1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/StringExtras.h"
43 // Forward declarations.
44 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
46 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
47 : TargetLowering(TM) {
48 Subtarget = &TM.getSubtarget<X86Subtarget>();
49 X86ScalarSSEf64 = Subtarget->hasSSE2();
50 X86ScalarSSEf32 = Subtarget->hasSSE1();
51 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
55 RegInfo = TM.getRegisterInfo();
57 // Set up the TargetLowering object.
59 // X86 is weird, it always uses i8 for shift amounts and setcc results.
60 setShiftAmountType(MVT::i8);
61 setSetCCResultContents(ZeroOrOneSetCCResult);
62 setSchedulingPreference(SchedulingForRegPressure);
63 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
64 setStackPointerRegisterToSaveRestore(X86StackPtr);
66 if (Subtarget->isTargetDarwin()) {
67 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
68 setUseUnderscoreSetJmp(false);
69 setUseUnderscoreLongJmp(false);
70 } else if (Subtarget->isTargetMingw()) {
71 // MS runtime is weird: it exports _setjmp, but longjmp!
72 setUseUnderscoreSetJmp(true);
73 setUseUnderscoreLongJmp(false);
75 setUseUnderscoreSetJmp(true);
76 setUseUnderscoreLongJmp(true);
79 // Set up the register classes.
80 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
81 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
82 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
83 if (Subtarget->is64Bit())
84 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
86 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
88 // We don't accept any truncstore of integer registers.
89 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
90 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
92 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
93 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
94 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
96 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
98 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
99 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
102 if (Subtarget->is64Bit()) {
103 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
104 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
107 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
108 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
110 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
113 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
115 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
116 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
117 // SSE has no i16 to fp conversion, only i32
118 if (X86ScalarSSEf32) {
119 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
120 // f32 and f64 cases are Legal, f80 case is not
121 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
123 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
124 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
127 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
128 // are Legal, f80 is custom lowered.
129 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
130 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
132 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
134 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
135 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
137 if (X86ScalarSSEf32) {
138 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
139 // f32 and f64 cases are Legal, f80 case is not
140 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
142 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
143 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
146 // Handle FP_TO_UINT by promoting the destination to a larger signed
148 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
149 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
150 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
152 if (Subtarget->is64Bit()) {
153 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
154 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
156 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
157 // Expand FP_TO_UINT into a select.
158 // FIXME: We would like to use a Custom expander here eventually to do
159 // the optimal thing for SSE vs. the default expansion in the legalizer.
160 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
162 // With SSE3 we can use fisttpll to convert to a signed i64.
163 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
166 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
167 if (!X86ScalarSSEf64) {
168 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
169 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
172 // Scalar integer divide and remainder are lowered to use operations that
173 // produce two results, to match the available instructions. This exposes
174 // the two-result form to trivial CSE, which is able to combine x/y and x%y
175 // into a single instruction.
177 // Scalar integer multiply-high is also lowered to use two-result
178 // operations, to match the available instructions. However, plain multiply
179 // (low) operations are left as Legal, as there are single-result
180 // instructions for this in x86. Using the two-result multiply instructions
181 // when both high and low results are needed must be arranged by dagcombine.
182 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
183 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
184 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
185 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
186 setOperationAction(ISD::SREM , MVT::i8 , Expand);
187 setOperationAction(ISD::UREM , MVT::i8 , Expand);
188 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
189 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
190 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
191 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
192 setOperationAction(ISD::SREM , MVT::i16 , Expand);
193 setOperationAction(ISD::UREM , MVT::i16 , Expand);
194 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
195 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
196 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
197 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
198 setOperationAction(ISD::SREM , MVT::i32 , Expand);
199 setOperationAction(ISD::UREM , MVT::i32 , Expand);
200 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
201 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
202 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
203 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::SREM , MVT::i64 , Expand);
205 setOperationAction(ISD::UREM , MVT::i64 , Expand);
207 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
208 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
209 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
210 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
211 if (Subtarget->is64Bit())
212 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
216 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
217 setOperationAction(ISD::FREM , MVT::f32 , Expand);
218 setOperationAction(ISD::FREM , MVT::f64 , Expand);
219 setOperationAction(ISD::FREM , MVT::f80 , Expand);
220 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
222 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
223 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
224 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
225 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
226 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
227 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
228 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
229 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
230 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
231 if (Subtarget->is64Bit()) {
232 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
233 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
234 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
237 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
238 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
240 // These should be promoted to a larger select which is supported.
241 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
242 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
243 // X86 wants to expand cmov itself.
244 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
245 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
246 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
248 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
249 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
251 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
252 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
254 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
257 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
259 // X86 ret instruction may pop stack.
260 setOperationAction(ISD::RET , MVT::Other, Custom);
261 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
264 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
265 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
266 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
267 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
268 if (Subtarget->is64Bit())
269 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
270 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
271 if (Subtarget->is64Bit()) {
272 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
273 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
274 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
275 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
277 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
278 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
279 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
280 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
281 if (Subtarget->is64Bit()) {
282 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
283 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
284 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
287 if (Subtarget->hasSSE1())
288 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
290 if (!Subtarget->hasSSE2())
291 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
293 // Expand certain atomics
294 setOperationAction(ISD::ATOMIC_CMP_SWAP_8 , MVT::i8, Custom);
295 setOperationAction(ISD::ATOMIC_CMP_SWAP_16, MVT::i16, Custom);
296 setOperationAction(ISD::ATOMIC_CMP_SWAP_32, MVT::i32, Custom);
297 setOperationAction(ISD::ATOMIC_CMP_SWAP_64, MVT::i64, Custom);
299 setOperationAction(ISD::ATOMIC_LOAD_SUB_8, MVT::i8, Expand);
300 setOperationAction(ISD::ATOMIC_LOAD_SUB_16, MVT::i16, Expand);
301 setOperationAction(ISD::ATOMIC_LOAD_SUB_32, MVT::i32, Expand);
302 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Expand);
304 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
305 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
306 // FIXME - use subtarget debug flags
307 if (!Subtarget->isTargetDarwin() &&
308 !Subtarget->isTargetELF() &&
309 !Subtarget->isTargetCygMing()) {
310 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
311 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
314 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
315 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
316 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
317 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
318 if (Subtarget->is64Bit()) {
319 setExceptionPointerRegister(X86::RAX);
320 setExceptionSelectorRegister(X86::RDX);
322 setExceptionPointerRegister(X86::EAX);
323 setExceptionSelectorRegister(X86::EDX);
325 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
326 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
328 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
330 setOperationAction(ISD::TRAP, MVT::Other, Legal);
332 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
333 setOperationAction(ISD::VASTART , MVT::Other, Custom);
334 setOperationAction(ISD::VAEND , MVT::Other, Expand);
335 if (Subtarget->is64Bit()) {
336 setOperationAction(ISD::VAARG , MVT::Other, Custom);
337 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
339 setOperationAction(ISD::VAARG , MVT::Other, Expand);
340 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
343 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
344 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
345 if (Subtarget->is64Bit())
346 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
347 if (Subtarget->isTargetCygMing())
348 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
350 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
352 if (X86ScalarSSEf64) {
353 // f32 and f64 use SSE.
354 // Set up the FP register classes.
355 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
356 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
358 // Use ANDPD to simulate FABS.
359 setOperationAction(ISD::FABS , MVT::f64, Custom);
360 setOperationAction(ISD::FABS , MVT::f32, Custom);
362 // Use XORP to simulate FNEG.
363 setOperationAction(ISD::FNEG , MVT::f64, Custom);
364 setOperationAction(ISD::FNEG , MVT::f32, Custom);
366 // Use ANDPD and ORPD to simulate FCOPYSIGN.
367 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
368 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
370 // We don't support sin/cos/fmod
371 setOperationAction(ISD::FSIN , MVT::f64, Expand);
372 setOperationAction(ISD::FCOS , MVT::f64, Expand);
373 setOperationAction(ISD::FSIN , MVT::f32, Expand);
374 setOperationAction(ISD::FCOS , MVT::f32, Expand);
376 // Expand FP immediates into loads from the stack, except for the special
378 addLegalFPImmediate(APFloat(+0.0)); // xorpd
379 addLegalFPImmediate(APFloat(+0.0f)); // xorps
381 // Floating truncations from f80 and extensions to f80 go through memory.
382 // If optimizing, we lie about this though and handle it in
383 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
385 setConvertAction(MVT::f32, MVT::f80, Expand);
386 setConvertAction(MVT::f64, MVT::f80, Expand);
387 setConvertAction(MVT::f80, MVT::f32, Expand);
388 setConvertAction(MVT::f80, MVT::f64, Expand);
390 } else if (X86ScalarSSEf32) {
391 // Use SSE for f32, x87 for f64.
392 // Set up the FP register classes.
393 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
394 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
396 // Use ANDPS to simulate FABS.
397 setOperationAction(ISD::FABS , MVT::f32, Custom);
399 // Use XORP to simulate FNEG.
400 setOperationAction(ISD::FNEG , MVT::f32, Custom);
402 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
404 // Use ANDPS and ORPS to simulate FCOPYSIGN.
405 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
406 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
408 // We don't support sin/cos/fmod
409 setOperationAction(ISD::FSIN , MVT::f32, Expand);
410 setOperationAction(ISD::FCOS , MVT::f32, Expand);
412 // Special cases we handle for FP constants.
413 addLegalFPImmediate(APFloat(+0.0f)); // xorps
414 addLegalFPImmediate(APFloat(+0.0)); // FLD0
415 addLegalFPImmediate(APFloat(+1.0)); // FLD1
416 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
417 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
419 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
420 // this though and handle it in InstructionSelectPreprocess so that
421 // dagcombine2 can hack on these.
423 setConvertAction(MVT::f32, MVT::f64, Expand);
424 setConvertAction(MVT::f32, MVT::f80, Expand);
425 setConvertAction(MVT::f80, MVT::f32, Expand);
426 setConvertAction(MVT::f64, MVT::f32, Expand);
427 // And x87->x87 truncations also.
428 setConvertAction(MVT::f80, MVT::f64, Expand);
432 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
433 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
436 // f32 and f64 in x87.
437 // Set up the FP register classes.
438 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
439 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
441 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
442 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
443 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
446 // Floating truncations go through memory. If optimizing, we lie about
447 // this though and handle it in InstructionSelectPreprocess so that
448 // dagcombine2 can hack on these.
450 setConvertAction(MVT::f80, MVT::f32, Expand);
451 setConvertAction(MVT::f64, MVT::f32, Expand);
452 setConvertAction(MVT::f80, MVT::f64, Expand);
456 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
457 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
459 addLegalFPImmediate(APFloat(+0.0)); // FLD0
460 addLegalFPImmediate(APFloat(+1.0)); // FLD1
461 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
462 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
463 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
464 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
465 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
466 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
469 // Long double always uses X87.
470 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
471 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
472 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
474 APFloat TmpFlt(+0.0);
475 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
476 addLegalFPImmediate(TmpFlt); // FLD0
478 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
479 APFloat TmpFlt2(+1.0);
480 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
481 addLegalFPImmediate(TmpFlt2); // FLD1
482 TmpFlt2.changeSign();
483 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
487 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
488 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
491 // Always use a library call for pow.
492 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
493 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
494 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
496 setOperationAction(ISD::FLOG, MVT::f32, Expand);
497 setOperationAction(ISD::FLOG, MVT::f64, Expand);
498 setOperationAction(ISD::FLOG, MVT::f80, Expand);
499 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
500 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
501 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
502 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
503 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
504 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
505 setOperationAction(ISD::FEXP, MVT::f32, Expand);
506 setOperationAction(ISD::FEXP, MVT::f64, Expand);
507 setOperationAction(ISD::FEXP, MVT::f80, Expand);
508 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
509 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
510 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
512 // First set operation action for all vector types to expand. Then we
513 // will selectively turn on ones that can be effectively codegen'd.
514 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
515 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
516 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
517 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
531 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
556 if (Subtarget->hasMMX()) {
557 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
558 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
559 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
560 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
561 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
563 // FIXME: add MMX packed arithmetics
565 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
566 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
567 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
568 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
570 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
571 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
572 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
573 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
575 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
576 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
578 setOperationAction(ISD::AND, MVT::v8i8, Promote);
579 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
580 setOperationAction(ISD::AND, MVT::v4i16, Promote);
581 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
582 setOperationAction(ISD::AND, MVT::v2i32, Promote);
583 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
584 setOperationAction(ISD::AND, MVT::v1i64, Legal);
586 setOperationAction(ISD::OR, MVT::v8i8, Promote);
587 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
588 setOperationAction(ISD::OR, MVT::v4i16, Promote);
589 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
590 setOperationAction(ISD::OR, MVT::v2i32, Promote);
591 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
592 setOperationAction(ISD::OR, MVT::v1i64, Legal);
594 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
595 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
596 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
597 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
598 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
599 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
600 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
602 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
603 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
604 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
605 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
606 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
607 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
608 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
609 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
610 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
612 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
613 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
614 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
615 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
616 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
618 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
619 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
620 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
621 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
623 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
624 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
625 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
626 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
628 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
631 if (Subtarget->hasSSE1()) {
632 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
634 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
635 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
636 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
637 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
638 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
639 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
640 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
641 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
643 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
644 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
645 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
648 if (Subtarget->hasSSE2()) {
649 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
650 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
651 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
652 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
653 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
655 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
656 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
657 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
658 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
659 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
660 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
661 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
662 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
663 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
664 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
665 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
666 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
667 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
668 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
669 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
671 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
672 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
673 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
674 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
678 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
679 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
682 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
683 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
684 MVT VT = (MVT::SimpleValueType)i;
685 // Do not attempt to custom lower non-power-of-2 vectors
686 if (!isPowerOf2_32(VT.getVectorNumElements()))
688 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
689 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
690 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
692 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
693 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
694 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
695 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
696 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
697 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
698 if (Subtarget->is64Bit()) {
699 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
703 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
704 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
705 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
706 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
707 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
708 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
709 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
710 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
711 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
712 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
713 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
714 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
717 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
719 // Custom lower v2i64 and v2f64 selects.
720 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
721 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
722 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
723 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
727 if (Subtarget->hasSSE41()) {
728 // FIXME: Do we need to handle scalar-to-vector here?
729 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
732 // i8 and i16 vectors are custom , because the source register and source
733 // source memory operand types are not the same width. f32 vectors are
734 // custom since the immediate controlling the insert encodes additional
736 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
741 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
742 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
744 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
746 if (Subtarget->is64Bit()) {
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
748 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
752 if (Subtarget->hasSSE42()) {
753 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
756 // We want to custom lower some of our intrinsics.
757 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
759 // We have target-specific dag combine patterns for the following nodes:
760 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
761 setTargetDAGCombine(ISD::BUILD_VECTOR);
762 setTargetDAGCombine(ISD::SELECT);
763 setTargetDAGCombine(ISD::STORE);
765 computeRegisterProperties();
767 // FIXME: These should be based on subtarget info. Plus, the values should
768 // be smaller when we are in optimizing for size mode.
769 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
770 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
771 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
772 allowUnalignedMemoryAccesses = true; // x86 supports it!
773 setPrefLoopAlignment(16);
777 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
782 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
783 /// the desired ByVal argument alignment.
784 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
787 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
788 if (VTy->getBitWidth() == 128)
790 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
791 unsigned EltAlign = 0;
792 getMaxByValAlign(ATy->getElementType(), EltAlign);
793 if (EltAlign > MaxAlign)
795 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
796 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
797 unsigned EltAlign = 0;
798 getMaxByValAlign(STy->getElementType(i), EltAlign);
799 if (EltAlign > MaxAlign)
808 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
809 /// function arguments in the caller parameter area. For X86, aggregates
810 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
811 /// are at 4-byte boundaries.
812 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
813 if (Subtarget->is64Bit()) {
814 // Max of 8 and alignment of type.
815 unsigned TyAlign = getTargetData()->getABITypeAlignment(Ty);
822 if (Subtarget->hasSSE1())
823 getMaxByValAlign(Ty, Align);
827 /// getOptimalMemOpType - Returns the target specific optimal type for load
828 /// and store operations as a result of memset, memcpy, and memmove
829 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
832 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
833 bool isSrcConst, bool isSrcStr) const {
834 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
836 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
838 if (Subtarget->is64Bit() && Size >= 8)
844 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
846 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
847 SelectionDAG &DAG) const {
848 if (usesGlobalOffsetTable())
849 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
850 if (!Subtarget->isPICStyleRIPRel())
851 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
855 //===----------------------------------------------------------------------===//
856 // Return Value Calling Convention Implementation
857 //===----------------------------------------------------------------------===//
859 #include "X86GenCallingConv.inc"
861 /// LowerRET - Lower an ISD::RET node.
862 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
863 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
865 SmallVector<CCValAssign, 16> RVLocs;
866 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
867 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
868 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
869 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
871 // If this is the first return lowered for this function, add the regs to the
872 // liveout set for the function.
873 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
874 for (unsigned i = 0; i != RVLocs.size(); ++i)
875 if (RVLocs[i].isRegLoc())
876 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
878 SDValue Chain = Op.getOperand(0);
880 // Handle tail call return.
881 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
882 if (Chain.getOpcode() == X86ISD::TAILCALL) {
883 SDValue TailCall = Chain;
884 SDValue TargetAddress = TailCall.getOperand(1);
885 SDValue StackAdjustment = TailCall.getOperand(2);
886 assert(((TargetAddress.getOpcode() == ISD::Register &&
887 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
888 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
889 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
890 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
891 "Expecting an global address, external symbol, or register");
892 assert(StackAdjustment.getOpcode() == ISD::Constant &&
893 "Expecting a const value");
895 SmallVector<SDValue,8> Operands;
896 Operands.push_back(Chain.getOperand(0));
897 Operands.push_back(TargetAddress);
898 Operands.push_back(StackAdjustment);
899 // Copy registers used by the call. Last operand is a flag so it is not
901 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
902 Operands.push_back(Chain.getOperand(i));
904 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
911 SmallVector<SDValue, 6> RetOps;
912 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
913 // Operand #1 = Bytes To Pop
914 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
916 // Copy the result values into the output registers.
917 for (unsigned i = 0; i != RVLocs.size(); ++i) {
918 CCValAssign &VA = RVLocs[i];
919 assert(VA.isRegLoc() && "Can only return in registers!");
920 SDValue ValToCopy = Op.getOperand(i*2+1);
922 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
923 // the RET instruction and handled by the FP Stackifier.
924 if (RVLocs[i].getLocReg() == X86::ST0 ||
925 RVLocs[i].getLocReg() == X86::ST1) {
926 // If this is a copy from an xmm register to ST(0), use an FPExtend to
927 // change the value to the FP stack register class.
928 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
929 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
930 RetOps.push_back(ValToCopy);
931 // Don't emit a copytoreg.
935 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
936 Flag = Chain.getValue(1);
939 // The x86-64 ABI for returning structs by value requires that we copy
940 // the sret argument into %rax for the return. We saved the argument into
941 // a virtual register in the entry block, so now we copy the value out
943 if (Subtarget->is64Bit() &&
944 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
945 MachineFunction &MF = DAG.getMachineFunction();
946 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
947 unsigned Reg = FuncInfo->getSRetReturnReg();
949 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
950 FuncInfo->setSRetReturnReg(Reg);
952 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
954 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
955 Flag = Chain.getValue(1);
958 RetOps[0] = Chain; // Update chain.
960 // Add the flag if we have it.
962 RetOps.push_back(Flag);
964 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
968 /// LowerCallResult - Lower the result values of an ISD::CALL into the
969 /// appropriate copies out of appropriate physical registers. This assumes that
970 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
971 /// being lowered. The returns a SDNode with the same number of values as the
973 SDNode *X86TargetLowering::
974 LowerCallResult(SDValue Chain, SDValue InFlag, SDNode *TheCall,
975 unsigned CallingConv, SelectionDAG &DAG) {
977 // Assign locations to each value returned by this call.
978 SmallVector<CCValAssign, 16> RVLocs;
979 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
980 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
981 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
983 SmallVector<SDValue, 8> ResultVals;
985 // Copy all of the result registers out of their specified physreg.
986 for (unsigned i = 0; i != RVLocs.size(); ++i) {
987 MVT CopyVT = RVLocs[i].getValVT();
989 // If this is a call to a function that returns an fp value on the floating
990 // point stack, but where we prefer to use the value in xmm registers, copy
991 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
992 if ((RVLocs[i].getLocReg() == X86::ST0 ||
993 RVLocs[i].getLocReg() == X86::ST1) &&
994 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
998 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
999 CopyVT, InFlag).getValue(1);
1000 SDValue Val = Chain.getValue(0);
1001 InFlag = Chain.getValue(2);
1003 if (CopyVT != RVLocs[i].getValVT()) {
1004 // Round the F80 the right size, which also moves to the appropriate xmm
1006 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1007 // This truncation won't change the value.
1008 DAG.getIntPtrConstant(1));
1011 ResultVals.push_back(Val);
1014 // Merge everything together with a MERGE_VALUES node.
1015 ResultVals.push_back(Chain);
1016 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1017 ResultVals.size()).getNode();
1021 //===----------------------------------------------------------------------===//
1022 // C & StdCall & Fast Calling Convention implementation
1023 //===----------------------------------------------------------------------===//
1024 // StdCall calling convention seems to be standard for many Windows' API
1025 // routines and around. It differs from C calling convention just a little:
1026 // callee should clean up the stack, not caller. Symbols should be also
1027 // decorated in some fancy way :) It doesn't support any vector arguments.
1028 // For info on fast calling convention see Fast Calling Convention (tail call)
1029 // implementation LowerX86_32FastCCCallTo.
1031 /// AddLiveIn - This helper function adds the specified physical register to the
1032 /// MachineFunction as a live in value. It also creates a corresponding virtual
1033 /// register for it.
1034 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1035 const TargetRegisterClass *RC) {
1036 assert(RC->contains(PReg) && "Not the correct regclass!");
1037 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1038 MF.getRegInfo().addLiveIn(PReg, VReg);
1042 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1044 static bool CallIsStructReturn(SDValue Op) {
1045 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1049 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
1052 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1053 /// return semantics.
1054 static bool ArgsAreStructReturn(SDValue Op) {
1055 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1059 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1062 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1063 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1065 bool X86TargetLowering::IsCalleePop(SDValue Op) {
1066 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1070 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1073 case CallingConv::X86_StdCall:
1074 return !Subtarget->is64Bit();
1075 case CallingConv::X86_FastCall:
1076 return !Subtarget->is64Bit();
1077 case CallingConv::Fast:
1078 return PerformTailCallOpt;
1082 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1083 /// FORMAL_ARGUMENTS node.
1084 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
1085 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1087 if (Subtarget->is64Bit()) {
1088 if (Subtarget->isTargetWin64())
1089 return CC_X86_Win64_C;
1090 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1091 return CC_X86_64_TailCall;
1096 if (CC == CallingConv::X86_FastCall)
1097 return CC_X86_32_FastCall;
1098 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1099 return CC_X86_32_TailCall;
1100 else if (CC == CallingConv::Fast)
1101 return CC_X86_32_FastCC;
1106 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1107 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1109 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1110 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1111 if (CC == CallingConv::X86_FastCall)
1113 else if (CC == CallingConv::X86_StdCall)
1119 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1120 /// in a register before calling.
1121 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1122 return !IsTailCall && !Is64Bit &&
1123 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1124 Subtarget->isPICStyleGOT();
1127 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1128 /// address to be loaded in a register.
1130 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1131 return !Is64Bit && IsTailCall &&
1132 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1133 Subtarget->isPICStyleGOT();
1136 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1137 /// by "Src" to address "Dst" with size and alignment information specified by
1138 /// the specific parameter attribute. The copy will be passed as a byval
1139 /// function parameter.
1141 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1142 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1143 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1144 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1145 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1148 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1149 const CCValAssign &VA,
1150 MachineFrameInfo *MFI,
1152 SDValue Root, unsigned i) {
1153 // Create the nodes corresponding to a load from this parameter slot.
1154 ISD::ArgFlagsTy Flags =
1155 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1156 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1157 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1159 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1160 // changed with more analysis.
1161 // In case of tail call optimization mark all arguments mutable. Since they
1162 // could be overwritten by lowering of arguments in case of a tail call.
1163 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1164 VA.getLocMemOffset(), isImmutable);
1165 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1166 if (Flags.isByVal())
1168 return DAG.getLoad(VA.getValVT(), Root, FIN,
1169 PseudoSourceValue::getFixedStack(FI), 0);
1173 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1174 MachineFunction &MF = DAG.getMachineFunction();
1175 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1177 const Function* Fn = MF.getFunction();
1178 if (Fn->hasExternalLinkage() &&
1179 Subtarget->isTargetCygMing() &&
1180 Fn->getName() == "main")
1181 FuncInfo->setForceFramePointer(true);
1183 // Decorate the function name.
1184 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1186 MachineFrameInfo *MFI = MF.getFrameInfo();
1187 SDValue Root = Op.getOperand(0);
1188 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1189 unsigned CC = MF.getFunction()->getCallingConv();
1190 bool Is64Bit = Subtarget->is64Bit();
1191 bool IsWin64 = Subtarget->isTargetWin64();
1193 assert(!(isVarArg && CC == CallingConv::Fast) &&
1194 "Var args not supported with calling convention fastcc");
1196 // Assign locations to all of the incoming arguments.
1197 SmallVector<CCValAssign, 16> ArgLocs;
1198 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1199 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(Op));
1201 SmallVector<SDValue, 8> ArgValues;
1202 unsigned LastVal = ~0U;
1203 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1204 CCValAssign &VA = ArgLocs[i];
1205 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1207 assert(VA.getValNo() != LastVal &&
1208 "Don't support value assigned to multiple locs yet");
1209 LastVal = VA.getValNo();
1211 if (VA.isRegLoc()) {
1212 MVT RegVT = VA.getLocVT();
1213 TargetRegisterClass *RC;
1214 if (RegVT == MVT::i32)
1215 RC = X86::GR32RegisterClass;
1216 else if (Is64Bit && RegVT == MVT::i64)
1217 RC = X86::GR64RegisterClass;
1218 else if (RegVT == MVT::f32)
1219 RC = X86::FR32RegisterClass;
1220 else if (RegVT == MVT::f64)
1221 RC = X86::FR64RegisterClass;
1222 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1223 RC = X86::VR128RegisterClass;
1224 else if (RegVT.isVector()) {
1225 assert(RegVT.getSizeInBits() == 64);
1227 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1229 // Darwin calling convention passes MMX values in either GPRs or
1230 // XMMs in x86-64. Other targets pass them in memory.
1231 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1232 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1235 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1240 assert(0 && "Unknown argument type!");
1243 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1244 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1246 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1247 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1249 if (VA.getLocInfo() == CCValAssign::SExt)
1250 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1251 DAG.getValueType(VA.getValVT()));
1252 else if (VA.getLocInfo() == CCValAssign::ZExt)
1253 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1254 DAG.getValueType(VA.getValVT()));
1256 if (VA.getLocInfo() != CCValAssign::Full)
1257 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1259 // Handle MMX values passed in GPRs.
1260 if (Is64Bit && RegVT != VA.getLocVT()) {
1261 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1262 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1263 else if (RC == X86::VR128RegisterClass) {
1264 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1265 DAG.getConstant(0, MVT::i64));
1266 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1270 ArgValues.push_back(ArgValue);
1272 assert(VA.isMemLoc());
1273 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1277 // The x86-64 ABI for returning structs by value requires that we copy
1278 // the sret argument into %rax for the return. Save the argument into
1279 // a virtual register so that we can access it from the return points.
1280 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1281 MachineFunction &MF = DAG.getMachineFunction();
1282 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1283 unsigned Reg = FuncInfo->getSRetReturnReg();
1285 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1286 FuncInfo->setSRetReturnReg(Reg);
1288 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1289 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1292 unsigned StackSize = CCInfo.getNextStackOffset();
1293 // align stack specially for tail calls
1294 if (PerformTailCallOpt && CC == CallingConv::Fast)
1295 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1297 // If the function takes variable number of arguments, make a frame index for
1298 // the start of the first vararg value... for expansion of llvm.va_start.
1300 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1301 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1304 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1306 // FIXME: We should really autogenerate these arrays
1307 static const unsigned GPR64ArgRegsWin64[] = {
1308 X86::RCX, X86::RDX, X86::R8, X86::R9
1310 static const unsigned XMMArgRegsWin64[] = {
1311 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1313 static const unsigned GPR64ArgRegs64Bit[] = {
1314 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1316 static const unsigned XMMArgRegs64Bit[] = {
1317 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1318 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1320 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1323 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1324 GPR64ArgRegs = GPR64ArgRegsWin64;
1325 XMMArgRegs = XMMArgRegsWin64;
1327 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1328 GPR64ArgRegs = GPR64ArgRegs64Bit;
1329 XMMArgRegs = XMMArgRegs64Bit;
1331 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1333 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1336 // For X86-64, if there are vararg parameters that are passed via
1337 // registers, then we must store them to their spots on the stack so they
1338 // may be loaded by deferencing the result of va_next.
1339 VarArgsGPOffset = NumIntRegs * 8;
1340 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1341 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1342 TotalNumXMMRegs * 16, 16);
1344 // Store the integer parameter registers.
1345 SmallVector<SDValue, 8> MemOps;
1346 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1347 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1348 DAG.getIntPtrConstant(VarArgsGPOffset));
1349 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1350 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1351 X86::GR64RegisterClass);
1352 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1354 DAG.getStore(Val.getValue(1), Val, FIN,
1355 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1356 MemOps.push_back(Store);
1357 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1358 DAG.getIntPtrConstant(8));
1361 // Now store the XMM (fp + vector) parameter registers.
1362 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1363 DAG.getIntPtrConstant(VarArgsFPOffset));
1364 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1365 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1366 X86::VR128RegisterClass);
1367 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1369 DAG.getStore(Val.getValue(1), Val, FIN,
1370 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1371 MemOps.push_back(Store);
1372 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1373 DAG.getIntPtrConstant(16));
1375 if (!MemOps.empty())
1376 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1377 &MemOps[0], MemOps.size());
1381 ArgValues.push_back(Root);
1383 // Some CCs need callee pop.
1384 if (IsCalleePop(Op)) {
1385 BytesToPopOnReturn = StackSize; // Callee pops everything.
1386 BytesCallerReserves = 0;
1388 BytesToPopOnReturn = 0; // Callee pops nothing.
1389 // If this is an sret function, the return should pop the hidden pointer.
1390 if (!Is64Bit && ArgsAreStructReturn(Op))
1391 BytesToPopOnReturn = 4;
1392 BytesCallerReserves = StackSize;
1396 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1397 if (CC == CallingConv::X86_FastCall)
1398 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1401 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1403 // Return the new list of results.
1404 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1405 ArgValues.size()).getValue(Op.getResNo());
1409 X86TargetLowering::LowerMemOpCallTo(SDValue Op, SelectionDAG &DAG,
1410 const SDValue &StackPtr,
1411 const CCValAssign &VA,
1414 unsigned LocMemOffset = VA.getLocMemOffset();
1415 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1416 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1417 ISD::ArgFlagsTy Flags =
1418 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1419 if (Flags.isByVal()) {
1420 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1422 return DAG.getStore(Chain, Arg, PtrOff,
1423 PseudoSourceValue::getStack(), LocMemOffset);
1426 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1427 /// optimization is performed and it is required.
1429 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1430 SDValue &OutRetAddr,
1435 if (!IsTailCall || FPDiff==0) return Chain;
1437 // Adjust the Return address stack slot.
1438 MVT VT = getPointerTy();
1439 OutRetAddr = getReturnAddressFrameIndex(DAG);
1440 // Load the "old" Return address.
1441 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1442 return SDValue(OutRetAddr.getNode(), 1);
1445 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1446 /// optimization is performed and it is required (FPDiff!=0).
1448 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1449 SDValue Chain, SDValue RetAddrFrIdx,
1450 bool Is64Bit, int FPDiff) {
1451 // Store the return address to the appropriate stack slot.
1452 if (!FPDiff) return Chain;
1453 // Calculate the new stack slot for the return address.
1454 int SlotSize = Is64Bit ? 8 : 4;
1455 int NewReturnAddrFI =
1456 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1457 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1458 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1459 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1460 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1464 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1465 MachineFunction &MF = DAG.getMachineFunction();
1466 SDValue Chain = Op.getOperand(0);
1467 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1468 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1469 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1470 && CC == CallingConv::Fast && PerformTailCallOpt;
1471 SDValue Callee = Op.getOperand(4);
1472 bool Is64Bit = Subtarget->is64Bit();
1473 bool IsStructRet = CallIsStructReturn(Op);
1475 assert(!(isVarArg && CC == CallingConv::Fast) &&
1476 "Var args not supported with calling convention fastcc");
1478 // Analyze operands of the call, assigning locations to each operand.
1479 SmallVector<CCValAssign, 16> ArgLocs;
1480 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1481 CCInfo.AnalyzeCallOperands(Op.getNode(), CCAssignFnForNode(Op));
1483 // Get a count of how many bytes are to be pushed on the stack.
1484 unsigned NumBytes = CCInfo.getNextStackOffset();
1486 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1490 // Lower arguments at fp - stackoffset + fpdiff.
1491 unsigned NumBytesCallerPushed =
1492 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1493 FPDiff = NumBytesCallerPushed - NumBytes;
1495 // Set the delta of movement of the returnaddr stackslot.
1496 // But only set if delta is greater than previous delta.
1497 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1498 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1501 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1503 SDValue RetAddrFrIdx;
1504 // Load return adress for tail calls.
1505 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1508 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1509 SmallVector<SDValue, 8> MemOpChains;
1512 // Walk the register/memloc assignments, inserting copies/loads. In the case
1513 // of tail call optimization arguments are handle later.
1514 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1515 CCValAssign &VA = ArgLocs[i];
1516 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1517 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1518 getArgFlags().isByVal();
1520 // Promote the value if needed.
1521 switch (VA.getLocInfo()) {
1522 default: assert(0 && "Unknown loc info!");
1523 case CCValAssign::Full: break;
1524 case CCValAssign::SExt:
1525 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1527 case CCValAssign::ZExt:
1528 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1530 case CCValAssign::AExt:
1531 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1535 if (VA.isRegLoc()) {
1537 MVT RegVT = VA.getLocVT();
1538 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1539 switch (VA.getLocReg()) {
1542 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1544 // Special case: passing MMX values in GPR registers.
1545 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1548 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1549 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1550 // Special case: passing MMX values in XMM registers.
1551 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1552 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1553 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1554 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1555 getMOVLMask(2, DAG));
1560 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1562 if (!IsTailCall || (IsTailCall && isByVal)) {
1563 assert(VA.isMemLoc());
1564 if (StackPtr.getNode() == 0)
1565 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1567 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1573 if (!MemOpChains.empty())
1574 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1575 &MemOpChains[0], MemOpChains.size());
1577 // Build a sequence of copy-to-reg nodes chained together with token chain
1578 // and flag operands which copy the outgoing args into registers.
1580 // Tail call byval lowering might overwrite argument registers so in case of
1581 // tail call optimization the copies to registers are lowered later.
1583 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1584 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1586 InFlag = Chain.getValue(1);
1589 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1591 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1592 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1593 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1595 InFlag = Chain.getValue(1);
1597 // If we are tail calling and generating PIC/GOT style code load the address
1598 // of the callee into ecx. The value in ecx is used as target of the tail
1599 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1600 // calls on PIC/GOT architectures. Normally we would just put the address of
1601 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1602 // restored (since ebx is callee saved) before jumping to the target@PLT.
1603 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1604 // Note: The actual moving to ecx is done further down.
1605 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1606 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1607 !G->getGlobal()->hasProtectedVisibility())
1608 Callee = LowerGlobalAddress(Callee, DAG);
1609 else if (isa<ExternalSymbolSDNode>(Callee))
1610 Callee = LowerExternalSymbol(Callee,DAG);
1613 if (Is64Bit && isVarArg) {
1614 // From AMD64 ABI document:
1615 // For calls that may call functions that use varargs or stdargs
1616 // (prototype-less calls or calls to functions containing ellipsis (...) in
1617 // the declaration) %al is used as hidden argument to specify the number
1618 // of SSE registers used. The contents of %al do not need to match exactly
1619 // the number of registers, but must be an ubound on the number of SSE
1620 // registers used and is in the range 0 - 8 inclusive.
1622 // FIXME: Verify this on Win64
1623 // Count the number of XMM registers allocated.
1624 static const unsigned XMMArgRegs[] = {
1625 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1626 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1628 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1630 Chain = DAG.getCopyToReg(Chain, X86::AL,
1631 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1632 InFlag = Chain.getValue(1);
1636 // For tail calls lower the arguments to the 'real' stack slot.
1638 SmallVector<SDValue, 8> MemOpChains2;
1641 // Do not flag preceeding copytoreg stuff together with the following stuff.
1643 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1644 CCValAssign &VA = ArgLocs[i];
1645 if (!VA.isRegLoc()) {
1646 assert(VA.isMemLoc());
1647 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1648 SDValue FlagsOp = Op.getOperand(6+2*VA.getValNo());
1649 ISD::ArgFlagsTy Flags =
1650 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1651 // Create frame index.
1652 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1653 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1654 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1655 FIN = DAG.getFrameIndex(FI, getPointerTy());
1657 if (Flags.isByVal()) {
1658 // Copy relative to framepointer.
1659 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1660 if (StackPtr.getNode() == 0)
1661 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1662 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1664 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1667 // Store relative to framepointer.
1668 MemOpChains2.push_back(
1669 DAG.getStore(Chain, Arg, FIN,
1670 PseudoSourceValue::getFixedStack(FI), 0));
1675 if (!MemOpChains2.empty())
1676 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1677 &MemOpChains2[0], MemOpChains2.size());
1679 // Copy arguments to their registers.
1680 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1681 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1683 InFlag = Chain.getValue(1);
1687 // Store the return address to the appropriate stack slot.
1688 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1692 // If the callee is a GlobalAddress node (quite common, every direct call is)
1693 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1694 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1695 // We should use extra load for direct calls to dllimported functions in
1697 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1698 getTargetMachine(), true))
1699 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1700 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1701 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1702 } else if (IsTailCall) {
1703 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1705 Chain = DAG.getCopyToReg(Chain,
1706 DAG.getRegister(Opc, getPointerTy()),
1708 Callee = DAG.getRegister(Opc, getPointerTy());
1709 // Add register as live out.
1710 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1713 // Returns a chain & a flag for retval copy to use.
1714 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1715 SmallVector<SDValue, 8> Ops;
1718 Ops.push_back(Chain);
1719 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1720 Ops.push_back(DAG.getIntPtrConstant(0));
1721 if (InFlag.getNode())
1722 Ops.push_back(InFlag);
1723 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1724 InFlag = Chain.getValue(1);
1726 // Returns a chain & a flag for retval copy to use.
1727 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1731 Ops.push_back(Chain);
1732 Ops.push_back(Callee);
1735 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1737 // Add argument registers to the end of the list so that they are known live
1739 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1740 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1741 RegsToPass[i].second.getValueType()));
1743 // Add an implicit use GOT pointer in EBX.
1744 if (!IsTailCall && !Is64Bit &&
1745 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1746 Subtarget->isPICStyleGOT())
1747 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1749 // Add an implicit use of AL for x86 vararg functions.
1750 if (Is64Bit && isVarArg)
1751 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1753 if (InFlag.getNode())
1754 Ops.push_back(InFlag);
1757 assert(InFlag.getNode() &&
1758 "Flag must be set. Depend on flag being set in LowerRET");
1759 Chain = DAG.getNode(X86ISD::TAILCALL,
1760 Op.getNode()->getVTList(), &Ops[0], Ops.size());
1762 return SDValue(Chain.getNode(), Op.getResNo());
1765 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1766 InFlag = Chain.getValue(1);
1768 // Create the CALLSEQ_END node.
1769 unsigned NumBytesForCalleeToPush;
1770 if (IsCalleePop(Op))
1771 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1772 else if (!Is64Bit && IsStructRet)
1773 // If this is is a call to a struct-return function, the callee
1774 // pops the hidden struct pointer, so we have to push it back.
1775 // This is common for Darwin/X86, Linux & Mingw32 targets.
1776 NumBytesForCalleeToPush = 4;
1778 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1780 // Returns a flag for retval copy to use.
1781 Chain = DAG.getCALLSEQ_END(Chain,
1782 DAG.getIntPtrConstant(NumBytes),
1783 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1785 InFlag = Chain.getValue(1);
1787 // Handle result values, copying them out of physregs into vregs that we
1789 return SDValue(LowerCallResult(Chain, InFlag, Op.getNode(), CC, DAG),
1794 //===----------------------------------------------------------------------===//
1795 // Fast Calling Convention (tail call) implementation
1796 //===----------------------------------------------------------------------===//
1798 // Like std call, callee cleans arguments, convention except that ECX is
1799 // reserved for storing the tail called function address. Only 2 registers are
1800 // free for argument passing (inreg). Tail call optimization is performed
1802 // * tailcallopt is enabled
1803 // * caller/callee are fastcc
1804 // On X86_64 architecture with GOT-style position independent code only local
1805 // (within module) calls are supported at the moment.
1806 // To keep the stack aligned according to platform abi the function
1807 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1808 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1809 // If a tail called function callee has more arguments than the caller the
1810 // caller needs to make sure that there is room to move the RETADDR to. This is
1811 // achieved by reserving an area the size of the argument delta right after the
1812 // original REtADDR, but before the saved framepointer or the spilled registers
1813 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1825 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1826 /// for a 16 byte align requirement.
1827 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1828 SelectionDAG& DAG) {
1829 MachineFunction &MF = DAG.getMachineFunction();
1830 const TargetMachine &TM = MF.getTarget();
1831 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1832 unsigned StackAlignment = TFI.getStackAlignment();
1833 uint64_t AlignMask = StackAlignment - 1;
1834 int64_t Offset = StackSize;
1835 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1836 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1837 // Number smaller than 12 so just add the difference.
1838 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1840 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1841 Offset = ((~AlignMask) & Offset) + StackAlignment +
1842 (StackAlignment-SlotSize);
1847 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1848 /// following the call is a return. A function is eligible if caller/callee
1849 /// calling conventions match, currently only fastcc supports tail calls, and
1850 /// the function CALL is immediatly followed by a RET.
1851 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1853 SelectionDAG& DAG) const {
1854 if (!PerformTailCallOpt)
1857 if (CheckTailCallReturnConstraints(Call, Ret)) {
1858 MachineFunction &MF = DAG.getMachineFunction();
1859 unsigned CallerCC = MF.getFunction()->getCallingConv();
1860 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1861 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1862 SDValue Callee = Call.getOperand(4);
1863 // On x86/32Bit PIC/GOT tail calls are supported.
1864 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1865 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1868 // Can only do local tail calls (in same module, hidden or protected) on
1869 // x86_64 PIC/GOT at the moment.
1870 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1871 return G->getGlobal()->hasHiddenVisibility()
1872 || G->getGlobal()->hasProtectedVisibility();
1880 X86TargetLowering::createFastISel(MachineFunction &mf,
1881 DenseMap<const Value *, unsigned> &vm,
1882 DenseMap<const BasicBlock *,
1883 MachineBasicBlock *> &bm) {
1884 return X86::createFastISel(mf, vm, bm);
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());
1912 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1913 /// specific condition code. It returns a false if it cannot do a direct
1914 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1916 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1917 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1918 SelectionDAG &DAG) {
1919 X86CC = X86::COND_INVALID;
1921 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1922 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1923 // X > -1 -> X == 0, jump !sign.
1924 RHS = DAG.getConstant(0, RHS.getValueType());
1925 X86CC = X86::COND_NS;
1927 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1928 // X < 0 -> X == 0, jump on sign.
1929 X86CC = X86::COND_S;
1931 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1933 RHS = DAG.getConstant(0, RHS.getValueType());
1934 X86CC = X86::COND_LE;
1939 switch (SetCCOpcode) {
1941 case ISD::SETEQ: X86CC = X86::COND_E; break;
1942 case ISD::SETGT: X86CC = X86::COND_G; break;
1943 case ISD::SETGE: X86CC = X86::COND_GE; break;
1944 case ISD::SETLT: X86CC = X86::COND_L; break;
1945 case ISD::SETLE: X86CC = X86::COND_LE; break;
1946 case ISD::SETNE: X86CC = X86::COND_NE; break;
1947 case ISD::SETULT: X86CC = X86::COND_B; break;
1948 case ISD::SETUGT: X86CC = X86::COND_A; break;
1949 case ISD::SETULE: X86CC = X86::COND_BE; break;
1950 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1953 // First determine if it requires or is profitable to flip the operands.
1955 switch (SetCCOpcode) {
1965 // If LHS is a foldable load, but RHS is not, flip the condition.
1967 (ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
1968 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
1969 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
1973 std::swap(LHS, RHS);
1975 // On a floating point condition, the flags are set as follows:
1977 // 0 | 0 | 0 | X > Y
1978 // 0 | 0 | 1 | X < Y
1979 // 1 | 0 | 0 | X == Y
1980 // 1 | 1 | 1 | unordered
1981 switch (SetCCOpcode) {
1985 X86CC = X86::COND_E;
1987 case ISD::SETOLT: // flipped
1990 X86CC = X86::COND_A;
1992 case ISD::SETOLE: // flipped
1995 X86CC = X86::COND_AE;
1997 case ISD::SETUGT: // flipped
2000 X86CC = X86::COND_B;
2002 case ISD::SETUGE: // flipped
2005 X86CC = X86::COND_BE;
2009 X86CC = X86::COND_NE;
2012 X86CC = X86::COND_P;
2015 X86CC = X86::COND_NP;
2020 return X86CC != X86::COND_INVALID;
2023 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2024 /// code. Current x86 isa includes the following FP cmov instructions:
2025 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2026 static bool hasFPCMov(unsigned X86CC) {
2042 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2043 /// true if Op is undef or if its value falls within the specified range (L, H].
2044 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2045 if (Op.getOpcode() == ISD::UNDEF)
2048 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2049 return (Val >= Low && Val < Hi);
2052 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2053 /// true if Op is undef or if its value equal to the specified value.
2054 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2055 if (Op.getOpcode() == ISD::UNDEF)
2057 return cast<ConstantSDNode>(Op)->getValue() == Val;
2060 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2061 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2062 bool X86::isPSHUFDMask(SDNode *N) {
2063 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2065 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2068 // Check if the value doesn't reference the second vector.
2069 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2070 SDValue Arg = N->getOperand(i);
2071 if (Arg.getOpcode() == ISD::UNDEF) continue;
2072 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2073 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2080 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2081 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2082 bool X86::isPSHUFHWMask(SDNode *N) {
2083 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2085 if (N->getNumOperands() != 8)
2088 // Lower quadword copied in order.
2089 for (unsigned i = 0; i != 4; ++i) {
2090 SDValue Arg = N->getOperand(i);
2091 if (Arg.getOpcode() == ISD::UNDEF) continue;
2092 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2093 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2097 // Upper quadword shuffled.
2098 for (unsigned i = 4; i != 8; ++i) {
2099 SDValue Arg = N->getOperand(i);
2100 if (Arg.getOpcode() == ISD::UNDEF) continue;
2101 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2102 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2103 if (Val < 4 || Val > 7)
2110 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2111 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2112 bool X86::isPSHUFLWMask(SDNode *N) {
2113 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2115 if (N->getNumOperands() != 8)
2118 // Upper quadword copied in order.
2119 for (unsigned i = 4; i != 8; ++i)
2120 if (!isUndefOrEqual(N->getOperand(i), i))
2123 // Lower quadword shuffled.
2124 for (unsigned i = 0; i != 4; ++i)
2125 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2131 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2132 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2133 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2134 if (NumElems != 2 && NumElems != 4) return false;
2136 unsigned Half = NumElems / 2;
2137 for (unsigned i = 0; i < Half; ++i)
2138 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2140 for (unsigned i = Half; i < NumElems; ++i)
2141 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2147 bool X86::isSHUFPMask(SDNode *N) {
2148 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2149 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2152 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2153 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2154 /// half elements to come from vector 1 (which would equal the dest.) and
2155 /// the upper half to come from vector 2.
2156 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2157 if (NumOps != 2 && NumOps != 4) return false;
2159 unsigned Half = NumOps / 2;
2160 for (unsigned i = 0; i < Half; ++i)
2161 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2163 for (unsigned i = Half; i < NumOps; ++i)
2164 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2169 static bool isCommutedSHUFP(SDNode *N) {
2170 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2171 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2174 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2175 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2176 bool X86::isMOVHLPSMask(SDNode *N) {
2177 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2179 if (N->getNumOperands() != 4)
2182 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2183 return isUndefOrEqual(N->getOperand(0), 6) &&
2184 isUndefOrEqual(N->getOperand(1), 7) &&
2185 isUndefOrEqual(N->getOperand(2), 2) &&
2186 isUndefOrEqual(N->getOperand(3), 3);
2189 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2190 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2192 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2193 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2195 if (N->getNumOperands() != 4)
2198 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2199 return isUndefOrEqual(N->getOperand(0), 2) &&
2200 isUndefOrEqual(N->getOperand(1), 3) &&
2201 isUndefOrEqual(N->getOperand(2), 2) &&
2202 isUndefOrEqual(N->getOperand(3), 3);
2205 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2206 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2207 bool X86::isMOVLPMask(SDNode *N) {
2208 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2210 unsigned NumElems = N->getNumOperands();
2211 if (NumElems != 2 && NumElems != 4)
2214 for (unsigned i = 0; i < NumElems/2; ++i)
2215 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2218 for (unsigned i = NumElems/2; i < NumElems; ++i)
2219 if (!isUndefOrEqual(N->getOperand(i), i))
2225 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2226 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2228 bool X86::isMOVHPMask(SDNode *N) {
2229 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2231 unsigned NumElems = N->getNumOperands();
2232 if (NumElems != 2 && NumElems != 4)
2235 for (unsigned i = 0; i < NumElems/2; ++i)
2236 if (!isUndefOrEqual(N->getOperand(i), i))
2239 for (unsigned i = 0; i < NumElems/2; ++i) {
2240 SDValue Arg = N->getOperand(i + NumElems/2);
2241 if (!isUndefOrEqual(Arg, i + NumElems))
2248 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2249 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2250 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2251 bool V2IsSplat = false) {
2252 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2255 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2256 SDValue BitI = Elts[i];
2257 SDValue BitI1 = Elts[i+1];
2258 if (!isUndefOrEqual(BitI, j))
2261 if (isUndefOrEqual(BitI1, NumElts))
2264 if (!isUndefOrEqual(BitI1, j + NumElts))
2272 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2273 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2274 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2277 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2278 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2279 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2280 bool V2IsSplat = false) {
2281 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2284 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2285 SDValue BitI = Elts[i];
2286 SDValue BitI1 = Elts[i+1];
2287 if (!isUndefOrEqual(BitI, j + NumElts/2))
2290 if (isUndefOrEqual(BitI1, NumElts))
2293 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2301 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2302 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2303 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2306 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2307 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2309 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2310 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2312 unsigned NumElems = N->getNumOperands();
2313 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2316 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2317 SDValue BitI = N->getOperand(i);
2318 SDValue BitI1 = N->getOperand(i+1);
2320 if (!isUndefOrEqual(BitI, j))
2322 if (!isUndefOrEqual(BitI1, j))
2329 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2330 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2332 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2333 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2335 unsigned NumElems = N->getNumOperands();
2336 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2339 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2340 SDValue BitI = N->getOperand(i);
2341 SDValue BitI1 = N->getOperand(i + 1);
2343 if (!isUndefOrEqual(BitI, j))
2345 if (!isUndefOrEqual(BitI1, j))
2352 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2353 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2354 /// MOVSD, and MOVD, i.e. setting the lowest element.
2355 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2356 if (NumElts != 2 && NumElts != 4)
2359 if (!isUndefOrEqual(Elts[0], NumElts))
2362 for (unsigned i = 1; i < NumElts; ++i) {
2363 if (!isUndefOrEqual(Elts[i], i))
2370 bool X86::isMOVLMask(SDNode *N) {
2371 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2372 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2375 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2376 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2377 /// element of vector 2 and the other elements to come from vector 1 in order.
2378 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2379 bool V2IsSplat = false,
2380 bool V2IsUndef = false) {
2381 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2384 if (!isUndefOrEqual(Ops[0], 0))
2387 for (unsigned i = 1; i < NumOps; ++i) {
2388 SDValue Arg = Ops[i];
2389 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2390 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2391 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2398 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2399 bool V2IsUndef = false) {
2400 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2401 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2402 V2IsSplat, V2IsUndef);
2405 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2406 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2407 bool X86::isMOVSHDUPMask(SDNode *N) {
2408 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2410 if (N->getNumOperands() != 4)
2413 // Expect 1, 1, 3, 3
2414 for (unsigned i = 0; i < 2; ++i) {
2415 SDValue Arg = N->getOperand(i);
2416 if (Arg.getOpcode() == ISD::UNDEF) continue;
2417 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2418 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2419 if (Val != 1) return false;
2423 for (unsigned i = 2; i < 4; ++i) {
2424 SDValue Arg = N->getOperand(i);
2425 if (Arg.getOpcode() == ISD::UNDEF) continue;
2426 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2427 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2428 if (Val != 3) return false;
2432 // Don't use movshdup if it can be done with a shufps.
2436 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2437 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2438 bool X86::isMOVSLDUPMask(SDNode *N) {
2439 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2441 if (N->getNumOperands() != 4)
2444 // Expect 0, 0, 2, 2
2445 for (unsigned i = 0; i < 2; ++i) {
2446 SDValue Arg = N->getOperand(i);
2447 if (Arg.getOpcode() == ISD::UNDEF) continue;
2448 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2449 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2450 if (Val != 0) return false;
2454 for (unsigned i = 2; i < 4; ++i) {
2455 SDValue Arg = N->getOperand(i);
2456 if (Arg.getOpcode() == ISD::UNDEF) continue;
2457 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2458 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2459 if (Val != 2) return false;
2463 // Don't use movshdup if it can be done with a shufps.
2467 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2468 /// specifies a identity operation on the LHS or RHS.
2469 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2470 unsigned NumElems = N->getNumOperands();
2471 for (unsigned i = 0; i < NumElems; ++i)
2472 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2477 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2478 /// a splat of a single element.
2479 static bool isSplatMask(SDNode *N) {
2480 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2482 // This is a splat operation if each element of the permute is the same, and
2483 // if the value doesn't reference the second vector.
2484 unsigned NumElems = N->getNumOperands();
2485 SDValue ElementBase;
2487 for (; i != NumElems; ++i) {
2488 SDValue Elt = N->getOperand(i);
2489 if (isa<ConstantSDNode>(Elt)) {
2495 if (!ElementBase.getNode())
2498 for (; i != NumElems; ++i) {
2499 SDValue Arg = N->getOperand(i);
2500 if (Arg.getOpcode() == ISD::UNDEF) continue;
2501 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2502 if (Arg != ElementBase) return false;
2505 // Make sure it is a splat of the first vector operand.
2506 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2509 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2510 /// a splat of a single element and it's a 2 or 4 element mask.
2511 bool X86::isSplatMask(SDNode *N) {
2512 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2514 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2515 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2517 return ::isSplatMask(N);
2520 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2521 /// specifies a splat of zero element.
2522 bool X86::isSplatLoMask(SDNode *N) {
2523 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2525 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2526 if (!isUndefOrEqual(N->getOperand(i), 0))
2531 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2532 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2534 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2535 unsigned NumOperands = N->getNumOperands();
2536 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2538 for (unsigned i = 0; i < NumOperands; ++i) {
2540 SDValue Arg = N->getOperand(NumOperands-i-1);
2541 if (Arg.getOpcode() != ISD::UNDEF)
2542 Val = cast<ConstantSDNode>(Arg)->getValue();
2543 if (Val >= NumOperands) Val -= NumOperands;
2545 if (i != NumOperands - 1)
2552 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2553 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2555 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2557 // 8 nodes, but we only care about the last 4.
2558 for (unsigned i = 7; i >= 4; --i) {
2560 SDValue Arg = N->getOperand(i);
2561 if (Arg.getOpcode() != ISD::UNDEF)
2562 Val = cast<ConstantSDNode>(Arg)->getValue();
2571 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2572 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2574 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2576 // 8 nodes, but we only care about the first 4.
2577 for (int i = 3; i >= 0; --i) {
2579 SDValue Arg = N->getOperand(i);
2580 if (Arg.getOpcode() != ISD::UNDEF)
2581 Val = cast<ConstantSDNode>(Arg)->getValue();
2590 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2591 /// specifies a 8 element shuffle that can be broken into a pair of
2592 /// PSHUFHW and PSHUFLW.
2593 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2594 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2596 if (N->getNumOperands() != 8)
2599 // Lower quadword shuffled.
2600 for (unsigned i = 0; i != 4; ++i) {
2601 SDValue Arg = N->getOperand(i);
2602 if (Arg.getOpcode() == ISD::UNDEF) continue;
2603 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2604 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2609 // Upper quadword shuffled.
2610 for (unsigned i = 4; i != 8; ++i) {
2611 SDValue Arg = N->getOperand(i);
2612 if (Arg.getOpcode() == ISD::UNDEF) continue;
2613 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2614 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2615 if (Val < 4 || Val > 7)
2622 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2623 /// values in ther permute mask.
2624 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2625 SDValue &V2, SDValue &Mask,
2626 SelectionDAG &DAG) {
2627 MVT VT = Op.getValueType();
2628 MVT MaskVT = Mask.getValueType();
2629 MVT EltVT = MaskVT.getVectorElementType();
2630 unsigned NumElems = Mask.getNumOperands();
2631 SmallVector<SDValue, 8> MaskVec;
2633 for (unsigned i = 0; i != NumElems; ++i) {
2634 SDValue Arg = Mask.getOperand(i);
2635 if (Arg.getOpcode() == ISD::UNDEF) {
2636 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2639 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2640 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2642 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2644 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2648 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2649 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2652 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2653 /// the two vector operands have swapped position.
2655 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2656 MVT MaskVT = Mask.getValueType();
2657 MVT EltVT = MaskVT.getVectorElementType();
2658 unsigned NumElems = Mask.getNumOperands();
2659 SmallVector<SDValue, 8> MaskVec;
2660 for (unsigned i = 0; i != NumElems; ++i) {
2661 SDValue Arg = Mask.getOperand(i);
2662 if (Arg.getOpcode() == ISD::UNDEF) {
2663 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2666 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2667 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2669 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2671 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2673 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2677 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2678 /// match movhlps. The lower half elements should come from upper half of
2679 /// V1 (and in order), and the upper half elements should come from the upper
2680 /// half of V2 (and in order).
2681 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2682 unsigned NumElems = Mask->getNumOperands();
2685 for (unsigned i = 0, e = 2; i != e; ++i)
2686 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2688 for (unsigned i = 2; i != 4; ++i)
2689 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2694 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2695 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2697 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2698 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2699 N = N->getOperand(0).getNode();
2700 if (ISD::isNON_EXTLoad(N)) {
2702 *LD = cast<LoadSDNode>(N);
2709 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2710 /// match movlp{s|d}. The lower half elements should come from lower half of
2711 /// V1 (and in order), and the upper half elements should come from the upper
2712 /// half of V2 (and in order). And since V1 will become the source of the
2713 /// MOVLP, it must be either a vector load or a scalar load to vector.
2714 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2715 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2717 // Is V2 is a vector load, don't do this transformation. We will try to use
2718 // load folding shufps op.
2719 if (ISD::isNON_EXTLoad(V2))
2722 unsigned NumElems = Mask->getNumOperands();
2723 if (NumElems != 2 && NumElems != 4)
2725 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2726 if (!isUndefOrEqual(Mask->getOperand(i), i))
2728 for (unsigned i = NumElems/2; i != NumElems; ++i)
2729 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2734 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2736 static bool isSplatVector(SDNode *N) {
2737 if (N->getOpcode() != ISD::BUILD_VECTOR)
2740 SDValue SplatValue = N->getOperand(0);
2741 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2742 if (N->getOperand(i) != SplatValue)
2747 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2749 static bool isUndefShuffle(SDNode *N) {
2750 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2753 SDValue V1 = N->getOperand(0);
2754 SDValue V2 = N->getOperand(1);
2755 SDValue Mask = N->getOperand(2);
2756 unsigned NumElems = Mask.getNumOperands();
2757 for (unsigned i = 0; i != NumElems; ++i) {
2758 SDValue Arg = Mask.getOperand(i);
2759 if (Arg.getOpcode() != ISD::UNDEF) {
2760 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2761 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2763 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2770 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2772 static inline bool isZeroNode(SDValue Elt) {
2773 return ((isa<ConstantSDNode>(Elt) &&
2774 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2775 (isa<ConstantFPSDNode>(Elt) &&
2776 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2779 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2780 /// to an zero vector.
2781 static bool isZeroShuffle(SDNode *N) {
2782 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2785 SDValue V1 = N->getOperand(0);
2786 SDValue V2 = N->getOperand(1);
2787 SDValue Mask = N->getOperand(2);
2788 unsigned NumElems = Mask.getNumOperands();
2789 for (unsigned i = 0; i != NumElems; ++i) {
2790 SDValue Arg = Mask.getOperand(i);
2791 if (Arg.getOpcode() == ISD::UNDEF)
2794 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2795 if (Idx < NumElems) {
2796 unsigned Opc = V1.getNode()->getOpcode();
2797 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2799 if (Opc != ISD::BUILD_VECTOR ||
2800 !isZeroNode(V1.getNode()->getOperand(Idx)))
2802 } else if (Idx >= NumElems) {
2803 unsigned Opc = V2.getNode()->getOpcode();
2804 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2806 if (Opc != ISD::BUILD_VECTOR ||
2807 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2814 /// getZeroVector - Returns a vector of specified type with all zero elements.
2816 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2817 assert(VT.isVector() && "Expected a vector type");
2819 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2820 // type. This ensures they get CSE'd.
2822 if (VT.getSizeInBits() == 64) { // MMX
2823 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2824 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2825 } else if (HasSSE2) { // SSE2
2826 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2827 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2829 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2830 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2832 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2835 /// getOnesVector - Returns a vector of specified type with all bits set.
2837 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2838 assert(VT.isVector() && "Expected a vector type");
2840 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2841 // type. This ensures they get CSE'd.
2842 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2844 if (VT.getSizeInBits() == 64) // MMX
2845 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2847 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2848 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2852 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2853 /// that point to V2 points to its first element.
2854 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2855 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2857 bool Changed = false;
2858 SmallVector<SDValue, 8> MaskVec;
2859 unsigned NumElems = Mask.getNumOperands();
2860 for (unsigned i = 0; i != NumElems; ++i) {
2861 SDValue Arg = Mask.getOperand(i);
2862 if (Arg.getOpcode() != ISD::UNDEF) {
2863 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2864 if (Val > NumElems) {
2865 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2869 MaskVec.push_back(Arg);
2873 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2874 &MaskVec[0], MaskVec.size());
2878 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2879 /// operation of specified width.
2880 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2881 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2882 MVT BaseVT = MaskVT.getVectorElementType();
2884 SmallVector<SDValue, 8> MaskVec;
2885 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2886 for (unsigned i = 1; i != NumElems; ++i)
2887 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2888 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2891 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2892 /// of specified width.
2893 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2894 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2895 MVT BaseVT = MaskVT.getVectorElementType();
2896 SmallVector<SDValue, 8> MaskVec;
2897 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2898 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2899 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2901 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2904 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2905 /// of specified width.
2906 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2907 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2908 MVT BaseVT = MaskVT.getVectorElementType();
2909 unsigned Half = NumElems/2;
2910 SmallVector<SDValue, 8> MaskVec;
2911 for (unsigned i = 0; i != Half; ++i) {
2912 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2913 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2915 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2918 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2919 /// element #0 of a vector with the specified index, leaving the rest of the
2920 /// elements in place.
2921 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2922 SelectionDAG &DAG) {
2923 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2924 MVT BaseVT = MaskVT.getVectorElementType();
2925 SmallVector<SDValue, 8> MaskVec;
2926 // Element #0 of the result gets the elt we are replacing.
2927 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2928 for (unsigned i = 1; i != NumElems; ++i)
2929 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2930 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2933 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2934 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2935 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2936 MVT VT = Op.getValueType();
2939 SDValue V1 = Op.getOperand(0);
2940 SDValue Mask = Op.getOperand(2);
2941 unsigned NumElems = Mask.getNumOperands();
2942 // Special handling of v4f32 -> v4i32.
2943 if (VT != MVT::v4f32) {
2944 Mask = getUnpacklMask(NumElems, DAG);
2945 while (NumElems > 4) {
2946 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2949 Mask = getZeroVector(MVT::v4i32, true, DAG);
2952 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2953 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2954 DAG.getNode(ISD::UNDEF, PVT), Mask);
2955 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2958 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2959 /// vector of zero or undef vector. This produces a shuffle where the low
2960 /// element of V2 is swizzled into the zero/undef vector, landing at element
2961 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2962 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2963 bool isZero, bool HasSSE2,
2964 SelectionDAG &DAG) {
2965 MVT VT = V2.getValueType();
2967 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
2968 unsigned NumElems = V2.getValueType().getVectorNumElements();
2969 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2970 MVT EVT = MaskVT.getVectorElementType();
2971 SmallVector<SDValue, 16> MaskVec;
2972 for (unsigned i = 0; i != NumElems; ++i)
2973 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2974 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2976 MaskVec.push_back(DAG.getConstant(i, EVT));
2977 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2978 &MaskVec[0], MaskVec.size());
2979 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2982 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2983 /// a shuffle that is zero.
2985 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
2986 unsigned NumElems, bool Low,
2987 SelectionDAG &DAG) {
2988 unsigned NumZeros = 0;
2989 for (unsigned i = 0; i < NumElems; ++i) {
2990 unsigned Index = Low ? i : NumElems-i-1;
2991 SDValue Idx = Mask.getOperand(Index);
2992 if (Idx.getOpcode() == ISD::UNDEF) {
2996 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
2997 if (Elt.getNode() && isZeroNode(Elt))
3005 /// isVectorShift - Returns true if the shuffle can be implemented as a
3006 /// logical left or right shift of a vector.
3007 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3008 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3009 unsigned NumElems = Mask.getNumOperands();
3012 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3015 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3020 bool SeenV1 = false;
3021 bool SeenV2 = false;
3022 for (unsigned i = NumZeros; i < NumElems; ++i) {
3023 unsigned Val = isLeft ? (i - NumZeros) : i;
3024 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3025 if (Idx.getOpcode() == ISD::UNDEF)
3027 unsigned Index = cast<ConstantSDNode>(Idx)->getValue();
3028 if (Index < NumElems)
3037 if (SeenV1 && SeenV2)
3040 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3046 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3048 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3049 unsigned NumNonZero, unsigned NumZero,
3050 SelectionDAG &DAG, TargetLowering &TLI) {
3056 for (unsigned i = 0; i < 16; ++i) {
3057 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3058 if (ThisIsNonZero && First) {
3060 V = getZeroVector(MVT::v8i16, true, DAG);
3062 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3067 SDValue ThisElt(0, 0), LastElt(0, 0);
3068 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3069 if (LastIsNonZero) {
3070 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3072 if (ThisIsNonZero) {
3073 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3074 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3075 ThisElt, DAG.getConstant(8, MVT::i8));
3077 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3081 if (ThisElt.getNode())
3082 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3083 DAG.getIntPtrConstant(i/2));
3087 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3090 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3092 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3093 unsigned NumNonZero, unsigned NumZero,
3094 SelectionDAG &DAG, TargetLowering &TLI) {
3100 for (unsigned i = 0; i < 8; ++i) {
3101 bool isNonZero = (NonZeros & (1 << i)) != 0;
3105 V = getZeroVector(MVT::v8i16, true, DAG);
3107 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3110 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3111 DAG.getIntPtrConstant(i));
3118 /// getVShift - Return a vector logical shift node.
3120 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3121 unsigned NumBits, SelectionDAG &DAG,
3122 const TargetLowering &TLI) {
3123 bool isMMX = VT.getSizeInBits() == 64;
3124 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3125 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3126 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3127 return DAG.getNode(ISD::BIT_CONVERT, VT,
3128 DAG.getNode(Opc, ShVT, SrcOp,
3129 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3133 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3134 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3135 if (ISD::isBuildVectorAllZeros(Op.getNode())
3136 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3137 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3138 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3139 // eliminated on x86-32 hosts.
3140 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3143 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3144 return getOnesVector(Op.getValueType(), DAG);
3145 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3148 MVT VT = Op.getValueType();
3149 MVT EVT = VT.getVectorElementType();
3150 unsigned EVTBits = EVT.getSizeInBits();
3152 unsigned NumElems = Op.getNumOperands();
3153 unsigned NumZero = 0;
3154 unsigned NumNonZero = 0;
3155 unsigned NonZeros = 0;
3156 bool IsAllConstants = true;
3157 SmallSet<SDValue, 8> Values;
3158 for (unsigned i = 0; i < NumElems; ++i) {
3159 SDValue Elt = Op.getOperand(i);
3160 if (Elt.getOpcode() == ISD::UNDEF)
3163 if (Elt.getOpcode() != ISD::Constant &&
3164 Elt.getOpcode() != ISD::ConstantFP)
3165 IsAllConstants = false;
3166 if (isZeroNode(Elt))
3169 NonZeros |= (1 << i);
3174 if (NumNonZero == 0) {
3175 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3176 return DAG.getNode(ISD::UNDEF, VT);
3179 // Special case for single non-zero, non-undef, element.
3180 if (NumNonZero == 1 && NumElems <= 4) {
3181 unsigned Idx = CountTrailingZeros_32(NonZeros);
3182 SDValue Item = Op.getOperand(Idx);
3184 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3185 // the value are obviously zero, truncate the value to i32 and do the
3186 // insertion that way. Only do this if the value is non-constant or if the
3187 // value is a constant being inserted into element 0. It is cheaper to do
3188 // a constant pool load than it is to do a movd + shuffle.
3189 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3190 (!IsAllConstants || Idx == 0)) {
3191 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3192 // Handle MMX and SSE both.
3193 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3194 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3196 // Truncate the value (which may itself be a constant) to i32, and
3197 // convert it to a vector with movd (S2V+shuffle to zero extend).
3198 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3199 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3200 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3201 Subtarget->hasSSE2(), DAG);
3203 // Now we have our 32-bit value zero extended in the low element of
3204 // a vector. If Idx != 0, swizzle it into place.
3207 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3208 getSwapEltZeroMask(VecElts, Idx, DAG)
3210 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3212 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3216 // If we have a constant or non-constant insertion into the low element of
3217 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3218 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3219 // depending on what the source datatype is. Because we can only get here
3220 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3222 // Don't do this for i64 values on x86-32.
3223 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3224 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3225 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3226 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3227 Subtarget->hasSSE2(), DAG);
3230 // Is it a vector logical left shift?
3231 if (NumElems == 2 && Idx == 1 &&
3232 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3233 unsigned NumBits = VT.getSizeInBits();
3234 return getVShift(true, VT,
3235 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3236 NumBits/2, DAG, *this);
3239 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3242 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3243 // is a non-constant being inserted into an element other than the low one,
3244 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3245 // movd/movss) to move this into the low element, then shuffle it into
3247 if (EVTBits == 32) {
3248 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3250 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3251 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3252 Subtarget->hasSSE2(), DAG);
3253 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3254 MVT MaskEVT = MaskVT.getVectorElementType();
3255 SmallVector<SDValue, 8> MaskVec;
3256 for (unsigned i = 0; i < NumElems; i++)
3257 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3258 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3259 &MaskVec[0], MaskVec.size());
3260 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3261 DAG.getNode(ISD::UNDEF, VT), Mask);
3265 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3266 if (Values.size() == 1)
3269 // A vector full of immediates; various special cases are already
3270 // handled, so this is best done with a single constant-pool load.
3274 // Let legalizer expand 2-wide build_vectors.
3275 if (EVTBits == 64) {
3276 if (NumNonZero == 1) {
3277 // One half is zero or undef.
3278 unsigned Idx = CountTrailingZeros_32(NonZeros);
3279 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3280 Op.getOperand(Idx));
3281 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3282 Subtarget->hasSSE2(), DAG);
3287 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3288 if (EVTBits == 8 && NumElems == 16) {
3289 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3291 if (V.getNode()) return V;
3294 if (EVTBits == 16 && NumElems == 8) {
3295 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3297 if (V.getNode()) return V;
3300 // If element VT is == 32 bits, turn it into a number of shuffles.
3301 SmallVector<SDValue, 8> V;
3303 if (NumElems == 4 && NumZero > 0) {
3304 for (unsigned i = 0; i < 4; ++i) {
3305 bool isZero = !(NonZeros & (1 << i));
3307 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3309 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3312 for (unsigned i = 0; i < 2; ++i) {
3313 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3316 V[i] = V[i*2]; // Must be a zero vector.
3319 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3320 getMOVLMask(NumElems, DAG));
3323 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3324 getMOVLMask(NumElems, DAG));
3327 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3328 getUnpacklMask(NumElems, DAG));
3333 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3334 MVT EVT = MaskVT.getVectorElementType();
3335 SmallVector<SDValue, 8> MaskVec;
3336 bool Reverse = (NonZeros & 0x3) == 2;
3337 for (unsigned i = 0; i < 2; ++i)
3339 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3341 MaskVec.push_back(DAG.getConstant(i, EVT));
3342 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3343 for (unsigned i = 0; i < 2; ++i)
3345 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3347 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3348 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3349 &MaskVec[0], MaskVec.size());
3350 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3353 if (Values.size() > 2) {
3354 // Expand into a number of unpckl*.
3356 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3357 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3358 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3359 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3360 for (unsigned i = 0; i < NumElems; ++i)
3361 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3363 while (NumElems != 0) {
3364 for (unsigned i = 0; i < NumElems; ++i)
3365 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3376 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3377 SDValue PermMask, SelectionDAG &DAG,
3378 TargetLowering &TLI) {
3380 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3381 MVT MaskEVT = MaskVT.getVectorElementType();
3382 MVT PtrVT = TLI.getPointerTy();
3383 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3384 PermMask.getNode()->op_end());
3386 // First record which half of which vector the low elements come from.
3387 SmallVector<unsigned, 4> LowQuad(4);
3388 for (unsigned i = 0; i < 4; ++i) {
3389 SDValue Elt = MaskElts[i];
3390 if (Elt.getOpcode() == ISD::UNDEF)
3392 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3393 int QuadIdx = EltIdx / 4;
3397 int BestLowQuad = -1;
3398 unsigned MaxQuad = 1;
3399 for (unsigned i = 0; i < 4; ++i) {
3400 if (LowQuad[i] > MaxQuad) {
3402 MaxQuad = LowQuad[i];
3406 // Record which half of which vector the high elements come from.
3407 SmallVector<unsigned, 4> HighQuad(4);
3408 for (unsigned i = 4; i < 8; ++i) {
3409 SDValue Elt = MaskElts[i];
3410 if (Elt.getOpcode() == ISD::UNDEF)
3412 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3413 int QuadIdx = EltIdx / 4;
3414 ++HighQuad[QuadIdx];
3417 int BestHighQuad = -1;
3419 for (unsigned i = 0; i < 4; ++i) {
3420 if (HighQuad[i] > MaxQuad) {
3422 MaxQuad = HighQuad[i];
3426 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3427 if (BestLowQuad != -1 || BestHighQuad != -1) {
3428 // First sort the 4 chunks in order using shufpd.
3429 SmallVector<SDValue, 8> MaskVec;
3431 if (BestLowQuad != -1)
3432 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3434 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3436 if (BestHighQuad != -1)
3437 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3439 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3441 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3442 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3443 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3444 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3445 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3447 // Now sort high and low parts separately.
3448 BitVector InOrder(8);
3449 if (BestLowQuad != -1) {
3450 // Sort lower half in order using PSHUFLW.
3452 bool AnyOutOrder = false;
3454 for (unsigned i = 0; i != 4; ++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, MaskEVT));
3466 // If this element is in the right place after this shuffle, then
3468 if ((int)(EltIdx / 4) == BestLowQuad)
3473 for (unsigned i = 4; i != 8; ++i)
3474 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3475 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3476 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3480 if (BestHighQuad != -1) {
3481 // Sort high half in order using PSHUFHW if possible.
3484 for (unsigned i = 0; i != 4; ++i)
3485 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3487 bool AnyOutOrder = false;
3488 for (unsigned i = 4; i != 8; ++i) {
3489 SDValue Elt = MaskElts[i];
3490 if (Elt.getOpcode() == ISD::UNDEF) {
3491 MaskVec.push_back(Elt);
3494 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3498 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3500 // If this element is in the right place after this shuffle, then
3502 if ((int)(EltIdx / 4) == BestHighQuad)
3508 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3509 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3513 // The other elements are put in the right place using pextrw and pinsrw.
3514 for (unsigned i = 0; i != 8; ++i) {
3517 SDValue Elt = MaskElts[i];
3518 if (Elt.getOpcode() == ISD::UNDEF)
3520 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3521 SDValue ExtOp = (EltIdx < 8)
3522 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3523 DAG.getConstant(EltIdx, PtrVT))
3524 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3525 DAG.getConstant(EltIdx - 8, PtrVT));
3526 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3527 DAG.getConstant(i, PtrVT));
3533 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3534 // few as possible. First, let's find out how many elements are already in the
3536 unsigned V1InOrder = 0;
3537 unsigned V1FromV1 = 0;
3538 unsigned V2InOrder = 0;
3539 unsigned V2FromV2 = 0;
3540 SmallVector<SDValue, 8> V1Elts;
3541 SmallVector<SDValue, 8> V2Elts;
3542 for (unsigned i = 0; i < 8; ++i) {
3543 SDValue Elt = MaskElts[i];
3544 if (Elt.getOpcode() == ISD::UNDEF) {
3545 V1Elts.push_back(Elt);
3546 V2Elts.push_back(Elt);
3551 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3553 V1Elts.push_back(Elt);
3554 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3556 } else if (EltIdx == i+8) {
3557 V1Elts.push_back(Elt);
3558 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3560 } else if (EltIdx < 8) {
3561 V1Elts.push_back(Elt);
3564 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3569 if (V2InOrder > V1InOrder) {
3570 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3572 std::swap(V1Elts, V2Elts);
3573 std::swap(V1FromV1, V2FromV2);
3576 if ((V1FromV1 + V1InOrder) != 8) {
3577 // Some elements are from V2.
3579 // If there are elements that are from V1 but out of place,
3580 // then first sort them in place
3581 SmallVector<SDValue, 8> MaskVec;
3582 for (unsigned i = 0; i < 8; ++i) {
3583 SDValue Elt = V1Elts[i];
3584 if (Elt.getOpcode() == ISD::UNDEF) {
3585 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3588 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3590 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3592 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3594 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3595 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3599 for (unsigned i = 0; i < 8; ++i) {
3600 SDValue Elt = V1Elts[i];
3601 if (Elt.getOpcode() == ISD::UNDEF)
3603 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3606 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3607 DAG.getConstant(EltIdx - 8, PtrVT));
3608 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3609 DAG.getConstant(i, PtrVT));
3613 // All elements are from V1.
3615 for (unsigned i = 0; i < 8; ++i) {
3616 SDValue Elt = V1Elts[i];
3617 if (Elt.getOpcode() == ISD::UNDEF)
3619 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3620 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3621 DAG.getConstant(EltIdx, PtrVT));
3622 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3623 DAG.getConstant(i, PtrVT));
3629 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3630 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3631 /// done when every pair / quad of shuffle mask elements point to elements in
3632 /// the right sequence. e.g.
3633 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3635 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3637 SDValue PermMask, SelectionDAG &DAG,
3638 TargetLowering &TLI) {
3639 unsigned NumElems = PermMask.getNumOperands();
3640 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3641 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3642 MVT MaskEltVT = MaskVT.getVectorElementType();
3644 switch (VT.getSimpleVT()) {
3645 default: assert(false && "Unexpected!");
3646 case MVT::v4f32: NewVT = MVT::v2f64; break;
3647 case MVT::v4i32: NewVT = MVT::v2i64; break;
3648 case MVT::v8i16: NewVT = MVT::v4i32; break;
3649 case MVT::v16i8: NewVT = MVT::v4i32; break;
3652 if (NewWidth == 2) {
3658 unsigned Scale = NumElems / NewWidth;
3659 SmallVector<SDValue, 8> MaskVec;
3660 for (unsigned i = 0; i < NumElems; i += Scale) {
3661 unsigned StartIdx = ~0U;
3662 for (unsigned j = 0; j < Scale; ++j) {
3663 SDValue Elt = PermMask.getOperand(i+j);
3664 if (Elt.getOpcode() == ISD::UNDEF)
3666 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3667 if (StartIdx == ~0U)
3668 StartIdx = EltIdx - (EltIdx % Scale);
3669 if (EltIdx != StartIdx + j)
3672 if (StartIdx == ~0U)
3673 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3675 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3678 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3679 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3680 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3681 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3682 &MaskVec[0], MaskVec.size()));
3685 /// getVZextMovL - Return a zero-extending vector move low node.
3687 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3688 SDValue SrcOp, SelectionDAG &DAG,
3689 const X86Subtarget *Subtarget) {
3690 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3691 LoadSDNode *LD = NULL;
3692 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3693 LD = dyn_cast<LoadSDNode>(SrcOp);
3695 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3697 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3698 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3699 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3700 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3701 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3703 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3704 return DAG.getNode(ISD::BIT_CONVERT, VT,
3705 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3706 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3713 return DAG.getNode(ISD::BIT_CONVERT, VT,
3714 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3715 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3718 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3721 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3722 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3723 MVT MaskVT = PermMask.getValueType();
3724 MVT MaskEVT = MaskVT.getVectorElementType();
3725 SmallVector<std::pair<int, int>, 8> Locs;
3727 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3730 for (unsigned i = 0; i != 4; ++i) {
3731 SDValue Elt = PermMask.getOperand(i);
3732 if (Elt.getOpcode() == ISD::UNDEF) {
3733 Locs[i] = std::make_pair(-1, -1);
3735 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3736 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3738 Locs[i] = std::make_pair(0, NumLo);
3742 Locs[i] = std::make_pair(1, NumHi);
3744 Mask1[2+NumHi] = Elt;
3750 if (NumLo <= 2 && NumHi <= 2) {
3751 // If no more than two elements come from either vector. This can be
3752 // implemented with two shuffles. First shuffle gather the elements.
3753 // The second shuffle, which takes the first shuffle as both of its
3754 // vector operands, put the elements into the right order.
3755 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3756 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3757 &Mask1[0], Mask1.size()));
3759 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3760 for (unsigned i = 0; i != 4; ++i) {
3761 if (Locs[i].first == -1)
3764 unsigned Idx = (i < 2) ? 0 : 4;
3765 Idx += Locs[i].first * 2 + Locs[i].second;
3766 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3770 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3771 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3772 &Mask2[0], Mask2.size()));
3773 } else if (NumLo == 3 || NumHi == 3) {
3774 // Otherwise, we must have three elements from one vector, call it X, and
3775 // one element from the other, call it Y. First, use a shufps to build an
3776 // intermediate vector with the one element from Y and the element from X
3777 // that will be in the same half in the final destination (the indexes don't
3778 // matter). Then, use a shufps to build the final vector, taking the half
3779 // containing the element from Y from the intermediate, and the other half
3782 // Normalize it so the 3 elements come from V1.
3783 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3787 // Find the element from V2.
3789 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3790 SDValue Elt = PermMask.getOperand(HiIndex);
3791 if (Elt.getOpcode() == ISD::UNDEF)
3793 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3798 Mask1[0] = PermMask.getOperand(HiIndex);
3799 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3800 Mask1[2] = PermMask.getOperand(HiIndex^1);
3801 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3802 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3803 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3806 Mask1[0] = PermMask.getOperand(0);
3807 Mask1[1] = PermMask.getOperand(1);
3808 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3809 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3810 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3811 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3813 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3814 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3815 Mask1[2] = PermMask.getOperand(2);
3816 Mask1[3] = PermMask.getOperand(3);
3817 if (Mask1[2].getOpcode() != ISD::UNDEF)
3818 Mask1[2] = DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getValue()+4,
3820 if (Mask1[3].getOpcode() != ISD::UNDEF)
3821 Mask1[3] = DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getValue()+4,
3823 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3824 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3828 // Break it into (shuffle shuffle_hi, shuffle_lo).
3830 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3831 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3832 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3833 unsigned MaskIdx = 0;
3836 for (unsigned i = 0; i != 4; ++i) {
3843 SDValue Elt = PermMask.getOperand(i);
3844 if (Elt.getOpcode() == ISD::UNDEF) {
3845 Locs[i] = std::make_pair(-1, -1);
3846 } else if (cast<ConstantSDNode>(Elt)->getValue() < 4) {
3847 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3848 (*MaskPtr)[LoIdx] = Elt;
3851 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3852 (*MaskPtr)[HiIdx] = Elt;
3857 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3858 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3859 &LoMask[0], LoMask.size()));
3860 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3861 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3862 &HiMask[0], HiMask.size()));
3863 SmallVector<SDValue, 8> MaskOps;
3864 for (unsigned i = 0; i != 4; ++i) {
3865 if (Locs[i].first == -1) {
3866 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3868 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3869 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3872 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3873 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3874 &MaskOps[0], MaskOps.size()));
3878 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3879 SDValue V1 = Op.getOperand(0);
3880 SDValue V2 = Op.getOperand(1);
3881 SDValue PermMask = Op.getOperand(2);
3882 MVT VT = Op.getValueType();
3883 unsigned NumElems = PermMask.getNumOperands();
3884 bool isMMX = VT.getSizeInBits() == 64;
3885 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3886 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3887 bool V1IsSplat = false;
3888 bool V2IsSplat = false;
3890 if (isUndefShuffle(Op.getNode()))
3891 return DAG.getNode(ISD::UNDEF, VT);
3893 if (isZeroShuffle(Op.getNode()))
3894 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3896 if (isIdentityMask(PermMask.getNode()))
3898 else if (isIdentityMask(PermMask.getNode(), true))
3901 if (isSplatMask(PermMask.getNode())) {
3902 if (isMMX || NumElems < 4) return Op;
3903 // Promote it to a v4{if}32 splat.
3904 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3907 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3909 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3910 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3911 if (NewOp.getNode())
3912 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3913 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3914 // FIXME: Figure out a cleaner way to do this.
3915 // Try to make use of movq to zero out the top part.
3916 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3917 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3919 if (NewOp.getNode()) {
3920 SDValue NewV1 = NewOp.getOperand(0);
3921 SDValue NewV2 = NewOp.getOperand(1);
3922 SDValue NewMask = NewOp.getOperand(2);
3923 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3924 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3925 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3928 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3929 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3931 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3932 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3937 // Check if this can be converted into a logical shift.
3938 bool isLeft = false;
3941 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
3942 if (isShift && ShVal.hasOneUse()) {
3943 // If the shifted value has multiple uses, it may be cheaper to use
3944 // v_set0 + movlhps or movhlps, etc.
3945 MVT EVT = VT.getVectorElementType();
3946 ShAmt *= EVT.getSizeInBits();
3947 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3950 if (X86::isMOVLMask(PermMask.getNode())) {
3953 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3954 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
3959 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
3960 X86::isMOVSLDUPMask(PermMask.getNode()) ||
3961 X86::isMOVHLPSMask(PermMask.getNode()) ||
3962 X86::isMOVHPMask(PermMask.getNode()) ||
3963 X86::isMOVLPMask(PermMask.getNode())))
3966 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
3967 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
3968 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3971 // No better options. Use a vshl / vsrl.
3972 MVT EVT = VT.getVectorElementType();
3973 ShAmt *= EVT.getSizeInBits();
3974 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3977 bool Commuted = false;
3978 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3979 // 1,1,1,1 -> v8i16 though.
3980 V1IsSplat = isSplatVector(V1.getNode());
3981 V2IsSplat = isSplatVector(V2.getNode());
3983 // Canonicalize the splat or undef, if present, to be on the RHS.
3984 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3985 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3986 std::swap(V1IsSplat, V2IsSplat);
3987 std::swap(V1IsUndef, V2IsUndef);
3991 // FIXME: Figure out a cleaner way to do this.
3992 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
3993 if (V2IsUndef) return V1;
3994 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3996 // V2 is a splat, so the mask may be malformed. That is, it may point
3997 // to any V2 element. The instruction selectior won't like this. Get
3998 // a corrected mask and commute to form a proper MOVS{S|D}.
3999 SDValue NewMask = getMOVLMask(NumElems, DAG);
4000 if (NewMask.getNode() != PermMask.getNode())
4001 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4006 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4007 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4008 X86::isUNPCKLMask(PermMask.getNode()) ||
4009 X86::isUNPCKHMask(PermMask.getNode()))
4013 // Normalize mask so all entries that point to V2 points to its first
4014 // element then try to match unpck{h|l} again. If match, return a
4015 // new vector_shuffle with the corrected mask.
4016 SDValue NewMask = NormalizeMask(PermMask, DAG);
4017 if (NewMask.getNode() != PermMask.getNode()) {
4018 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4019 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4020 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4021 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4022 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4023 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4028 // Normalize the node to match x86 shuffle ops if needed
4029 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4030 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4033 // Commute is back and try unpck* again.
4034 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4035 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4036 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4037 X86::isUNPCKLMask(PermMask.getNode()) ||
4038 X86::isUNPCKHMask(PermMask.getNode()))
4042 // Try PSHUF* first, then SHUFP*.
4043 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4044 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4045 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4046 if (V2.getOpcode() != ISD::UNDEF)
4047 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4048 DAG.getNode(ISD::UNDEF, VT), PermMask);
4053 if (Subtarget->hasSSE2() &&
4054 (X86::isPSHUFDMask(PermMask.getNode()) ||
4055 X86::isPSHUFHWMask(PermMask.getNode()) ||
4056 X86::isPSHUFLWMask(PermMask.getNode()))) {
4058 if (VT == MVT::v4f32) {
4060 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4061 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4062 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4063 } else if (V2.getOpcode() != ISD::UNDEF)
4064 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4065 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4067 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4071 // Binary or unary shufps.
4072 if (X86::isSHUFPMask(PermMask.getNode()) ||
4073 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4077 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4078 if (VT == MVT::v8i16) {
4079 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4080 if (NewOp.getNode())
4084 // Handle all 4 wide cases with a number of shuffles except for MMX.
4085 if (NumElems == 4 && !isMMX)
4086 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4092 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4093 SelectionDAG &DAG) {
4094 MVT VT = Op.getValueType();
4095 if (VT.getSizeInBits() == 8) {
4096 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4097 Op.getOperand(0), Op.getOperand(1));
4098 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4099 DAG.getValueType(VT));
4100 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4101 } else if (VT.getSizeInBits() == 16) {
4102 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4103 Op.getOperand(0), Op.getOperand(1));
4104 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4105 DAG.getValueType(VT));
4106 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4107 } else if (VT == MVT::f32) {
4108 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4109 // the result back to FR32 register. It's only worth matching if the
4110 // result has a single use which is a store or a bitcast to i32.
4111 if (!Op.hasOneUse())
4113 SDNode *User = *Op.getNode()->use_begin();
4114 if (User->getOpcode() != ISD::STORE &&
4115 (User->getOpcode() != ISD::BIT_CONVERT ||
4116 User->getValueType(0) != MVT::i32))
4118 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4119 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4121 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4128 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4129 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4132 if (Subtarget->hasSSE41()) {
4133 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4138 MVT VT = Op.getValueType();
4139 // TODO: handle v16i8.
4140 if (VT.getSizeInBits() == 16) {
4141 SDValue Vec = Op.getOperand(0);
4142 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4144 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4145 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4146 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4148 // Transform it so it match pextrw which produces a 32-bit result.
4149 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4150 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4151 Op.getOperand(0), Op.getOperand(1));
4152 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4153 DAG.getValueType(VT));
4154 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4155 } else if (VT.getSizeInBits() == 32) {
4156 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4159 // SHUFPS the element to the lowest double word, then movss.
4160 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4161 SmallVector<SDValue, 8> IdxVec;
4163 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4165 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4167 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4169 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4170 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4171 &IdxVec[0], IdxVec.size());
4172 SDValue Vec = Op.getOperand(0);
4173 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4174 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4175 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4176 DAG.getIntPtrConstant(0));
4177 } else if (VT.getSizeInBits() == 64) {
4178 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4179 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4180 // to match extract_elt for f64.
4181 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4185 // UNPCKHPD the element to the lowest double word, then movsd.
4186 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4187 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4188 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4189 SmallVector<SDValue, 8> IdxVec;
4190 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4192 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4193 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4194 &IdxVec[0], IdxVec.size());
4195 SDValue Vec = Op.getOperand(0);
4196 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4197 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4198 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4199 DAG.getIntPtrConstant(0));
4206 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4207 MVT VT = Op.getValueType();
4208 MVT EVT = VT.getVectorElementType();
4210 SDValue N0 = Op.getOperand(0);
4211 SDValue N1 = Op.getOperand(1);
4212 SDValue N2 = Op.getOperand(2);
4214 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4215 isa<ConstantSDNode>(N2)) {
4216 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4218 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4220 if (N1.getValueType() != MVT::i32)
4221 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4222 if (N2.getValueType() != MVT::i32)
4223 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4224 return DAG.getNode(Opc, VT, N0, N1, N2);
4225 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4226 // Bits [7:6] of the constant are the source select. This will always be
4227 // zero here. The DAG Combiner may combine an extract_elt index into these
4228 // bits. For example (insert (extract, 3), 2) could be matched by putting
4229 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4230 // Bits [5:4] of the constant are the destination select. This is the
4231 // value of the incoming immediate.
4232 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4233 // combine either bitwise AND or insert of float 0.0 to set these bits.
4234 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
4235 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4241 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4242 MVT VT = Op.getValueType();
4243 MVT EVT = VT.getVectorElementType();
4245 if (Subtarget->hasSSE41())
4246 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4251 SDValue N0 = Op.getOperand(0);
4252 SDValue N1 = Op.getOperand(1);
4253 SDValue N2 = Op.getOperand(2);
4255 if (EVT.getSizeInBits() == 16) {
4256 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4257 // as its second argument.
4258 if (N1.getValueType() != MVT::i32)
4259 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4260 if (N2.getValueType() != MVT::i32)
4261 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4262 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4268 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4269 if (Op.getValueType() == MVT::v2f32)
4270 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4271 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4272 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4273 Op.getOperand(0))));
4275 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4276 MVT VT = MVT::v2i32;
4277 switch (Op.getValueType().getSimpleVT()) {
4284 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4285 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4288 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4289 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4290 // one of the above mentioned nodes. It has to be wrapped because otherwise
4291 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4292 // be used to form addressing mode. These wrapped nodes will be selected
4295 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4296 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4297 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4299 CP->getAlignment());
4300 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4301 // With PIC, the address is actually $g + Offset.
4302 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4303 !Subtarget->isPICStyleRIPRel()) {
4304 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4305 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4313 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4314 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4315 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4316 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4317 // With PIC, the address is actually $g + Offset.
4318 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4319 !Subtarget->isPICStyleRIPRel()) {
4320 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4321 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4325 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4326 // load the value at address GV, not the value of GV itself. This means that
4327 // the GlobalAddress must be in the base or index register of the address, not
4328 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4329 // The same applies for external symbols during PIC codegen
4330 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4331 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4332 PseudoSourceValue::getGOT(), 0);
4337 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4339 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4342 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4343 DAG.getNode(X86ISD::GlobalBaseReg,
4345 InFlag = Chain.getValue(1);
4347 // emit leal symbol@TLSGD(,%ebx,1), %eax
4348 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4349 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4350 GA->getValueType(0),
4352 SDValue Ops[] = { Chain, TGA, InFlag };
4353 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4354 InFlag = Result.getValue(2);
4355 Chain = Result.getValue(1);
4357 // call ___tls_get_addr. This function receives its argument in
4358 // the register EAX.
4359 Chain = DAG.getCopyToReg(Chain, X86::EAX, 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::EAX, PtrVT),
4367 DAG.getRegister(X86::EBX, PtrVT),
4369 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4370 InFlag = Chain.getValue(1);
4372 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4375 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4377 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4379 SDValue InFlag, Chain;
4381 // emit leaq symbol@TLSGD(%rip), %rdi
4382 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4383 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4384 GA->getValueType(0),
4386 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4387 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4388 Chain = Result.getValue(1);
4389 InFlag = Result.getValue(2);
4391 // call __tls_get_addr. This function receives its argument in
4392 // the register RDI.
4393 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4394 InFlag = Chain.getValue(1);
4396 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4397 SDValue Ops1[] = { Chain,
4398 DAG.getTargetExternalSymbol("__tls_get_addr",
4400 DAG.getRegister(X86::RDI, PtrVT),
4402 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4403 InFlag = Chain.getValue(1);
4405 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4408 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4409 // "local exec" model.
4410 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4412 // Get the Thread Pointer
4413 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4414 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4416 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4417 GA->getValueType(0),
4419 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4421 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4422 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4423 PseudoSourceValue::getGOT(), 0);
4425 // The address of the thread local variable is the add of the thread
4426 // pointer with the offset of the variable.
4427 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4431 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4432 // TODO: implement the "local dynamic" model
4433 // TODO: implement the "initial exec"model for pic executables
4434 assert(Subtarget->isTargetELF() &&
4435 "TLS not implemented for non-ELF targets");
4436 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4437 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4438 // otherwise use the "Local Exec"TLS Model
4439 if (Subtarget->is64Bit()) {
4440 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4442 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4443 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4445 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4450 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4451 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4452 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4453 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4454 // With PIC, the address is actually $g + Offset.
4455 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4456 !Subtarget->isPICStyleRIPRel()) {
4457 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4458 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4465 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4466 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4467 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4468 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4469 // With PIC, the address is actually $g + Offset.
4470 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4471 !Subtarget->isPICStyleRIPRel()) {
4472 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4473 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4480 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4481 /// take a 2 x i32 value to shift plus a shift amount.
4482 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4483 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4484 MVT VT = Op.getValueType();
4485 unsigned VTBits = VT.getSizeInBits();
4486 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4487 SDValue ShOpLo = Op.getOperand(0);
4488 SDValue ShOpHi = Op.getOperand(1);
4489 SDValue ShAmt = Op.getOperand(2);
4490 SDValue Tmp1 = isSRA ?
4491 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4492 DAG.getConstant(0, VT);
4495 if (Op.getOpcode() == ISD::SHL_PARTS) {
4496 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4497 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4499 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4500 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4503 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4504 DAG.getConstant(VTBits, MVT::i8));
4505 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4506 AndNode, DAG.getConstant(0, MVT::i8));
4509 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4510 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4511 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4513 if (Op.getOpcode() == ISD::SHL_PARTS) {
4514 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4515 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4517 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4518 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4521 SDValue Ops[2] = { Lo, Hi };
4522 return DAG.getMergeValues(Ops, 2);
4525 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4526 MVT SrcVT = Op.getOperand(0).getValueType();
4527 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4528 "Unknown SINT_TO_FP to lower!");
4530 // These are really Legal; caller falls through into that case.
4531 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4533 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4534 Subtarget->is64Bit())
4537 unsigned Size = SrcVT.getSizeInBits()/8;
4538 MachineFunction &MF = DAG.getMachineFunction();
4539 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4540 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4541 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4543 PseudoSourceValue::getFixedStack(SSFI), 0);
4547 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4549 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4551 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4552 SmallVector<SDValue, 8> Ops;
4553 Ops.push_back(Chain);
4554 Ops.push_back(StackSlot);
4555 Ops.push_back(DAG.getValueType(SrcVT));
4556 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4557 Tys, &Ops[0], Ops.size());
4560 Chain = Result.getValue(1);
4561 SDValue InFlag = Result.getValue(2);
4563 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4564 // shouldn't be necessary except that RFP cannot be live across
4565 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4566 MachineFunction &MF = DAG.getMachineFunction();
4567 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4568 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4569 Tys = DAG.getVTList(MVT::Other);
4570 SmallVector<SDValue, 8> Ops;
4571 Ops.push_back(Chain);
4572 Ops.push_back(Result);
4573 Ops.push_back(StackSlot);
4574 Ops.push_back(DAG.getValueType(Op.getValueType()));
4575 Ops.push_back(InFlag);
4576 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4577 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4578 PseudoSourceValue::getFixedStack(SSFI), 0);
4584 std::pair<SDValue,SDValue> X86TargetLowering::
4585 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4586 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4587 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4588 "Unknown FP_TO_SINT to lower!");
4590 // These are really Legal.
4591 if (Op.getValueType() == MVT::i32 &&
4592 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4593 return std::make_pair(SDValue(), SDValue());
4594 if (Subtarget->is64Bit() &&
4595 Op.getValueType() == MVT::i64 &&
4596 Op.getOperand(0).getValueType() != MVT::f80)
4597 return std::make_pair(SDValue(), SDValue());
4599 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4601 MachineFunction &MF = DAG.getMachineFunction();
4602 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4603 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4604 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4606 switch (Op.getValueType().getSimpleVT()) {
4607 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4608 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4609 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4610 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4613 SDValue Chain = DAG.getEntryNode();
4614 SDValue Value = Op.getOperand(0);
4615 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4616 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4617 Chain = DAG.getStore(Chain, Value, StackSlot,
4618 PseudoSourceValue::getFixedStack(SSFI), 0);
4619 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4621 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4623 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4624 Chain = Value.getValue(1);
4625 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4626 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4629 // Build the FP_TO_INT*_IN_MEM
4630 SDValue Ops[] = { Chain, Value, StackSlot };
4631 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4633 return std::make_pair(FIST, StackSlot);
4636 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4637 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4638 SDValue FIST = Vals.first, StackSlot = Vals.second;
4639 if (FIST.getNode() == 0) return SDValue();
4642 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4645 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4646 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4647 SDValue FIST = Vals.first, StackSlot = Vals.second;
4648 if (FIST.getNode() == 0) return 0;
4650 MVT VT = N->getValueType(0);
4652 // Return a load from the stack slot.
4653 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4655 // Use MERGE_VALUES to drop the chain result value and get a node with one
4656 // result. This requires turning off getMergeValues simplification, since
4657 // otherwise it will give us Res back.
4658 return DAG.getMergeValues(&Res, 1, false).getNode();
4661 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4662 MVT VT = Op.getValueType();
4665 EltVT = VT.getVectorElementType();
4666 std::vector<Constant*> CV;
4667 if (EltVT == MVT::f64) {
4668 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4672 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4678 Constant *C = ConstantVector::get(CV);
4679 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4680 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4681 PseudoSourceValue::getConstantPool(), 0,
4683 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4686 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4687 MVT VT = Op.getValueType();
4689 unsigned EltNum = 1;
4690 if (VT.isVector()) {
4691 EltVT = VT.getVectorElementType();
4692 EltNum = VT.getVectorNumElements();
4694 std::vector<Constant*> CV;
4695 if (EltVT == MVT::f64) {
4696 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4700 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4706 Constant *C = ConstantVector::get(CV);
4707 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4708 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4709 PseudoSourceValue::getConstantPool(), 0,
4711 if (VT.isVector()) {
4712 return DAG.getNode(ISD::BIT_CONVERT, VT,
4713 DAG.getNode(ISD::XOR, MVT::v2i64,
4714 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4715 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4717 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4721 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4722 SDValue Op0 = Op.getOperand(0);
4723 SDValue Op1 = Op.getOperand(1);
4724 MVT VT = Op.getValueType();
4725 MVT SrcVT = Op1.getValueType();
4727 // If second operand is smaller, extend it first.
4728 if (SrcVT.bitsLT(VT)) {
4729 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4732 // And if it is bigger, shrink it first.
4733 if (SrcVT.bitsGT(VT)) {
4734 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4738 // At this point the operands and the result should have the same
4739 // type, and that won't be f80 since that is not custom lowered.
4741 // First get the sign bit of second operand.
4742 std::vector<Constant*> CV;
4743 if (SrcVT == MVT::f64) {
4744 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4745 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4747 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4748 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4749 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4750 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4752 Constant *C = ConstantVector::get(CV);
4753 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4754 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4755 PseudoSourceValue::getConstantPool(), 0,
4757 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4759 // Shift sign bit right or left if the two operands have different types.
4760 if (SrcVT.bitsGT(VT)) {
4761 // Op0 is MVT::f32, Op1 is MVT::f64.
4762 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4763 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4764 DAG.getConstant(32, MVT::i32));
4765 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4766 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4767 DAG.getIntPtrConstant(0));
4770 // Clear first operand sign bit.
4772 if (VT == MVT::f64) {
4773 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4774 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4776 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4777 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4778 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4779 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4781 C = ConstantVector::get(CV);
4782 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4783 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4784 PseudoSourceValue::getConstantPool(), 0,
4786 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4788 // Or the value with the sign bit.
4789 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4792 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4793 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4795 SDValue Op0 = Op.getOperand(0);
4796 SDValue Op1 = Op.getOperand(1);
4797 SDValue CC = Op.getOperand(2);
4798 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4799 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4802 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4804 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4805 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4806 DAG.getConstant(X86CC, MVT::i8), Cond);
4809 assert(isFP && "Illegal integer SetCC!");
4811 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4812 switch (SetCCOpcode) {
4813 default: assert(false && "Illegal floating point SetCC!");
4814 case ISD::SETOEQ: { // !PF & ZF
4815 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4816 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4817 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4818 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4819 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4821 case ISD::SETUNE: { // PF | !ZF
4822 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4823 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4824 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4825 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4826 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4831 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4833 SDValue Op0 = Op.getOperand(0);
4834 SDValue Op1 = Op.getOperand(1);
4835 SDValue CC = Op.getOperand(2);
4836 MVT VT = Op.getValueType();
4837 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4838 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4842 MVT VT0 = Op0.getValueType();
4843 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4844 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4847 switch (SetCCOpcode) {
4850 case ISD::SETEQ: SSECC = 0; break;
4852 case ISD::SETGT: Swap = true; // Fallthrough
4854 case ISD::SETOLT: SSECC = 1; break;
4856 case ISD::SETGE: Swap = true; // Fallthrough
4858 case ISD::SETOLE: SSECC = 2; break;
4859 case ISD::SETUO: SSECC = 3; break;
4861 case ISD::SETNE: SSECC = 4; break;
4862 case ISD::SETULE: Swap = true;
4863 case ISD::SETUGE: SSECC = 5; break;
4864 case ISD::SETULT: Swap = true;
4865 case ISD::SETUGT: SSECC = 6; break;
4866 case ISD::SETO: SSECC = 7; break;
4869 std::swap(Op0, Op1);
4871 // In the two special cases we can't handle, emit two comparisons.
4873 if (SetCCOpcode == ISD::SETUEQ) {
4875 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4876 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4877 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4879 else if (SetCCOpcode == ISD::SETONE) {
4881 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4882 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4883 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4885 assert(0 && "Illegal FP comparison");
4887 // Handle all other FP comparisons here.
4888 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4891 // We are handling one of the integer comparisons here. Since SSE only has
4892 // GT and EQ comparisons for integer, swapping operands and multiple
4893 // operations may be required for some comparisons.
4894 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4895 bool Swap = false, Invert = false, FlipSigns = false;
4897 switch (VT.getSimpleVT()) {
4899 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4900 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4901 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4902 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4905 switch (SetCCOpcode) {
4907 case ISD::SETNE: Invert = true;
4908 case ISD::SETEQ: Opc = EQOpc; break;
4909 case ISD::SETLT: Swap = true;
4910 case ISD::SETGT: Opc = GTOpc; break;
4911 case ISD::SETGE: Swap = true;
4912 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4913 case ISD::SETULT: Swap = true;
4914 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4915 case ISD::SETUGE: Swap = true;
4916 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4919 std::swap(Op0, Op1);
4921 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4922 // bits of the inputs before performing those operations.
4924 MVT EltVT = VT.getVectorElementType();
4925 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4926 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4927 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4929 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
4930 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
4933 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
4935 // If the logical-not of the result is required, perform that now.
4937 MVT EltVT = VT.getVectorElementType();
4938 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
4939 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
4940 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
4942 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
4947 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
4948 bool addTest = true;
4949 SDValue Cond = Op.getOperand(0);
4952 if (Cond.getOpcode() == ISD::SETCC)
4953 Cond = LowerSETCC(Cond, DAG);
4955 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4956 // setting operand in place of the X86ISD::SETCC.
4957 if (Cond.getOpcode() == X86ISD::SETCC) {
4958 CC = Cond.getOperand(0);
4960 SDValue Cmp = Cond.getOperand(1);
4961 unsigned Opc = Cmp.getOpcode();
4962 MVT VT = Op.getValueType();
4964 bool IllegalFPCMov = false;
4965 if (VT.isFloatingPoint() && !VT.isVector() &&
4966 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4967 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4969 if ((Opc == X86ISD::CMP ||
4970 Opc == X86ISD::COMI ||
4971 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4978 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4979 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4982 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4984 SmallVector<SDValue, 4> Ops;
4985 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4986 // condition is true.
4987 Ops.push_back(Op.getOperand(2));
4988 Ops.push_back(Op.getOperand(1));
4990 Ops.push_back(Cond);
4991 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4994 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
4995 bool addTest = true;
4996 SDValue Chain = Op.getOperand(0);
4997 SDValue Cond = Op.getOperand(1);
4998 SDValue Dest = Op.getOperand(2);
5001 if (Cond.getOpcode() == ISD::SETCC)
5002 Cond = LowerSETCC(Cond, DAG);
5004 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5005 // setting operand in place of the X86ISD::SETCC.
5006 if (Cond.getOpcode() == X86ISD::SETCC) {
5007 CC = Cond.getOperand(0);
5009 SDValue Cmp = Cond.getOperand(1);
5010 unsigned Opc = Cmp.getOpcode();
5011 if (Opc == X86ISD::CMP ||
5012 Opc == X86ISD::COMI ||
5013 Opc == X86ISD::UCOMI) {
5020 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5021 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5023 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5024 Chain, Op.getOperand(2), CC, Cond);
5028 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5029 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5030 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5031 // that the guard pages used by the OS virtual memory manager are allocated in
5032 // correct sequence.
5034 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5035 SelectionDAG &DAG) {
5036 assert(Subtarget->isTargetCygMing() &&
5037 "This should be used only on Cygwin/Mingw targets");
5040 SDValue Chain = Op.getOperand(0);
5041 SDValue Size = Op.getOperand(1);
5042 // FIXME: Ensure alignment here
5046 MVT IntPtr = getPointerTy();
5047 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5049 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0));
5051 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5052 Flag = Chain.getValue(1);
5054 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5055 SDValue Ops[] = { Chain,
5056 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5057 DAG.getRegister(X86::EAX, IntPtr),
5058 DAG.getRegister(X86StackPtr, SPTy),
5060 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5061 Flag = Chain.getValue(1);
5063 Chain = DAG.getCALLSEQ_END(Chain,
5064 DAG.getIntPtrConstant(0),
5065 DAG.getIntPtrConstant(0),
5068 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5070 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5071 return DAG.getMergeValues(Ops1, 2);
5075 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5077 SDValue Dst, SDValue Src,
5078 SDValue Size, unsigned Align,
5079 const Value *DstSV, uint64_t DstSVOff) {
5080 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5082 /// If not DWORD aligned or size is more than the threshold, call the library.
5083 /// The libc version is likely to be faster for these cases. It can use the
5084 /// address value and run time information about the CPU.
5085 if ((Align & 3) != 0 ||
5087 ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
5088 SDValue InFlag(0, 0);
5090 // Check to see if there is a specialized entry-point for memory zeroing.
5091 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5092 if (const char *bzeroEntry =
5093 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5094 MVT IntPtr = getPointerTy();
5095 const Type *IntPtrTy = getTargetData()->getIntPtrType();
5096 TargetLowering::ArgListTy Args;
5097 TargetLowering::ArgListEntry Entry;
5099 Entry.Ty = IntPtrTy;
5100 Args.push_back(Entry);
5102 Args.push_back(Entry);
5103 std::pair<SDValue,SDValue> CallResult =
5104 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
5105 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
5107 return CallResult.second;
5110 // Otherwise have the target-independent code call memset.
5114 uint64_t SizeVal = ConstantSize->getValue();
5115 SDValue InFlag(0, 0);
5118 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5119 unsigned BytesLeft = 0;
5120 bool TwoRepStos = false;
5123 uint64_t Val = ValC->getValue() & 255;
5125 // If the value is a constant, then we can potentially use larger sets.
5126 switch (Align & 3) {
5127 case 2: // WORD aligned
5130 Val = (Val << 8) | Val;
5132 case 0: // DWORD aligned
5135 Val = (Val << 8) | Val;
5136 Val = (Val << 16) | Val;
5137 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5140 Val = (Val << 32) | Val;
5143 default: // Byte aligned
5146 Count = DAG.getIntPtrConstant(SizeVal);
5150 if (AVT.bitsGT(MVT::i8)) {
5151 unsigned UBytes = AVT.getSizeInBits() / 8;
5152 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5153 BytesLeft = SizeVal % UBytes;
5156 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5158 InFlag = Chain.getValue(1);
5161 Count = DAG.getIntPtrConstant(SizeVal);
5162 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5163 InFlag = Chain.getValue(1);
5166 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5168 InFlag = Chain.getValue(1);
5169 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5171 InFlag = Chain.getValue(1);
5173 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5174 SmallVector<SDValue, 8> Ops;
5175 Ops.push_back(Chain);
5176 Ops.push_back(DAG.getValueType(AVT));
5177 Ops.push_back(InFlag);
5178 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5181 InFlag = Chain.getValue(1);
5183 MVT CVT = Count.getValueType();
5184 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5185 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5186 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5188 InFlag = Chain.getValue(1);
5189 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5191 Ops.push_back(Chain);
5192 Ops.push_back(DAG.getValueType(MVT::i8));
5193 Ops.push_back(InFlag);
5194 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5195 } else if (BytesLeft) {
5196 // Handle the last 1 - 7 bytes.
5197 unsigned Offset = SizeVal - BytesLeft;
5198 MVT AddrVT = Dst.getValueType();
5199 MVT SizeVT = Size.getValueType();
5201 Chain = DAG.getMemset(Chain,
5202 DAG.getNode(ISD::ADD, AddrVT, Dst,
5203 DAG.getConstant(Offset, AddrVT)),
5205 DAG.getConstant(BytesLeft, SizeVT),
5206 Align, DstSV, DstSVOff + Offset);
5209 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5214 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5215 SDValue Chain, SDValue Dst, SDValue Src,
5216 SDValue Size, unsigned Align,
5218 const Value *DstSV, uint64_t DstSVOff,
5219 const Value *SrcSV, uint64_t SrcSVOff) {
5220 // This requires the copy size to be a constant, preferrably
5221 // within a subtarget-specific limit.
5222 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5225 uint64_t SizeVal = ConstantSize->getValue();
5226 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5229 /// If not DWORD aligned, call the library.
5230 if ((Align & 3) != 0)
5235 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5238 unsigned UBytes = AVT.getSizeInBits() / 8;
5239 unsigned CountVal = SizeVal / UBytes;
5240 SDValue Count = DAG.getIntPtrConstant(CountVal);
5241 unsigned BytesLeft = SizeVal % UBytes;
5243 SDValue InFlag(0, 0);
5244 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5246 InFlag = Chain.getValue(1);
5247 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5249 InFlag = Chain.getValue(1);
5250 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5252 InFlag = Chain.getValue(1);
5254 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5255 SmallVector<SDValue, 8> Ops;
5256 Ops.push_back(Chain);
5257 Ops.push_back(DAG.getValueType(AVT));
5258 Ops.push_back(InFlag);
5259 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5261 SmallVector<SDValue, 4> Results;
5262 Results.push_back(RepMovs);
5264 // Handle the last 1 - 7 bytes.
5265 unsigned Offset = SizeVal - BytesLeft;
5266 MVT DstVT = Dst.getValueType();
5267 MVT SrcVT = Src.getValueType();
5268 MVT SizeVT = Size.getValueType();
5269 Results.push_back(DAG.getMemcpy(Chain,
5270 DAG.getNode(ISD::ADD, DstVT, Dst,
5271 DAG.getConstant(Offset, DstVT)),
5272 DAG.getNode(ISD::ADD, SrcVT, Src,
5273 DAG.getConstant(Offset, SrcVT)),
5274 DAG.getConstant(BytesLeft, SizeVT),
5275 Align, AlwaysInline,
5276 DstSV, DstSVOff + Offset,
5277 SrcSV, SrcSVOff + Offset));
5280 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5283 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5284 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5285 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5286 SDValue TheChain = N->getOperand(0);
5287 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5288 if (Subtarget->is64Bit()) {
5289 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5290 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5291 MVT::i64, rax.getValue(2));
5292 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5293 DAG.getConstant(32, MVT::i8));
5295 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5298 return DAG.getMergeValues(Ops, 2).getNode();
5301 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5302 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5303 MVT::i32, eax.getValue(2));
5304 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5305 SDValue Ops[] = { eax, edx };
5306 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5308 // Use a MERGE_VALUES to return the value and chain.
5309 Ops[1] = edx.getValue(1);
5310 return DAG.getMergeValues(Ops, 2).getNode();
5313 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5314 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5316 if (!Subtarget->is64Bit()) {
5317 // vastart just stores the address of the VarArgsFrameIndex slot into the
5318 // memory location argument.
5319 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5320 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5324 // gp_offset (0 - 6 * 8)
5325 // fp_offset (48 - 48 + 8 * 16)
5326 // overflow_arg_area (point to parameters coming in memory).
5328 SmallVector<SDValue, 8> MemOps;
5329 SDValue FIN = Op.getOperand(1);
5331 SDValue Store = DAG.getStore(Op.getOperand(0),
5332 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5334 MemOps.push_back(Store);
5337 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5338 Store = DAG.getStore(Op.getOperand(0),
5339 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5341 MemOps.push_back(Store);
5343 // Store ptr to overflow_arg_area
5344 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5345 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5346 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5347 MemOps.push_back(Store);
5349 // Store ptr to reg_save_area.
5350 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5351 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5352 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5353 MemOps.push_back(Store);
5354 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5357 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5358 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5359 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5360 SDValue Chain = Op.getOperand(0);
5361 SDValue SrcPtr = Op.getOperand(1);
5362 SDValue SrcSV = Op.getOperand(2);
5364 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5369 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5370 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5371 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5372 SDValue Chain = Op.getOperand(0);
5373 SDValue DstPtr = Op.getOperand(1);
5374 SDValue SrcPtr = Op.getOperand(2);
5375 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5376 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5378 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5379 DAG.getIntPtrConstant(24), 8, false,
5380 DstSV, 0, SrcSV, 0);
5384 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5385 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5387 default: return SDValue(); // Don't custom lower most intrinsics.
5388 // Comparison intrinsics.
5389 case Intrinsic::x86_sse_comieq_ss:
5390 case Intrinsic::x86_sse_comilt_ss:
5391 case Intrinsic::x86_sse_comile_ss:
5392 case Intrinsic::x86_sse_comigt_ss:
5393 case Intrinsic::x86_sse_comige_ss:
5394 case Intrinsic::x86_sse_comineq_ss:
5395 case Intrinsic::x86_sse_ucomieq_ss:
5396 case Intrinsic::x86_sse_ucomilt_ss:
5397 case Intrinsic::x86_sse_ucomile_ss:
5398 case Intrinsic::x86_sse_ucomigt_ss:
5399 case Intrinsic::x86_sse_ucomige_ss:
5400 case Intrinsic::x86_sse_ucomineq_ss:
5401 case Intrinsic::x86_sse2_comieq_sd:
5402 case Intrinsic::x86_sse2_comilt_sd:
5403 case Intrinsic::x86_sse2_comile_sd:
5404 case Intrinsic::x86_sse2_comigt_sd:
5405 case Intrinsic::x86_sse2_comige_sd:
5406 case Intrinsic::x86_sse2_comineq_sd:
5407 case Intrinsic::x86_sse2_ucomieq_sd:
5408 case Intrinsic::x86_sse2_ucomilt_sd:
5409 case Intrinsic::x86_sse2_ucomile_sd:
5410 case Intrinsic::x86_sse2_ucomigt_sd:
5411 case Intrinsic::x86_sse2_ucomige_sd:
5412 case Intrinsic::x86_sse2_ucomineq_sd: {
5414 ISD::CondCode CC = ISD::SETCC_INVALID;
5417 case Intrinsic::x86_sse_comieq_ss:
5418 case Intrinsic::x86_sse2_comieq_sd:
5422 case Intrinsic::x86_sse_comilt_ss:
5423 case Intrinsic::x86_sse2_comilt_sd:
5427 case Intrinsic::x86_sse_comile_ss:
5428 case Intrinsic::x86_sse2_comile_sd:
5432 case Intrinsic::x86_sse_comigt_ss:
5433 case Intrinsic::x86_sse2_comigt_sd:
5437 case Intrinsic::x86_sse_comige_ss:
5438 case Intrinsic::x86_sse2_comige_sd:
5442 case Intrinsic::x86_sse_comineq_ss:
5443 case Intrinsic::x86_sse2_comineq_sd:
5447 case Intrinsic::x86_sse_ucomieq_ss:
5448 case Intrinsic::x86_sse2_ucomieq_sd:
5449 Opc = X86ISD::UCOMI;
5452 case Intrinsic::x86_sse_ucomilt_ss:
5453 case Intrinsic::x86_sse2_ucomilt_sd:
5454 Opc = X86ISD::UCOMI;
5457 case Intrinsic::x86_sse_ucomile_ss:
5458 case Intrinsic::x86_sse2_ucomile_sd:
5459 Opc = X86ISD::UCOMI;
5462 case Intrinsic::x86_sse_ucomigt_ss:
5463 case Intrinsic::x86_sse2_ucomigt_sd:
5464 Opc = X86ISD::UCOMI;
5467 case Intrinsic::x86_sse_ucomige_ss:
5468 case Intrinsic::x86_sse2_ucomige_sd:
5469 Opc = X86ISD::UCOMI;
5472 case Intrinsic::x86_sse_ucomineq_ss:
5473 case Intrinsic::x86_sse2_ucomineq_sd:
5474 Opc = X86ISD::UCOMI;
5480 SDValue LHS = Op.getOperand(1);
5481 SDValue RHS = Op.getOperand(2);
5482 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5484 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5485 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5486 DAG.getConstant(X86CC, MVT::i8), Cond);
5487 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5490 // Fix vector shift instructions where the last operand is a non-immediate
5492 case Intrinsic::x86_sse2_pslli_w:
5493 case Intrinsic::x86_sse2_pslli_d:
5494 case Intrinsic::x86_sse2_pslli_q:
5495 case Intrinsic::x86_sse2_psrli_w:
5496 case Intrinsic::x86_sse2_psrli_d:
5497 case Intrinsic::x86_sse2_psrli_q:
5498 case Intrinsic::x86_sse2_psrai_w:
5499 case Intrinsic::x86_sse2_psrai_d:
5500 case Intrinsic::x86_mmx_pslli_w:
5501 case Intrinsic::x86_mmx_pslli_d:
5502 case Intrinsic::x86_mmx_pslli_q:
5503 case Intrinsic::x86_mmx_psrli_w:
5504 case Intrinsic::x86_mmx_psrli_d:
5505 case Intrinsic::x86_mmx_psrli_q:
5506 case Intrinsic::x86_mmx_psrai_w:
5507 case Intrinsic::x86_mmx_psrai_d: {
5508 SDValue ShAmt = Op.getOperand(2);
5509 if (isa<ConstantSDNode>(ShAmt))
5512 unsigned NewIntNo = 0;
5513 MVT ShAmtVT = MVT::v4i32;
5515 case Intrinsic::x86_sse2_pslli_w:
5516 NewIntNo = Intrinsic::x86_sse2_psll_w;
5518 case Intrinsic::x86_sse2_pslli_d:
5519 NewIntNo = Intrinsic::x86_sse2_psll_d;
5521 case Intrinsic::x86_sse2_pslli_q:
5522 NewIntNo = Intrinsic::x86_sse2_psll_q;
5524 case Intrinsic::x86_sse2_psrli_w:
5525 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5527 case Intrinsic::x86_sse2_psrli_d:
5528 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5530 case Intrinsic::x86_sse2_psrli_q:
5531 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5533 case Intrinsic::x86_sse2_psrai_w:
5534 NewIntNo = Intrinsic::x86_sse2_psra_w;
5536 case Intrinsic::x86_sse2_psrai_d:
5537 NewIntNo = Intrinsic::x86_sse2_psra_d;
5540 ShAmtVT = MVT::v2i32;
5542 case Intrinsic::x86_mmx_pslli_w:
5543 NewIntNo = Intrinsic::x86_mmx_psll_w;
5545 case Intrinsic::x86_mmx_pslli_d:
5546 NewIntNo = Intrinsic::x86_mmx_psll_d;
5548 case Intrinsic::x86_mmx_pslli_q:
5549 NewIntNo = Intrinsic::x86_mmx_psll_q;
5551 case Intrinsic::x86_mmx_psrli_w:
5552 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5554 case Intrinsic::x86_mmx_psrli_d:
5555 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5557 case Intrinsic::x86_mmx_psrli_q:
5558 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5560 case Intrinsic::x86_mmx_psrai_w:
5561 NewIntNo = Intrinsic::x86_mmx_psra_w;
5563 case Intrinsic::x86_mmx_psrai_d:
5564 NewIntNo = Intrinsic::x86_mmx_psra_d;
5566 default: abort(); // Can't reach here.
5571 MVT VT = Op.getValueType();
5572 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5573 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5574 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5575 DAG.getConstant(NewIntNo, MVT::i32),
5576 Op.getOperand(1), ShAmt);
5581 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5582 // Depths > 0 not supported yet!
5583 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5586 // Just load the return address
5587 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5588 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5591 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5592 // Depths > 0 not supported yet!
5593 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5596 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5597 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5598 DAG.getIntPtrConstant(Subtarget->is64Bit() ? 8 : 4));
5601 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5602 SelectionDAG &DAG) {
5603 return DAG.getIntPtrConstant(Subtarget->is64Bit() ? 16 : 8);
5606 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5608 MachineFunction &MF = DAG.getMachineFunction();
5609 SDValue Chain = Op.getOperand(0);
5610 SDValue Offset = Op.getOperand(1);
5611 SDValue Handler = Op.getOperand(2);
5613 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5615 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5617 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5618 DAG.getIntPtrConstant(Subtarget->is64Bit() ?
5620 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5621 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5622 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5623 MF.getRegInfo().addLiveOut(StoreAddrReg);
5625 return DAG.getNode(X86ISD::EH_RETURN,
5627 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5630 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5631 SelectionDAG &DAG) {
5632 SDValue Root = Op.getOperand(0);
5633 SDValue Trmp = Op.getOperand(1); // trampoline
5634 SDValue FPtr = Op.getOperand(2); // nested function
5635 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5637 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5639 const X86InstrInfo *TII =
5640 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5642 if (Subtarget->is64Bit()) {
5643 SDValue OutChains[6];
5645 // Large code-model.
5647 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5648 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5650 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5651 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5653 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5655 // Load the pointer to the nested function into R11.
5656 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5657 SDValue Addr = Trmp;
5658 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5661 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5662 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5664 // Load the 'nest' parameter value into R10.
5665 // R10 is specified in X86CallingConv.td
5666 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5667 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5668 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5671 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5672 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5674 // Jump to the nested function.
5675 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5676 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5677 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5680 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5681 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5682 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5686 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5687 return DAG.getMergeValues(Ops, 2);
5689 const Function *Func =
5690 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5691 unsigned CC = Func->getCallingConv();
5696 assert(0 && "Unsupported calling convention");
5697 case CallingConv::C:
5698 case CallingConv::X86_StdCall: {
5699 // Pass 'nest' parameter in ECX.
5700 // Must be kept in sync with X86CallingConv.td
5703 // Check that ECX wasn't needed by an 'inreg' parameter.
5704 const FunctionType *FTy = Func->getFunctionType();
5705 const PAListPtr &Attrs = Func->getParamAttrs();
5707 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5708 unsigned InRegCount = 0;
5711 for (FunctionType::param_iterator I = FTy->param_begin(),
5712 E = FTy->param_end(); I != E; ++I, ++Idx)
5713 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5714 // FIXME: should only count parameters that are lowered to integers.
5715 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5717 if (InRegCount > 2) {
5718 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5724 case CallingConv::X86_FastCall:
5725 // Pass 'nest' parameter in EAX.
5726 // Must be kept in sync with X86CallingConv.td
5731 SDValue OutChains[4];
5734 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5735 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5737 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5738 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5739 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5742 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5743 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5745 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5746 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5747 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5748 TrmpAddr, 5, false, 1);
5750 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5751 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5754 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5755 return DAG.getMergeValues(Ops, 2);
5759 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5761 The rounding mode is in bits 11:10 of FPSR, and has the following
5768 FLT_ROUNDS, on the other hand, expects the following:
5775 To perform the conversion, we do:
5776 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5779 MachineFunction &MF = DAG.getMachineFunction();
5780 const TargetMachine &TM = MF.getTarget();
5781 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5782 unsigned StackAlignment = TFI.getStackAlignment();
5783 MVT VT = Op.getValueType();
5785 // Save FP Control Word to stack slot
5786 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5787 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5789 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5790 DAG.getEntryNode(), StackSlot);
5792 // Load FP Control Word from stack slot
5793 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5795 // Transform as necessary
5797 DAG.getNode(ISD::SRL, MVT::i16,
5798 DAG.getNode(ISD::AND, MVT::i16,
5799 CWD, DAG.getConstant(0x800, MVT::i16)),
5800 DAG.getConstant(11, MVT::i8));
5802 DAG.getNode(ISD::SRL, MVT::i16,
5803 DAG.getNode(ISD::AND, MVT::i16,
5804 CWD, DAG.getConstant(0x400, MVT::i16)),
5805 DAG.getConstant(9, MVT::i8));
5808 DAG.getNode(ISD::AND, MVT::i16,
5809 DAG.getNode(ISD::ADD, MVT::i16,
5810 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5811 DAG.getConstant(1, MVT::i16)),
5812 DAG.getConstant(3, MVT::i16));
5815 return DAG.getNode((VT.getSizeInBits() < 16 ?
5816 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5819 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5820 MVT VT = Op.getValueType();
5822 unsigned NumBits = VT.getSizeInBits();
5824 Op = Op.getOperand(0);
5825 if (VT == MVT::i8) {
5826 // Zero extend to i32 since there is not an i8 bsr.
5828 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5831 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5832 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5833 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5835 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5836 SmallVector<SDValue, 4> Ops;
5838 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5839 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5840 Ops.push_back(Op.getValue(1));
5841 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5843 // Finally xor with NumBits-1.
5844 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5847 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5851 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5852 MVT VT = Op.getValueType();
5854 unsigned NumBits = VT.getSizeInBits();
5856 Op = Op.getOperand(0);
5857 if (VT == MVT::i8) {
5859 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5862 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5863 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5864 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5866 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5867 SmallVector<SDValue, 4> Ops;
5869 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5870 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5871 Ops.push_back(Op.getValue(1));
5872 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5875 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5879 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5880 MVT T = Op.getValueType();
5883 switch(T.getSimpleVT()) {
5885 assert(false && "Invalid value type!");
5886 case MVT::i8: Reg = X86::AL; size = 1; break;
5887 case MVT::i16: Reg = X86::AX; size = 2; break;
5888 case MVT::i32: Reg = X86::EAX; size = 4; break;
5890 if (Subtarget->is64Bit()) {
5891 Reg = X86::RAX; size = 8;
5892 } else //Should go away when LowerType stuff lands
5893 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5896 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5897 Op.getOperand(3), SDValue());
5898 SDValue Ops[] = { cpIn.getValue(0),
5901 DAG.getTargetConstant(size, MVT::i8),
5903 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5904 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5906 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5910 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5911 SelectionDAG &DAG) {
5912 MVT T = Op->getValueType(0);
5913 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5914 SDValue cpInL, cpInH;
5915 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5916 DAG.getConstant(0, MVT::i32));
5917 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5918 DAG.getConstant(1, MVT::i32));
5919 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5921 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5922 cpInH, cpInL.getValue(1));
5923 SDValue swapInL, swapInH;
5924 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5925 DAG.getConstant(0, MVT::i32));
5926 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5927 DAG.getConstant(1, MVT::i32));
5928 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5929 swapInL, cpInH.getValue(1));
5930 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5931 swapInH, swapInL.getValue(1));
5932 SDValue Ops[] = { swapInH.getValue(0),
5934 swapInH.getValue(1)};
5935 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5936 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5937 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5938 Result.getValue(1));
5939 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5940 cpOutL.getValue(2));
5941 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5942 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5943 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
5944 return DAG.getMergeValues(Vals, 2).getNode();
5947 SDNode* X86TargetLowering::ExpandATOMIC_LOAD_SUB(SDNode* Op,
5948 SelectionDAG &DAG) {
5949 MVT T = Op->getValueType(0);
5950 SDValue negOp = DAG.getNode(ISD::SUB, T,
5951 DAG.getConstant(0, T), Op->getOperand(2));
5952 return DAG.getAtomic((T==MVT::i8 ? ISD::ATOMIC_LOAD_ADD_8:
5953 T==MVT::i16 ? ISD::ATOMIC_LOAD_ADD_16:
5954 T==MVT::i32 ? ISD::ATOMIC_LOAD_ADD_32:
5955 T==MVT::i64 ? ISD::ATOMIC_LOAD_ADD_64: 0),
5956 Op->getOperand(0), Op->getOperand(1), negOp,
5957 cast<AtomicSDNode>(Op)->getSrcValue(),
5958 cast<AtomicSDNode>(Op)->getAlignment()).getNode();
5961 /// LowerOperation - Provide custom lowering hooks for some operations.
5963 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5964 switch (Op.getOpcode()) {
5965 default: assert(0 && "Should not custom lower this!");
5966 case ISD::ATOMIC_CMP_SWAP_8: return LowerCMP_SWAP(Op,DAG);
5967 case ISD::ATOMIC_CMP_SWAP_16: return LowerCMP_SWAP(Op,DAG);
5968 case ISD::ATOMIC_CMP_SWAP_32: return LowerCMP_SWAP(Op,DAG);
5969 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
5970 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5971 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5972 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5973 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5974 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5975 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5976 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5977 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5978 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5979 case ISD::SHL_PARTS:
5980 case ISD::SRA_PARTS:
5981 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5982 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5983 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5984 case ISD::FABS: return LowerFABS(Op, DAG);
5985 case ISD::FNEG: return LowerFNEG(Op, DAG);
5986 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5987 case ISD::SETCC: return LowerSETCC(Op, DAG);
5988 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
5989 case ISD::SELECT: return LowerSELECT(Op, DAG);
5990 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5991 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5992 case ISD::CALL: return LowerCALL(Op, DAG);
5993 case ISD::RET: return LowerRET(Op, DAG);
5994 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5995 case ISD::VASTART: return LowerVASTART(Op, DAG);
5996 case ISD::VAARG: return LowerVAARG(Op, DAG);
5997 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5998 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5999 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6000 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6001 case ISD::FRAME_TO_ARGS_OFFSET:
6002 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6003 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6004 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6005 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6006 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6007 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6008 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6010 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
6011 case ISD::READCYCLECOUNTER:
6012 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
6016 /// ReplaceNodeResults - Replace a node with an illegal result type
6017 /// with a new node built out of custom code.
6018 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
6019 switch (N->getOpcode()) {
6020 default: assert(0 && "Should not custom lower this!");
6021 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
6022 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
6023 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
6024 case ISD::ATOMIC_LOAD_SUB_8: return ExpandATOMIC_LOAD_SUB(N,DAG);
6025 case ISD::ATOMIC_LOAD_SUB_16: return ExpandATOMIC_LOAD_SUB(N,DAG);
6026 case ISD::ATOMIC_LOAD_SUB_32: return ExpandATOMIC_LOAD_SUB(N,DAG);
6027 case ISD::ATOMIC_LOAD_SUB_64: return ExpandATOMIC_LOAD_SUB(N,DAG);
6031 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6033 default: return NULL;
6034 case X86ISD::BSF: return "X86ISD::BSF";
6035 case X86ISD::BSR: return "X86ISD::BSR";
6036 case X86ISD::SHLD: return "X86ISD::SHLD";
6037 case X86ISD::SHRD: return "X86ISD::SHRD";
6038 case X86ISD::FAND: return "X86ISD::FAND";
6039 case X86ISD::FOR: return "X86ISD::FOR";
6040 case X86ISD::FXOR: return "X86ISD::FXOR";
6041 case X86ISD::FSRL: return "X86ISD::FSRL";
6042 case X86ISD::FILD: return "X86ISD::FILD";
6043 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6044 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6045 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6046 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6047 case X86ISD::FLD: return "X86ISD::FLD";
6048 case X86ISD::FST: return "X86ISD::FST";
6049 case X86ISD::CALL: return "X86ISD::CALL";
6050 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6051 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6052 case X86ISD::CMP: return "X86ISD::CMP";
6053 case X86ISD::COMI: return "X86ISD::COMI";
6054 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6055 case X86ISD::SETCC: return "X86ISD::SETCC";
6056 case X86ISD::CMOV: return "X86ISD::CMOV";
6057 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6058 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6059 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6060 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6061 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6062 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6063 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6064 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6065 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6066 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6067 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6068 case X86ISD::FMAX: return "X86ISD::FMAX";
6069 case X86ISD::FMIN: return "X86ISD::FMIN";
6070 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6071 case X86ISD::FRCP: return "X86ISD::FRCP";
6072 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6073 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6074 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6075 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6076 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6077 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6078 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6079 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6080 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6081 case X86ISD::VSHL: return "X86ISD::VSHL";
6082 case X86ISD::VSRL: return "X86ISD::VSRL";
6083 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6084 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6085 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6086 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6087 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6088 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6089 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6090 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6091 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6092 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6096 // isLegalAddressingMode - Return true if the addressing mode represented
6097 // by AM is legal for this target, for a load/store of the specified type.
6098 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6099 const Type *Ty) const {
6100 // X86 supports extremely general addressing modes.
6102 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6103 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6107 // We can only fold this if we don't need an extra load.
6108 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6111 // X86-64 only supports addr of globals in small code model.
6112 if (Subtarget->is64Bit()) {
6113 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6115 // If lower 4G is not available, then we must use rip-relative addressing.
6116 if (AM.BaseOffs || AM.Scale > 1)
6127 // These scales always work.
6132 // These scales are formed with basereg+scalereg. Only accept if there is
6137 default: // Other stuff never works.
6145 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6146 if (!Ty1->isInteger() || !Ty2->isInteger())
6148 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6149 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6150 if (NumBits1 <= NumBits2)
6152 return Subtarget->is64Bit() || NumBits1 < 64;
6155 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6156 if (!VT1.isInteger() || !VT2.isInteger())
6158 unsigned NumBits1 = VT1.getSizeInBits();
6159 unsigned NumBits2 = VT2.getSizeInBits();
6160 if (NumBits1 <= NumBits2)
6162 return Subtarget->is64Bit() || NumBits1 < 64;
6165 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6166 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6167 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6168 /// are assumed to be legal.
6170 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6171 // Only do shuffles on 128-bit vector types for now.
6172 if (VT.getSizeInBits() == 64) return false;
6173 return (Mask.getNode()->getNumOperands() <= 4 ||
6174 isIdentityMask(Mask.getNode()) ||
6175 isIdentityMask(Mask.getNode(), true) ||
6176 isSplatMask(Mask.getNode()) ||
6177 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6178 X86::isUNPCKLMask(Mask.getNode()) ||
6179 X86::isUNPCKHMask(Mask.getNode()) ||
6180 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6181 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6185 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6186 MVT EVT, SelectionDAG &DAG) const {
6187 unsigned NumElts = BVOps.size();
6188 // Only do shuffles on 128-bit vector types for now.
6189 if (EVT.getSizeInBits() * NumElts == 64) return false;
6190 if (NumElts == 2) return true;
6192 return (isMOVLMask(&BVOps[0], 4) ||
6193 isCommutedMOVL(&BVOps[0], 4, true) ||
6194 isSHUFPMask(&BVOps[0], 4) ||
6195 isCommutedSHUFP(&BVOps[0], 4));
6200 //===----------------------------------------------------------------------===//
6201 // X86 Scheduler Hooks
6202 //===----------------------------------------------------------------------===//
6204 // private utility function
6206 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6207 MachineBasicBlock *MBB,
6215 TargetRegisterClass *RC,
6217 // For the atomic bitwise operator, we generate
6220 // ld t1 = [bitinstr.addr]
6221 // op t2 = t1, [bitinstr.val]
6223 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6225 // fallthrough -->nextMBB
6226 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6227 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6228 MachineFunction::iterator MBBIter = MBB;
6231 /// First build the CFG
6232 MachineFunction *F = MBB->getParent();
6233 MachineBasicBlock *thisMBB = MBB;
6234 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6235 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6236 F->insert(MBBIter, newMBB);
6237 F->insert(MBBIter, nextMBB);
6239 // Move all successors to thisMBB to nextMBB
6240 nextMBB->transferSuccessors(thisMBB);
6242 // Update thisMBB to fall through to newMBB
6243 thisMBB->addSuccessor(newMBB);
6245 // newMBB jumps to itself and fall through to nextMBB
6246 newMBB->addSuccessor(nextMBB);
6247 newMBB->addSuccessor(newMBB);
6249 // Insert instructions into newMBB based on incoming instruction
6250 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6251 MachineOperand& destOper = bInstr->getOperand(0);
6252 MachineOperand* argOpers[6];
6253 int numArgs = bInstr->getNumOperands() - 1;
6254 for (int i=0; i < numArgs; ++i)
6255 argOpers[i] = &bInstr->getOperand(i+1);
6257 // x86 address has 4 operands: base, index, scale, and displacement
6258 int lastAddrIndx = 3; // [0,3]
6261 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6262 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6263 for (int i=0; i <= lastAddrIndx; ++i)
6264 (*MIB).addOperand(*argOpers[i]);
6266 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6268 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6273 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6274 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6275 && "invalid operand");
6276 if (argOpers[valArgIndx]->isReg())
6277 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6279 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6281 (*MIB).addOperand(*argOpers[valArgIndx]);
6283 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6286 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6287 for (int i=0; i <= lastAddrIndx; ++i)
6288 (*MIB).addOperand(*argOpers[i]);
6290 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6291 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6293 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6297 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6299 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6303 // private utility function
6305 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6306 MachineBasicBlock *MBB,
6308 // For the atomic min/max operator, we generate
6311 // ld t1 = [min/max.addr]
6312 // mov t2 = [min/max.val]
6314 // cmov[cond] t2 = t1
6316 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6318 // fallthrough -->nextMBB
6320 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6321 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6322 MachineFunction::iterator MBBIter = MBB;
6325 /// First build the CFG
6326 MachineFunction *F = MBB->getParent();
6327 MachineBasicBlock *thisMBB = MBB;
6328 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6329 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6330 F->insert(MBBIter, newMBB);
6331 F->insert(MBBIter, nextMBB);
6333 // Move all successors to thisMBB to nextMBB
6334 nextMBB->transferSuccessors(thisMBB);
6336 // Update thisMBB to fall through to newMBB
6337 thisMBB->addSuccessor(newMBB);
6339 // newMBB jumps to newMBB and fall through to nextMBB
6340 newMBB->addSuccessor(nextMBB);
6341 newMBB->addSuccessor(newMBB);
6343 // Insert instructions into newMBB based on incoming instruction
6344 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6345 MachineOperand& destOper = mInstr->getOperand(0);
6346 MachineOperand* argOpers[6];
6347 int numArgs = mInstr->getNumOperands() - 1;
6348 for (int i=0; i < numArgs; ++i)
6349 argOpers[i] = &mInstr->getOperand(i+1);
6351 // x86 address has 4 operands: base, index, scale, and displacement
6352 int lastAddrIndx = 3; // [0,3]
6355 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6356 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6357 for (int i=0; i <= lastAddrIndx; ++i)
6358 (*MIB).addOperand(*argOpers[i]);
6360 // We only support register and immediate values
6361 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6362 && "invalid operand");
6364 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6365 if (argOpers[valArgIndx]->isReg())
6366 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6368 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6369 (*MIB).addOperand(*argOpers[valArgIndx]);
6371 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6374 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6379 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6380 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6384 // Cmp and exchange if none has modified the memory location
6385 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6386 for (int i=0; i <= lastAddrIndx; ++i)
6387 (*MIB).addOperand(*argOpers[i]);
6389 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6390 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6392 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6393 MIB.addReg(X86::EAX);
6396 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6398 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6404 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6405 MachineBasicBlock *BB) {
6406 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6407 switch (MI->getOpcode()) {
6408 default: assert(false && "Unexpected instr type to insert");
6409 case X86::CMOV_FR32:
6410 case X86::CMOV_FR64:
6411 case X86::CMOV_V4F32:
6412 case X86::CMOV_V2F64:
6413 case X86::CMOV_V2I64: {
6414 // To "insert" a SELECT_CC instruction, we actually have to insert the
6415 // diamond control-flow pattern. The incoming instruction knows the
6416 // destination vreg to set, the condition code register to branch on, the
6417 // true/false values to select between, and a branch opcode to use.
6418 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6419 MachineFunction::iterator It = BB;
6425 // cmpTY ccX, r1, r2
6427 // fallthrough --> copy0MBB
6428 MachineBasicBlock *thisMBB = BB;
6429 MachineFunction *F = BB->getParent();
6430 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6431 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6433 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6434 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6435 F->insert(It, copy0MBB);
6436 F->insert(It, sinkMBB);
6437 // Update machine-CFG edges by transferring all successors of the current
6438 // block to the new block which will contain the Phi node for the select.
6439 sinkMBB->transferSuccessors(BB);
6441 // Add the true and fallthrough blocks as its successors.
6442 BB->addSuccessor(copy0MBB);
6443 BB->addSuccessor(sinkMBB);
6446 // %FalseValue = ...
6447 // # fallthrough to sinkMBB
6450 // Update machine-CFG edges
6451 BB->addSuccessor(sinkMBB);
6454 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6457 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6458 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6459 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6461 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6465 case X86::FP32_TO_INT16_IN_MEM:
6466 case X86::FP32_TO_INT32_IN_MEM:
6467 case X86::FP32_TO_INT64_IN_MEM:
6468 case X86::FP64_TO_INT16_IN_MEM:
6469 case X86::FP64_TO_INT32_IN_MEM:
6470 case X86::FP64_TO_INT64_IN_MEM:
6471 case X86::FP80_TO_INT16_IN_MEM:
6472 case X86::FP80_TO_INT32_IN_MEM:
6473 case X86::FP80_TO_INT64_IN_MEM: {
6474 // Change the floating point control register to use "round towards zero"
6475 // mode when truncating to an integer value.
6476 MachineFunction *F = BB->getParent();
6477 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6478 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6480 // Load the old value of the high byte of the control word...
6482 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6483 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6485 // Set the high part to be round to zero...
6486 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6489 // Reload the modified control word now...
6490 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6492 // Restore the memory image of control word to original value
6493 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6496 // Get the X86 opcode to use.
6498 switch (MI->getOpcode()) {
6499 default: assert(0 && "illegal opcode!");
6500 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6501 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6502 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6503 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6504 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6505 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6506 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6507 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6508 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6512 MachineOperand &Op = MI->getOperand(0);
6513 if (Op.isRegister()) {
6514 AM.BaseType = X86AddressMode::RegBase;
6515 AM.Base.Reg = Op.getReg();
6517 AM.BaseType = X86AddressMode::FrameIndexBase;
6518 AM.Base.FrameIndex = Op.getIndex();
6520 Op = MI->getOperand(1);
6521 if (Op.isImmediate())
6522 AM.Scale = Op.getImm();
6523 Op = MI->getOperand(2);
6524 if (Op.isImmediate())
6525 AM.IndexReg = Op.getImm();
6526 Op = MI->getOperand(3);
6527 if (Op.isGlobalAddress()) {
6528 AM.GV = Op.getGlobal();
6530 AM.Disp = Op.getImm();
6532 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6533 .addReg(MI->getOperand(4).getReg());
6535 // Reload the original control word now.
6536 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6538 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6541 case X86::ATOMAND32:
6542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6543 X86::AND32ri, X86::MOV32rm,
6544 X86::LCMPXCHG32, X86::MOV32rr,
6545 X86::NOT32r, X86::EAX,
6546 X86::GR32RegisterClass);
6548 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6549 X86::OR32ri, X86::MOV32rm,
6550 X86::LCMPXCHG32, X86::MOV32rr,
6551 X86::NOT32r, X86::EAX,
6552 X86::GR32RegisterClass);
6553 case X86::ATOMXOR32:
6554 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6555 X86::XOR32ri, X86::MOV32rm,
6556 X86::LCMPXCHG32, X86::MOV32rr,
6557 X86::NOT32r, X86::EAX,
6558 X86::GR32RegisterClass);
6559 case X86::ATOMNAND32:
6560 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6561 X86::AND32ri, X86::MOV32rm,
6562 X86::LCMPXCHG32, X86::MOV32rr,
6563 X86::NOT32r, X86::EAX,
6564 X86::GR32RegisterClass, true);
6565 case X86::ATOMMIN32:
6566 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6567 case X86::ATOMMAX32:
6568 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6569 case X86::ATOMUMIN32:
6570 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6571 case X86::ATOMUMAX32:
6572 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6574 case X86::ATOMAND16:
6575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6576 X86::AND16ri, X86::MOV16rm,
6577 X86::LCMPXCHG16, X86::MOV16rr,
6578 X86::NOT16r, X86::AX,
6579 X86::GR16RegisterClass);
6581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6582 X86::OR16ri, X86::MOV16rm,
6583 X86::LCMPXCHG16, X86::MOV16rr,
6584 X86::NOT16r, X86::AX,
6585 X86::GR16RegisterClass);
6586 case X86::ATOMXOR16:
6587 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6588 X86::XOR16ri, X86::MOV16rm,
6589 X86::LCMPXCHG16, X86::MOV16rr,
6590 X86::NOT16r, X86::AX,
6591 X86::GR16RegisterClass);
6592 case X86::ATOMNAND16:
6593 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6594 X86::AND16ri, X86::MOV16rm,
6595 X86::LCMPXCHG16, X86::MOV16rr,
6596 X86::NOT16r, X86::AX,
6597 X86::GR16RegisterClass, true);
6598 case X86::ATOMMIN16:
6599 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6600 case X86::ATOMMAX16:
6601 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6602 case X86::ATOMUMIN16:
6603 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6604 case X86::ATOMUMAX16:
6605 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6608 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6609 X86::AND8ri, X86::MOV8rm,
6610 X86::LCMPXCHG8, X86::MOV8rr,
6611 X86::NOT8r, X86::AL,
6612 X86::GR8RegisterClass);
6614 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6615 X86::OR8ri, X86::MOV8rm,
6616 X86::LCMPXCHG8, X86::MOV8rr,
6617 X86::NOT8r, X86::AL,
6618 X86::GR8RegisterClass);
6620 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6621 X86::XOR8ri, X86::MOV8rm,
6622 X86::LCMPXCHG8, X86::MOV8rr,
6623 X86::NOT8r, X86::AL,
6624 X86::GR8RegisterClass);
6625 case X86::ATOMNAND8:
6626 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6627 X86::AND8ri, X86::MOV8rm,
6628 X86::LCMPXCHG8, X86::MOV8rr,
6629 X86::NOT8r, X86::AL,
6630 X86::GR8RegisterClass, true);
6631 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6632 case X86::ATOMAND64:
6633 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6634 X86::AND64ri32, X86::MOV64rm,
6635 X86::LCMPXCHG64, X86::MOV64rr,
6636 X86::NOT64r, X86::RAX,
6637 X86::GR64RegisterClass);
6639 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6640 X86::OR64ri32, X86::MOV64rm,
6641 X86::LCMPXCHG64, X86::MOV64rr,
6642 X86::NOT64r, X86::RAX,
6643 X86::GR64RegisterClass);
6644 case X86::ATOMXOR64:
6645 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6646 X86::XOR64ri32, X86::MOV64rm,
6647 X86::LCMPXCHG64, X86::MOV64rr,
6648 X86::NOT64r, X86::RAX,
6649 X86::GR64RegisterClass);
6650 case X86::ATOMNAND64:
6651 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6652 X86::AND64ri32, X86::MOV64rm,
6653 X86::LCMPXCHG64, X86::MOV64rr,
6654 X86::NOT64r, X86::RAX,
6655 X86::GR64RegisterClass, true);
6656 case X86::ATOMMIN64:
6657 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6658 case X86::ATOMMAX64:
6659 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6660 case X86::ATOMUMIN64:
6661 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6662 case X86::ATOMUMAX64:
6663 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6667 //===----------------------------------------------------------------------===//
6668 // X86 Optimization Hooks
6669 //===----------------------------------------------------------------------===//
6671 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6675 const SelectionDAG &DAG,
6676 unsigned Depth) const {
6677 unsigned Opc = Op.getOpcode();
6678 assert((Opc >= ISD::BUILTIN_OP_END ||
6679 Opc == ISD::INTRINSIC_WO_CHAIN ||
6680 Opc == ISD::INTRINSIC_W_CHAIN ||
6681 Opc == ISD::INTRINSIC_VOID) &&
6682 "Should use MaskedValueIsZero if you don't know whether Op"
6683 " is a target node!");
6685 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
6689 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
6690 Mask.getBitWidth() - 1);
6695 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
6696 /// node is a GlobalAddress + offset.
6697 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
6698 GlobalValue* &GA, int64_t &Offset) const{
6699 if (N->getOpcode() == X86ISD::Wrapper) {
6700 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
6701 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
6705 return TargetLowering::isGAPlusOffset(N, GA, Offset);
6708 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
6709 const TargetLowering &TLI) {
6712 if (TLI.isGAPlusOffset(Base, GV, Offset))
6713 return (GV->getAlignment() >= N && (Offset % N) == 0);
6714 // DAG combine handles the stack object case.
6718 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
6719 unsigned NumElems, MVT EVT,
6721 SelectionDAG &DAG, MachineFrameInfo *MFI,
6722 const TargetLowering &TLI) {
6724 for (unsigned i = 0; i < NumElems; ++i) {
6725 SDValue Idx = PermMask.getOperand(i);
6726 if (Idx.getOpcode() == ISD::UNDEF) {
6732 SDValue Elt = DAG.getShuffleScalarElt(N, i);
6733 if (!Elt.getNode() ||
6734 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6737 Base = Elt.getNode();
6738 if (Base->getOpcode() == ISD::UNDEF)
6742 if (Elt.getOpcode() == ISD::UNDEF)
6745 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
6746 EVT.getSizeInBits()/8, i, MFI))
6752 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6753 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6754 /// if the load addresses are consecutive, non-overlapping, and in the right
6756 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6757 const TargetLowering &TLI) {
6758 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6759 MVT VT = N->getValueType(0);
6760 MVT EVT = VT.getVectorElementType();
6761 SDValue PermMask = N->getOperand(2);
6762 unsigned NumElems = PermMask.getNumOperands();
6763 SDNode *Base = NULL;
6764 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
6768 LoadSDNode *LD = cast<LoadSDNode>(Base);
6769 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
6770 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6771 LD->getSrcValueOffset(), LD->isVolatile());
6772 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6773 LD->getSrcValueOffset(), LD->isVolatile(),
6774 LD->getAlignment());
6777 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
6778 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
6779 const X86Subtarget *Subtarget,
6780 const TargetLowering &TLI) {
6781 unsigned NumOps = N->getNumOperands();
6783 // Ignore single operand BUILD_VECTOR.
6787 MVT VT = N->getValueType(0);
6788 MVT EVT = VT.getVectorElementType();
6789 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
6790 // We are looking for load i64 and zero extend. We want to transform
6791 // it before legalizer has a chance to expand it. Also look for i64
6792 // BUILD_PAIR bit casted to f64.
6794 // This must be an insertion into a zero vector.
6795 SDValue HighElt = N->getOperand(1);
6796 if (!isZeroNode(HighElt))
6799 // Value must be a load.
6800 SDNode *Base = N->getOperand(0).getNode();
6801 if (!isa<LoadSDNode>(Base)) {
6802 if (Base->getOpcode() != ISD::BIT_CONVERT)
6804 Base = Base->getOperand(0).getNode();
6805 if (!isa<LoadSDNode>(Base))
6809 // Transform it into VZEXT_LOAD addr.
6810 LoadSDNode *LD = cast<LoadSDNode>(Base);
6812 // Load must not be an extload.
6813 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
6816 return DAG.getNode(X86ISD::VZEXT_LOAD, VT, LD->getChain(), LD->getBasePtr());
6819 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6820 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6821 const X86Subtarget *Subtarget) {
6822 SDValue Cond = N->getOperand(0);
6824 // If we have SSE[12] support, try to form min/max nodes.
6825 if (Subtarget->hasSSE2() &&
6826 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6827 if (Cond.getOpcode() == ISD::SETCC) {
6828 // Get the LHS/RHS of the select.
6829 SDValue LHS = N->getOperand(1);
6830 SDValue RHS = N->getOperand(2);
6831 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6833 unsigned Opcode = 0;
6834 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6837 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6840 if (!UnsafeFPMath) break;
6842 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6844 Opcode = X86ISD::FMIN;
6847 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6850 if (!UnsafeFPMath) break;
6852 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6854 Opcode = X86ISD::FMAX;
6857 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6860 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6863 if (!UnsafeFPMath) break;
6865 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6867 Opcode = X86ISD::FMIN;
6870 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6873 if (!UnsafeFPMath) break;
6875 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6877 Opcode = X86ISD::FMAX;
6883 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6891 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6892 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
6893 const X86Subtarget *Subtarget) {
6894 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6895 // the FP state in cases where an emms may be missing.
6896 // A preferable solution to the general problem is to figure out the right
6897 // places to insert EMMS. This qualifies as a quick hack.
6898 StoreSDNode *St = cast<StoreSDNode>(N);
6899 if (St->getValue().getValueType().isVector() &&
6900 St->getValue().getValueType().getSizeInBits() == 64 &&
6901 isa<LoadSDNode>(St->getValue()) &&
6902 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6903 St->getChain().hasOneUse() && !St->isVolatile()) {
6904 SDNode* LdVal = St->getValue().getNode();
6906 int TokenFactorIndex = -1;
6907 SmallVector<SDValue, 8> Ops;
6908 SDNode* ChainVal = St->getChain().getNode();
6909 // Must be a store of a load. We currently handle two cases: the load
6910 // is a direct child, and it's under an intervening TokenFactor. It is
6911 // possible to dig deeper under nested TokenFactors.
6912 if (ChainVal == LdVal)
6913 Ld = cast<LoadSDNode>(St->getChain());
6914 else if (St->getValue().hasOneUse() &&
6915 ChainVal->getOpcode() == ISD::TokenFactor) {
6916 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6917 if (ChainVal->getOperand(i).getNode() == LdVal) {
6918 TokenFactorIndex = i;
6919 Ld = cast<LoadSDNode>(St->getValue());
6921 Ops.push_back(ChainVal->getOperand(i));
6925 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6926 if (Subtarget->is64Bit()) {
6927 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6928 Ld->getBasePtr(), Ld->getSrcValue(),
6929 Ld->getSrcValueOffset(), Ld->isVolatile(),
6930 Ld->getAlignment());
6931 SDValue NewChain = NewLd.getValue(1);
6932 if (TokenFactorIndex != -1) {
6933 Ops.push_back(NewChain);
6934 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6937 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6938 St->getSrcValue(), St->getSrcValueOffset(),
6939 St->isVolatile(), St->getAlignment());
6942 // Otherwise, lower to two 32-bit copies.
6943 SDValue LoAddr = Ld->getBasePtr();
6944 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6945 DAG.getConstant(4, MVT::i32));
6947 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6948 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6949 Ld->isVolatile(), Ld->getAlignment());
6950 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6951 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6953 MinAlign(Ld->getAlignment(), 4));
6955 SDValue NewChain = LoLd.getValue(1);
6956 if (TokenFactorIndex != -1) {
6957 Ops.push_back(LoLd);
6958 Ops.push_back(HiLd);
6959 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6963 LoAddr = St->getBasePtr();
6964 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6965 DAG.getConstant(4, MVT::i32));
6967 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6968 St->getSrcValue(), St->getSrcValueOffset(),
6969 St->isVolatile(), St->getAlignment());
6970 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6972 St->getSrcValueOffset() + 4,
6974 MinAlign(St->getAlignment(), 4));
6975 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6981 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6982 /// X86ISD::FXOR nodes.
6983 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6984 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6985 // F[X]OR(0.0, x) -> x
6986 // F[X]OR(x, 0.0) -> x
6987 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6988 if (C->getValueAPF().isPosZero())
6989 return N->getOperand(1);
6990 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6991 if (C->getValueAPF().isPosZero())
6992 return N->getOperand(0);
6996 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
6997 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
6998 // FAND(0.0, x) -> 0.0
6999 // FAND(x, 0.0) -> 0.0
7000 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7001 if (C->getValueAPF().isPosZero())
7002 return N->getOperand(0);
7003 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7004 if (C->getValueAPF().isPosZero())
7005 return N->getOperand(1);
7010 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7011 DAGCombinerInfo &DCI) const {
7012 SelectionDAG &DAG = DCI.DAG;
7013 switch (N->getOpcode()) {
7015 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7016 case ISD::BUILD_VECTOR:
7017 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7018 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7019 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7021 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7022 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7028 //===----------------------------------------------------------------------===//
7029 // X86 Inline Assembly Support
7030 //===----------------------------------------------------------------------===//
7032 /// getConstraintType - Given a constraint letter, return the type of
7033 /// constraint it is for this target.
7034 X86TargetLowering::ConstraintType
7035 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7036 if (Constraint.size() == 1) {
7037 switch (Constraint[0]) {
7048 return C_RegisterClass;
7053 return TargetLowering::getConstraintType(Constraint);
7056 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7057 /// with another that has more specific requirements based on the type of the
7058 /// corresponding operand.
7059 const char *X86TargetLowering::
7060 LowerXConstraint(MVT ConstraintVT) const {
7061 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7062 // 'f' like normal targets.
7063 if (ConstraintVT.isFloatingPoint()) {
7064 if (Subtarget->hasSSE2())
7066 if (Subtarget->hasSSE1())
7070 return TargetLowering::LowerXConstraint(ConstraintVT);
7073 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7074 /// vector. If it is invalid, don't add anything to Ops.
7075 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7077 std::vector<SDValue>&Ops,
7078 SelectionDAG &DAG) const {
7079 SDValue Result(0, 0);
7081 switch (Constraint) {
7084 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7085 if (C->getValue() <= 31) {
7086 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7092 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7093 if (C->getValue() <= 255) {
7094 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7100 // Literal immediates are always ok.
7101 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7102 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
7106 // If we are in non-pic codegen mode, we allow the address of a global (with
7107 // an optional displacement) to be used with 'i'.
7108 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7111 // Match either (GA) or (GA+C)
7113 Offset = GA->getOffset();
7114 } else if (Op.getOpcode() == ISD::ADD) {
7115 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7116 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7118 Offset = GA->getOffset()+C->getValue();
7120 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7121 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7123 Offset = GA->getOffset()+C->getValue();
7130 // If addressing this global requires a load (e.g. in PIC mode), we can't
7132 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
7136 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7142 // Otherwise, not valid for this mode.
7147 if (Result.getNode()) {
7148 Ops.push_back(Result);
7151 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7154 std::vector<unsigned> X86TargetLowering::
7155 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7157 if (Constraint.size() == 1) {
7158 // FIXME: not handling fp-stack yet!
7159 switch (Constraint[0]) { // GCC X86 Constraint Letters
7160 default: break; // Unknown constraint letter
7161 case 'A': // EAX/EDX
7162 if (VT == MVT::i32 || VT == MVT::i64)
7163 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7165 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7168 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7169 else if (VT == MVT::i16)
7170 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7171 else if (VT == MVT::i8)
7172 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7173 else if (VT == MVT::i64)
7174 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7179 return std::vector<unsigned>();
7182 std::pair<unsigned, const TargetRegisterClass*>
7183 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7185 // First, see if this is a constraint that directly corresponds to an LLVM
7187 if (Constraint.size() == 1) {
7188 // GCC Constraint Letters
7189 switch (Constraint[0]) {
7191 case 'r': // GENERAL_REGS
7192 case 'R': // LEGACY_REGS
7193 case 'l': // INDEX_REGS
7194 if (VT == MVT::i64 && Subtarget->is64Bit())
7195 return std::make_pair(0U, X86::GR64RegisterClass);
7197 return std::make_pair(0U, X86::GR32RegisterClass);
7198 else if (VT == MVT::i16)
7199 return std::make_pair(0U, X86::GR16RegisterClass);
7200 else if (VT == MVT::i8)
7201 return std::make_pair(0U, X86::GR8RegisterClass);
7203 case 'f': // FP Stack registers.
7204 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7205 // value to the correct fpstack register class.
7206 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7207 return std::make_pair(0U, X86::RFP32RegisterClass);
7208 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7209 return std::make_pair(0U, X86::RFP64RegisterClass);
7210 return std::make_pair(0U, X86::RFP80RegisterClass);
7211 case 'y': // MMX_REGS if MMX allowed.
7212 if (!Subtarget->hasMMX()) break;
7213 return std::make_pair(0U, X86::VR64RegisterClass);
7215 case 'Y': // SSE_REGS if SSE2 allowed
7216 if (!Subtarget->hasSSE2()) break;
7218 case 'x': // SSE_REGS if SSE1 allowed
7219 if (!Subtarget->hasSSE1()) break;
7221 switch (VT.getSimpleVT()) {
7223 // Scalar SSE types.
7226 return std::make_pair(0U, X86::FR32RegisterClass);
7229 return std::make_pair(0U, X86::FR64RegisterClass);
7237 return std::make_pair(0U, X86::VR128RegisterClass);
7243 // Use the default implementation in TargetLowering to convert the register
7244 // constraint into a member of a register class.
7245 std::pair<unsigned, const TargetRegisterClass*> Res;
7246 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7248 // Not found as a standard register?
7249 if (Res.second == 0) {
7250 // GCC calls "st(0)" just plain "st".
7251 if (StringsEqualNoCase("{st}", Constraint)) {
7252 Res.first = X86::ST0;
7253 Res.second = X86::RFP80RegisterClass;
7259 // Otherwise, check to see if this is a register class of the wrong value
7260 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7261 // turn into {ax},{dx}.
7262 if (Res.second->hasType(VT))
7263 return Res; // Correct type already, nothing to do.
7265 // All of the single-register GCC register classes map their values onto
7266 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7267 // really want an 8-bit or 32-bit register, map to the appropriate register
7268 // class and return the appropriate register.
7269 if (Res.second == X86::GR16RegisterClass) {
7270 if (VT == MVT::i8) {
7271 unsigned DestReg = 0;
7272 switch (Res.first) {
7274 case X86::AX: DestReg = X86::AL; break;
7275 case X86::DX: DestReg = X86::DL; break;
7276 case X86::CX: DestReg = X86::CL; break;
7277 case X86::BX: DestReg = X86::BL; break;
7280 Res.first = DestReg;
7281 Res.second = Res.second = X86::GR8RegisterClass;
7283 } else if (VT == MVT::i32) {
7284 unsigned DestReg = 0;
7285 switch (Res.first) {
7287 case X86::AX: DestReg = X86::EAX; break;
7288 case X86::DX: DestReg = X86::EDX; break;
7289 case X86::CX: DestReg = X86::ECX; break;
7290 case X86::BX: DestReg = X86::EBX; break;
7291 case X86::SI: DestReg = X86::ESI; break;
7292 case X86::DI: DestReg = X86::EDI; break;
7293 case X86::BP: DestReg = X86::EBP; break;
7294 case X86::SP: DestReg = X86::ESP; break;
7297 Res.first = DestReg;
7298 Res.second = Res.second = X86::GR32RegisterClass;
7300 } else if (VT == MVT::i64) {
7301 unsigned DestReg = 0;
7302 switch (Res.first) {
7304 case X86::AX: DestReg = X86::RAX; break;
7305 case X86::DX: DestReg = X86::RDX; break;
7306 case X86::CX: DestReg = X86::RCX; break;
7307 case X86::BX: DestReg = X86::RBX; break;
7308 case X86::SI: DestReg = X86::RSI; break;
7309 case X86::DI: DestReg = X86::RDI; break;
7310 case X86::BP: DestReg = X86::RBP; break;
7311 case X86::SP: DestReg = X86::RSP; break;
7314 Res.first = DestReg;
7315 Res.second = Res.second = X86::GR64RegisterClass;
7318 } else if (Res.second == X86::FR32RegisterClass ||
7319 Res.second == X86::FR64RegisterClass ||
7320 Res.second == X86::VR128RegisterClass) {
7321 // Handle references to XMM physical registers that got mapped into the
7322 // wrong class. This can happen with constraints like {xmm0} where the
7323 // target independent register mapper will just pick the first match it can
7324 // find, ignoring the required type.
7326 Res.second = X86::FR32RegisterClass;
7327 else if (VT == MVT::f64)
7328 Res.second = X86::FR64RegisterClass;
7329 else if (X86::VR128RegisterClass->hasType(VT))
7330 Res.second = X86::VR128RegisterClass;