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::i64 , Expand);
117 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
119 if (X86ScalarSSEf64) {
120 // We have an impenetrably clever algorithm for ui64->double only.
121 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
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);
129 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
131 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
132 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
133 // SSE has no i16 to fp conversion, only i32
134 if (X86ScalarSSEf32) {
135 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
136 // f32 and f64 cases are Legal, f80 case is not
137 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
139 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
140 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
143 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
144 // are Legal, f80 is custom lowered.
145 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
146 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
148 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
150 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
151 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
153 if (X86ScalarSSEf32) {
154 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
155 // f32 and f64 cases are Legal, f80 case is not
156 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
158 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
159 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
162 // Handle FP_TO_UINT by promoting the destination to a larger signed
164 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
165 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
166 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
168 if (Subtarget->is64Bit()) {
169 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
170 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
172 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
173 // Expand FP_TO_UINT into a select.
174 // FIXME: We would like to use a Custom expander here eventually to do
175 // the optimal thing for SSE vs. the default expansion in the legalizer.
176 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
178 // With SSE3 we can use fisttpll to convert to a signed i64.
179 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
182 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
183 if (!X86ScalarSSEf64) {
184 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
185 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
188 // Scalar integer divide and remainder are lowered to use operations that
189 // produce two results, to match the available instructions. This exposes
190 // the two-result form to trivial CSE, which is able to combine x/y and x%y
191 // into a single instruction.
193 // Scalar integer multiply-high is also lowered to use two-result
194 // operations, to match the available instructions. However, plain multiply
195 // (low) operations are left as Legal, as there are single-result
196 // instructions for this in x86. Using the two-result multiply instructions
197 // when both high and low results are needed must be arranged by dagcombine.
198 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
199 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
200 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
201 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
202 setOperationAction(ISD::SREM , MVT::i8 , Expand);
203 setOperationAction(ISD::UREM , MVT::i8 , Expand);
204 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
205 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
206 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
207 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
208 setOperationAction(ISD::SREM , MVT::i16 , Expand);
209 setOperationAction(ISD::UREM , MVT::i16 , Expand);
210 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
211 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
212 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
213 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
214 setOperationAction(ISD::SREM , MVT::i32 , Expand);
215 setOperationAction(ISD::UREM , MVT::i32 , Expand);
216 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
217 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
218 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
219 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
220 setOperationAction(ISD::SREM , MVT::i64 , Expand);
221 setOperationAction(ISD::UREM , MVT::i64 , Expand);
223 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
224 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
225 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
226 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
227 if (Subtarget->is64Bit())
228 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
229 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
230 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
231 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
232 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
233 setOperationAction(ISD::FREM , MVT::f32 , Expand);
234 setOperationAction(ISD::FREM , MVT::f64 , Expand);
235 setOperationAction(ISD::FREM , MVT::f80 , Expand);
236 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
238 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
239 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
240 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
241 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
242 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
243 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
244 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
245 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
246 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
247 if (Subtarget->is64Bit()) {
248 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
249 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
250 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
253 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
254 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
256 // These should be promoted to a larger select which is supported.
257 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
258 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
259 // X86 wants to expand cmov itself.
260 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
261 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
262 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
263 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
264 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
265 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
266 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
267 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
268 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
269 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
270 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
271 if (Subtarget->is64Bit()) {
272 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
275 // X86 ret instruction may pop stack.
276 setOperationAction(ISD::RET , MVT::Other, Custom);
277 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
280 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
281 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
282 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
283 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
284 if (Subtarget->is64Bit())
285 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
286 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
289 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
290 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
291 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
293 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
294 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
295 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
296 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
297 if (Subtarget->is64Bit()) {
298 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
299 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
300 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
303 if (Subtarget->hasSSE1())
304 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
306 if (!Subtarget->hasSSE2())
307 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
309 // Expand certain atomics
310 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
311 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
312 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
313 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
315 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
316 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
317 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
318 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
320 if (!Subtarget->is64Bit()) {
321 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
327 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
330 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
331 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
332 // FIXME - use subtarget debug flags
333 if (!Subtarget->isTargetDarwin() &&
334 !Subtarget->isTargetELF() &&
335 !Subtarget->isTargetCygMing()) {
336 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
337 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
340 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
341 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
342 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
343 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
344 if (Subtarget->is64Bit()) {
345 setExceptionPointerRegister(X86::RAX);
346 setExceptionSelectorRegister(X86::RDX);
348 setExceptionPointerRegister(X86::EAX);
349 setExceptionSelectorRegister(X86::EDX);
351 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
352 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
354 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
356 setOperationAction(ISD::TRAP, MVT::Other, Legal);
358 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
359 setOperationAction(ISD::VASTART , MVT::Other, Custom);
360 setOperationAction(ISD::VAEND , MVT::Other, Expand);
361 if (Subtarget->is64Bit()) {
362 setOperationAction(ISD::VAARG , MVT::Other, Custom);
363 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
365 setOperationAction(ISD::VAARG , MVT::Other, Expand);
366 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
369 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
370 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
371 if (Subtarget->is64Bit())
372 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
373 if (Subtarget->isTargetCygMing())
374 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
376 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
378 if (!UseSoftFloat && X86ScalarSSEf64) {
379 // f32 and f64 use SSE.
380 // Set up the FP register classes.
381 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
382 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
384 // Use ANDPD to simulate FABS.
385 setOperationAction(ISD::FABS , MVT::f64, Custom);
386 setOperationAction(ISD::FABS , MVT::f32, Custom);
388 // Use XORP to simulate FNEG.
389 setOperationAction(ISD::FNEG , MVT::f64, Custom);
390 setOperationAction(ISD::FNEG , MVT::f32, Custom);
392 // Use ANDPD and ORPD to simulate FCOPYSIGN.
393 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
394 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
396 // We don't support sin/cos/fmod
397 setOperationAction(ISD::FSIN , MVT::f64, Expand);
398 setOperationAction(ISD::FCOS , MVT::f64, Expand);
399 setOperationAction(ISD::FSIN , MVT::f32, Expand);
400 setOperationAction(ISD::FCOS , MVT::f32, Expand);
402 // Expand FP immediates into loads from the stack, except for the special
404 addLegalFPImmediate(APFloat(+0.0)); // xorpd
405 addLegalFPImmediate(APFloat(+0.0f)); // xorps
407 // Floating truncations from f80 and extensions to f80 go through memory.
408 // If optimizing, we lie about this though and handle it in
409 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
411 setConvertAction(MVT::f32, MVT::f80, Expand);
412 setConvertAction(MVT::f64, MVT::f80, Expand);
413 setConvertAction(MVT::f80, MVT::f32, Expand);
414 setConvertAction(MVT::f80, MVT::f64, Expand);
416 } else if (!UseSoftFloat && X86ScalarSSEf32) {
417 // Use SSE for f32, x87 for f64.
418 // Set up the FP register classes.
419 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
420 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
422 // Use ANDPS to simulate FABS.
423 setOperationAction(ISD::FABS , MVT::f32, Custom);
425 // Use XORP to simulate FNEG.
426 setOperationAction(ISD::FNEG , MVT::f32, Custom);
428 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
430 // Use ANDPS and ORPS to simulate FCOPYSIGN.
431 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
432 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
434 // We don't support sin/cos/fmod
435 setOperationAction(ISD::FSIN , MVT::f32, Expand);
436 setOperationAction(ISD::FCOS , MVT::f32, Expand);
438 // Special cases we handle for FP constants.
439 addLegalFPImmediate(APFloat(+0.0f)); // xorps
440 addLegalFPImmediate(APFloat(+0.0)); // FLD0
441 addLegalFPImmediate(APFloat(+1.0)); // FLD1
442 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
443 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
445 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
446 // this though and handle it in InstructionSelectPreprocess so that
447 // dagcombine2 can hack on these.
449 setConvertAction(MVT::f32, MVT::f64, Expand);
450 setConvertAction(MVT::f32, MVT::f80, Expand);
451 setConvertAction(MVT::f80, MVT::f32, Expand);
452 setConvertAction(MVT::f64, MVT::f32, Expand);
453 // And x87->x87 truncations also.
454 setConvertAction(MVT::f80, MVT::f64, Expand);
458 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
459 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
461 } else if (!UseSoftFloat) {
462 // f32 and f64 in x87.
463 // Set up the FP register classes.
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
465 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
467 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
468 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
472 // Floating truncations go through memory. If optimizing, we lie about
473 // this though and handle it in InstructionSelectPreprocess so that
474 // dagcombine2 can hack on these.
476 setConvertAction(MVT::f80, MVT::f32, Expand);
477 setConvertAction(MVT::f64, MVT::f32, Expand);
478 setConvertAction(MVT::f80, MVT::f64, Expand);
482 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
483 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
485 addLegalFPImmediate(APFloat(+0.0)); // FLD0
486 addLegalFPImmediate(APFloat(+1.0)); // FLD1
487 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
488 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
489 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
490 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
491 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
492 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
495 // Long double always uses X87.
497 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
498 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
502 APFloat TmpFlt(+0.0);
503 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
505 addLegalFPImmediate(TmpFlt); // FLD0
507 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
508 APFloat TmpFlt2(+1.0);
509 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
511 addLegalFPImmediate(TmpFlt2); // FLD1
512 TmpFlt2.changeSign();
513 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
517 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
518 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
522 // Always use a library call for pow.
523 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
524 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
525 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
527 setOperationAction(ISD::FLOG, MVT::f80, Expand);
528 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
529 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
530 setOperationAction(ISD::FEXP, MVT::f80, Expand);
531 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
533 // First set operation action for all vector types to either promote
534 // (for widening) or expand (for scalarization). Then we will selectively
535 // turn on ones that can be effectively codegen'd.
536 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
537 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
538 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
553 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
583 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
584 // with -msoft-float, disable use of MMX as well.
585 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
586 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
587 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
588 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
589 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
590 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
592 // FIXME: add MMX packed arithmetics
594 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
595 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
596 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
597 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
599 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
600 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
601 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
602 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
604 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
605 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
607 setOperationAction(ISD::AND, MVT::v8i8, Promote);
608 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
609 setOperationAction(ISD::AND, MVT::v4i16, Promote);
610 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
611 setOperationAction(ISD::AND, MVT::v2i32, Promote);
612 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
613 setOperationAction(ISD::AND, MVT::v1i64, Legal);
615 setOperationAction(ISD::OR, MVT::v8i8, Promote);
616 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
617 setOperationAction(ISD::OR, MVT::v4i16, Promote);
618 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
619 setOperationAction(ISD::OR, MVT::v2i32, Promote);
620 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
621 setOperationAction(ISD::OR, MVT::v1i64, Legal);
623 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
624 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
625 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
626 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
627 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
628 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
629 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
631 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
632 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
633 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
634 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
635 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
636 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
637 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
638 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
639 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
641 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
642 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
643 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
644 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
645 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
647 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
648 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
649 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
650 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
652 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
653 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
654 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
655 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
659 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
660 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
661 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
662 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
663 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
664 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
667 if (!UseSoftFloat && Subtarget->hasSSE1()) {
668 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
670 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
671 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
672 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
673 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
674 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
675 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
676 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
679 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
680 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
681 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
684 if (!UseSoftFloat && Subtarget->hasSSE2()) {
685 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
687 // FIXME: Unfortunately -soft-float means XMM registers cannot be used even
688 // for integer operations.
689 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
690 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
691 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
692 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
694 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
695 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
696 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
697 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
698 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
699 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
700 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
701 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
702 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
703 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
704 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
705 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
706 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
707 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
708 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
709 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
711 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
712 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
714 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
716 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
717 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
718 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
719 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
720 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
722 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
723 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
724 MVT VT = (MVT::SimpleValueType)i;
725 // Do not attempt to custom lower non-power-of-2 vectors
726 if (!isPowerOf2_32(VT.getVectorNumElements()))
728 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
729 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
732 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
733 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
734 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
735 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
736 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
737 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
738 if (Subtarget->is64Bit()) {
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
740 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
743 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
744 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
745 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
746 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
747 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
748 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
749 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
750 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
751 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
752 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
753 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
754 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
757 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
759 // Custom lower v2i64 and v2f64 selects.
760 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
761 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
762 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
763 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
767 if (Subtarget->hasSSE41()) {
768 // FIXME: Do we need to handle scalar-to-vector here?
769 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
771 // i8 and i16 vectors are custom , because the source register and source
772 // source memory operand types are not the same width. f32 vectors are
773 // custom since the immediate controlling the insert encodes additional
775 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
776 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
777 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
778 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
780 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
781 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
785 if (Subtarget->is64Bit()) {
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
791 if (Subtarget->hasSSE42()) {
792 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
795 // We want to custom lower some of our intrinsics.
796 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
798 // Add/Sub/Mul with overflow operations are custom lowered.
799 setOperationAction(ISD::SADDO, MVT::i32, Custom);
800 setOperationAction(ISD::SADDO, MVT::i64, Custom);
801 setOperationAction(ISD::UADDO, MVT::i32, Custom);
802 setOperationAction(ISD::UADDO, MVT::i64, Custom);
803 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
804 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
805 setOperationAction(ISD::USUBO, MVT::i32, Custom);
806 setOperationAction(ISD::USUBO, MVT::i64, Custom);
807 setOperationAction(ISD::SMULO, MVT::i32, Custom);
808 setOperationAction(ISD::SMULO, MVT::i64, Custom);
809 setOperationAction(ISD::UMULO, MVT::i32, Custom);
810 setOperationAction(ISD::UMULO, MVT::i64, Custom);
812 // We have target-specific dag combine patterns for the following nodes:
813 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
814 setTargetDAGCombine(ISD::BUILD_VECTOR);
815 setTargetDAGCombine(ISD::SELECT);
816 setTargetDAGCombine(ISD::SHL);
817 setTargetDAGCombine(ISD::SRA);
818 setTargetDAGCombine(ISD::SRL);
819 setTargetDAGCombine(ISD::STORE);
821 computeRegisterProperties();
823 // FIXME: These should be based on subtarget info. Plus, the values should
824 // be smaller when we are in optimizing for size mode.
825 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
826 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
827 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
828 allowUnalignedMemoryAccesses = true; // x86 supports it!
829 setPrefLoopAlignment(16);
833 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
838 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
839 /// the desired ByVal argument alignment.
840 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
843 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
844 if (VTy->getBitWidth() == 128)
846 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
847 unsigned EltAlign = 0;
848 getMaxByValAlign(ATy->getElementType(), EltAlign);
849 if (EltAlign > MaxAlign)
851 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
852 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
853 unsigned EltAlign = 0;
854 getMaxByValAlign(STy->getElementType(i), EltAlign);
855 if (EltAlign > MaxAlign)
864 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
865 /// function arguments in the caller parameter area. For X86, aggregates
866 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
867 /// are at 4-byte boundaries.
868 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
869 if (Subtarget->is64Bit()) {
870 // Max of 8 and alignment of type.
871 unsigned TyAlign = TD->getABITypeAlignment(Ty);
878 if (Subtarget->hasSSE1())
879 getMaxByValAlign(Ty, Align);
883 /// getOptimalMemOpType - Returns the target specific optimal type for load
884 /// and store operations as a result of memset, memcpy, and memmove
885 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
888 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
889 bool isSrcConst, bool isSrcStr) const {
890 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
891 // linux. This is because the stack realignment code can't handle certain
892 // cases like PR2962. This should be removed when PR2962 is fixed.
893 if (Subtarget->getStackAlignment() >= 16) {
894 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
896 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
899 if (Subtarget->is64Bit() && Size >= 8)
905 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
907 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
908 SelectionDAG &DAG) const {
909 if (usesGlobalOffsetTable())
910 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
911 if (!Subtarget->isPICStyleRIPRel())
912 // This doesn't have DebugLoc associated with it, but is not really the
913 // same as a Register.
914 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
919 //===----------------------------------------------------------------------===//
920 // Return Value Calling Convention Implementation
921 //===----------------------------------------------------------------------===//
923 #include "X86GenCallingConv.inc"
925 /// LowerRET - Lower an ISD::RET node.
926 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
927 DebugLoc dl = Op.getDebugLoc();
928 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
930 SmallVector<CCValAssign, 16> RVLocs;
931 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
932 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
933 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
934 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
936 // If this is the first return lowered for this function, add the regs to the
937 // liveout set for the function.
938 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
939 for (unsigned i = 0; i != RVLocs.size(); ++i)
940 if (RVLocs[i].isRegLoc())
941 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
943 SDValue Chain = Op.getOperand(0);
945 // Handle tail call return.
946 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
947 if (Chain.getOpcode() == X86ISD::TAILCALL) {
948 SDValue TailCall = Chain;
949 SDValue TargetAddress = TailCall.getOperand(1);
950 SDValue StackAdjustment = TailCall.getOperand(2);
951 assert(((TargetAddress.getOpcode() == ISD::Register &&
952 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
953 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
954 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
955 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
956 "Expecting an global address, external symbol, or register");
957 assert(StackAdjustment.getOpcode() == ISD::Constant &&
958 "Expecting a const value");
960 SmallVector<SDValue,8> Operands;
961 Operands.push_back(Chain.getOperand(0));
962 Operands.push_back(TargetAddress);
963 Operands.push_back(StackAdjustment);
964 // Copy registers used by the call. Last operand is a flag so it is not
966 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
967 Operands.push_back(Chain.getOperand(i));
969 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
976 SmallVector<SDValue, 6> RetOps;
977 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
978 // Operand #1 = Bytes To Pop
979 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
981 // Copy the result values into the output registers.
982 for (unsigned i = 0; i != RVLocs.size(); ++i) {
983 CCValAssign &VA = RVLocs[i];
984 assert(VA.isRegLoc() && "Can only return in registers!");
985 SDValue ValToCopy = Op.getOperand(i*2+1);
987 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
988 // the RET instruction and handled by the FP Stackifier.
989 if (VA.getLocReg() == X86::ST0 ||
990 VA.getLocReg() == X86::ST1) {
991 // If this is a copy from an xmm register to ST(0), use an FPExtend to
992 // change the value to the FP stack register class.
993 if (isScalarFPTypeInSSEReg(VA.getValVT()))
994 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
995 RetOps.push_back(ValToCopy);
996 // Don't emit a copytoreg.
1000 // 64-bit vector (MMX) values are returned in RAX.
1001 if (Subtarget->is64Bit()) {
1002 MVT ValVT = ValToCopy.getValueType();
1003 if (VA.getLocReg() == X86::RAX &&
1004 ValVT.isVector() && ValVT.getSizeInBits() == 64)
1005 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1008 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1009 Flag = Chain.getValue(1);
1012 // The x86-64 ABI for returning structs by value requires that we copy
1013 // the sret argument into %rax for the return. We saved the argument into
1014 // a virtual register in the entry block, so now we copy the value out
1016 if (Subtarget->is64Bit() &&
1017 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1018 MachineFunction &MF = DAG.getMachineFunction();
1019 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1020 unsigned Reg = FuncInfo->getSRetReturnReg();
1022 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1023 FuncInfo->setSRetReturnReg(Reg);
1025 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1027 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1028 Flag = Chain.getValue(1);
1031 RetOps[0] = Chain; // Update chain.
1033 // Add the flag if we have it.
1035 RetOps.push_back(Flag);
1037 return DAG.getNode(X86ISD::RET_FLAG, dl,
1038 MVT::Other, &RetOps[0], RetOps.size());
1042 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1043 /// appropriate copies out of appropriate physical registers. This assumes that
1044 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1045 /// being lowered. The returns a SDNode with the same number of values as the
1047 SDNode *X86TargetLowering::
1048 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1049 unsigned CallingConv, SelectionDAG &DAG) {
1051 DebugLoc dl = TheCall->getDebugLoc();
1052 // Assign locations to each value returned by this call.
1053 SmallVector<CCValAssign, 16> RVLocs;
1054 bool isVarArg = TheCall->isVarArg();
1055 bool Is64Bit = Subtarget->is64Bit();
1056 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1057 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1059 SmallVector<SDValue, 8> ResultVals;
1061 // Copy all of the result registers out of their specified physreg.
1062 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1063 CCValAssign &VA = RVLocs[i];
1064 MVT CopyVT = VA.getValVT();
1066 // If this is x86-64, and we disabled SSE, we can't return FP values
1067 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1068 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1069 cerr << "SSE register return with SSE disabled\n";
1073 // If this is a call to a function that returns an fp value on the floating
1074 // point stack, but where we prefer to use the value in xmm registers, copy
1075 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1076 if ((VA.getLocReg() == X86::ST0 ||
1077 VA.getLocReg() == X86::ST1) &&
1078 isScalarFPTypeInSSEReg(VA.getValVT())) {
1083 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1084 // For x86-64, MMX values are returned in RAX.
1085 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1086 MVT::i64, InFlag).getValue(1);
1087 Val = Chain.getValue(0);
1088 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1090 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1091 CopyVT, InFlag).getValue(1);
1092 Val = Chain.getValue(0);
1094 InFlag = Chain.getValue(2);
1096 if (CopyVT != VA.getValVT()) {
1097 // Round the F80 the right size, which also moves to the appropriate xmm
1099 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1100 // This truncation won't change the value.
1101 DAG.getIntPtrConstant(1));
1104 ResultVals.push_back(Val);
1107 // Merge everything together with a MERGE_VALUES node.
1108 ResultVals.push_back(Chain);
1109 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1110 &ResultVals[0], ResultVals.size()).getNode();
1114 //===----------------------------------------------------------------------===//
1115 // C & StdCall & Fast Calling Convention implementation
1116 //===----------------------------------------------------------------------===//
1117 // StdCall calling convention seems to be standard for many Windows' API
1118 // routines and around. It differs from C calling convention just a little:
1119 // callee should clean up the stack, not caller. Symbols should be also
1120 // decorated in some fancy way :) It doesn't support any vector arguments.
1121 // For info on fast calling convention see Fast Calling Convention (tail call)
1122 // implementation LowerX86_32FastCCCallTo.
1124 /// AddLiveIn - This helper function adds the specified physical register to the
1125 /// MachineFunction as a live in value. It also creates a corresponding virtual
1126 /// register for it.
1127 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1128 const TargetRegisterClass *RC) {
1129 assert(RC->contains(PReg) && "Not the correct regclass!");
1130 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1131 MF.getRegInfo().addLiveIn(PReg, VReg);
1135 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1137 static bool CallIsStructReturn(CallSDNode *TheCall) {
1138 unsigned NumOps = TheCall->getNumArgs();
1142 return TheCall->getArgFlags(0).isSRet();
1145 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1146 /// return semantics.
1147 static bool ArgsAreStructReturn(SDValue Op) {
1148 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1152 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1155 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1156 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1158 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1162 switch (CallingConv) {
1165 case CallingConv::X86_StdCall:
1166 return !Subtarget->is64Bit();
1167 case CallingConv::X86_FastCall:
1168 return !Subtarget->is64Bit();
1169 case CallingConv::Fast:
1170 return PerformTailCallOpt;
1174 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1175 /// given CallingConvention value.
1176 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1177 if (Subtarget->is64Bit()) {
1178 if (Subtarget->isTargetWin64())
1179 return CC_X86_Win64_C;
1180 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1181 return CC_X86_64_TailCall;
1186 if (CC == CallingConv::X86_FastCall)
1187 return CC_X86_32_FastCall;
1188 else if (CC == CallingConv::Fast)
1189 return CC_X86_32_FastCC;
1194 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1195 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1197 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1198 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1199 if (CC == CallingConv::X86_FastCall)
1201 else if (CC == CallingConv::X86_StdCall)
1207 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1208 /// in a register before calling.
1209 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1210 return !IsTailCall && !Is64Bit &&
1211 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1212 Subtarget->isPICStyleGOT();
1215 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1216 /// address to be loaded in a register.
1218 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1219 return !Is64Bit && IsTailCall &&
1220 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1221 Subtarget->isPICStyleGOT();
1224 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1225 /// by "Src" to address "Dst" with size and alignment information specified by
1226 /// the specific parameter attribute. The copy will be passed as a byval
1227 /// function parameter.
1229 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1230 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1232 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1233 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1234 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1237 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1238 const CCValAssign &VA,
1239 MachineFrameInfo *MFI,
1241 SDValue Root, unsigned i) {
1242 // Create the nodes corresponding to a load from this parameter slot.
1243 ISD::ArgFlagsTy Flags =
1244 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1245 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1246 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1248 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1249 // changed with more analysis.
1250 // In case of tail call optimization mark all arguments mutable. Since they
1251 // could be overwritten by lowering of arguments in case of a tail call.
1252 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1253 VA.getLocMemOffset(), isImmutable);
1254 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1255 if (Flags.isByVal())
1257 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1258 PseudoSourceValue::getFixedStack(FI), 0);
1262 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1263 MachineFunction &MF = DAG.getMachineFunction();
1264 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1265 DebugLoc dl = Op.getDebugLoc();
1267 const Function* Fn = MF.getFunction();
1268 if (Fn->hasExternalLinkage() &&
1269 Subtarget->isTargetCygMing() &&
1270 Fn->getName() == "main")
1271 FuncInfo->setForceFramePointer(true);
1273 // Decorate the function name.
1274 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1276 MachineFrameInfo *MFI = MF.getFrameInfo();
1277 SDValue Root = Op.getOperand(0);
1278 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1279 unsigned CC = MF.getFunction()->getCallingConv();
1280 bool Is64Bit = Subtarget->is64Bit();
1281 bool IsWin64 = Subtarget->isTargetWin64();
1283 assert(!(isVarArg && CC == CallingConv::Fast) &&
1284 "Var args not supported with calling convention fastcc");
1286 // Assign locations to all of the incoming arguments.
1287 SmallVector<CCValAssign, 16> ArgLocs;
1288 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1289 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1291 SmallVector<SDValue, 8> ArgValues;
1292 unsigned LastVal = ~0U;
1293 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1294 CCValAssign &VA = ArgLocs[i];
1295 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1297 assert(VA.getValNo() != LastVal &&
1298 "Don't support value assigned to multiple locs yet");
1299 LastVal = VA.getValNo();
1301 if (VA.isRegLoc()) {
1302 MVT RegVT = VA.getLocVT();
1303 TargetRegisterClass *RC = NULL;
1304 if (RegVT == MVT::i32)
1305 RC = X86::GR32RegisterClass;
1306 else if (Is64Bit && RegVT == MVT::i64)
1307 RC = X86::GR64RegisterClass;
1308 else if (RegVT == MVT::f32)
1309 RC = X86::FR32RegisterClass;
1310 else if (RegVT == MVT::f64)
1311 RC = X86::FR64RegisterClass;
1312 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1313 RC = X86::VR128RegisterClass;
1314 else if (RegVT.isVector()) {
1315 assert(RegVT.getSizeInBits() == 64);
1317 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1319 // Darwin calling convention passes MMX values in either GPRs or
1320 // XMMs in x86-64. Other targets pass them in memory.
1321 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1322 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1325 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1330 assert(0 && "Unknown argument type!");
1333 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1334 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1336 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1337 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1339 if (VA.getLocInfo() == CCValAssign::SExt)
1340 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1341 DAG.getValueType(VA.getValVT()));
1342 else if (VA.getLocInfo() == CCValAssign::ZExt)
1343 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1344 DAG.getValueType(VA.getValVT()));
1346 if (VA.getLocInfo() != CCValAssign::Full)
1347 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1349 // Handle MMX values passed in GPRs.
1350 if (Is64Bit && RegVT != VA.getLocVT()) {
1351 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1352 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1353 else if (RC == X86::VR128RegisterClass) {
1354 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1355 ArgValue, DAG.getConstant(0, MVT::i64));
1356 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1360 ArgValues.push_back(ArgValue);
1362 assert(VA.isMemLoc());
1363 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1367 // The x86-64 ABI for returning structs by value requires that we copy
1368 // the sret argument into %rax for the return. Save the argument into
1369 // a virtual register so that we can access it from the return points.
1370 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1371 MachineFunction &MF = DAG.getMachineFunction();
1372 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1373 unsigned Reg = FuncInfo->getSRetReturnReg();
1375 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1376 FuncInfo->setSRetReturnReg(Reg);
1378 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1379 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1382 unsigned StackSize = CCInfo.getNextStackOffset();
1383 // align stack specially for tail calls
1384 if (PerformTailCallOpt && CC == CallingConv::Fast)
1385 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1387 // If the function takes variable number of arguments, make a frame index for
1388 // the start of the first vararg value... for expansion of llvm.va_start.
1390 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1391 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1394 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1396 // FIXME: We should really autogenerate these arrays
1397 static const unsigned GPR64ArgRegsWin64[] = {
1398 X86::RCX, X86::RDX, X86::R8, X86::R9
1400 static const unsigned XMMArgRegsWin64[] = {
1401 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1403 static const unsigned GPR64ArgRegs64Bit[] = {
1404 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1406 static const unsigned XMMArgRegs64Bit[] = {
1407 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1408 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1410 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1413 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1414 GPR64ArgRegs = GPR64ArgRegsWin64;
1415 XMMArgRegs = XMMArgRegsWin64;
1417 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1418 GPR64ArgRegs = GPR64ArgRegs64Bit;
1419 XMMArgRegs = XMMArgRegs64Bit;
1421 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1423 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1426 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1427 "SSE register cannot be used when SSE is disabled!");
1428 assert(!(NumXMMRegs && UseSoftFloat) &&
1429 "SSE register cannot be used when SSE is disabled!");
1430 if (UseSoftFloat || !Subtarget->hasSSE1()) {
1431 // Kernel mode asks for SSE to be disabled, so don't push them
1433 TotalNumXMMRegs = 0;
1435 // For X86-64, if there are vararg parameters that are passed via
1436 // registers, then we must store them to their spots on the stack so they
1437 // may be loaded by deferencing the result of va_next.
1438 VarArgsGPOffset = NumIntRegs * 8;
1439 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1440 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1441 TotalNumXMMRegs * 16, 16);
1443 // Store the integer parameter registers.
1444 SmallVector<SDValue, 8> MemOps;
1445 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1446 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1447 DAG.getIntPtrConstant(VarArgsGPOffset));
1448 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1449 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1450 X86::GR64RegisterClass);
1451 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1453 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1454 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1455 MemOps.push_back(Store);
1456 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1457 DAG.getIntPtrConstant(8));
1460 // Now store the XMM (fp + vector) parameter registers.
1461 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1462 DAG.getIntPtrConstant(VarArgsFPOffset));
1463 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1464 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1465 X86::VR128RegisterClass);
1466 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1468 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1469 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1470 MemOps.push_back(Store);
1471 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1472 DAG.getIntPtrConstant(16));
1474 if (!MemOps.empty())
1475 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1476 &MemOps[0], MemOps.size());
1480 ArgValues.push_back(Root);
1482 // Some CCs need callee pop.
1483 if (IsCalleePop(isVarArg, CC)) {
1484 BytesToPopOnReturn = StackSize; // Callee pops everything.
1485 BytesCallerReserves = 0;
1487 BytesToPopOnReturn = 0; // Callee pops nothing.
1488 // If this is an sret function, the return should pop the hidden pointer.
1489 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1490 BytesToPopOnReturn = 4;
1491 BytesCallerReserves = StackSize;
1495 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1496 if (CC == CallingConv::X86_FastCall)
1497 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1500 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1502 // Return the new list of results.
1503 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1504 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1508 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1509 const SDValue &StackPtr,
1510 const CCValAssign &VA,
1512 SDValue Arg, ISD::ArgFlagsTy Flags) {
1513 DebugLoc dl = TheCall->getDebugLoc();
1514 unsigned LocMemOffset = VA.getLocMemOffset();
1515 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1516 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1517 if (Flags.isByVal()) {
1518 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1520 return DAG.getStore(Chain, dl, Arg, PtrOff,
1521 PseudoSourceValue::getStack(), LocMemOffset);
1524 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1525 /// optimization is performed and it is required.
1527 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1528 SDValue &OutRetAddr,
1534 if (!IsTailCall || FPDiff==0) return Chain;
1536 // Adjust the Return address stack slot.
1537 MVT VT = getPointerTy();
1538 OutRetAddr = getReturnAddressFrameIndex(DAG);
1540 // Load the "old" Return address.
1541 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1542 return SDValue(OutRetAddr.getNode(), 1);
1545 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1546 /// optimization is performed and it is required (FPDiff!=0).
1548 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1549 SDValue Chain, SDValue RetAddrFrIdx,
1550 bool Is64Bit, int FPDiff, DebugLoc dl) {
1551 // Store the return address to the appropriate stack slot.
1552 if (!FPDiff) return Chain;
1553 // Calculate the new stack slot for the return address.
1554 int SlotSize = Is64Bit ? 8 : 4;
1555 int NewReturnAddrFI =
1556 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1557 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1558 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1559 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1560 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1564 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1565 MachineFunction &MF = DAG.getMachineFunction();
1566 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1567 SDValue Chain = TheCall->getChain();
1568 unsigned CC = TheCall->getCallingConv();
1569 bool isVarArg = TheCall->isVarArg();
1570 bool IsTailCall = TheCall->isTailCall() &&
1571 CC == CallingConv::Fast && PerformTailCallOpt;
1572 SDValue Callee = TheCall->getCallee();
1573 bool Is64Bit = Subtarget->is64Bit();
1574 bool IsStructRet = CallIsStructReturn(TheCall);
1575 DebugLoc dl = TheCall->getDebugLoc();
1577 assert(!(isVarArg && CC == CallingConv::Fast) &&
1578 "Var args not supported with calling convention fastcc");
1580 // Analyze operands of the call, assigning locations to each operand.
1581 SmallVector<CCValAssign, 16> ArgLocs;
1582 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1583 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1585 // Get a count of how many bytes are to be pushed on the stack.
1586 unsigned NumBytes = CCInfo.getNextStackOffset();
1587 if (PerformTailCallOpt && CC == CallingConv::Fast)
1588 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1592 // Lower arguments at fp - stackoffset + fpdiff.
1593 unsigned NumBytesCallerPushed =
1594 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1595 FPDiff = NumBytesCallerPushed - NumBytes;
1597 // Set the delta of movement of the returnaddr stackslot.
1598 // But only set if delta is greater than previous delta.
1599 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1600 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1603 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1605 SDValue RetAddrFrIdx;
1606 // Load return adress for tail calls.
1607 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1610 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1611 SmallVector<SDValue, 8> MemOpChains;
1614 // Walk the register/memloc assignments, inserting copies/loads. In the case
1615 // of tail call optimization arguments are handle later.
1616 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1617 CCValAssign &VA = ArgLocs[i];
1618 SDValue Arg = TheCall->getArg(i);
1619 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1620 bool isByVal = Flags.isByVal();
1622 // Promote the value if needed.
1623 switch (VA.getLocInfo()) {
1624 default: assert(0 && "Unknown loc info!");
1625 case CCValAssign::Full: break;
1626 case CCValAssign::SExt:
1627 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1629 case CCValAssign::ZExt:
1630 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1632 case CCValAssign::AExt:
1633 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1637 if (VA.isRegLoc()) {
1639 MVT RegVT = VA.getLocVT();
1640 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1641 switch (VA.getLocReg()) {
1644 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1646 // Special case: passing MMX values in GPR registers.
1647 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1650 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1651 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1652 // Special case: passing MMX values in XMM registers.
1653 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1654 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1655 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
1656 DAG.getUNDEF(MVT::v2i64), Arg,
1657 getMOVLMask(2, DAG, dl));
1662 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1664 if (!IsTailCall || (IsTailCall && isByVal)) {
1665 assert(VA.isMemLoc());
1666 if (StackPtr.getNode() == 0)
1667 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1669 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1670 Chain, Arg, Flags));
1675 if (!MemOpChains.empty())
1676 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1677 &MemOpChains[0], MemOpChains.size());
1679 // Build a sequence of copy-to-reg nodes chained together with token chain
1680 // and flag operands which copy the outgoing args into registers.
1682 // Tail call byval lowering might overwrite argument registers so in case of
1683 // tail call optimization the copies to registers are lowered later.
1685 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1686 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1687 RegsToPass[i].second, InFlag);
1688 InFlag = Chain.getValue(1);
1691 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1693 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1694 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1695 DAG.getNode(X86ISD::GlobalBaseReg,
1696 DebugLoc::getUnknownLoc(),
1699 InFlag = Chain.getValue(1);
1701 // If we are tail calling and generating PIC/GOT style code load the address
1702 // of the callee into ecx. The value in ecx is used as target of the tail
1703 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1704 // calls on PIC/GOT architectures. Normally we would just put the address of
1705 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1706 // restored (since ebx is callee saved) before jumping to the target@PLT.
1707 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1708 // Note: The actual moving to ecx is done further down.
1709 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1710 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1711 !G->getGlobal()->hasProtectedVisibility())
1712 Callee = LowerGlobalAddress(Callee, DAG);
1713 else if (isa<ExternalSymbolSDNode>(Callee))
1714 Callee = LowerExternalSymbol(Callee,DAG);
1717 if (Is64Bit && isVarArg) {
1718 // From AMD64 ABI document:
1719 // For calls that may call functions that use varargs or stdargs
1720 // (prototype-less calls or calls to functions containing ellipsis (...) in
1721 // the declaration) %al is used as hidden argument to specify the number
1722 // of SSE registers used. The contents of %al do not need to match exactly
1723 // the number of registers, but must be an ubound on the number of SSE
1724 // registers used and is in the range 0 - 8 inclusive.
1726 // FIXME: Verify this on Win64
1727 // Count the number of XMM registers allocated.
1728 static const unsigned XMMArgRegs[] = {
1729 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1730 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1732 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1733 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1734 && "SSE registers cannot be used when SSE is disabled");
1736 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1737 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1738 InFlag = Chain.getValue(1);
1742 // For tail calls lower the arguments to the 'real' stack slot.
1744 SmallVector<SDValue, 8> MemOpChains2;
1747 // Do not flag preceeding copytoreg stuff together with the following stuff.
1749 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1750 CCValAssign &VA = ArgLocs[i];
1751 if (!VA.isRegLoc()) {
1752 assert(VA.isMemLoc());
1753 SDValue Arg = TheCall->getArg(i);
1754 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1755 // Create frame index.
1756 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1757 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1758 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1759 FIN = DAG.getFrameIndex(FI, getPointerTy());
1761 if (Flags.isByVal()) {
1762 // Copy relative to framepointer.
1763 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1764 if (StackPtr.getNode() == 0)
1765 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1767 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1769 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1772 // Store relative to framepointer.
1773 MemOpChains2.push_back(
1774 DAG.getStore(Chain, dl, Arg, FIN,
1775 PseudoSourceValue::getFixedStack(FI), 0));
1780 if (!MemOpChains2.empty())
1781 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1782 &MemOpChains2[0], MemOpChains2.size());
1784 // Copy arguments to their registers.
1785 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1786 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1787 RegsToPass[i].second, InFlag);
1788 InFlag = Chain.getValue(1);
1792 // Store the return address to the appropriate stack slot.
1793 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1797 // If the callee is a GlobalAddress node (quite common, every direct call is)
1798 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1799 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1800 // We should use extra load for direct calls to dllimported functions in
1802 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1803 getTargetMachine(), true))
1804 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1806 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1807 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1808 } else if (IsTailCall) {
1809 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1811 Chain = DAG.getCopyToReg(Chain, dl,
1812 DAG.getRegister(Opc, getPointerTy()),
1814 Callee = DAG.getRegister(Opc, getPointerTy());
1815 // Add register as live out.
1816 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1819 // Returns a chain & a flag for retval copy to use.
1820 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1821 SmallVector<SDValue, 8> Ops;
1824 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1825 DAG.getIntPtrConstant(0, true), InFlag);
1826 InFlag = Chain.getValue(1);
1828 // Returns a chain & a flag for retval copy to use.
1829 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1833 Ops.push_back(Chain);
1834 Ops.push_back(Callee);
1837 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1839 // Add argument registers to the end of the list so that they are known live
1841 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1842 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1843 RegsToPass[i].second.getValueType()));
1845 // Add an implicit use GOT pointer in EBX.
1846 if (!IsTailCall && !Is64Bit &&
1847 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1848 Subtarget->isPICStyleGOT())
1849 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1851 // Add an implicit use of AL for x86 vararg functions.
1852 if (Is64Bit && isVarArg)
1853 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1855 if (InFlag.getNode())
1856 Ops.push_back(InFlag);
1859 assert(InFlag.getNode() &&
1860 "Flag must be set. Depend on flag being set in LowerRET");
1861 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1862 TheCall->getVTList(), &Ops[0], Ops.size());
1864 return SDValue(Chain.getNode(), Op.getResNo());
1867 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1868 InFlag = Chain.getValue(1);
1870 // Create the CALLSEQ_END node.
1871 unsigned NumBytesForCalleeToPush;
1872 if (IsCalleePop(isVarArg, CC))
1873 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1874 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1875 // If this is is a call to a struct-return function, the callee
1876 // pops the hidden struct pointer, so we have to push it back.
1877 // This is common for Darwin/X86, Linux & Mingw32 targets.
1878 NumBytesForCalleeToPush = 4;
1880 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1882 // Returns a flag for retval copy to use.
1883 Chain = DAG.getCALLSEQ_END(Chain,
1884 DAG.getIntPtrConstant(NumBytes, true),
1885 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1888 InFlag = Chain.getValue(1);
1890 // Handle result values, copying them out of physregs into vregs that we
1892 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1897 //===----------------------------------------------------------------------===//
1898 // Fast Calling Convention (tail call) implementation
1899 //===----------------------------------------------------------------------===//
1901 // Like std call, callee cleans arguments, convention except that ECX is
1902 // reserved for storing the tail called function address. Only 2 registers are
1903 // free for argument passing (inreg). Tail call optimization is performed
1905 // * tailcallopt is enabled
1906 // * caller/callee are fastcc
1907 // On X86_64 architecture with GOT-style position independent code only local
1908 // (within module) calls are supported at the moment.
1909 // To keep the stack aligned according to platform abi the function
1910 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1911 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1912 // If a tail called function callee has more arguments than the caller the
1913 // caller needs to make sure that there is room to move the RETADDR to. This is
1914 // achieved by reserving an area the size of the argument delta right after the
1915 // original REtADDR, but before the saved framepointer or the spilled registers
1916 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1928 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1929 /// for a 16 byte align requirement.
1930 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1931 SelectionDAG& DAG) {
1932 MachineFunction &MF = DAG.getMachineFunction();
1933 const TargetMachine &TM = MF.getTarget();
1934 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1935 unsigned StackAlignment = TFI.getStackAlignment();
1936 uint64_t AlignMask = StackAlignment - 1;
1937 int64_t Offset = StackSize;
1938 uint64_t SlotSize = TD->getPointerSize();
1939 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1940 // Number smaller than 12 so just add the difference.
1941 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1943 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1944 Offset = ((~AlignMask) & Offset) + StackAlignment +
1945 (StackAlignment-SlotSize);
1950 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1951 /// following the call is a return. A function is eligible if caller/callee
1952 /// calling conventions match, currently only fastcc supports tail calls, and
1953 /// the function CALL is immediatly followed by a RET.
1954 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1956 SelectionDAG& DAG) const {
1957 if (!PerformTailCallOpt)
1960 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1961 MachineFunction &MF = DAG.getMachineFunction();
1962 unsigned CallerCC = MF.getFunction()->getCallingConv();
1963 unsigned CalleeCC= TheCall->getCallingConv();
1964 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1965 SDValue Callee = TheCall->getCallee();
1966 // On x86/32Bit PIC/GOT tail calls are supported.
1967 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1968 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1971 // Can only do local tail calls (in same module, hidden or protected) on
1972 // x86_64 PIC/GOT at the moment.
1973 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1974 return G->getGlobal()->hasHiddenVisibility()
1975 || G->getGlobal()->hasProtectedVisibility();
1983 X86TargetLowering::createFastISel(MachineFunction &mf,
1984 MachineModuleInfo *mmo,
1986 DenseMap<const Value *, unsigned> &vm,
1987 DenseMap<const BasicBlock *,
1988 MachineBasicBlock *> &bm,
1989 DenseMap<const AllocaInst *, int> &am
1991 , SmallSet<Instruction*, 8> &cil
1994 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2002 //===----------------------------------------------------------------------===//
2003 // Other Lowering Hooks
2004 //===----------------------------------------------------------------------===//
2007 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2008 MachineFunction &MF = DAG.getMachineFunction();
2009 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2010 int ReturnAddrIndex = FuncInfo->getRAIndex();
2012 if (ReturnAddrIndex == 0) {
2013 // Set up a frame object for the return address.
2014 uint64_t SlotSize = TD->getPointerSize();
2015 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2016 FuncInfo->setRAIndex(ReturnAddrIndex);
2019 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2023 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2024 /// specific condition code, returning the condition code and the LHS/RHS of the
2025 /// comparison to make.
2026 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2027 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2029 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2030 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2031 // X > -1 -> X == 0, jump !sign.
2032 RHS = DAG.getConstant(0, RHS.getValueType());
2033 return X86::COND_NS;
2034 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2035 // X < 0 -> X == 0, jump on sign.
2037 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2039 RHS = DAG.getConstant(0, RHS.getValueType());
2040 return X86::COND_LE;
2044 switch (SetCCOpcode) {
2045 default: assert(0 && "Invalid integer condition!");
2046 case ISD::SETEQ: return X86::COND_E;
2047 case ISD::SETGT: return X86::COND_G;
2048 case ISD::SETGE: return X86::COND_GE;
2049 case ISD::SETLT: return X86::COND_L;
2050 case ISD::SETLE: return X86::COND_LE;
2051 case ISD::SETNE: return X86::COND_NE;
2052 case ISD::SETULT: return X86::COND_B;
2053 case ISD::SETUGT: return X86::COND_A;
2054 case ISD::SETULE: return X86::COND_BE;
2055 case ISD::SETUGE: return X86::COND_AE;
2059 // First determine if it is required or is profitable to flip the operands.
2061 // If LHS is a foldable load, but RHS is not, flip the condition.
2062 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2063 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2064 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2065 std::swap(LHS, RHS);
2068 switch (SetCCOpcode) {
2074 std::swap(LHS, RHS);
2078 // On a floating point condition, the flags are set as follows:
2080 // 0 | 0 | 0 | X > Y
2081 // 0 | 0 | 1 | X < Y
2082 // 1 | 0 | 0 | X == Y
2083 // 1 | 1 | 1 | unordered
2084 switch (SetCCOpcode) {
2085 default: assert(0 && "Condcode should be pre-legalized away");
2087 case ISD::SETEQ: return X86::COND_E;
2088 case ISD::SETOLT: // flipped
2090 case ISD::SETGT: return X86::COND_A;
2091 case ISD::SETOLE: // flipped
2093 case ISD::SETGE: return X86::COND_AE;
2094 case ISD::SETUGT: // flipped
2096 case ISD::SETLT: return X86::COND_B;
2097 case ISD::SETUGE: // flipped
2099 case ISD::SETLE: return X86::COND_BE;
2101 case ISD::SETNE: return X86::COND_NE;
2102 case ISD::SETUO: return X86::COND_P;
2103 case ISD::SETO: return X86::COND_NP;
2107 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2108 /// code. Current x86 isa includes the following FP cmov instructions:
2109 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2110 static bool hasFPCMov(unsigned X86CC) {
2126 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2127 /// true if Op is undef or if its value falls within the specified range (L, H].
2128 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2129 if (Op.getOpcode() == ISD::UNDEF)
2132 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2133 return (Val >= Low && Val < Hi);
2136 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2137 /// true if Op is undef or if its value equal to the specified value.
2138 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2139 if (Op.getOpcode() == ISD::UNDEF)
2141 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2144 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2145 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2146 bool X86::isPSHUFDMask(SDNode *N) {
2147 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2149 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2152 // Check if the value doesn't reference the second vector.
2153 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2154 SDValue Arg = N->getOperand(i);
2155 if (Arg.getOpcode() == ISD::UNDEF) continue;
2156 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2157 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2164 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2165 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2166 bool X86::isPSHUFHWMask(SDNode *N) {
2167 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2169 if (N->getNumOperands() != 8)
2172 // Lower quadword copied in order.
2173 for (unsigned i = 0; i != 4; ++i) {
2174 SDValue Arg = N->getOperand(i);
2175 if (Arg.getOpcode() == ISD::UNDEF) continue;
2176 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2177 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2181 // Upper quadword shuffled.
2182 for (unsigned i = 4; i != 8; ++i) {
2183 SDValue Arg = N->getOperand(i);
2184 if (Arg.getOpcode() == ISD::UNDEF) continue;
2185 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2186 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2187 if (Val < 4 || Val > 7)
2194 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2195 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2196 bool X86::isPSHUFLWMask(SDNode *N) {
2197 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2199 if (N->getNumOperands() != 8)
2202 // Upper quadword copied in order.
2203 for (unsigned i = 4; i != 8; ++i)
2204 if (!isUndefOrEqual(N->getOperand(i), i))
2207 // Lower quadword shuffled.
2208 for (unsigned i = 0; i != 4; ++i)
2209 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2215 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2216 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2217 template<class SDOperand>
2218 static bool isSHUFPMask(SDOperand *Elems, unsigned NumElems) {
2219 if (NumElems != 2 && NumElems != 4) return false;
2221 unsigned Half = NumElems / 2;
2222 for (unsigned i = 0; i < Half; ++i)
2223 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2225 for (unsigned i = Half; i < NumElems; ++i)
2226 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2232 bool X86::isSHUFPMask(SDNode *N) {
2233 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2234 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2237 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2238 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2239 /// half elements to come from vector 1 (which would equal the dest.) and
2240 /// the upper half to come from vector 2.
2241 template<class SDOperand>
2242 static bool isCommutedSHUFP(SDOperand *Ops, unsigned NumOps) {
2243 if (NumOps != 2 && NumOps != 4) return false;
2245 unsigned Half = NumOps / 2;
2246 for (unsigned i = 0; i < Half; ++i)
2247 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2249 for (unsigned i = Half; i < NumOps; ++i)
2250 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2255 static bool isCommutedSHUFP(SDNode *N) {
2256 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2257 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2260 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2261 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2262 bool X86::isMOVHLPSMask(SDNode *N) {
2263 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2265 if (N->getNumOperands() != 4)
2268 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2269 return isUndefOrEqual(N->getOperand(0), 6) &&
2270 isUndefOrEqual(N->getOperand(1), 7) &&
2271 isUndefOrEqual(N->getOperand(2), 2) &&
2272 isUndefOrEqual(N->getOperand(3), 3);
2275 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2276 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2278 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2279 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2281 if (N->getNumOperands() != 4)
2284 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2285 return isUndefOrEqual(N->getOperand(0), 2) &&
2286 isUndefOrEqual(N->getOperand(1), 3) &&
2287 isUndefOrEqual(N->getOperand(2), 2) &&
2288 isUndefOrEqual(N->getOperand(3), 3);
2291 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2292 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2293 bool X86::isMOVLPMask(SDNode *N) {
2294 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2296 unsigned NumElems = N->getNumOperands();
2297 if (NumElems != 2 && NumElems != 4)
2300 for (unsigned i = 0; i < NumElems/2; ++i)
2301 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2304 for (unsigned i = NumElems/2; i < NumElems; ++i)
2305 if (!isUndefOrEqual(N->getOperand(i), i))
2311 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2312 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2314 bool X86::isMOVHPMask(SDNode *N) {
2315 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2317 unsigned NumElems = N->getNumOperands();
2318 if (NumElems != 2 && NumElems != 4)
2321 for (unsigned i = 0; i < NumElems/2; ++i)
2322 if (!isUndefOrEqual(N->getOperand(i), i))
2325 for (unsigned i = 0; i < NumElems/2; ++i) {
2326 SDValue Arg = N->getOperand(i + NumElems/2);
2327 if (!isUndefOrEqual(Arg, i + NumElems))
2334 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2335 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2336 template<class SDOperand>
2337 bool static isUNPCKLMask(SDOperand *Elts, unsigned NumElts,
2338 bool V2IsSplat = false) {
2339 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2342 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2343 SDValue BitI = Elts[i];
2344 SDValue BitI1 = Elts[i+1];
2345 if (!isUndefOrEqual(BitI, j))
2348 if (!isUndefOrEqual(BitI1, NumElts))
2351 if (!isUndefOrEqual(BitI1, j + NumElts))
2359 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2360 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2361 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2364 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2365 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2366 template<class SDOperand>
2367 bool static isUNPCKHMask(SDOperand *Elts, unsigned NumElts,
2368 bool V2IsSplat = false) {
2369 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2372 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2373 SDValue BitI = Elts[i];
2374 SDValue BitI1 = Elts[i+1];
2375 if (!isUndefOrEqual(BitI, j + NumElts/2))
2378 if (isUndefOrEqual(BitI1, NumElts))
2381 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2389 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2390 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2391 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2394 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2395 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2397 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2398 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2400 unsigned NumElems = N->getNumOperands();
2401 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2404 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2405 SDValue BitI = N->getOperand(i);
2406 SDValue BitI1 = N->getOperand(i+1);
2408 if (!isUndefOrEqual(BitI, j))
2410 if (!isUndefOrEqual(BitI1, j))
2417 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2418 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2420 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2421 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2423 unsigned NumElems = N->getNumOperands();
2424 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2427 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2428 SDValue BitI = N->getOperand(i);
2429 SDValue BitI1 = N->getOperand(i + 1);
2431 if (!isUndefOrEqual(BitI, j))
2433 if (!isUndefOrEqual(BitI1, j))
2440 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2441 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2442 /// MOVSD, and MOVD, i.e. setting the lowest element.
2443 template<class SDOperand>
2444 static bool isMOVLMask(SDOperand *Elts, unsigned NumElts) {
2445 if (NumElts != 2 && NumElts != 4)
2448 if (!isUndefOrEqual(Elts[0], NumElts))
2451 for (unsigned i = 1; i < NumElts; ++i) {
2452 if (!isUndefOrEqual(Elts[i], i))
2459 bool X86::isMOVLMask(SDNode *N) {
2460 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2461 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2464 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2465 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2466 /// element of vector 2 and the other elements to come from vector 1 in order.
2467 template<class SDOperand>
2468 static bool isCommutedMOVL(SDOperand *Ops, unsigned NumOps,
2469 bool V2IsSplat = false,
2470 bool V2IsUndef = false) {
2471 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2474 if (!isUndefOrEqual(Ops[0], 0))
2477 for (unsigned i = 1; i < NumOps; ++i) {
2478 SDValue Arg = Ops[i];
2479 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2480 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2481 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2488 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2489 bool V2IsUndef = false) {
2490 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2491 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2492 V2IsSplat, V2IsUndef);
2495 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2496 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2497 bool X86::isMOVSHDUPMask(SDNode *N) {
2498 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2500 if (N->getNumOperands() != 4)
2503 // Expect 1, 1, 3, 3
2504 for (unsigned i = 0; i < 2; ++i) {
2505 SDValue Arg = N->getOperand(i);
2506 if (Arg.getOpcode() == ISD::UNDEF) continue;
2507 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2508 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2509 if (Val != 1) return false;
2513 for (unsigned i = 2; i < 4; ++i) {
2514 SDValue Arg = N->getOperand(i);
2515 if (Arg.getOpcode() == ISD::UNDEF) continue;
2516 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2517 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2518 if (Val != 3) return false;
2522 // Don't use movshdup if it can be done with a shufps.
2526 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2527 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2528 bool X86::isMOVSLDUPMask(SDNode *N) {
2529 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2531 if (N->getNumOperands() != 4)
2534 // Expect 0, 0, 2, 2
2535 for (unsigned i = 0; i < 2; ++i) {
2536 SDValue Arg = N->getOperand(i);
2537 if (Arg.getOpcode() == ISD::UNDEF) continue;
2538 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2539 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2540 if (Val != 0) return false;
2544 for (unsigned i = 2; i < 4; ++i) {
2545 SDValue Arg = N->getOperand(i);
2546 if (Arg.getOpcode() == ISD::UNDEF) continue;
2547 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2548 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2549 if (Val != 2) return false;
2553 // Don't use movshdup if it can be done with a shufps.
2557 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2558 /// specifies a identity operation on the LHS or RHS.
2559 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2560 unsigned NumElems = N->getNumOperands();
2561 for (unsigned i = 0; i < NumElems; ++i)
2562 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2567 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2568 /// a splat of a single element.
2569 static bool isSplatMask(SDNode *N) {
2570 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2572 // This is a splat operation if each element of the permute is the same, and
2573 // if the value doesn't reference the second vector.
2574 unsigned NumElems = N->getNumOperands();
2575 SDValue ElementBase;
2577 for (; i != NumElems; ++i) {
2578 SDValue Elt = N->getOperand(i);
2579 if (isa<ConstantSDNode>(Elt)) {
2585 if (!ElementBase.getNode())
2588 for (; i != NumElems; ++i) {
2589 SDValue Arg = N->getOperand(i);
2590 if (Arg.getOpcode() == ISD::UNDEF) continue;
2591 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2592 if (Arg != ElementBase) return false;
2595 // Make sure it is a splat of the first vector operand.
2596 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2599 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2600 /// we want to splat.
2601 static SDValue getSplatMaskEltNo(SDNode *N) {
2602 assert(isSplatMask(N) && "Not a splat mask");
2603 unsigned NumElems = N->getNumOperands();
2604 SDValue ElementBase;
2606 for (; i != NumElems; ++i) {
2607 SDValue Elt = N->getOperand(i);
2608 if (isa<ConstantSDNode>(Elt))
2611 assert(0 && " No splat value found!");
2616 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2617 /// a splat of a single element and it's a 2 or 4 element mask.
2618 bool X86::isSplatMask(SDNode *N) {
2619 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2621 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2622 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2624 return ::isSplatMask(N);
2627 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2628 /// specifies a splat of zero element.
2629 bool X86::isSplatLoMask(SDNode *N) {
2630 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2632 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2633 if (!isUndefOrEqual(N->getOperand(i), 0))
2638 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2639 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2640 bool X86::isMOVDDUPMask(SDNode *N) {
2641 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2643 unsigned e = N->getNumOperands() / 2;
2644 for (unsigned i = 0; i < e; ++i)
2645 if (!isUndefOrEqual(N->getOperand(i), i))
2647 for (unsigned i = 0; i < e; ++i)
2648 if (!isUndefOrEqual(N->getOperand(e+i), i))
2653 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2654 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2656 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2657 unsigned NumOperands = N->getNumOperands();
2658 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2660 for (unsigned i = 0; i < NumOperands; ++i) {
2662 SDValue Arg = N->getOperand(NumOperands-i-1);
2663 if (Arg.getOpcode() != ISD::UNDEF)
2664 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2665 if (Val >= NumOperands) Val -= NumOperands;
2667 if (i != NumOperands - 1)
2674 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2675 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2677 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2679 // 8 nodes, but we only care about the last 4.
2680 for (unsigned i = 7; i >= 4; --i) {
2682 SDValue Arg = N->getOperand(i);
2683 if (Arg.getOpcode() != ISD::UNDEF) {
2684 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2694 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2695 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2697 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2699 // 8 nodes, but we only care about the first 4.
2700 for (int i = 3; i >= 0; --i) {
2702 SDValue Arg = N->getOperand(i);
2703 if (Arg.getOpcode() != ISD::UNDEF)
2704 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2713 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2714 /// specifies a 8 element shuffle that can be broken into a pair of
2715 /// PSHUFHW and PSHUFLW.
2716 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2717 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2719 if (N->getNumOperands() != 8)
2722 // Lower quadword shuffled.
2723 for (unsigned i = 0; i != 4; ++i) {
2724 SDValue Arg = N->getOperand(i);
2725 if (Arg.getOpcode() == ISD::UNDEF) continue;
2726 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2727 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2732 // Upper quadword shuffled.
2733 for (unsigned i = 4; i != 8; ++i) {
2734 SDValue Arg = N->getOperand(i);
2735 if (Arg.getOpcode() == ISD::UNDEF) continue;
2736 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2737 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2738 if (Val < 4 || Val > 7)
2745 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2746 /// values in ther permute mask.
2747 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2748 SDValue &V2, SDValue &Mask,
2749 SelectionDAG &DAG) {
2750 MVT VT = Op.getValueType();
2751 MVT MaskVT = Mask.getValueType();
2752 MVT EltVT = MaskVT.getVectorElementType();
2753 unsigned NumElems = Mask.getNumOperands();
2754 SmallVector<SDValue, 8> MaskVec;
2755 DebugLoc dl = Op.getDebugLoc();
2757 for (unsigned i = 0; i != NumElems; ++i) {
2758 SDValue Arg = Mask.getOperand(i);
2759 if (Arg.getOpcode() == ISD::UNDEF) {
2760 MaskVec.push_back(DAG.getUNDEF(EltVT));
2763 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2764 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2766 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2768 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2772 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2773 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
2776 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2777 /// the two vector operands have swapped position.
2779 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG, DebugLoc dl) {
2780 MVT MaskVT = Mask.getValueType();
2781 MVT EltVT = MaskVT.getVectorElementType();
2782 unsigned NumElems = Mask.getNumOperands();
2783 SmallVector<SDValue, 8> MaskVec;
2784 for (unsigned i = 0; i != NumElems; ++i) {
2785 SDValue Arg = Mask.getOperand(i);
2786 if (Arg.getOpcode() == ISD::UNDEF) {
2787 MaskVec.push_back(DAG.getUNDEF(EltVT));
2790 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2791 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2793 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2795 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2797 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2801 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2802 /// match movhlps. The lower half elements should come from upper half of
2803 /// V1 (and in order), and the upper half elements should come from the upper
2804 /// half of V2 (and in order).
2805 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2806 unsigned NumElems = Mask->getNumOperands();
2809 for (unsigned i = 0, e = 2; i != e; ++i)
2810 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2812 for (unsigned i = 2; i != 4; ++i)
2813 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2818 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2819 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2821 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2822 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2824 N = N->getOperand(0).getNode();
2825 if (!ISD::isNON_EXTLoad(N))
2828 *LD = cast<LoadSDNode>(N);
2832 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2833 /// match movlp{s|d}. The lower half elements should come from lower half of
2834 /// V1 (and in order), and the upper half elements should come from the upper
2835 /// half of V2 (and in order). And since V1 will become the source of the
2836 /// MOVLP, it must be either a vector load or a scalar load to vector.
2837 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2838 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2840 // Is V2 is a vector load, don't do this transformation. We will try to use
2841 // load folding shufps op.
2842 if (ISD::isNON_EXTLoad(V2))
2845 unsigned NumElems = Mask->getNumOperands();
2846 if (NumElems != 2 && NumElems != 4)
2848 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2849 if (!isUndefOrEqual(Mask->getOperand(i), i))
2851 for (unsigned i = NumElems/2; i != NumElems; ++i)
2852 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2857 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2859 static bool isSplatVector(SDNode *N) {
2860 if (N->getOpcode() != ISD::BUILD_VECTOR)
2863 SDValue SplatValue = N->getOperand(0);
2864 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2865 if (N->getOperand(i) != SplatValue)
2870 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2872 static bool isUndefShuffle(SDNode *N) {
2873 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2876 SDValue V1 = N->getOperand(0);
2877 SDValue V2 = N->getOperand(1);
2878 SDValue Mask = N->getOperand(2);
2879 unsigned NumElems = Mask.getNumOperands();
2880 for (unsigned i = 0; i != NumElems; ++i) {
2881 SDValue Arg = Mask.getOperand(i);
2882 if (Arg.getOpcode() != ISD::UNDEF) {
2883 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2884 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2886 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2893 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2895 static inline bool isZeroNode(SDValue Elt) {
2896 return ((isa<ConstantSDNode>(Elt) &&
2897 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2898 (isa<ConstantFPSDNode>(Elt) &&
2899 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2902 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2903 /// to an zero vector.
2904 static bool isZeroShuffle(SDNode *N) {
2905 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2908 SDValue V1 = N->getOperand(0);
2909 SDValue V2 = N->getOperand(1);
2910 SDValue Mask = N->getOperand(2);
2911 unsigned NumElems = Mask.getNumOperands();
2912 for (unsigned i = 0; i != NumElems; ++i) {
2913 SDValue Arg = Mask.getOperand(i);
2914 if (Arg.getOpcode() == ISD::UNDEF)
2917 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2918 if (Idx < NumElems) {
2919 unsigned Opc = V1.getNode()->getOpcode();
2920 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2922 if (Opc != ISD::BUILD_VECTOR ||
2923 !isZeroNode(V1.getNode()->getOperand(Idx)))
2925 } else if (Idx >= NumElems) {
2926 unsigned Opc = V2.getNode()->getOpcode();
2927 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2929 if (Opc != ISD::BUILD_VECTOR ||
2930 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2937 /// getZeroVector - Returns a vector of specified type with all zero elements.
2939 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2941 assert(VT.isVector() && "Expected a vector type");
2943 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2944 // type. This ensures they get CSE'd.
2946 if (VT.getSizeInBits() == 64) { // MMX
2947 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2948 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2949 } else if (HasSSE2) { // SSE2
2950 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2951 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2953 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2954 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2956 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2959 /// getOnesVector - Returns a vector of specified type with all bits set.
2961 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2962 assert(VT.isVector() && "Expected a vector type");
2964 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2965 // type. This ensures they get CSE'd.
2966 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2968 if (VT.getSizeInBits() == 64) // MMX
2969 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2971 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2972 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2976 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2977 /// that point to V2 points to its first element.
2978 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2979 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2981 bool Changed = false;
2982 SmallVector<SDValue, 8> MaskVec;
2983 unsigned NumElems = Mask.getNumOperands();
2984 for (unsigned i = 0; i != NumElems; ++i) {
2985 SDValue Arg = Mask.getOperand(i);
2986 if (Arg.getOpcode() != ISD::UNDEF) {
2987 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2988 if (Val > NumElems) {
2989 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2993 MaskVec.push_back(Arg);
2997 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getDebugLoc(),
2998 Mask.getValueType(),
2999 &MaskVec[0], MaskVec.size());
3003 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3004 /// operation of specified width.
3005 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl) {
3006 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3007 MVT BaseVT = MaskVT.getVectorElementType();
3009 SmallVector<SDValue, 8> MaskVec;
3010 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
3011 for (unsigned i = 1; i != NumElems; ++i)
3012 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3013 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3014 &MaskVec[0], MaskVec.size());
3017 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
3018 /// of specified width.
3019 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG,
3021 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3022 MVT BaseVT = MaskVT.getVectorElementType();
3023 SmallVector<SDValue, 8> MaskVec;
3024 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3025 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3026 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3028 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3029 &MaskVec[0], MaskVec.size());
3032 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3033 /// of specified width.
3034 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG,
3036 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3037 MVT BaseVT = MaskVT.getVectorElementType();
3038 unsigned Half = NumElems/2;
3039 SmallVector<SDValue, 8> MaskVec;
3040 for (unsigned i = 0; i != Half; ++i) {
3041 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3042 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3044 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3045 &MaskVec[0], MaskVec.size());
3048 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
3049 /// element #0 of a vector with the specified index, leaving the rest of the
3050 /// elements in place.
3051 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
3052 SelectionDAG &DAG, DebugLoc dl) {
3053 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3054 MVT BaseVT = MaskVT.getVectorElementType();
3055 SmallVector<SDValue, 8> MaskVec;
3056 // Element #0 of the result gets the elt we are replacing.
3057 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
3058 for (unsigned i = 1; i != NumElems; ++i)
3059 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
3060 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3061 &MaskVec[0], MaskVec.size());
3064 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3065 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
3066 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
3067 MVT VT = Op.getValueType();
3070 SDValue V1 = Op.getOperand(0);
3071 SDValue Mask = Op.getOperand(2);
3072 unsigned MaskNumElems = Mask.getNumOperands();
3073 unsigned NumElems = MaskNumElems;
3074 DebugLoc dl = Op.getDebugLoc();
3075 // Special handling of v4f32 -> v4i32.
3076 if (VT != MVT::v4f32) {
3077 // Find which element we want to splat.
3078 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3079 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3080 // unpack elements to the correct location
3081 while (NumElems > 4) {
3082 if (EltNo < NumElems/2) {
3083 Mask = getUnpacklMask(MaskNumElems, DAG, dl);
3085 Mask = getUnpackhMask(MaskNumElems, DAG, dl);
3086 EltNo -= NumElems/2;
3088 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1, Mask);
3091 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3092 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3095 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3096 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3097 DAG.getUNDEF(PVT), Mask);
3098 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3101 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3102 /// load that's promoted to vector, or a load bitcasted.
3103 static bool isVectorLoad(SDValue Op) {
3104 assert(Op.getValueType().isVector() && "Expected a vector type");
3105 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3106 Op.getOpcode() == ISD::BIT_CONVERT) {
3107 return isa<LoadSDNode>(Op.getOperand(0));
3109 return isa<LoadSDNode>(Op);
3113 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3115 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3116 SelectionDAG &DAG, bool HasSSE3) {
3117 // If we have sse3 and shuffle has more than one use or input is a load, then
3118 // use movddup. Otherwise, use movlhps.
3119 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3120 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3121 MVT VT = Op.getValueType();
3124 DebugLoc dl = Op.getDebugLoc();
3125 unsigned NumElems = PVT.getVectorNumElements();
3126 if (NumElems == 2) {
3127 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3128 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3130 assert(NumElems == 4);
3131 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3132 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3133 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3134 Cst0, Cst1, Cst0, Cst1);
3137 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3138 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3139 DAG.getUNDEF(PVT), Mask);
3140 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3143 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3144 /// vector of zero or undef vector. This produces a shuffle where the low
3145 /// element of V2 is swizzled into the zero/undef vector, landing at element
3146 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3147 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3148 bool isZero, bool HasSSE2,
3149 SelectionDAG &DAG) {
3150 DebugLoc dl = V2.getDebugLoc();
3151 MVT VT = V2.getValueType();
3153 ? getZeroVector(VT, HasSSE2, DAG, dl) : DAG.getUNDEF(VT);
3154 unsigned NumElems = V2.getValueType().getVectorNumElements();
3155 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3156 MVT EVT = MaskVT.getVectorElementType();
3157 SmallVector<SDValue, 16> MaskVec;
3158 for (unsigned i = 0; i != NumElems; ++i)
3159 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3160 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3162 MaskVec.push_back(DAG.getConstant(i, EVT));
3163 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3164 &MaskVec[0], MaskVec.size());
3165 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
3168 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3169 /// a shuffle that is zero.
3171 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3172 unsigned NumElems, bool Low,
3173 SelectionDAG &DAG) {
3174 unsigned NumZeros = 0;
3175 for (unsigned i = 0; i < NumElems; ++i) {
3176 unsigned Index = Low ? i : NumElems-i-1;
3177 SDValue Idx = Mask.getOperand(Index);
3178 if (Idx.getOpcode() == ISD::UNDEF) {
3182 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3183 if (Elt.getNode() && isZeroNode(Elt))
3191 /// isVectorShift - Returns true if the shuffle can be implemented as a
3192 /// logical left or right shift of a vector.
3193 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3194 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3195 unsigned NumElems = Mask.getNumOperands();
3198 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3201 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3206 bool SeenV1 = false;
3207 bool SeenV2 = false;
3208 for (unsigned i = NumZeros; i < NumElems; ++i) {
3209 unsigned Val = isLeft ? (i - NumZeros) : i;
3210 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3211 if (Idx.getOpcode() == ISD::UNDEF)
3213 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3214 if (Index < NumElems)
3223 if (SeenV1 && SeenV2)
3226 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3232 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3234 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3235 unsigned NumNonZero, unsigned NumZero,
3236 SelectionDAG &DAG, TargetLowering &TLI) {
3240 DebugLoc dl = Op.getDebugLoc();
3243 for (unsigned i = 0; i < 16; ++i) {
3244 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3245 if (ThisIsNonZero && First) {
3247 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3249 V = DAG.getUNDEF(MVT::v8i16);
3254 SDValue ThisElt(0, 0), LastElt(0, 0);
3255 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3256 if (LastIsNonZero) {
3257 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3258 MVT::i16, Op.getOperand(i-1));
3260 if (ThisIsNonZero) {
3261 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3262 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3263 ThisElt, DAG.getConstant(8, MVT::i8));
3265 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3269 if (ThisElt.getNode())
3270 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3271 DAG.getIntPtrConstant(i/2));
3275 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3278 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3280 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3281 unsigned NumNonZero, unsigned NumZero,
3282 SelectionDAG &DAG, TargetLowering &TLI) {
3286 DebugLoc dl = Op.getDebugLoc();
3289 for (unsigned i = 0; i < 8; ++i) {
3290 bool isNonZero = (NonZeros & (1 << i)) != 0;
3294 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3296 V = DAG.getUNDEF(MVT::v8i16);
3299 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3300 MVT::v8i16, V, Op.getOperand(i),
3301 DAG.getIntPtrConstant(i));
3308 /// getVShift - Return a vector logical shift node.
3310 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3311 unsigned NumBits, SelectionDAG &DAG,
3312 const TargetLowering &TLI, DebugLoc dl) {
3313 bool isMMX = VT.getSizeInBits() == 64;
3314 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3315 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3316 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3317 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3318 DAG.getNode(Opc, dl, ShVT, SrcOp,
3319 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3323 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3324 DebugLoc dl = Op.getDebugLoc();
3325 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3326 if (ISD::isBuildVectorAllZeros(Op.getNode())
3327 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3328 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3329 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3330 // eliminated on x86-32 hosts.
3331 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3334 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3335 return getOnesVector(Op.getValueType(), DAG, dl);
3336 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3339 MVT VT = Op.getValueType();
3340 MVT EVT = VT.getVectorElementType();
3341 unsigned EVTBits = EVT.getSizeInBits();
3343 unsigned NumElems = Op.getNumOperands();
3344 unsigned NumZero = 0;
3345 unsigned NumNonZero = 0;
3346 unsigned NonZeros = 0;
3347 bool IsAllConstants = true;
3348 SmallSet<SDValue, 8> Values;
3349 for (unsigned i = 0; i < NumElems; ++i) {
3350 SDValue Elt = Op.getOperand(i);
3351 if (Elt.getOpcode() == ISD::UNDEF)
3354 if (Elt.getOpcode() != ISD::Constant &&
3355 Elt.getOpcode() != ISD::ConstantFP)
3356 IsAllConstants = false;
3357 if (isZeroNode(Elt))
3360 NonZeros |= (1 << i);
3365 if (NumNonZero == 0) {
3366 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3367 return DAG.getUNDEF(VT);
3370 // Special case for single non-zero, non-undef, element.
3371 if (NumNonZero == 1 && NumElems <= 4) {
3372 unsigned Idx = CountTrailingZeros_32(NonZeros);
3373 SDValue Item = Op.getOperand(Idx);
3375 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3376 // the value are obviously zero, truncate the value to i32 and do the
3377 // insertion that way. Only do this if the value is non-constant or if the
3378 // value is a constant being inserted into element 0. It is cheaper to do
3379 // a constant pool load than it is to do a movd + shuffle.
3380 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3381 (!IsAllConstants || Idx == 0)) {
3382 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3383 // Handle MMX and SSE both.
3384 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3385 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3387 // Truncate the value (which may itself be a constant) to i32, and
3388 // convert it to a vector with movd (S2V+shuffle to zero extend).
3389 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3390 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3391 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3392 Subtarget->hasSSE2(), DAG);
3394 // Now we have our 32-bit value zero extended in the low element of
3395 // a vector. If Idx != 0, swizzle it into place.
3398 Item, DAG.getUNDEF(Item.getValueType()),
3399 getSwapEltZeroMask(VecElts, Idx, DAG, dl)
3401 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VecVT, Ops, 3);
3403 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3407 // If we have a constant or non-constant insertion into the low element of
3408 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3409 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3410 // depending on what the source datatype is. Because we can only get here
3411 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3413 // Don't do this for i64 values on x86-32.
3414 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3415 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3416 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3417 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3418 Subtarget->hasSSE2(), DAG);
3421 // Is it a vector logical left shift?
3422 if (NumElems == 2 && Idx == 1 &&
3423 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3424 unsigned NumBits = VT.getSizeInBits();
3425 return getVShift(true, VT,
3426 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3427 VT, Op.getOperand(1)),
3428 NumBits/2, DAG, *this, dl);
3431 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3434 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3435 // is a non-constant being inserted into an element other than the low one,
3436 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3437 // movd/movss) to move this into the low element, then shuffle it into
3439 if (EVTBits == 32) {
3440 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3442 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3443 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3444 Subtarget->hasSSE2(), DAG);
3445 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3446 MVT MaskEVT = MaskVT.getVectorElementType();
3447 SmallVector<SDValue, 8> MaskVec;
3448 for (unsigned i = 0; i < NumElems; i++)
3449 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3450 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3451 &MaskVec[0], MaskVec.size());
3452 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, Item,
3453 DAG.getUNDEF(VT), Mask);
3457 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3458 if (Values.size() == 1)
3461 // A vector full of immediates; various special cases are already
3462 // handled, so this is best done with a single constant-pool load.
3466 // Let legalizer expand 2-wide build_vectors.
3467 if (EVTBits == 64) {
3468 if (NumNonZero == 1) {
3469 // One half is zero or undef.
3470 unsigned Idx = CountTrailingZeros_32(NonZeros);
3471 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3472 Op.getOperand(Idx));
3473 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3474 Subtarget->hasSSE2(), DAG);
3479 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3480 if (EVTBits == 8 && NumElems == 16) {
3481 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3483 if (V.getNode()) return V;
3486 if (EVTBits == 16 && NumElems == 8) {
3487 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3489 if (V.getNode()) return V;
3492 // If element VT is == 32 bits, turn it into a number of shuffles.
3493 SmallVector<SDValue, 8> V;
3495 if (NumElems == 4 && NumZero > 0) {
3496 for (unsigned i = 0; i < 4; ++i) {
3497 bool isZero = !(NonZeros & (1 << i));
3499 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3501 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3504 for (unsigned i = 0; i < 2; ++i) {
3505 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3508 V[i] = V[i*2]; // Must be a zero vector.
3511 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2+1], V[i*2],
3512 getMOVLMask(NumElems, DAG, dl));
3515 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3516 getMOVLMask(NumElems, DAG, dl));
3519 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3520 getUnpacklMask(NumElems, DAG, dl));
3525 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3526 MVT EVT = MaskVT.getVectorElementType();
3527 SmallVector<SDValue, 8> MaskVec;
3528 bool Reverse = (NonZeros & 0x3) == 2;
3529 for (unsigned i = 0; i < 2; ++i)
3531 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3533 MaskVec.push_back(DAG.getConstant(i, EVT));
3534 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3535 for (unsigned i = 0; i < 2; ++i)
3537 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3539 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3540 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3541 &MaskVec[0], MaskVec.size());
3542 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[0], V[1], ShufMask);
3545 if (Values.size() > 2) {
3546 // Expand into a number of unpckl*.
3548 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3549 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3550 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3551 SDValue UnpckMask = getUnpacklMask(NumElems, DAG, dl);
3552 for (unsigned i = 0; i < NumElems; ++i)
3553 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3555 while (NumElems != 0) {
3556 for (unsigned i = 0; i < NumElems; ++i)
3557 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i], V[i + NumElems],
3568 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3569 SDValue PermMask, SelectionDAG &DAG,
3570 TargetLowering &TLI, DebugLoc dl) {
3572 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3573 MVT MaskEVT = MaskVT.getVectorElementType();
3574 MVT PtrVT = TLI.getPointerTy();
3575 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3576 PermMask.getNode()->op_end());
3578 // First record which half of which vector the low elements come from.
3579 SmallVector<unsigned, 4> LowQuad(4);
3580 for (unsigned i = 0; i < 4; ++i) {
3581 SDValue Elt = MaskElts[i];
3582 if (Elt.getOpcode() == ISD::UNDEF)
3584 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3585 int QuadIdx = EltIdx / 4;
3589 int BestLowQuad = -1;
3590 unsigned MaxQuad = 1;
3591 for (unsigned i = 0; i < 4; ++i) {
3592 if (LowQuad[i] > MaxQuad) {
3594 MaxQuad = LowQuad[i];
3598 // Record which half of which vector the high elements come from.
3599 SmallVector<unsigned, 4> HighQuad(4);
3600 for (unsigned i = 4; i < 8; ++i) {
3601 SDValue Elt = MaskElts[i];
3602 if (Elt.getOpcode() == ISD::UNDEF)
3604 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3605 int QuadIdx = EltIdx / 4;
3606 ++HighQuad[QuadIdx];
3609 int BestHighQuad = -1;
3611 for (unsigned i = 0; i < 4; ++i) {
3612 if (HighQuad[i] > MaxQuad) {
3614 MaxQuad = HighQuad[i];
3618 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3619 if (BestLowQuad != -1 || BestHighQuad != -1) {
3620 // First sort the 4 chunks in order using shufpd.
3621 SmallVector<SDValue, 8> MaskVec;
3623 if (BestLowQuad != -1)
3624 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3626 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3628 if (BestHighQuad != -1)
3629 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3631 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3633 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, &MaskVec[0],2);
3634 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
3635 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3636 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), Mask);
3637 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3639 // Now sort high and low parts separately.
3640 BitVector InOrder(8);
3641 if (BestLowQuad != -1) {
3642 // Sort lower half in order using PSHUFLW.
3644 bool AnyOutOrder = false;
3646 for (unsigned i = 0; i != 4; ++i) {
3647 SDValue Elt = MaskElts[i];
3648 if (Elt.getOpcode() == ISD::UNDEF) {
3649 MaskVec.push_back(Elt);
3652 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3656 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3658 // If this element is in the right place after this shuffle, then
3660 if ((int)(EltIdx / 4) == BestLowQuad)
3665 for (unsigned i = 4; i != 8; ++i)
3666 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3667 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3669 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16,
3674 if (BestHighQuad != -1) {
3675 // Sort high half in order using PSHUFHW if possible.
3678 for (unsigned i = 0; i != 4; ++i)
3679 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3681 bool AnyOutOrder = false;
3682 for (unsigned i = 4; i != 8; ++i) {
3683 SDValue Elt = MaskElts[i];
3684 if (Elt.getOpcode() == ISD::UNDEF) {
3685 MaskVec.push_back(Elt);
3688 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3692 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3694 // If this element is in the right place after this shuffle, then
3696 if ((int)(EltIdx / 4) == BestHighQuad)
3702 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl,
3703 MaskVT, &MaskVec[0], 8);
3704 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16,
3709 // The other elements are put in the right place using pextrw and pinsrw.
3710 for (unsigned i = 0; i != 8; ++i) {
3713 SDValue Elt = MaskElts[i];
3714 if (Elt.getOpcode() == ISD::UNDEF)
3716 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3717 SDValue ExtOp = (EltIdx < 8)
3718 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3719 DAG.getConstant(EltIdx, PtrVT))
3720 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3721 DAG.getConstant(EltIdx - 8, PtrVT));
3722 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3723 DAG.getConstant(i, PtrVT));
3729 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3730 // few as possible. First, let's find out how many elements are already in the
3732 unsigned V1InOrder = 0;
3733 unsigned V1FromV1 = 0;
3734 unsigned V2InOrder = 0;
3735 unsigned V2FromV2 = 0;
3736 SmallVector<SDValue, 8> V1Elts;
3737 SmallVector<SDValue, 8> V2Elts;
3738 for (unsigned i = 0; i < 8; ++i) {
3739 SDValue Elt = MaskElts[i];
3740 if (Elt.getOpcode() == ISD::UNDEF) {
3741 V1Elts.push_back(Elt);
3742 V2Elts.push_back(Elt);
3747 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3749 V1Elts.push_back(Elt);
3750 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3752 } else if (EltIdx == i+8) {
3753 V1Elts.push_back(Elt);
3754 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3756 } else if (EltIdx < 8) {
3757 V1Elts.push_back(Elt);
3758 V2Elts.push_back(DAG.getConstant(EltIdx+8, MaskEVT));
3761 V1Elts.push_back(Elt);
3762 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3767 if (V2InOrder > V1InOrder) {
3768 PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
3770 std::swap(V1Elts, V2Elts);
3771 std::swap(V1FromV1, V2FromV2);
3774 if ((V1FromV1 + V1InOrder) != 8) {
3775 // Some elements are from V2.
3777 // If there are elements that are from V1 but out of place,
3778 // then first sort them in place
3779 SmallVector<SDValue, 8> MaskVec;
3780 for (unsigned i = 0; i < 8; ++i) {
3781 SDValue Elt = V1Elts[i];
3782 if (Elt.getOpcode() == ISD::UNDEF) {
3783 MaskVec.push_back(DAG.getUNDEF(MaskEVT));
3786 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3788 MaskVec.push_back(DAG.getUNDEF(MaskEVT));
3790 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3792 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], 8);
3793 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, V1, V1, Mask);
3797 for (unsigned i = 0; i < 8; ++i) {
3798 SDValue Elt = V1Elts[i];
3799 if (Elt.getOpcode() == ISD::UNDEF)
3801 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3804 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3805 DAG.getConstant(EltIdx - 8, PtrVT));
3806 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3807 DAG.getConstant(i, PtrVT));
3811 // All elements are from V1.
3813 for (unsigned i = 0; i < 8; ++i) {
3814 SDValue Elt = V1Elts[i];
3815 if (Elt.getOpcode() == ISD::UNDEF)
3817 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3818 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3819 DAG.getConstant(EltIdx, PtrVT));
3820 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3821 DAG.getConstant(i, PtrVT));
3827 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3828 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3829 /// done when every pair / quad of shuffle mask elements point to elements in
3830 /// the right sequence. e.g.
3831 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3833 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3835 SDValue PermMask, SelectionDAG &DAG,
3836 TargetLowering &TLI, DebugLoc dl) {
3837 unsigned NumElems = PermMask.getNumOperands();
3838 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3839 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3840 MVT MaskEltVT = MaskVT.getVectorElementType();
3842 switch (VT.getSimpleVT()) {
3843 default: assert(false && "Unexpected!");
3844 case MVT::v4f32: NewVT = MVT::v2f64; break;
3845 case MVT::v4i32: NewVT = MVT::v2i64; break;
3846 case MVT::v8i16: NewVT = MVT::v4i32; break;
3847 case MVT::v16i8: NewVT = MVT::v4i32; break;
3850 if (NewWidth == 2) {
3856 unsigned Scale = NumElems / NewWidth;
3857 SmallVector<SDValue, 8> MaskVec;
3858 for (unsigned i = 0; i < NumElems; i += Scale) {
3859 unsigned StartIdx = ~0U;
3860 for (unsigned j = 0; j < Scale; ++j) {
3861 SDValue Elt = PermMask.getOperand(i+j);
3862 if (Elt.getOpcode() == ISD::UNDEF)
3864 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3865 if (StartIdx == ~0U)
3866 StartIdx = EltIdx - (EltIdx % Scale);
3867 if (EltIdx != StartIdx + j)
3870 if (StartIdx == ~0U)
3871 MaskVec.push_back(DAG.getUNDEF(MaskEltVT));
3873 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3876 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3877 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3878 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, NewVT, V1, V2,
3879 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3880 &MaskVec[0], MaskVec.size()));
3883 /// getVZextMovL - Return a zero-extending vector move low node.
3885 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3886 SDValue SrcOp, SelectionDAG &DAG,
3887 const X86Subtarget *Subtarget, DebugLoc dl) {
3888 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3889 LoadSDNode *LD = NULL;
3890 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3891 LD = dyn_cast<LoadSDNode>(SrcOp);
3893 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3895 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3896 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3897 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3898 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3899 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3901 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3902 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3903 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3904 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3912 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3913 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3914 DAG.getNode(ISD::BIT_CONVERT, dl,
3918 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3921 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3922 SDValue PermMask, MVT VT, SelectionDAG &DAG,
3924 MVT MaskVT = PermMask.getValueType();
3925 MVT MaskEVT = MaskVT.getVectorElementType();
3926 SmallVector<std::pair<int, int>, 8> Locs;
3928 SmallVector<SDValue, 8> Mask1(4, DAG.getUNDEF(MaskEVT));
3931 for (unsigned i = 0; i != 4; ++i) {
3932 SDValue Elt = PermMask.getOperand(i);
3933 if (Elt.getOpcode() == ISD::UNDEF) {
3934 Locs[i] = std::make_pair(-1, -1);
3936 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3937 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3939 Locs[i] = std::make_pair(0, NumLo);
3943 Locs[i] = std::make_pair(1, NumHi);
3945 Mask1[2+NumHi] = Elt;
3951 if (NumLo <= 2 && NumHi <= 2) {
3952 // If no more than two elements come from either vector. This can be
3953 // implemented with two shuffles. First shuffle gather the elements.
3954 // The second shuffle, which takes the first shuffle as both of its
3955 // vector operands, put the elements into the right order.
3956 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
3957 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3958 &Mask1[0], Mask1.size()));
3960 SmallVector<SDValue, 8> Mask2(4, DAG.getUNDEF(MaskEVT));
3961 for (unsigned i = 0; i != 4; ++i) {
3962 if (Locs[i].first == -1)
3965 unsigned Idx = (i < 2) ? 0 : 4;
3966 Idx += Locs[i].first * 2 + Locs[i].second;
3967 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3971 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1,
3972 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3973 &Mask2[0], Mask2.size()));
3974 } else if (NumLo == 3 || NumHi == 3) {
3975 // Otherwise, we must have three elements from one vector, call it X, and
3976 // one element from the other, call it Y. First, use a shufps to build an
3977 // intermediate vector with the one element from Y and the element from X
3978 // that will be in the same half in the final destination (the indexes don't
3979 // matter). Then, use a shufps to build the final vector, taking the half
3980 // containing the element from Y from the intermediate, and the other half
3983 // Normalize it so the 3 elements come from V1.
3984 PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
3988 // Find the element from V2.
3990 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3991 SDValue Elt = PermMask.getOperand(HiIndex);
3992 if (Elt.getOpcode() == ISD::UNDEF)
3994 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3999 Mask1[0] = PermMask.getOperand(HiIndex);
4000 Mask1[1] = DAG.getUNDEF(MaskEVT);
4001 Mask1[2] = PermMask.getOperand(HiIndex^1);
4002 Mask1[3] = DAG.getUNDEF(MaskEVT);
4003 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4004 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &Mask1[0], 4));
4007 Mask1[0] = PermMask.getOperand(0);
4008 Mask1[1] = PermMask.getOperand(1);
4009 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
4010 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
4011 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4012 DAG.getNode(ISD::BUILD_VECTOR, dl,
4013 MaskVT, &Mask1[0], 4));
4015 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
4016 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
4017 Mask1[2] = PermMask.getOperand(2);
4018 Mask1[3] = PermMask.getOperand(3);
4019 if (Mask1[2].getOpcode() != ISD::UNDEF)
4021 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
4023 if (Mask1[3].getOpcode() != ISD::UNDEF)
4025 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
4027 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V2, V1,
4028 DAG.getNode(ISD::BUILD_VECTOR, dl,
4029 MaskVT, &Mask1[0], 4));
4033 // Break it into (shuffle shuffle_hi, shuffle_lo).
4035 SmallVector<SDValue,8> LoMask(4, DAG.getUNDEF(MaskEVT));
4036 SmallVector<SDValue,8> HiMask(4, DAG.getUNDEF(MaskEVT));
4037 SmallVector<SDValue,8> *MaskPtr = &LoMask;
4038 unsigned MaskIdx = 0;
4041 for (unsigned i = 0; i != 4; ++i) {
4048 SDValue Elt = PermMask.getOperand(i);
4049 if (Elt.getOpcode() == ISD::UNDEF) {
4050 Locs[i] = std::make_pair(-1, -1);
4051 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
4052 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4053 (*MaskPtr)[LoIdx] = Elt;
4056 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4057 (*MaskPtr)[HiIdx] = Elt;
4062 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4063 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4064 &LoMask[0], LoMask.size()));
4065 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4066 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4067 &HiMask[0], HiMask.size()));
4068 SmallVector<SDValue, 8> MaskOps;
4069 for (unsigned i = 0; i != 4; ++i) {
4070 if (Locs[i].first == -1) {
4071 MaskOps.push_back(DAG.getUNDEF(MaskEVT));
4073 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4074 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
4077 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, LoShuffle, HiShuffle,
4078 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4079 &MaskOps[0], MaskOps.size()));
4083 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4084 SDValue V1 = Op.getOperand(0);
4085 SDValue V2 = Op.getOperand(1);
4086 SDValue PermMask = Op.getOperand(2);
4087 MVT VT = Op.getValueType();
4088 DebugLoc dl = Op.getDebugLoc();
4089 unsigned NumElems = PermMask.getNumOperands();
4090 bool isMMX = VT.getSizeInBits() == 64;
4091 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4092 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4093 bool V1IsSplat = false;
4094 bool V2IsSplat = false;
4096 if (isUndefShuffle(Op.getNode()))
4097 return DAG.getUNDEF(VT);
4099 if (isZeroShuffle(Op.getNode()))
4100 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4102 if (isIdentityMask(PermMask.getNode()))
4104 else if (isIdentityMask(PermMask.getNode(), true))
4107 // Canonicalize movddup shuffles.
4108 if (V2IsUndef && Subtarget->hasSSE2() &&
4109 VT.getSizeInBits() == 128 &&
4110 X86::isMOVDDUPMask(PermMask.getNode()))
4111 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4113 if (isSplatMask(PermMask.getNode())) {
4114 if (isMMX || NumElems < 4) return Op;
4115 // Promote it to a v4{if}32 splat.
4116 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4119 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4121 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4122 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG,
4124 if (NewOp.getNode())
4125 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4126 LowerVECTOR_SHUFFLE(NewOp, DAG));
4127 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4128 // FIXME: Figure out a cleaner way to do this.
4129 // Try to make use of movq to zero out the top part.
4130 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4131 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4133 if (NewOp.getNode()) {
4134 SDValue NewV1 = NewOp.getOperand(0);
4135 SDValue NewV2 = NewOp.getOperand(1);
4136 SDValue NewMask = NewOp.getOperand(2);
4137 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4138 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4139 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget,
4143 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4144 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4146 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4147 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4148 DAG, Subtarget, dl);
4152 // Check if this can be converted into a logical shift.
4153 bool isLeft = false;
4156 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4157 if (isShift && ShVal.hasOneUse()) {
4158 // If the shifted value has multiple uses, it may be cheaper to use
4159 // v_set0 + movlhps or movhlps, etc.
4160 MVT EVT = VT.getVectorElementType();
4161 ShAmt *= EVT.getSizeInBits();
4162 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4165 if (X86::isMOVLMask(PermMask.getNode())) {
4168 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4169 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4174 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4175 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4176 X86::isMOVHLPSMask(PermMask.getNode()) ||
4177 X86::isMOVHPMask(PermMask.getNode()) ||
4178 X86::isMOVLPMask(PermMask.getNode())))
4181 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4182 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4183 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4186 // No better options. Use a vshl / vsrl.
4187 MVT EVT = VT.getVectorElementType();
4188 ShAmt *= EVT.getSizeInBits();
4189 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4192 bool Commuted = false;
4193 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4194 // 1,1,1,1 -> v8i16 though.
4195 V1IsSplat = isSplatVector(V1.getNode());
4196 V2IsSplat = isSplatVector(V2.getNode());
4198 // Canonicalize the splat or undef, if present, to be on the RHS.
4199 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4200 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4201 std::swap(V1IsSplat, V2IsSplat);
4202 std::swap(V1IsUndef, V2IsUndef);
4206 // FIXME: Figure out a cleaner way to do this.
4207 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4208 if (V2IsUndef) return V1;
4209 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4211 // V2 is a splat, so the mask may be malformed. That is, it may point
4212 // to any V2 element. The instruction selectior won't like this. Get
4213 // a corrected mask and commute to form a proper MOVS{S|D}.
4214 SDValue NewMask = getMOVLMask(NumElems, DAG, dl);
4215 if (NewMask.getNode() != PermMask.getNode())
4216 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4221 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4222 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4223 X86::isUNPCKLMask(PermMask.getNode()) ||
4224 X86::isUNPCKHMask(PermMask.getNode()))
4228 // Normalize mask so all entries that point to V2 points to its first
4229 // element then try to match unpck{h|l} again. If match, return a
4230 // new vector_shuffle with the corrected mask.
4231 SDValue NewMask = NormalizeMask(PermMask, DAG);
4232 if (NewMask.getNode() != PermMask.getNode()) {
4233 if (X86::isUNPCKLMask(NewMask.getNode(), true)) {
4234 SDValue NewMask = getUnpacklMask(NumElems, DAG, dl);
4235 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4236 } else if (X86::isUNPCKHMask(NewMask.getNode(), true)) {
4237 SDValue NewMask = getUnpackhMask(NumElems, DAG, dl);
4238 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4243 // Normalize the node to match x86 shuffle ops if needed
4244 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4245 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4248 // Commute is back and try unpck* again.
4249 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4250 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4251 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4252 X86::isUNPCKLMask(PermMask.getNode()) ||
4253 X86::isUNPCKHMask(PermMask.getNode()))
4257 // Try PSHUF* first, then SHUFP*.
4258 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4259 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4260 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4261 if (V2.getOpcode() != ISD::UNDEF)
4262 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1,
4263 DAG.getUNDEF(VT), PermMask);
4268 if (Subtarget->hasSSE2() &&
4269 (X86::isPSHUFDMask(PermMask.getNode()) ||
4270 X86::isPSHUFHWMask(PermMask.getNode()) ||
4271 X86::isPSHUFLWMask(PermMask.getNode()))) {
4273 if (VT == MVT::v4f32) {
4275 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT,
4276 DAG.getNode(ISD::BIT_CONVERT, dl, RVT, V1),
4277 DAG.getUNDEF(RVT), PermMask);
4278 } else if (V2.getOpcode() != ISD::UNDEF)
4279 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT, V1,
4280 DAG.getUNDEF(RVT), PermMask);
4282 Op = DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
4286 // Binary or unary shufps.
4287 if (X86::isSHUFPMask(PermMask.getNode()) ||
4288 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4292 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4293 if (VT == MVT::v8i16) {
4294 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this, dl);
4295 if (NewOp.getNode())
4299 // Handle all 4 wide cases with a number of shuffles except for MMX.
4300 if (NumElems == 4 && !isMMX)
4301 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG, dl);
4307 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4308 SelectionDAG &DAG) {
4309 MVT VT = Op.getValueType();
4310 DebugLoc dl = Op.getDebugLoc();
4311 if (VT.getSizeInBits() == 8) {
4312 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4313 Op.getOperand(0), Op.getOperand(1));
4314 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4315 DAG.getValueType(VT));
4316 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4317 } else if (VT.getSizeInBits() == 16) {
4318 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4319 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4321 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4322 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4323 DAG.getNode(ISD::BIT_CONVERT, dl,
4327 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4328 Op.getOperand(0), Op.getOperand(1));
4329 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4330 DAG.getValueType(VT));
4331 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4332 } else if (VT == MVT::f32) {
4333 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4334 // the result back to FR32 register. It's only worth matching if the
4335 // result has a single use which is a store or a bitcast to i32. And in
4336 // the case of a store, it's not worth it if the index is a constant 0,
4337 // because a MOVSSmr can be used instead, which is smaller and faster.
4338 if (!Op.hasOneUse())
4340 SDNode *User = *Op.getNode()->use_begin();
4341 if ((User->getOpcode() != ISD::STORE ||
4342 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4343 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4344 (User->getOpcode() != ISD::BIT_CONVERT ||
4345 User->getValueType(0) != MVT::i32))
4347 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4348 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4351 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4352 } else if (VT == MVT::i32) {
4353 // ExtractPS works with constant index.
4354 if (isa<ConstantSDNode>(Op.getOperand(1)))
4362 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4363 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4366 if (Subtarget->hasSSE41()) {
4367 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4372 MVT VT = Op.getValueType();
4373 DebugLoc dl = Op.getDebugLoc();
4374 // TODO: handle v16i8.
4375 if (VT.getSizeInBits() == 16) {
4376 SDValue Vec = Op.getOperand(0);
4377 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4379 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4380 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4381 DAG.getNode(ISD::BIT_CONVERT, dl,
4384 // Transform it so it match pextrw which produces a 32-bit result.
4385 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4386 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4387 Op.getOperand(0), Op.getOperand(1));
4388 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4389 DAG.getValueType(VT));
4390 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4391 } else if (VT.getSizeInBits() == 32) {
4392 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4395 // SHUFPS the element to the lowest double word, then movss.
4396 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4397 SmallVector<SDValue, 8> IdxVec;
4399 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4401 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4403 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4405 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4406 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4407 &IdxVec[0], IdxVec.size());
4408 SDValue Vec = Op.getOperand(0);
4409 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4410 Vec, DAG.getUNDEF(Vec.getValueType()), Mask);
4411 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4412 DAG.getIntPtrConstant(0));
4413 } else if (VT.getSizeInBits() == 64) {
4414 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4415 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4416 // to match extract_elt for f64.
4417 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4421 // UNPCKHPD the element to the lowest double word, then movsd.
4422 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4423 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4424 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4425 SmallVector<SDValue, 8> IdxVec;
4426 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4428 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4429 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4430 &IdxVec[0], IdxVec.size());
4431 SDValue Vec = Op.getOperand(0);
4432 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4433 Vec, DAG.getUNDEF(Vec.getValueType()),
4435 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4436 DAG.getIntPtrConstant(0));
4443 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4444 MVT VT = Op.getValueType();
4445 MVT EVT = VT.getVectorElementType();
4446 DebugLoc dl = Op.getDebugLoc();
4448 SDValue N0 = Op.getOperand(0);
4449 SDValue N1 = Op.getOperand(1);
4450 SDValue N2 = Op.getOperand(2);
4452 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4453 isa<ConstantSDNode>(N2)) {
4454 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4456 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4458 if (N1.getValueType() != MVT::i32)
4459 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4460 if (N2.getValueType() != MVT::i32)
4461 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4462 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4463 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4464 // Bits [7:6] of the constant are the source select. This will always be
4465 // zero here. The DAG Combiner may combine an extract_elt index into these
4466 // bits. For example (insert (extract, 3), 2) could be matched by putting
4467 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4468 // Bits [5:4] of the constant are the destination select. This is the
4469 // value of the incoming immediate.
4470 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4471 // combine either bitwise AND or insert of float 0.0 to set these bits.
4472 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4473 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4474 } else if (EVT == MVT::i32) {
4475 // InsertPS works with constant index.
4476 if (isa<ConstantSDNode>(N2))
4483 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4484 MVT VT = Op.getValueType();
4485 MVT EVT = VT.getVectorElementType();
4487 if (Subtarget->hasSSE41())
4488 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4493 DebugLoc dl = Op.getDebugLoc();
4494 SDValue N0 = Op.getOperand(0);
4495 SDValue N1 = Op.getOperand(1);
4496 SDValue N2 = Op.getOperand(2);
4498 if (EVT.getSizeInBits() == 16) {
4499 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4500 // as its second argument.
4501 if (N1.getValueType() != MVT::i32)
4502 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4503 if (N2.getValueType() != MVT::i32)
4504 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4505 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4511 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4512 DebugLoc dl = Op.getDebugLoc();
4513 if (Op.getValueType() == MVT::v2f32)
4514 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4515 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4516 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4517 Op.getOperand(0))));
4519 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4520 MVT VT = MVT::v2i32;
4521 switch (Op.getValueType().getSimpleVT()) {
4528 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4529 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4532 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4533 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4534 // one of the above mentioned nodes. It has to be wrapped because otherwise
4535 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4536 // be used to form addressing mode. These wrapped nodes will be selected
4539 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4540 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4541 // FIXME there isn't really any debug info here, should come from the parent
4542 DebugLoc dl = CP->getDebugLoc();
4543 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4545 CP->getAlignment());
4546 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4547 // With PIC, the address is actually $g + Offset.
4548 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4549 !Subtarget->isPICStyleRIPRel()) {
4550 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4551 DAG.getNode(X86ISD::GlobalBaseReg,
4552 DebugLoc::getUnknownLoc(),
4561 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4563 SelectionDAG &DAG) const {
4564 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4565 bool ExtraLoadRequired =
4566 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4568 // Create the TargetGlobalAddress node, folding in the constant
4569 // offset if it is legal.
4571 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4572 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4575 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4576 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4578 // With PIC, the address is actually $g + Offset.
4579 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4580 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4581 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4585 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4586 // load the value at address GV, not the value of GV itself. This means that
4587 // the GlobalAddress must be in the base or index register of the address, not
4588 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4589 // The same applies for external symbols during PIC codegen
4590 if (ExtraLoadRequired)
4591 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4592 PseudoSourceValue::getGOT(), 0);
4594 // If there was a non-zero offset that we didn't fold, create an explicit
4597 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4598 DAG.getConstant(Offset, getPointerTy()));
4604 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4605 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4606 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4607 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4610 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4612 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4615 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4616 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4617 DAG.getNode(X86ISD::GlobalBaseReg,
4618 DebugLoc::getUnknownLoc(),
4620 InFlag = Chain.getValue(1);
4622 // emit leal symbol@TLSGD(,%ebx,1), %eax
4623 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4624 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4625 GA->getValueType(0),
4627 SDValue Ops[] = { Chain, TGA, InFlag };
4628 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4629 InFlag = Result.getValue(2);
4630 Chain = Result.getValue(1);
4632 // call ___tls_get_addr. This function receives its argument in
4633 // the register EAX.
4634 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Result, InFlag);
4635 InFlag = Chain.getValue(1);
4637 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4638 SDValue Ops1[] = { Chain,
4639 DAG.getTargetExternalSymbol("___tls_get_addr",
4641 DAG.getRegister(X86::EAX, PtrVT),
4642 DAG.getRegister(X86::EBX, PtrVT),
4644 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
4645 InFlag = Chain.getValue(1);
4647 return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
4650 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4652 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4654 SDValue InFlag, Chain;
4655 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4657 // emit leaq symbol@TLSGD(%rip), %rdi
4658 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4659 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4660 GA->getValueType(0),
4662 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4663 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4664 Chain = Result.getValue(1);
4665 InFlag = Result.getValue(2);
4667 // call __tls_get_addr. This function receives its argument in
4668 // the register RDI.
4669 Chain = DAG.getCopyToReg(Chain, dl, X86::RDI, Result, InFlag);
4670 InFlag = Chain.getValue(1);
4672 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4673 SDValue Ops1[] = { Chain,
4674 DAG.getTargetExternalSymbol("__tls_get_addr",
4676 DAG.getRegister(X86::RDI, PtrVT),
4678 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
4679 InFlag = Chain.getValue(1);
4681 return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
4684 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4685 // "local exec" model.
4686 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4688 DebugLoc dl = GA->getDebugLoc();
4689 // Get the Thread Pointer
4690 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER,
4691 DebugLoc::getUnknownLoc(), PtrVT);
4692 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4694 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4695 GA->getValueType(0),
4697 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4699 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4700 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4701 PseudoSourceValue::getGOT(), 0);
4703 // The address of the thread local variable is the add of the thread
4704 // pointer with the offset of the variable.
4705 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4709 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4710 // TODO: implement the "local dynamic" model
4711 // TODO: implement the "initial exec"model for pic executables
4712 assert(Subtarget->isTargetELF() &&
4713 "TLS not implemented for non-ELF targets");
4714 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4715 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4716 // otherwise use the "Local Exec"TLS Model
4717 if (Subtarget->is64Bit()) {
4718 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4720 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4721 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4723 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4728 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4729 // FIXME there isn't really any debug info here
4730 DebugLoc dl = Op.getDebugLoc();
4731 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4732 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4733 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4734 // With PIC, the address is actually $g + Offset.
4735 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4736 !Subtarget->isPICStyleRIPRel()) {
4737 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4738 DAG.getNode(X86ISD::GlobalBaseReg,
4739 DebugLoc::getUnknownLoc(),
4747 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4748 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4749 // FIXME there isn't really any debug into here
4750 DebugLoc dl = JT->getDebugLoc();
4751 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4752 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4753 // With PIC, the address is actually $g + Offset.
4754 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4755 !Subtarget->isPICStyleRIPRel()) {
4756 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4757 DAG.getNode(X86ISD::GlobalBaseReg,
4758 DebugLoc::getUnknownLoc(),
4766 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4767 /// take a 2 x i32 value to shift plus a shift amount.
4768 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4769 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4770 MVT VT = Op.getValueType();
4771 unsigned VTBits = VT.getSizeInBits();
4772 DebugLoc dl = Op.getDebugLoc();
4773 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4774 SDValue ShOpLo = Op.getOperand(0);
4775 SDValue ShOpHi = Op.getOperand(1);
4776 SDValue ShAmt = Op.getOperand(2);
4777 SDValue Tmp1 = isSRA ?
4778 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4779 DAG.getConstant(VTBits - 1, MVT::i8)) :
4780 DAG.getConstant(0, VT);
4783 if (Op.getOpcode() == ISD::SHL_PARTS) {
4784 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4785 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4787 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4788 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4791 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4792 DAG.getConstant(VTBits, MVT::i8));
4793 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4794 AndNode, DAG.getConstant(0, MVT::i8));
4797 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4798 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4799 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4801 if (Op.getOpcode() == ISD::SHL_PARTS) {
4802 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4803 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4805 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4806 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4809 SDValue Ops[2] = { Lo, Hi };
4810 return DAG.getMergeValues(Ops, 2, dl);
4813 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4814 MVT SrcVT = Op.getOperand(0).getValueType();
4815 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4816 "Unknown SINT_TO_FP to lower!");
4818 // These are really Legal; caller falls through into that case.
4819 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4821 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4822 Subtarget->is64Bit())
4825 DebugLoc dl = Op.getDebugLoc();
4826 unsigned Size = SrcVT.getSizeInBits()/8;
4827 MachineFunction &MF = DAG.getMachineFunction();
4828 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4829 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4830 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4832 PseudoSourceValue::getFixedStack(SSFI), 0);
4836 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4838 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4840 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4841 SmallVector<SDValue, 8> Ops;
4842 Ops.push_back(Chain);
4843 Ops.push_back(StackSlot);
4844 Ops.push_back(DAG.getValueType(SrcVT));
4845 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4846 Tys, &Ops[0], Ops.size());
4849 Chain = Result.getValue(1);
4850 SDValue InFlag = Result.getValue(2);
4852 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4853 // shouldn't be necessary except that RFP cannot be live across
4854 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4855 MachineFunction &MF = DAG.getMachineFunction();
4856 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4857 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4858 Tys = DAG.getVTList(MVT::Other);
4859 SmallVector<SDValue, 8> Ops;
4860 Ops.push_back(Chain);
4861 Ops.push_back(Result);
4862 Ops.push_back(StackSlot);
4863 Ops.push_back(DAG.getValueType(Op.getValueType()));
4864 Ops.push_back(InFlag);
4865 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4866 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4867 PseudoSourceValue::getFixedStack(SSFI), 0);
4873 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4874 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4875 // This algorithm is not obvious. Here it is in C code, more or less:
4877 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4878 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4879 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4881 // Copy ints to xmm registers.
4882 __m128i xh = _mm_cvtsi32_si128( hi );
4883 __m128i xl = _mm_cvtsi32_si128( lo );
4885 // Combine into low half of a single xmm register.
4886 __m128i x = _mm_unpacklo_epi32( xh, xl );
4890 // Merge in appropriate exponents to give the integer bits the right
4892 x = _mm_unpacklo_epi32( x, exp );
4894 // Subtract away the biases to deal with the IEEE-754 double precision
4896 d = _mm_sub_pd( (__m128d) x, bias );
4898 // All conversions up to here are exact. The correctly rounded result is
4899 // calculated using the current rounding mode using the following
4901 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4902 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4903 // store doesn't really need to be here (except
4904 // maybe to zero the other double)
4909 DebugLoc dl = Op.getDebugLoc();
4911 // Build some magic constants.
4912 std::vector<Constant*> CV0;
4913 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4914 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4915 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4916 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4917 Constant *C0 = ConstantVector::get(CV0);
4918 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 4);
4920 std::vector<Constant*> CV1;
4921 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4922 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4923 Constant *C1 = ConstantVector::get(CV1);
4924 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 4);
4926 SmallVector<SDValue, 4> MaskVec;
4927 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
4928 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
4929 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
4930 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
4931 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
4932 &MaskVec[0], MaskVec.size());
4933 SmallVector<SDValue, 4> MaskVec2;
4934 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
4935 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
4936 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32,
4937 &MaskVec2[0], MaskVec2.size());
4939 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4940 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4942 DAG.getIntPtrConstant(1)));
4943 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4944 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4946 DAG.getIntPtrConstant(0)));
4947 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
4948 XR1, XR2, UnpcklMask);
4949 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4950 PseudoSourceValue::getConstantPool(), 0,
4952 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
4953 Unpck1, CLod0, UnpcklMask);
4954 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4955 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4956 PseudoSourceValue::getConstantPool(), 0,
4958 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4960 // Add the halves; easiest way is to swap them into another reg first.
4961 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2f64,
4962 Sub, Sub, ShufMask);
4963 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4964 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4965 DAG.getIntPtrConstant(0));
4968 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4969 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4970 DebugLoc dl = Op.getDebugLoc();
4971 // FP constant to bias correct the final result.
4972 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4975 // Load the 32-bit value into an XMM register.
4976 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4977 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4979 DAG.getIntPtrConstant(0)));
4981 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4982 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4983 DAG.getIntPtrConstant(0));
4985 // Or the load with the bias.
4986 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4987 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4988 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4990 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4991 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4992 MVT::v2f64, Bias)));
4993 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4994 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4995 DAG.getIntPtrConstant(0));
4997 // Subtract the bias.
4998 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5000 // Handle final rounding.
5001 MVT DestVT = Op.getValueType();
5003 if (DestVT.bitsLT(MVT::f64)) {
5004 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5005 DAG.getIntPtrConstant(0));
5006 } else if (DestVT.bitsGT(MVT::f64)) {
5007 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5010 // Handle final rounding.
5014 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5015 SDValue N0 = Op.getOperand(0);
5016 DebugLoc dl = Op.getDebugLoc();
5018 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5019 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5020 // the optimization here.
5021 if (DAG.SignBitIsZero(N0))
5022 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5024 MVT SrcVT = N0.getValueType();
5025 if (SrcVT == MVT::i64) {
5026 // We only handle SSE2 f64 target here; caller can handle the rest.
5027 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5030 return LowerUINT_TO_FP_i64(Op, DAG);
5031 } else if (SrcVT == MVT::i32) {
5032 return LowerUINT_TO_FP_i32(Op, DAG);
5035 assert(0 && "Unknown UINT_TO_FP to lower!");
5039 std::pair<SDValue,SDValue> X86TargetLowering::
5040 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
5041 DebugLoc dl = Op.getDebugLoc();
5042 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
5043 Op.getValueType().getSimpleVT() >= MVT::i16 &&
5044 "Unknown FP_TO_SINT to lower!");
5046 // These are really Legal.
5047 if (Op.getValueType() == MVT::i32 &&
5048 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5049 return std::make_pair(SDValue(), SDValue());
5050 if (Subtarget->is64Bit() &&
5051 Op.getValueType() == MVT::i64 &&
5052 Op.getOperand(0).getValueType() != MVT::f80)
5053 return std::make_pair(SDValue(), SDValue());
5055 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5057 MachineFunction &MF = DAG.getMachineFunction();
5058 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
5059 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5060 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5062 switch (Op.getValueType().getSimpleVT()) {
5063 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5064 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5065 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5066 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5069 SDValue Chain = DAG.getEntryNode();
5070 SDValue Value = Op.getOperand(0);
5071 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5072 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5073 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5074 PseudoSourceValue::getFixedStack(SSFI), 0);
5075 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5077 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5079 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5080 Chain = Value.getValue(1);
5081 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5082 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5085 // Build the FP_TO_INT*_IN_MEM
5086 SDValue Ops[] = { Chain, Value, StackSlot };
5087 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5089 return std::make_pair(FIST, StackSlot);
5092 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5093 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
5094 SDValue FIST = Vals.first, StackSlot = Vals.second;
5095 if (FIST.getNode() == 0) return SDValue();
5098 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5099 FIST, StackSlot, NULL, 0);
5102 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5103 DebugLoc dl = Op.getDebugLoc();
5104 MVT VT = Op.getValueType();
5107 EltVT = VT.getVectorElementType();
5108 std::vector<Constant*> CV;
5109 if (EltVT == MVT::f64) {
5110 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5114 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5120 Constant *C = ConstantVector::get(CV);
5121 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5122 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5123 PseudoSourceValue::getConstantPool(), 0,
5125 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5128 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5129 DebugLoc dl = Op.getDebugLoc();
5130 MVT VT = Op.getValueType();
5132 unsigned EltNum = 1;
5133 if (VT.isVector()) {
5134 EltVT = VT.getVectorElementType();
5135 EltNum = VT.getVectorNumElements();
5137 std::vector<Constant*> CV;
5138 if (EltVT == MVT::f64) {
5139 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5143 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5149 Constant *C = ConstantVector::get(CV);
5150 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5151 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5152 PseudoSourceValue::getConstantPool(), 0,
5154 if (VT.isVector()) {
5155 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5156 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5157 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5159 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5161 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5165 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5166 SDValue Op0 = Op.getOperand(0);
5167 SDValue Op1 = Op.getOperand(1);
5168 DebugLoc dl = Op.getDebugLoc();
5169 MVT VT = Op.getValueType();
5170 MVT SrcVT = Op1.getValueType();
5172 // If second operand is smaller, extend it first.
5173 if (SrcVT.bitsLT(VT)) {
5174 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5177 // And if it is bigger, shrink it first.
5178 if (SrcVT.bitsGT(VT)) {
5179 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5183 // At this point the operands and the result should have the same
5184 // type, and that won't be f80 since that is not custom lowered.
5186 // First get the sign bit of second operand.
5187 std::vector<Constant*> CV;
5188 if (SrcVT == MVT::f64) {
5189 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5190 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5192 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5193 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5194 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5195 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5197 Constant *C = ConstantVector::get(CV);
5198 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5199 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5200 PseudoSourceValue::getConstantPool(), 0,
5202 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5204 // Shift sign bit right or left if the two operands have different types.
5205 if (SrcVT.bitsGT(VT)) {
5206 // Op0 is MVT::f32, Op1 is MVT::f64.
5207 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5208 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5209 DAG.getConstant(32, MVT::i32));
5210 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5211 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5212 DAG.getIntPtrConstant(0));
5215 // Clear first operand sign bit.
5217 if (VT == MVT::f64) {
5218 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5219 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5221 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5222 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5223 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5224 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5226 C = ConstantVector::get(CV);
5227 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5228 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5229 PseudoSourceValue::getConstantPool(), 0,
5231 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5233 // Or the value with the sign bit.
5234 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5237 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5238 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5239 SDValue Op0 = Op.getOperand(0);
5240 SDValue Op1 = Op.getOperand(1);
5241 DebugLoc dl = Op.getDebugLoc();
5242 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5244 // Lower (X & (1 << N)) == 0 to BT(X, N).
5245 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5246 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5247 if (Op0.getOpcode() == ISD::AND &&
5249 Op1.getOpcode() == ISD::Constant &&
5250 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5251 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5253 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5254 if (ConstantSDNode *Op010C =
5255 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5256 if (Op010C->getZExtValue() == 1) {
5257 LHS = Op0.getOperand(0);
5258 RHS = Op0.getOperand(1).getOperand(1);
5260 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5261 if (ConstantSDNode *Op000C =
5262 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5263 if (Op000C->getZExtValue() == 1) {
5264 LHS = Op0.getOperand(1);
5265 RHS = Op0.getOperand(0).getOperand(1);
5267 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5268 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5269 SDValue AndLHS = Op0.getOperand(0);
5270 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5271 LHS = AndLHS.getOperand(0);
5272 RHS = AndLHS.getOperand(1);
5276 if (LHS.getNode()) {
5277 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5278 // instruction. Since the shift amount is in-range-or-undefined, we know
5279 // that doing a bittest on the i16 value is ok. We extend to i32 because
5280 // the encoding for the i16 version is larger than the i32 version.
5281 if (LHS.getValueType() == MVT::i8)
5282 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5284 // If the operand types disagree, extend the shift amount to match. Since
5285 // BT ignores high bits (like shifts) we can use anyextend.
5286 if (LHS.getValueType() != RHS.getValueType())
5287 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5289 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5290 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5291 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5292 DAG.getConstant(Cond, MVT::i8), BT);
5296 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5297 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5299 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5300 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5301 DAG.getConstant(X86CC, MVT::i8), Cond);
5304 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5306 SDValue Op0 = Op.getOperand(0);
5307 SDValue Op1 = Op.getOperand(1);
5308 SDValue CC = Op.getOperand(2);
5309 MVT VT = Op.getValueType();
5310 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5311 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5312 DebugLoc dl = Op.getDebugLoc();
5316 MVT VT0 = Op0.getValueType();
5317 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5318 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5321 switch (SetCCOpcode) {
5324 case ISD::SETEQ: SSECC = 0; break;
5326 case ISD::SETGT: Swap = true; // Fallthrough
5328 case ISD::SETOLT: SSECC = 1; break;
5330 case ISD::SETGE: Swap = true; // Fallthrough
5332 case ISD::SETOLE: SSECC = 2; break;
5333 case ISD::SETUO: SSECC = 3; break;
5335 case ISD::SETNE: SSECC = 4; break;
5336 case ISD::SETULE: Swap = true;
5337 case ISD::SETUGE: SSECC = 5; break;
5338 case ISD::SETULT: Swap = true;
5339 case ISD::SETUGT: SSECC = 6; break;
5340 case ISD::SETO: SSECC = 7; break;
5343 std::swap(Op0, Op1);
5345 // In the two special cases we can't handle, emit two comparisons.
5347 if (SetCCOpcode == ISD::SETUEQ) {
5349 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5350 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5351 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5353 else if (SetCCOpcode == ISD::SETONE) {
5355 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5356 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5357 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5359 assert(0 && "Illegal FP comparison");
5361 // Handle all other FP comparisons here.
5362 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5365 // We are handling one of the integer comparisons here. Since SSE only has
5366 // GT and EQ comparisons for integer, swapping operands and multiple
5367 // operations may be required for some comparisons.
5368 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5369 bool Swap = false, Invert = false, FlipSigns = false;
5371 switch (VT.getSimpleVT()) {
5373 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5374 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5375 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5376 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5379 switch (SetCCOpcode) {
5381 case ISD::SETNE: Invert = true;
5382 case ISD::SETEQ: Opc = EQOpc; break;
5383 case ISD::SETLT: Swap = true;
5384 case ISD::SETGT: Opc = GTOpc; break;
5385 case ISD::SETGE: Swap = true;
5386 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5387 case ISD::SETULT: Swap = true;
5388 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5389 case ISD::SETUGE: Swap = true;
5390 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5393 std::swap(Op0, Op1);
5395 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5396 // bits of the inputs before performing those operations.
5398 MVT EltVT = VT.getVectorElementType();
5399 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5401 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5402 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5404 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5405 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5408 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5410 // If the logical-not of the result is required, perform that now.
5412 Result = DAG.getNOT(dl, Result, VT);
5417 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5418 static bool isX86LogicalCmp(unsigned Opc) {
5419 return Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI;
5422 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5423 bool addTest = true;
5424 SDValue Cond = Op.getOperand(0);
5425 DebugLoc dl = Op.getDebugLoc();
5428 if (Cond.getOpcode() == ISD::SETCC)
5429 Cond = LowerSETCC(Cond, DAG);
5431 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5432 // setting operand in place of the X86ISD::SETCC.
5433 if (Cond.getOpcode() == X86ISD::SETCC) {
5434 CC = Cond.getOperand(0);
5436 SDValue Cmp = Cond.getOperand(1);
5437 unsigned Opc = Cmp.getOpcode();
5438 MVT VT = Op.getValueType();
5440 bool IllegalFPCMov = false;
5441 if (VT.isFloatingPoint() && !VT.isVector() &&
5442 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5443 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5445 if ((isX86LogicalCmp(Opc) && !IllegalFPCMov) || Opc == X86ISD::BT) { // FIXME
5452 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5453 Cond= DAG.getNode(X86ISD::CMP, dl, MVT::i32, Cond,
5454 DAG.getConstant(0, MVT::i8));
5457 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5459 SmallVector<SDValue, 4> Ops;
5460 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5461 // condition is true.
5462 Ops.push_back(Op.getOperand(2));
5463 Ops.push_back(Op.getOperand(1));
5465 Ops.push_back(Cond);
5466 return DAG.getNode(X86ISD::CMOV, dl, VTs, 2, &Ops[0], Ops.size());
5469 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5470 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5471 // from the AND / OR.
5472 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5473 Opc = Op.getOpcode();
5474 if (Opc != ISD::OR && Opc != ISD::AND)
5476 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5477 Op.getOperand(0).hasOneUse() &&
5478 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5479 Op.getOperand(1).hasOneUse());
5482 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5483 // 1 and that the SETCC node has a single use.
5484 static bool isXor1OfSetCC(SDValue Op) {
5485 if (Op.getOpcode() != ISD::XOR)
5487 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5488 if (N1C && N1C->getAPIntValue() == 1) {
5489 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5490 Op.getOperand(0).hasOneUse();
5495 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5496 bool addTest = true;
5497 SDValue Chain = Op.getOperand(0);
5498 SDValue Cond = Op.getOperand(1);
5499 SDValue Dest = Op.getOperand(2);
5500 DebugLoc dl = Op.getDebugLoc();
5503 if (Cond.getOpcode() == ISD::SETCC)
5504 Cond = LowerSETCC(Cond, DAG);
5506 // FIXME: LowerXALUO doesn't handle these!!
5507 else if (Cond.getOpcode() == X86ISD::ADD ||
5508 Cond.getOpcode() == X86ISD::SUB ||
5509 Cond.getOpcode() == X86ISD::SMUL ||
5510 Cond.getOpcode() == X86ISD::UMUL)
5511 Cond = LowerXALUO(Cond, DAG);
5514 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5515 // setting operand in place of the X86ISD::SETCC.
5516 if (Cond.getOpcode() == X86ISD::SETCC) {
5517 CC = Cond.getOperand(0);
5519 SDValue Cmp = Cond.getOperand(1);
5520 unsigned Opc = Cmp.getOpcode();
5521 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5522 if (isX86LogicalCmp(Opc) || Opc == X86ISD::BT) {
5526 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5530 // These can only come from an arithmetic instruction with overflow,
5531 // e.g. SADDO, UADDO.
5532 Cond = Cond.getNode()->getOperand(1);
5539 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5540 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5541 unsigned Opc = Cmp.getOpcode();
5542 if (CondOpc == ISD::OR) {
5543 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5544 // two branches instead of an explicit OR instruction with a
5546 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5547 isX86LogicalCmp(Opc)) {
5548 CC = Cond.getOperand(0).getOperand(0);
5549 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5550 Chain, Dest, CC, Cmp);
5551 CC = Cond.getOperand(1).getOperand(0);
5555 } else { // ISD::AND
5556 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5557 // two branches instead of an explicit AND instruction with a
5558 // separate test. However, we only do this if this block doesn't
5559 // have a fall-through edge, because this requires an explicit
5560 // jmp when the condition is false.
5561 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5562 isX86LogicalCmp(Opc) &&
5563 Op.getNode()->hasOneUse()) {
5564 X86::CondCode CCode =
5565 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5566 CCode = X86::GetOppositeBranchCondition(CCode);
5567 CC = DAG.getConstant(CCode, MVT::i8);
5568 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5569 // Look for an unconditional branch following this conditional branch.
5570 // We need this because we need to reverse the successors in order
5571 // to implement FCMP_OEQ.
5572 if (User.getOpcode() == ISD::BR) {
5573 SDValue FalseBB = User.getOperand(1);
5575 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5576 assert(NewBR == User);
5579 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5580 Chain, Dest, CC, Cmp);
5581 X86::CondCode CCode =
5582 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5583 CCode = X86::GetOppositeBranchCondition(CCode);
5584 CC = DAG.getConstant(CCode, MVT::i8);
5590 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5591 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5592 // It should be transformed during dag combiner except when the condition
5593 // is set by a arithmetics with overflow node.
5594 X86::CondCode CCode =
5595 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5596 CCode = X86::GetOppositeBranchCondition(CCode);
5597 CC = DAG.getConstant(CCode, MVT::i8);
5598 Cond = Cond.getOperand(0).getOperand(1);
5604 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5605 Cond= DAG.getNode(X86ISD::CMP, dl, MVT::i32, Cond,
5606 DAG.getConstant(0, MVT::i8));
5608 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5609 Chain, Dest, CC, Cond);
5613 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5614 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5615 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5616 // that the guard pages used by the OS virtual memory manager are allocated in
5617 // correct sequence.
5619 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5620 SelectionDAG &DAG) {
5621 assert(Subtarget->isTargetCygMing() &&
5622 "This should be used only on Cygwin/Mingw targets");
5623 DebugLoc dl = Op.getDebugLoc();
5626 SDValue Chain = Op.getOperand(0);
5627 SDValue Size = Op.getOperand(1);
5628 // FIXME: Ensure alignment here
5632 MVT IntPtr = getPointerTy();
5633 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5635 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5637 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5638 Flag = Chain.getValue(1);
5640 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5641 SDValue Ops[] = { Chain,
5642 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5643 DAG.getRegister(X86::EAX, IntPtr),
5644 DAG.getRegister(X86StackPtr, SPTy),
5646 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5647 Flag = Chain.getValue(1);
5649 Chain = DAG.getCALLSEQ_END(Chain,
5650 DAG.getIntPtrConstant(0, true),
5651 DAG.getIntPtrConstant(0, true),
5654 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5656 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5657 return DAG.getMergeValues(Ops1, 2, dl);
5661 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5663 SDValue Dst, SDValue Src,
5664 SDValue Size, unsigned Align,
5666 uint64_t DstSVOff) {
5667 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5669 // If not DWORD aligned or size is more than the threshold, call the library.
5670 // The libc version is likely to be faster for these cases. It can use the
5671 // address value and run time information about the CPU.
5672 if ((Align & 3) != 0 ||
5674 ConstantSize->getZExtValue() >
5675 getSubtarget()->getMaxInlineSizeThreshold()) {
5676 SDValue InFlag(0, 0);
5678 // Check to see if there is a specialized entry-point for memory zeroing.
5679 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5681 if (const char *bzeroEntry = V &&
5682 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5683 MVT IntPtr = getPointerTy();
5684 const Type *IntPtrTy = TD->getIntPtrType();
5685 TargetLowering::ArgListTy Args;
5686 TargetLowering::ArgListEntry Entry;
5688 Entry.Ty = IntPtrTy;
5689 Args.push_back(Entry);
5691 Args.push_back(Entry);
5692 std::pair<SDValue,SDValue> CallResult =
5693 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5694 CallingConv::C, false,
5695 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5696 return CallResult.second;
5699 // Otherwise have the target-independent code call memset.
5703 uint64_t SizeVal = ConstantSize->getZExtValue();
5704 SDValue InFlag(0, 0);
5707 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5708 unsigned BytesLeft = 0;
5709 bool TwoRepStos = false;
5712 uint64_t Val = ValC->getZExtValue() & 255;
5714 // If the value is a constant, then we can potentially use larger sets.
5715 switch (Align & 3) {
5716 case 2: // WORD aligned
5719 Val = (Val << 8) | Val;
5721 case 0: // DWORD aligned
5724 Val = (Val << 8) | Val;
5725 Val = (Val << 16) | Val;
5726 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5729 Val = (Val << 32) | Val;
5732 default: // Byte aligned
5735 Count = DAG.getIntPtrConstant(SizeVal);
5739 if (AVT.bitsGT(MVT::i8)) {
5740 unsigned UBytes = AVT.getSizeInBits() / 8;
5741 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5742 BytesLeft = SizeVal % UBytes;
5745 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5747 InFlag = Chain.getValue(1);
5750 Count = DAG.getIntPtrConstant(SizeVal);
5751 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5752 InFlag = Chain.getValue(1);
5755 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5758 InFlag = Chain.getValue(1);
5759 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5762 InFlag = Chain.getValue(1);
5764 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5765 SmallVector<SDValue, 8> Ops;
5766 Ops.push_back(Chain);
5767 Ops.push_back(DAG.getValueType(AVT));
5768 Ops.push_back(InFlag);
5769 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5772 InFlag = Chain.getValue(1);
5774 MVT CVT = Count.getValueType();
5775 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5776 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5777 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5780 InFlag = Chain.getValue(1);
5781 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5783 Ops.push_back(Chain);
5784 Ops.push_back(DAG.getValueType(MVT::i8));
5785 Ops.push_back(InFlag);
5786 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5787 } else if (BytesLeft) {
5788 // Handle the last 1 - 7 bytes.
5789 unsigned Offset = SizeVal - BytesLeft;
5790 MVT AddrVT = Dst.getValueType();
5791 MVT SizeVT = Size.getValueType();
5793 Chain = DAG.getMemset(Chain, dl,
5794 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5795 DAG.getConstant(Offset, AddrVT)),
5797 DAG.getConstant(BytesLeft, SizeVT),
5798 Align, DstSV, DstSVOff + Offset);
5801 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5806 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5807 SDValue Chain, SDValue Dst, SDValue Src,
5808 SDValue Size, unsigned Align,
5810 const Value *DstSV, uint64_t DstSVOff,
5811 const Value *SrcSV, uint64_t SrcSVOff) {
5812 // This requires the copy size to be a constant, preferrably
5813 // within a subtarget-specific limit.
5814 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5817 uint64_t SizeVal = ConstantSize->getZExtValue();
5818 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5821 /// If not DWORD aligned, call the library.
5822 if ((Align & 3) != 0)
5827 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5830 unsigned UBytes = AVT.getSizeInBits() / 8;
5831 unsigned CountVal = SizeVal / UBytes;
5832 SDValue Count = DAG.getIntPtrConstant(CountVal);
5833 unsigned BytesLeft = SizeVal % UBytes;
5835 SDValue InFlag(0, 0);
5836 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5839 InFlag = Chain.getValue(1);
5840 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5843 InFlag = Chain.getValue(1);
5844 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5847 InFlag = Chain.getValue(1);
5849 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5850 SmallVector<SDValue, 8> Ops;
5851 Ops.push_back(Chain);
5852 Ops.push_back(DAG.getValueType(AVT));
5853 Ops.push_back(InFlag);
5854 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5856 SmallVector<SDValue, 4> Results;
5857 Results.push_back(RepMovs);
5859 // Handle the last 1 - 7 bytes.
5860 unsigned Offset = SizeVal - BytesLeft;
5861 MVT DstVT = Dst.getValueType();
5862 MVT SrcVT = Src.getValueType();
5863 MVT SizeVT = Size.getValueType();
5864 Results.push_back(DAG.getMemcpy(Chain, dl,
5865 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5866 DAG.getConstant(Offset, DstVT)),
5867 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5868 DAG.getConstant(Offset, SrcVT)),
5869 DAG.getConstant(BytesLeft, SizeVT),
5870 Align, AlwaysInline,
5871 DstSV, DstSVOff + Offset,
5872 SrcSV, SrcSVOff + Offset));
5875 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5876 &Results[0], Results.size());
5879 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5880 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5881 DebugLoc dl = Op.getDebugLoc();
5883 if (!Subtarget->is64Bit()) {
5884 // vastart just stores the address of the VarArgsFrameIndex slot into the
5885 // memory location argument.
5886 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5887 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
5891 // gp_offset (0 - 6 * 8)
5892 // fp_offset (48 - 48 + 8 * 16)
5893 // overflow_arg_area (point to parameters coming in memory).
5895 SmallVector<SDValue, 8> MemOps;
5896 SDValue FIN = Op.getOperand(1);
5898 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
5899 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5901 MemOps.push_back(Store);
5904 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5905 FIN, DAG.getIntPtrConstant(4));
5906 Store = DAG.getStore(Op.getOperand(0), dl,
5907 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5909 MemOps.push_back(Store);
5911 // Store ptr to overflow_arg_area
5912 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5913 FIN, DAG.getIntPtrConstant(4));
5914 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5915 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
5916 MemOps.push_back(Store);
5918 // Store ptr to reg_save_area.
5919 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5920 FIN, DAG.getIntPtrConstant(8));
5921 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5922 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
5923 MemOps.push_back(Store);
5924 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5925 &MemOps[0], MemOps.size());
5928 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5929 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5930 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5931 SDValue Chain = Op.getOperand(0);
5932 SDValue SrcPtr = Op.getOperand(1);
5933 SDValue SrcSV = Op.getOperand(2);
5935 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5940 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5941 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5942 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5943 SDValue Chain = Op.getOperand(0);
5944 SDValue DstPtr = Op.getOperand(1);
5945 SDValue SrcPtr = Op.getOperand(2);
5946 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5947 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5948 DebugLoc dl = Op.getDebugLoc();
5950 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
5951 DAG.getIntPtrConstant(24), 8, false,
5952 DstSV, 0, SrcSV, 0);
5956 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5957 DebugLoc dl = Op.getDebugLoc();
5958 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5960 default: return SDValue(); // Don't custom lower most intrinsics.
5961 // Comparison intrinsics.
5962 case Intrinsic::x86_sse_comieq_ss:
5963 case Intrinsic::x86_sse_comilt_ss:
5964 case Intrinsic::x86_sse_comile_ss:
5965 case Intrinsic::x86_sse_comigt_ss:
5966 case Intrinsic::x86_sse_comige_ss:
5967 case Intrinsic::x86_sse_comineq_ss:
5968 case Intrinsic::x86_sse_ucomieq_ss:
5969 case Intrinsic::x86_sse_ucomilt_ss:
5970 case Intrinsic::x86_sse_ucomile_ss:
5971 case Intrinsic::x86_sse_ucomigt_ss:
5972 case Intrinsic::x86_sse_ucomige_ss:
5973 case Intrinsic::x86_sse_ucomineq_ss:
5974 case Intrinsic::x86_sse2_comieq_sd:
5975 case Intrinsic::x86_sse2_comilt_sd:
5976 case Intrinsic::x86_sse2_comile_sd:
5977 case Intrinsic::x86_sse2_comigt_sd:
5978 case Intrinsic::x86_sse2_comige_sd:
5979 case Intrinsic::x86_sse2_comineq_sd:
5980 case Intrinsic::x86_sse2_ucomieq_sd:
5981 case Intrinsic::x86_sse2_ucomilt_sd:
5982 case Intrinsic::x86_sse2_ucomile_sd:
5983 case Intrinsic::x86_sse2_ucomigt_sd:
5984 case Intrinsic::x86_sse2_ucomige_sd:
5985 case Intrinsic::x86_sse2_ucomineq_sd: {
5987 ISD::CondCode CC = ISD::SETCC_INVALID;
5990 case Intrinsic::x86_sse_comieq_ss:
5991 case Intrinsic::x86_sse2_comieq_sd:
5995 case Intrinsic::x86_sse_comilt_ss:
5996 case Intrinsic::x86_sse2_comilt_sd:
6000 case Intrinsic::x86_sse_comile_ss:
6001 case Intrinsic::x86_sse2_comile_sd:
6005 case Intrinsic::x86_sse_comigt_ss:
6006 case Intrinsic::x86_sse2_comigt_sd:
6010 case Intrinsic::x86_sse_comige_ss:
6011 case Intrinsic::x86_sse2_comige_sd:
6015 case Intrinsic::x86_sse_comineq_ss:
6016 case Intrinsic::x86_sse2_comineq_sd:
6020 case Intrinsic::x86_sse_ucomieq_ss:
6021 case Intrinsic::x86_sse2_ucomieq_sd:
6022 Opc = X86ISD::UCOMI;
6025 case Intrinsic::x86_sse_ucomilt_ss:
6026 case Intrinsic::x86_sse2_ucomilt_sd:
6027 Opc = X86ISD::UCOMI;
6030 case Intrinsic::x86_sse_ucomile_ss:
6031 case Intrinsic::x86_sse2_ucomile_sd:
6032 Opc = X86ISD::UCOMI;
6035 case Intrinsic::x86_sse_ucomigt_ss:
6036 case Intrinsic::x86_sse2_ucomigt_sd:
6037 Opc = X86ISD::UCOMI;
6040 case Intrinsic::x86_sse_ucomige_ss:
6041 case Intrinsic::x86_sse2_ucomige_sd:
6042 Opc = X86ISD::UCOMI;
6045 case Intrinsic::x86_sse_ucomineq_ss:
6046 case Intrinsic::x86_sse2_ucomineq_sd:
6047 Opc = X86ISD::UCOMI;
6052 SDValue LHS = Op.getOperand(1);
6053 SDValue RHS = Op.getOperand(2);
6054 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6055 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6056 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6057 DAG.getConstant(X86CC, MVT::i8), Cond);
6058 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6061 // Fix vector shift instructions where the last operand is a non-immediate
6063 case Intrinsic::x86_sse2_pslli_w:
6064 case Intrinsic::x86_sse2_pslli_d:
6065 case Intrinsic::x86_sse2_pslli_q:
6066 case Intrinsic::x86_sse2_psrli_w:
6067 case Intrinsic::x86_sse2_psrli_d:
6068 case Intrinsic::x86_sse2_psrli_q:
6069 case Intrinsic::x86_sse2_psrai_w:
6070 case Intrinsic::x86_sse2_psrai_d:
6071 case Intrinsic::x86_mmx_pslli_w:
6072 case Intrinsic::x86_mmx_pslli_d:
6073 case Intrinsic::x86_mmx_pslli_q:
6074 case Intrinsic::x86_mmx_psrli_w:
6075 case Intrinsic::x86_mmx_psrli_d:
6076 case Intrinsic::x86_mmx_psrli_q:
6077 case Intrinsic::x86_mmx_psrai_w:
6078 case Intrinsic::x86_mmx_psrai_d: {
6079 SDValue ShAmt = Op.getOperand(2);
6080 if (isa<ConstantSDNode>(ShAmt))
6083 unsigned NewIntNo = 0;
6084 MVT ShAmtVT = MVT::v4i32;
6086 case Intrinsic::x86_sse2_pslli_w:
6087 NewIntNo = Intrinsic::x86_sse2_psll_w;
6089 case Intrinsic::x86_sse2_pslli_d:
6090 NewIntNo = Intrinsic::x86_sse2_psll_d;
6092 case Intrinsic::x86_sse2_pslli_q:
6093 NewIntNo = Intrinsic::x86_sse2_psll_q;
6095 case Intrinsic::x86_sse2_psrli_w:
6096 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6098 case Intrinsic::x86_sse2_psrli_d:
6099 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6101 case Intrinsic::x86_sse2_psrli_q:
6102 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6104 case Intrinsic::x86_sse2_psrai_w:
6105 NewIntNo = Intrinsic::x86_sse2_psra_w;
6107 case Intrinsic::x86_sse2_psrai_d:
6108 NewIntNo = Intrinsic::x86_sse2_psra_d;
6111 ShAmtVT = MVT::v2i32;
6113 case Intrinsic::x86_mmx_pslli_w:
6114 NewIntNo = Intrinsic::x86_mmx_psll_w;
6116 case Intrinsic::x86_mmx_pslli_d:
6117 NewIntNo = Intrinsic::x86_mmx_psll_d;
6119 case Intrinsic::x86_mmx_pslli_q:
6120 NewIntNo = Intrinsic::x86_mmx_psll_q;
6122 case Intrinsic::x86_mmx_psrli_w:
6123 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6125 case Intrinsic::x86_mmx_psrli_d:
6126 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6128 case Intrinsic::x86_mmx_psrli_q:
6129 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6131 case Intrinsic::x86_mmx_psrai_w:
6132 NewIntNo = Intrinsic::x86_mmx_psra_w;
6134 case Intrinsic::x86_mmx_psrai_d:
6135 NewIntNo = Intrinsic::x86_mmx_psra_d;
6137 default: abort(); // Can't reach here.
6142 MVT VT = Op.getValueType();
6143 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6144 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6145 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6146 DAG.getConstant(NewIntNo, MVT::i32),
6147 Op.getOperand(1), ShAmt);
6152 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6153 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6154 DebugLoc dl = Op.getDebugLoc();
6157 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6159 DAG.getConstant(TD->getPointerSize(),
6160 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6161 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6162 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6167 // Just load the return address.
6168 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6169 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6170 RetAddrFI, NULL, 0);
6173 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6174 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6175 MFI->setFrameAddressIsTaken(true);
6176 MVT VT = Op.getValueType();
6177 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6178 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6179 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6180 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6182 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6186 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6187 SelectionDAG &DAG) {
6188 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6191 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6193 MachineFunction &MF = DAG.getMachineFunction();
6194 SDValue Chain = Op.getOperand(0);
6195 SDValue Offset = Op.getOperand(1);
6196 SDValue Handler = Op.getOperand(2);
6197 DebugLoc dl = Op.getDebugLoc();
6199 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6201 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6203 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6204 DAG.getIntPtrConstant(-TD->getPointerSize()));
6205 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6206 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6207 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6208 MF.getRegInfo().addLiveOut(StoreAddrReg);
6210 return DAG.getNode(X86ISD::EH_RETURN, dl,
6212 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6215 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6216 SelectionDAG &DAG) {
6217 SDValue Root = Op.getOperand(0);
6218 SDValue Trmp = Op.getOperand(1); // trampoline
6219 SDValue FPtr = Op.getOperand(2); // nested function
6220 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6221 DebugLoc dl = Op.getDebugLoc();
6223 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6225 const X86InstrInfo *TII =
6226 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6228 if (Subtarget->is64Bit()) {
6229 SDValue OutChains[6];
6231 // Large code-model.
6233 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6234 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6236 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6237 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6239 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6241 // Load the pointer to the nested function into R11.
6242 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6243 SDValue Addr = Trmp;
6244 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6247 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6248 DAG.getConstant(2, MVT::i64));
6249 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6251 // Load the 'nest' parameter value into R10.
6252 // R10 is specified in X86CallingConv.td
6253 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6254 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6255 DAG.getConstant(10, MVT::i64));
6256 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6257 Addr, TrmpAddr, 10);
6259 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6260 DAG.getConstant(12, MVT::i64));
6261 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6263 // Jump to the nested function.
6264 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6265 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6266 DAG.getConstant(20, MVT::i64));
6267 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6268 Addr, TrmpAddr, 20);
6270 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6271 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6272 DAG.getConstant(22, MVT::i64));
6273 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6277 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6278 return DAG.getMergeValues(Ops, 2, dl);
6280 const Function *Func =
6281 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6282 unsigned CC = Func->getCallingConv();
6287 assert(0 && "Unsupported calling convention");
6288 case CallingConv::C:
6289 case CallingConv::X86_StdCall: {
6290 // Pass 'nest' parameter in ECX.
6291 // Must be kept in sync with X86CallingConv.td
6294 // Check that ECX wasn't needed by an 'inreg' parameter.
6295 const FunctionType *FTy = Func->getFunctionType();
6296 const AttrListPtr &Attrs = Func->getAttributes();
6298 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6299 unsigned InRegCount = 0;
6302 for (FunctionType::param_iterator I = FTy->param_begin(),
6303 E = FTy->param_end(); I != E; ++I, ++Idx)
6304 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6305 // FIXME: should only count parameters that are lowered to integers.
6306 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6308 if (InRegCount > 2) {
6309 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6315 case CallingConv::X86_FastCall:
6316 case CallingConv::Fast:
6317 // Pass 'nest' parameter in EAX.
6318 // Must be kept in sync with X86CallingConv.td
6323 SDValue OutChains[4];
6326 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6327 DAG.getConstant(10, MVT::i32));
6328 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6330 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6331 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6332 OutChains[0] = DAG.getStore(Root, dl,
6333 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6336 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6337 DAG.getConstant(1, MVT::i32));
6338 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6340 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6341 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6342 DAG.getConstant(5, MVT::i32));
6343 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6344 TrmpAddr, 5, false, 1);
6346 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6347 DAG.getConstant(6, MVT::i32));
6348 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6351 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6352 return DAG.getMergeValues(Ops, 2, dl);
6356 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6358 The rounding mode is in bits 11:10 of FPSR, and has the following
6365 FLT_ROUNDS, on the other hand, expects the following:
6372 To perform the conversion, we do:
6373 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6376 MachineFunction &MF = DAG.getMachineFunction();
6377 const TargetMachine &TM = MF.getTarget();
6378 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6379 unsigned StackAlignment = TFI.getStackAlignment();
6380 MVT VT = Op.getValueType();
6381 DebugLoc dl = Op.getDebugLoc();
6383 // Save FP Control Word to stack slot
6384 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6385 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6387 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6388 DAG.getEntryNode(), StackSlot);
6390 // Load FP Control Word from stack slot
6391 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6393 // Transform as necessary
6395 DAG.getNode(ISD::SRL, dl, MVT::i16,
6396 DAG.getNode(ISD::AND, dl, MVT::i16,
6397 CWD, DAG.getConstant(0x800, MVT::i16)),
6398 DAG.getConstant(11, MVT::i8));
6400 DAG.getNode(ISD::SRL, dl, MVT::i16,
6401 DAG.getNode(ISD::AND, dl, MVT::i16,
6402 CWD, DAG.getConstant(0x400, MVT::i16)),
6403 DAG.getConstant(9, MVT::i8));
6406 DAG.getNode(ISD::AND, dl, MVT::i16,
6407 DAG.getNode(ISD::ADD, dl, MVT::i16,
6408 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6409 DAG.getConstant(1, MVT::i16)),
6410 DAG.getConstant(3, MVT::i16));
6413 return DAG.getNode((VT.getSizeInBits() < 16 ?
6414 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6417 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6418 MVT VT = Op.getValueType();
6420 unsigned NumBits = VT.getSizeInBits();
6421 DebugLoc dl = Op.getDebugLoc();
6423 Op = Op.getOperand(0);
6424 if (VT == MVT::i8) {
6425 // Zero extend to i32 since there is not an i8 bsr.
6427 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6430 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6431 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6432 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6434 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6435 SmallVector<SDValue, 4> Ops;
6437 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6438 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6439 Ops.push_back(Op.getValue(1));
6440 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6442 // Finally xor with NumBits-1.
6443 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6446 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6450 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6451 MVT VT = Op.getValueType();
6453 unsigned NumBits = VT.getSizeInBits();
6454 DebugLoc dl = Op.getDebugLoc();
6456 Op = Op.getOperand(0);
6457 if (VT == MVT::i8) {
6459 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6462 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6463 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6464 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6466 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6467 SmallVector<SDValue, 4> Ops;
6469 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6470 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6471 Ops.push_back(Op.getValue(1));
6472 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6475 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6479 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6480 MVT VT = Op.getValueType();
6481 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6482 DebugLoc dl = Op.getDebugLoc();
6484 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6485 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6486 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6487 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6488 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6490 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6491 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6492 // return AloBlo + AloBhi + AhiBlo;
6494 SDValue A = Op.getOperand(0);
6495 SDValue B = Op.getOperand(1);
6497 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6498 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6499 A, DAG.getConstant(32, MVT::i32));
6500 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6501 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6502 B, DAG.getConstant(32, MVT::i32));
6503 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6504 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6506 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6507 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6509 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6510 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6512 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6513 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6514 AloBhi, DAG.getConstant(32, MVT::i32));
6515 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6516 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6517 AhiBlo, DAG.getConstant(32, MVT::i32));
6518 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6519 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6524 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6525 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6526 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6527 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6528 // has only one use.
6529 SDNode *N = Op.getNode();
6530 SDValue LHS = N->getOperand(0);
6531 SDValue RHS = N->getOperand(1);
6532 unsigned BaseOp = 0;
6534 DebugLoc dl = Op.getDebugLoc();
6536 switch (Op.getOpcode()) {
6537 default: assert(0 && "Unknown ovf instruction!");
6539 BaseOp = X86ISD::ADD;
6543 BaseOp = X86ISD::ADD;
6547 BaseOp = X86ISD::SUB;
6551 BaseOp = X86ISD::SUB;
6555 BaseOp = X86ISD::SMUL;
6559 BaseOp = X86ISD::UMUL;
6564 // Also sets EFLAGS.
6565 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6566 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6569 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6570 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6572 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6576 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6577 MVT T = Op.getValueType();
6578 DebugLoc dl = Op.getDebugLoc();
6581 switch(T.getSimpleVT()) {
6583 assert(false && "Invalid value type!");
6584 case MVT::i8: Reg = X86::AL; size = 1; break;
6585 case MVT::i16: Reg = X86::AX; size = 2; break;
6586 case MVT::i32: Reg = X86::EAX; size = 4; break;
6588 assert(Subtarget->is64Bit() && "Node not type legal!");
6589 Reg = X86::RAX; size = 8;
6592 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6593 Op.getOperand(2), SDValue());
6594 SDValue Ops[] = { cpIn.getValue(0),
6597 DAG.getTargetConstant(size, MVT::i8),
6599 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6600 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6602 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6606 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6607 SelectionDAG &DAG) {
6608 assert(Subtarget->is64Bit() && "Result not type legalized?");
6609 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6610 SDValue TheChain = Op.getOperand(0);
6611 DebugLoc dl = Op.getDebugLoc();
6612 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6613 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6614 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6616 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6617 DAG.getConstant(32, MVT::i8));
6619 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6622 return DAG.getMergeValues(Ops, 2, dl);
6625 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6626 SDNode *Node = Op.getNode();
6627 DebugLoc dl = Node->getDebugLoc();
6628 MVT T = Node->getValueType(0);
6629 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6630 DAG.getConstant(0, T), Node->getOperand(2));
6631 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6632 cast<AtomicSDNode>(Node)->getMemoryVT(),
6633 Node->getOperand(0),
6634 Node->getOperand(1), negOp,
6635 cast<AtomicSDNode>(Node)->getSrcValue(),
6636 cast<AtomicSDNode>(Node)->getAlignment());
6639 /// LowerOperation - Provide custom lowering hooks for some operations.
6641 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6642 switch (Op.getOpcode()) {
6643 default: assert(0 && "Should not custom lower this!");
6644 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6645 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6646 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6647 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6648 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6649 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6650 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6651 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6652 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6653 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6654 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6655 case ISD::SHL_PARTS:
6656 case ISD::SRA_PARTS:
6657 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6658 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6659 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6660 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6661 case ISD::FABS: return LowerFABS(Op, DAG);
6662 case ISD::FNEG: return LowerFNEG(Op, DAG);
6663 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6664 case ISD::SETCC: return LowerSETCC(Op, DAG);
6665 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6666 case ISD::SELECT: return LowerSELECT(Op, DAG);
6667 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6668 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6669 case ISD::CALL: return LowerCALL(Op, DAG);
6670 case ISD::RET: return LowerRET(Op, DAG);
6671 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6672 case ISD::VASTART: return LowerVASTART(Op, DAG);
6673 case ISD::VAARG: return LowerVAARG(Op, DAG);
6674 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6675 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6676 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6677 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6678 case ISD::FRAME_TO_ARGS_OFFSET:
6679 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6680 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6681 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6682 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6683 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6684 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6685 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6686 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6692 case ISD::UMULO: return LowerXALUO(Op, DAG);
6693 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6697 void X86TargetLowering::
6698 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6699 SelectionDAG &DAG, unsigned NewOp) {
6700 MVT T = Node->getValueType(0);
6701 DebugLoc dl = Node->getDebugLoc();
6702 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6704 SDValue Chain = Node->getOperand(0);
6705 SDValue In1 = Node->getOperand(1);
6706 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6707 Node->getOperand(2), DAG.getIntPtrConstant(0));
6708 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6709 Node->getOperand(2), DAG.getIntPtrConstant(1));
6710 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6711 // have a MemOperand. Pass the info through as a normal operand.
6712 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6713 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6714 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6715 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6716 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6717 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6718 Results.push_back(Result.getValue(2));
6721 /// ReplaceNodeResults - Replace a node with an illegal result type
6722 /// with a new node built out of custom code.
6723 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6724 SmallVectorImpl<SDValue>&Results,
6725 SelectionDAG &DAG) {
6726 DebugLoc dl = N->getDebugLoc();
6727 switch (N->getOpcode()) {
6729 assert(false && "Do not know how to custom type legalize this operation!");
6731 case ISD::FP_TO_SINT: {
6732 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
6733 SDValue FIST = Vals.first, StackSlot = Vals.second;
6734 if (FIST.getNode() != 0) {
6735 MVT VT = N->getValueType(0);
6736 // Return a load from the stack slot.
6737 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6741 case ISD::READCYCLECOUNTER: {
6742 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6743 SDValue TheChain = N->getOperand(0);
6744 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6745 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6747 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6749 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6750 SDValue Ops[] = { eax, edx };
6751 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6752 Results.push_back(edx.getValue(1));
6755 case ISD::ATOMIC_CMP_SWAP: {
6756 MVT T = N->getValueType(0);
6757 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6758 SDValue cpInL, cpInH;
6759 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6760 DAG.getConstant(0, MVT::i32));
6761 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6762 DAG.getConstant(1, MVT::i32));
6763 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6764 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6766 SDValue swapInL, swapInH;
6767 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6768 DAG.getConstant(0, MVT::i32));
6769 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6770 DAG.getConstant(1, MVT::i32));
6771 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6773 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6774 swapInL.getValue(1));
6775 SDValue Ops[] = { swapInH.getValue(0),
6777 swapInH.getValue(1) };
6778 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6779 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6780 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6781 MVT::i32, Result.getValue(1));
6782 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6783 MVT::i32, cpOutL.getValue(2));
6784 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6785 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6786 Results.push_back(cpOutH.getValue(1));
6789 case ISD::ATOMIC_LOAD_ADD:
6790 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6792 case ISD::ATOMIC_LOAD_AND:
6793 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6795 case ISD::ATOMIC_LOAD_NAND:
6796 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6798 case ISD::ATOMIC_LOAD_OR:
6799 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6801 case ISD::ATOMIC_LOAD_SUB:
6802 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6804 case ISD::ATOMIC_LOAD_XOR:
6805 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6807 case ISD::ATOMIC_SWAP:
6808 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6813 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6815 default: return NULL;
6816 case X86ISD::BSF: return "X86ISD::BSF";
6817 case X86ISD::BSR: return "X86ISD::BSR";
6818 case X86ISD::SHLD: return "X86ISD::SHLD";
6819 case X86ISD::SHRD: return "X86ISD::SHRD";
6820 case X86ISD::FAND: return "X86ISD::FAND";
6821 case X86ISD::FOR: return "X86ISD::FOR";
6822 case X86ISD::FXOR: return "X86ISD::FXOR";
6823 case X86ISD::FSRL: return "X86ISD::FSRL";
6824 case X86ISD::FILD: return "X86ISD::FILD";
6825 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6826 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6827 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6828 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6829 case X86ISD::FLD: return "X86ISD::FLD";
6830 case X86ISD::FST: return "X86ISD::FST";
6831 case X86ISD::CALL: return "X86ISD::CALL";
6832 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6833 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6834 case X86ISD::BT: return "X86ISD::BT";
6835 case X86ISD::CMP: return "X86ISD::CMP";
6836 case X86ISD::COMI: return "X86ISD::COMI";
6837 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6838 case X86ISD::SETCC: return "X86ISD::SETCC";
6839 case X86ISD::CMOV: return "X86ISD::CMOV";
6840 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6841 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6842 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6843 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6844 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6845 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6846 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6847 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6848 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6849 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6850 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6851 case X86ISD::FMAX: return "X86ISD::FMAX";
6852 case X86ISD::FMIN: return "X86ISD::FMIN";
6853 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6854 case X86ISD::FRCP: return "X86ISD::FRCP";
6855 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6856 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6857 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6858 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6859 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6860 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6861 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6862 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6863 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6864 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6865 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6866 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6867 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6868 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6869 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6870 case X86ISD::VSHL: return "X86ISD::VSHL";
6871 case X86ISD::VSRL: return "X86ISD::VSRL";
6872 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6873 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6874 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6875 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6876 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6877 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6878 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6879 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6880 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6881 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6882 case X86ISD::ADD: return "X86ISD::ADD";
6883 case X86ISD::SUB: return "X86ISD::SUB";
6884 case X86ISD::SMUL: return "X86ISD::SMUL";
6885 case X86ISD::UMUL: return "X86ISD::UMUL";
6889 // isLegalAddressingMode - Return true if the addressing mode represented
6890 // by AM is legal for this target, for a load/store of the specified type.
6891 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6892 const Type *Ty) const {
6893 // X86 supports extremely general addressing modes.
6895 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6896 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6900 // We can only fold this if we don't need an extra load.
6901 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6903 // If BaseGV requires a register, we cannot also have a BaseReg.
6904 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6908 // X86-64 only supports addr of globals in small code model.
6909 if (Subtarget->is64Bit()) {
6910 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6912 // If lower 4G is not available, then we must use rip-relative addressing.
6913 if (AM.BaseOffs || AM.Scale > 1)
6924 // These scales always work.
6929 // These scales are formed with basereg+scalereg. Only accept if there is
6934 default: // Other stuff never works.
6942 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6943 if (!Ty1->isInteger() || !Ty2->isInteger())
6945 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6946 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6947 if (NumBits1 <= NumBits2)
6949 return Subtarget->is64Bit() || NumBits1 < 64;
6952 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6953 if (!VT1.isInteger() || !VT2.isInteger())
6955 unsigned NumBits1 = VT1.getSizeInBits();
6956 unsigned NumBits2 = VT2.getSizeInBits();
6957 if (NumBits1 <= NumBits2)
6959 return Subtarget->is64Bit() || NumBits1 < 64;
6962 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6963 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6964 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6965 /// are assumed to be legal.
6967 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6968 // Only do shuffles on 128-bit vector types for now.
6969 if (VT.getSizeInBits() == 64) return false;
6970 return (Mask.getNode()->getNumOperands() <= 4 ||
6971 isIdentityMask(Mask.getNode()) ||
6972 isIdentityMask(Mask.getNode(), true) ||
6973 isSplatMask(Mask.getNode()) ||
6974 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6975 X86::isUNPCKLMask(Mask.getNode()) ||
6976 X86::isUNPCKHMask(Mask.getNode()) ||
6977 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6978 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6982 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6983 MVT EVT, SelectionDAG &DAG) const {
6984 unsigned NumElts = BVOps.size();
6985 // Only do shuffles on 128-bit vector types for now.
6986 if (EVT.getSizeInBits() * NumElts == 64) return false;
6987 if (NumElts == 2) return true;
6989 return (isMOVLMask(&BVOps[0], 4) ||
6990 isCommutedMOVL(&BVOps[0], 4, true) ||
6991 isSHUFPMask(&BVOps[0], 4) ||
6992 isCommutedSHUFP(&BVOps[0], 4));
6997 //===----------------------------------------------------------------------===//
6998 // X86 Scheduler Hooks
6999 //===----------------------------------------------------------------------===//
7001 // private utility function
7003 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7004 MachineBasicBlock *MBB,
7012 TargetRegisterClass *RC,
7013 bool invSrc) const {
7014 // For the atomic bitwise operator, we generate
7017 // ld t1 = [bitinstr.addr]
7018 // op t2 = t1, [bitinstr.val]
7020 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7022 // fallthrough -->nextMBB
7023 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7024 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7025 MachineFunction::iterator MBBIter = MBB;
7028 /// First build the CFG
7029 MachineFunction *F = MBB->getParent();
7030 MachineBasicBlock *thisMBB = MBB;
7031 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7032 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7033 F->insert(MBBIter, newMBB);
7034 F->insert(MBBIter, nextMBB);
7036 // Move all successors to thisMBB to nextMBB
7037 nextMBB->transferSuccessors(thisMBB);
7039 // Update thisMBB to fall through to newMBB
7040 thisMBB->addSuccessor(newMBB);
7042 // newMBB jumps to itself and fall through to nextMBB
7043 newMBB->addSuccessor(nextMBB);
7044 newMBB->addSuccessor(newMBB);
7046 // Insert instructions into newMBB based on incoming instruction
7047 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
7048 DebugLoc dl = bInstr->getDebugLoc();
7049 MachineOperand& destOper = bInstr->getOperand(0);
7050 MachineOperand* argOpers[6];
7051 int numArgs = bInstr->getNumOperands() - 1;
7052 for (int i=0; i < numArgs; ++i)
7053 argOpers[i] = &bInstr->getOperand(i+1);
7055 // x86 address has 4 operands: base, index, scale, and displacement
7056 int lastAddrIndx = 3; // [0,3]
7059 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7060 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7061 for (int i=0; i <= lastAddrIndx; ++i)
7062 (*MIB).addOperand(*argOpers[i]);
7064 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7066 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7071 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7072 assert((argOpers[valArgIndx]->isReg() ||
7073 argOpers[valArgIndx]->isImm()) &&
7075 if (argOpers[valArgIndx]->isReg())
7076 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7078 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7080 (*MIB).addOperand(*argOpers[valArgIndx]);
7082 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7085 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7086 for (int i=0; i <= lastAddrIndx; ++i)
7087 (*MIB).addOperand(*argOpers[i]);
7089 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7090 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7092 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7096 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7098 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7102 // private utility function: 64 bit atomics on 32 bit host.
7104 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7105 MachineBasicBlock *MBB,
7110 bool invSrc) const {
7111 // For the atomic bitwise operator, we generate
7112 // thisMBB (instructions are in pairs, except cmpxchg8b)
7113 // ld t1,t2 = [bitinstr.addr]
7115 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7116 // op t5, t6 <- out1, out2, [bitinstr.val]
7117 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7118 // mov ECX, EBX <- t5, t6
7119 // mov EAX, EDX <- t1, t2
7120 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7121 // mov t3, t4 <- EAX, EDX
7123 // result in out1, out2
7124 // fallthrough -->nextMBB
7126 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7127 const unsigned LoadOpc = X86::MOV32rm;
7128 const unsigned copyOpc = X86::MOV32rr;
7129 const unsigned NotOpc = X86::NOT32r;
7130 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7131 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7132 MachineFunction::iterator MBBIter = MBB;
7135 /// First build the CFG
7136 MachineFunction *F = MBB->getParent();
7137 MachineBasicBlock *thisMBB = MBB;
7138 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7139 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7140 F->insert(MBBIter, newMBB);
7141 F->insert(MBBIter, nextMBB);
7143 // Move all successors to thisMBB to nextMBB
7144 nextMBB->transferSuccessors(thisMBB);
7146 // Update thisMBB to fall through to newMBB
7147 thisMBB->addSuccessor(newMBB);
7149 // newMBB jumps to itself and fall through to nextMBB
7150 newMBB->addSuccessor(nextMBB);
7151 newMBB->addSuccessor(newMBB);
7153 DebugLoc dl = bInstr->getDebugLoc();
7154 // Insert instructions into newMBB based on incoming instruction
7155 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7156 assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
7157 MachineOperand& dest1Oper = bInstr->getOperand(0);
7158 MachineOperand& dest2Oper = bInstr->getOperand(1);
7159 MachineOperand* argOpers[6];
7160 for (int i=0; i < 6; ++i)
7161 argOpers[i] = &bInstr->getOperand(i+2);
7163 // x86 address has 4 operands: base, index, scale, and displacement
7164 int lastAddrIndx = 3; // [0,3]
7166 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7167 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7168 for (int i=0; i <= lastAddrIndx; ++i)
7169 (*MIB).addOperand(*argOpers[i]);
7170 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7171 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7172 // add 4 to displacement.
7173 for (int i=0; i <= lastAddrIndx-1; ++i)
7174 (*MIB).addOperand(*argOpers[i]);
7175 MachineOperand newOp3 = *(argOpers[3]);
7177 newOp3.setImm(newOp3.getImm()+4);
7179 newOp3.setOffset(newOp3.getOffset()+4);
7180 (*MIB).addOperand(newOp3);
7182 // t3/4 are defined later, at the bottom of the loop
7183 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7184 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7185 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7186 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7187 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7188 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7190 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7191 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7193 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7194 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7200 assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
7202 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7203 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7204 if (argOpers[4]->isReg())
7205 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7207 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7208 if (regOpcL != X86::MOV32rr)
7210 (*MIB).addOperand(*argOpers[4]);
7211 assert(argOpers[5]->isReg() == argOpers[4]->isReg());
7212 assert(argOpers[5]->isImm() == argOpers[4]->isImm());
7213 if (argOpers[5]->isReg())
7214 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7216 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7217 if (regOpcH != X86::MOV32rr)
7219 (*MIB).addOperand(*argOpers[5]);
7221 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7223 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7226 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7228 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7231 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7232 for (int i=0; i <= lastAddrIndx; ++i)
7233 (*MIB).addOperand(*argOpers[i]);
7235 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7236 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7238 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7239 MIB.addReg(X86::EAX);
7240 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7241 MIB.addReg(X86::EDX);
7244 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7246 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7250 // private utility function
7252 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7253 MachineBasicBlock *MBB,
7254 unsigned cmovOpc) const {
7255 // For the atomic min/max operator, we generate
7258 // ld t1 = [min/max.addr]
7259 // mov t2 = [min/max.val]
7261 // cmov[cond] t2 = t1
7263 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7265 // fallthrough -->nextMBB
7267 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7268 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7269 MachineFunction::iterator MBBIter = MBB;
7272 /// First build the CFG
7273 MachineFunction *F = MBB->getParent();
7274 MachineBasicBlock *thisMBB = MBB;
7275 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7276 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7277 F->insert(MBBIter, newMBB);
7278 F->insert(MBBIter, nextMBB);
7280 // Move all successors to thisMBB to nextMBB
7281 nextMBB->transferSuccessors(thisMBB);
7283 // Update thisMBB to fall through to newMBB
7284 thisMBB->addSuccessor(newMBB);
7286 // newMBB jumps to newMBB and fall through to nextMBB
7287 newMBB->addSuccessor(nextMBB);
7288 newMBB->addSuccessor(newMBB);
7290 DebugLoc dl = mInstr->getDebugLoc();
7291 // Insert instructions into newMBB based on incoming instruction
7292 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
7293 MachineOperand& destOper = mInstr->getOperand(0);
7294 MachineOperand* argOpers[6];
7295 int numArgs = mInstr->getNumOperands() - 1;
7296 for (int i=0; i < numArgs; ++i)
7297 argOpers[i] = &mInstr->getOperand(i+1);
7299 // x86 address has 4 operands: base, index, scale, and displacement
7300 int lastAddrIndx = 3; // [0,3]
7303 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7304 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7305 for (int i=0; i <= lastAddrIndx; ++i)
7306 (*MIB).addOperand(*argOpers[i]);
7308 // We only support register and immediate values
7309 assert((argOpers[valArgIndx]->isReg() ||
7310 argOpers[valArgIndx]->isImm()) &&
7313 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7314 if (argOpers[valArgIndx]->isReg())
7315 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7317 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7318 (*MIB).addOperand(*argOpers[valArgIndx]);
7320 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7323 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7328 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7329 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7333 // Cmp and exchange if none has modified the memory location
7334 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7335 for (int i=0; i <= lastAddrIndx; ++i)
7336 (*MIB).addOperand(*argOpers[i]);
7338 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7339 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7341 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7342 MIB.addReg(X86::EAX);
7345 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7347 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7353 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7354 MachineBasicBlock *BB) const {
7355 DebugLoc dl = MI->getDebugLoc();
7356 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7357 switch (MI->getOpcode()) {
7358 default: assert(false && "Unexpected instr type to insert");
7359 case X86::CMOV_V1I64:
7360 case X86::CMOV_FR32:
7361 case X86::CMOV_FR64:
7362 case X86::CMOV_V4F32:
7363 case X86::CMOV_V2F64:
7364 case X86::CMOV_V2I64: {
7365 // To "insert" a SELECT_CC instruction, we actually have to insert the
7366 // diamond control-flow pattern. The incoming instruction knows the
7367 // destination vreg to set, the condition code register to branch on, the
7368 // true/false values to select between, and a branch opcode to use.
7369 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7370 MachineFunction::iterator It = BB;
7376 // cmpTY ccX, r1, r2
7378 // fallthrough --> copy0MBB
7379 MachineBasicBlock *thisMBB = BB;
7380 MachineFunction *F = BB->getParent();
7381 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7382 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7384 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7385 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7386 F->insert(It, copy0MBB);
7387 F->insert(It, sinkMBB);
7388 // Update machine-CFG edges by transferring all successors of the current
7389 // block to the new block which will contain the Phi node for the select.
7390 sinkMBB->transferSuccessors(BB);
7392 // Add the true and fallthrough blocks as its successors.
7393 BB->addSuccessor(copy0MBB);
7394 BB->addSuccessor(sinkMBB);
7397 // %FalseValue = ...
7398 // # fallthrough to sinkMBB
7401 // Update machine-CFG edges
7402 BB->addSuccessor(sinkMBB);
7405 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7408 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7409 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7410 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7412 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7416 case X86::FP32_TO_INT16_IN_MEM:
7417 case X86::FP32_TO_INT32_IN_MEM:
7418 case X86::FP32_TO_INT64_IN_MEM:
7419 case X86::FP64_TO_INT16_IN_MEM:
7420 case X86::FP64_TO_INT32_IN_MEM:
7421 case X86::FP64_TO_INT64_IN_MEM:
7422 case X86::FP80_TO_INT16_IN_MEM:
7423 case X86::FP80_TO_INT32_IN_MEM:
7424 case X86::FP80_TO_INT64_IN_MEM: {
7425 // Change the floating point control register to use "round towards zero"
7426 // mode when truncating to an integer value.
7427 MachineFunction *F = BB->getParent();
7428 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7429 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7431 // Load the old value of the high byte of the control word...
7433 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7434 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7437 // Set the high part to be round to zero...
7438 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7441 // Reload the modified control word now...
7442 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7444 // Restore the memory image of control word to original value
7445 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7448 // Get the X86 opcode to use.
7450 switch (MI->getOpcode()) {
7451 default: assert(0 && "illegal opcode!");
7452 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7453 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7454 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7455 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7456 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7457 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7458 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7459 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7460 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7464 MachineOperand &Op = MI->getOperand(0);
7466 AM.BaseType = X86AddressMode::RegBase;
7467 AM.Base.Reg = Op.getReg();
7469 AM.BaseType = X86AddressMode::FrameIndexBase;
7470 AM.Base.FrameIndex = Op.getIndex();
7472 Op = MI->getOperand(1);
7474 AM.Scale = Op.getImm();
7475 Op = MI->getOperand(2);
7477 AM.IndexReg = Op.getImm();
7478 Op = MI->getOperand(3);
7479 if (Op.isGlobal()) {
7480 AM.GV = Op.getGlobal();
7482 AM.Disp = Op.getImm();
7484 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7485 .addReg(MI->getOperand(4).getReg());
7487 // Reload the original control word now.
7488 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7490 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7493 case X86::ATOMAND32:
7494 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7495 X86::AND32ri, X86::MOV32rm,
7496 X86::LCMPXCHG32, X86::MOV32rr,
7497 X86::NOT32r, X86::EAX,
7498 X86::GR32RegisterClass);
7500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7501 X86::OR32ri, X86::MOV32rm,
7502 X86::LCMPXCHG32, X86::MOV32rr,
7503 X86::NOT32r, X86::EAX,
7504 X86::GR32RegisterClass);
7505 case X86::ATOMXOR32:
7506 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7507 X86::XOR32ri, X86::MOV32rm,
7508 X86::LCMPXCHG32, X86::MOV32rr,
7509 X86::NOT32r, X86::EAX,
7510 X86::GR32RegisterClass);
7511 case X86::ATOMNAND32:
7512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7513 X86::AND32ri, X86::MOV32rm,
7514 X86::LCMPXCHG32, X86::MOV32rr,
7515 X86::NOT32r, X86::EAX,
7516 X86::GR32RegisterClass, true);
7517 case X86::ATOMMIN32:
7518 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7519 case X86::ATOMMAX32:
7520 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7521 case X86::ATOMUMIN32:
7522 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7523 case X86::ATOMUMAX32:
7524 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7526 case X86::ATOMAND16:
7527 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7528 X86::AND16ri, X86::MOV16rm,
7529 X86::LCMPXCHG16, X86::MOV16rr,
7530 X86::NOT16r, X86::AX,
7531 X86::GR16RegisterClass);
7533 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7534 X86::OR16ri, X86::MOV16rm,
7535 X86::LCMPXCHG16, X86::MOV16rr,
7536 X86::NOT16r, X86::AX,
7537 X86::GR16RegisterClass);
7538 case X86::ATOMXOR16:
7539 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7540 X86::XOR16ri, X86::MOV16rm,
7541 X86::LCMPXCHG16, X86::MOV16rr,
7542 X86::NOT16r, X86::AX,
7543 X86::GR16RegisterClass);
7544 case X86::ATOMNAND16:
7545 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7546 X86::AND16ri, X86::MOV16rm,
7547 X86::LCMPXCHG16, X86::MOV16rr,
7548 X86::NOT16r, X86::AX,
7549 X86::GR16RegisterClass, true);
7550 case X86::ATOMMIN16:
7551 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7552 case X86::ATOMMAX16:
7553 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7554 case X86::ATOMUMIN16:
7555 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7556 case X86::ATOMUMAX16:
7557 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7560 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7561 X86::AND8ri, X86::MOV8rm,
7562 X86::LCMPXCHG8, X86::MOV8rr,
7563 X86::NOT8r, X86::AL,
7564 X86::GR8RegisterClass);
7566 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7567 X86::OR8ri, X86::MOV8rm,
7568 X86::LCMPXCHG8, X86::MOV8rr,
7569 X86::NOT8r, X86::AL,
7570 X86::GR8RegisterClass);
7572 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7573 X86::XOR8ri, X86::MOV8rm,
7574 X86::LCMPXCHG8, X86::MOV8rr,
7575 X86::NOT8r, X86::AL,
7576 X86::GR8RegisterClass);
7577 case X86::ATOMNAND8:
7578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7579 X86::AND8ri, X86::MOV8rm,
7580 X86::LCMPXCHG8, X86::MOV8rr,
7581 X86::NOT8r, X86::AL,
7582 X86::GR8RegisterClass, true);
7583 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7584 // This group is for 64-bit host.
7585 case X86::ATOMAND64:
7586 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7587 X86::AND64ri32, X86::MOV64rm,
7588 X86::LCMPXCHG64, X86::MOV64rr,
7589 X86::NOT64r, X86::RAX,
7590 X86::GR64RegisterClass);
7592 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7593 X86::OR64ri32, X86::MOV64rm,
7594 X86::LCMPXCHG64, X86::MOV64rr,
7595 X86::NOT64r, X86::RAX,
7596 X86::GR64RegisterClass);
7597 case X86::ATOMXOR64:
7598 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7599 X86::XOR64ri32, X86::MOV64rm,
7600 X86::LCMPXCHG64, X86::MOV64rr,
7601 X86::NOT64r, X86::RAX,
7602 X86::GR64RegisterClass);
7603 case X86::ATOMNAND64:
7604 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7605 X86::AND64ri32, X86::MOV64rm,
7606 X86::LCMPXCHG64, X86::MOV64rr,
7607 X86::NOT64r, X86::RAX,
7608 X86::GR64RegisterClass, true);
7609 case X86::ATOMMIN64:
7610 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7611 case X86::ATOMMAX64:
7612 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7613 case X86::ATOMUMIN64:
7614 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7615 case X86::ATOMUMAX64:
7616 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7618 // This group does 64-bit operations on a 32-bit host.
7619 case X86::ATOMAND6432:
7620 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7621 X86::AND32rr, X86::AND32rr,
7622 X86::AND32ri, X86::AND32ri,
7624 case X86::ATOMOR6432:
7625 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7626 X86::OR32rr, X86::OR32rr,
7627 X86::OR32ri, X86::OR32ri,
7629 case X86::ATOMXOR6432:
7630 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7631 X86::XOR32rr, X86::XOR32rr,
7632 X86::XOR32ri, X86::XOR32ri,
7634 case X86::ATOMNAND6432:
7635 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7636 X86::AND32rr, X86::AND32rr,
7637 X86::AND32ri, X86::AND32ri,
7639 case X86::ATOMADD6432:
7640 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7641 X86::ADD32rr, X86::ADC32rr,
7642 X86::ADD32ri, X86::ADC32ri,
7644 case X86::ATOMSUB6432:
7645 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7646 X86::SUB32rr, X86::SBB32rr,
7647 X86::SUB32ri, X86::SBB32ri,
7649 case X86::ATOMSWAP6432:
7650 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7651 X86::MOV32rr, X86::MOV32rr,
7652 X86::MOV32ri, X86::MOV32ri,
7657 //===----------------------------------------------------------------------===//
7658 // X86 Optimization Hooks
7659 //===----------------------------------------------------------------------===//
7661 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7665 const SelectionDAG &DAG,
7666 unsigned Depth) const {
7667 unsigned Opc = Op.getOpcode();
7668 assert((Opc >= ISD::BUILTIN_OP_END ||
7669 Opc == ISD::INTRINSIC_WO_CHAIN ||
7670 Opc == ISD::INTRINSIC_W_CHAIN ||
7671 Opc == ISD::INTRINSIC_VOID) &&
7672 "Should use MaskedValueIsZero if you don't know whether Op"
7673 " is a target node!");
7675 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7682 // These nodes' second result is a boolean.
7683 if (Op.getResNo() == 0)
7687 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7688 Mask.getBitWidth() - 1);
7693 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7694 /// node is a GlobalAddress + offset.
7695 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7696 GlobalValue* &GA, int64_t &Offset) const{
7697 if (N->getOpcode() == X86ISD::Wrapper) {
7698 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7699 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7700 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7704 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7707 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7708 const TargetLowering &TLI) {
7711 if (TLI.isGAPlusOffset(Base, GV, Offset))
7712 return (GV->getAlignment() >= N && (Offset % N) == 0);
7713 // DAG combine handles the stack object case.
7717 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
7718 unsigned NumElems, MVT EVT,
7720 SelectionDAG &DAG, MachineFrameInfo *MFI,
7721 const TargetLowering &TLI) {
7723 for (unsigned i = 0; i < NumElems; ++i) {
7724 SDValue Idx = PermMask.getOperand(i);
7725 if (Idx.getOpcode() == ISD::UNDEF) {
7731 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7732 if (!Elt.getNode() ||
7733 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7736 Base = Elt.getNode();
7737 if (Base->getOpcode() == ISD::UNDEF)
7741 if (Elt.getOpcode() == ISD::UNDEF)
7744 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7745 EVT.getSizeInBits()/8, i, MFI))
7751 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7752 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7753 /// if the load addresses are consecutive, non-overlapping, and in the right
7755 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7756 const TargetLowering &TLI) {
7757 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7758 DebugLoc dl = N->getDebugLoc();
7759 MVT VT = N->getValueType(0);
7760 MVT EVT = VT.getVectorElementType();
7761 SDValue PermMask = N->getOperand(2);
7762 unsigned NumElems = PermMask.getNumOperands();
7763 SDNode *Base = NULL;
7764 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
7768 LoadSDNode *LD = cast<LoadSDNode>(Base);
7769 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7770 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7771 LD->getSrcValue(), LD->getSrcValueOffset(),
7773 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7774 LD->getSrcValue(), LD->getSrcValueOffset(),
7775 LD->isVolatile(), LD->getAlignment());
7778 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7779 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7780 TargetLowering::DAGCombinerInfo &DCI,
7781 const X86Subtarget *Subtarget,
7782 const TargetLowering &TLI) {
7783 unsigned NumOps = N->getNumOperands();
7784 DebugLoc dl = N->getDebugLoc();
7786 // Ignore single operand BUILD_VECTOR.
7790 MVT VT = N->getValueType(0);
7791 MVT EVT = VT.getVectorElementType();
7792 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7793 // We are looking for load i64 and zero extend. We want to transform
7794 // it before legalizer has a chance to expand it. Also look for i64
7795 // BUILD_PAIR bit casted to f64.
7797 // This must be an insertion into a zero vector.
7798 SDValue HighElt = N->getOperand(1);
7799 if (!isZeroNode(HighElt))
7802 // Value must be a load.
7803 SDNode *Base = N->getOperand(0).getNode();
7804 if (!isa<LoadSDNode>(Base)) {
7805 if (Base->getOpcode() != ISD::BIT_CONVERT)
7807 Base = Base->getOperand(0).getNode();
7808 if (!isa<LoadSDNode>(Base))
7812 // Transform it into VZEXT_LOAD addr.
7813 LoadSDNode *LD = cast<LoadSDNode>(Base);
7815 // Load must not be an extload.
7816 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7819 // Load type should legal type so we don't have to legalize it.
7820 if (!TLI.isTypeLegal(VT))
7823 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7824 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7825 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7826 TargetLowering::TargetLoweringOpt TLO(DAG);
7827 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
7828 DCI.CommitTargetLoweringOpt(TLO);
7832 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7833 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7834 const X86Subtarget *Subtarget) {
7835 DebugLoc dl = N->getDebugLoc();
7836 SDValue Cond = N->getOperand(0);
7838 // If we have SSE[12] support, try to form min/max nodes.
7839 if (Subtarget->hasSSE2() &&
7840 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
7841 if (Cond.getOpcode() == ISD::SETCC) {
7842 // Get the LHS/RHS of the select.
7843 SDValue LHS = N->getOperand(1);
7844 SDValue RHS = N->getOperand(2);
7845 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7847 unsigned Opcode = 0;
7848 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7851 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7854 if (!UnsafeFPMath) break;
7856 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7858 Opcode = X86ISD::FMIN;
7861 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7864 if (!UnsafeFPMath) break;
7866 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7868 Opcode = X86ISD::FMAX;
7871 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7874 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7877 if (!UnsafeFPMath) break;
7879 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7881 Opcode = X86ISD::FMIN;
7884 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7887 if (!UnsafeFPMath) break;
7889 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7891 Opcode = X86ISD::FMAX;
7897 return DAG.getNode(Opcode, dl, N->getValueType(0), LHS, RHS);
7905 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
7907 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
7908 const X86Subtarget *Subtarget) {
7909 // On X86 with SSE2 support, we can transform this to a vector shift if
7910 // all elements are shifted by the same amount. We can't do this in legalize
7911 // because the a constant vector is typically transformed to a constant pool
7912 // so we have no knowledge of the shift amount.
7913 if (!Subtarget->hasSSE2())
7916 MVT VT = N->getValueType(0);
7917 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
7920 SDValue ShAmtOp = N->getOperand(1);
7921 MVT EltVT = VT.getVectorElementType();
7922 DebugLoc dl = N->getDebugLoc();
7924 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
7925 unsigned NumElts = VT.getVectorNumElements();
7927 for (; i != NumElts; ++i) {
7928 SDValue Arg = ShAmtOp.getOperand(i);
7929 if (Arg.getOpcode() == ISD::UNDEF) continue;
7933 for (; i != NumElts; ++i) {
7934 SDValue Arg = ShAmtOp.getOperand(i);
7935 if (Arg.getOpcode() == ISD::UNDEF) continue;
7936 if (Arg != BaseShAmt) {
7940 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
7941 isSplatMask(ShAmtOp.getOperand(2).getNode())) {
7942 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ShAmtOp,
7943 DAG.getIntPtrConstant(0));
7947 if (EltVT.bitsGT(MVT::i32))
7948 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
7949 else if (EltVT.bitsLT(MVT::i32))
7950 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BaseShAmt);
7952 // The shift amount is identical so we can do a vector shift.
7953 SDValue ValOp = N->getOperand(0);
7954 switch (N->getOpcode()) {
7956 assert(0 && "Unknown shift opcode!");
7959 if (VT == MVT::v2i64)
7960 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7961 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7963 if (VT == MVT::v4i32)
7964 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7965 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7967 if (VT == MVT::v8i16)
7968 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7969 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
7973 if (VT == MVT::v4i32)
7974 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7975 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
7977 if (VT == MVT::v8i16)
7978 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7979 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
7983 if (VT == MVT::v2i64)
7984 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7985 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7987 if (VT == MVT::v4i32)
7988 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7989 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
7991 if (VT == MVT::v8i16)
7992 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7993 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8000 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8001 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8002 const X86Subtarget *Subtarget) {
8003 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8004 // the FP state in cases where an emms may be missing.
8005 // A preferable solution to the general problem is to figure out the right
8006 // places to insert EMMS. This qualifies as a quick hack.
8007 StoreSDNode *St = cast<StoreSDNode>(N);
8008 if (St->getValue().getValueType().isVector() &&
8009 St->getValue().getValueType().getSizeInBits() == 64 &&
8010 isa<LoadSDNode>(St->getValue()) &&
8011 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8012 St->getChain().hasOneUse() && !St->isVolatile()) {
8013 SDNode* LdVal = St->getValue().getNode();
8015 int TokenFactorIndex = -1;
8016 SmallVector<SDValue, 8> Ops;
8017 SDNode* ChainVal = St->getChain().getNode();
8018 // Must be a store of a load. We currently handle two cases: the load
8019 // is a direct child, and it's under an intervening TokenFactor. It is
8020 // possible to dig deeper under nested TokenFactors.
8021 if (ChainVal == LdVal)
8022 Ld = cast<LoadSDNode>(St->getChain());
8023 else if (St->getValue().hasOneUse() &&
8024 ChainVal->getOpcode() == ISD::TokenFactor) {
8025 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8026 if (ChainVal->getOperand(i).getNode() == LdVal) {
8027 TokenFactorIndex = i;
8028 Ld = cast<LoadSDNode>(St->getValue());
8030 Ops.push_back(ChainVal->getOperand(i));
8034 DebugLoc dl = N->getDebugLoc();
8035 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8036 if (Subtarget->is64Bit()) {
8037 SDValue NewLd = DAG.getLoad(MVT::i64, dl, Ld->getChain(),
8038 Ld->getBasePtr(), Ld->getSrcValue(),
8039 Ld->getSrcValueOffset(), Ld->isVolatile(),
8040 Ld->getAlignment());
8041 SDValue NewChain = NewLd.getValue(1);
8042 if (TokenFactorIndex != -1) {
8043 Ops.push_back(NewChain);
8044 NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Ops[0],
8047 return DAG.getStore(NewChain, dl, NewLd, St->getBasePtr(),
8048 St->getSrcValue(), St->getSrcValueOffset(),
8049 St->isVolatile(), St->getAlignment());
8052 // Otherwise, lower to two 32-bit copies.
8053 SDValue LoAddr = Ld->getBasePtr();
8054 SDValue HiAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, LoAddr,
8055 DAG.getConstant(4, MVT::i32));
8057 SDValue LoLd = DAG.getLoad(MVT::i32, dl, Ld->getChain(), LoAddr,
8058 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8059 Ld->isVolatile(), Ld->getAlignment());
8060 SDValue HiLd = DAG.getLoad(MVT::i32, dl, Ld->getChain(), HiAddr,
8061 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8063 MinAlign(Ld->getAlignment(), 4));
8065 SDValue NewChain = LoLd.getValue(1);
8066 if (TokenFactorIndex != -1) {
8067 Ops.push_back(LoLd);
8068 Ops.push_back(HiLd);
8069 NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Ops[0],
8073 LoAddr = St->getBasePtr();
8074 HiAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, LoAddr,
8075 DAG.getConstant(4, MVT::i32));
8077 SDValue LoSt = DAG.getStore(NewChain, dl, LoLd, LoAddr,
8078 St->getSrcValue(), St->getSrcValueOffset(),
8079 St->isVolatile(), St->getAlignment());
8080 SDValue HiSt = DAG.getStore(NewChain, dl, HiLd, HiAddr,
8082 St->getSrcValueOffset() + 4,
8084 MinAlign(St->getAlignment(), 4));
8085 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoSt, HiSt);
8091 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8092 /// X86ISD::FXOR nodes.
8093 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8094 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8095 // F[X]OR(0.0, x) -> x
8096 // F[X]OR(x, 0.0) -> x
8097 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8098 if (C->getValueAPF().isPosZero())
8099 return N->getOperand(1);
8100 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8101 if (C->getValueAPF().isPosZero())
8102 return N->getOperand(0);
8106 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8107 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8108 // FAND(0.0, x) -> 0.0
8109 // FAND(x, 0.0) -> 0.0
8110 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8111 if (C->getValueAPF().isPosZero())
8112 return N->getOperand(0);
8113 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8114 if (C->getValueAPF().isPosZero())
8115 return N->getOperand(1);
8119 static SDValue PerformBTCombine(SDNode *N,
8121 TargetLowering::DAGCombinerInfo &DCI) {
8122 // BT ignores high bits in the bit index operand.
8123 SDValue Op1 = N->getOperand(1);
8124 if (Op1.hasOneUse()) {
8125 unsigned BitWidth = Op1.getValueSizeInBits();
8126 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8127 APInt KnownZero, KnownOne;
8128 TargetLowering::TargetLoweringOpt TLO(DAG);
8129 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8130 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8131 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8132 DCI.CommitTargetLoweringOpt(TLO);
8137 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8138 DAGCombinerInfo &DCI) const {
8139 SelectionDAG &DAG = DCI.DAG;
8140 switch (N->getOpcode()) {
8142 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8143 case ISD::BUILD_VECTOR:
8144 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8145 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8148 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8149 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8151 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8152 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8153 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8159 //===----------------------------------------------------------------------===//
8160 // X86 Inline Assembly Support
8161 //===----------------------------------------------------------------------===//
8163 /// getConstraintType - Given a constraint letter, return the type of
8164 /// constraint it is for this target.
8165 X86TargetLowering::ConstraintType
8166 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8167 if (Constraint.size() == 1) {
8168 switch (Constraint[0]) {
8180 return C_RegisterClass;
8188 return TargetLowering::getConstraintType(Constraint);
8191 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8192 /// with another that has more specific requirements based on the type of the
8193 /// corresponding operand.
8194 const char *X86TargetLowering::
8195 LowerXConstraint(MVT ConstraintVT) const {
8196 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8197 // 'f' like normal targets.
8198 if (ConstraintVT.isFloatingPoint()) {
8199 if (Subtarget->hasSSE2())
8201 if (Subtarget->hasSSE1())
8205 return TargetLowering::LowerXConstraint(ConstraintVT);
8208 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8209 /// vector. If it is invalid, don't add anything to Ops.
8210 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8213 std::vector<SDValue>&Ops,
8214 SelectionDAG &DAG) const {
8215 SDValue Result(0, 0);
8217 switch (Constraint) {
8220 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8221 if (C->getZExtValue() <= 31) {
8222 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8228 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8229 if (C->getZExtValue() <= 63) {
8230 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8237 if (C->getZExtValue() <= 255) {
8238 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8244 // 32-bit signed value
8245 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8246 const ConstantInt *CI = C->getConstantIntValue();
8247 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8248 // Widen to 64 bits here to get it sign extended.
8249 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8252 // FIXME gcc accepts some relocatable values here too, but only in certain
8253 // memory models; it's complicated.
8258 // 32-bit unsigned value
8259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8260 const ConstantInt *CI = C->getConstantIntValue();
8261 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8262 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8266 // FIXME gcc accepts some relocatable values here too, but only in certain
8267 // memory models; it's complicated.
8271 // Literal immediates are always ok.
8272 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8273 // Widen to 64 bits here to get it sign extended.
8274 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8278 // If we are in non-pic codegen mode, we allow the address of a global (with
8279 // an optional displacement) to be used with 'i'.
8280 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8283 // Match either (GA) or (GA+C)
8285 Offset = GA->getOffset();
8286 } else if (Op.getOpcode() == ISD::ADD) {
8287 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8288 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8290 Offset = GA->getOffset()+C->getZExtValue();
8292 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8293 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8295 Offset = GA->getOffset()+C->getZExtValue();
8303 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8306 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8312 // Otherwise, not valid for this mode.
8317 if (Result.getNode()) {
8318 Ops.push_back(Result);
8321 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8325 std::vector<unsigned> X86TargetLowering::
8326 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8328 if (Constraint.size() == 1) {
8329 // FIXME: not handling fp-stack yet!
8330 switch (Constraint[0]) { // GCC X86 Constraint Letters
8331 default: break; // Unknown constraint letter
8332 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8335 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8336 else if (VT == MVT::i16)
8337 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8338 else if (VT == MVT::i8)
8339 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8340 else if (VT == MVT::i64)
8341 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8346 return std::vector<unsigned>();
8349 std::pair<unsigned, const TargetRegisterClass*>
8350 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8352 // First, see if this is a constraint that directly corresponds to an LLVM
8354 if (Constraint.size() == 1) {
8355 // GCC Constraint Letters
8356 switch (Constraint[0]) {
8358 case 'r': // GENERAL_REGS
8359 case 'R': // LEGACY_REGS
8360 case 'l': // INDEX_REGS
8362 return std::make_pair(0U, X86::GR8RegisterClass);
8364 return std::make_pair(0U, X86::GR16RegisterClass);
8365 if (VT == MVT::i32 || !Subtarget->is64Bit())
8366 return std::make_pair(0U, X86::GR32RegisterClass);
8367 return std::make_pair(0U, X86::GR64RegisterClass);
8368 case 'f': // FP Stack registers.
8369 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8370 // value to the correct fpstack register class.
8371 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8372 return std::make_pair(0U, X86::RFP32RegisterClass);
8373 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8374 return std::make_pair(0U, X86::RFP64RegisterClass);
8375 return std::make_pair(0U, X86::RFP80RegisterClass);
8376 case 'y': // MMX_REGS if MMX allowed.
8377 if (!Subtarget->hasMMX()) break;
8378 return std::make_pair(0U, X86::VR64RegisterClass);
8379 case 'Y': // SSE_REGS if SSE2 allowed
8380 if (!Subtarget->hasSSE2()) break;
8382 case 'x': // SSE_REGS if SSE1 allowed
8383 if (!Subtarget->hasSSE1()) break;
8385 switch (VT.getSimpleVT()) {
8387 // Scalar SSE types.
8390 return std::make_pair(0U, X86::FR32RegisterClass);
8393 return std::make_pair(0U, X86::FR64RegisterClass);
8401 return std::make_pair(0U, X86::VR128RegisterClass);
8407 // Use the default implementation in TargetLowering to convert the register
8408 // constraint into a member of a register class.
8409 std::pair<unsigned, const TargetRegisterClass*> Res;
8410 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8412 // Not found as a standard register?
8413 if (Res.second == 0) {
8414 // GCC calls "st(0)" just plain "st".
8415 if (StringsEqualNoCase("{st}", Constraint)) {
8416 Res.first = X86::ST0;
8417 Res.second = X86::RFP80RegisterClass;
8419 // 'A' means EAX + EDX.
8420 if (Constraint == "A") {
8421 Res.first = X86::EAX;
8422 Res.second = X86::GRADRegisterClass;
8427 // Otherwise, check to see if this is a register class of the wrong value
8428 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8429 // turn into {ax},{dx}.
8430 if (Res.second->hasType(VT))
8431 return Res; // Correct type already, nothing to do.
8433 // All of the single-register GCC register classes map their values onto
8434 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8435 // really want an 8-bit or 32-bit register, map to the appropriate register
8436 // class and return the appropriate register.
8437 if (Res.second == X86::GR16RegisterClass) {
8438 if (VT == MVT::i8) {
8439 unsigned DestReg = 0;
8440 switch (Res.first) {
8442 case X86::AX: DestReg = X86::AL; break;
8443 case X86::DX: DestReg = X86::DL; break;
8444 case X86::CX: DestReg = X86::CL; break;
8445 case X86::BX: DestReg = X86::BL; break;
8448 Res.first = DestReg;
8449 Res.second = Res.second = X86::GR8RegisterClass;
8451 } else if (VT == MVT::i32) {
8452 unsigned DestReg = 0;
8453 switch (Res.first) {
8455 case X86::AX: DestReg = X86::EAX; break;
8456 case X86::DX: DestReg = X86::EDX; break;
8457 case X86::CX: DestReg = X86::ECX; break;
8458 case X86::BX: DestReg = X86::EBX; break;
8459 case X86::SI: DestReg = X86::ESI; break;
8460 case X86::DI: DestReg = X86::EDI; break;
8461 case X86::BP: DestReg = X86::EBP; break;
8462 case X86::SP: DestReg = X86::ESP; break;
8465 Res.first = DestReg;
8466 Res.second = Res.second = X86::GR32RegisterClass;
8468 } else if (VT == MVT::i64) {
8469 unsigned DestReg = 0;
8470 switch (Res.first) {
8472 case X86::AX: DestReg = X86::RAX; break;
8473 case X86::DX: DestReg = X86::RDX; break;
8474 case X86::CX: DestReg = X86::RCX; break;
8475 case X86::BX: DestReg = X86::RBX; break;
8476 case X86::SI: DestReg = X86::RSI; break;
8477 case X86::DI: DestReg = X86::RDI; break;
8478 case X86::BP: DestReg = X86::RBP; break;
8479 case X86::SP: DestReg = X86::RSP; break;
8482 Res.first = DestReg;
8483 Res.second = Res.second = X86::GR64RegisterClass;
8486 } else if (Res.second == X86::FR32RegisterClass ||
8487 Res.second == X86::FR64RegisterClass ||
8488 Res.second == X86::VR128RegisterClass) {
8489 // Handle references to XMM physical registers that got mapped into the
8490 // wrong class. This can happen with constraints like {xmm0} where the
8491 // target independent register mapper will just pick the first match it can
8492 // find, ignoring the required type.
8494 Res.second = X86::FR32RegisterClass;
8495 else if (VT == MVT::f64)
8496 Res.second = X86::FR64RegisterClass;
8497 else if (X86::VR128RegisterClass->hasType(VT))
8498 Res.second = X86::VR128RegisterClass;
8504 //===----------------------------------------------------------------------===//
8505 // X86 Widen vector type
8506 //===----------------------------------------------------------------------===//
8508 /// getWidenVectorType: given a vector type, returns the type to widen
8509 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8510 /// If there is no vector type that we want to widen to, returns MVT::Other
8511 /// When and where to widen is target dependent based on the cost of
8512 /// scalarizing vs using the wider vector type.
8514 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8515 assert(VT.isVector());
8516 if (isTypeLegal(VT))
8519 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8520 // type based on element type. This would speed up our search (though
8521 // it may not be worth it since the size of the list is relatively
8523 MVT EltVT = VT.getVectorElementType();
8524 unsigned NElts = VT.getVectorNumElements();
8526 // On X86, it make sense to widen any vector wider than 1
8530 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8531 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8532 MVT SVT = (MVT::SimpleValueType)nVT;
8534 if (isTypeLegal(SVT) &&
8535 SVT.getVectorElementType() == EltVT &&
8536 SVT.getVectorNumElements() > NElts)