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 "X86FastISel.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/ADT/BitVector.h"
28 #include "llvm/ADT/VectorExtras.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/SmallSet.h"
41 #include "llvm/ADT/StringExtras.h"
44 // Forward declarations.
45 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
47 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
48 : TargetLowering(TM) {
49 Subtarget = &TM.getSubtarget<X86Subtarget>();
50 X86ScalarSSEf64 = Subtarget->hasSSE2();
51 X86ScalarSSEf32 = Subtarget->hasSSE1();
52 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
56 RegInfo = TM.getRegisterInfo();
58 // Set up the TargetLowering object.
60 // X86 is weird, it always uses i8 for shift amounts and setcc results.
61 setShiftAmountType(MVT::i8);
62 setSetCCResultContents(ZeroOrOneSetCCResult);
63 setSchedulingPreference(SchedulingForRegPressure);
64 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
65 setStackPointerRegisterToSaveRestore(X86StackPtr);
67 if (Subtarget->isTargetDarwin()) {
68 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
69 setUseUnderscoreSetJmp(false);
70 setUseUnderscoreLongJmp(false);
71 } else if (Subtarget->isTargetMingw()) {
72 // MS runtime is weird: it exports _setjmp, but longjmp!
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(false);
76 setUseUnderscoreSetJmp(true);
77 setUseUnderscoreLongJmp(true);
80 // Set up the register classes.
81 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
82 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
83 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
84 if (Subtarget->is64Bit())
85 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
87 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 // We don't accept any truncstore of integer registers.
90 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
92 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
93 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
94 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
95 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
97 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
99 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
101 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
103 if (Subtarget->is64Bit()) {
104 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
105 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
108 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
109 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
111 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
114 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
116 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
117 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
118 // SSE has no i16 to fp conversion, only i32
119 if (X86ScalarSSEf32) {
120 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
121 // f32 and f64 cases are Legal, f80 case is not
122 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
124 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
125 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
128 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
129 // are Legal, f80 is custom lowered.
130 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
131 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
133 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
135 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
136 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
138 if (X86ScalarSSEf32) {
139 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
140 // f32 and f64 cases are Legal, f80 case is not
141 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
143 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
144 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
147 // Handle FP_TO_UINT by promoting the destination to a larger signed
149 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
150 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
151 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
153 if (Subtarget->is64Bit()) {
154 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
155 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
158 // Expand FP_TO_UINT into a select.
159 // FIXME: We would like to use a Custom expander here eventually to do
160 // the optimal thing for SSE vs. the default expansion in the legalizer.
161 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
163 // With SSE3 we can use fisttpll to convert to a signed i64.
164 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
167 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
168 if (!X86ScalarSSEf64) {
169 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
170 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
173 // Scalar integer divide and remainder are lowered to use operations that
174 // produce two results, to match the available instructions. This exposes
175 // the two-result form to trivial CSE, which is able to combine x/y and x%y
176 // into a single instruction.
178 // Scalar integer multiply-high is also lowered to use two-result
179 // operations, to match the available instructions. However, plain multiply
180 // (low) operations are left as Legal, as there are single-result
181 // instructions for this in x86. Using the two-result multiply instructions
182 // when both high and low results are needed must be arranged by dagcombine.
183 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
184 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
185 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
186 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
187 setOperationAction(ISD::SREM , MVT::i8 , Expand);
188 setOperationAction(ISD::UREM , MVT::i8 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
193 setOperationAction(ISD::SREM , MVT::i16 , Expand);
194 setOperationAction(ISD::UREM , MVT::i16 , Expand);
195 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
196 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
197 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
198 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
199 setOperationAction(ISD::SREM , MVT::i32 , Expand);
200 setOperationAction(ISD::UREM , MVT::i32 , Expand);
201 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
202 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
203 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
205 setOperationAction(ISD::SREM , MVT::i64 , Expand);
206 setOperationAction(ISD::UREM , MVT::i64 , Expand);
208 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
209 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
210 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
211 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
212 if (Subtarget->is64Bit())
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
216 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
217 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
218 setOperationAction(ISD::FREM , MVT::f32 , Expand);
219 setOperationAction(ISD::FREM , MVT::f64 , Expand);
220 setOperationAction(ISD::FREM , MVT::f80 , Expand);
221 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
223 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
224 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
225 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
226 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
227 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
228 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
229 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
230 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
231 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
232 if (Subtarget->is64Bit()) {
233 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
234 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
235 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
238 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
239 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
241 // These should be promoted to a larger select which is supported.
242 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
243 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
244 // X86 wants to expand cmov itself.
245 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
246 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
248 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
249 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
251 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
252 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
254 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
255 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
258 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
260 // X86 ret instruction may pop stack.
261 setOperationAction(ISD::RET , MVT::Other, Custom);
262 if (!Subtarget->is64Bit())
263 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
266 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
267 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
268 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
269 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
270 if (Subtarget->is64Bit())
271 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
272 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
273 if (Subtarget->is64Bit()) {
274 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
275 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
276 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
277 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
279 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
280 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
281 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
282 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
283 if (Subtarget->is64Bit()) {
284 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
285 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
286 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
289 if (Subtarget->hasSSE1())
290 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
292 if (!Subtarget->hasSSE2())
293 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
295 // Expand certain atomics
296 setOperationAction(ISD::ATOMIC_CMP_SWAP_8 , MVT::i8, Custom);
297 setOperationAction(ISD::ATOMIC_CMP_SWAP_16, MVT::i16, Custom);
298 setOperationAction(ISD::ATOMIC_CMP_SWAP_32, MVT::i32, Custom);
299 setOperationAction(ISD::ATOMIC_CMP_SWAP_64, MVT::i64, Custom);
301 setOperationAction(ISD::ATOMIC_LOAD_SUB_8, MVT::i8, Expand);
302 setOperationAction(ISD::ATOMIC_LOAD_SUB_16, MVT::i16, Expand);
303 setOperationAction(ISD::ATOMIC_LOAD_SUB_32, MVT::i32, Expand);
304 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Expand);
306 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
307 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
308 // FIXME - use subtarget debug flags
309 if (!Subtarget->isTargetDarwin() &&
310 !Subtarget->isTargetELF() &&
311 !Subtarget->isTargetCygMing()) {
312 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
313 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
316 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
317 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
318 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
319 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
320 if (Subtarget->is64Bit()) {
322 setExceptionPointerRegister(X86::RAX);
323 setExceptionSelectorRegister(X86::RDX);
325 setExceptionPointerRegister(X86::EAX);
326 setExceptionSelectorRegister(X86::EDX);
328 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
330 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
332 setOperationAction(ISD::TRAP, MVT::Other, Legal);
334 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
335 setOperationAction(ISD::VASTART , MVT::Other, Custom);
336 setOperationAction(ISD::VAEND , MVT::Other, Expand);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::VAARG , MVT::Other, Custom);
339 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
341 setOperationAction(ISD::VAARG , MVT::Other, Expand);
342 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
345 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
346 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
347 if (Subtarget->is64Bit())
348 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
349 if (Subtarget->isTargetCygMing())
350 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
352 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
354 if (X86ScalarSSEf64) {
355 // f32 and f64 use SSE.
356 // Set up the FP register classes.
357 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
358 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
360 // Use ANDPD to simulate FABS.
361 setOperationAction(ISD::FABS , MVT::f64, Custom);
362 setOperationAction(ISD::FABS , MVT::f32, Custom);
364 // Use XORP to simulate FNEG.
365 setOperationAction(ISD::FNEG , MVT::f64, Custom);
366 setOperationAction(ISD::FNEG , MVT::f32, Custom);
368 // Use ANDPD and ORPD to simulate FCOPYSIGN.
369 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
370 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
372 // We don't support sin/cos/fmod
373 setOperationAction(ISD::FSIN , MVT::f64, Expand);
374 setOperationAction(ISD::FCOS , MVT::f64, Expand);
375 setOperationAction(ISD::FSIN , MVT::f32, Expand);
376 setOperationAction(ISD::FCOS , MVT::f32, Expand);
378 // Expand FP immediates into loads from the stack, except for the special
380 addLegalFPImmediate(APFloat(+0.0)); // xorpd
381 addLegalFPImmediate(APFloat(+0.0f)); // xorps
383 // Floating truncations from f80 and extensions to f80 go through memory.
384 // If optimizing, we lie about this though and handle it in
385 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
387 setConvertAction(MVT::f32, MVT::f80, Expand);
388 setConvertAction(MVT::f64, MVT::f80, Expand);
389 setConvertAction(MVT::f80, MVT::f32, Expand);
390 setConvertAction(MVT::f80, MVT::f64, Expand);
392 } else if (X86ScalarSSEf32) {
393 // Use SSE for f32, x87 for f64.
394 // Set up the FP register classes.
395 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
396 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
398 // Use ANDPS to simulate FABS.
399 setOperationAction(ISD::FABS , MVT::f32, Custom);
401 // Use XORP to simulate FNEG.
402 setOperationAction(ISD::FNEG , MVT::f32, Custom);
404 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
406 // Use ANDPS and ORPS to simulate FCOPYSIGN.
407 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
408 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
410 // We don't support sin/cos/fmod
411 setOperationAction(ISD::FSIN , MVT::f32, Expand);
412 setOperationAction(ISD::FCOS , MVT::f32, Expand);
414 // Special cases we handle for FP constants.
415 addLegalFPImmediate(APFloat(+0.0f)); // xorps
416 addLegalFPImmediate(APFloat(+0.0)); // FLD0
417 addLegalFPImmediate(APFloat(+1.0)); // FLD1
418 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
419 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
421 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
422 // this though and handle it in InstructionSelectPreprocess so that
423 // dagcombine2 can hack on these.
425 setConvertAction(MVT::f32, MVT::f64, Expand);
426 setConvertAction(MVT::f32, MVT::f80, Expand);
427 setConvertAction(MVT::f80, MVT::f32, Expand);
428 setConvertAction(MVT::f64, MVT::f32, Expand);
429 // And x87->x87 truncations also.
430 setConvertAction(MVT::f80, MVT::f64, Expand);
434 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
435 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
438 // f32 and f64 in x87.
439 // Set up the FP register classes.
440 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
441 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
443 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
444 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
446 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
448 // Floating truncations go through memory. If optimizing, we lie about
449 // this though and handle it in InstructionSelectPreprocess so that
450 // dagcombine2 can hack on these.
452 setConvertAction(MVT::f80, MVT::f32, Expand);
453 setConvertAction(MVT::f64, MVT::f32, Expand);
454 setConvertAction(MVT::f80, MVT::f64, Expand);
458 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
459 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
461 addLegalFPImmediate(APFloat(+0.0)); // FLD0
462 addLegalFPImmediate(APFloat(+1.0)); // FLD1
463 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
464 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
465 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
466 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
467 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
468 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
471 // Long double always uses X87.
472 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
473 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
474 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
476 APFloat TmpFlt(+0.0);
477 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
478 addLegalFPImmediate(TmpFlt); // FLD0
480 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
481 APFloat TmpFlt2(+1.0);
482 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
483 addLegalFPImmediate(TmpFlt2); // FLD1
484 TmpFlt2.changeSign();
485 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
489 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
490 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
493 // Always use a library call for pow.
494 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
495 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
496 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
498 // First set operation action for all vector types to expand. Then we
499 // will selectively turn on ones that can be effectively codegen'd.
500 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
501 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
502 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
503 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
504 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
505 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
506 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
507 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
508 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
509 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
510 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
511 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
512 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
513 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
514 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
515 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
516 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
517 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
542 if (Subtarget->hasMMX()) {
543 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
544 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
545 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
546 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
547 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
549 // FIXME: add MMX packed arithmetics
551 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
552 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
553 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
554 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
556 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
557 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
558 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
559 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
561 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
562 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
564 setOperationAction(ISD::AND, MVT::v8i8, Promote);
565 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
566 setOperationAction(ISD::AND, MVT::v4i16, Promote);
567 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
568 setOperationAction(ISD::AND, MVT::v2i32, Promote);
569 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
570 setOperationAction(ISD::AND, MVT::v1i64, Legal);
572 setOperationAction(ISD::OR, MVT::v8i8, Promote);
573 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
574 setOperationAction(ISD::OR, MVT::v4i16, Promote);
575 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
576 setOperationAction(ISD::OR, MVT::v2i32, Promote);
577 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
578 setOperationAction(ISD::OR, MVT::v1i64, Legal);
580 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
581 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
582 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
583 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
584 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
585 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
586 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
588 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
589 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
590 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
591 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
592 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
593 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
594 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
595 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
596 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
598 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
599 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
600 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
601 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
602 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
604 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
605 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
606 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
607 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
609 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
610 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
611 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
612 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
614 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
617 if (Subtarget->hasSSE1()) {
618 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
620 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
621 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
622 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
623 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
624 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
625 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
626 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
627 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
628 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
629 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
630 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
631 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
634 if (Subtarget->hasSSE2()) {
635 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
636 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
637 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
638 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
639 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
641 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
642 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
643 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
644 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
645 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
646 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
647 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
648 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
649 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
650 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
651 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
652 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
653 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
654 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
655 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
657 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
658 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
659 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
660 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
662 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
663 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
664 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
665 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
666 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
668 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
669 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
670 MVT VT = (MVT::SimpleValueType)i;
671 // Do not attempt to custom lower non-power-of-2 vectors
672 if (!isPowerOf2_32(VT.getVectorNumElements()))
674 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
676 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
680 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
681 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
682 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
683 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
684 if (Subtarget->is64Bit()) {
685 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
686 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
689 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
690 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
691 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
692 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
693 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
694 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
695 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
696 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
697 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
698 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
699 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
700 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
703 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
705 // Custom lower v2i64 and v2f64 selects.
706 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
707 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
708 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
709 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
713 if (Subtarget->hasSSE41()) {
714 // FIXME: Do we need to handle scalar-to-vector here?
715 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
716 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
718 // i8 and i16 vectors are custom , because the source register and source
719 // source memory operand types are not the same width. f32 vectors are
720 // custom since the immediate controlling the insert encodes additional
722 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
723 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
724 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
727 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
728 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
729 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
732 if (Subtarget->is64Bit()) {
733 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
734 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
738 if (Subtarget->hasSSE42()) {
739 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
742 // We want to custom lower some of our intrinsics.
743 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
745 // We have target-specific dag combine patterns for the following nodes:
746 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
747 setTargetDAGCombine(ISD::BUILD_VECTOR);
748 setTargetDAGCombine(ISD::SELECT);
749 setTargetDAGCombine(ISD::STORE);
751 computeRegisterProperties();
753 // FIXME: These should be based on subtarget info. Plus, the values should
754 // be smaller when we are in optimizing for size mode.
755 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
756 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
757 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
758 allowUnalignedMemoryAccesses = true; // x86 supports it!
759 setPrefLoopAlignment(16);
763 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
768 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
769 /// the desired ByVal argument alignment.
770 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
773 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
774 if (VTy->getBitWidth() == 128)
776 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
777 unsigned EltAlign = 0;
778 getMaxByValAlign(ATy->getElementType(), EltAlign);
779 if (EltAlign > MaxAlign)
781 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
782 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
783 unsigned EltAlign = 0;
784 getMaxByValAlign(STy->getElementType(i), EltAlign);
785 if (EltAlign > MaxAlign)
794 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
795 /// function arguments in the caller parameter area. For X86, aggregates
796 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
797 /// are at 4-byte boundaries.
798 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
799 if (Subtarget->is64Bit()) {
800 // Max of 8 and alignment of type.
801 unsigned TyAlign = getTargetData()->getABITypeAlignment(Ty);
808 if (Subtarget->hasSSE1())
809 getMaxByValAlign(Ty, Align);
813 /// getOptimalMemOpType - Returns the target specific optimal type for load
814 /// and store operations as a result of memset, memcpy, and memmove
815 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
818 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
819 bool isSrcConst, bool isSrcStr) const {
820 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
822 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
824 if (Subtarget->is64Bit() && Size >= 8)
830 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
832 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
833 SelectionDAG &DAG) const {
834 if (usesGlobalOffsetTable())
835 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
836 if (!Subtarget->isPICStyleRIPRel())
837 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
841 //===----------------------------------------------------------------------===//
842 // Return Value Calling Convention Implementation
843 //===----------------------------------------------------------------------===//
845 #include "X86GenCallingConv.inc"
847 /// LowerRET - Lower an ISD::RET node.
848 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
849 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
851 SmallVector<CCValAssign, 16> RVLocs;
852 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
853 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
854 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
855 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
857 // If this is the first return lowered for this function, add the regs to the
858 // liveout set for the function.
859 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
860 for (unsigned i = 0; i != RVLocs.size(); ++i)
861 if (RVLocs[i].isRegLoc())
862 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
864 SDValue Chain = Op.getOperand(0);
866 // Handle tail call return.
867 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
868 if (Chain.getOpcode() == X86ISD::TAILCALL) {
869 SDValue TailCall = Chain;
870 SDValue TargetAddress = TailCall.getOperand(1);
871 SDValue StackAdjustment = TailCall.getOperand(2);
872 assert(((TargetAddress.getOpcode() == ISD::Register &&
873 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
874 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
875 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
876 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
877 "Expecting an global address, external symbol, or register");
878 assert(StackAdjustment.getOpcode() == ISD::Constant &&
879 "Expecting a const value");
881 SmallVector<SDValue,8> Operands;
882 Operands.push_back(Chain.getOperand(0));
883 Operands.push_back(TargetAddress);
884 Operands.push_back(StackAdjustment);
885 // Copy registers used by the call. Last operand is a flag so it is not
887 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
888 Operands.push_back(Chain.getOperand(i));
890 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
897 SmallVector<SDValue, 6> RetOps;
898 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
899 // Operand #1 = Bytes To Pop
900 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
902 // Copy the result values into the output registers.
903 for (unsigned i = 0; i != RVLocs.size(); ++i) {
904 CCValAssign &VA = RVLocs[i];
905 assert(VA.isRegLoc() && "Can only return in registers!");
906 SDValue ValToCopy = Op.getOperand(i*2+1);
908 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
909 // the RET instruction and handled by the FP Stackifier.
910 if (RVLocs[i].getLocReg() == X86::ST0 ||
911 RVLocs[i].getLocReg() == X86::ST1) {
912 // If this is a copy from an xmm register to ST(0), use an FPExtend to
913 // change the value to the FP stack register class.
914 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
915 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
916 RetOps.push_back(ValToCopy);
917 // Don't emit a copytoreg.
921 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
922 Flag = Chain.getValue(1);
925 // The x86-64 ABI for returning structs by value requires that we copy
926 // the sret argument into %rax for the return. We saved the argument into
927 // a virtual register in the entry block, so now we copy the value out
929 if (Subtarget->is64Bit() &&
930 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
931 MachineFunction &MF = DAG.getMachineFunction();
932 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
933 unsigned Reg = FuncInfo->getSRetReturnReg();
935 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
936 FuncInfo->setSRetReturnReg(Reg);
938 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
940 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
941 Flag = Chain.getValue(1);
944 RetOps[0] = Chain; // Update chain.
946 // Add the flag if we have it.
948 RetOps.push_back(Flag);
950 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
954 /// LowerCallResult - Lower the result values of an ISD::CALL into the
955 /// appropriate copies out of appropriate physical registers. This assumes that
956 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
957 /// being lowered. The returns a SDNode with the same number of values as the
959 SDNode *X86TargetLowering::
960 LowerCallResult(SDValue Chain, SDValue InFlag, SDNode *TheCall,
961 unsigned CallingConv, SelectionDAG &DAG) {
963 // Assign locations to each value returned by this call.
964 SmallVector<CCValAssign, 16> RVLocs;
965 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
966 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
967 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
969 SmallVector<SDValue, 8> ResultVals;
971 // Copy all of the result registers out of their specified physreg.
972 for (unsigned i = 0; i != RVLocs.size(); ++i) {
973 MVT CopyVT = RVLocs[i].getValVT();
975 // If this is a call to a function that returns an fp value on the floating
976 // point stack, but where we prefer to use the value in xmm registers, copy
977 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
978 if ((RVLocs[i].getLocReg() == X86::ST0 ||
979 RVLocs[i].getLocReg() == X86::ST1) &&
980 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
984 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
985 CopyVT, InFlag).getValue(1);
986 SDValue Val = Chain.getValue(0);
987 InFlag = Chain.getValue(2);
989 if (CopyVT != RVLocs[i].getValVT()) {
990 // Round the F80 the right size, which also moves to the appropriate xmm
992 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
993 // This truncation won't change the value.
994 DAG.getIntPtrConstant(1));
997 ResultVals.push_back(Val);
1000 // Merge everything together with a MERGE_VALUES node.
1001 ResultVals.push_back(Chain);
1002 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1003 ResultVals.size()).getNode();
1007 //===----------------------------------------------------------------------===//
1008 // C & StdCall & Fast Calling Convention implementation
1009 //===----------------------------------------------------------------------===//
1010 // StdCall calling convention seems to be standard for many Windows' API
1011 // routines and around. It differs from C calling convention just a little:
1012 // callee should clean up the stack, not caller. Symbols should be also
1013 // decorated in some fancy way :) It doesn't support any vector arguments.
1014 // For info on fast calling convention see Fast Calling Convention (tail call)
1015 // implementation LowerX86_32FastCCCallTo.
1017 /// AddLiveIn - This helper function adds the specified physical register to the
1018 /// MachineFunction as a live in value. It also creates a corresponding virtual
1019 /// register for it.
1020 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1021 const TargetRegisterClass *RC) {
1022 assert(RC->contains(PReg) && "Not the correct regclass!");
1023 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1024 MF.getRegInfo().addLiveIn(PReg, VReg);
1028 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1030 static bool CallIsStructReturn(SDValue Op) {
1031 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1035 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
1038 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1039 /// return semantics.
1040 static bool ArgsAreStructReturn(SDValue Op) {
1041 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1045 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1048 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1049 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1051 bool X86TargetLowering::IsCalleePop(SDValue Op) {
1052 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1056 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1059 case CallingConv::X86_StdCall:
1060 return !Subtarget->is64Bit();
1061 case CallingConv::X86_FastCall:
1062 return !Subtarget->is64Bit();
1063 case CallingConv::Fast:
1064 return PerformTailCallOpt;
1068 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1069 /// FORMAL_ARGUMENTS node.
1070 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
1071 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1073 if (Subtarget->is64Bit()) {
1074 if (Subtarget->isTargetWin64())
1075 return CC_X86_Win64_C;
1077 if (CC == CallingConv::Fast && PerformTailCallOpt)
1078 return CC_X86_64_TailCall;
1084 if (CC == CallingConv::X86_FastCall)
1085 return CC_X86_32_FastCall;
1086 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1087 return CC_X86_32_TailCall;
1092 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1093 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1095 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1096 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1097 if (CC == CallingConv::X86_FastCall)
1099 else if (CC == CallingConv::X86_StdCall)
1105 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1106 /// in a register before calling.
1107 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1108 return !IsTailCall && !Is64Bit &&
1109 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1110 Subtarget->isPICStyleGOT();
1113 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1114 /// address to be loaded in a register.
1116 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1117 return !Is64Bit && IsTailCall &&
1118 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1119 Subtarget->isPICStyleGOT();
1122 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1123 /// by "Src" to address "Dst" with size and alignment information specified by
1124 /// the specific parameter attribute. The copy will be passed as a byval
1125 /// function parameter.
1127 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1128 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1129 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1130 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1131 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1134 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1135 const CCValAssign &VA,
1136 MachineFrameInfo *MFI,
1138 SDValue Root, unsigned i) {
1139 // Create the nodes corresponding to a load from this parameter slot.
1140 ISD::ArgFlagsTy Flags =
1141 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1142 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1143 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1145 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1146 // changed with more analysis.
1147 // In case of tail call optimization mark all arguments mutable. Since they
1148 // could be overwritten by lowering of arguments in case of a tail call.
1149 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1150 VA.getLocMemOffset(), isImmutable);
1151 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1152 if (Flags.isByVal())
1154 return DAG.getLoad(VA.getValVT(), Root, FIN,
1155 PseudoSourceValue::getFixedStack(FI), 0);
1159 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1160 MachineFunction &MF = DAG.getMachineFunction();
1161 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1163 const Function* Fn = MF.getFunction();
1164 if (Fn->hasExternalLinkage() &&
1165 Subtarget->isTargetCygMing() &&
1166 Fn->getName() == "main")
1167 FuncInfo->setForceFramePointer(true);
1169 // Decorate the function name.
1170 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1172 MachineFrameInfo *MFI = MF.getFrameInfo();
1173 SDValue Root = Op.getOperand(0);
1174 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1175 unsigned CC = MF.getFunction()->getCallingConv();
1176 bool Is64Bit = Subtarget->is64Bit();
1177 bool IsWin64 = Subtarget->isTargetWin64();
1179 assert(!(isVarArg && CC == CallingConv::Fast) &&
1180 "Var args not supported with calling convention fastcc");
1182 // Assign locations to all of the incoming arguments.
1183 SmallVector<CCValAssign, 16> ArgLocs;
1184 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1185 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(Op));
1187 SmallVector<SDValue, 8> ArgValues;
1188 unsigned LastVal = ~0U;
1189 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1190 CCValAssign &VA = ArgLocs[i];
1191 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1193 assert(VA.getValNo() != LastVal &&
1194 "Don't support value assigned to multiple locs yet");
1195 LastVal = VA.getValNo();
1197 if (VA.isRegLoc()) {
1198 MVT RegVT = VA.getLocVT();
1199 TargetRegisterClass *RC;
1200 if (RegVT == MVT::i32)
1201 RC = X86::GR32RegisterClass;
1202 else if (Is64Bit && RegVT == MVT::i64)
1203 RC = X86::GR64RegisterClass;
1204 else if (RegVT == MVT::f32)
1205 RC = X86::FR32RegisterClass;
1206 else if (RegVT == MVT::f64)
1207 RC = X86::FR64RegisterClass;
1208 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1209 RC = X86::VR128RegisterClass;
1210 else if (RegVT.isVector()) {
1211 assert(RegVT.getSizeInBits() == 64);
1213 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1215 // Darwin calling convention passes MMX values in either GPRs or
1216 // XMMs in x86-64. Other targets pass them in memory.
1217 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1218 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1221 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1226 assert(0 && "Unknown argument type!");
1229 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1230 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1232 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1233 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1235 if (VA.getLocInfo() == CCValAssign::SExt)
1236 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1237 DAG.getValueType(VA.getValVT()));
1238 else if (VA.getLocInfo() == CCValAssign::ZExt)
1239 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1240 DAG.getValueType(VA.getValVT()));
1242 if (VA.getLocInfo() != CCValAssign::Full)
1243 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1245 // Handle MMX values passed in GPRs.
1246 if (Is64Bit && RegVT != VA.getLocVT()) {
1247 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1248 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1249 else if (RC == X86::VR128RegisterClass) {
1250 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1251 DAG.getConstant(0, MVT::i64));
1252 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1256 ArgValues.push_back(ArgValue);
1258 assert(VA.isMemLoc());
1259 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1263 // The x86-64 ABI for returning structs by value requires that we copy
1264 // the sret argument into %rax for the return. Save the argument into
1265 // a virtual register so that we can access it from the return points.
1266 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1267 MachineFunction &MF = DAG.getMachineFunction();
1268 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1269 unsigned Reg = FuncInfo->getSRetReturnReg();
1271 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1272 FuncInfo->setSRetReturnReg(Reg);
1274 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1275 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1278 unsigned StackSize = CCInfo.getNextStackOffset();
1279 // align stack specially for tail calls
1280 if (CC == CallingConv::Fast)
1281 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1283 // If the function takes variable number of arguments, make a frame index for
1284 // the start of the first vararg value... for expansion of llvm.va_start.
1286 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1287 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1290 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1292 // FIXME: We should really autogenerate these arrays
1293 static const unsigned GPR64ArgRegsWin64[] = {
1294 X86::RCX, X86::RDX, X86::R8, X86::R9
1296 static const unsigned XMMArgRegsWin64[] = {
1297 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1299 static const unsigned GPR64ArgRegs64Bit[] = {
1300 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1302 static const unsigned XMMArgRegs64Bit[] = {
1303 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1304 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1306 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1309 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1310 GPR64ArgRegs = GPR64ArgRegsWin64;
1311 XMMArgRegs = XMMArgRegsWin64;
1313 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1314 GPR64ArgRegs = GPR64ArgRegs64Bit;
1315 XMMArgRegs = XMMArgRegs64Bit;
1317 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1319 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1322 // For X86-64, if there are vararg parameters that are passed via
1323 // registers, then we must store them to their spots on the stack so they
1324 // may be loaded by deferencing the result of va_next.
1325 VarArgsGPOffset = NumIntRegs * 8;
1326 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1327 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1328 TotalNumXMMRegs * 16, 16);
1330 // Store the integer parameter registers.
1331 SmallVector<SDValue, 8> MemOps;
1332 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1333 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1334 DAG.getIntPtrConstant(VarArgsGPOffset));
1335 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1336 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1337 X86::GR64RegisterClass);
1338 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1340 DAG.getStore(Val.getValue(1), Val, FIN,
1341 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1342 MemOps.push_back(Store);
1343 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1344 DAG.getIntPtrConstant(8));
1347 // Now store the XMM (fp + vector) parameter registers.
1348 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1349 DAG.getIntPtrConstant(VarArgsFPOffset));
1350 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1351 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1352 X86::VR128RegisterClass);
1353 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1355 DAG.getStore(Val.getValue(1), Val, FIN,
1356 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1357 MemOps.push_back(Store);
1358 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1359 DAG.getIntPtrConstant(16));
1361 if (!MemOps.empty())
1362 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1363 &MemOps[0], MemOps.size());
1367 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1368 // arguments and the arguments after the retaddr has been pushed are
1370 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1371 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1372 (StackSize & 7) == 0)
1375 ArgValues.push_back(Root);
1377 // Some CCs need callee pop.
1378 if (IsCalleePop(Op)) {
1379 BytesToPopOnReturn = StackSize; // Callee pops everything.
1380 BytesCallerReserves = 0;
1382 BytesToPopOnReturn = 0; // Callee pops nothing.
1383 // If this is an sret function, the return should pop the hidden pointer.
1384 if (!Is64Bit && ArgsAreStructReturn(Op))
1385 BytesToPopOnReturn = 4;
1386 BytesCallerReserves = StackSize;
1390 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1391 if (CC == CallingConv::X86_FastCall)
1392 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1395 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1397 // Return the new list of results.
1398 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1399 ArgValues.size()).getValue(Op.getResNo());
1403 X86TargetLowering::LowerMemOpCallTo(SDValue Op, SelectionDAG &DAG,
1404 const SDValue &StackPtr,
1405 const CCValAssign &VA,
1408 unsigned LocMemOffset = VA.getLocMemOffset();
1409 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1410 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1411 ISD::ArgFlagsTy Flags =
1412 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1413 if (Flags.isByVal()) {
1414 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1416 return DAG.getStore(Chain, Arg, PtrOff,
1417 PseudoSourceValue::getStack(), LocMemOffset);
1420 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1421 /// optimization is performed and it is required.
1423 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1424 SDValue &OutRetAddr,
1429 if (!IsTailCall || FPDiff==0) return Chain;
1431 // Adjust the Return address stack slot.
1432 MVT VT = getPointerTy();
1433 OutRetAddr = getReturnAddressFrameIndex(DAG);
1434 // Load the "old" Return address.
1435 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1436 return SDValue(OutRetAddr.getNode(), 1);
1439 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1440 /// optimization is performed and it is required (FPDiff!=0).
1442 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1443 SDValue Chain, SDValue RetAddrFrIdx,
1444 bool Is64Bit, int FPDiff) {
1445 // Store the return address to the appropriate stack slot.
1446 if (!FPDiff) return Chain;
1447 // Calculate the new stack slot for the return address.
1448 int SlotSize = Is64Bit ? 8 : 4;
1449 int NewReturnAddrFI =
1450 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1451 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1452 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1453 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1454 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1458 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1459 MachineFunction &MF = DAG.getMachineFunction();
1460 SDValue Chain = Op.getOperand(0);
1461 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1462 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1463 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1464 && CC == CallingConv::Fast && PerformTailCallOpt;
1465 SDValue Callee = Op.getOperand(4);
1466 bool Is64Bit = Subtarget->is64Bit();
1467 bool IsStructRet = CallIsStructReturn(Op);
1469 assert(!(isVarArg && CC == CallingConv::Fast) &&
1470 "Var args not supported with calling convention fastcc");
1472 // Analyze operands of the call, assigning locations to each operand.
1473 SmallVector<CCValAssign, 16> ArgLocs;
1474 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1475 CCInfo.AnalyzeCallOperands(Op.getNode(), CCAssignFnForNode(Op));
1477 // Get a count of how many bytes are to be pushed on the stack.
1478 unsigned NumBytes = CCInfo.getNextStackOffset();
1479 if (CC == CallingConv::Fast)
1480 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1482 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1483 // arguments and the arguments after the retaddr has been pushed are aligned.
1484 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1485 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1486 (NumBytes & 7) == 0)
1491 // Lower arguments at fp - stackoffset + fpdiff.
1492 unsigned NumBytesCallerPushed =
1493 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1494 FPDiff = NumBytesCallerPushed - NumBytes;
1496 // Set the delta of movement of the returnaddr stackslot.
1497 // But only set if delta is greater than previous delta.
1498 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1499 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1502 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1504 SDValue RetAddrFrIdx;
1505 // Load return adress for tail calls.
1506 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1509 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1510 SmallVector<SDValue, 8> MemOpChains;
1513 // Walk the register/memloc assignments, inserting copies/loads. In the case
1514 // of tail call optimization arguments are handle later.
1515 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1516 CCValAssign &VA = ArgLocs[i];
1517 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1518 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1519 getArgFlags().isByVal();
1521 // Promote the value if needed.
1522 switch (VA.getLocInfo()) {
1523 default: assert(0 && "Unknown loc info!");
1524 case CCValAssign::Full: break;
1525 case CCValAssign::SExt:
1526 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1528 case CCValAssign::ZExt:
1529 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1531 case CCValAssign::AExt:
1532 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1536 if (VA.isRegLoc()) {
1538 MVT RegVT = VA.getLocVT();
1539 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1540 switch (VA.getLocReg()) {
1543 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1545 // Special case: passing MMX values in GPR registers.
1546 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1549 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1550 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1551 // Special case: passing MMX values in XMM registers.
1552 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1553 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1554 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1555 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1556 getMOVLMask(2, DAG));
1561 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1563 if (!IsTailCall || (IsTailCall && isByVal)) {
1564 assert(VA.isMemLoc());
1565 if (StackPtr.getNode() == 0)
1566 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1568 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1574 if (!MemOpChains.empty())
1575 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1576 &MemOpChains[0], MemOpChains.size());
1578 // Build a sequence of copy-to-reg nodes chained together with token chain
1579 // and flag operands which copy the outgoing args into registers.
1581 // Tail call byval lowering might overwrite argument registers so in case of
1582 // tail call optimization the copies to registers are lowered later.
1584 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1585 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1587 InFlag = Chain.getValue(1);
1590 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1592 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1593 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1594 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1596 InFlag = Chain.getValue(1);
1598 // If we are tail calling and generating PIC/GOT style code load the address
1599 // of the callee into ecx. The value in ecx is used as target of the tail
1600 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1601 // calls on PIC/GOT architectures. Normally we would just put the address of
1602 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1603 // restored (since ebx is callee saved) before jumping to the target@PLT.
1604 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1605 // Note: The actual moving to ecx is done further down.
1606 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1607 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1608 !G->getGlobal()->hasProtectedVisibility())
1609 Callee = LowerGlobalAddress(Callee, DAG);
1610 else if (isa<ExternalSymbolSDNode>(Callee))
1611 Callee = LowerExternalSymbol(Callee,DAG);
1614 if (Is64Bit && isVarArg) {
1615 // From AMD64 ABI document:
1616 // For calls that may call functions that use varargs or stdargs
1617 // (prototype-less calls or calls to functions containing ellipsis (...) in
1618 // the declaration) %al is used as hidden argument to specify the number
1619 // of SSE registers used. The contents of %al do not need to match exactly
1620 // the number of registers, but must be an ubound on the number of SSE
1621 // registers used and is in the range 0 - 8 inclusive.
1623 // FIXME: Verify this on Win64
1624 // Count the number of XMM registers allocated.
1625 static const unsigned XMMArgRegs[] = {
1626 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1627 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1629 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1631 Chain = DAG.getCopyToReg(Chain, X86::AL,
1632 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1633 InFlag = Chain.getValue(1);
1637 // For tail calls lower the arguments to the 'real' stack slot.
1639 SmallVector<SDValue, 8> MemOpChains2;
1642 // Do not flag preceeding copytoreg stuff together with the following stuff.
1644 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1645 CCValAssign &VA = ArgLocs[i];
1646 if (!VA.isRegLoc()) {
1647 assert(VA.isMemLoc());
1648 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1649 SDValue FlagsOp = Op.getOperand(6+2*VA.getValNo());
1650 ISD::ArgFlagsTy Flags =
1651 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1652 // Create frame index.
1653 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1654 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1655 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1656 FIN = DAG.getFrameIndex(FI, getPointerTy());
1658 if (Flags.isByVal()) {
1659 // Copy relative to framepointer.
1660 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1661 if (StackPtr.getNode() == 0)
1662 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1663 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1665 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1668 // Store relative to framepointer.
1669 MemOpChains2.push_back(
1670 DAG.getStore(Chain, Arg, FIN,
1671 PseudoSourceValue::getFixedStack(FI), 0));
1676 if (!MemOpChains2.empty())
1677 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1678 &MemOpChains2[0], MemOpChains2.size());
1680 // Copy arguments to their registers.
1681 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1682 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1684 InFlag = Chain.getValue(1);
1688 // Store the return address to the appropriate stack slot.
1689 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1693 // If the callee is a GlobalAddress node (quite common, every direct call is)
1694 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1695 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1696 // We should use extra load for direct calls to dllimported functions in
1698 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1699 getTargetMachine(), true))
1700 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1701 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1702 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1703 } else if (IsTailCall) {
1704 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1706 Chain = DAG.getCopyToReg(Chain,
1707 DAG.getRegister(Opc, getPointerTy()),
1709 Callee = DAG.getRegister(Opc, getPointerTy());
1710 // Add register as live out.
1711 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1714 // Returns a chain & a flag for retval copy to use.
1715 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1716 SmallVector<SDValue, 8> Ops;
1719 Ops.push_back(Chain);
1720 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1721 Ops.push_back(DAG.getIntPtrConstant(0));
1722 if (InFlag.getNode())
1723 Ops.push_back(InFlag);
1724 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1725 InFlag = Chain.getValue(1);
1727 // Returns a chain & a flag for retval copy to use.
1728 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1732 Ops.push_back(Chain);
1733 Ops.push_back(Callee);
1736 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1738 // Add argument registers to the end of the list so that they are known live
1740 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1741 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1742 RegsToPass[i].second.getValueType()));
1744 // Add an implicit use GOT pointer in EBX.
1745 if (!IsTailCall && !Is64Bit &&
1746 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1747 Subtarget->isPICStyleGOT())
1748 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1750 // Add an implicit use of AL for x86 vararg functions.
1751 if (Is64Bit && isVarArg)
1752 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1754 if (InFlag.getNode())
1755 Ops.push_back(InFlag);
1758 assert(InFlag.getNode() &&
1759 "Flag must be set. Depend on flag being set in LowerRET");
1760 Chain = DAG.getNode(X86ISD::TAILCALL,
1761 Op.getNode()->getVTList(), &Ops[0], Ops.size());
1763 return SDValue(Chain.getNode(), Op.getResNo());
1766 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1767 InFlag = Chain.getValue(1);
1769 // Create the CALLSEQ_END node.
1770 unsigned NumBytesForCalleeToPush;
1771 if (IsCalleePop(Op))
1772 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1773 else if (!Is64Bit && IsStructRet)
1774 // If this is is a call to a struct-return function, the callee
1775 // pops the hidden struct pointer, so we have to push it back.
1776 // This is common for Darwin/X86, Linux & Mingw32 targets.
1777 NumBytesForCalleeToPush = 4;
1779 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1781 // Returns a flag for retval copy to use.
1782 Chain = DAG.getCALLSEQ_END(Chain,
1783 DAG.getIntPtrConstant(NumBytes),
1784 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1786 InFlag = Chain.getValue(1);
1788 // Handle result values, copying them out of physregs into vregs that we
1790 return SDValue(LowerCallResult(Chain, InFlag, Op.getNode(), CC, DAG),
1795 //===----------------------------------------------------------------------===//
1796 // Fast Calling Convention (tail call) implementation
1797 //===----------------------------------------------------------------------===//
1799 // Like std call, callee cleans arguments, convention except that ECX is
1800 // reserved for storing the tail called function address. Only 2 registers are
1801 // free for argument passing (inreg). Tail call optimization is performed
1803 // * tailcallopt is enabled
1804 // * caller/callee are fastcc
1805 // On X86_64 architecture with GOT-style position independent code only local
1806 // (within module) calls are supported at the moment.
1807 // To keep the stack aligned according to platform abi the function
1808 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1809 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1810 // If a tail called function callee has more arguments than the caller the
1811 // caller needs to make sure that there is room to move the RETADDR to. This is
1812 // achieved by reserving an area the size of the argument delta right after the
1813 // original REtADDR, but before the saved framepointer or the spilled registers
1814 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1826 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1827 /// for a 16 byte align requirement.
1828 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1829 SelectionDAG& DAG) {
1830 if (PerformTailCallOpt) {
1831 MachineFunction &MF = DAG.getMachineFunction();
1832 const TargetMachine &TM = MF.getTarget();
1833 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1834 unsigned StackAlignment = TFI.getStackAlignment();
1835 uint64_t AlignMask = StackAlignment - 1;
1836 int64_t Offset = StackSize;
1837 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1838 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1839 // Number smaller than 12 so just add the difference.
1840 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1842 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1843 Offset = ((~AlignMask) & Offset) + StackAlignment +
1844 (StackAlignment-SlotSize);
1851 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1852 /// following the call is a return. A function is eligible if caller/callee
1853 /// calling conventions match, currently only fastcc supports tail calls, and
1854 /// the function CALL is immediatly followed by a RET.
1855 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1857 SelectionDAG& DAG) const {
1858 if (!PerformTailCallOpt)
1861 if (CheckTailCallReturnConstraints(Call, Ret)) {
1862 MachineFunction &MF = DAG.getMachineFunction();
1863 unsigned CallerCC = MF.getFunction()->getCallingConv();
1864 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1865 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1866 SDValue Callee = Call.getOperand(4);
1867 // On x86/32Bit PIC/GOT tail calls are supported.
1868 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1869 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1872 // Can only do local tail calls (in same module, hidden or protected) on
1873 // x86_64 PIC/GOT at the moment.
1874 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1875 return G->getGlobal()->hasHiddenVisibility()
1876 || G->getGlobal()->hasProtectedVisibility();
1883 FastISel *X86TargetLowering::createFastISel(MachineFunction &mf) {
1884 return X86::createFastISel(mf);
1888 //===----------------------------------------------------------------------===//
1889 // Other Lowering Hooks
1890 //===----------------------------------------------------------------------===//
1893 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1894 MachineFunction &MF = DAG.getMachineFunction();
1895 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1896 int ReturnAddrIndex = FuncInfo->getRAIndex();
1898 if (ReturnAddrIndex == 0) {
1899 // Set up a frame object for the return address.
1900 if (Subtarget->is64Bit())
1901 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1903 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1905 FuncInfo->setRAIndex(ReturnAddrIndex);
1908 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1913 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1914 /// specific condition code. It returns a false if it cannot do a direct
1915 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1917 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1918 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1919 SelectionDAG &DAG) {
1920 X86CC = X86::COND_INVALID;
1922 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1923 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1924 // X > -1 -> X == 0, jump !sign.
1925 RHS = DAG.getConstant(0, RHS.getValueType());
1926 X86CC = X86::COND_NS;
1928 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1929 // X < 0 -> X == 0, jump on sign.
1930 X86CC = X86::COND_S;
1932 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1934 RHS = DAG.getConstant(0, RHS.getValueType());
1935 X86CC = X86::COND_LE;
1940 switch (SetCCOpcode) {
1942 case ISD::SETEQ: X86CC = X86::COND_E; break;
1943 case ISD::SETGT: X86CC = X86::COND_G; break;
1944 case ISD::SETGE: X86CC = X86::COND_GE; break;
1945 case ISD::SETLT: X86CC = X86::COND_L; break;
1946 case ISD::SETLE: X86CC = X86::COND_LE; break;
1947 case ISD::SETNE: X86CC = X86::COND_NE; break;
1948 case ISD::SETULT: X86CC = X86::COND_B; break;
1949 case ISD::SETUGT: X86CC = X86::COND_A; break;
1950 case ISD::SETULE: X86CC = X86::COND_BE; break;
1951 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1954 // On a floating point condition, the flags are set as follows:
1956 // 0 | 0 | 0 | X > Y
1957 // 0 | 0 | 1 | X < Y
1958 // 1 | 0 | 0 | X == Y
1959 // 1 | 1 | 1 | unordered
1961 switch (SetCCOpcode) {
1964 case ISD::SETEQ: X86CC = X86::COND_E; break;
1965 case ISD::SETOLT: Flip = true; // Fallthrough
1967 case ISD::SETGT: X86CC = X86::COND_A; break;
1968 case ISD::SETOLE: Flip = true; // Fallthrough
1970 case ISD::SETGE: X86CC = X86::COND_AE; break;
1971 case ISD::SETUGT: Flip = true; // Fallthrough
1973 case ISD::SETLT: X86CC = X86::COND_B; break;
1974 case ISD::SETUGE: Flip = true; // Fallthrough
1976 case ISD::SETLE: X86CC = X86::COND_BE; break;
1978 case ISD::SETNE: X86CC = X86::COND_NE; break;
1979 case ISD::SETUO: X86CC = X86::COND_P; break;
1980 case ISD::SETO: X86CC = X86::COND_NP; break;
1983 std::swap(LHS, RHS);
1986 return X86CC != X86::COND_INVALID;
1989 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1990 /// code. Current x86 isa includes the following FP cmov instructions:
1991 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1992 static bool hasFPCMov(unsigned X86CC) {
2008 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2009 /// true if Op is undef or if its value falls within the specified range (L, H].
2010 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2011 if (Op.getOpcode() == ISD::UNDEF)
2014 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2015 return (Val >= Low && Val < Hi);
2018 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2019 /// true if Op is undef or if its value equal to the specified value.
2020 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2021 if (Op.getOpcode() == ISD::UNDEF)
2023 return cast<ConstantSDNode>(Op)->getValue() == Val;
2026 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2027 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2028 bool X86::isPSHUFDMask(SDNode *N) {
2029 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2031 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2034 // Check if the value doesn't reference the second vector.
2035 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2036 SDValue Arg = N->getOperand(i);
2037 if (Arg.getOpcode() == ISD::UNDEF) continue;
2038 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2039 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2046 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2047 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2048 bool X86::isPSHUFHWMask(SDNode *N) {
2049 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2051 if (N->getNumOperands() != 8)
2054 // Lower quadword copied in order.
2055 for (unsigned i = 0; i != 4; ++i) {
2056 SDValue Arg = N->getOperand(i);
2057 if (Arg.getOpcode() == ISD::UNDEF) continue;
2058 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2059 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2063 // Upper quadword shuffled.
2064 for (unsigned i = 4; i != 8; ++i) {
2065 SDValue Arg = N->getOperand(i);
2066 if (Arg.getOpcode() == ISD::UNDEF) continue;
2067 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2068 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2069 if (Val < 4 || Val > 7)
2076 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2077 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2078 bool X86::isPSHUFLWMask(SDNode *N) {
2079 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2081 if (N->getNumOperands() != 8)
2084 // Upper quadword copied in order.
2085 for (unsigned i = 4; i != 8; ++i)
2086 if (!isUndefOrEqual(N->getOperand(i), i))
2089 // Lower quadword shuffled.
2090 for (unsigned i = 0; i != 4; ++i)
2091 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2097 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2098 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2099 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2100 if (NumElems != 2 && NumElems != 4) return false;
2102 unsigned Half = NumElems / 2;
2103 for (unsigned i = 0; i < Half; ++i)
2104 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2106 for (unsigned i = Half; i < NumElems; ++i)
2107 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2113 bool X86::isSHUFPMask(SDNode *N) {
2114 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2115 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2118 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2119 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2120 /// half elements to come from vector 1 (which would equal the dest.) and
2121 /// the upper half to come from vector 2.
2122 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2123 if (NumOps != 2 && NumOps != 4) return false;
2125 unsigned Half = NumOps / 2;
2126 for (unsigned i = 0; i < Half; ++i)
2127 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2129 for (unsigned i = Half; i < NumOps; ++i)
2130 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2135 static bool isCommutedSHUFP(SDNode *N) {
2136 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2137 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2140 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2141 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2142 bool X86::isMOVHLPSMask(SDNode *N) {
2143 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2145 if (N->getNumOperands() != 4)
2148 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2149 return isUndefOrEqual(N->getOperand(0), 6) &&
2150 isUndefOrEqual(N->getOperand(1), 7) &&
2151 isUndefOrEqual(N->getOperand(2), 2) &&
2152 isUndefOrEqual(N->getOperand(3), 3);
2155 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2156 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2158 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2159 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2161 if (N->getNumOperands() != 4)
2164 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2165 return isUndefOrEqual(N->getOperand(0), 2) &&
2166 isUndefOrEqual(N->getOperand(1), 3) &&
2167 isUndefOrEqual(N->getOperand(2), 2) &&
2168 isUndefOrEqual(N->getOperand(3), 3);
2171 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2172 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2173 bool X86::isMOVLPMask(SDNode *N) {
2174 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2176 unsigned NumElems = N->getNumOperands();
2177 if (NumElems != 2 && NumElems != 4)
2180 for (unsigned i = 0; i < NumElems/2; ++i)
2181 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2184 for (unsigned i = NumElems/2; i < NumElems; ++i)
2185 if (!isUndefOrEqual(N->getOperand(i), i))
2191 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2192 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2194 bool X86::isMOVHPMask(SDNode *N) {
2195 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2197 unsigned NumElems = N->getNumOperands();
2198 if (NumElems != 2 && NumElems != 4)
2201 for (unsigned i = 0; i < NumElems/2; ++i)
2202 if (!isUndefOrEqual(N->getOperand(i), i))
2205 for (unsigned i = 0; i < NumElems/2; ++i) {
2206 SDValue Arg = N->getOperand(i + NumElems/2);
2207 if (!isUndefOrEqual(Arg, i + NumElems))
2214 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2215 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2216 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2217 bool V2IsSplat = false) {
2218 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2221 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2222 SDValue BitI = Elts[i];
2223 SDValue BitI1 = Elts[i+1];
2224 if (!isUndefOrEqual(BitI, j))
2227 if (isUndefOrEqual(BitI1, NumElts))
2230 if (!isUndefOrEqual(BitI1, j + NumElts))
2238 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2239 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2240 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2243 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2244 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2245 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2246 bool V2IsSplat = false) {
2247 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2250 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2251 SDValue BitI = Elts[i];
2252 SDValue BitI1 = Elts[i+1];
2253 if (!isUndefOrEqual(BitI, j + NumElts/2))
2256 if (isUndefOrEqual(BitI1, NumElts))
2259 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2267 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2268 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2269 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2272 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2273 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2275 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2276 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2278 unsigned NumElems = N->getNumOperands();
2279 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2282 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2283 SDValue BitI = N->getOperand(i);
2284 SDValue BitI1 = N->getOperand(i+1);
2286 if (!isUndefOrEqual(BitI, j))
2288 if (!isUndefOrEqual(BitI1, j))
2295 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2296 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2298 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2299 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2301 unsigned NumElems = N->getNumOperands();
2302 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2305 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2306 SDValue BitI = N->getOperand(i);
2307 SDValue BitI1 = N->getOperand(i + 1);
2309 if (!isUndefOrEqual(BitI, j))
2311 if (!isUndefOrEqual(BitI1, j))
2318 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2319 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2320 /// MOVSD, and MOVD, i.e. setting the lowest element.
2321 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2322 if (NumElts != 2 && NumElts != 4)
2325 if (!isUndefOrEqual(Elts[0], NumElts))
2328 for (unsigned i = 1; i < NumElts; ++i) {
2329 if (!isUndefOrEqual(Elts[i], i))
2336 bool X86::isMOVLMask(SDNode *N) {
2337 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2338 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2341 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2342 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2343 /// element of vector 2 and the other elements to come from vector 1 in order.
2344 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2345 bool V2IsSplat = false,
2346 bool V2IsUndef = false) {
2347 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2350 if (!isUndefOrEqual(Ops[0], 0))
2353 for (unsigned i = 1; i < NumOps; ++i) {
2354 SDValue Arg = Ops[i];
2355 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2356 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2357 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2364 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2365 bool V2IsUndef = false) {
2366 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2367 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2368 V2IsSplat, V2IsUndef);
2371 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2372 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2373 bool X86::isMOVSHDUPMask(SDNode *N) {
2374 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2376 if (N->getNumOperands() != 4)
2379 // Expect 1, 1, 3, 3
2380 for (unsigned i = 0; i < 2; ++i) {
2381 SDValue Arg = N->getOperand(i);
2382 if (Arg.getOpcode() == ISD::UNDEF) continue;
2383 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2384 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2385 if (Val != 1) return false;
2389 for (unsigned i = 2; i < 4; ++i) {
2390 SDValue Arg = N->getOperand(i);
2391 if (Arg.getOpcode() == ISD::UNDEF) continue;
2392 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2393 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2394 if (Val != 3) return false;
2398 // Don't use movshdup if it can be done with a shufps.
2402 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2403 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2404 bool X86::isMOVSLDUPMask(SDNode *N) {
2405 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2407 if (N->getNumOperands() != 4)
2410 // Expect 0, 0, 2, 2
2411 for (unsigned i = 0; i < 2; ++i) {
2412 SDValue Arg = N->getOperand(i);
2413 if (Arg.getOpcode() == ISD::UNDEF) continue;
2414 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2415 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2416 if (Val != 0) return false;
2420 for (unsigned i = 2; i < 4; ++i) {
2421 SDValue Arg = N->getOperand(i);
2422 if (Arg.getOpcode() == ISD::UNDEF) continue;
2423 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2424 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2425 if (Val != 2) return false;
2429 // Don't use movshdup if it can be done with a shufps.
2433 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2434 /// specifies a identity operation on the LHS or RHS.
2435 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2436 unsigned NumElems = N->getNumOperands();
2437 for (unsigned i = 0; i < NumElems; ++i)
2438 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2443 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2444 /// a splat of a single element.
2445 static bool isSplatMask(SDNode *N) {
2446 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2448 // This is a splat operation if each element of the permute is the same, and
2449 // if the value doesn't reference the second vector.
2450 unsigned NumElems = N->getNumOperands();
2451 SDValue ElementBase;
2453 for (; i != NumElems; ++i) {
2454 SDValue Elt = N->getOperand(i);
2455 if (isa<ConstantSDNode>(Elt)) {
2461 if (!ElementBase.getNode())
2464 for (; i != NumElems; ++i) {
2465 SDValue Arg = N->getOperand(i);
2466 if (Arg.getOpcode() == ISD::UNDEF) continue;
2467 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2468 if (Arg != ElementBase) return false;
2471 // Make sure it is a splat of the first vector operand.
2472 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2475 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2476 /// a splat of a single element and it's a 2 or 4 element mask.
2477 bool X86::isSplatMask(SDNode *N) {
2478 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2480 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2481 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2483 return ::isSplatMask(N);
2486 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2487 /// specifies a splat of zero element.
2488 bool X86::isSplatLoMask(SDNode *N) {
2489 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2491 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2492 if (!isUndefOrEqual(N->getOperand(i), 0))
2497 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2498 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2500 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2501 unsigned NumOperands = N->getNumOperands();
2502 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2504 for (unsigned i = 0; i < NumOperands; ++i) {
2506 SDValue Arg = N->getOperand(NumOperands-i-1);
2507 if (Arg.getOpcode() != ISD::UNDEF)
2508 Val = cast<ConstantSDNode>(Arg)->getValue();
2509 if (Val >= NumOperands) Val -= NumOperands;
2511 if (i != NumOperands - 1)
2518 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2519 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2521 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2523 // 8 nodes, but we only care about the last 4.
2524 for (unsigned i = 7; i >= 4; --i) {
2526 SDValue Arg = N->getOperand(i);
2527 if (Arg.getOpcode() != ISD::UNDEF)
2528 Val = cast<ConstantSDNode>(Arg)->getValue();
2537 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2538 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2540 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2542 // 8 nodes, but we only care about the first 4.
2543 for (int i = 3; i >= 0; --i) {
2545 SDValue Arg = N->getOperand(i);
2546 if (Arg.getOpcode() != ISD::UNDEF)
2547 Val = cast<ConstantSDNode>(Arg)->getValue();
2556 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2557 /// specifies a 8 element shuffle that can be broken into a pair of
2558 /// PSHUFHW and PSHUFLW.
2559 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2560 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2562 if (N->getNumOperands() != 8)
2565 // Lower quadword shuffled.
2566 for (unsigned i = 0; i != 4; ++i) {
2567 SDValue Arg = N->getOperand(i);
2568 if (Arg.getOpcode() == ISD::UNDEF) continue;
2569 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2570 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2575 // Upper quadword shuffled.
2576 for (unsigned i = 4; i != 8; ++i) {
2577 SDValue Arg = N->getOperand(i);
2578 if (Arg.getOpcode() == ISD::UNDEF) continue;
2579 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2580 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2581 if (Val < 4 || Val > 7)
2588 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2589 /// values in ther permute mask.
2590 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2591 SDValue &V2, SDValue &Mask,
2592 SelectionDAG &DAG) {
2593 MVT VT = Op.getValueType();
2594 MVT MaskVT = Mask.getValueType();
2595 MVT EltVT = MaskVT.getVectorElementType();
2596 unsigned NumElems = Mask.getNumOperands();
2597 SmallVector<SDValue, 8> MaskVec;
2599 for (unsigned i = 0; i != NumElems; ++i) {
2600 SDValue Arg = Mask.getOperand(i);
2601 if (Arg.getOpcode() == ISD::UNDEF) {
2602 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2605 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2606 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2608 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2610 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2614 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2615 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2618 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2619 /// the two vector operands have swapped position.
2621 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2622 MVT MaskVT = Mask.getValueType();
2623 MVT EltVT = MaskVT.getVectorElementType();
2624 unsigned NumElems = Mask.getNumOperands();
2625 SmallVector<SDValue, 8> MaskVec;
2626 for (unsigned i = 0; i != NumElems; ++i) {
2627 SDValue Arg = Mask.getOperand(i);
2628 if (Arg.getOpcode() == ISD::UNDEF) {
2629 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2632 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2633 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2635 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2637 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2639 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2643 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2644 /// match movhlps. The lower half elements should come from upper half of
2645 /// V1 (and in order), and the upper half elements should come from the upper
2646 /// half of V2 (and in order).
2647 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2648 unsigned NumElems = Mask->getNumOperands();
2651 for (unsigned i = 0, e = 2; i != e; ++i)
2652 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2654 for (unsigned i = 2; i != 4; ++i)
2655 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2660 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2661 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2663 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2664 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2665 N = N->getOperand(0).getNode();
2666 if (ISD::isNON_EXTLoad(N)) {
2668 *LD = cast<LoadSDNode>(N);
2675 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2676 /// match movlp{s|d}. The lower half elements should come from lower half of
2677 /// V1 (and in order), and the upper half elements should come from the upper
2678 /// half of V2 (and in order). And since V1 will become the source of the
2679 /// MOVLP, it must be either a vector load or a scalar load to vector.
2680 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2681 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2683 // Is V2 is a vector load, don't do this transformation. We will try to use
2684 // load folding shufps op.
2685 if (ISD::isNON_EXTLoad(V2))
2688 unsigned NumElems = Mask->getNumOperands();
2689 if (NumElems != 2 && NumElems != 4)
2691 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2692 if (!isUndefOrEqual(Mask->getOperand(i), i))
2694 for (unsigned i = NumElems/2; i != NumElems; ++i)
2695 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2700 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2702 static bool isSplatVector(SDNode *N) {
2703 if (N->getOpcode() != ISD::BUILD_VECTOR)
2706 SDValue SplatValue = N->getOperand(0);
2707 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2708 if (N->getOperand(i) != SplatValue)
2713 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2715 static bool isUndefShuffle(SDNode *N) {
2716 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2719 SDValue V1 = N->getOperand(0);
2720 SDValue V2 = N->getOperand(1);
2721 SDValue Mask = N->getOperand(2);
2722 unsigned NumElems = Mask.getNumOperands();
2723 for (unsigned i = 0; i != NumElems; ++i) {
2724 SDValue Arg = Mask.getOperand(i);
2725 if (Arg.getOpcode() != ISD::UNDEF) {
2726 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2727 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2729 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2736 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2738 static inline bool isZeroNode(SDValue Elt) {
2739 return ((isa<ConstantSDNode>(Elt) &&
2740 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2741 (isa<ConstantFPSDNode>(Elt) &&
2742 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2745 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2746 /// to an zero vector.
2747 static bool isZeroShuffle(SDNode *N) {
2748 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2751 SDValue V1 = N->getOperand(0);
2752 SDValue V2 = N->getOperand(1);
2753 SDValue Mask = N->getOperand(2);
2754 unsigned NumElems = Mask.getNumOperands();
2755 for (unsigned i = 0; i != NumElems; ++i) {
2756 SDValue Arg = Mask.getOperand(i);
2757 if (Arg.getOpcode() == ISD::UNDEF)
2760 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2761 if (Idx < NumElems) {
2762 unsigned Opc = V1.getNode()->getOpcode();
2763 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2765 if (Opc != ISD::BUILD_VECTOR ||
2766 !isZeroNode(V1.getNode()->getOperand(Idx)))
2768 } else if (Idx >= NumElems) {
2769 unsigned Opc = V2.getNode()->getOpcode();
2770 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2772 if (Opc != ISD::BUILD_VECTOR ||
2773 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2780 /// getZeroVector - Returns a vector of specified type with all zero elements.
2782 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2783 assert(VT.isVector() && "Expected a vector type");
2785 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2786 // type. This ensures they get CSE'd.
2788 if (VT.getSizeInBits() == 64) { // MMX
2789 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2790 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2791 } else if (HasSSE2) { // SSE2
2792 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2793 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2795 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2796 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2798 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2801 /// getOnesVector - Returns a vector of specified type with all bits set.
2803 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2804 assert(VT.isVector() && "Expected a vector type");
2806 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2807 // type. This ensures they get CSE'd.
2808 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2810 if (VT.getSizeInBits() == 64) // MMX
2811 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2813 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2814 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2818 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2819 /// that point to V2 points to its first element.
2820 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2821 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2823 bool Changed = false;
2824 SmallVector<SDValue, 8> MaskVec;
2825 unsigned NumElems = Mask.getNumOperands();
2826 for (unsigned i = 0; i != NumElems; ++i) {
2827 SDValue Arg = Mask.getOperand(i);
2828 if (Arg.getOpcode() != ISD::UNDEF) {
2829 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2830 if (Val > NumElems) {
2831 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2835 MaskVec.push_back(Arg);
2839 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2840 &MaskVec[0], MaskVec.size());
2844 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2845 /// operation of specified width.
2846 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2847 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2848 MVT BaseVT = MaskVT.getVectorElementType();
2850 SmallVector<SDValue, 8> MaskVec;
2851 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2852 for (unsigned i = 1; i != NumElems; ++i)
2853 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2854 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2857 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2858 /// of specified width.
2859 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2860 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2861 MVT BaseVT = MaskVT.getVectorElementType();
2862 SmallVector<SDValue, 8> MaskVec;
2863 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2864 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2865 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2867 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2870 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2871 /// of specified width.
2872 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2873 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2874 MVT BaseVT = MaskVT.getVectorElementType();
2875 unsigned Half = NumElems/2;
2876 SmallVector<SDValue, 8> MaskVec;
2877 for (unsigned i = 0; i != Half; ++i) {
2878 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2879 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2881 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2884 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2885 /// element #0 of a vector with the specified index, leaving the rest of the
2886 /// elements in place.
2887 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2888 SelectionDAG &DAG) {
2889 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2890 MVT BaseVT = MaskVT.getVectorElementType();
2891 SmallVector<SDValue, 8> MaskVec;
2892 // Element #0 of the result gets the elt we are replacing.
2893 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2894 for (unsigned i = 1; i != NumElems; ++i)
2895 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2896 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2899 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2900 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2901 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2902 MVT VT = Op.getValueType();
2905 SDValue V1 = Op.getOperand(0);
2906 SDValue Mask = Op.getOperand(2);
2907 unsigned NumElems = Mask.getNumOperands();
2908 // Special handling of v4f32 -> v4i32.
2909 if (VT != MVT::v4f32) {
2910 Mask = getUnpacklMask(NumElems, DAG);
2911 while (NumElems > 4) {
2912 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2915 Mask = getZeroVector(MVT::v4i32, true, DAG);
2918 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2919 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2920 DAG.getNode(ISD::UNDEF, PVT), Mask);
2921 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2924 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2925 /// vector of zero or undef vector. This produces a shuffle where the low
2926 /// element of V2 is swizzled into the zero/undef vector, landing at element
2927 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2928 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2929 bool isZero, bool HasSSE2,
2930 SelectionDAG &DAG) {
2931 MVT VT = V2.getValueType();
2933 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
2934 unsigned NumElems = V2.getValueType().getVectorNumElements();
2935 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2936 MVT EVT = MaskVT.getVectorElementType();
2937 SmallVector<SDValue, 16> MaskVec;
2938 for (unsigned i = 0; i != NumElems; ++i)
2939 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2940 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2942 MaskVec.push_back(DAG.getConstant(i, EVT));
2943 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2944 &MaskVec[0], MaskVec.size());
2945 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2948 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2949 /// a shuffle that is zero.
2951 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
2952 unsigned NumElems, bool Low,
2953 SelectionDAG &DAG) {
2954 unsigned NumZeros = 0;
2955 for (unsigned i = 0; i < NumElems; ++i) {
2956 unsigned Index = Low ? i : NumElems-i-1;
2957 SDValue Idx = Mask.getOperand(Index);
2958 if (Idx.getOpcode() == ISD::UNDEF) {
2962 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
2963 if (Elt.getNode() && isZeroNode(Elt))
2971 /// isVectorShift - Returns true if the shuffle can be implemented as a
2972 /// logical left or right shift of a vector.
2973 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
2974 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
2975 unsigned NumElems = Mask.getNumOperands();
2978 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
2981 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
2986 bool SeenV1 = false;
2987 bool SeenV2 = false;
2988 for (unsigned i = NumZeros; i < NumElems; ++i) {
2989 unsigned Val = isLeft ? (i - NumZeros) : i;
2990 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
2991 if (Idx.getOpcode() == ISD::UNDEF)
2993 unsigned Index = cast<ConstantSDNode>(Idx)->getValue();
2994 if (Index < NumElems)
3003 if (SeenV1 && SeenV2)
3006 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3012 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3014 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3015 unsigned NumNonZero, unsigned NumZero,
3016 SelectionDAG &DAG, TargetLowering &TLI) {
3022 for (unsigned i = 0; i < 16; ++i) {
3023 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3024 if (ThisIsNonZero && First) {
3026 V = getZeroVector(MVT::v8i16, true, DAG);
3028 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3033 SDValue ThisElt(0, 0), LastElt(0, 0);
3034 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3035 if (LastIsNonZero) {
3036 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3038 if (ThisIsNonZero) {
3039 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3040 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3041 ThisElt, DAG.getConstant(8, MVT::i8));
3043 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3047 if (ThisElt.getNode())
3048 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3049 DAG.getIntPtrConstant(i/2));
3053 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3056 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3058 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3059 unsigned NumNonZero, unsigned NumZero,
3060 SelectionDAG &DAG, TargetLowering &TLI) {
3066 for (unsigned i = 0; i < 8; ++i) {
3067 bool isNonZero = (NonZeros & (1 << i)) != 0;
3071 V = getZeroVector(MVT::v8i16, true, DAG);
3073 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3076 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3077 DAG.getIntPtrConstant(i));
3084 /// getVShift - Return a vector logical shift node.
3086 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3087 unsigned NumBits, SelectionDAG &DAG,
3088 const TargetLowering &TLI) {
3089 bool isMMX = VT.getSizeInBits() == 64;
3090 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3091 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3092 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3093 return DAG.getNode(ISD::BIT_CONVERT, VT,
3094 DAG.getNode(Opc, ShVT, SrcOp,
3095 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3099 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3100 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3101 if (ISD::isBuildVectorAllZeros(Op.getNode())
3102 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3103 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3104 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3105 // eliminated on x86-32 hosts.
3106 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3109 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3110 return getOnesVector(Op.getValueType(), DAG);
3111 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3114 MVT VT = Op.getValueType();
3115 MVT EVT = VT.getVectorElementType();
3116 unsigned EVTBits = EVT.getSizeInBits();
3118 unsigned NumElems = Op.getNumOperands();
3119 unsigned NumZero = 0;
3120 unsigned NumNonZero = 0;
3121 unsigned NonZeros = 0;
3122 bool IsAllConstants = true;
3123 SmallSet<SDValue, 8> Values;
3124 for (unsigned i = 0; i < NumElems; ++i) {
3125 SDValue Elt = Op.getOperand(i);
3126 if (Elt.getOpcode() == ISD::UNDEF)
3129 if (Elt.getOpcode() != ISD::Constant &&
3130 Elt.getOpcode() != ISD::ConstantFP)
3131 IsAllConstants = false;
3132 if (isZeroNode(Elt))
3135 NonZeros |= (1 << i);
3140 if (NumNonZero == 0) {
3141 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3142 return DAG.getNode(ISD::UNDEF, VT);
3145 // Special case for single non-zero, non-undef, element.
3146 if (NumNonZero == 1 && NumElems <= 4) {
3147 unsigned Idx = CountTrailingZeros_32(NonZeros);
3148 SDValue Item = Op.getOperand(Idx);
3150 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3151 // the value are obviously zero, truncate the value to i32 and do the
3152 // insertion that way. Only do this if the value is non-constant or if the
3153 // value is a constant being inserted into element 0. It is cheaper to do
3154 // a constant pool load than it is to do a movd + shuffle.
3155 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3156 (!IsAllConstants || Idx == 0)) {
3157 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3158 // Handle MMX and SSE both.
3159 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3160 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3162 // Truncate the value (which may itself be a constant) to i32, and
3163 // convert it to a vector with movd (S2V+shuffle to zero extend).
3164 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3165 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3166 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3167 Subtarget->hasSSE2(), DAG);
3169 // Now we have our 32-bit value zero extended in the low element of
3170 // a vector. If Idx != 0, swizzle it into place.
3173 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3174 getSwapEltZeroMask(VecElts, Idx, DAG)
3176 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3178 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3182 // If we have a constant or non-constant insertion into the low element of
3183 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3184 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3185 // depending on what the source datatype is. Because we can only get here
3186 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3188 // Don't do this for i64 values on x86-32.
3189 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3190 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3191 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3192 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3193 Subtarget->hasSSE2(), DAG);
3196 // Is it a vector logical left shift?
3197 if (NumElems == 2 && Idx == 1 &&
3198 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3199 unsigned NumBits = VT.getSizeInBits();
3200 return getVShift(true, VT,
3201 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3202 NumBits/2, DAG, *this);
3205 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3208 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3209 // is a non-constant being inserted into an element other than the low one,
3210 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3211 // movd/movss) to move this into the low element, then shuffle it into
3213 if (EVTBits == 32) {
3214 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3216 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3217 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3218 Subtarget->hasSSE2(), DAG);
3219 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3220 MVT MaskEVT = MaskVT.getVectorElementType();
3221 SmallVector<SDValue, 8> MaskVec;
3222 for (unsigned i = 0; i < NumElems; i++)
3223 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3224 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3225 &MaskVec[0], MaskVec.size());
3226 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3227 DAG.getNode(ISD::UNDEF, VT), Mask);
3231 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3232 if (Values.size() == 1)
3235 // A vector full of immediates; various special cases are already
3236 // handled, so this is best done with a single constant-pool load.
3240 // Let legalizer expand 2-wide build_vectors.
3241 if (EVTBits == 64) {
3242 if (NumNonZero == 1) {
3243 // One half is zero or undef.
3244 unsigned Idx = CountTrailingZeros_32(NonZeros);
3245 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3246 Op.getOperand(Idx));
3247 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3248 Subtarget->hasSSE2(), DAG);
3253 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3254 if (EVTBits == 8 && NumElems == 16) {
3255 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3257 if (V.getNode()) return V;
3260 if (EVTBits == 16 && NumElems == 8) {
3261 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3263 if (V.getNode()) return V;
3266 // If element VT is == 32 bits, turn it into a number of shuffles.
3267 SmallVector<SDValue, 8> V;
3269 if (NumElems == 4 && NumZero > 0) {
3270 for (unsigned i = 0; i < 4; ++i) {
3271 bool isZero = !(NonZeros & (1 << i));
3273 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3275 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3278 for (unsigned i = 0; i < 2; ++i) {
3279 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3282 V[i] = V[i*2]; // Must be a zero vector.
3285 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3286 getMOVLMask(NumElems, DAG));
3289 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3290 getMOVLMask(NumElems, DAG));
3293 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3294 getUnpacklMask(NumElems, DAG));
3299 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3300 MVT EVT = MaskVT.getVectorElementType();
3301 SmallVector<SDValue, 8> MaskVec;
3302 bool Reverse = (NonZeros & 0x3) == 2;
3303 for (unsigned i = 0; i < 2; ++i)
3305 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3307 MaskVec.push_back(DAG.getConstant(i, EVT));
3308 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3309 for (unsigned i = 0; i < 2; ++i)
3311 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3313 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3314 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3315 &MaskVec[0], MaskVec.size());
3316 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3319 if (Values.size() > 2) {
3320 // Expand into a number of unpckl*.
3322 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3323 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3324 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3325 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3326 for (unsigned i = 0; i < NumElems; ++i)
3327 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3329 while (NumElems != 0) {
3330 for (unsigned i = 0; i < NumElems; ++i)
3331 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3342 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3343 SDValue PermMask, SelectionDAG &DAG,
3344 TargetLowering &TLI) {
3346 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3347 MVT MaskEVT = MaskVT.getVectorElementType();
3348 MVT PtrVT = TLI.getPointerTy();
3349 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3350 PermMask.getNode()->op_end());
3352 // First record which half of which vector the low elements come from.
3353 SmallVector<unsigned, 4> LowQuad(4);
3354 for (unsigned i = 0; i < 4; ++i) {
3355 SDValue Elt = MaskElts[i];
3356 if (Elt.getOpcode() == ISD::UNDEF)
3358 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3359 int QuadIdx = EltIdx / 4;
3363 int BestLowQuad = -1;
3364 unsigned MaxQuad = 1;
3365 for (unsigned i = 0; i < 4; ++i) {
3366 if (LowQuad[i] > MaxQuad) {
3368 MaxQuad = LowQuad[i];
3372 // Record which half of which vector the high elements come from.
3373 SmallVector<unsigned, 4> HighQuad(4);
3374 for (unsigned i = 4; i < 8; ++i) {
3375 SDValue Elt = MaskElts[i];
3376 if (Elt.getOpcode() == ISD::UNDEF)
3378 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3379 int QuadIdx = EltIdx / 4;
3380 ++HighQuad[QuadIdx];
3383 int BestHighQuad = -1;
3385 for (unsigned i = 0; i < 4; ++i) {
3386 if (HighQuad[i] > MaxQuad) {
3388 MaxQuad = HighQuad[i];
3392 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3393 if (BestLowQuad != -1 || BestHighQuad != -1) {
3394 // First sort the 4 chunks in order using shufpd.
3395 SmallVector<SDValue, 8> MaskVec;
3397 if (BestLowQuad != -1)
3398 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3400 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3402 if (BestHighQuad != -1)
3403 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3405 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3407 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3408 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3409 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3410 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3411 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3413 // Now sort high and low parts separately.
3414 BitVector InOrder(8);
3415 if (BestLowQuad != -1) {
3416 // Sort lower half in order using PSHUFLW.
3418 bool AnyOutOrder = false;
3420 for (unsigned i = 0; i != 4; ++i) {
3421 SDValue Elt = MaskElts[i];
3422 if (Elt.getOpcode() == ISD::UNDEF) {
3423 MaskVec.push_back(Elt);
3426 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3430 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3432 // If this element is in the right place after this shuffle, then
3434 if ((int)(EltIdx / 4) == BestLowQuad)
3439 for (unsigned i = 4; i != 8; ++i)
3440 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3441 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3442 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3446 if (BestHighQuad != -1) {
3447 // Sort high half in order using PSHUFHW if possible.
3450 for (unsigned i = 0; i != 4; ++i)
3451 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3453 bool AnyOutOrder = false;
3454 for (unsigned i = 4; i != 8; ++i) {
3455 SDValue Elt = MaskElts[i];
3456 if (Elt.getOpcode() == ISD::UNDEF) {
3457 MaskVec.push_back(Elt);
3460 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3464 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3466 // If this element is in the right place after this shuffle, then
3468 if ((int)(EltIdx / 4) == BestHighQuad)
3474 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3475 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3479 // The other elements are put in the right place using pextrw and pinsrw.
3480 for (unsigned i = 0; i != 8; ++i) {
3483 SDValue Elt = MaskElts[i];
3484 if (Elt.getOpcode() == ISD::UNDEF)
3486 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3487 SDValue ExtOp = (EltIdx < 8)
3488 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3489 DAG.getConstant(EltIdx, PtrVT))
3490 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3491 DAG.getConstant(EltIdx - 8, PtrVT));
3492 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3493 DAG.getConstant(i, PtrVT));
3499 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3500 // few as possible. First, let's find out how many elements are already in the
3502 unsigned V1InOrder = 0;
3503 unsigned V1FromV1 = 0;
3504 unsigned V2InOrder = 0;
3505 unsigned V2FromV2 = 0;
3506 SmallVector<SDValue, 8> V1Elts;
3507 SmallVector<SDValue, 8> V2Elts;
3508 for (unsigned i = 0; i < 8; ++i) {
3509 SDValue Elt = MaskElts[i];
3510 if (Elt.getOpcode() == ISD::UNDEF) {
3511 V1Elts.push_back(Elt);
3512 V2Elts.push_back(Elt);
3517 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3519 V1Elts.push_back(Elt);
3520 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3522 } else if (EltIdx == i+8) {
3523 V1Elts.push_back(Elt);
3524 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3526 } else if (EltIdx < 8) {
3527 V1Elts.push_back(Elt);
3530 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3535 if (V2InOrder > V1InOrder) {
3536 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3538 std::swap(V1Elts, V2Elts);
3539 std::swap(V1FromV1, V2FromV2);
3542 if ((V1FromV1 + V1InOrder) != 8) {
3543 // Some elements are from V2.
3545 // If there are elements that are from V1 but out of place,
3546 // then first sort them in place
3547 SmallVector<SDValue, 8> MaskVec;
3548 for (unsigned i = 0; i < 8; ++i) {
3549 SDValue Elt = V1Elts[i];
3550 if (Elt.getOpcode() == ISD::UNDEF) {
3551 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3554 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3556 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3558 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3560 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3561 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3565 for (unsigned i = 0; i < 8; ++i) {
3566 SDValue Elt = V1Elts[i];
3567 if (Elt.getOpcode() == ISD::UNDEF)
3569 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3572 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3573 DAG.getConstant(EltIdx - 8, PtrVT));
3574 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3575 DAG.getConstant(i, PtrVT));
3579 // All elements are from V1.
3581 for (unsigned i = 0; i < 8; ++i) {
3582 SDValue Elt = V1Elts[i];
3583 if (Elt.getOpcode() == ISD::UNDEF)
3585 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3586 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3587 DAG.getConstant(EltIdx, PtrVT));
3588 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3589 DAG.getConstant(i, PtrVT));
3595 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3596 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3597 /// done when every pair / quad of shuffle mask elements point to elements in
3598 /// the right sequence. e.g.
3599 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3601 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3603 SDValue PermMask, SelectionDAG &DAG,
3604 TargetLowering &TLI) {
3605 unsigned NumElems = PermMask.getNumOperands();
3606 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3607 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3608 MVT MaskEltVT = MaskVT.getVectorElementType();
3610 switch (VT.getSimpleVT()) {
3611 default: assert(false && "Unexpected!");
3612 case MVT::v4f32: NewVT = MVT::v2f64; break;
3613 case MVT::v4i32: NewVT = MVT::v2i64; break;
3614 case MVT::v8i16: NewVT = MVT::v4i32; break;
3615 case MVT::v16i8: NewVT = MVT::v4i32; break;
3618 if (NewWidth == 2) {
3624 unsigned Scale = NumElems / NewWidth;
3625 SmallVector<SDValue, 8> MaskVec;
3626 for (unsigned i = 0; i < NumElems; i += Scale) {
3627 unsigned StartIdx = ~0U;
3628 for (unsigned j = 0; j < Scale; ++j) {
3629 SDValue Elt = PermMask.getOperand(i+j);
3630 if (Elt.getOpcode() == ISD::UNDEF)
3632 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3633 if (StartIdx == ~0U)
3634 StartIdx = EltIdx - (EltIdx % Scale);
3635 if (EltIdx != StartIdx + j)
3638 if (StartIdx == ~0U)
3639 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3641 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3644 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3645 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3646 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3647 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3648 &MaskVec[0], MaskVec.size()));
3651 /// getVZextMovL - Return a zero-extending vector move low node.
3653 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3654 SDValue SrcOp, SelectionDAG &DAG,
3655 const X86Subtarget *Subtarget) {
3656 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3657 LoadSDNode *LD = NULL;
3658 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3659 LD = dyn_cast<LoadSDNode>(SrcOp);
3661 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3663 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3664 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3665 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3666 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3667 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3669 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3670 return DAG.getNode(ISD::BIT_CONVERT, VT,
3671 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3672 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3679 return DAG.getNode(ISD::BIT_CONVERT, VT,
3680 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3681 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3684 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3687 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3688 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3689 MVT MaskVT = PermMask.getValueType();
3690 MVT MaskEVT = MaskVT.getVectorElementType();
3691 SmallVector<std::pair<int, int>, 8> Locs;
3693 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3696 for (unsigned i = 0; i != 4; ++i) {
3697 SDValue Elt = PermMask.getOperand(i);
3698 if (Elt.getOpcode() == ISD::UNDEF) {
3699 Locs[i] = std::make_pair(-1, -1);
3701 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3702 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3704 Locs[i] = std::make_pair(0, NumLo);
3708 Locs[i] = std::make_pair(1, NumHi);
3710 Mask1[2+NumHi] = Elt;
3716 if (NumLo <= 2 && NumHi <= 2) {
3717 // If no more than two elements come from either vector. This can be
3718 // implemented with two shuffles. First shuffle gather the elements.
3719 // The second shuffle, which takes the first shuffle as both of its
3720 // vector operands, put the elements into the right order.
3721 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3722 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3723 &Mask1[0], Mask1.size()));
3725 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3726 for (unsigned i = 0; i != 4; ++i) {
3727 if (Locs[i].first == -1)
3730 unsigned Idx = (i < 2) ? 0 : 4;
3731 Idx += Locs[i].first * 2 + Locs[i].second;
3732 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3736 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3737 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3738 &Mask2[0], Mask2.size()));
3739 } else if (NumLo == 3 || NumHi == 3) {
3740 // Otherwise, we must have three elements from one vector, call it X, and
3741 // one element from the other, call it Y. First, use a shufps to build an
3742 // intermediate vector with the one element from Y and the element from X
3743 // that will be in the same half in the final destination (the indexes don't
3744 // matter). Then, use a shufps to build the final vector, taking the half
3745 // containing the element from Y from the intermediate, and the other half
3748 // Normalize it so the 3 elements come from V1.
3749 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3753 // Find the element from V2.
3755 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3756 SDValue Elt = PermMask.getOperand(HiIndex);
3757 if (Elt.getOpcode() == ISD::UNDEF)
3759 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3764 Mask1[0] = PermMask.getOperand(HiIndex);
3765 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3766 Mask1[2] = PermMask.getOperand(HiIndex^1);
3767 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3768 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3769 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3772 Mask1[0] = PermMask.getOperand(0);
3773 Mask1[1] = PermMask.getOperand(1);
3774 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3775 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3776 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3777 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3779 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3780 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3781 Mask1[2] = PermMask.getOperand(2);
3782 Mask1[3] = PermMask.getOperand(3);
3783 if (Mask1[2].getOpcode() != ISD::UNDEF)
3784 Mask1[2] = DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getValue()+4,
3786 if (Mask1[3].getOpcode() != ISD::UNDEF)
3787 Mask1[3] = DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getValue()+4,
3789 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3790 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3794 // Break it into (shuffle shuffle_hi, shuffle_lo).
3796 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3797 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3798 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3799 unsigned MaskIdx = 0;
3802 for (unsigned i = 0; i != 4; ++i) {
3809 SDValue Elt = PermMask.getOperand(i);
3810 if (Elt.getOpcode() == ISD::UNDEF) {
3811 Locs[i] = std::make_pair(-1, -1);
3812 } else if (cast<ConstantSDNode>(Elt)->getValue() < 4) {
3813 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3814 (*MaskPtr)[LoIdx] = Elt;
3817 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3818 (*MaskPtr)[HiIdx] = Elt;
3823 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3824 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3825 &LoMask[0], LoMask.size()));
3826 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3827 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3828 &HiMask[0], HiMask.size()));
3829 SmallVector<SDValue, 8> MaskOps;
3830 for (unsigned i = 0; i != 4; ++i) {
3831 if (Locs[i].first == -1) {
3832 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3834 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3835 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3838 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3839 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3840 &MaskOps[0], MaskOps.size()));
3844 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3845 SDValue V1 = Op.getOperand(0);
3846 SDValue V2 = Op.getOperand(1);
3847 SDValue PermMask = Op.getOperand(2);
3848 MVT VT = Op.getValueType();
3849 unsigned NumElems = PermMask.getNumOperands();
3850 bool isMMX = VT.getSizeInBits() == 64;
3851 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3852 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3853 bool V1IsSplat = false;
3854 bool V2IsSplat = false;
3856 if (isUndefShuffle(Op.getNode()))
3857 return DAG.getNode(ISD::UNDEF, VT);
3859 if (isZeroShuffle(Op.getNode()))
3860 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3862 if (isIdentityMask(PermMask.getNode()))
3864 else if (isIdentityMask(PermMask.getNode(), true))
3867 if (isSplatMask(PermMask.getNode())) {
3868 if (isMMX || NumElems < 4) return Op;
3869 // Promote it to a v4{if}32 splat.
3870 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3873 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3875 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3876 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3877 if (NewOp.getNode())
3878 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3879 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3880 // FIXME: Figure out a cleaner way to do this.
3881 // Try to make use of movq to zero out the top part.
3882 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3883 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3885 if (NewOp.getNode()) {
3886 SDValue NewV1 = NewOp.getOperand(0);
3887 SDValue NewV2 = NewOp.getOperand(1);
3888 SDValue NewMask = NewOp.getOperand(2);
3889 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3890 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3891 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3894 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3895 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3897 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3898 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3903 // Check if this can be converted into a logical shift.
3904 bool isLeft = false;
3907 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
3908 if (isShift && ShVal.hasOneUse()) {
3909 // If the shifted value has multiple uses, it may be cheaper to use
3910 // v_set0 + movlhps or movhlps, etc.
3911 MVT EVT = VT.getVectorElementType();
3912 ShAmt *= EVT.getSizeInBits();
3913 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3916 if (X86::isMOVLMask(PermMask.getNode())) {
3919 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3920 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
3925 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
3926 X86::isMOVSLDUPMask(PermMask.getNode()) ||
3927 X86::isMOVHLPSMask(PermMask.getNode()) ||
3928 X86::isMOVHPMask(PermMask.getNode()) ||
3929 X86::isMOVLPMask(PermMask.getNode())))
3932 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
3933 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
3934 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3937 // No better options. Use a vshl / vsrl.
3938 MVT EVT = VT.getVectorElementType();
3939 ShAmt *= EVT.getSizeInBits();
3940 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3943 bool Commuted = false;
3944 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3945 // 1,1,1,1 -> v8i16 though.
3946 V1IsSplat = isSplatVector(V1.getNode());
3947 V2IsSplat = isSplatVector(V2.getNode());
3949 // Canonicalize the splat or undef, if present, to be on the RHS.
3950 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3951 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3952 std::swap(V1IsSplat, V2IsSplat);
3953 std::swap(V1IsUndef, V2IsUndef);
3957 // FIXME: Figure out a cleaner way to do this.
3958 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
3959 if (V2IsUndef) return V1;
3960 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3962 // V2 is a splat, so the mask may be malformed. That is, it may point
3963 // to any V2 element. The instruction selectior won't like this. Get
3964 // a corrected mask and commute to form a proper MOVS{S|D}.
3965 SDValue NewMask = getMOVLMask(NumElems, DAG);
3966 if (NewMask.getNode() != PermMask.getNode())
3967 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3972 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
3973 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
3974 X86::isUNPCKLMask(PermMask.getNode()) ||
3975 X86::isUNPCKHMask(PermMask.getNode()))
3979 // Normalize mask so all entries that point to V2 points to its first
3980 // element then try to match unpck{h|l} again. If match, return a
3981 // new vector_shuffle with the corrected mask.
3982 SDValue NewMask = NormalizeMask(PermMask, DAG);
3983 if (NewMask.getNode() != PermMask.getNode()) {
3984 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
3985 SDValue NewMask = getUnpacklMask(NumElems, DAG);
3986 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3987 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
3988 SDValue NewMask = getUnpackhMask(NumElems, DAG);
3989 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3994 // Normalize the node to match x86 shuffle ops if needed
3995 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
3996 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3999 // Commute is back and try unpck* again.
4000 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4001 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4002 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4003 X86::isUNPCKLMask(PermMask.getNode()) ||
4004 X86::isUNPCKHMask(PermMask.getNode()))
4008 // Try PSHUF* first, then SHUFP*.
4009 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4010 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4011 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4012 if (V2.getOpcode() != ISD::UNDEF)
4013 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4014 DAG.getNode(ISD::UNDEF, VT), PermMask);
4019 if (Subtarget->hasSSE2() &&
4020 (X86::isPSHUFDMask(PermMask.getNode()) ||
4021 X86::isPSHUFHWMask(PermMask.getNode()) ||
4022 X86::isPSHUFLWMask(PermMask.getNode()))) {
4024 if (VT == MVT::v4f32) {
4026 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4027 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4028 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4029 } else if (V2.getOpcode() != ISD::UNDEF)
4030 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4031 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4033 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4037 // Binary or unary shufps.
4038 if (X86::isSHUFPMask(PermMask.getNode()) ||
4039 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4043 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4044 if (VT == MVT::v8i16) {
4045 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4046 if (NewOp.getNode())
4050 // Handle all 4 wide cases with a number of shuffles except for MMX.
4051 if (NumElems == 4 && !isMMX)
4052 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4058 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4059 SelectionDAG &DAG) {
4060 MVT VT = Op.getValueType();
4061 if (VT.getSizeInBits() == 8) {
4062 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4063 Op.getOperand(0), Op.getOperand(1));
4064 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4065 DAG.getValueType(VT));
4066 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4067 } else if (VT.getSizeInBits() == 16) {
4068 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4069 Op.getOperand(0), Op.getOperand(1));
4070 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4071 DAG.getValueType(VT));
4072 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4073 } else if (VT == MVT::f32) {
4074 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4075 // the result back to FR32 register. It's only worth matching if the
4076 // result has a single use which is a store or a bitcast to i32.
4077 if (!Op.hasOneUse())
4079 SDNode *User = *Op.getNode()->use_begin();
4080 if (User->getOpcode() != ISD::STORE &&
4081 (User->getOpcode() != ISD::BIT_CONVERT ||
4082 User->getValueType(0) != MVT::i32))
4084 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4085 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4087 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4094 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4095 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4098 if (Subtarget->hasSSE41()) {
4099 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4104 MVT VT = Op.getValueType();
4105 // TODO: handle v16i8.
4106 if (VT.getSizeInBits() == 16) {
4107 SDValue Vec = Op.getOperand(0);
4108 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4110 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4111 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4112 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4114 // Transform it so it match pextrw which produces a 32-bit result.
4115 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4116 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4117 Op.getOperand(0), Op.getOperand(1));
4118 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4119 DAG.getValueType(VT));
4120 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4121 } else if (VT.getSizeInBits() == 32) {
4122 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4125 // SHUFPS the element to the lowest double word, then movss.
4126 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4127 SmallVector<SDValue, 8> IdxVec;
4129 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4131 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4133 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4135 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4136 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4137 &IdxVec[0], IdxVec.size());
4138 SDValue Vec = Op.getOperand(0);
4139 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4140 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4141 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4142 DAG.getIntPtrConstant(0));
4143 } else if (VT.getSizeInBits() == 64) {
4144 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4145 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4146 // to match extract_elt for f64.
4147 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4151 // UNPCKHPD the element to the lowest double word, then movsd.
4152 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4153 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4154 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4155 SmallVector<SDValue, 8> IdxVec;
4156 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4158 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4159 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4160 &IdxVec[0], IdxVec.size());
4161 SDValue Vec = Op.getOperand(0);
4162 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4163 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4164 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4165 DAG.getIntPtrConstant(0));
4172 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4173 MVT VT = Op.getValueType();
4174 MVT EVT = VT.getVectorElementType();
4176 SDValue N0 = Op.getOperand(0);
4177 SDValue N1 = Op.getOperand(1);
4178 SDValue N2 = Op.getOperand(2);
4180 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4181 isa<ConstantSDNode>(N2)) {
4182 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4184 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4186 if (N1.getValueType() != MVT::i32)
4187 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4188 if (N2.getValueType() != MVT::i32)
4189 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4190 return DAG.getNode(Opc, VT, N0, N1, N2);
4191 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4192 // Bits [7:6] of the constant are the source select. This will always be
4193 // zero here. The DAG Combiner may combine an extract_elt index into these
4194 // bits. For example (insert (extract, 3), 2) could be matched by putting
4195 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4196 // Bits [5:4] of the constant are the destination select. This is the
4197 // value of the incoming immediate.
4198 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4199 // combine either bitwise AND or insert of float 0.0 to set these bits.
4200 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
4201 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4207 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4208 MVT VT = Op.getValueType();
4209 MVT EVT = VT.getVectorElementType();
4211 if (Subtarget->hasSSE41())
4212 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4217 SDValue N0 = Op.getOperand(0);
4218 SDValue N1 = Op.getOperand(1);
4219 SDValue N2 = Op.getOperand(2);
4221 if (EVT.getSizeInBits() == 16) {
4222 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4223 // as its second argument.
4224 if (N1.getValueType() != MVT::i32)
4225 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4226 if (N2.getValueType() != MVT::i32)
4227 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4228 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4234 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4235 if (Op.getValueType() == MVT::v2f32)
4236 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4237 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4238 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4239 Op.getOperand(0))));
4241 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4242 MVT VT = MVT::v2i32;
4243 switch (Op.getValueType().getSimpleVT()) {
4250 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4251 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4254 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4255 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4256 // one of the above mentioned nodes. It has to be wrapped because otherwise
4257 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4258 // be used to form addressing mode. These wrapped nodes will be selected
4261 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4262 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4263 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4265 CP->getAlignment());
4266 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4267 // With PIC, the address is actually $g + Offset.
4268 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4269 !Subtarget->isPICStyleRIPRel()) {
4270 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4271 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4279 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4280 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4281 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4282 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4283 // With PIC, the address is actually $g + Offset.
4284 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4285 !Subtarget->isPICStyleRIPRel()) {
4286 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4287 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4291 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4292 // load the value at address GV, not the value of GV itself. This means that
4293 // the GlobalAddress must be in the base or index register of the address, not
4294 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4295 // The same applies for external symbols during PIC codegen
4296 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4297 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4298 PseudoSourceValue::getGOT(), 0);
4303 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4305 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4308 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4309 DAG.getNode(X86ISD::GlobalBaseReg,
4311 InFlag = Chain.getValue(1);
4313 // emit leal symbol@TLSGD(,%ebx,1), %eax
4314 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4315 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4316 GA->getValueType(0),
4318 SDValue Ops[] = { Chain, TGA, InFlag };
4319 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4320 InFlag = Result.getValue(2);
4321 Chain = Result.getValue(1);
4323 // call ___tls_get_addr. This function receives its argument in
4324 // the register EAX.
4325 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4326 InFlag = Chain.getValue(1);
4328 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4329 SDValue Ops1[] = { Chain,
4330 DAG.getTargetExternalSymbol("___tls_get_addr",
4332 DAG.getRegister(X86::EAX, PtrVT),
4333 DAG.getRegister(X86::EBX, PtrVT),
4335 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4336 InFlag = Chain.getValue(1);
4338 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4341 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4343 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4345 SDValue InFlag, Chain;
4347 // emit leaq symbol@TLSGD(%rip), %rdi
4348 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4349 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4350 GA->getValueType(0),
4352 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4353 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4354 Chain = Result.getValue(1);
4355 InFlag = Result.getValue(2);
4357 // call __tls_get_addr. This function receives its argument in
4358 // the register RDI.
4359 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4360 InFlag = Chain.getValue(1);
4362 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4363 SDValue Ops1[] = { Chain,
4364 DAG.getTargetExternalSymbol("__tls_get_addr",
4366 DAG.getRegister(X86::RDI, PtrVT),
4368 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4369 InFlag = Chain.getValue(1);
4371 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4374 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4375 // "local exec" model.
4376 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4378 // Get the Thread Pointer
4379 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4380 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4382 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4383 GA->getValueType(0),
4385 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4387 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4388 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4389 PseudoSourceValue::getGOT(), 0);
4391 // The address of the thread local variable is the add of the thread
4392 // pointer with the offset of the variable.
4393 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4397 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4398 // TODO: implement the "local dynamic" model
4399 // TODO: implement the "initial exec"model for pic executables
4400 assert(Subtarget->isTargetELF() &&
4401 "TLS not implemented for non-ELF targets");
4402 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4403 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4404 // otherwise use the "Local Exec"TLS Model
4405 if (Subtarget->is64Bit()) {
4406 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4408 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4409 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4411 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4416 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4417 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4418 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4419 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4420 // With PIC, the address is actually $g + Offset.
4421 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4422 !Subtarget->isPICStyleRIPRel()) {
4423 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4424 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4431 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4432 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4433 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4434 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4435 // With PIC, the address is actually $g + Offset.
4436 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4437 !Subtarget->isPICStyleRIPRel()) {
4438 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4439 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4446 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4447 /// take a 2 x i32 value to shift plus a shift amount.
4448 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4449 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4450 MVT VT = Op.getValueType();
4451 unsigned VTBits = VT.getSizeInBits();
4452 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4453 SDValue ShOpLo = Op.getOperand(0);
4454 SDValue ShOpHi = Op.getOperand(1);
4455 SDValue ShAmt = Op.getOperand(2);
4456 SDValue Tmp1 = isSRA ?
4457 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4458 DAG.getConstant(0, VT);
4461 if (Op.getOpcode() == ISD::SHL_PARTS) {
4462 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4463 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4465 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4466 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4469 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4470 DAG.getConstant(VTBits, MVT::i8));
4471 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4472 AndNode, DAG.getConstant(0, MVT::i8));
4475 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4476 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4477 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4479 if (Op.getOpcode() == ISD::SHL_PARTS) {
4480 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4481 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4483 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4484 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4487 SDValue Ops[2] = { Lo, Hi };
4488 return DAG.getMergeValues(Ops, 2);
4491 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4492 MVT SrcVT = Op.getOperand(0).getValueType();
4493 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4494 "Unknown SINT_TO_FP to lower!");
4496 // These are really Legal; caller falls through into that case.
4497 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4499 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4500 Subtarget->is64Bit())
4503 unsigned Size = SrcVT.getSizeInBits()/8;
4504 MachineFunction &MF = DAG.getMachineFunction();
4505 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4506 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4507 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4509 PseudoSourceValue::getFixedStack(SSFI), 0);
4513 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4515 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4517 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4518 SmallVector<SDValue, 8> Ops;
4519 Ops.push_back(Chain);
4520 Ops.push_back(StackSlot);
4521 Ops.push_back(DAG.getValueType(SrcVT));
4522 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4523 Tys, &Ops[0], Ops.size());
4526 Chain = Result.getValue(1);
4527 SDValue InFlag = Result.getValue(2);
4529 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4530 // shouldn't be necessary except that RFP cannot be live across
4531 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4532 MachineFunction &MF = DAG.getMachineFunction();
4533 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4534 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4535 Tys = DAG.getVTList(MVT::Other);
4536 SmallVector<SDValue, 8> Ops;
4537 Ops.push_back(Chain);
4538 Ops.push_back(Result);
4539 Ops.push_back(StackSlot);
4540 Ops.push_back(DAG.getValueType(Op.getValueType()));
4541 Ops.push_back(InFlag);
4542 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4543 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4544 PseudoSourceValue::getFixedStack(SSFI), 0);
4550 std::pair<SDValue,SDValue> X86TargetLowering::
4551 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4552 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4553 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4554 "Unknown FP_TO_SINT to lower!");
4556 // These are really Legal.
4557 if (Op.getValueType() == MVT::i32 &&
4558 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4559 return std::make_pair(SDValue(), SDValue());
4560 if (Subtarget->is64Bit() &&
4561 Op.getValueType() == MVT::i64 &&
4562 Op.getOperand(0).getValueType() != MVT::f80)
4563 return std::make_pair(SDValue(), SDValue());
4565 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4567 MachineFunction &MF = DAG.getMachineFunction();
4568 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4569 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4570 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4572 switch (Op.getValueType().getSimpleVT()) {
4573 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4574 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4575 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4576 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4579 SDValue Chain = DAG.getEntryNode();
4580 SDValue Value = Op.getOperand(0);
4581 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4582 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4583 Chain = DAG.getStore(Chain, Value, StackSlot,
4584 PseudoSourceValue::getFixedStack(SSFI), 0);
4585 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4587 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4589 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4590 Chain = Value.getValue(1);
4591 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4592 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4595 // Build the FP_TO_INT*_IN_MEM
4596 SDValue Ops[] = { Chain, Value, StackSlot };
4597 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4599 return std::make_pair(FIST, StackSlot);
4602 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4603 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4604 SDValue FIST = Vals.first, StackSlot = Vals.second;
4605 if (FIST.getNode() == 0) return SDValue();
4608 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4611 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4612 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4613 SDValue FIST = Vals.first, StackSlot = Vals.second;
4614 if (FIST.getNode() == 0) return 0;
4616 MVT VT = N->getValueType(0);
4618 // Return a load from the stack slot.
4619 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4621 // Use MERGE_VALUES to drop the chain result value and get a node with one
4622 // result. This requires turning off getMergeValues simplification, since
4623 // otherwise it will give us Res back.
4624 return DAG.getMergeValues(&Res, 1, false).getNode();
4627 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4628 MVT VT = Op.getValueType();
4631 EltVT = VT.getVectorElementType();
4632 std::vector<Constant*> CV;
4633 if (EltVT == MVT::f64) {
4634 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4638 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4644 Constant *C = ConstantVector::get(CV);
4645 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4646 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4647 PseudoSourceValue::getConstantPool(), 0,
4649 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4652 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4653 MVT VT = Op.getValueType();
4655 unsigned EltNum = 1;
4656 if (VT.isVector()) {
4657 EltVT = VT.getVectorElementType();
4658 EltNum = VT.getVectorNumElements();
4660 std::vector<Constant*> CV;
4661 if (EltVT == MVT::f64) {
4662 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4666 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4672 Constant *C = ConstantVector::get(CV);
4673 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4674 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4675 PseudoSourceValue::getConstantPool(), 0,
4677 if (VT.isVector()) {
4678 return DAG.getNode(ISD::BIT_CONVERT, VT,
4679 DAG.getNode(ISD::XOR, MVT::v2i64,
4680 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4681 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4683 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4687 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4688 SDValue Op0 = Op.getOperand(0);
4689 SDValue Op1 = Op.getOperand(1);
4690 MVT VT = Op.getValueType();
4691 MVT SrcVT = Op1.getValueType();
4693 // If second operand is smaller, extend it first.
4694 if (SrcVT.bitsLT(VT)) {
4695 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4698 // And if it is bigger, shrink it first.
4699 if (SrcVT.bitsGT(VT)) {
4700 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4704 // At this point the operands and the result should have the same
4705 // type, and that won't be f80 since that is not custom lowered.
4707 // First get the sign bit of second operand.
4708 std::vector<Constant*> CV;
4709 if (SrcVT == MVT::f64) {
4710 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4711 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4713 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4714 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4715 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4716 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4718 Constant *C = ConstantVector::get(CV);
4719 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4720 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4721 PseudoSourceValue::getConstantPool(), 0,
4723 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4725 // Shift sign bit right or left if the two operands have different types.
4726 if (SrcVT.bitsGT(VT)) {
4727 // Op0 is MVT::f32, Op1 is MVT::f64.
4728 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4729 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4730 DAG.getConstant(32, MVT::i32));
4731 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4732 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4733 DAG.getIntPtrConstant(0));
4736 // Clear first operand sign bit.
4738 if (VT == MVT::f64) {
4739 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4740 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4742 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4743 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4744 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4745 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4747 C = ConstantVector::get(CV);
4748 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4749 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4750 PseudoSourceValue::getConstantPool(), 0,
4752 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4754 // Or the value with the sign bit.
4755 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4758 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4759 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4761 SDValue Op0 = Op.getOperand(0);
4762 SDValue Op1 = Op.getOperand(1);
4763 SDValue CC = Op.getOperand(2);
4764 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4765 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4768 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4770 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4771 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4772 DAG.getConstant(X86CC, MVT::i8), Cond);
4775 assert(isFP && "Illegal integer SetCC!");
4777 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4778 switch (SetCCOpcode) {
4779 default: assert(false && "Illegal floating point SetCC!");
4780 case ISD::SETOEQ: { // !PF & ZF
4781 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4782 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4783 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4784 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4785 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4787 case ISD::SETUNE: { // PF | !ZF
4788 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4789 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4790 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4791 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4792 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4797 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4799 SDValue Op0 = Op.getOperand(0);
4800 SDValue Op1 = Op.getOperand(1);
4801 SDValue CC = Op.getOperand(2);
4802 MVT VT = Op.getValueType();
4803 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4804 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4808 MVT VT0 = Op0.getValueType();
4809 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4810 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4813 switch (SetCCOpcode) {
4816 case ISD::SETEQ: SSECC = 0; break;
4818 case ISD::SETGT: Swap = true; // Fallthrough
4820 case ISD::SETOLT: SSECC = 1; break;
4822 case ISD::SETGE: Swap = true; // Fallthrough
4824 case ISD::SETOLE: SSECC = 2; break;
4825 case ISD::SETUO: SSECC = 3; break;
4827 case ISD::SETNE: SSECC = 4; break;
4828 case ISD::SETULE: Swap = true;
4829 case ISD::SETUGE: SSECC = 5; break;
4830 case ISD::SETULT: Swap = true;
4831 case ISD::SETUGT: SSECC = 6; break;
4832 case ISD::SETO: SSECC = 7; break;
4835 std::swap(Op0, Op1);
4837 // In the two special cases we can't handle, emit two comparisons.
4839 if (SetCCOpcode == ISD::SETUEQ) {
4841 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4842 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4843 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4845 else if (SetCCOpcode == ISD::SETONE) {
4847 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4848 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4849 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4851 assert(0 && "Illegal FP comparison");
4853 // Handle all other FP comparisons here.
4854 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4857 // We are handling one of the integer comparisons here. Since SSE only has
4858 // GT and EQ comparisons for integer, swapping operands and multiple
4859 // operations may be required for some comparisons.
4860 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4861 bool Swap = false, Invert = false, FlipSigns = false;
4863 switch (VT.getSimpleVT()) {
4865 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4866 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4867 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4868 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4871 switch (SetCCOpcode) {
4873 case ISD::SETNE: Invert = true;
4874 case ISD::SETEQ: Opc = EQOpc; break;
4875 case ISD::SETLT: Swap = true;
4876 case ISD::SETGT: Opc = GTOpc; break;
4877 case ISD::SETGE: Swap = true;
4878 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4879 case ISD::SETULT: Swap = true;
4880 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4881 case ISD::SETUGE: Swap = true;
4882 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4885 std::swap(Op0, Op1);
4887 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4888 // bits of the inputs before performing those operations.
4890 MVT EltVT = VT.getVectorElementType();
4891 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4892 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4893 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4895 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
4896 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
4899 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
4901 // If the logical-not of the result is required, perform that now.
4903 MVT EltVT = VT.getVectorElementType();
4904 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
4905 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
4906 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
4908 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
4913 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
4914 bool addTest = true;
4915 SDValue Cond = Op.getOperand(0);
4918 if (Cond.getOpcode() == ISD::SETCC)
4919 Cond = LowerSETCC(Cond, DAG);
4921 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4922 // setting operand in place of the X86ISD::SETCC.
4923 if (Cond.getOpcode() == X86ISD::SETCC) {
4924 CC = Cond.getOperand(0);
4926 SDValue Cmp = Cond.getOperand(1);
4927 unsigned Opc = Cmp.getOpcode();
4928 MVT VT = Op.getValueType();
4930 bool IllegalFPCMov = false;
4931 if (VT.isFloatingPoint() && !VT.isVector() &&
4932 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4933 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4935 if ((Opc == X86ISD::CMP ||
4936 Opc == X86ISD::COMI ||
4937 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4944 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4945 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4948 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4950 SmallVector<SDValue, 4> Ops;
4951 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4952 // condition is true.
4953 Ops.push_back(Op.getOperand(2));
4954 Ops.push_back(Op.getOperand(1));
4956 Ops.push_back(Cond);
4957 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4960 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
4961 bool addTest = true;
4962 SDValue Chain = Op.getOperand(0);
4963 SDValue Cond = Op.getOperand(1);
4964 SDValue Dest = Op.getOperand(2);
4967 if (Cond.getOpcode() == ISD::SETCC)
4968 Cond = LowerSETCC(Cond, DAG);
4970 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4971 // setting operand in place of the X86ISD::SETCC.
4972 if (Cond.getOpcode() == X86ISD::SETCC) {
4973 CC = Cond.getOperand(0);
4975 SDValue Cmp = Cond.getOperand(1);
4976 unsigned Opc = Cmp.getOpcode();
4977 if (Opc == X86ISD::CMP ||
4978 Opc == X86ISD::COMI ||
4979 Opc == X86ISD::UCOMI) {
4986 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4987 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4989 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4990 Chain, Op.getOperand(2), CC, Cond);
4994 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4995 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4996 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4997 // that the guard pages used by the OS virtual memory manager are allocated in
4998 // correct sequence.
5000 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5001 SelectionDAG &DAG) {
5002 assert(Subtarget->isTargetCygMing() &&
5003 "This should be used only on Cygwin/Mingw targets");
5006 SDValue Chain = Op.getOperand(0);
5007 SDValue Size = Op.getOperand(1);
5008 // FIXME: Ensure alignment here
5012 MVT IntPtr = getPointerTy();
5013 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5015 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0));
5017 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5018 Flag = Chain.getValue(1);
5020 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5021 SDValue Ops[] = { Chain,
5022 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5023 DAG.getRegister(X86::EAX, IntPtr),
5024 DAG.getRegister(X86StackPtr, SPTy),
5026 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5027 Flag = Chain.getValue(1);
5029 Chain = DAG.getCALLSEQ_END(Chain,
5030 DAG.getIntPtrConstant(0),
5031 DAG.getIntPtrConstant(0),
5034 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5036 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5037 return DAG.getMergeValues(Ops1, 2);
5041 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5043 SDValue Dst, SDValue Src,
5044 SDValue Size, unsigned Align,
5045 const Value *DstSV, uint64_t DstSVOff) {
5046 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5048 /// If not DWORD aligned or size is more than the threshold, call the library.
5049 /// The libc version is likely to be faster for these cases. It can use the
5050 /// address value and run time information about the CPU.
5051 if ((Align & 3) != 0 ||
5053 ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
5054 SDValue InFlag(0, 0);
5056 // Check to see if there is a specialized entry-point for memory zeroing.
5057 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5058 if (const char *bzeroEntry =
5059 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5060 MVT IntPtr = getPointerTy();
5061 const Type *IntPtrTy = getTargetData()->getIntPtrType();
5062 TargetLowering::ArgListTy Args;
5063 TargetLowering::ArgListEntry Entry;
5065 Entry.Ty = IntPtrTy;
5066 Args.push_back(Entry);
5068 Args.push_back(Entry);
5069 std::pair<SDValue,SDValue> CallResult =
5070 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
5071 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
5073 return CallResult.second;
5076 // Otherwise have the target-independent code call memset.
5080 uint64_t SizeVal = ConstantSize->getValue();
5081 SDValue InFlag(0, 0);
5084 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5085 unsigned BytesLeft = 0;
5086 bool TwoRepStos = false;
5089 uint64_t Val = ValC->getValue() & 255;
5091 // If the value is a constant, then we can potentially use larger sets.
5092 switch (Align & 3) {
5093 case 2: // WORD aligned
5096 Val = (Val << 8) | Val;
5098 case 0: // DWORD aligned
5101 Val = (Val << 8) | Val;
5102 Val = (Val << 16) | Val;
5103 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5106 Val = (Val << 32) | Val;
5109 default: // Byte aligned
5112 Count = DAG.getIntPtrConstant(SizeVal);
5116 if (AVT.bitsGT(MVT::i8)) {
5117 unsigned UBytes = AVT.getSizeInBits() / 8;
5118 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5119 BytesLeft = SizeVal % UBytes;
5122 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5124 InFlag = Chain.getValue(1);
5127 Count = DAG.getIntPtrConstant(SizeVal);
5128 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5129 InFlag = Chain.getValue(1);
5132 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5134 InFlag = Chain.getValue(1);
5135 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5137 InFlag = Chain.getValue(1);
5139 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5140 SmallVector<SDValue, 8> Ops;
5141 Ops.push_back(Chain);
5142 Ops.push_back(DAG.getValueType(AVT));
5143 Ops.push_back(InFlag);
5144 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5147 InFlag = Chain.getValue(1);
5149 MVT CVT = Count.getValueType();
5150 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5151 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5152 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5154 InFlag = Chain.getValue(1);
5155 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5157 Ops.push_back(Chain);
5158 Ops.push_back(DAG.getValueType(MVT::i8));
5159 Ops.push_back(InFlag);
5160 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5161 } else if (BytesLeft) {
5162 // Handle the last 1 - 7 bytes.
5163 unsigned Offset = SizeVal - BytesLeft;
5164 MVT AddrVT = Dst.getValueType();
5165 MVT SizeVT = Size.getValueType();
5167 Chain = DAG.getMemset(Chain,
5168 DAG.getNode(ISD::ADD, AddrVT, Dst,
5169 DAG.getConstant(Offset, AddrVT)),
5171 DAG.getConstant(BytesLeft, SizeVT),
5172 Align, DstSV, DstSVOff + Offset);
5175 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5180 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5181 SDValue Chain, SDValue Dst, SDValue Src,
5182 SDValue Size, unsigned Align,
5184 const Value *DstSV, uint64_t DstSVOff,
5185 const Value *SrcSV, uint64_t SrcSVOff) {
5186 // This requires the copy size to be a constant, preferrably
5187 // within a subtarget-specific limit.
5188 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5191 uint64_t SizeVal = ConstantSize->getValue();
5192 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5195 /// If not DWORD aligned, call the library.
5196 if ((Align & 3) != 0)
5201 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5204 unsigned UBytes = AVT.getSizeInBits() / 8;
5205 unsigned CountVal = SizeVal / UBytes;
5206 SDValue Count = DAG.getIntPtrConstant(CountVal);
5207 unsigned BytesLeft = SizeVal % UBytes;
5209 SDValue InFlag(0, 0);
5210 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5212 InFlag = Chain.getValue(1);
5213 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5215 InFlag = Chain.getValue(1);
5216 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5218 InFlag = Chain.getValue(1);
5220 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5221 SmallVector<SDValue, 8> Ops;
5222 Ops.push_back(Chain);
5223 Ops.push_back(DAG.getValueType(AVT));
5224 Ops.push_back(InFlag);
5225 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5227 SmallVector<SDValue, 4> Results;
5228 Results.push_back(RepMovs);
5230 // Handle the last 1 - 7 bytes.
5231 unsigned Offset = SizeVal - BytesLeft;
5232 MVT DstVT = Dst.getValueType();
5233 MVT SrcVT = Src.getValueType();
5234 MVT SizeVT = Size.getValueType();
5235 Results.push_back(DAG.getMemcpy(Chain,
5236 DAG.getNode(ISD::ADD, DstVT, Dst,
5237 DAG.getConstant(Offset, DstVT)),
5238 DAG.getNode(ISD::ADD, SrcVT, Src,
5239 DAG.getConstant(Offset, SrcVT)),
5240 DAG.getConstant(BytesLeft, SizeVT),
5241 Align, AlwaysInline,
5242 DstSV, DstSVOff + Offset,
5243 SrcSV, SrcSVOff + Offset));
5246 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5249 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5250 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5251 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5252 SDValue TheChain = N->getOperand(0);
5253 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5254 if (Subtarget->is64Bit()) {
5255 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5256 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5257 MVT::i64, rax.getValue(2));
5258 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5259 DAG.getConstant(32, MVT::i8));
5261 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5264 return DAG.getMergeValues(Ops, 2).getNode();
5267 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5268 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5269 MVT::i32, eax.getValue(2));
5270 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5271 SDValue Ops[] = { eax, edx };
5272 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5274 // Use a MERGE_VALUES to return the value and chain.
5275 Ops[1] = edx.getValue(1);
5276 return DAG.getMergeValues(Ops, 2).getNode();
5279 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5280 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5282 if (!Subtarget->is64Bit()) {
5283 // vastart just stores the address of the VarArgsFrameIndex slot into the
5284 // memory location argument.
5285 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5286 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5290 // gp_offset (0 - 6 * 8)
5291 // fp_offset (48 - 48 + 8 * 16)
5292 // overflow_arg_area (point to parameters coming in memory).
5294 SmallVector<SDValue, 8> MemOps;
5295 SDValue FIN = Op.getOperand(1);
5297 SDValue Store = DAG.getStore(Op.getOperand(0),
5298 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5300 MemOps.push_back(Store);
5303 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5304 Store = DAG.getStore(Op.getOperand(0),
5305 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5307 MemOps.push_back(Store);
5309 // Store ptr to overflow_arg_area
5310 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5311 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5312 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5313 MemOps.push_back(Store);
5315 // Store ptr to reg_save_area.
5316 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5317 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5318 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5319 MemOps.push_back(Store);
5320 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5323 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5324 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5325 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5326 SDValue Chain = Op.getOperand(0);
5327 SDValue SrcPtr = Op.getOperand(1);
5328 SDValue SrcSV = Op.getOperand(2);
5330 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5335 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5336 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5337 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5338 SDValue Chain = Op.getOperand(0);
5339 SDValue DstPtr = Op.getOperand(1);
5340 SDValue SrcPtr = Op.getOperand(2);
5341 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5342 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5344 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5345 DAG.getIntPtrConstant(24), 8, false,
5346 DstSV, 0, SrcSV, 0);
5350 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5351 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5353 default: return SDValue(); // Don't custom lower most intrinsics.
5354 // Comparison intrinsics.
5355 case Intrinsic::x86_sse_comieq_ss:
5356 case Intrinsic::x86_sse_comilt_ss:
5357 case Intrinsic::x86_sse_comile_ss:
5358 case Intrinsic::x86_sse_comigt_ss:
5359 case Intrinsic::x86_sse_comige_ss:
5360 case Intrinsic::x86_sse_comineq_ss:
5361 case Intrinsic::x86_sse_ucomieq_ss:
5362 case Intrinsic::x86_sse_ucomilt_ss:
5363 case Intrinsic::x86_sse_ucomile_ss:
5364 case Intrinsic::x86_sse_ucomigt_ss:
5365 case Intrinsic::x86_sse_ucomige_ss:
5366 case Intrinsic::x86_sse_ucomineq_ss:
5367 case Intrinsic::x86_sse2_comieq_sd:
5368 case Intrinsic::x86_sse2_comilt_sd:
5369 case Intrinsic::x86_sse2_comile_sd:
5370 case Intrinsic::x86_sse2_comigt_sd:
5371 case Intrinsic::x86_sse2_comige_sd:
5372 case Intrinsic::x86_sse2_comineq_sd:
5373 case Intrinsic::x86_sse2_ucomieq_sd:
5374 case Intrinsic::x86_sse2_ucomilt_sd:
5375 case Intrinsic::x86_sse2_ucomile_sd:
5376 case Intrinsic::x86_sse2_ucomigt_sd:
5377 case Intrinsic::x86_sse2_ucomige_sd:
5378 case Intrinsic::x86_sse2_ucomineq_sd: {
5380 ISD::CondCode CC = ISD::SETCC_INVALID;
5383 case Intrinsic::x86_sse_comieq_ss:
5384 case Intrinsic::x86_sse2_comieq_sd:
5388 case Intrinsic::x86_sse_comilt_ss:
5389 case Intrinsic::x86_sse2_comilt_sd:
5393 case Intrinsic::x86_sse_comile_ss:
5394 case Intrinsic::x86_sse2_comile_sd:
5398 case Intrinsic::x86_sse_comigt_ss:
5399 case Intrinsic::x86_sse2_comigt_sd:
5403 case Intrinsic::x86_sse_comige_ss:
5404 case Intrinsic::x86_sse2_comige_sd:
5408 case Intrinsic::x86_sse_comineq_ss:
5409 case Intrinsic::x86_sse2_comineq_sd:
5413 case Intrinsic::x86_sse_ucomieq_ss:
5414 case Intrinsic::x86_sse2_ucomieq_sd:
5415 Opc = X86ISD::UCOMI;
5418 case Intrinsic::x86_sse_ucomilt_ss:
5419 case Intrinsic::x86_sse2_ucomilt_sd:
5420 Opc = X86ISD::UCOMI;
5423 case Intrinsic::x86_sse_ucomile_ss:
5424 case Intrinsic::x86_sse2_ucomile_sd:
5425 Opc = X86ISD::UCOMI;
5428 case Intrinsic::x86_sse_ucomigt_ss:
5429 case Intrinsic::x86_sse2_ucomigt_sd:
5430 Opc = X86ISD::UCOMI;
5433 case Intrinsic::x86_sse_ucomige_ss:
5434 case Intrinsic::x86_sse2_ucomige_sd:
5435 Opc = X86ISD::UCOMI;
5438 case Intrinsic::x86_sse_ucomineq_ss:
5439 case Intrinsic::x86_sse2_ucomineq_sd:
5440 Opc = X86ISD::UCOMI;
5446 SDValue LHS = Op.getOperand(1);
5447 SDValue RHS = Op.getOperand(2);
5448 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5450 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5451 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5452 DAG.getConstant(X86CC, MVT::i8), Cond);
5453 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5456 // Fix vector shift instructions where the last operand is a non-immediate
5458 case Intrinsic::x86_sse2_pslli_w:
5459 case Intrinsic::x86_sse2_pslli_d:
5460 case Intrinsic::x86_sse2_pslli_q:
5461 case Intrinsic::x86_sse2_psrli_w:
5462 case Intrinsic::x86_sse2_psrli_d:
5463 case Intrinsic::x86_sse2_psrli_q:
5464 case Intrinsic::x86_sse2_psrai_w:
5465 case Intrinsic::x86_sse2_psrai_d:
5466 case Intrinsic::x86_mmx_pslli_w:
5467 case Intrinsic::x86_mmx_pslli_d:
5468 case Intrinsic::x86_mmx_pslli_q:
5469 case Intrinsic::x86_mmx_psrli_w:
5470 case Intrinsic::x86_mmx_psrli_d:
5471 case Intrinsic::x86_mmx_psrli_q:
5472 case Intrinsic::x86_mmx_psrai_w:
5473 case Intrinsic::x86_mmx_psrai_d: {
5474 SDValue ShAmt = Op.getOperand(2);
5475 if (isa<ConstantSDNode>(ShAmt))
5478 unsigned NewIntNo = 0;
5479 MVT ShAmtVT = MVT::v4i32;
5481 case Intrinsic::x86_sse2_pslli_w:
5482 NewIntNo = Intrinsic::x86_sse2_psll_w;
5484 case Intrinsic::x86_sse2_pslli_d:
5485 NewIntNo = Intrinsic::x86_sse2_psll_d;
5487 case Intrinsic::x86_sse2_pslli_q:
5488 NewIntNo = Intrinsic::x86_sse2_psll_q;
5490 case Intrinsic::x86_sse2_psrli_w:
5491 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5493 case Intrinsic::x86_sse2_psrli_d:
5494 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5496 case Intrinsic::x86_sse2_psrli_q:
5497 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5499 case Intrinsic::x86_sse2_psrai_w:
5500 NewIntNo = Intrinsic::x86_sse2_psra_w;
5502 case Intrinsic::x86_sse2_psrai_d:
5503 NewIntNo = Intrinsic::x86_sse2_psra_d;
5506 ShAmtVT = MVT::v2i32;
5508 case Intrinsic::x86_mmx_pslli_w:
5509 NewIntNo = Intrinsic::x86_mmx_psll_w;
5511 case Intrinsic::x86_mmx_pslli_d:
5512 NewIntNo = Intrinsic::x86_mmx_psll_d;
5514 case Intrinsic::x86_mmx_pslli_q:
5515 NewIntNo = Intrinsic::x86_mmx_psll_q;
5517 case Intrinsic::x86_mmx_psrli_w:
5518 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5520 case Intrinsic::x86_mmx_psrli_d:
5521 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5523 case Intrinsic::x86_mmx_psrli_q:
5524 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5526 case Intrinsic::x86_mmx_psrai_w:
5527 NewIntNo = Intrinsic::x86_mmx_psra_w;
5529 case Intrinsic::x86_mmx_psrai_d:
5530 NewIntNo = Intrinsic::x86_mmx_psra_d;
5532 default: abort(); // Can't reach here.
5537 MVT VT = Op.getValueType();
5538 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5539 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5540 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5541 DAG.getConstant(NewIntNo, MVT::i32),
5542 Op.getOperand(1), ShAmt);
5547 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5548 // Depths > 0 not supported yet!
5549 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5552 // Just load the return address
5553 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5554 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5557 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5558 // Depths > 0 not supported yet!
5559 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5562 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5563 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5564 DAG.getIntPtrConstant(!Subtarget->is64Bit() ? 4 : 8));
5567 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5568 SelectionDAG &DAG) {
5569 // Is not yet supported on x86-64
5570 if (Subtarget->is64Bit())
5573 return DAG.getIntPtrConstant(8);
5576 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5578 assert(!Subtarget->is64Bit() &&
5579 "Lowering of eh_return builtin is not supported yet on x86-64");
5581 MachineFunction &MF = DAG.getMachineFunction();
5582 SDValue Chain = Op.getOperand(0);
5583 SDValue Offset = Op.getOperand(1);
5584 SDValue Handler = Op.getOperand(2);
5586 SDValue Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
5589 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5590 DAG.getIntPtrConstant(-4UL));
5591 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5592 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5593 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5594 MF.getRegInfo().addLiveOut(X86::ECX);
5596 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5597 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5600 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5601 SelectionDAG &DAG) {
5602 SDValue Root = Op.getOperand(0);
5603 SDValue Trmp = Op.getOperand(1); // trampoline
5604 SDValue FPtr = Op.getOperand(2); // nested function
5605 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5607 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5609 const X86InstrInfo *TII =
5610 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5612 if (Subtarget->is64Bit()) {
5613 SDValue OutChains[6];
5615 // Large code-model.
5617 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5618 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5620 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5621 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5623 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5625 // Load the pointer to the nested function into R11.
5626 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5627 SDValue Addr = Trmp;
5628 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5631 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5632 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5634 // Load the 'nest' parameter value into R10.
5635 // R10 is specified in X86CallingConv.td
5636 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5637 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5638 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5641 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5642 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5644 // Jump to the nested function.
5645 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5646 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5647 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5650 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5651 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5652 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5656 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5657 return DAG.getMergeValues(Ops, 2);
5659 const Function *Func =
5660 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5661 unsigned CC = Func->getCallingConv();
5666 assert(0 && "Unsupported calling convention");
5667 case CallingConv::C:
5668 case CallingConv::X86_StdCall: {
5669 // Pass 'nest' parameter in ECX.
5670 // Must be kept in sync with X86CallingConv.td
5673 // Check that ECX wasn't needed by an 'inreg' parameter.
5674 const FunctionType *FTy = Func->getFunctionType();
5675 const PAListPtr &Attrs = Func->getParamAttrs();
5677 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5678 unsigned InRegCount = 0;
5681 for (FunctionType::param_iterator I = FTy->param_begin(),
5682 E = FTy->param_end(); I != E; ++I, ++Idx)
5683 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5684 // FIXME: should only count parameters that are lowered to integers.
5685 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5687 if (InRegCount > 2) {
5688 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5694 case CallingConv::X86_FastCall:
5695 // Pass 'nest' parameter in EAX.
5696 // Must be kept in sync with X86CallingConv.td
5701 SDValue OutChains[4];
5704 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5705 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5707 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5708 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5709 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5712 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5713 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5715 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5716 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5717 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5718 TrmpAddr, 5, false, 1);
5720 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5721 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5724 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5725 return DAG.getMergeValues(Ops, 2);
5729 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5731 The rounding mode is in bits 11:10 of FPSR, and has the following
5738 FLT_ROUNDS, on the other hand, expects the following:
5745 To perform the conversion, we do:
5746 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5749 MachineFunction &MF = DAG.getMachineFunction();
5750 const TargetMachine &TM = MF.getTarget();
5751 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5752 unsigned StackAlignment = TFI.getStackAlignment();
5753 MVT VT = Op.getValueType();
5755 // Save FP Control Word to stack slot
5756 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5757 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5759 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5760 DAG.getEntryNode(), StackSlot);
5762 // Load FP Control Word from stack slot
5763 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5765 // Transform as necessary
5767 DAG.getNode(ISD::SRL, MVT::i16,
5768 DAG.getNode(ISD::AND, MVT::i16,
5769 CWD, DAG.getConstant(0x800, MVT::i16)),
5770 DAG.getConstant(11, MVT::i8));
5772 DAG.getNode(ISD::SRL, MVT::i16,
5773 DAG.getNode(ISD::AND, MVT::i16,
5774 CWD, DAG.getConstant(0x400, MVT::i16)),
5775 DAG.getConstant(9, MVT::i8));
5778 DAG.getNode(ISD::AND, MVT::i16,
5779 DAG.getNode(ISD::ADD, MVT::i16,
5780 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5781 DAG.getConstant(1, MVT::i16)),
5782 DAG.getConstant(3, MVT::i16));
5785 return DAG.getNode((VT.getSizeInBits() < 16 ?
5786 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5789 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5790 MVT VT = Op.getValueType();
5792 unsigned NumBits = VT.getSizeInBits();
5794 Op = Op.getOperand(0);
5795 if (VT == MVT::i8) {
5796 // Zero extend to i32 since there is not an i8 bsr.
5798 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5801 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5802 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5803 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5805 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5806 SmallVector<SDValue, 4> Ops;
5808 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5809 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5810 Ops.push_back(Op.getValue(1));
5811 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5813 // Finally xor with NumBits-1.
5814 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5817 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5821 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5822 MVT VT = Op.getValueType();
5824 unsigned NumBits = VT.getSizeInBits();
5826 Op = Op.getOperand(0);
5827 if (VT == MVT::i8) {
5829 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5832 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5833 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5834 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5836 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5837 SmallVector<SDValue, 4> Ops;
5839 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5840 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5841 Ops.push_back(Op.getValue(1));
5842 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5845 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5849 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5850 MVT T = Op.getValueType();
5853 switch(T.getSimpleVT()) {
5855 assert(false && "Invalid value type!");
5856 case MVT::i8: Reg = X86::AL; size = 1; break;
5857 case MVT::i16: Reg = X86::AX; size = 2; break;
5858 case MVT::i32: Reg = X86::EAX; size = 4; break;
5860 if (Subtarget->is64Bit()) {
5861 Reg = X86::RAX; size = 8;
5862 } else //Should go away when LowerType stuff lands
5863 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5866 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5867 Op.getOperand(3), SDValue());
5868 SDValue Ops[] = { cpIn.getValue(0),
5871 DAG.getTargetConstant(size, MVT::i8),
5873 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5874 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5876 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5880 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5881 SelectionDAG &DAG) {
5882 MVT T = Op->getValueType(0);
5883 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5884 SDValue cpInL, cpInH;
5885 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5886 DAG.getConstant(0, MVT::i32));
5887 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5888 DAG.getConstant(1, MVT::i32));
5889 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5891 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5892 cpInH, cpInL.getValue(1));
5893 SDValue swapInL, swapInH;
5894 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5895 DAG.getConstant(0, MVT::i32));
5896 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5897 DAG.getConstant(1, MVT::i32));
5898 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5899 swapInL, cpInH.getValue(1));
5900 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5901 swapInH, swapInL.getValue(1));
5902 SDValue Ops[] = { swapInH.getValue(0),
5904 swapInH.getValue(1)};
5905 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5906 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5907 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5908 Result.getValue(1));
5909 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5910 cpOutL.getValue(2));
5911 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5912 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5913 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
5914 return DAG.getMergeValues(Vals, 2).getNode();
5917 SDNode* X86TargetLowering::ExpandATOMIC_LOAD_SUB(SDNode* Op,
5918 SelectionDAG &DAG) {
5919 MVT T = Op->getValueType(0);
5920 SDValue negOp = DAG.getNode(ISD::SUB, T,
5921 DAG.getConstant(0, T), Op->getOperand(2));
5922 return DAG.getAtomic((T==MVT::i8 ? ISD::ATOMIC_LOAD_ADD_8:
5923 T==MVT::i16 ? ISD::ATOMIC_LOAD_ADD_16:
5924 T==MVT::i32 ? ISD::ATOMIC_LOAD_ADD_32:
5925 T==MVT::i64 ? ISD::ATOMIC_LOAD_ADD_64: 0),
5926 Op->getOperand(0), Op->getOperand(1), negOp,
5927 cast<AtomicSDNode>(Op)->getSrcValue(),
5928 cast<AtomicSDNode>(Op)->getAlignment()).getNode();
5931 /// LowerOperation - Provide custom lowering hooks for some operations.
5933 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5934 switch (Op.getOpcode()) {
5935 default: assert(0 && "Should not custom lower this!");
5936 case ISD::ATOMIC_CMP_SWAP_8: return LowerCMP_SWAP(Op,DAG);
5937 case ISD::ATOMIC_CMP_SWAP_16: return LowerCMP_SWAP(Op,DAG);
5938 case ISD::ATOMIC_CMP_SWAP_32: return LowerCMP_SWAP(Op,DAG);
5939 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
5940 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5941 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5942 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5943 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5944 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5945 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5946 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5947 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5948 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5949 case ISD::SHL_PARTS:
5950 case ISD::SRA_PARTS:
5951 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5952 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5953 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5954 case ISD::FABS: return LowerFABS(Op, DAG);
5955 case ISD::FNEG: return LowerFNEG(Op, DAG);
5956 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5957 case ISD::SETCC: return LowerSETCC(Op, DAG);
5958 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
5959 case ISD::SELECT: return LowerSELECT(Op, DAG);
5960 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5961 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5962 case ISD::CALL: return LowerCALL(Op, DAG);
5963 case ISD::RET: return LowerRET(Op, DAG);
5964 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5965 case ISD::VASTART: return LowerVASTART(Op, DAG);
5966 case ISD::VAARG: return LowerVAARG(Op, DAG);
5967 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5968 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5969 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5970 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5971 case ISD::FRAME_TO_ARGS_OFFSET:
5972 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5973 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5974 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5975 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5976 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5977 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5978 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5980 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5981 case ISD::READCYCLECOUNTER:
5982 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
5986 /// ReplaceNodeResults - Replace a node with an illegal result type
5987 /// with a new node built out of custom code.
5988 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
5989 switch (N->getOpcode()) {
5990 default: assert(0 && "Should not custom lower this!");
5991 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5992 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5993 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
5994 case ISD::ATOMIC_LOAD_SUB_8: return ExpandATOMIC_LOAD_SUB(N,DAG);
5995 case ISD::ATOMIC_LOAD_SUB_16: return ExpandATOMIC_LOAD_SUB(N,DAG);
5996 case ISD::ATOMIC_LOAD_SUB_32: return ExpandATOMIC_LOAD_SUB(N,DAG);
5997 case ISD::ATOMIC_LOAD_SUB_64: return ExpandATOMIC_LOAD_SUB(N,DAG);
6001 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6003 default: return NULL;
6004 case X86ISD::BSF: return "X86ISD::BSF";
6005 case X86ISD::BSR: return "X86ISD::BSR";
6006 case X86ISD::SHLD: return "X86ISD::SHLD";
6007 case X86ISD::SHRD: return "X86ISD::SHRD";
6008 case X86ISD::FAND: return "X86ISD::FAND";
6009 case X86ISD::FOR: return "X86ISD::FOR";
6010 case X86ISD::FXOR: return "X86ISD::FXOR";
6011 case X86ISD::FSRL: return "X86ISD::FSRL";
6012 case X86ISD::FILD: return "X86ISD::FILD";
6013 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6014 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6015 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6016 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6017 case X86ISD::FLD: return "X86ISD::FLD";
6018 case X86ISD::FST: return "X86ISD::FST";
6019 case X86ISD::CALL: return "X86ISD::CALL";
6020 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6021 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6022 case X86ISD::CMP: return "X86ISD::CMP";
6023 case X86ISD::COMI: return "X86ISD::COMI";
6024 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6025 case X86ISD::SETCC: return "X86ISD::SETCC";
6026 case X86ISD::CMOV: return "X86ISD::CMOV";
6027 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6028 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6029 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6030 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6031 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6032 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6033 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6034 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6035 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6036 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6037 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6038 case X86ISD::FMAX: return "X86ISD::FMAX";
6039 case X86ISD::FMIN: return "X86ISD::FMIN";
6040 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6041 case X86ISD::FRCP: return "X86ISD::FRCP";
6042 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6043 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6044 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6045 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6046 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6047 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6048 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6049 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6050 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6051 case X86ISD::VSHL: return "X86ISD::VSHL";
6052 case X86ISD::VSRL: return "X86ISD::VSRL";
6053 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6054 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6055 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6056 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6057 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6058 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6059 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6060 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6061 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6062 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6066 // isLegalAddressingMode - Return true if the addressing mode represented
6067 // by AM is legal for this target, for a load/store of the specified type.
6068 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6069 const Type *Ty) const {
6070 // X86 supports extremely general addressing modes.
6072 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6073 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6077 // We can only fold this if we don't need an extra load.
6078 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6081 // X86-64 only supports addr of globals in small code model.
6082 if (Subtarget->is64Bit()) {
6083 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6085 // If lower 4G is not available, then we must use rip-relative addressing.
6086 if (AM.BaseOffs || AM.Scale > 1)
6097 // These scales always work.
6102 // These scales are formed with basereg+scalereg. Only accept if there is
6107 default: // Other stuff never works.
6115 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6116 if (!Ty1->isInteger() || !Ty2->isInteger())
6118 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6119 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6120 if (NumBits1 <= NumBits2)
6122 return Subtarget->is64Bit() || NumBits1 < 64;
6125 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6126 if (!VT1.isInteger() || !VT2.isInteger())
6128 unsigned NumBits1 = VT1.getSizeInBits();
6129 unsigned NumBits2 = VT2.getSizeInBits();
6130 if (NumBits1 <= NumBits2)
6132 return Subtarget->is64Bit() || NumBits1 < 64;
6135 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6136 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6137 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6138 /// are assumed to be legal.
6140 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6141 // Only do shuffles on 128-bit vector types for now.
6142 if (VT.getSizeInBits() == 64) return false;
6143 return (Mask.getNode()->getNumOperands() <= 4 ||
6144 isIdentityMask(Mask.getNode()) ||
6145 isIdentityMask(Mask.getNode(), true) ||
6146 isSplatMask(Mask.getNode()) ||
6147 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6148 X86::isUNPCKLMask(Mask.getNode()) ||
6149 X86::isUNPCKHMask(Mask.getNode()) ||
6150 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6151 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6155 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6156 MVT EVT, SelectionDAG &DAG) const {
6157 unsigned NumElts = BVOps.size();
6158 // Only do shuffles on 128-bit vector types for now.
6159 if (EVT.getSizeInBits() * NumElts == 64) return false;
6160 if (NumElts == 2) return true;
6162 return (isMOVLMask(&BVOps[0], 4) ||
6163 isCommutedMOVL(&BVOps[0], 4, true) ||
6164 isSHUFPMask(&BVOps[0], 4) ||
6165 isCommutedSHUFP(&BVOps[0], 4));
6170 //===----------------------------------------------------------------------===//
6171 // X86 Scheduler Hooks
6172 //===----------------------------------------------------------------------===//
6174 // private utility function
6176 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6177 MachineBasicBlock *MBB,
6185 TargetRegisterClass *RC,
6187 // For the atomic bitwise operator, we generate
6190 // ld t1 = [bitinstr.addr]
6191 // op t2 = t1, [bitinstr.val]
6193 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6195 // fallthrough -->nextMBB
6196 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6197 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6198 MachineFunction::iterator MBBIter = MBB;
6201 /// First build the CFG
6202 MachineFunction *F = MBB->getParent();
6203 MachineBasicBlock *thisMBB = MBB;
6204 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6205 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6206 F->insert(MBBIter, newMBB);
6207 F->insert(MBBIter, nextMBB);
6209 // Move all successors to thisMBB to nextMBB
6210 nextMBB->transferSuccessors(thisMBB);
6212 // Update thisMBB to fall through to newMBB
6213 thisMBB->addSuccessor(newMBB);
6215 // newMBB jumps to itself and fall through to nextMBB
6216 newMBB->addSuccessor(nextMBB);
6217 newMBB->addSuccessor(newMBB);
6219 // Insert instructions into newMBB based on incoming instruction
6220 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6221 MachineOperand& destOper = bInstr->getOperand(0);
6222 MachineOperand* argOpers[6];
6223 int numArgs = bInstr->getNumOperands() - 1;
6224 for (int i=0; i < numArgs; ++i)
6225 argOpers[i] = &bInstr->getOperand(i+1);
6227 // x86 address has 4 operands: base, index, scale, and displacement
6228 int lastAddrIndx = 3; // [0,3]
6231 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6232 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6233 for (int i=0; i <= lastAddrIndx; ++i)
6234 (*MIB).addOperand(*argOpers[i]);
6236 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6238 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6243 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6244 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6245 && "invalid operand");
6246 if (argOpers[valArgIndx]->isReg())
6247 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6249 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6251 (*MIB).addOperand(*argOpers[valArgIndx]);
6253 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6256 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6257 for (int i=0; i <= lastAddrIndx; ++i)
6258 (*MIB).addOperand(*argOpers[i]);
6260 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6261 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6263 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6267 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6269 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6273 // private utility function
6275 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6276 MachineBasicBlock *MBB,
6278 // For the atomic min/max operator, we generate
6281 // ld t1 = [min/max.addr]
6282 // mov t2 = [min/max.val]
6284 // cmov[cond] t2 = t1
6286 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6288 // fallthrough -->nextMBB
6290 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6291 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6292 MachineFunction::iterator MBBIter = MBB;
6295 /// First build the CFG
6296 MachineFunction *F = MBB->getParent();
6297 MachineBasicBlock *thisMBB = MBB;
6298 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6299 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6300 F->insert(MBBIter, newMBB);
6301 F->insert(MBBIter, nextMBB);
6303 // Move all successors to thisMBB to nextMBB
6304 nextMBB->transferSuccessors(thisMBB);
6306 // Update thisMBB to fall through to newMBB
6307 thisMBB->addSuccessor(newMBB);
6309 // newMBB jumps to newMBB and fall through to nextMBB
6310 newMBB->addSuccessor(nextMBB);
6311 newMBB->addSuccessor(newMBB);
6313 // Insert instructions into newMBB based on incoming instruction
6314 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6315 MachineOperand& destOper = mInstr->getOperand(0);
6316 MachineOperand* argOpers[6];
6317 int numArgs = mInstr->getNumOperands() - 1;
6318 for (int i=0; i < numArgs; ++i)
6319 argOpers[i] = &mInstr->getOperand(i+1);
6321 // x86 address has 4 operands: base, index, scale, and displacement
6322 int lastAddrIndx = 3; // [0,3]
6325 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6326 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6327 for (int i=0; i <= lastAddrIndx; ++i)
6328 (*MIB).addOperand(*argOpers[i]);
6330 // We only support register and immediate values
6331 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6332 && "invalid operand");
6334 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6335 if (argOpers[valArgIndx]->isReg())
6336 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6338 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6339 (*MIB).addOperand(*argOpers[valArgIndx]);
6341 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6344 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6349 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6350 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6354 // Cmp and exchange if none has modified the memory location
6355 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6356 for (int i=0; i <= lastAddrIndx; ++i)
6357 (*MIB).addOperand(*argOpers[i]);
6359 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6360 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6362 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6363 MIB.addReg(X86::EAX);
6366 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6368 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6374 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6375 MachineBasicBlock *BB) {
6376 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6377 switch (MI->getOpcode()) {
6378 default: assert(false && "Unexpected instr type to insert");
6379 case X86::CMOV_FR32:
6380 case X86::CMOV_FR64:
6381 case X86::CMOV_V4F32:
6382 case X86::CMOV_V2F64:
6383 case X86::CMOV_V2I64: {
6384 // To "insert" a SELECT_CC instruction, we actually have to insert the
6385 // diamond control-flow pattern. The incoming instruction knows the
6386 // destination vreg to set, the condition code register to branch on, the
6387 // true/false values to select between, and a branch opcode to use.
6388 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6389 MachineFunction::iterator It = BB;
6395 // cmpTY ccX, r1, r2
6397 // fallthrough --> copy0MBB
6398 MachineBasicBlock *thisMBB = BB;
6399 MachineFunction *F = BB->getParent();
6400 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6401 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6403 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6404 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6405 F->insert(It, copy0MBB);
6406 F->insert(It, sinkMBB);
6407 // Update machine-CFG edges by transferring all successors of the current
6408 // block to the new block which will contain the Phi node for the select.
6409 sinkMBB->transferSuccessors(BB);
6411 // Add the true and fallthrough blocks as its successors.
6412 BB->addSuccessor(copy0MBB);
6413 BB->addSuccessor(sinkMBB);
6416 // %FalseValue = ...
6417 // # fallthrough to sinkMBB
6420 // Update machine-CFG edges
6421 BB->addSuccessor(sinkMBB);
6424 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6427 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6428 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6429 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6431 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6435 case X86::FP32_TO_INT16_IN_MEM:
6436 case X86::FP32_TO_INT32_IN_MEM:
6437 case X86::FP32_TO_INT64_IN_MEM:
6438 case X86::FP64_TO_INT16_IN_MEM:
6439 case X86::FP64_TO_INT32_IN_MEM:
6440 case X86::FP64_TO_INT64_IN_MEM:
6441 case X86::FP80_TO_INT16_IN_MEM:
6442 case X86::FP80_TO_INT32_IN_MEM:
6443 case X86::FP80_TO_INT64_IN_MEM: {
6444 // Change the floating point control register to use "round towards zero"
6445 // mode when truncating to an integer value.
6446 MachineFunction *F = BB->getParent();
6447 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6448 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6450 // Load the old value of the high byte of the control word...
6452 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6453 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6455 // Set the high part to be round to zero...
6456 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6459 // Reload the modified control word now...
6460 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6462 // Restore the memory image of control word to original value
6463 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6466 // Get the X86 opcode to use.
6468 switch (MI->getOpcode()) {
6469 default: assert(0 && "illegal opcode!");
6470 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6471 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6472 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6473 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6474 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6475 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6476 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6477 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6478 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6482 MachineOperand &Op = MI->getOperand(0);
6483 if (Op.isRegister()) {
6484 AM.BaseType = X86AddressMode::RegBase;
6485 AM.Base.Reg = Op.getReg();
6487 AM.BaseType = X86AddressMode::FrameIndexBase;
6488 AM.Base.FrameIndex = Op.getIndex();
6490 Op = MI->getOperand(1);
6491 if (Op.isImmediate())
6492 AM.Scale = Op.getImm();
6493 Op = MI->getOperand(2);
6494 if (Op.isImmediate())
6495 AM.IndexReg = Op.getImm();
6496 Op = MI->getOperand(3);
6497 if (Op.isGlobalAddress()) {
6498 AM.GV = Op.getGlobal();
6500 AM.Disp = Op.getImm();
6502 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6503 .addReg(MI->getOperand(4).getReg());
6505 // Reload the original control word now.
6506 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6508 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6511 case X86::ATOMAND32:
6512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6513 X86::AND32ri, X86::MOV32rm,
6514 X86::LCMPXCHG32, X86::MOV32rr,
6515 X86::NOT32r, X86::EAX,
6516 X86::GR32RegisterClass);
6518 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6519 X86::OR32ri, X86::MOV32rm,
6520 X86::LCMPXCHG32, X86::MOV32rr,
6521 X86::NOT32r, X86::EAX,
6522 X86::GR32RegisterClass);
6523 case X86::ATOMXOR32:
6524 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6525 X86::XOR32ri, X86::MOV32rm,
6526 X86::LCMPXCHG32, X86::MOV32rr,
6527 X86::NOT32r, X86::EAX,
6528 X86::GR32RegisterClass);
6529 case X86::ATOMNAND32:
6530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6531 X86::AND32ri, X86::MOV32rm,
6532 X86::LCMPXCHG32, X86::MOV32rr,
6533 X86::NOT32r, X86::EAX,
6534 X86::GR32RegisterClass, true);
6535 case X86::ATOMMIN32:
6536 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6537 case X86::ATOMMAX32:
6538 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6539 case X86::ATOMUMIN32:
6540 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6541 case X86::ATOMUMAX32:
6542 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6544 case X86::ATOMAND16:
6545 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6546 X86::AND16ri, X86::MOV16rm,
6547 X86::LCMPXCHG16, X86::MOV16rr,
6548 X86::NOT16r, X86::AX,
6549 X86::GR16RegisterClass);
6551 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6552 X86::OR16ri, X86::MOV16rm,
6553 X86::LCMPXCHG16, X86::MOV16rr,
6554 X86::NOT16r, X86::AX,
6555 X86::GR16RegisterClass);
6556 case X86::ATOMXOR16:
6557 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6558 X86::XOR16ri, X86::MOV16rm,
6559 X86::LCMPXCHG16, X86::MOV16rr,
6560 X86::NOT16r, X86::AX,
6561 X86::GR16RegisterClass);
6562 case X86::ATOMNAND16:
6563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6564 X86::AND16ri, X86::MOV16rm,
6565 X86::LCMPXCHG16, X86::MOV16rr,
6566 X86::NOT16r, X86::AX,
6567 X86::GR16RegisterClass, true);
6568 case X86::ATOMMIN16:
6569 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6570 case X86::ATOMMAX16:
6571 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6572 case X86::ATOMUMIN16:
6573 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6574 case X86::ATOMUMAX16:
6575 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6579 X86::AND8ri, X86::MOV8rm,
6580 X86::LCMPXCHG8, X86::MOV8rr,
6581 X86::NOT8r, X86::AL,
6582 X86::GR8RegisterClass);
6584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6585 X86::OR8ri, X86::MOV8rm,
6586 X86::LCMPXCHG8, X86::MOV8rr,
6587 X86::NOT8r, X86::AL,
6588 X86::GR8RegisterClass);
6590 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6591 X86::XOR8ri, X86::MOV8rm,
6592 X86::LCMPXCHG8, X86::MOV8rr,
6593 X86::NOT8r, X86::AL,
6594 X86::GR8RegisterClass);
6595 case X86::ATOMNAND8:
6596 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6597 X86::AND8ri, X86::MOV8rm,
6598 X86::LCMPXCHG8, X86::MOV8rr,
6599 X86::NOT8r, X86::AL,
6600 X86::GR8RegisterClass, true);
6601 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6602 case X86::ATOMAND64:
6603 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6604 X86::AND64ri32, X86::MOV64rm,
6605 X86::LCMPXCHG64, X86::MOV64rr,
6606 X86::NOT64r, X86::RAX,
6607 X86::GR64RegisterClass);
6609 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6610 X86::OR64ri32, X86::MOV64rm,
6611 X86::LCMPXCHG64, X86::MOV64rr,
6612 X86::NOT64r, X86::RAX,
6613 X86::GR64RegisterClass);
6614 case X86::ATOMXOR64:
6615 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6616 X86::XOR64ri32, X86::MOV64rm,
6617 X86::LCMPXCHG64, X86::MOV64rr,
6618 X86::NOT64r, X86::RAX,
6619 X86::GR64RegisterClass);
6620 case X86::ATOMNAND64:
6621 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6622 X86::AND64ri32, X86::MOV64rm,
6623 X86::LCMPXCHG64, X86::MOV64rr,
6624 X86::NOT64r, X86::RAX,
6625 X86::GR64RegisterClass, true);
6626 case X86::ATOMMIN64:
6627 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6628 case X86::ATOMMAX64:
6629 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6630 case X86::ATOMUMIN64:
6631 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6632 case X86::ATOMUMAX64:
6633 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6637 //===----------------------------------------------------------------------===//
6638 // X86 Optimization Hooks
6639 //===----------------------------------------------------------------------===//
6641 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6645 const SelectionDAG &DAG,
6646 unsigned Depth) const {
6647 unsigned Opc = Op.getOpcode();
6648 assert((Opc >= ISD::BUILTIN_OP_END ||
6649 Opc == ISD::INTRINSIC_WO_CHAIN ||
6650 Opc == ISD::INTRINSIC_W_CHAIN ||
6651 Opc == ISD::INTRINSIC_VOID) &&
6652 "Should use MaskedValueIsZero if you don't know whether Op"
6653 " is a target node!");
6655 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
6659 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
6660 Mask.getBitWidth() - 1);
6665 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
6666 /// node is a GlobalAddress + offset.
6667 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
6668 GlobalValue* &GA, int64_t &Offset) const{
6669 if (N->getOpcode() == X86ISD::Wrapper) {
6670 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
6671 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
6675 return TargetLowering::isGAPlusOffset(N, GA, Offset);
6678 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
6679 const TargetLowering &TLI) {
6682 if (TLI.isGAPlusOffset(Base, GV, Offset))
6683 return (GV->getAlignment() >= N && (Offset % N) == 0);
6684 // DAG combine handles the stack object case.
6688 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
6689 unsigned NumElems, MVT EVT,
6691 SelectionDAG &DAG, MachineFrameInfo *MFI,
6692 const TargetLowering &TLI) {
6694 for (unsigned i = 0; i < NumElems; ++i) {
6695 SDValue Idx = PermMask.getOperand(i);
6696 if (Idx.getOpcode() == ISD::UNDEF) {
6702 SDValue Elt = DAG.getShuffleScalarElt(N, i);
6703 if (!Elt.getNode() ||
6704 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6707 Base = Elt.getNode();
6708 if (Base->getOpcode() == ISD::UNDEF)
6712 if (Elt.getOpcode() == ISD::UNDEF)
6715 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
6716 EVT.getSizeInBits()/8, i, MFI))
6722 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6723 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6724 /// if the load addresses are consecutive, non-overlapping, and in the right
6726 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6727 const TargetLowering &TLI) {
6728 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6729 MVT VT = N->getValueType(0);
6730 MVT EVT = VT.getVectorElementType();
6731 SDValue PermMask = N->getOperand(2);
6732 unsigned NumElems = PermMask.getNumOperands();
6733 SDNode *Base = NULL;
6734 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
6738 LoadSDNode *LD = cast<LoadSDNode>(Base);
6739 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
6740 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6741 LD->getSrcValueOffset(), LD->isVolatile());
6742 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6743 LD->getSrcValueOffset(), LD->isVolatile(),
6744 LD->getAlignment());
6747 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
6748 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
6749 const X86Subtarget *Subtarget,
6750 const TargetLowering &TLI) {
6751 unsigned NumOps = N->getNumOperands();
6753 // Ignore single operand BUILD_VECTOR.
6757 MVT VT = N->getValueType(0);
6758 MVT EVT = VT.getVectorElementType();
6759 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
6760 // We are looking for load i64 and zero extend. We want to transform
6761 // it before legalizer has a chance to expand it. Also look for i64
6762 // BUILD_PAIR bit casted to f64.
6764 // This must be an insertion into a zero vector.
6765 SDValue HighElt = N->getOperand(1);
6766 if (!isZeroNode(HighElt))
6769 // Value must be a load.
6770 SDNode *Base = N->getOperand(0).getNode();
6771 if (!isa<LoadSDNode>(Base)) {
6772 if (Base->getOpcode() != ISD::BIT_CONVERT)
6774 Base = Base->getOperand(0).getNode();
6775 if (!isa<LoadSDNode>(Base))
6779 // Transform it into VZEXT_LOAD addr.
6780 LoadSDNode *LD = cast<LoadSDNode>(Base);
6782 // Load must not be an extload.
6783 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
6786 return DAG.getNode(X86ISD::VZEXT_LOAD, VT, LD->getChain(), LD->getBasePtr());
6789 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6790 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6791 const X86Subtarget *Subtarget) {
6792 SDValue Cond = N->getOperand(0);
6794 // If we have SSE[12] support, try to form min/max nodes.
6795 if (Subtarget->hasSSE2() &&
6796 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6797 if (Cond.getOpcode() == ISD::SETCC) {
6798 // Get the LHS/RHS of the select.
6799 SDValue LHS = N->getOperand(1);
6800 SDValue RHS = N->getOperand(2);
6801 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6803 unsigned Opcode = 0;
6804 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6807 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6810 if (!UnsafeFPMath) break;
6812 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6814 Opcode = X86ISD::FMIN;
6817 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6820 if (!UnsafeFPMath) break;
6822 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6824 Opcode = X86ISD::FMAX;
6827 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6830 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6833 if (!UnsafeFPMath) break;
6835 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6837 Opcode = X86ISD::FMIN;
6840 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6843 if (!UnsafeFPMath) break;
6845 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6847 Opcode = X86ISD::FMAX;
6853 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6861 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6862 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
6863 const X86Subtarget *Subtarget) {
6864 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6865 // the FP state in cases where an emms may be missing.
6866 // A preferable solution to the general problem is to figure out the right
6867 // places to insert EMMS. This qualifies as a quick hack.
6868 StoreSDNode *St = cast<StoreSDNode>(N);
6869 if (St->getValue().getValueType().isVector() &&
6870 St->getValue().getValueType().getSizeInBits() == 64 &&
6871 isa<LoadSDNode>(St->getValue()) &&
6872 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6873 St->getChain().hasOneUse() && !St->isVolatile()) {
6874 SDNode* LdVal = St->getValue().getNode();
6876 int TokenFactorIndex = -1;
6877 SmallVector<SDValue, 8> Ops;
6878 SDNode* ChainVal = St->getChain().getNode();
6879 // Must be a store of a load. We currently handle two cases: the load
6880 // is a direct child, and it's under an intervening TokenFactor. It is
6881 // possible to dig deeper under nested TokenFactors.
6882 if (ChainVal == LdVal)
6883 Ld = cast<LoadSDNode>(St->getChain());
6884 else if (St->getValue().hasOneUse() &&
6885 ChainVal->getOpcode() == ISD::TokenFactor) {
6886 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6887 if (ChainVal->getOperand(i).getNode() == LdVal) {
6888 TokenFactorIndex = i;
6889 Ld = cast<LoadSDNode>(St->getValue());
6891 Ops.push_back(ChainVal->getOperand(i));
6895 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6896 if (Subtarget->is64Bit()) {
6897 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6898 Ld->getBasePtr(), Ld->getSrcValue(),
6899 Ld->getSrcValueOffset(), Ld->isVolatile(),
6900 Ld->getAlignment());
6901 SDValue NewChain = NewLd.getValue(1);
6902 if (TokenFactorIndex != -1) {
6903 Ops.push_back(NewChain);
6904 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6907 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6908 St->getSrcValue(), St->getSrcValueOffset(),
6909 St->isVolatile(), St->getAlignment());
6912 // Otherwise, lower to two 32-bit copies.
6913 SDValue LoAddr = Ld->getBasePtr();
6914 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6915 DAG.getConstant(4, MVT::i32));
6917 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6918 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6919 Ld->isVolatile(), Ld->getAlignment());
6920 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6921 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6923 MinAlign(Ld->getAlignment(), 4));
6925 SDValue NewChain = LoLd.getValue(1);
6926 if (TokenFactorIndex != -1) {
6927 Ops.push_back(LoLd);
6928 Ops.push_back(HiLd);
6929 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6933 LoAddr = St->getBasePtr();
6934 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6935 DAG.getConstant(4, MVT::i32));
6937 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6938 St->getSrcValue(), St->getSrcValueOffset(),
6939 St->isVolatile(), St->getAlignment());
6940 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6942 St->getSrcValueOffset() + 4,
6944 MinAlign(St->getAlignment(), 4));
6945 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6951 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6952 /// X86ISD::FXOR nodes.
6953 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6954 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6955 // F[X]OR(0.0, x) -> x
6956 // F[X]OR(x, 0.0) -> x
6957 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6958 if (C->getValueAPF().isPosZero())
6959 return N->getOperand(1);
6960 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6961 if (C->getValueAPF().isPosZero())
6962 return N->getOperand(0);
6966 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6967 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6968 // FAND(0.0, x) -> 0.0
6969 // FAND(x, 0.0) -> 0.0
6970 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6971 if (C->getValueAPF().isPosZero())
6972 return N->getOperand(0);
6973 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6974 if (C->getValueAPF().isPosZero())
6975 return N->getOperand(1);
6980 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
6981 DAGCombinerInfo &DCI) const {
6982 SelectionDAG &DAG = DCI.DAG;
6983 switch (N->getOpcode()) {
6985 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
6986 case ISD::BUILD_VECTOR:
6987 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
6988 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
6989 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
6991 case X86ISD::FOR: return PerformFORCombine(N, DAG);
6992 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
6998 //===----------------------------------------------------------------------===//
6999 // X86 Inline Assembly Support
7000 //===----------------------------------------------------------------------===//
7002 /// getConstraintType - Given a constraint letter, return the type of
7003 /// constraint it is for this target.
7004 X86TargetLowering::ConstraintType
7005 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7006 if (Constraint.size() == 1) {
7007 switch (Constraint[0]) {
7018 return C_RegisterClass;
7023 return TargetLowering::getConstraintType(Constraint);
7026 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7027 /// with another that has more specific requirements based on the type of the
7028 /// corresponding operand.
7029 const char *X86TargetLowering::
7030 LowerXConstraint(MVT ConstraintVT) const {
7031 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7032 // 'f' like normal targets.
7033 if (ConstraintVT.isFloatingPoint()) {
7034 if (Subtarget->hasSSE2())
7036 if (Subtarget->hasSSE1())
7040 return TargetLowering::LowerXConstraint(ConstraintVT);
7043 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7044 /// vector. If it is invalid, don't add anything to Ops.
7045 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7047 std::vector<SDValue>&Ops,
7048 SelectionDAG &DAG) const {
7049 SDValue Result(0, 0);
7051 switch (Constraint) {
7054 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7055 if (C->getValue() <= 31) {
7056 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7062 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7063 if (C->getValue() <= 255) {
7064 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7070 // Literal immediates are always ok.
7071 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7072 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
7076 // If we are in non-pic codegen mode, we allow the address of a global (with
7077 // an optional displacement) to be used with 'i'.
7078 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7081 // Match either (GA) or (GA+C)
7083 Offset = GA->getOffset();
7084 } else if (Op.getOpcode() == ISD::ADD) {
7085 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7086 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7088 Offset = GA->getOffset()+C->getValue();
7090 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7091 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7093 Offset = GA->getOffset()+C->getValue();
7100 // If addressing this global requires a load (e.g. in PIC mode), we can't
7102 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
7106 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7112 // Otherwise, not valid for this mode.
7117 if (Result.getNode()) {
7118 Ops.push_back(Result);
7121 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7124 std::vector<unsigned> X86TargetLowering::
7125 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7127 if (Constraint.size() == 1) {
7128 // FIXME: not handling fp-stack yet!
7129 switch (Constraint[0]) { // GCC X86 Constraint Letters
7130 default: break; // Unknown constraint letter
7131 case 'A': // EAX/EDX
7132 if (VT == MVT::i32 || VT == MVT::i64)
7133 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7135 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7138 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7139 else if (VT == MVT::i16)
7140 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7141 else if (VT == MVT::i8)
7142 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7143 else if (VT == MVT::i64)
7144 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7149 return std::vector<unsigned>();
7152 std::pair<unsigned, const TargetRegisterClass*>
7153 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7155 // First, see if this is a constraint that directly corresponds to an LLVM
7157 if (Constraint.size() == 1) {
7158 // GCC Constraint Letters
7159 switch (Constraint[0]) {
7161 case 'r': // GENERAL_REGS
7162 case 'R': // LEGACY_REGS
7163 case 'l': // INDEX_REGS
7164 if (VT == MVT::i64 && Subtarget->is64Bit())
7165 return std::make_pair(0U, X86::GR64RegisterClass);
7167 return std::make_pair(0U, X86::GR32RegisterClass);
7168 else if (VT == MVT::i16)
7169 return std::make_pair(0U, X86::GR16RegisterClass);
7170 else if (VT == MVT::i8)
7171 return std::make_pair(0U, X86::GR8RegisterClass);
7173 case 'f': // FP Stack registers.
7174 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7175 // value to the correct fpstack register class.
7176 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7177 return std::make_pair(0U, X86::RFP32RegisterClass);
7178 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7179 return std::make_pair(0U, X86::RFP64RegisterClass);
7180 return std::make_pair(0U, X86::RFP80RegisterClass);
7181 case 'y': // MMX_REGS if MMX allowed.
7182 if (!Subtarget->hasMMX()) break;
7183 return std::make_pair(0U, X86::VR64RegisterClass);
7185 case 'Y': // SSE_REGS if SSE2 allowed
7186 if (!Subtarget->hasSSE2()) break;
7188 case 'x': // SSE_REGS if SSE1 allowed
7189 if (!Subtarget->hasSSE1()) break;
7191 switch (VT.getSimpleVT()) {
7193 // Scalar SSE types.
7196 return std::make_pair(0U, X86::FR32RegisterClass);
7199 return std::make_pair(0U, X86::FR64RegisterClass);
7207 return std::make_pair(0U, X86::VR128RegisterClass);
7213 // Use the default implementation in TargetLowering to convert the register
7214 // constraint into a member of a register class.
7215 std::pair<unsigned, const TargetRegisterClass*> Res;
7216 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7218 // Not found as a standard register?
7219 if (Res.second == 0) {
7220 // GCC calls "st(0)" just plain "st".
7221 if (StringsEqualNoCase("{st}", Constraint)) {
7222 Res.first = X86::ST0;
7223 Res.second = X86::RFP80RegisterClass;
7229 // Otherwise, check to see if this is a register class of the wrong value
7230 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7231 // turn into {ax},{dx}.
7232 if (Res.second->hasType(VT))
7233 return Res; // Correct type already, nothing to do.
7235 // All of the single-register GCC register classes map their values onto
7236 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7237 // really want an 8-bit or 32-bit register, map to the appropriate register
7238 // class and return the appropriate register.
7239 if (Res.second == X86::GR16RegisterClass) {
7240 if (VT == MVT::i8) {
7241 unsigned DestReg = 0;
7242 switch (Res.first) {
7244 case X86::AX: DestReg = X86::AL; break;
7245 case X86::DX: DestReg = X86::DL; break;
7246 case X86::CX: DestReg = X86::CL; break;
7247 case X86::BX: DestReg = X86::BL; break;
7250 Res.first = DestReg;
7251 Res.second = Res.second = X86::GR8RegisterClass;
7253 } else if (VT == MVT::i32) {
7254 unsigned DestReg = 0;
7255 switch (Res.first) {
7257 case X86::AX: DestReg = X86::EAX; break;
7258 case X86::DX: DestReg = X86::EDX; break;
7259 case X86::CX: DestReg = X86::ECX; break;
7260 case X86::BX: DestReg = X86::EBX; break;
7261 case X86::SI: DestReg = X86::ESI; break;
7262 case X86::DI: DestReg = X86::EDI; break;
7263 case X86::BP: DestReg = X86::EBP; break;
7264 case X86::SP: DestReg = X86::ESP; break;
7267 Res.first = DestReg;
7268 Res.second = Res.second = X86::GR32RegisterClass;
7270 } else if (VT == MVT::i64) {
7271 unsigned DestReg = 0;
7272 switch (Res.first) {
7274 case X86::AX: DestReg = X86::RAX; break;
7275 case X86::DX: DestReg = X86::RDX; break;
7276 case X86::CX: DestReg = X86::RCX; break;
7277 case X86::BX: DestReg = X86::RBX; break;
7278 case X86::SI: DestReg = X86::RSI; break;
7279 case X86::DI: DestReg = X86::RDI; break;
7280 case X86::BP: DestReg = X86::RBP; break;
7281 case X86::SP: DestReg = X86::RSP; break;
7284 Res.first = DestReg;
7285 Res.second = Res.second = X86::GR64RegisterClass;
7288 } else if (Res.second == X86::FR32RegisterClass ||
7289 Res.second == X86::FR64RegisterClass ||
7290 Res.second == X86::VR128RegisterClass) {
7291 // Handle references to XMM physical registers that got mapped into the
7292 // wrong class. This can happen with constraints like {xmm0} where the
7293 // target independent register mapper will just pick the first match it can
7294 // find, ignoring the required type.
7296 Res.second = X86::FR32RegisterClass;
7297 else if (VT == MVT::f64)
7298 Res.second = X86::FR64RegisterClass;
7299 else if (X86::VR128RegisterClass->hasType(VT))
7300 Res.second = X86::VR128RegisterClass;