOperand(1));
4527 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4528 DAG.getValueType(VT));
4529 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4530 } else if (VT.getSizeInBits() == 32) {
4531 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4534 // SHUFPS the element to the lowest double word, then movss.
4535 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4536 SmallVector<SDValue, 8> IdxVec;
4538 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4540 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4542 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4544 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4545 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4546 &IdxVec[0], IdxVec.size());
4547 SDValue Vec = Op.getOperand(0);
4548 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4549 Vec, DAG.getUNDEF(Vec.getValueType()), Mask);
4550 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4551 DAG.getIntPtrConstant(0));
4552 } else if (VT.getSizeInBits() == 64) {
4553 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4554 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4555 // to match extract_elt for f64.
4556 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4560 // UNPCKHPD the element to the lowest double word, then movsd.
4561 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4562 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4563 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4564 SmallVector<SDValue, 8> IdxVec;
4565 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4567 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4568 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4569 &IdxVec[0], IdxVec.size());
4570 SDValue Vec = Op.getOperand(0);
4571 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4572 Vec, DAG.getUNDEF(Vec.getValueType()),
4574 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4575 DAG.getIntPtrConstant(0));
4582 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4583 MVT VT = Op.getValueType();
4584 MVT EVT = VT.getVectorElementType();
4585 DebugLoc dl = Op.getDebugLoc();
4587 SDValue N0 = Op.getOperand(0);
4588 SDValue N1 = Op.getOperand(1);
4589 SDValue N2 = Op.getOperand(2);
4591 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4592 isa<ConstantSDNode>(N2)) {
4593 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4595 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4597 if (N1.getValueType() != MVT::i32)
4598 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4599 if (N2.getValueType() != MVT::i32)
4600 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4601 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4602 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4603 // Bits [7:6] of the constant are the source select. This will always be
4604 // zero here. The DAG Combiner may combine an extract_elt index into these
4605 // bits. For example (insert (extract, 3), 2) could be matched by putting
4606 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4607 // Bits [5:4] of the constant are the destination select. This is the
4608 // value of the incoming immediate.
4609 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4610 // combine either bitwise AND or insert of float 0.0 to set these bits.
4611 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4612 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4613 } else if (EVT == MVT::i32) {
4614 // InsertPS works with constant index.
4615 if (isa<ConstantSDNode>(N2))
4622 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4623 MVT VT = Op.getValueType();
4624 MVT EVT = VT.getVectorElementType();
4626 if (Subtarget->hasSSE41())
4627 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4632 DebugLoc dl = Op.getDebugLoc();
4633 SDValue N0 = Op.getOperand(0);
4634 SDValue N1 = Op.getOperand(1);
4635 SDValue N2 = Op.getOperand(2);
4637 if (EVT.getSizeInBits() == 16) {
4638 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4639 // as its second argument.
4640 if (N1.getValueType() != MVT::i32)
4641 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4642 if (N2.getValueType() != MVT::i32)
4643 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4644 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4650 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4651 DebugLoc dl = Op.getDebugLoc();
4652 if (Op.getValueType() == MVT::v2f32)
4653 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4654 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4655 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4656 Op.getOperand(0))));
4658 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4659 MVT VT = MVT::v2i32;
4660 switch (Op.getValueType().getSimpleVT()) {
4667 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4668 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4671 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4672 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4673 // one of the above mentioned nodes. It has to be wrapped because otherwise
4674 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4675 // be used to form addressing mode. These wrapped nodes will be selected
4678 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4679 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4680 // FIXME there isn't really any debug info here, should come from the parent
4681 DebugLoc dl = CP->getDebugLoc();
4682 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4683 CP->getAlignment());
4684 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4685 // With PIC, the address is actually $g + Offset.
4686 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4687 !Subtarget->isPICStyleRIPRel()) {
4688 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4689 DAG.getNode(X86ISD::GlobalBaseReg,
4690 DebugLoc::getUnknownLoc(),
4699 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4701 SelectionDAG &DAG) const {
4702 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4703 bool ExtraLoadRequired =
4704 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4706 // Create the TargetGlobalAddress node, folding in the constant
4707 // offset if it is legal.
4709 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4710 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4713 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4714 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4716 // With PIC, the address is actually $g + Offset.
4717 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4718 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4719 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4723 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4724 // load the value at address GV, not the value of GV itself. This means that
4725 // the GlobalAddress must be in the base or index register of the address, not
4726 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4727 // The same applies for external symbols during PIC codegen
4728 if (ExtraLoadRequired)
4729 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4730 PseudoSourceValue::getGOT(), 0);
4732 // If there was a non-zero offset that we didn't fold, create an explicit
4735 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4736 DAG.getConstant(Offset, getPointerTy()));
4742 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4743 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4744 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4745 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4748 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4750 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4753 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4754 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4755 DAG.getNode(X86ISD::GlobalBaseReg,
4756 DebugLoc::getUnknownLoc(),
4758 InFlag = Chain.getValue(1);
4760 // emit leal symbol@TLSGD(,%ebx,1), %eax
4761 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4762 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4763 GA->getValueType(0),
4765 SDValue Ops[] = { Chain, TGA, InFlag };
4766 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4767 InFlag = Result.getValue(2);
4768 Chain = Result.getValue(1);
4770 // call ___tls_get_addr. This function receives its argument in
4771 // the register EAX.
4772 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Result, InFlag);
4773 InFlag = Chain.getValue(1);
4775 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4776 SDValue Ops1[] = { Chain,
4777 DAG.getTargetExternalSymbol("___tls_get_addr",
4779 DAG.getRegister(X86::EAX, PtrVT),
4780 DAG.getRegister(X86::EBX, PtrVT),
4782 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
4783 InFlag = Chain.getValue(1);
4785 return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
4788 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4790 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4792 SDValue InFlag, Chain;
4793 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4795 // emit leaq symbol@TLSGD(%rip), %rdi
4796 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4797 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4798 GA->getValueType(0),
4800 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4801 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4802 Chain = Result.getValue(1);
4803 InFlag = Result.getValue(2);
4805 // call __tls_get_addr. This function receives its argument in
4806 // the register RDI.
4807 Chain = DAG.getCopyToReg(Chain, dl, X86::RDI, Result, InFlag);
4808 InFlag = Chain.getValue(1);
4810 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4811 SDValue Ops1[] = { Chain,
4812 DAG.getTargetExternalSymbol("__tls_get_addr",
4814 DAG.getRegister(X86::RDI, PtrVT),
4816 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
4817 InFlag = Chain.getValue(1);
4819 return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
4822 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4823 // "local exec" model.
4824 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4825 const MVT PtrVT, TLSModel::Model model) {
4826 DebugLoc dl = GA->getDebugLoc();
4827 // Get the Thread Pointer
4828 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER,
4829 DebugLoc::getUnknownLoc(), PtrVT);
4830 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4832 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4833 GA->getValueType(0),
4835 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4837 if (model == TLSModel::InitialExec)
4838 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4839 PseudoSourceValue::getGOT(), 0);
4841 // The address of the thread local variable is the add of the thread
4842 // pointer with the offset of the variable.
4843 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4847 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4848 // TODO: implement the "local dynamic" model
4849 // TODO: implement the "initial exec"model for pic executables
4850 assert(Subtarget->isTargetELF() &&
4851 "TLS not implemented for non-ELF targets");
4852 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4853 GlobalValue *GV = GA->getGlobal();
4854 TLSModel::Model model =
4855 getTLSModel (GV, getTargetMachine().getRelocationModel());
4856 if (Subtarget->is64Bit()) {
4858 case TLSModel::GeneralDynamic:
4859 case TLSModel::LocalDynamic: // not implemented
4860 case TLSModel::InitialExec: // not implemented
4861 case TLSModel::LocalExec: // not implemented
4862 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4864 assert (0 && "Unknown TLS model");
4868 case TLSModel::GeneralDynamic:
4869 case TLSModel::LocalDynamic: // not implemented
4870 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4872 case TLSModel::InitialExec:
4873 case TLSModel::LocalExec:
4874 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model);
4876 assert (0 && "Unknown TLS model");
4882 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4883 // FIXME there isn't really any debug info here
4884 DebugLoc dl = Op.getDebugLoc();
4885 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4886 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4887 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4888 // With PIC, the address is actually $g + Offset.
4889 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4890 !Subtarget->isPICStyleRIPRel()) {
4891 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4892 DAG.getNode(X86ISD::GlobalBaseReg,
4893 DebugLoc::getUnknownLoc(),
4901 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4902 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4903 // FIXME there isn't really any debug into here
4904 DebugLoc dl = JT->getDebugLoc();
4905 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4906 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4907 // With PIC, the address is actually $g + Offset.
4908 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4909 !Subtarget->isPICStyleRIPRel()) {
4910 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4911 DAG.getNode(X86ISD::GlobalBaseReg,
4912 DebugLoc::getUnknownLoc(),
4920 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4921 /// take a 2 x i32 value to shift plus a shift amount.
4922 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4923 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4924 MVT VT = Op.getValueType();
4925 unsigned VTBits = VT.getSizeInBits();
4926 DebugLoc dl = Op.getDebugLoc();
4927 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4928 SDValue ShOpLo = Op.getOperand(0);
4929 SDValue ShOpHi = Op.getOperand(1);
4930 SDValue ShAmt = Op.getOperand(2);
4931 SDValue Tmp1 = isSRA ?
4932 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4933 DAG.getConstant(VTBits - 1, MVT::i8)) :
4934 DAG.getConstant(0, VT);
4937 if (Op.getOpcode() == ISD::SHL_PARTS) {
4938 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4939 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4941 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4942 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4945 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4946 DAG.getConstant(VTBits, MVT::i8));
4947 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4948 AndNode, DAG.getConstant(0, MVT::i8));
4951 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4952 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4953 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4955 if (Op.getOpcode() == ISD::SHL_PARTS) {
4956 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4957 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4959 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4960 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4963 SDValue Ops[2] = { Lo, Hi };
4964 return DAG.getMergeValues(Ops, 2, dl);
4967 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4968 MVT SrcVT = Op.getOperand(0).getValueType();
4969 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4970 "Unknown SINT_TO_FP to lower!");
4972 // These are really Legal; caller falls through into that case.
4973 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4975 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4976 Subtarget->is64Bit())
4979 DebugLoc dl = Op.getDebugLoc();
4980 unsigned Size = SrcVT.getSizeInBits()/8;
4981 MachineFunction &MF = DAG.getMachineFunction();
4982 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4983 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4984 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4986 PseudoSourceValue::getFixedStack(SSFI), 0);
4990 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4992 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4994 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4995 SmallVector<SDValue, 8> Ops;
4996 Ops.push_back(Chain);
4997 Ops.push_back(StackSlot);
4998 Ops.push_back(DAG.getValueType(SrcVT));
4999 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5000 Tys, &Ops[0], Ops.size());
5003 Chain = Result.getValue(1);
5004 SDValue InFlag = Result.getValue(2);
5006 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5007 // shouldn't be necessary except that RFP cannot be live across
5008 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5009 MachineFunction &MF = DAG.getMachineFunction();
5010 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
5011 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5012 Tys = DAG.getVTList(MVT::Other);
5013 SmallVector<SDValue, 8> Ops;
5014 Ops.push_back(Chain);
5015 Ops.push_back(Result);
5016 Ops.push_back(StackSlot);
5017 Ops.push_back(DAG.getValueType(Op.getValueType()));
5018 Ops.push_back(InFlag);
5019 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5020 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5021 PseudoSourceValue::getFixedStack(SSFI), 0);
5027 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5028 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5029 // This algorithm is not obvious. Here it is in C code, more or less:
5031 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5032 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5033 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5035 // Copy ints to xmm registers.
5036 __m128i xh = _mm_cvtsi32_si128( hi );
5037 __m128i xl = _mm_cvtsi32_si128( lo );
5039 // Combine into low half of a single xmm register.
5040 __m128i x = _mm_unpacklo_epi32( xh, xl );
5044 // Merge in appropriate exponents to give the integer bits the right
5046 x = _mm_unpacklo_epi32( x, exp );
5048 // Subtract away the biases to deal with the IEEE-754 double precision
5050 d = _mm_sub_pd( (__m128d) x, bias );
5052 // All conversions up to here are exact. The correctly rounded result is
5053 // calculated using the current rounding mode using the following
5055 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5056 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5057 // store doesn't really need to be here (except
5058 // maybe to zero the other double)
5063 DebugLoc dl = Op.getDebugLoc();
5065 // Build some magic constants.
5066 std::vector<Constant*> CV0;
5067 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
5068 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
5069 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5070 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5071 Constant *C0 = ConstantVector::get(CV0);
5072 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5074 std::vector<Constant*> CV1;
5075 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
5076 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
5077 Constant *C1 = ConstantVector::get(CV1);
5078 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5080 SmallVector<SDValue, 4> MaskVec;
5081 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
5082 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
5083 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
5084 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
5085 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
5086 &MaskVec[0], MaskVec.size());
5087 SmallVector<SDValue, 4> MaskVec2;
5088 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
5089 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
5090 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32,
5091 &MaskVec2[0], MaskVec2.size());
5093 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5094 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5096 DAG.getIntPtrConstant(1)));
5097 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5098 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5100 DAG.getIntPtrConstant(0)));
5101 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5102 XR1, XR2, UnpcklMask);
5103 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5104 PseudoSourceValue::getConstantPool(), 0,
5106 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5107 Unpck1, CLod0, UnpcklMask);
5108 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5109 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5110 PseudoSourceValue::getConstantPool(), 0,
5112 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5114 // Add the halves; easiest way is to swap them into another reg first.
5115 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2f64,
5116 Sub, Sub, ShufMask);
5117 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5118 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5119 DAG.getIntPtrConstant(0));
5122 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5123 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5124 DebugLoc dl = Op.getDebugLoc();
5125 // FP constant to bias correct the final result.
5126 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5129 // Load the 32-bit value into an XMM register.
5130 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5131 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5133 DAG.getIntPtrConstant(0)));
5135 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5136 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5137 DAG.getIntPtrConstant(0));
5139 // Or the load with the bias.
5140 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5141 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5142 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5144 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5145 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5146 MVT::v2f64, Bias)));
5147 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5148 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5149 DAG.getIntPtrConstant(0));
5151 // Subtract the bias.
5152 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5154 // Handle final rounding.
5155 MVT DestVT = Op.getValueType();
5157 if (DestVT.bitsLT(MVT::f64)) {
5158 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5159 DAG.getIntPtrConstant(0));
5160 } else if (DestVT.bitsGT(MVT::f64)) {
5161 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5164 // Handle final rounding.
5168 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5169 SDValue N0 = Op.getOperand(0);
5170 DebugLoc dl = Op.getDebugLoc();
5172 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5173 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5174 // the optimization here.
5175 if (DAG.SignBitIsZero(N0))
5176 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5178 MVT SrcVT = N0.getValueType();
5179 if (SrcVT == MVT::i64) {
5180 // We only handle SSE2 f64 target here; caller can handle the rest.
5181 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5184 return LowerUINT_TO_FP_i64(Op, DAG);
5185 } else if (SrcVT == MVT::i32) {
5186 return LowerUINT_TO_FP_i32(Op, DAG);
5189 assert(0 && "Unknown UINT_TO_FP to lower!");
5193 std::pair<SDValue,SDValue> X86TargetLowering::
5194 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
5195 DebugLoc dl = Op.getDebugLoc();
5196 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
5197 Op.getValueType().getSimpleVT() >= MVT::i16 &&
5198 "Unknown FP_TO_SINT to lower!");
5200 // These are really Legal.
5201 if (Op.getValueType() == MVT::i32 &&
5202 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5203 return std::make_pair(SDValue(), SDValue());
5204 if (Subtarget->is64Bit() &&
5205 Op.getValueType() == MVT::i64 &&
5206 Op.getOperand(0).getValueType() != MVT::f80)
5207 return std::make_pair(SDValue(), SDValue());
5209 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5211 MachineFunction &MF = DAG.getMachineFunction();
5212 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
5213 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5214 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5216 switch (Op.getValueType().getSimpleVT()) {
5217 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5218 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5219 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5220 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5223 SDValue Chain = DAG.getEntryNode();
5224 SDValue Value = Op.getOperand(0);
5225 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5226 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5227 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5228 PseudoSourceValue::getFixedStack(SSFI), 0);
5229 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5231 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5233 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5234 Chain = Value.getValue(1);
5235 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5236 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5239 // Build the FP_TO_INT*_IN_MEM
5240 SDValue Ops[] = { Chain, Value, StackSlot };
5241 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5243 return std::make_pair(FIST, StackSlot);
5246 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5247 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
5248 SDValue FIST = Vals.first, StackSlot = Vals.second;
5249 if (FIST.getNode() == 0) return SDValue();
5252 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5253 FIST, StackSlot, NULL, 0);
5256 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5257 DebugLoc dl = Op.getDebugLoc();
5258 MVT VT = Op.getValueType();
5261 EltVT = VT.getVectorElementType();
5262 std::vector<Constant*> CV;
5263 if (EltVT == MVT::f64) {
5264 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5268 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5274 Constant *C = ConstantVector::get(CV);
5275 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5276 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5277 PseudoSourceValue::getConstantPool(), 0,
5279 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5282 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5283 DebugLoc dl = Op.getDebugLoc();
5284 MVT VT = Op.getValueType();
5286 unsigned EltNum = 1;
5287 if (VT.isVector()) {
5288 EltVT = VT.getVectorElementType();
5289 EltNum = VT.getVectorNumElements();
5291 std::vector<Constant*> CV;
5292 if (EltVT == MVT::f64) {
5293 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5297 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5303 Constant *C = ConstantVector::get(CV);
5304 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5305 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5306 PseudoSourceValue::getConstantPool(), 0,
5308 if (VT.isVector()) {
5309 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5310 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5311 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5313 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5315 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5319 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5320 SDValue Op0 = Op.getOperand(0);
5321 SDValue Op1 = Op.getOperand(1);
5322 DebugLoc dl = Op.getDebugLoc();
5323 MVT VT = Op.getValueType();
5324 MVT SrcVT = Op1.getValueType();
5326 // If second operand is smaller, extend it first.
5327 if (SrcVT.bitsLT(VT)) {
5328 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5331 // And if it is bigger, shrink it first.
5332 if (SrcVT.bitsGT(VT)) {
5333 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5337 // At this point the operands and the result should have the same
5338 // type, and that won't be f80 since that is not custom lowered.
5340 // First get the sign bit of second operand.
5341 std::vector<Constant*> CV;
5342 if (SrcVT == MVT::f64) {
5343 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5344 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5346 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5347 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5348 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5349 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5351 Constant *C = ConstantVector::get(CV);
5352 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5353 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5354 PseudoSourceValue::getConstantPool(), 0,
5356 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5358 // Shift sign bit right or left if the two operands have different types.
5359 if (SrcVT.bitsGT(VT)) {
5360 // Op0 is MVT::f32, Op1 is MVT::f64.
5361 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5362 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5363 DAG.getConstant(32, MVT::i32));
5364 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5365 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5366 DAG.getIntPtrConstant(0));
5369 // Clear first operand sign bit.
5371 if (VT == MVT::f64) {
5372 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5373 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5375 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5376 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5377 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5378 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5380 C = ConstantVector::get(CV);
5381 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5382 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5383 PseudoSourceValue::getConstantPool(), 0,
5385 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5387 // Or the value with the sign bit.
5388 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5391 /// Emit nodes that will be selected as "test Op0,Op0", or something
5393 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5394 SelectionDAG &DAG) {
5395 DebugLoc dl = Op.getDebugLoc();
5397 // CF and OF aren't always set the way we want. Determine which
5398 // of these we need.
5399 bool NeedCF = false;
5400 bool NeedOF = false;
5402 case X86::COND_A: case X86::COND_AE:
5403 case X86::COND_B: case X86::COND_BE:
5406 case X86::COND_G: case X86::COND_GE:
5407 case X86::COND_L: case X86::COND_LE:
5408 case X86::COND_O: case X86::COND_NO:
5414 // See if we can use the EFLAGS value from the operand instead of
5415 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5416 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5417 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5418 unsigned Opcode = 0;
5419 unsigned NumOperands = 0;
5420 switch (Op.getNode()->getOpcode()) {
5422 // Due to an isel shortcoming, be conservative if this add is likely to
5423 // be selected as part of a load-modify-store instruction. When the root
5424 // node in a match is a store, isel doesn't know how to remap non-chain
5425 // non-flag uses of other nodes in the match, such as the ADD in this
5426 // case. This leads to the ADD being left around and reselected, with
5427 // the result being two adds in the output.
5428 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5429 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5430 if (UI->getOpcode() == ISD::STORE)
5432 if (ConstantSDNode *C =
5433 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5434 // An add of one will be selected as an INC.
5435 if (C->getAPIntValue() == 1) {
5436 Opcode = X86ISD::INC;
5440 // An add of negative one (subtract of one) will be selected as a DEC.
5441 if (C->getAPIntValue().isAllOnesValue()) {
5442 Opcode = X86ISD::DEC;
5447 // Otherwise use a regular EFLAGS-setting add.
5448 Opcode = X86ISD::ADD;
5452 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5453 // likely to be selected as part of a load-modify-store instruction.
5454 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5455 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5456 if (UI->getOpcode() == ISD::STORE)
5458 // Otherwise use a regular EFLAGS-setting sub.
5459 Opcode = X86ISD::SUB;
5466 return SDValue(Op.getNode(), 1);
5472 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::i32);
5473 SmallVector<SDValue, 4> Ops;
5474 for (unsigned i = 0; i != NumOperands; ++i)
5475 Ops.push_back(Op.getOperand(i));
5476 SDValue New = DAG.getNode(Opcode, dl, VTs, 2, &Ops[0], NumOperands);
5477 DAG.ReplaceAllUsesWith(Op, New);
5478 return SDValue(New.getNode(), 1);
5482 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5483 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5484 DAG.getConstant(0, Op.getValueType()));
5487 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5489 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5490 SelectionDAG &DAG) {
5491 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5492 if (C->getAPIntValue() == 0)
5493 return EmitTest(Op0, X86CC, DAG);
5495 DebugLoc dl = Op0.getDebugLoc();
5496 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5499 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5500 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5501 SDValue Op0 = Op.getOperand(0);
5502 SDValue Op1 = Op.getOperand(1);
5503 DebugLoc dl = Op.getDebugLoc();
5504 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5506 // Lower (X & (1 << N)) == 0 to BT(X, N).
5507 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5508 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5509 if (Op0.getOpcode() == ISD::AND &&
5511 Op1.getOpcode() == ISD::Constant &&
5512 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5513 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5515 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5516 if (ConstantSDNode *Op010C =
5517 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5518 if (Op010C->getZExtValue() == 1) {
5519 LHS = Op0.getOperand(0);
5520 RHS = Op0.getOperand(1).getOperand(1);
5522 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5523 if (ConstantSDNode *Op000C =
5524 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5525 if (Op000C->getZExtValue() == 1) {
5526 LHS = Op0.getOperand(1);
5527 RHS = Op0.getOperand(0).getOperand(1);
5529 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5530 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5531 SDValue AndLHS = Op0.getOperand(0);
5532 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5533 LHS = AndLHS.getOperand(0);
5534 RHS = AndLHS.getOperand(1);
5538 if (LHS.getNode()) {
5539 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5540 // instruction. Since the shift amount is in-range-or-undefined, we know
5541 // that doing a bittest on the i16 value is ok. We extend to i32 because
5542 // the encoding for the i16 version is larger than the i32 version.
5543 if (LHS.getValueType() == MVT::i8)
5544 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5546 // If the operand types disagree, extend the shift amount to match. Since
5547 // BT ignores high bits (like shifts) we can use anyextend.
5548 if (LHS.getValueType() != RHS.getValueType())
5549 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5551 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5552 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5553 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5554 DAG.getConstant(Cond, MVT::i8), BT);
5558 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5559 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5561 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5562 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5563 DAG.getConstant(X86CC, MVT::i8), Cond);
5566 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5568 SDValue Op0 = Op.getOperand(0);
5569 SDValue Op1 = Op.getOperand(1);
5570 SDValue CC = Op.getOperand(2);
5571 MVT VT = Op.getValueType();
5572 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5573 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5574 DebugLoc dl = Op.getDebugLoc();
5578 MVT VT0 = Op0.getValueType();
5579 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5580 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5583 switch (SetCCOpcode) {
5586 case ISD::SETEQ: SSECC = 0; break;
5588 case ISD::SETGT: Swap = true; // Fallthrough
5590 case ISD::SETOLT: SSECC = 1; break;
5592 case ISD::SETGE: Swap = true; // Fallthrough
5594 case ISD::SETOLE: SSECC = 2; break;
5595 case ISD::SETUO: SSECC = 3; break;
5597 case ISD::SETNE: SSECC = 4; break;
5598 case ISD::SETULE: Swap = true;
5599 case ISD::SETUGE: SSECC = 5; break;
5600 case ISD::SETULT: Swap = true;
5601 case ISD::SETUGT: SSECC = 6; break;
5602 case ISD::SETO: SSECC = 7; break;
5605 std::swap(Op0, Op1);
5607 // In the two special cases we can't handle, emit two comparisons.
5609 if (SetCCOpcode == ISD::SETUEQ) {
5611 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5612 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5613 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5615 else if (SetCCOpcode == ISD::SETONE) {
5617 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5618 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5619 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5621 assert(0 && "Illegal FP comparison");
5623 // Handle all other FP comparisons here.
5624 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5627 // We are handling one of the integer comparisons here. Since SSE only has
5628 // GT and EQ comparisons for integer, swapping operands and multiple
5629 // operations may be required for some comparisons.
5630 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5631 bool Swap = false, Invert = false, FlipSigns = false;
5633 switch (VT.getSimpleVT()) {
5635 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5636 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5637 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5638 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5641 switch (SetCCOpcode) {
5643 case ISD::SETNE: Invert = true;
5644 case ISD::SETEQ: Opc = EQOpc; break;
5645 case ISD::SETLT: Swap = true;
5646 case ISD::SETGT: Opc = GTOpc; break;
5647 case ISD::SETGE: Swap = true;
5648 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5649 case ISD::SETULT: Swap = true;
5650 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5651 case ISD::SETUGE: Swap = true;
5652 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5655 std::swap(Op0, Op1);
5657 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5658 // bits of the inputs before performing those operations.
5660 MVT EltVT = VT.getVectorElementType();
5661 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5663 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5664 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5666 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5667 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5670 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5672 // If the logical-not of the result is required, perform that now.
5674 Result = DAG.getNOT(dl, Result, VT);
5679 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5680 static bool isX86LogicalCmp(SDValue Op) {
5681 unsigned Opc = Op.getNode()->getOpcode();
5682 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5684 if (Op.getResNo() == 1 &&
5685 (Opc == X86ISD::ADD ||
5686 Opc == X86ISD::SUB ||
5687 Opc == X86ISD::SMUL ||
5688 Opc == X86ISD::UMUL ||
5689 Opc == X86ISD::INC ||
5690 Opc == X86ISD::DEC))
5696 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5697 bool addTest = true;
5698 SDValue Cond = Op.getOperand(0);
5699 DebugLoc dl = Op.getDebugLoc();
5702 if (Cond.getOpcode() == ISD::SETCC)
5703 Cond = LowerSETCC(Cond, DAG);
5705 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5706 // setting operand in place of the X86ISD::SETCC.
5707 if (Cond.getOpcode() == X86ISD::SETCC) {
5708 CC = Cond.getOperand(0);
5710 SDValue Cmp = Cond.getOperand(1);
5711 unsigned Opc = Cmp.getOpcode();
5712 MVT VT = Op.getValueType();
5714 bool IllegalFPCMov = false;
5715 if (VT.isFloatingPoint() && !VT.isVector() &&
5716 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5717 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5719 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5720 Opc == X86ISD::BT) { // FIXME
5727 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5728 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5731 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5733 SmallVector<SDValue, 4> Ops;
5734 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5735 // condition is true.
5736 Ops.push_back(Op.getOperand(2));
5737 Ops.push_back(Op.getOperand(1));
5739 Ops.push_back(Cond);
5740 return DAG.getNode(X86ISD::CMOV, dl, VTs, 2, &Ops[0], Ops.size());
5743 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5744 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5745 // from the AND / OR.
5746 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5747 Opc = Op.getOpcode();
5748 if (Opc != ISD::OR && Opc != ISD::AND)
5750 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5751 Op.getOperand(0).hasOneUse() &&
5752 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5753 Op.getOperand(1).hasOneUse());
5756 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5757 // 1 and that the SETCC node has a single use.
5758 static bool isXor1OfSetCC(SDValue Op) {
5759 if (Op.getOpcode() != ISD::XOR)
5761 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5762 if (N1C && N1C->getAPIntValue() == 1) {
5763 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5764 Op.getOperand(0).hasOneUse();
5769 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5770 bool addTest = true;
5771 SDValue Chain = Op.getOperand(0);
5772 SDValue Cond = Op.getOperand(1);
5773 SDValue Dest = Op.getOperand(2);
5774 DebugLoc dl = Op.getDebugLoc();
5777 if (Cond.getOpcode() == ISD::SETCC)
5778 Cond = LowerSETCC(Cond, DAG);
5780 // FIXME: LowerXALUO doesn't handle these!!
5781 else if (Cond.getOpcode() == X86ISD::ADD ||
5782 Cond.getOpcode() == X86ISD::SUB ||
5783 Cond.getOpcode() == X86ISD::SMUL ||
5784 Cond.getOpcode() == X86ISD::UMUL)
5785 Cond = LowerXALUO(Cond, DAG);
5788 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5789 // setting operand in place of the X86ISD::SETCC.
5790 if (Cond.getOpcode() == X86ISD::SETCC) {
5791 CC = Cond.getOperand(0);
5793 SDValue Cmp = Cond.getOperand(1);
5794 unsigned Opc = Cmp.getOpcode();
5795 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5796 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5800 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5804 // These can only come from an arithmetic instruction with overflow,
5805 // e.g. SADDO, UADDO.
5806 Cond = Cond.getNode()->getOperand(1);
5813 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5814 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5815 if (CondOpc == ISD::OR) {
5816 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5817 // two branches instead of an explicit OR instruction with a
5819 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5820 isX86LogicalCmp(Cmp)) {
5821 CC = Cond.getOperand(0).getOperand(0);
5822 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5823 Chain, Dest, CC, Cmp);
5824 CC = Cond.getOperand(1).getOperand(0);
5828 } else { // ISD::AND
5829 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5830 // two branches instead of an explicit AND instruction with a
5831 // separate test. However, we only do this if this block doesn't
5832 // have a fall-through edge, because this requires an explicit
5833 // jmp when the condition is false.
5834 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5835 isX86LogicalCmp(Cmp) &&
5836 Op.getNode()->hasOneUse()) {
5837 X86::CondCode CCode =
5838 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5839 CCode = X86::GetOppositeBranchCondition(CCode);
5840 CC = DAG.getConstant(CCode, MVT::i8);
5841 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5842 // Look for an unconditional branch following this conditional branch.
5843 // We need this because we need to reverse the successors in order
5844 // to implement FCMP_OEQ.
5845 if (User.getOpcode() == ISD::BR) {
5846 SDValue FalseBB = User.getOperand(1);
5848 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5849 assert(NewBR == User);
5852 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5853 Chain, Dest, CC, Cmp);
5854 X86::CondCode CCode =
5855 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5856 CCode = X86::GetOppositeBranchCondition(CCode);
5857 CC = DAG.getConstant(CCode, MVT::i8);
5863 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5864 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5865 // It should be transformed during dag combiner except when the condition
5866 // is set by a arithmetics with overflow node.
5867 X86::CondCode CCode =
5868 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5869 CCode = X86::GetOppositeBranchCondition(CCode);
5870 CC = DAG.getConstant(CCode, MVT::i8);
5871 Cond = Cond.getOperand(0).getOperand(1);
5877 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5878 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5880 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5881 Chain, Dest, CC, Cond);
5885 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5886 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5887 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5888 // that the guard pages used by the OS virtual memory manager are allocated in
5889 // correct sequence.
5891 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5892 SelectionDAG &DAG) {
5893 assert(Subtarget->isTargetCygMing() &&
5894 "This should be used only on Cygwin/Mingw targets");
5895 DebugLoc dl = Op.getDebugLoc();
5898 SDValue Chain = Op.getOperand(0);
5899 SDValue Size = Op.getOperand(1);
5900 // FIXME: Ensure alignment here
5904 MVT IntPtr = getPointerTy();
5905 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5907 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5909 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5910 Flag = Chain.getValue(1);
5912 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5913 SDValue Ops[] = { Chain,
5914 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5915 DAG.getRegister(X86::EAX, IntPtr),
5916 DAG.getRegister(X86StackPtr, SPTy),
5918 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5919 Flag = Chain.getValue(1);
5921 Chain = DAG.getCALLSEQ_END(Chain,
5922 DAG.getIntPtrConstant(0, true),
5923 DAG.getIntPtrConstant(0, true),
5926 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5928 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5929 return DAG.getMergeValues(Ops1, 2, dl);
5933 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5935 SDValue Dst, SDValue Src,
5936 SDValue Size, unsigned Align,
5938 uint64_t DstSVOff) {
5939 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5941 // If not DWORD aligned or size is more than the threshold, call the library.
5942 // The libc version is likely to be faster for these cases. It can use the
5943 // address value and run time information about the CPU.
5944 if ((Align & 3) != 0 ||
5946 ConstantSize->getZExtValue() >
5947 getSubtarget()->getMaxInlineSizeThreshold()) {
5948 SDValue InFlag(0, 0);
5950 // Check to see if there is a specialized entry-point for memory zeroing.
5951 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5953 if (const char *bzeroEntry = V &&
5954 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5955 MVT IntPtr = getPointerTy();
5956 const Type *IntPtrTy = TD->getIntPtrType();
5957 TargetLowering::ArgListTy Args;
5958 TargetLowering::ArgListEntry Entry;
5960 Entry.Ty = IntPtrTy;
5961 Args.push_back(Entry);
5963 Args.push_back(Entry);
5964 std::pair<SDValue,SDValue> CallResult =
5965 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5966 CallingConv::C, false,
5967 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5968 return CallResult.second;
5971 // Otherwise have the target-independent code call memset.
5975 uint64_t SizeVal = ConstantSize->getZExtValue();
5976 SDValue InFlag(0, 0);
5979 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5980 unsigned BytesLeft = 0;
5981 bool TwoRepStos = false;
5984 uint64_t Val = ValC->getZExtValue() & 255;
5986 // If the value is a constant, then we can potentially use larger sets.
5987 switch (Align & 3) {
5988 case 2: // WORD aligned
5991 Val = (Val << 8) | Val;
5993 case 0: // DWORD aligned
5996 Val = (Val << 8) | Val;
5997 Val = (Val << 16) | Val;
5998 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6001 Val = (Val << 32) | Val;
6004 default: // Byte aligned
6007 Count = DAG.getIntPtrConstant(SizeVal);
6011 if (AVT.bitsGT(MVT::i8)) {
6012 unsigned UBytes = AVT.getSizeInBits() / 8;
6013 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6014 BytesLeft = SizeVal % UBytes;
6017 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6019 InFlag = Chain.getValue(1);
6022 Count = DAG.getIntPtrConstant(SizeVal);
6023 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6024 InFlag = Chain.getValue(1);
6027 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6030 InFlag = Chain.getValue(1);
6031 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6034 InFlag = Chain.getValue(1);
6036 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6037 SmallVector<SDValue, 8> Ops;
6038 Ops.push_back(Chain);
6039 Ops.push_back(DAG.getValueType(AVT));
6040 Ops.push_back(InFlag);
6041 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6044 InFlag = Chain.getValue(1);
6046 MVT CVT = Count.getValueType();
6047 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6048 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6049 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6052 InFlag = Chain.getValue(1);
6053 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6055 Ops.push_back(Chain);
6056 Ops.push_back(DAG.getValueType(MVT::i8));
6057 Ops.push_back(InFlag);
6058 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6059 } else if (BytesLeft) {
6060 // Handle the last 1 - 7 bytes.
6061 unsigned Offset = SizeVal - BytesLeft;
6062 MVT AddrVT = Dst.getValueType();
6063 MVT SizeVT = Size.getValueType();
6065 Chain = DAG.getMemset(Chain, dl,
6066 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6067 DAG.getConstant(Offset, AddrVT)),
6069 DAG.getConstant(BytesLeft, SizeVT),
6070 Align, DstSV, DstSVOff + Offset);
6073 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6078 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6079 SDValue Chain, SDValue Dst, SDValue Src,
6080 SDValue Size, unsigned Align,
6082 const Value *DstSV, uint64_t DstSVOff,
6083 const Value *SrcSV, uint64_t SrcSVOff) {
6084 // This requires the copy size to be a constant, preferrably
6085 // within a subtarget-specific limit.
6086 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6089 uint64_t SizeVal = ConstantSize->getZExtValue();
6090 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6093 /// If not DWORD aligned, call the library.
6094 if ((Align & 3) != 0)
6099 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6102 unsigned UBytes = AVT.getSizeInBits() / 8;
6103 unsigned CountVal = SizeVal / UBytes;
6104 SDValue Count = DAG.getIntPtrConstant(CountVal);
6105 unsigned BytesLeft = SizeVal % UBytes;
6107 SDValue InFlag(0, 0);
6108 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6111 InFlag = Chain.getValue(1);
6112 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6115 InFlag = Chain.getValue(1);
6116 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6119 InFlag = Chain.getValue(1);
6121 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6122 SmallVector<SDValue, 8> Ops;
6123 Ops.push_back(Chain);
6124 Ops.push_back(DAG.getValueType(AVT));
6125 Ops.push_back(InFlag);
6126 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6128 SmallVector<SDValue, 4> Results;
6129 Results.push_back(RepMovs);
6131 // Handle the last 1 - 7 bytes.
6132 unsigned Offset = SizeVal - BytesLeft;
6133 MVT DstVT = Dst.getValueType();
6134 MVT SrcVT = Src.getValueType();
6135 MVT SizeVT = Size.getValueType();
6136 Results.push_back(DAG.getMemcpy(Chain, dl,
6137 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6138 DAG.getConstant(Offset, DstVT)),
6139 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6140 DAG.getConstant(Offset, SrcVT)),
6141 DAG.getConstant(BytesLeft, SizeVT),
6142 Align, AlwaysInline,
6143 DstSV, DstSVOff + Offset,
6144 SrcSV, SrcSVOff + Offset));
6147 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6148 &Results[0], Results.size());
6151 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6152 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6153 DebugLoc dl = Op.getDebugLoc();
6155 if (!Subtarget->is64Bit()) {
6156 // vastart just stores the address of the VarArgsFrameIndex slot into the
6157 // memory location argument.
6158 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6159 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6163 // gp_offset (0 - 6 * 8)
6164 // fp_offset (48 - 48 + 8 * 16)
6165 // overflow_arg_area (point to parameters coming in memory).
6167 SmallVector<SDValue, 8> MemOps;
6168 SDValue FIN = Op.getOperand(1);
6170 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6171 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6173 MemOps.push_back(Store);
6176 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6177 FIN, DAG.getIntPtrConstant(4));
6178 Store = DAG.getStore(Op.getOperand(0), dl,
6179 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6181 MemOps.push_back(Store);
6183 // Store ptr to overflow_arg_area
6184 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6185 FIN, DAG.getIntPtrConstant(4));
6186 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6187 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6188 MemOps.push_back(Store);
6190 // Store ptr to reg_save_area.
6191 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6192 FIN, DAG.getIntPtrConstant(8));
6193 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6194 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6195 MemOps.push_back(Store);
6196 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6197 &MemOps[0], MemOps.size());
6200 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6201 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6202 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6203 SDValue Chain = Op.getOperand(0);
6204 SDValue SrcPtr = Op.getOperand(1);
6205 SDValue SrcSV = Op.getOperand(2);
6207 assert(0 && "VAArgInst is not yet implemented for x86-64!");
6212 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6213 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6214 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6215 SDValue Chain = Op.getOperand(0);
6216 SDValue DstPtr = Op.getOperand(1);
6217 SDValue SrcPtr = Op.getOperand(2);
6218 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6219 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6220 DebugLoc dl = Op.getDebugLoc();
6222 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6223 DAG.getIntPtrConstant(24), 8, false,
6224 DstSV, 0, SrcSV, 0);
6228 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6229 DebugLoc dl = Op.getDebugLoc();
6230 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6232 default: return SDValue(); // Don't custom lower most intrinsics.
6233 // Comparison intrinsics.
6234 case Intrinsic::x86_sse_comieq_ss:
6235 case Intrinsic::x86_sse_comilt_ss:
6236 case Intrinsic::x86_sse_comile_ss:
6237 case Intrinsic::x86_sse_comigt_ss:
6238 case Intrinsic::x86_sse_comige_ss:
6239 case Intrinsic::x86_sse_comineq_ss:
6240 case Intrinsic::x86_sse_ucomieq_ss:
6241 case Intrinsic::x86_sse_ucomilt_ss:
6242 case Intrinsic::x86_sse_ucomile_ss:
6243 case Intrinsic::x86_sse_ucomigt_ss:
6244 case Intrinsic::x86_sse_ucomige_ss:
6245 case Intrinsic::x86_sse_ucomineq_ss:
6246 case Intrinsic::x86_sse2_comieq_sd:
6247 case Intrinsic::x86_sse2_comilt_sd:
6248 case Intrinsic::x86_sse2_comile_sd:
6249 case Intrinsic::x86_sse2_comigt_sd:
6250 case Intrinsic::x86_sse2_comige_sd:
6251 case Intrinsic::x86_sse2_comineq_sd:
6252 case Intrinsic::x86_sse2_ucomieq_sd:
6253 case Intrinsic::x86_sse2_ucomilt_sd:
6254 case Intrinsic::x86_sse2_ucomile_sd:
6255 case Intrinsic::x86_sse2_ucomigt_sd:
6256 case Intrinsic::x86_sse2_ucomige_sd:
6257 case Intrinsic::x86_sse2_ucomineq_sd: {
6259 ISD::CondCode CC = ISD::SETCC_INVALID;
6262 case Intrinsic::x86_sse_comieq_ss:
6263 case Intrinsic::x86_sse2_comieq_sd:
6267 case Intrinsic::x86_sse_comilt_ss:
6268 case Intrinsic::x86_sse2_comilt_sd:
6272 case Intrinsic::x86_sse_comile_ss:
6273 case Intrinsic::x86_sse2_comile_sd:
6277 case Intrinsic::x86_sse_comigt_ss:
6278 case Intrinsic::x86_sse2_comigt_sd:
6282 case Intrinsic::x86_sse_comige_ss:
6283 case Intrinsic::x86_sse2_comige_sd:
6287 case Intrinsic::x86_sse_comineq_ss:
6288 case Intrinsic::x86_sse2_comineq_sd:
6292 case Intrinsic::x86_sse_ucomieq_ss:
6293 case Intrinsic::x86_sse2_ucomieq_sd:
6294 Opc = X86ISD::UCOMI;
6297 case Intrinsic::x86_sse_ucomilt_ss:
6298 case Intrinsic::x86_sse2_ucomilt_sd:
6299 Opc = X86ISD::UCOMI;
6302 case Intrinsic::x86_sse_ucomile_ss:
6303 case Intrinsic::x86_sse2_ucomile_sd:
6304 Opc = X86ISD::UCOMI;
6307 case Intrinsic::x86_sse_ucomigt_ss:
6308 case Intrinsic::x86_sse2_ucomigt_sd:
6309 Opc = X86ISD::UCOMI;
6312 case Intrinsic::x86_sse_ucomige_ss:
6313 case Intrinsic::x86_sse2_ucomige_sd:
6314 Opc = X86ISD::UCOMI;
6317 case Intrinsic::x86_sse_ucomineq_ss:
6318 case Intrinsic::x86_sse2_ucomineq_sd:
6319 Opc = X86ISD::UCOMI;
6324 SDValue LHS = Op.getOperand(1);
6325 SDValue RHS = Op.getOperand(2);
6326 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6327 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6328 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6329 DAG.getConstant(X86CC, MVT::i8), Cond);
6330 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6333 // Fix vector shift instructions where the last operand is a non-immediate
6335 case Intrinsic::x86_sse2_pslli_w:
6336 case Intrinsic::x86_sse2_pslli_d:
6337 case Intrinsic::x86_sse2_pslli_q:
6338 case Intrinsic::x86_sse2_psrli_w:
6339 case Intrinsic::x86_sse2_psrli_d:
6340 case Intrinsic::x86_sse2_psrli_q:
6341 case Intrinsic::x86_sse2_psrai_w:
6342 case Intrinsic::x86_sse2_psrai_d:
6343 case Intrinsic::x86_mmx_pslli_w:
6344 case Intrinsic::x86_mmx_pslli_d:
6345 case Intrinsic::x86_mmx_pslli_q:
6346 case Intrinsic::x86_mmx_psrli_w:
6347 case Intrinsic::x86_mmx_psrli_d:
6348 case Intrinsic::x86_mmx_psrli_q:
6349 case Intrinsic::x86_mmx_psrai_w:
6350 case Intrinsic::x86_mmx_psrai_d: {
6351 SDValue ShAmt = Op.getOperand(2);
6352 if (isa<ConstantSDNode>(ShAmt))
6355 unsigned NewIntNo = 0;
6356 MVT ShAmtVT = MVT::v4i32;
6358 case Intrinsic::x86_sse2_pslli_w:
6359 NewIntNo = Intrinsic::x86_sse2_psll_w;
6361 case Intrinsic::x86_sse2_pslli_d:
6362 NewIntNo = Intrinsic::x86_sse2_psll_d;
6364 case Intrinsic::x86_sse2_pslli_q:
6365 NewIntNo = Intrinsic::x86_sse2_psll_q;
6367 case Intrinsic::x86_sse2_psrli_w:
6368 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6370 case Intrinsic::x86_sse2_psrli_d:
6371 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6373 case Intrinsic::x86_sse2_psrli_q:
6374 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6376 case Intrinsic::x86_sse2_psrai_w:
6377 NewIntNo = Intrinsic::x86_sse2_psra_w;
6379 case Intrinsic::x86_sse2_psrai_d:
6380 NewIntNo = Intrinsic::x86_sse2_psra_d;
6383 ShAmtVT = MVT::v2i32;
6385 case Intrinsic::x86_mmx_pslli_w:
6386 NewIntNo = Intrinsic::x86_mmx_psll_w;
6388 case Intrinsic::x86_mmx_pslli_d:
6389 NewIntNo = Intrinsic::x86_mmx_psll_d;
6391 case Intrinsic::x86_mmx_pslli_q:
6392 NewIntNo = Intrinsic::x86_mmx_psll_q;
6394 case Intrinsic::x86_mmx_psrli_w:
6395 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6397 case Intrinsic::x86_mmx_psrli_d:
6398 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6400 case Intrinsic::x86_mmx_psrli_q:
6401 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6403 case Intrinsic::x86_mmx_psrai_w:
6404 NewIntNo = Intrinsic::x86_mmx_psra_w;
6406 case Intrinsic::x86_mmx_psrai_d:
6407 NewIntNo = Intrinsic::x86_mmx_psra_d;
6409 default: abort(); // Can't reach here.
6414 MVT VT = Op.getValueType();
6415 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6416 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6417 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6418 DAG.getConstant(NewIntNo, MVT::i32),
6419 Op.getOperand(1), ShAmt);
6424 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6425 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6426 DebugLoc dl = Op.getDebugLoc();
6429 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6431 DAG.getConstant(TD->getPointerSize(),
6432 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6433 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6434 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6439 // Just load the return address.
6440 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6441 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6442 RetAddrFI, NULL, 0);
6445 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6446 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6447 MFI->setFrameAddressIsTaken(true);
6448 MVT VT = Op.getValueType();
6449 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6450 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6451 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6452 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6454 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6458 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6459 SelectionDAG &DAG) {
6460 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6463 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6465 MachineFunction &MF = DAG.getMachineFunction();
6466 SDValue Chain = Op.getOperand(0);
6467 SDValue Offset = Op.getOperand(1);
6468 SDValue Handler = Op.getOperand(2);
6469 DebugLoc dl = Op.getDebugLoc();
6471 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6473 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6475 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6476 DAG.getIntPtrConstant(-TD->getPointerSize()));
6477 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6478 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6479 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6480 MF.getRegInfo().addLiveOut(StoreAddrReg);
6482 return DAG.getNode(X86ISD::EH_RETURN, dl,
6484 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6487 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6488 SelectionDAG &DAG) {
6489 SDValue Root = Op.getOperand(0);
6490 SDValue Trmp = Op.getOperand(1); // trampoline
6491 SDValue FPtr = Op.getOperand(2); // nested function
6492 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6493 DebugLoc dl = Op.getDebugLoc();
6495 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6497 const X86InstrInfo *TII =
6498 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6500 if (Subtarget->is64Bit()) {
6501 SDValue OutChains[6];
6503 // Large code-model.
6505 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6506 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6508 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6509 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6511 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6513 // Load the pointer to the nested function into R11.
6514 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6515 SDValue Addr = Trmp;
6516 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6519 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6520 DAG.getConstant(2, MVT::i64));
6521 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6523 // Load the 'nest' parameter value into R10.
6524 // R10 is specified in X86CallingConv.td
6525 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6526 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6527 DAG.getConstant(10, MVT::i64));
6528 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6529 Addr, TrmpAddr, 10);
6531 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6532 DAG.getConstant(12, MVT::i64));
6533 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6535 // Jump to the nested function.
6536 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6537 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6538 DAG.getConstant(20, MVT::i64));
6539 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6540 Addr, TrmpAddr, 20);
6542 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6543 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6544 DAG.getConstant(22, MVT::i64));
6545 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6549 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6550 return DAG.getMergeValues(Ops, 2, dl);
6552 const Function *Func =
6553 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6554 unsigned CC = Func->getCallingConv();
6559 assert(0 && "Unsupported calling convention");
6560 case CallingConv::C:
6561 case CallingConv::X86_StdCall: {
6562 // Pass 'nest' parameter in ECX.
6563 // Must be kept in sync with X86CallingConv.td
6566 // Check that ECX wasn't needed by an 'inreg' parameter.
6567 const FunctionType *FTy = Func->getFunctionType();
6568 const AttrListPtr &Attrs = Func->getAttributes();
6570 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6571 unsigned InRegCount = 0;
6574 for (FunctionType::param_iterator I = FTy->param_begin(),
6575 E = FTy->param_end(); I != E; ++I, ++Idx)
6576 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6577 // FIXME: should only count parameters that are lowered to integers.
6578 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6580 if (InRegCount > 2) {
6581 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6587 case CallingConv::X86_FastCall:
6588 case CallingConv::Fast:
6589 // Pass 'nest' parameter in EAX.
6590 // Must be kept in sync with X86CallingConv.td
6595 SDValue OutChains[4];
6598 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6599 DAG.getConstant(10, MVT::i32));
6600 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6602 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6603 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6604 OutChains[0] = DAG.getStore(Root, dl,
6605 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6608 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6609 DAG.getConstant(1, MVT::i32));
6610 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6612 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6613 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6614 DAG.getConstant(5, MVT::i32));
6615 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6616 TrmpAddr, 5, false, 1);
6618 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6619 DAG.getConstant(6, MVT::i32));
6620 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6623 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6624 return DAG.getMergeValues(Ops, 2, dl);
6628 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6630 The rounding mode is in bits 11:10 of FPSR, and has the following
6637 FLT_ROUNDS, on the other hand, expects the following:
6644 To perform the conversion, we do:
6645 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6648 MachineFunction &MF = DAG.getMachineFunction();
6649 const TargetMachine &TM = MF.getTarget();
6650 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6651 unsigned StackAlignment = TFI.getStackAlignment();
6652 MVT VT = Op.getValueType();
6653 DebugLoc dl = Op.getDebugLoc();
6655 // Save FP Control Word to stack slot
6656 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6657 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6659 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6660 DAG.getEntryNode(), StackSlot);
6662 // Load FP Control Word from stack slot
6663 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6665 // Transform as necessary
6667 DAG.getNode(ISD::SRL, dl, MVT::i16,
6668 DAG.getNode(ISD::AND, dl, MVT::i16,
6669 CWD, DAG.getConstant(0x800, MVT::i16)),
6670 DAG.getConstant(11, MVT::i8));
6672 DAG.getNode(ISD::SRL, dl, MVT::i16,
6673 DAG.getNode(ISD::AND, dl, MVT::i16,
6674 CWD, DAG.getConstant(0x400, MVT::i16)),
6675 DAG.getConstant(9, MVT::i8));
6678 DAG.getNode(ISD::AND, dl, MVT::i16,
6679 DAG.getNode(ISD::ADD, dl, MVT::i16,
6680 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6681 DAG.getConstant(1, MVT::i16)),
6682 DAG.getConstant(3, MVT::i16));
6685 return DAG.getNode((VT.getSizeInBits() < 16 ?
6686 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6689 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6690 MVT VT = Op.getValueType();
6692 unsigned NumBits = VT.getSizeInBits();
6693 DebugLoc dl = Op.getDebugLoc();
6695 Op = Op.getOperand(0);
6696 if (VT == MVT::i8) {
6697 // Zero extend to i32 since there is not an i8 bsr.
6699 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6702 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6703 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6704 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6706 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6707 SmallVector<SDValue, 4> Ops;
6709 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6710 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6711 Ops.push_back(Op.getValue(1));
6712 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6714 // Finally xor with NumBits-1.
6715 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6718 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6722 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6723 MVT VT = Op.getValueType();
6725 unsigned NumBits = VT.getSizeInBits();
6726 DebugLoc dl = Op.getDebugLoc();
6728 Op = Op.getOperand(0);
6729 if (VT == MVT::i8) {
6731 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6734 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6735 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6736 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6738 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6739 SmallVector<SDValue, 4> Ops;
6741 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6742 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6743 Ops.push_back(Op.getValue(1));
6744 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6747 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6751 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6752 MVT VT = Op.getValueType();
6753 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6754 DebugLoc dl = Op.getDebugLoc();
6756 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6757 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6758 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6759 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6760 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6762 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6763 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6764 // return AloBlo + AloBhi + AhiBlo;
6766 SDValue A = Op.getOperand(0);
6767 SDValue B = Op.getOperand(1);
6769 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6770 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6771 A, DAG.getConstant(32, MVT::i32));
6772 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6773 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6774 B, DAG.getConstant(32, MVT::i32));
6775 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6776 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6778 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6779 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6781 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6782 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6784 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6785 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6786 AloBhi, DAG.getConstant(32, MVT::i32));
6787 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6788 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6789 AhiBlo, DAG.getConstant(32, MVT::i32));
6790 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6791 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6796 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6797 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6798 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6799 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6800 // has only one use.
6801 SDNode *N = Op.getNode();
6802 SDValue LHS = N->getOperand(0);
6803 SDValue RHS = N->getOperand(1);
6804 unsigned BaseOp = 0;
6806 DebugLoc dl = Op.getDebugLoc();
6808 switch (Op.getOpcode()) {
6809 default: assert(0 && "Unknown ovf instruction!");
6811 // A subtract of one will be selected as a INC. Note that INC doesn't
6812 // set CF, so we can't do this for UADDO.
6813 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6814 if (C->getAPIntValue() == 1) {
6815 BaseOp = X86ISD::INC;
6819 BaseOp = X86ISD::ADD;
6823 BaseOp = X86ISD::ADD;
6827 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6828 // set CF, so we can't do this for USUBO.
6829 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6830 if (C->getAPIntValue() == 1) {
6831 BaseOp = X86ISD::DEC;
6835 BaseOp = X86ISD::SUB;
6839 BaseOp = X86ISD::SUB;
6843 BaseOp = X86ISD::SMUL;
6847 BaseOp = X86ISD::UMUL;
6852 // Also sets EFLAGS.
6853 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6854 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6857 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6858 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6860 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6864 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6865 MVT T = Op.getValueType();
6866 DebugLoc dl = Op.getDebugLoc();
6869 switch(T.getSimpleVT()) {
6871 assert(false && "Invalid value type!");
6872 case MVT::i8: Reg = X86::AL; size = 1; break;
6873 case MVT::i16: Reg = X86::AX; size = 2; break;
6874 case MVT::i32: Reg = X86::EAX; size = 4; break;
6876 assert(Subtarget->is64Bit() && "Node not type legal!");
6877 Reg = X86::RAX; size = 8;
6880 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6881 Op.getOperand(2), SDValue());
6882 SDValue Ops[] = { cpIn.getValue(0),
6885 DAG.getTargetConstant(size, MVT::i8),
6887 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6888 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6890 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6894 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6895 SelectionDAG &DAG) {
6896 assert(Subtarget->is64Bit() && "Result not type legalized?");
6897 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6898 SDValue TheChain = Op.getOperand(0);
6899 DebugLoc dl = Op.getDebugLoc();
6900 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6901 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6902 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6904 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6905 DAG.getConstant(32, MVT::i8));
6907 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6910 return DAG.getMergeValues(Ops, 2, dl);
6913 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6914 SDNode *Node = Op.getNode();
6915 DebugLoc dl = Node->getDebugLoc();
6916 MVT T = Node->getValueType(0);
6917 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6918 DAG.getConstant(0, T), Node->getOperand(2));
6919 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6920 cast<AtomicSDNode>(Node)->getMemoryVT(),
6921 Node->getOperand(0),
6922 Node->getOperand(1), negOp,
6923 cast<AtomicSDNode>(Node)->getSrcValue(),
6924 cast<AtomicSDNode>(Node)->getAlignment());
6927 /// LowerOperation - Provide custom lowering hooks for some operations.
6929 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6930 switch (Op.getOpcode()) {
6931 default: assert(0 && "Should not custom lower this!");
6932 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6933 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6934 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6935 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6936 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6937 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6938 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6939 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6940 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6941 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6942 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6943 case ISD::SHL_PARTS:
6944 case ISD::SRA_PARTS:
6945 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6946 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6947 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6948 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6949 case ISD::FABS: return LowerFABS(Op, DAG);
6950 case ISD::FNEG: return LowerFNEG(Op, DAG);
6951 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6952 case ISD::SETCC: return LowerSETCC(Op, DAG);
6953 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6954 case ISD::SELECT: return LowerSELECT(Op, DAG);
6955 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6956 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6957 case ISD::CALL: return LowerCALL(Op, DAG);
6958 case ISD::RET: return LowerRET(Op, DAG);
6959 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6960 case ISD::VASTART: return LowerVASTART(Op, DAG);
6961 case ISD::VAARG: return LowerVAARG(Op, DAG);
6962 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6963 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6964 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6965 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6966 case ISD::FRAME_TO_ARGS_OFFSET:
6967 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6968 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6969 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6970 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6971 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6972 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6973 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6974 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6980 case ISD::UMULO: return LowerXALUO(Op, DAG);
6981 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6985 void X86TargetLowering::
6986 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6987 SelectionDAG &DAG, unsigned NewOp) {
6988 MVT T = Node->getValueType(0);
6989 DebugLoc dl = Node->getDebugLoc();
6990 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6992 SDValue Chain = Node->getOperand(0);
6993 SDValue In1 = Node->getOperand(1);
6994 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6995 Node->getOperand(2), DAG.getIntPtrConstant(0));
6996 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6997 Node->getOperand(2), DAG.getIntPtrConstant(1));
6998 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6999 // have a MemOperand. Pass the info through as a normal operand.
7000 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
7001 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
7002 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7003 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
7004 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7005 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7006 Results.push_back(Result.getValue(2));
7009 /// ReplaceNodeResults - Replace a node with an illegal result type
7010 /// with a new node built out of custom code.
7011 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7012 SmallVectorImpl<SDValue>&Results,
7013 SelectionDAG &DAG) {
7014 DebugLoc dl = N->getDebugLoc();
7015 switch (N->getOpcode()) {
7017 assert(false && "Do not know how to custom type legalize this operation!");
7019 case ISD::FP_TO_SINT: {
7020 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
7021 SDValue FIST = Vals.first, StackSlot = Vals.second;
7022 if (FIST.getNode() != 0) {
7023 MVT VT = N->getValueType(0);
7024 // Return a load from the stack slot.
7025 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7029 case ISD::READCYCLECOUNTER: {
7030 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7031 SDValue TheChain = N->getOperand(0);
7032 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7033 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7035 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7037 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7038 SDValue Ops[] = { eax, edx };
7039 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7040 Results.push_back(edx.getValue(1));
7043 case ISD::ATOMIC_CMP_SWAP: {
7044 MVT T = N->getValueType(0);
7045 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7046 SDValue cpInL, cpInH;
7047 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7048 DAG.getConstant(0, MVT::i32));
7049 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7050 DAG.getConstant(1, MVT::i32));
7051 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7052 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7054 SDValue swapInL, swapInH;
7055 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7056 DAG.getConstant(0, MVT::i32));
7057 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7058 DAG.getConstant(1, MVT::i32));
7059 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7061 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7062 swapInL.getValue(1));
7063 SDValue Ops[] = { swapInH.getValue(0),
7065 swapInH.getValue(1) };
7066 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7067 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7068 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7069 MVT::i32, Result.getValue(1));
7070 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7071 MVT::i32, cpOutL.getValue(2));
7072 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7073 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7074 Results.push_back(cpOutH.getValue(1));
7077 case ISD::ATOMIC_LOAD_ADD:
7078 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7080 case ISD::ATOMIC_LOAD_AND:
7081 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7083 case ISD::ATOMIC_LOAD_NAND:
7084 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7086 case ISD::ATOMIC_LOAD_OR:
7087 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7089 case ISD::ATOMIC_LOAD_SUB:
7090 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7092 case ISD::ATOMIC_LOAD_XOR:
7093 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7095 case ISD::ATOMIC_SWAP:
7096 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7101 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7103 default: return NULL;
7104 case X86ISD::BSF: return "X86ISD::BSF";
7105 case X86ISD::BSR: return "X86ISD::BSR";
7106 case X86ISD::SHLD: return "X86ISD::SHLD";
7107 case X86ISD::SHRD: return "X86ISD::SHRD";
7108 case X86ISD::FAND: return "X86ISD::FAND";
7109 case X86ISD::FOR: return "X86ISD::FOR";
7110 case X86ISD::FXOR: return "X86ISD::FXOR";
7111 case X86ISD::FSRL: return "X86ISD::FSRL";
7112 case X86ISD::FILD: return "X86ISD::FILD";
7113 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7114 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7115 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7116 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7117 case X86ISD::FLD: return "X86ISD::FLD";
7118 case X86ISD::FST: return "X86ISD::FST";
7119 case X86ISD::CALL: return "X86ISD::CALL";
7120 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
7121 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7122 case X86ISD::BT: return "X86ISD::BT";
7123 case X86ISD::CMP: return "X86ISD::CMP";
7124 case X86ISD::COMI: return "X86ISD::COMI";
7125 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7126 case X86ISD::SETCC: return "X86ISD::SETCC";
7127 case X86ISD::CMOV: return "X86ISD::CMOV";
7128 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7129 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7130 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7131 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7132 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7133 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7134 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7135 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7136 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7137 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7138 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7139 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7140 case X86ISD::FMAX: return "X86ISD::FMAX";
7141 case X86ISD::FMIN: return "X86ISD::FMIN";
7142 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7143 case X86ISD::FRCP: return "X86ISD::FRCP";
7144 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7145 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
7146 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7147 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7148 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7149 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7150 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7151 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7152 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7153 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7154 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7155 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7156 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7157 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7158 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7159 case X86ISD::VSHL: return "X86ISD::VSHL";
7160 case X86ISD::VSRL: return "X86ISD::VSRL";
7161 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7162 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7163 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7164 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7165 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7166 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7167 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7168 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7169 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7170 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7171 case X86ISD::ADD: return "X86ISD::ADD";
7172 case X86ISD::SUB: return "X86ISD::SUB";
7173 case X86ISD::SMUL: return "X86ISD::SMUL";
7174 case X86ISD::UMUL: return "X86ISD::UMUL";
7175 case X86ISD::INC: return "X86ISD::INC";
7176 case X86ISD::DEC: return "X86ISD::DEC";
7180 // isLegalAddressingMode - Return true if the addressing mode represented
7181 // by AM is legal for this target, for a load/store of the specified type.
7182 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7183 const Type *Ty) const {
7184 // X86 supports extremely general addressing modes.
7186 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7187 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7191 // We can only fold this if we don't need an extra load.
7192 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
7194 // If BaseGV requires a register, we cannot also have a BaseReg.
7195 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
7199 // X86-64 only supports addr of globals in small code model.
7200 if (Subtarget->is64Bit()) {
7201 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7203 // If lower 4G is not available, then we must use rip-relative addressing.
7204 if (AM.BaseOffs || AM.Scale > 1)
7215 // These scales always work.
7220 // These scales are formed with basereg+scalereg. Only accept if there is
7225 default: // Other stuff never works.
7233 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7234 if (!Ty1->isInteger() || !Ty2->isInteger())
7236 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7237 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7238 if (NumBits1 <= NumBits2)
7240 return Subtarget->is64Bit() || NumBits1 < 64;
7243 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7244 if (!VT1.isInteger() || !VT2.isInteger())
7246 unsigned NumBits1 = VT1.getSizeInBits();
7247 unsigned NumBits2 = VT2.getSizeInBits();
7248 if (NumBits1 <= NumBits2)
7250 return Subtarget->is64Bit() || NumBits1 < 64;
7253 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7254 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7255 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7256 /// are assumed to be legal.
7258 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
7259 // Only do shuffles on 128-bit vector types for now.
7260 // FIXME: pshufb, blends
7261 if (VT.getSizeInBits() == 64) return false;
7262 return (Mask.getNode()->getNumOperands() <= 4 ||
7263 isIdentityMask(Mask.getNode()) ||
7264 isIdentityMask(Mask.getNode(), true) ||
7265 isSplatMask(Mask.getNode()) ||
7266 X86::isPSHUFHWMask(Mask.getNode()) ||
7267 X86::isPSHUFLWMask(Mask.getNode()) ||
7268 X86::isUNPCKLMask(Mask.getNode()) ||
7269 X86::isUNPCKHMask(Mask.getNode()) ||
7270 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
7271 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
7275 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
7276 MVT EVT, SelectionDAG &DAG) const {
7277 unsigned NumElts = BVOps.size();
7278 // Only do shuffles on 128-bit vector types for now.
7279 if (EVT.getSizeInBits() * NumElts == 64) return false;
7280 if (NumElts == 2) return true;
7282 return (isMOVLMask(&BVOps[0], 4) ||
7283 isCommutedMOVL(&BVOps[0], 4, true) ||
7284 isSHUFPMask(&BVOps[0], 4) ||
7285 isCommutedSHUFP(&BVOps[0], 4));
7290 //===----------------------------------------------------------------------===//
7291 // X86 Scheduler Hooks
7292 //===----------------------------------------------------------------------===//
7294 // private utility function
7296 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7297 MachineBasicBlock *MBB,
7305 TargetRegisterClass *RC,
7306 bool invSrc) const {
7307 // For the atomic bitwise operator, we generate
7310 // ld t1 = [bitinstr.addr]
7311 // op t2 = t1, [bitinstr.val]
7313 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7315 // fallthrough -->nextMBB
7316 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7317 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7318 MachineFunction::iterator MBBIter = MBB;
7321 /// First build the CFG
7322 MachineFunction *F = MBB->getParent();
7323 MachineBasicBlock *thisMBB = MBB;
7324 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7325 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7326 F->insert(MBBIter, newMBB);
7327 F->insert(MBBIter, nextMBB);
7329 // Move all successors to thisMBB to nextMBB
7330 nextMBB->transferSuccessors(thisMBB);
7332 // Update thisMBB to fall through to newMBB
7333 thisMBB->addSuccessor(newMBB);
7335 // newMBB jumps to itself and fall through to nextMBB
7336 newMBB->addSuccessor(nextMBB);
7337 newMBB->addSuccessor(newMBB);
7339 // Insert instructions into newMBB based on incoming instruction
7340 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
7341 DebugLoc dl = bInstr->getDebugLoc();
7342 MachineOperand& destOper = bInstr->getOperand(0);
7343 MachineOperand* argOpers[6];
7344 int numArgs = bInstr->getNumOperands() - 1;
7345 for (int i=0; i < numArgs; ++i)
7346 argOpers[i] = &bInstr->getOperand(i+1);
7348 // x86 address has 4 operands: base, index, scale, and displacement
7349 int lastAddrIndx = 3; // [0,3]
7352 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7353 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7354 for (int i=0; i <= lastAddrIndx; ++i)
7355 (*MIB).addOperand(*argOpers[i]);
7357 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7359 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7364 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7365 assert((argOpers[valArgIndx]->isReg() ||
7366 argOpers[valArgIndx]->isImm()) &&
7368 if (argOpers[valArgIndx]->isReg())
7369 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7371 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7373 (*MIB).addOperand(*argOpers[valArgIndx]);
7375 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7378 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7379 for (int i=0; i <= lastAddrIndx; ++i)
7380 (*MIB).addOperand(*argOpers[i]);
7382 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7383 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7385 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7389 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7391 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7395 // private utility function: 64 bit atomics on 32 bit host.
7397 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7398 MachineBasicBlock *MBB,
7403 bool invSrc) const {
7404 // For the atomic bitwise operator, we generate
7405 // thisMBB (instructions are in pairs, except cmpxchg8b)
7406 // ld t1,t2 = [bitinstr.addr]
7408 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7409 // op t5, t6 <- out1, out2, [bitinstr.val]
7410 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7411 // mov ECX, EBX <- t5, t6
7412 // mov EAX, EDX <- t1, t2
7413 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7414 // mov t3, t4 <- EAX, EDX
7416 // result in out1, out2
7417 // fallthrough -->nextMBB
7419 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7420 const unsigned LoadOpc = X86::MOV32rm;
7421 const unsigned copyOpc = X86::MOV32rr;
7422 const unsigned NotOpc = X86::NOT32r;
7423 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7424 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7425 MachineFunction::iterator MBBIter = MBB;
7428 /// First build the CFG
7429 MachineFunction *F = MBB->getParent();
7430 MachineBasicBlock *thisMBB = MBB;
7431 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7432 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7433 F->insert(MBBIter, newMBB);
7434 F->insert(MBBIter, nextMBB);
7436 // Move all successors to thisMBB to nextMBB
7437 nextMBB->transferSuccessors(thisMBB);
7439 // Update thisMBB to fall through to newMBB
7440 thisMBB->addSuccessor(newMBB);
7442 // newMBB jumps to itself and fall through to nextMBB
7443 newMBB->addSuccessor(nextMBB);
7444 newMBB->addSuccessor(newMBB);
7446 DebugLoc dl = bInstr->getDebugLoc();
7447 // Insert instructions into newMBB based on incoming instruction
7448 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7449 assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
7450 MachineOperand& dest1Oper = bInstr->getOperand(0);
7451 MachineOperand& dest2Oper = bInstr->getOperand(1);
7452 MachineOperand* argOpers[6];
7453 for (int i=0; i < 6; ++i)
7454 argOpers[i] = &bInstr->getOperand(i+2);
7456 // x86 address has 4 operands: base, index, scale, and displacement
7457 int lastAddrIndx = 3; // [0,3]
7459 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7460 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7461 for (int i=0; i <= lastAddrIndx; ++i)
7462 (*MIB).addOperand(*argOpers[i]);
7463 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7464 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7465 // add 4 to displacement.
7466 for (int i=0; i <= lastAddrIndx-1; ++i)
7467 (*MIB).addOperand(*argOpers[i]);
7468 MachineOperand newOp3 = *(argOpers[3]);
7470 newOp3.setImm(newOp3.getImm()+4);
7472 newOp3.setOffset(newOp3.getOffset()+4);
7473 (*MIB).addOperand(newOp3);
7475 // t3/4 are defined later, at the bottom of the loop
7476 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7477 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7478 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7479 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7480 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7481 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7483 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7484 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7486 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7487 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7493 assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
7495 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7496 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7497 if (argOpers[4]->isReg())
7498 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7500 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7501 if (regOpcL != X86::MOV32rr)
7503 (*MIB).addOperand(*argOpers[4]);
7504 assert(argOpers[5]->isReg() == argOpers[4]->isReg());
7505 assert(argOpers[5]->isImm() == argOpers[4]->isImm());
7506 if (argOpers[5]->isReg())
7507 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7509 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7510 if (regOpcH != X86::MOV32rr)
7512 (*MIB).addOperand(*argOpers[5]);
7514 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7516 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7519 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7521 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7524 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7525 for (int i=0; i <= lastAddrIndx; ++i)
7526 (*MIB).addOperand(*argOpers[i]);
7528 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7529 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7531 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7532 MIB.addReg(X86::EAX);
7533 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7534 MIB.addReg(X86::EDX);
7537 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7539 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7543 // private utility function
7545 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7546 MachineBasicBlock *MBB,
7547 unsigned cmovOpc) const {
7548 // For the atomic min/max operator, we generate
7551 // ld t1 = [min/max.addr]
7552 // mov t2 = [min/max.val]
7554 // cmov[cond] t2 = t1
7556 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7558 // fallthrough -->nextMBB
7560 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7561 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7562 MachineFunction::iterator MBBIter = MBB;
7565 /// First build the CFG
7566 MachineFunction *F = MBB->getParent();
7567 MachineBasicBlock *thisMBB = MBB;
7568 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7569 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7570 F->insert(MBBIter, newMBB);
7571 F->insert(MBBIter, nextMBB);
7573 // Move all successors to thisMBB to nextMBB
7574 nextMBB->transferSuccessors(thisMBB);
7576 // Update thisMBB to fall through to newMBB
7577 thisMBB->addSuccessor(newMBB);
7579 // newMBB jumps to newMBB and fall through to nextMBB
7580 newMBB->addSuccessor(nextMBB);
7581 newMBB->addSuccessor(newMBB);
7583 DebugLoc dl = mInstr->getDebugLoc();
7584 // Insert instructions into newMBB based on incoming instruction
7585 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
7586 MachineOperand& destOper = mInstr->getOperand(0);
7587 MachineOperand* argOpers[6];
7588 int numArgs = mInstr->getNumOperands() - 1;
7589 for (int i=0; i < numArgs; ++i)
7590 argOpers[i] = &mInstr->getOperand(i+1);
7592 // x86 address has 4 operands: base, index, scale, and displacement
7593 int lastAddrIndx = 3; // [0,3]
7596 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7597 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7598 for (int i=0; i <= lastAddrIndx; ++i)
7599 (*MIB).addOperand(*argOpers[i]);
7601 // We only support register and immediate values
7602 assert((argOpers[valArgIndx]->isReg() ||
7603 argOpers[valArgIndx]->isImm()) &&
7606 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7607 if (argOpers[valArgIndx]->isReg())
7608 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7610 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7611 (*MIB).addOperand(*argOpers[valArgIndx]);
7613 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7616 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7621 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7622 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7626 // Cmp and exchange if none has modified the memory location
7627 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7628 for (int i=0; i <= lastAddrIndx; ++i)
7629 (*MIB).addOperand(*argOpers[i]);
7631 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7632 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7634 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7635 MIB.addReg(X86::EAX);
7638 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7640 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7646 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7647 MachineBasicBlock *BB) const {
7648 DebugLoc dl = MI->getDebugLoc();
7649 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7650 switch (MI->getOpcode()) {
7651 default: assert(false && "Unexpected instr type to insert");
7652 case X86::CMOV_V1I64:
7653 case X86::CMOV_FR32:
7654 case X86::CMOV_FR64:
7655 case X86::CMOV_V4F32:
7656 case X86::CMOV_V2F64:
7657 case X86::CMOV_V2I64: {
7658 // To "insert" a SELECT_CC instruction, we actually have to insert the
7659 // diamond control-flow pattern. The incoming instruction knows the
7660 // destination vreg to set, the condition code register to branch on, the
7661 // true/false values to select between, and a branch opcode to use.
7662 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7663 MachineFunction::iterator It = BB;
7669 // cmpTY ccX, r1, r2
7671 // fallthrough --> copy0MBB
7672 MachineBasicBlock *thisMBB = BB;
7673 MachineFunction *F = BB->getParent();
7674 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7675 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7677 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7678 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7679 F->insert(It, copy0MBB);
7680 F->insert(It, sinkMBB);
7681 // Update machine-CFG edges by transferring all successors of the current
7682 // block to the new block which will contain the Phi node for the select.
7683 sinkMBB->transferSuccessors(BB);
7685 // Add the true and fallthrough blocks as its successors.
7686 BB->addSuccessor(copy0MBB);
7687 BB->addSuccessor(sinkMBB);
7690 // %FalseValue = ...
7691 // # fallthrough to sinkMBB
7694 // Update machine-CFG edges
7695 BB->addSuccessor(sinkMBB);
7698 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7701 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7702 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7703 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7705 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7709 case X86::FP32_TO_INT16_IN_MEM:
7710 case X86::FP32_TO_INT32_IN_MEM:
7711 case X86::FP32_TO_INT64_IN_MEM:
7712 case X86::FP64_TO_INT16_IN_MEM:
7713 case X86::FP64_TO_INT32_IN_MEM:
7714 case X86::FP64_TO_INT64_IN_MEM:
7715 case X86::FP80_TO_INT16_IN_MEM:
7716 case X86::FP80_TO_INT32_IN_MEM:
7717 case X86::FP80_TO_INT64_IN_MEM: {
7718 // Change the floating point control register to use "round towards zero"
7719 // mode when truncating to an integer value.
7720 MachineFunction *F = BB->getParent();
7721 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7722 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7724 // Load the old value of the high byte of the control word...
7726 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7727 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7730 // Set the high part to be round to zero...
7731 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7734 // Reload the modified control word now...
7735 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7737 // Restore the memory image of control word to original value
7738 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7741 // Get the X86 opcode to use.
7743 switch (MI->getOpcode()) {
7744 default: assert(0 && "illegal opcode!");
7745 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7746 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7747 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7748 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7749 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7750 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7751 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7752 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7753 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7757 MachineOperand &Op = MI->getOperand(0);
7759 AM.BaseType = X86AddressMode::RegBase;
7760 AM.Base.Reg = Op.getReg();
7762 AM.BaseType = X86AddressMode::FrameIndexBase;
7763 AM.Base.FrameIndex = Op.getIndex();
7765 Op = MI->getOperand(1);
7767 AM.Scale = Op.getImm();
7768 Op = MI->getOperand(2);
7770 AM.IndexReg = Op.getImm();
7771 Op = MI->getOperand(3);
7772 if (Op.isGlobal()) {
7773 AM.GV = Op.getGlobal();
7775 AM.Disp = Op.getImm();
7777 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7778 .addReg(MI->getOperand(4).getReg());
7780 // Reload the original control word now.
7781 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7783 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7786 case X86::ATOMAND32:
7787 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7788 X86::AND32ri, X86::MOV32rm,
7789 X86::LCMPXCHG32, X86::MOV32rr,
7790 X86::NOT32r, X86::EAX,
7791 X86::GR32RegisterClass);
7793 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7794 X86::OR32ri, X86::MOV32rm,
7795 X86::LCMPXCHG32, X86::MOV32rr,
7796 X86::NOT32r, X86::EAX,
7797 X86::GR32RegisterClass);
7798 case X86::ATOMXOR32:
7799 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7800 X86::XOR32ri, X86::MOV32rm,
7801 X86::LCMPXCHG32, X86::MOV32rr,
7802 X86::NOT32r, X86::EAX,
7803 X86::GR32RegisterClass);
7804 case X86::ATOMNAND32:
7805 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7806 X86::AND32ri, X86::MOV32rm,
7807 X86::LCMPXCHG32, X86::MOV32rr,
7808 X86::NOT32r, X86::EAX,
7809 X86::GR32RegisterClass, true);
7810 case X86::ATOMMIN32:
7811 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7812 case X86::ATOMMAX32:
7813 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7814 case X86::ATOMUMIN32:
7815 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7816 case X86::ATOMUMAX32:
7817 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7819 case X86::ATOMAND16:
7820 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7821 X86::AND16ri, X86::MOV16rm,
7822 X86::LCMPXCHG16, X86::MOV16rr,
7823 X86::NOT16r, X86::AX,
7824 X86::GR16RegisterClass);
7826 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7827 X86::OR16ri, X86::MOV16rm,
7828 X86::LCMPXCHG16, X86::MOV16rr,
7829 X86::NOT16r, X86::AX,
7830 X86::GR16RegisterClass);
7831 case X86::ATOMXOR16:
7832 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7833 X86::XOR16ri, X86::MOV16rm,
7834 X86::LCMPXCHG16, X86::MOV16rr,
7835 X86::NOT16r, X86::AX,
7836 X86::GR16RegisterClass);
7837 case X86::ATOMNAND16:
7838 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7839 X86::AND16ri, X86::MOV16rm,
7840 X86::LCMPXCHG16, X86::MOV16rr,
7841 X86::NOT16r, X86::AX,
7842 X86::GR16RegisterClass, true);
7843 case X86::ATOMMIN16:
7844 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7845 case X86::ATOMMAX16:
7846 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7847 case X86::ATOMUMIN16:
7848 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7849 case X86::ATOMUMAX16:
7850 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7853 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7854 X86::AND8ri, X86::MOV8rm,
7855 X86::LCMPXCHG8, X86::MOV8rr,
7856 X86::NOT8r, X86::AL,
7857 X86::GR8RegisterClass);
7859 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7860 X86::OR8ri, X86::MOV8rm,
7861 X86::LCMPXCHG8, X86::MOV8rr,
7862 X86::NOT8r, X86::AL,
7863 X86::GR8RegisterClass);
7865 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7866 X86::XOR8ri, X86::MOV8rm,
7867 X86::LCMPXCHG8, X86::MOV8rr,
7868 X86::NOT8r, X86::AL,
7869 X86::GR8RegisterClass);
7870 case X86::ATOMNAND8:
7871 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7872 X86::AND8ri, X86::MOV8rm,
7873 X86::LCMPXCHG8, X86::MOV8rr,
7874 X86::NOT8r, X86::AL,
7875 X86::GR8RegisterClass, true);
7876 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7877 // This group is for 64-bit host.
7878 case X86::ATOMAND64:
7879 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7880 X86::AND64ri32, X86::MOV64rm,
7881 X86::LCMPXCHG64, X86::MOV64rr,
7882 X86::NOT64r, X86::RAX,
7883 X86::GR64RegisterClass);
7885 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7886 X86::OR64ri32, X86::MOV64rm,
7887 X86::LCMPXCHG64, X86::MOV64rr,
7888 X86::NOT64r, X86::RAX,
7889 X86::GR64RegisterClass);
7890 case X86::ATOMXOR64:
7891 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7892 X86::XOR64ri32, X86::MOV64rm,
7893 X86::LCMPXCHG64, X86::MOV64rr,
7894 X86::NOT64r, X86::RAX,
7895 X86::GR64RegisterClass);
7896 case X86::ATOMNAND64:
7897 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7898 X86::AND64ri32, X86::MOV64rm,
7899 X86::LCMPXCHG64, X86::MOV64rr,
7900 X86::NOT64r, X86::RAX,
7901 X86::GR64RegisterClass, true);
7902 case X86::ATOMMIN64:
7903 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7904 case X86::ATOMMAX64:
7905 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7906 case X86::ATOMUMIN64:
7907 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7908 case X86::ATOMUMAX64:
7909 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7911 // This group does 64-bit operations on a 32-bit host.
7912 case X86::ATOMAND6432:
7913 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7914 X86::AND32rr, X86::AND32rr,
7915 X86::AND32ri, X86::AND32ri,
7917 case X86::ATOMOR6432:
7918 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7919 X86::OR32rr, X86::OR32rr,
7920 X86::OR32ri, X86::OR32ri,
7922 case X86::ATOMXOR6432:
7923 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7924 X86::XOR32rr, X86::XOR32rr,
7925 X86::XOR32ri, X86::XOR32ri,
7927 case X86::ATOMNAND6432:
7928 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7929 X86::AND32rr, X86::AND32rr,
7930 X86::AND32ri, X86::AND32ri,
7932 case X86::ATOMADD6432:
7933 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7934 X86::ADD32rr, X86::ADC32rr,
7935 X86::ADD32ri, X86::ADC32ri,
7937 case X86::ATOMSUB6432:
7938 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7939 X86::SUB32rr, X86::SBB32rr,
7940 X86::SUB32ri, X86::SBB32ri,
7942 case X86::ATOMSWAP6432:
7943 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7944 X86::MOV32rr, X86::MOV32rr,
7945 X86::MOV32ri, X86::MOV32ri,
7950 //===----------------------------------------------------------------------===//
7951 // X86 Optimization Hooks
7952 //===----------------------------------------------------------------------===//
7954 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7958 const SelectionDAG &DAG,
7959 unsigned Depth) const {
7960 unsigned Opc = Op.getOpcode();
7961 assert((Opc >= ISD::BUILTIN_OP_END ||
7962 Opc == ISD::INTRINSIC_WO_CHAIN ||
7963 Opc == ISD::INTRINSIC_W_CHAIN ||
7964 Opc == ISD::INTRINSIC_VOID) &&
7965 "Should use MaskedValueIsZero if you don't know whether Op"
7966 " is a target node!");
7968 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7977 // These nodes' second result is a boolean.
7978 if (Op.getResNo() == 0)
7982 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7983 Mask.getBitWidth() - 1);
7988 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7989 /// node is a GlobalAddress + offset.
7990 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7991 GlobalValue* &GA, int64_t &Offset) const{
7992 if (N->getOpcode() == X86ISD::Wrapper) {
7993 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7994 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7995 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7999 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8002 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8003 const TargetLowering &TLI) {
8006 if (TLI.isGAPlusOffset(Base, GV, Offset))
8007 return (GV->getAlignment() >= N && (Offset % N) == 0);
8008 // DAG combine handles the stack object case.
8012 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
8013 unsigned NumElems, MVT EVT,
8015 SelectionDAG &DAG, MachineFrameInfo *MFI,
8016 const TargetLowering &TLI) {
8018 for (unsigned i = 0; i < NumElems; ++i) {
8019 SDValue Idx = PermMask.getOperand(i);
8020 if (Idx.getOpcode() == ISD::UNDEF) {
8026 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8027 if (!Elt.getNode() ||
8028 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8031 Base = Elt.getNode();
8032 if (Base->getOpcode() == ISD::UNDEF)
8036 if (Elt.getOpcode() == ISD::UNDEF)
8039 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
8040 EVT.getSizeInBits()/8, i, MFI))
8046 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8047 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8048 /// if the load addresses are consecutive, non-overlapping, and in the right
8050 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8051 const TargetLowering &TLI) {
8052 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8053 DebugLoc dl = N->getDebugLoc();
8054 MVT VT = N->getValueType(0);
8055 MVT EVT = VT.getVectorElementType();
8056 SDValue PermMask = N->getOperand(2);
8057 unsigned NumElems = PermMask.getNumOperands();
8058 SDNode *Base = NULL;
8059 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
8063 LoadSDNode *LD = cast<LoadSDNode>(Base);
8064 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
8065 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8066 LD->getSrcValue(), LD->getSrcValueOffset(),
8068 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8069 LD->getSrcValue(), LD->getSrcValueOffset(),
8070 LD->isVolatile(), LD->getAlignment());
8073 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
8074 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
8075 TargetLowering::DAGCombinerInfo &DCI,
8076 const X86Subtarget *Subtarget,
8077 const TargetLowering &TLI) {
8078 unsigned NumOps = N->getNumOperands();
8079 DebugLoc dl = N->getDebugLoc();
8081 // Ignore single operand BUILD_VECTOR.
8085 MVT VT = N->getValueType(0);
8086 MVT EVT = VT.getVectorElementType();
8087 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
8088 // We are looking for load i64 and zero extend. We want to transform
8089 // it before legalizer has a chance to expand it. Also look for i64
8090 // BUILD_PAIR bit casted to f64.
8092 // This must be an insertion into a zero vector.
8093 SDValue HighElt = N->getOperand(1);
8094 if (!isZeroNode(HighElt))
8097 // Value must be a load.
8098 SDNode *Base = N->getOperand(0).getNode();
8099 if (!isa<LoadSDNode>(Base)) {
8100 if (Base->getOpcode() != ISD::BIT_CONVERT)
8102 Base = Base->getOperand(0).getNode();
8103 if (!isa<LoadSDNode>(Base))
8107 // Transform it into VZEXT_LOAD addr.
8108 LoadSDNode *LD = cast<LoadSDNode>(Base);
8110 // Load must not be an extload.
8111 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
8114 // Load type should legal type so we don't have to legalize it.
8115 if (!TLI.isTypeLegal(VT))
8118 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
8119 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8120 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8121 TargetLowering::TargetLoweringOpt TLO(DAG);
8122 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
8123 DCI.CommitTargetLoweringOpt(TLO);
8127 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8128 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8129 const X86Subtarget *Subtarget) {
8130 DebugLoc DL = N->getDebugLoc();
8131 SDValue Cond = N->getOperand(0);
8132 // Get the LHS/RHS of the select.
8133 SDValue LHS = N->getOperand(1);
8134 SDValue RHS = N->getOperand(2);
8136 // If we have SSE[12] support, try to form min/max nodes.
8137 if (Subtarget->hasSSE2() &&
8138 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8139 Cond.getOpcode() == ISD::SETCC) {
8140 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8142 unsigned Opcode = 0;
8143 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8146 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8149 if (!UnsafeFPMath) break;
8151 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8153 Opcode = X86ISD::FMIN;
8156 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8159 if (!UnsafeFPMath) break;
8161 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8163 Opcode = X86ISD::FMAX;
8166 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8169 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8172 if (!UnsafeFPMath) break;
8174 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8176 Opcode = X86ISD::FMIN;
8179 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8182 if (!UnsafeFPMath) break;
8184 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8186 Opcode = X86ISD::FMAX;
8192 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8195 // If this is a select between two integer constants, try to do some
8197 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8198 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8199 // Don't do this for crazy integer types.
8200 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8201 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8202 // so that TrueC (the true value) is larger than FalseC.
8203 bool NeedsCondInvert = false;
8205 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8206 // Efficiently invertible.
8207 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8208 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8209 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8210 NeedsCondInvert = true;
8211 std::swap(TrueC, FalseC);
8214 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8215 if (FalseC->getAPIntValue() == 0 &&
8216 TrueC->getAPIntValue().isPowerOf2()) {
8217 if (NeedsCondInvert) // Invert the condition if needed.
8218 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8219 DAG.getConstant(1, Cond.getValueType()));
8221 // Zero extend the condition if needed.
8222 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8224 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8225 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8226 DAG.getConstant(ShAmt, MVT::i8));
8229 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8230 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8231 if (NeedsCondInvert) // Invert the condition if needed.
8232 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8233 DAG.getConstant(1, Cond.getValueType()));
8235 // Zero extend the condition if needed.
8236 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8237 FalseC->getValueType(0), Cond);
8238 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8239 SDValue(FalseC, 0));
8242 // Optimize cases that will turn into an LEA instruction. This requires
8243 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8244 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8245 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8246 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8248 bool isFastMultiplier = false;
8250 switch ((unsigned char)Diff) {
8252 case 1: // result = add base, cond
8253 case 2: // result = lea base( , cond*2)
8254 case 3: // result = lea base(cond, cond*2)
8255 case 4: // result = lea base( , cond*4)
8256 case 5: // result = lea base(cond, cond*4)
8257 case 8: // result = lea base( , cond*8)
8258 case 9: // result = lea base(cond, cond*8)
8259 isFastMultiplier = true;
8264 if (isFastMultiplier) {
8265 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8266 if (NeedsCondInvert) // Invert the condition if needed.
8267 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8268 DAG.getConstant(1, Cond.getValueType()));
8270 // Zero extend the condition if needed.
8271 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8273 // Scale the condition by the difference.
8275 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8276 DAG.getConstant(Diff, Cond.getValueType()));
8278 // Add the base if non-zero.
8279 if (FalseC->getAPIntValue() != 0)
8280 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8281 SDValue(FalseC, 0));
8291 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8292 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8293 TargetLowering::DAGCombinerInfo &DCI) {
8294 DebugLoc DL = N->getDebugLoc();
8296 // If the flag operand isn't dead, don't touch this CMOV.
8297 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8300 // If this is a select between two integer constants, try to do some
8301 // optimizations. Note that the operands are ordered the opposite of SELECT
8303 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8304 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8305 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8306 // larger than FalseC (the false value).
8307 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8309 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8310 CC = X86::GetOppositeBranchCondition(CC);
8311 std::swap(TrueC, FalseC);
8314 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8315 // This is efficient for any integer data type (including i8/i16) and
8317 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8318 SDValue Cond = N->getOperand(3);
8319 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8320 DAG.getConstant(CC, MVT::i8), Cond);
8322 // Zero extend the condition if needed.
8323 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8325 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8326 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8327 DAG.getConstant(ShAmt, MVT::i8));
8328 if (N->getNumValues() == 2) // Dead flag value?
8329 return DCI.CombineTo(N, Cond, SDValue());
8333 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8334 // for any integer data type, including i8/i16.
8335 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8336 SDValue Cond = N->getOperand(3);
8337 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8338 DAG.getConstant(CC, MVT::i8), Cond);
8340 // Zero extend the condition if needed.
8341 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8342 FalseC->getValueType(0), Cond);
8343 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8344 SDValue(FalseC, 0));
8346 if (N->getNumValues() == 2) // Dead flag value?
8347 return DCI.CombineTo(N, Cond, SDValue());
8351 // Optimize cases that will turn into an LEA instruction. This requires
8352 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8353 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8354 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8355 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8357 bool isFastMultiplier = false;
8359 switch ((unsigned char)Diff) {
8361 case 1: // result = add base, cond
8362 case 2: // result = lea base( , cond*2)
8363 case 3: // result = lea base(cond, cond*2)
8364 case 4: // result = lea base( , cond*4)
8365 case 5: // result = lea base(cond, cond*4)
8366 case 8: // result = lea base( , cond*8)
8367 case 9: // result = lea base(cond, cond*8)
8368 isFastMultiplier = true;
8373 if (isFastMultiplier) {
8374 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8375 SDValue Cond = N->getOperand(3);
8376 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8377 DAG.getConstant(CC, MVT::i8), Cond);
8378 // Zero extend the condition if needed.
8379 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8381 // Scale the condition by the difference.
8383 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8384 DAG.getConstant(Diff, Cond.getValueType()));
8386 // Add the base if non-zero.
8387 if (FalseC->getAPIntValue() != 0)
8388 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8389 SDValue(FalseC, 0));
8390 if (N->getNumValues() == 2) // Dead flag value?
8391 return DCI.CombineTo(N, Cond, SDValue());
8401 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8403 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8404 const X86Subtarget *Subtarget) {
8405 // On X86 with SSE2 support, we can transform this to a vector shift if
8406 // all elements are shifted by the same amount. We can't do this in legalize
8407 // because the a constant vector is typically transformed to a constant pool
8408 // so we have no knowledge of the shift amount.
8409 if (!Subtarget->hasSSE2())
8412 MVT VT = N->getValueType(0);
8413 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8416 SDValue ShAmtOp = N->getOperand(1);
8417 MVT EltVT = VT.getVectorElementType();
8418 DebugLoc DL = N->getDebugLoc();
8420 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8421 unsigned NumElts = VT.getVectorNumElements();
8423 for (; i != NumElts; ++i) {
8424 SDValue Arg = ShAmtOp.getOperand(i);
8425 if (Arg.getOpcode() == ISD::UNDEF) continue;
8429 for (; i != NumElts; ++i) {
8430 SDValue Arg = ShAmtOp.getOperand(i);
8431 if (Arg.getOpcode() == ISD::UNDEF) continue;
8432 if (Arg != BaseShAmt) {
8436 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8437 isSplatMask(ShAmtOp.getOperand(2).getNode())) {
8438 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8439 DAG.getIntPtrConstant(0));
8443 if (EltVT.bitsGT(MVT::i32))
8444 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8445 else if (EltVT.bitsLT(MVT::i32))
8446 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8448 // The shift amount is identical so we can do a vector shift.
8449 SDValue ValOp = N->getOperand(0);
8450 switch (N->getOpcode()) {
8452 assert(0 && "Unknown shift opcode!");
8455 if (VT == MVT::v2i64)
8456 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8457 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8459 if (VT == MVT::v4i32)
8460 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8461 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8463 if (VT == MVT::v8i16)
8464 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8465 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8469 if (VT == MVT::v4i32)
8470 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8471 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8473 if (VT == MVT::v8i16)
8474 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8475 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8479 if (VT == MVT::v2i64)
8480 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8481 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8483 if (VT == MVT::v4i32)
8484 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8485 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8487 if (VT == MVT::v8i16)
8488 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8489 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8496 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8497 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8498 const X86Subtarget *Subtarget) {
8499 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8500 // the FP state in cases where an emms may be missing.
8501 // A preferable solution to the general problem is to figure out the right
8502 // places to insert EMMS. This qualifies as a quick hack.
8504 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8505 StoreSDNode *St = cast<StoreSDNode>(N);
8506 MVT VT = St->getValue().getValueType();
8507 if (VT.getSizeInBits() != 64)
8510 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2();
8511 if ((VT.isVector() ||
8512 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8513 isa<LoadSDNode>(St->getValue()) &&
8514 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8515 St->getChain().hasOneUse() && !St->isVolatile()) {
8516 SDNode* LdVal = St->getValue().getNode();
8518 int TokenFactorIndex = -1;
8519 SmallVector<SDValue, 8> Ops;
8520 SDNode* ChainVal = St->getChain().getNode();
8521 // Must be a store of a load. We currently handle two cases: the load
8522 // is a direct child, and it's under an intervening TokenFactor. It is
8523 // possible to dig deeper under nested TokenFactors.
8524 if (ChainVal == LdVal)
8525 Ld = cast<LoadSDNode>(St->getChain());
8526 else if (St->getValue().hasOneUse() &&
8527 ChainVal->getOpcode() == ISD::TokenFactor) {
8528 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8529 if (ChainVal->getOperand(i).getNode() == LdVal) {
8530 TokenFactorIndex = i;
8531 Ld = cast<LoadSDNode>(St->getValue());
8533 Ops.push_back(ChainVal->getOperand(i));
8537 if (!Ld || !ISD::isNormalLoad(Ld))
8540 // If this is not the MMX case, i.e. we are just turning i64 load/store
8541 // into f64 load/store, avoid the transformation if there are multiple
8542 // uses of the loaded value.
8543 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8546 DebugLoc LdDL = Ld->getDebugLoc();
8547 DebugLoc StDL = N->getDebugLoc();
8548 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8549 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8551 if (Subtarget->is64Bit() || F64IsLegal) {
8552 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8553 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8554 Ld->getBasePtr(), Ld->getSrcValue(),
8555 Ld->getSrcValueOffset(), Ld->isVolatile(),
8556 Ld->getAlignment());
8557 SDValue NewChain = NewLd.getValue(1);
8558 if (TokenFactorIndex != -1) {
8559 Ops.push_back(NewChain);
8560 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8563 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8564 St->getSrcValue(), St->getSrcValueOffset(),
8565 St->isVolatile(), St->getAlignment());
8568 // Otherwise, lower to two pairs of 32-bit loads / stores.
8569 SDValue LoAddr = Ld->getBasePtr();
8570 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8571 DAG.getConstant(4, MVT::i32));
8573 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8574 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8575 Ld->isVolatile(), Ld->getAlignment());
8576 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8577 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8579 MinAlign(Ld->getAlignment(), 4));
8581 SDValue NewChain = LoLd.getValue(1);
8582 if (TokenFactorIndex != -1) {
8583 Ops.push_back(LoLd);
8584 Ops.push_back(HiLd);
8585 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8589 LoAddr = St->getBasePtr();
8590 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8591 DAG.getConstant(4, MVT::i32));
8593 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8594 St->getSrcValue(), St->getSrcValueOffset(),
8595 St->isVolatile(), St->getAlignment());
8596 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8598 St->getSrcValueOffset() + 4,
8600 MinAlign(St->getAlignment(), 4));
8601 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8606 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8607 /// X86ISD::FXOR nodes.
8608 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8609 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8610 // F[X]OR(0.0, x) -> x
8611 // F[X]OR(x, 0.0) -> x
8612 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8613 if (C->getValueAPF().isPosZero())
8614 return N->getOperand(1);
8615 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8616 if (C->getValueAPF().isPosZero())
8617 return N->getOperand(0);
8621 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8622 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8623 // FAND(0.0, x) -> 0.0
8624 // FAND(x, 0.0) -> 0.0
8625 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8626 if (C->getValueAPF().isPosZero())
8627 return N->getOperand(0);
8628 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8629 if (C->getValueAPF().isPosZero())
8630 return N->getOperand(1);
8634 static SDValue PerformBTCombine(SDNode *N,
8636 TargetLowering::DAGCombinerInfo &DCI) {
8637 // BT ignores high bits in the bit index operand.
8638 SDValue Op1 = N->getOperand(1);
8639 if (Op1.hasOneUse()) {
8640 unsigned BitWidth = Op1.getValueSizeInBits();
8641 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8642 APInt KnownZero, KnownOne;
8643 TargetLowering::TargetLoweringOpt TLO(DAG);
8644 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8645 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8646 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8647 DCI.CommitTargetLoweringOpt(TLO);
8652 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8653 DAGCombinerInfo &DCI) const {
8654 SelectionDAG &DAG = DCI.DAG;
8655 switch (N->getOpcode()) {
8657 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8658 case ISD::BUILD_VECTOR:
8659 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8660 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8661 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8664 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8665 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8667 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8668 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8669 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8675 //===----------------------------------------------------------------------===//
8676 // X86 Inline Assembly Support
8677 //===----------------------------------------------------------------------===//
8679 /// getConstraintType - Given a constraint letter, return the type of
8680 /// constraint it is for this target.
8681 X86TargetLowering::ConstraintType
8682 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8683 if (Constraint.size() == 1) {
8684 switch (Constraint[0]) {
8696 return C_RegisterClass;
8704 return TargetLowering::getConstraintType(Constraint);
8707 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8708 /// with another that has more specific requirements based on the type of the
8709 /// corresponding operand.
8710 const char *X86TargetLowering::
8711 LowerXConstraint(MVT ConstraintVT) const {
8712 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8713 // 'f' like normal targets.
8714 if (ConstraintVT.isFloatingPoint()) {
8715 if (Subtarget->hasSSE2())
8717 if (Subtarget->hasSSE1())
8721 return TargetLowering::LowerXConstraint(ConstraintVT);
8724 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8725 /// vector. If it is invalid, don't add anything to Ops.
8726 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8729 std::vector<SDValue>&Ops,
8730 SelectionDAG &DAG) const {
8731 SDValue Result(0, 0);
8733 switch (Constraint) {
8736 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8737 if (C->getZExtValue() <= 31) {
8738 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8744 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8745 if (C->getZExtValue() <= 63) {
8746 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8752 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8753 if (C->getZExtValue() <= 255) {
8754 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8760 // 32-bit signed value
8761 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8762 const ConstantInt *CI = C->getConstantIntValue();
8763 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8764 // Widen to 64 bits here to get it sign extended.
8765 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8768 // FIXME gcc accepts some relocatable values here too, but only in certain
8769 // memory models; it's complicated.
8774 // 32-bit unsigned value
8775 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8776 const ConstantInt *CI = C->getConstantIntValue();
8777 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8778 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8782 // FIXME gcc accepts some relocatable values here too, but only in certain
8783 // memory models; it's complicated.
8787 // Literal immediates are always ok.
8788 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8789 // Widen to 64 bits here to get it sign extended.
8790 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8794 // If we are in non-pic codegen mode, we allow the address of a global (with
8795 // an optional displacement) to be used with 'i'.
8796 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8799 // Match either (GA) or (GA+C)
8801 Offset = GA->getOffset();
8802 } else if (Op.getOpcode() == ISD::ADD) {
8803 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8804 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8806 Offset = GA->getOffset()+C->getZExtValue();
8808 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8809 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8811 Offset = GA->getOffset()+C->getZExtValue();
8819 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8822 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8828 // Otherwise, not valid for this mode.
8833 if (Result.getNode()) {
8834 Ops.push_back(Result);
8837 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8841 std::vector<unsigned> X86TargetLowering::
8842 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8844 if (Constraint.size() == 1) {
8845 // FIXME: not handling fp-stack yet!
8846 switch (Constraint[0]) { // GCC X86 Constraint Letters
8847 default: break; // Unknown constraint letter
8848 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8851 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8852 else if (VT == MVT::i16)
8853 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8854 else if (VT == MVT::i8)
8855 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8856 else if (VT == MVT::i64)
8857 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8862 return std::vector<unsigned>();
8865 std::pair<unsigned, const TargetRegisterClass*>
8866 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8868 // First, see if this is a constraint that directly corresponds to an LLVM
8870 if (Constraint.size() == 1) {
8871 // GCC Constraint Letters
8872 switch (Constraint[0]) {
8874 case 'r': // GENERAL_REGS
8875 case 'R': // LEGACY_REGS
8876 case 'l': // INDEX_REGS
8878 return std::make_pair(0U, X86::GR8RegisterClass);
8880 return std::make_pair(0U, X86::GR16RegisterClass);
8881 if (VT == MVT::i32 || !Subtarget->is64Bit())
8882 return std::make_pair(0U, X86::GR32RegisterClass);
8883 return std::make_pair(0U, X86::GR64RegisterClass);
8884 case 'f': // FP Stack registers.
8885 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8886 // value to the correct fpstack register class.
8887 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8888 return std::make_pair(0U, X86::RFP32RegisterClass);
8889 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8890 return std::make_pair(0U, X86::RFP64RegisterClass);
8891 return std::make_pair(0U, X86::RFP80RegisterClass);
8892 case 'y': // MMX_REGS if MMX allowed.
8893 if (!Subtarget->hasMMX()) break;
8894 return std::make_pair(0U, X86::VR64RegisterClass);
8895 case 'Y': // SSE_REGS if SSE2 allowed
8896 if (!Subtarget->hasSSE2()) break;
8898 case 'x': // SSE_REGS if SSE1 allowed
8899 if (!Subtarget->hasSSE1()) break;
8901 switch (VT.getSimpleVT()) {
8903 // Scalar SSE types.
8906 return std::make_pair(0U, X86::FR32RegisterClass);
8909 return std::make_pair(0U, X86::FR64RegisterClass);
8917 return std::make_pair(0U, X86::VR128RegisterClass);
8923 // Use the default implementation in TargetLowering to convert the register
8924 // constraint into a member of a register class.
8925 std::pair<unsigned, const TargetRegisterClass*> Res;
8926 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8928 // Not found as a standard register?
8929 if (Res.second == 0) {
8930 // GCC calls "st(0)" just plain "st".
8931 if (StringsEqualNoCase("{st}", Constraint)) {
8932 Res.first = X86::ST0;
8933 Res.second = X86::RFP80RegisterClass;
8935 // 'A' means EAX + EDX.
8936 if (Constraint == "A") {
8937 Res.first = X86::EAX;
8938 Res.second = X86::GRADRegisterClass;
8943 // Otherwise, check to see if this is a register class of the wrong value
8944 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8945 // turn into {ax},{dx}.
8946 if (Res.second->hasType(VT))
8947 return Res; // Correct type already, nothing to do.
8949 // All of the single-register GCC register classes map their values onto
8950 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8951 // really want an 8-bit or 32-bit register, map to the appropriate register
8952 // class and return the appropriate register.
8953 if (Res.second == X86::GR16RegisterClass) {
8954 if (VT == MVT::i8) {
8955 unsigned DestReg = 0;
8956 switch (Res.first) {
8958 case X86::AX: DestReg = X86::AL; break;
8959 case X86::DX: DestReg = X86::DL; break;
8960 case X86::CX: DestReg = X86::CL; break;
8961 case X86::BX: DestReg = X86::BL; break;
8964 Res.first = DestReg;
8965 Res.second = Res.second = X86::GR8RegisterClass;
8967 } else if (VT == MVT::i32) {
8968 unsigned DestReg = 0;
8969 switch (Res.first) {
8971 case X86::AX: DestReg = X86::EAX; break;
8972 case X86::DX: DestReg = X86::EDX; break;
8973 case X86::CX: DestReg = X86::ECX; break;
8974 case X86::BX: DestReg = X86::EBX; break;
8975 case X86::SI: DestReg = X86::ESI; break;
8976 case X86::DI: DestReg = X86::EDI; break;
8977 case X86::BP: DestReg = X86::EBP; break;
8978 case X86::SP: DestReg = X86::ESP; break;
8981 Res.first = DestReg;
8982 Res.second = Res.second = X86::GR32RegisterClass;
8984 } else if (VT == MVT::i64) {
8985 unsigned DestReg = 0;
8986 switch (Res.first) {
8988 case X86::AX: DestReg = X86::RAX; break;
8989 case X86::DX: DestReg = X86::RDX; break;
8990 case X86::CX: DestReg = X86::RCX; break;
8991 case X86::BX: DestReg = X86::RBX; break;
8992 case X86::SI: DestReg = X86::RSI; break;
8993 case X86::DI: DestReg = X86::RDI; break;
8994 case X86::BP: DestReg = X86::RBP; break;
8995 case X86::SP: DestReg = X86::RSP; break;
8998 Res.first = DestReg;
8999 Res.second = Res.second = X86::GR64RegisterClass;
9002 } else if (Res.second == X86::FR32RegisterClass ||
9003 Res.second == X86::FR64RegisterClass ||
9004 Res.second == X86::VR128RegisterClass) {
9005 // Handle references to XMM physical registers that got mapped into the
9006 // wrong class. This can happen with constraints like {xmm0} where the
9007 // target independent register mapper will just pick the first match it can
9008 // find, ignoring the required type.
9010 Res.second = X86::FR32RegisterClass;
9011 else if (VT == MVT::f64)
9012 Res.second = X86::FR64RegisterClass;
9013 else if (X86::VR128RegisterClass->hasType(VT))
9014 Res.second = X86::VR128RegisterClass;
9020 //===----------------------------------------------------------------------===//
9021 // X86 Widen vector type
9022 //===----------------------------------------------------------------------===//
9024 /// getWidenVectorType: given a vector type, returns the type to widen
9025 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9026 /// If there is no vector type that we want to widen to, returns MVT::Other
9027 /// When and where to widen is target dependent based on the cost of
9028 /// scalarizing vs using the wider vector type.
9030 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9031 assert(VT.isVector());
9032 if (isTypeLegal(VT))
9035 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9036 // type based on element type. This would speed up our search (though
9037 // it may not be worth it since the size of the list is relatively
9039 MVT EltVT = VT.getVectorElementType();
9040 unsigned NElts = VT.getVectorNumElements();
9042 // On X86, it make sense to widen any vector wider than 1
9046 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9047 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9048 MVT SVT = (MVT::SimpleValueType)nVT;
9050 if (isTypeLegal(SVT) &&
9051 SVT.getVectorElementType() == EltVT &&
9052 SVT.getVectorNumElements() > NElts)