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();
58 // Set up the TargetLowering object.
60 // X86 is weird, it always uses i8 for shift amounts and setcc results.
61 setShiftAmountType(MVT::i8);
62 setSetCCResultContents(ZeroOrOneSetCCResult);
63 setSchedulingPreference(SchedulingForRegPressure);
64 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
65 setStackPointerRegisterToSaveRestore(X86StackPtr);
67 if (Subtarget->isTargetDarwin()) {
68 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
69 setUseUnderscoreSetJmp(false);
70 setUseUnderscoreLongJmp(false);
71 } else if (Subtarget->isTargetMingw()) {
72 // MS runtime is weird: it exports _setjmp, but longjmp!
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(false);
76 setUseUnderscoreSetJmp(true);
77 setUseUnderscoreLongJmp(true);
80 // Set up the register classes.
81 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
82 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
83 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
84 if (Subtarget->is64Bit())
85 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
87 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 // We don't accept any truncstore of integer registers.
90 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
92 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
93 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
94 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
95 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
97 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
99 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
101 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
103 if (Subtarget->is64Bit()) {
104 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
105 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
108 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
109 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
111 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
114 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
116 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
117 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
118 // SSE has no i16 to fp conversion, only i32
119 if (X86ScalarSSEf32) {
120 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
121 // f32 and f64 cases are Legal, f80 case is not
122 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
124 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
125 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
128 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
129 // are Legal, f80 is custom lowered.
130 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
131 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
133 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
135 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
136 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
138 if (X86ScalarSSEf32) {
139 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
140 // f32 and f64 cases are Legal, f80 case is not
141 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
143 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
144 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
147 // Handle FP_TO_UINT by promoting the destination to a larger signed
149 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
150 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
151 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
153 if (Subtarget->is64Bit()) {
154 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
155 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
158 // Expand FP_TO_UINT into a select.
159 // FIXME: We would like to use a Custom expander here eventually to do
160 // the optimal thing for SSE vs. the default expansion in the legalizer.
161 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
163 // With SSE3 we can use fisttpll to convert to a signed i64.
164 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
167 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
168 if (!X86ScalarSSEf64) {
169 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
170 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
173 // Scalar integer divide and remainder are lowered to use operations that
174 // produce two results, to match the available instructions. This exposes
175 // the two-result form to trivial CSE, which is able to combine x/y and x%y
176 // into a single instruction.
178 // Scalar integer multiply-high is also lowered to use two-result
179 // operations, to match the available instructions. However, plain multiply
180 // (low) operations are left as Legal, as there are single-result
181 // instructions for this in x86. Using the two-result multiply instructions
182 // when both high and low results are needed must be arranged by dagcombine.
183 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
184 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
185 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
186 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
187 setOperationAction(ISD::SREM , MVT::i8 , Expand);
188 setOperationAction(ISD::UREM , MVT::i8 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
193 setOperationAction(ISD::SREM , MVT::i16 , Expand);
194 setOperationAction(ISD::UREM , MVT::i16 , Expand);
195 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
196 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
197 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
198 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
199 setOperationAction(ISD::SREM , MVT::i32 , Expand);
200 setOperationAction(ISD::UREM , MVT::i32 , Expand);
201 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
202 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
203 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
205 setOperationAction(ISD::SREM , MVT::i64 , Expand);
206 setOperationAction(ISD::UREM , MVT::i64 , Expand);
208 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
209 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
210 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
211 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
212 if (Subtarget->is64Bit())
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
216 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
217 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
218 setOperationAction(ISD::FREM , MVT::f32 , Expand);
219 setOperationAction(ISD::FREM , MVT::f64 , Expand);
220 setOperationAction(ISD::FREM , MVT::f80 , Expand);
221 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
223 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
224 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
225 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
226 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
227 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
228 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
229 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
230 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
231 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
232 if (Subtarget->is64Bit()) {
233 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
234 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
235 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
238 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
239 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
241 // These should be promoted to a larger select which is supported.
242 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
243 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
244 // X86 wants to expand cmov itself.
245 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
246 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
248 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
249 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
251 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
252 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
254 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
255 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
258 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
260 // X86 ret instruction may pop stack.
261 setOperationAction(ISD::RET , MVT::Other, Custom);
262 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
265 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
266 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
267 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
268 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
269 if (Subtarget->is64Bit())
270 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
271 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
272 if (Subtarget->is64Bit()) {
273 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
274 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
275 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
276 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
278 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
279 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
280 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
281 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
282 if (Subtarget->is64Bit()) {
283 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
284 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
285 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
288 if (Subtarget->hasSSE1())
289 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
291 if (!Subtarget->hasSSE2())
292 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
294 // Expand certain atomics
295 setOperationAction(ISD::ATOMIC_CMP_SWAP_8 , MVT::i8, Custom);
296 setOperationAction(ISD::ATOMIC_CMP_SWAP_16, MVT::i16, Custom);
297 setOperationAction(ISD::ATOMIC_CMP_SWAP_32, MVT::i32, Custom);
298 setOperationAction(ISD::ATOMIC_CMP_SWAP_64, MVT::i64, Custom);
300 setOperationAction(ISD::ATOMIC_LOAD_SUB_8, MVT::i8, Expand);
301 setOperationAction(ISD::ATOMIC_LOAD_SUB_16, MVT::i16, Expand);
302 setOperationAction(ISD::ATOMIC_LOAD_SUB_32, MVT::i32, Expand);
303 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Expand);
305 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
306 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
307 // FIXME - use subtarget debug flags
308 if (!Subtarget->isTargetDarwin() &&
309 !Subtarget->isTargetELF() &&
310 !Subtarget->isTargetCygMing()) {
311 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
312 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
315 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
316 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
317 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
318 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
319 if (Subtarget->is64Bit()) {
320 setExceptionPointerRegister(X86::RAX);
321 setExceptionSelectorRegister(X86::RDX);
323 setExceptionPointerRegister(X86::EAX);
324 setExceptionSelectorRegister(X86::EDX);
326 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
327 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
329 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
331 setOperationAction(ISD::TRAP, MVT::Other, Legal);
333 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
334 setOperationAction(ISD::VASTART , MVT::Other, Custom);
335 setOperationAction(ISD::VAEND , MVT::Other, Expand);
336 if (Subtarget->is64Bit()) {
337 setOperationAction(ISD::VAARG , MVT::Other, Custom);
338 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
340 setOperationAction(ISD::VAARG , MVT::Other, Expand);
341 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
344 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
345 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
346 if (Subtarget->is64Bit())
347 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
348 if (Subtarget->isTargetCygMing())
349 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
351 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
353 if (X86ScalarSSEf64) {
354 // f32 and f64 use SSE.
355 // Set up the FP register classes.
356 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
357 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
359 // Use ANDPD to simulate FABS.
360 setOperationAction(ISD::FABS , MVT::f64, Custom);
361 setOperationAction(ISD::FABS , MVT::f32, Custom);
363 // Use XORP to simulate FNEG.
364 setOperationAction(ISD::FNEG , MVT::f64, Custom);
365 setOperationAction(ISD::FNEG , MVT::f32, Custom);
367 // Use ANDPD and ORPD to simulate FCOPYSIGN.
368 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
369 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
371 // We don't support sin/cos/fmod
372 setOperationAction(ISD::FSIN , MVT::f64, Expand);
373 setOperationAction(ISD::FCOS , MVT::f64, Expand);
374 setOperationAction(ISD::FSIN , MVT::f32, Expand);
375 setOperationAction(ISD::FCOS , MVT::f32, Expand);
377 // Expand FP immediates into loads from the stack, except for the special
379 addLegalFPImmediate(APFloat(+0.0)); // xorpd
380 addLegalFPImmediate(APFloat(+0.0f)); // xorps
382 // Floating truncations from f80 and extensions to f80 go through memory.
383 // If optimizing, we lie about this though and handle it in
384 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
386 setConvertAction(MVT::f32, MVT::f80, Expand);
387 setConvertAction(MVT::f64, MVT::f80, Expand);
388 setConvertAction(MVT::f80, MVT::f32, Expand);
389 setConvertAction(MVT::f80, MVT::f64, Expand);
391 } else if (X86ScalarSSEf32) {
392 // Use SSE for f32, x87 for f64.
393 // Set up the FP register classes.
394 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
395 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
397 // Use ANDPS to simulate FABS.
398 setOperationAction(ISD::FABS , MVT::f32, Custom);
400 // Use XORP to simulate FNEG.
401 setOperationAction(ISD::FNEG , MVT::f32, Custom);
403 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
405 // Use ANDPS and ORPS to simulate FCOPYSIGN.
406 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
407 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
409 // We don't support sin/cos/fmod
410 setOperationAction(ISD::FSIN , MVT::f32, Expand);
411 setOperationAction(ISD::FCOS , MVT::f32, Expand);
413 // Special cases we handle for FP constants.
414 addLegalFPImmediate(APFloat(+0.0f)); // xorps
415 addLegalFPImmediate(APFloat(+0.0)); // FLD0
416 addLegalFPImmediate(APFloat(+1.0)); // FLD1
417 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
418 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
420 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
421 // this though and handle it in InstructionSelectPreprocess so that
422 // dagcombine2 can hack on these.
424 setConvertAction(MVT::f32, MVT::f64, Expand);
425 setConvertAction(MVT::f32, MVT::f80, Expand);
426 setConvertAction(MVT::f80, MVT::f32, Expand);
427 setConvertAction(MVT::f64, MVT::f32, Expand);
428 // And x87->x87 truncations also.
429 setConvertAction(MVT::f80, MVT::f64, Expand);
433 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
434 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
437 // f32 and f64 in x87.
438 // Set up the FP register classes.
439 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
440 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
442 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
443 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
447 // Floating truncations go through memory. If optimizing, we lie about
448 // this though and handle it in InstructionSelectPreprocess so that
449 // dagcombine2 can hack on these.
451 setConvertAction(MVT::f80, MVT::f32, Expand);
452 setConvertAction(MVT::f64, MVT::f32, Expand);
453 setConvertAction(MVT::f80, MVT::f64, Expand);
457 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
458 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
460 addLegalFPImmediate(APFloat(+0.0)); // FLD0
461 addLegalFPImmediate(APFloat(+1.0)); // FLD1
462 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
463 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
464 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
465 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
466 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
467 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
470 // Long double always uses X87.
471 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
472 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
475 APFloat TmpFlt(+0.0);
476 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
477 addLegalFPImmediate(TmpFlt); // FLD0
479 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
480 APFloat TmpFlt2(+1.0);
481 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
482 addLegalFPImmediate(TmpFlt2); // FLD1
483 TmpFlt2.changeSign();
484 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
488 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
489 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
492 // Always use a library call for pow.
493 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
494 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
495 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
497 setOperationAction(ISD::FLOG, MVT::f32, Expand);
498 setOperationAction(ISD::FLOG, MVT::f64, Expand);
499 setOperationAction(ISD::FLOG, MVT::f80, Expand);
500 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
501 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
502 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
503 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
504 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
505 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
506 setOperationAction(ISD::FEXP, MVT::f32, Expand);
507 setOperationAction(ISD::FEXP, MVT::f64, Expand);
508 setOperationAction(ISD::FEXP, MVT::f80, Expand);
509 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
510 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
511 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
513 // First set operation action for all vector types to expand. Then we
514 // will selectively turn on ones that can be effectively codegen'd.
515 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
516 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
517 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
532 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
562 if (Subtarget->hasMMX()) {
563 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
564 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
565 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
566 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
567 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
569 // FIXME: add MMX packed arithmetics
571 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
572 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
573 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
574 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
576 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
577 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
578 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
579 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
581 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
582 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
584 setOperationAction(ISD::AND, MVT::v8i8, Promote);
585 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
586 setOperationAction(ISD::AND, MVT::v4i16, Promote);
587 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
588 setOperationAction(ISD::AND, MVT::v2i32, Promote);
589 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
590 setOperationAction(ISD::AND, MVT::v1i64, Legal);
592 setOperationAction(ISD::OR, MVT::v8i8, Promote);
593 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
594 setOperationAction(ISD::OR, MVT::v4i16, Promote);
595 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
596 setOperationAction(ISD::OR, MVT::v2i32, Promote);
597 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
598 setOperationAction(ISD::OR, MVT::v1i64, Legal);
600 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
601 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
602 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
603 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
604 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
605 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
606 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
608 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
609 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
610 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
611 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
612 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
613 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
614 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
615 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
616 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
619 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
620 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
621 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
622 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
624 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
625 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
626 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
629 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
630 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
631 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
632 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
634 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
637 if (Subtarget->hasSSE1()) {
638 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
640 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
641 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
642 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
643 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
644 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
645 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
646 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
648 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
649 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
650 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
651 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
654 if (Subtarget->hasSSE2()) {
655 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
656 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
657 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
658 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
659 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
661 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
662 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
663 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
664 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
665 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
666 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
667 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
668 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
669 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
670 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
671 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
672 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
673 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
674 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
675 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
677 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
678 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
679 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
680 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
683 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
685 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
686 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
688 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
689 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
690 MVT VT = (MVT::SimpleValueType)i;
691 // Do not attempt to custom lower non-power-of-2 vectors
692 if (!isPowerOf2_32(VT.getVectorNumElements()))
694 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
695 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
696 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
698 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
699 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
700 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
701 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
702 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
703 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
704 if (Subtarget->is64Bit()) {
705 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
706 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
709 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
710 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
711 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
712 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
713 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
714 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
715 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
716 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
717 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
718 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
719 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
720 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
723 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
725 // Custom lower v2i64 and v2f64 selects.
726 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
727 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
728 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
729 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
733 if (Subtarget->hasSSE41()) {
734 // FIXME: Do we need to handle scalar-to-vector here?
735 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
736 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
738 // i8 and i16 vectors are custom , because the source register and source
739 // source memory operand types are not the same width. f32 vectors are
740 // custom since the immediate controlling the insert encodes additional
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
743 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
744 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
745 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
747 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
748 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
749 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
750 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
752 if (Subtarget->is64Bit()) {
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
754 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
758 if (Subtarget->hasSSE42()) {
759 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
762 // We want to custom lower some of our intrinsics.
763 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
765 // We have target-specific dag combine patterns for the following nodes:
766 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
767 setTargetDAGCombine(ISD::BUILD_VECTOR);
768 setTargetDAGCombine(ISD::SELECT);
769 setTargetDAGCombine(ISD::STORE);
771 computeRegisterProperties();
773 // FIXME: These should be based on subtarget info. Plus, the values should
774 // be smaller when we are in optimizing for size mode.
775 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
776 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
777 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
778 allowUnalignedMemoryAccesses = true; // x86 supports it!
779 setPrefLoopAlignment(16);
783 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
788 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
789 /// the desired ByVal argument alignment.
790 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
793 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
794 if (VTy->getBitWidth() == 128)
796 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
797 unsigned EltAlign = 0;
798 getMaxByValAlign(ATy->getElementType(), EltAlign);
799 if (EltAlign > MaxAlign)
801 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
802 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
803 unsigned EltAlign = 0;
804 getMaxByValAlign(STy->getElementType(i), EltAlign);
805 if (EltAlign > MaxAlign)
814 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
815 /// function arguments in the caller parameter area. For X86, aggregates
816 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
817 /// are at 4-byte boundaries.
818 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
819 if (Subtarget->is64Bit()) {
820 // Max of 8 and alignment of type.
821 unsigned TyAlign = TD->getABITypeAlignment(Ty);
828 if (Subtarget->hasSSE1())
829 getMaxByValAlign(Ty, Align);
833 /// getOptimalMemOpType - Returns the target specific optimal type for load
834 /// and store operations as a result of memset, memcpy, and memmove
835 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
838 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
839 bool isSrcConst, bool isSrcStr) const {
840 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
842 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
844 if (Subtarget->is64Bit() && Size >= 8)
850 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
852 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
853 SelectionDAG &DAG) const {
854 if (usesGlobalOffsetTable())
855 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
856 if (!Subtarget->isPICStyleRIPRel())
857 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
861 //===----------------------------------------------------------------------===//
862 // Return Value Calling Convention Implementation
863 //===----------------------------------------------------------------------===//
865 #include "X86GenCallingConv.inc"
867 /// LowerRET - Lower an ISD::RET node.
868 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
869 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
871 SmallVector<CCValAssign, 16> RVLocs;
872 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
873 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
874 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
875 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
877 // If this is the first return lowered for this function, add the regs to the
878 // liveout set for the function.
879 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
880 for (unsigned i = 0; i != RVLocs.size(); ++i)
881 if (RVLocs[i].isRegLoc())
882 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
884 SDValue Chain = Op.getOperand(0);
886 // Handle tail call return.
887 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
888 if (Chain.getOpcode() == X86ISD::TAILCALL) {
889 SDValue TailCall = Chain;
890 SDValue TargetAddress = TailCall.getOperand(1);
891 SDValue StackAdjustment = TailCall.getOperand(2);
892 assert(((TargetAddress.getOpcode() == ISD::Register &&
893 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
894 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
895 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
896 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
897 "Expecting an global address, external symbol, or register");
898 assert(StackAdjustment.getOpcode() == ISD::Constant &&
899 "Expecting a const value");
901 SmallVector<SDValue,8> Operands;
902 Operands.push_back(Chain.getOperand(0));
903 Operands.push_back(TargetAddress);
904 Operands.push_back(StackAdjustment);
905 // Copy registers used by the call. Last operand is a flag so it is not
907 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
908 Operands.push_back(Chain.getOperand(i));
910 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
917 SmallVector<SDValue, 6> RetOps;
918 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
919 // Operand #1 = Bytes To Pop
920 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
922 // Copy the result values into the output registers.
923 for (unsigned i = 0; i != RVLocs.size(); ++i) {
924 CCValAssign &VA = RVLocs[i];
925 assert(VA.isRegLoc() && "Can only return in registers!");
926 SDValue ValToCopy = Op.getOperand(i*2+1);
928 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
929 // the RET instruction and handled by the FP Stackifier.
930 if (RVLocs[i].getLocReg() == X86::ST0 ||
931 RVLocs[i].getLocReg() == X86::ST1) {
932 // If this is a copy from an xmm register to ST(0), use an FPExtend to
933 // change the value to the FP stack register class.
934 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
935 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
936 RetOps.push_back(ValToCopy);
937 // Don't emit a copytoreg.
941 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
942 Flag = Chain.getValue(1);
945 // The x86-64 ABI for returning structs by value requires that we copy
946 // the sret argument into %rax for the return. We saved the argument into
947 // a virtual register in the entry block, so now we copy the value out
949 if (Subtarget->is64Bit() &&
950 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
951 MachineFunction &MF = DAG.getMachineFunction();
952 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
953 unsigned Reg = FuncInfo->getSRetReturnReg();
955 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
956 FuncInfo->setSRetReturnReg(Reg);
958 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
960 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
961 Flag = Chain.getValue(1);
964 RetOps[0] = Chain; // Update chain.
966 // Add the flag if we have it.
968 RetOps.push_back(Flag);
970 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
974 /// LowerCallResult - Lower the result values of an ISD::CALL into the
975 /// appropriate copies out of appropriate physical registers. This assumes that
976 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
977 /// being lowered. The returns a SDNode with the same number of values as the
979 SDNode *X86TargetLowering::
980 LowerCallResult(SDValue Chain, SDValue InFlag, SDNode *TheCall,
981 unsigned CallingConv, SelectionDAG &DAG) {
983 // Assign locations to each value returned by this call.
984 SmallVector<CCValAssign, 16> RVLocs;
985 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
986 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
987 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
989 SmallVector<SDValue, 8> ResultVals;
991 // Copy all of the result registers out of their specified physreg.
992 for (unsigned i = 0; i != RVLocs.size(); ++i) {
993 MVT CopyVT = RVLocs[i].getValVT();
995 // If this is a call to a function that returns an fp value on the floating
996 // point stack, but where we prefer to use the value in xmm registers, copy
997 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
998 if ((RVLocs[i].getLocReg() == X86::ST0 ||
999 RVLocs[i].getLocReg() == X86::ST1) &&
1000 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
1004 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1005 CopyVT, InFlag).getValue(1);
1006 SDValue Val = Chain.getValue(0);
1007 InFlag = Chain.getValue(2);
1009 if (CopyVT != RVLocs[i].getValVT()) {
1010 // Round the F80 the right size, which also moves to the appropriate xmm
1012 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1013 // This truncation won't change the value.
1014 DAG.getIntPtrConstant(1));
1017 ResultVals.push_back(Val);
1020 // Merge everything together with a MERGE_VALUES node.
1021 ResultVals.push_back(Chain);
1022 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1023 ResultVals.size()).getNode();
1027 //===----------------------------------------------------------------------===//
1028 // C & StdCall & Fast Calling Convention implementation
1029 //===----------------------------------------------------------------------===//
1030 // StdCall calling convention seems to be standard for many Windows' API
1031 // routines and around. It differs from C calling convention just a little:
1032 // callee should clean up the stack, not caller. Symbols should be also
1033 // decorated in some fancy way :) It doesn't support any vector arguments.
1034 // For info on fast calling convention see Fast Calling Convention (tail call)
1035 // implementation LowerX86_32FastCCCallTo.
1037 /// AddLiveIn - This helper function adds the specified physical register to the
1038 /// MachineFunction as a live in value. It also creates a corresponding virtual
1039 /// register for it.
1040 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1041 const TargetRegisterClass *RC) {
1042 assert(RC->contains(PReg) && "Not the correct regclass!");
1043 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1044 MF.getRegInfo().addLiveIn(PReg, VReg);
1048 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1050 static bool CallIsStructReturn(SDValue Op) {
1051 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1055 return cast<ARG_FLAGSSDNode>(Op.getOperand(6))->getArgFlags().isSRet();
1058 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1059 /// return semantics.
1060 static bool ArgsAreStructReturn(SDValue Op) {
1061 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1065 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1068 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1069 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1071 bool X86TargetLowering::IsCalleePop(SDValue Op) {
1072 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1076 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1079 case CallingConv::X86_StdCall:
1080 return !Subtarget->is64Bit();
1081 case CallingConv::X86_FastCall:
1082 return !Subtarget->is64Bit();
1083 case CallingConv::Fast:
1084 return PerformTailCallOpt;
1088 /// CCAssignFnForNode - Selects the correct CCAssignFn for a CALL or
1089 /// FORMAL_ARGUMENTS node.
1090 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDValue Op) const {
1091 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1093 if (Subtarget->is64Bit()) {
1094 if (Subtarget->isTargetWin64())
1095 return CC_X86_Win64_C;
1096 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1097 return CC_X86_64_TailCall;
1102 if (CC == CallingConv::X86_FastCall)
1103 return CC_X86_32_FastCall;
1104 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1105 return CC_X86_32_TailCall;
1106 else if (CC == CallingConv::Fast)
1107 return CC_X86_32_FastCC;
1112 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1113 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1115 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1116 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1117 if (CC == CallingConv::X86_FastCall)
1119 else if (CC == CallingConv::X86_StdCall)
1125 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1126 /// in a register before calling.
1127 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1128 return !IsTailCall && !Is64Bit &&
1129 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1130 Subtarget->isPICStyleGOT();
1133 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1134 /// address to be loaded in a register.
1136 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1137 return !Is64Bit && IsTailCall &&
1138 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT();
1142 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1143 /// by "Src" to address "Dst" with size and alignment information specified by
1144 /// the specific parameter attribute. The copy will be passed as a byval
1145 /// function parameter.
1147 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1148 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1149 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1150 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1151 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1154 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1155 const CCValAssign &VA,
1156 MachineFrameInfo *MFI,
1158 SDValue Root, unsigned i) {
1159 // Create the nodes corresponding to a load from this parameter slot.
1160 ISD::ArgFlagsTy Flags =
1161 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1162 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1163 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1165 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1166 // changed with more analysis.
1167 // In case of tail call optimization mark all arguments mutable. Since they
1168 // could be overwritten by lowering of arguments in case of a tail call.
1169 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1170 VA.getLocMemOffset(), isImmutable);
1171 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1172 if (Flags.isByVal())
1174 return DAG.getLoad(VA.getValVT(), Root, FIN,
1175 PseudoSourceValue::getFixedStack(FI), 0);
1179 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1180 MachineFunction &MF = DAG.getMachineFunction();
1181 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1183 const Function* Fn = MF.getFunction();
1184 if (Fn->hasExternalLinkage() &&
1185 Subtarget->isTargetCygMing() &&
1186 Fn->getName() == "main")
1187 FuncInfo->setForceFramePointer(true);
1189 // Decorate the function name.
1190 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1192 MachineFrameInfo *MFI = MF.getFrameInfo();
1193 SDValue Root = Op.getOperand(0);
1194 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1195 unsigned CC = MF.getFunction()->getCallingConv();
1196 bool Is64Bit = Subtarget->is64Bit();
1197 bool IsWin64 = Subtarget->isTargetWin64();
1199 assert(!(isVarArg && CC == CallingConv::Fast) &&
1200 "Var args not supported with calling convention fastcc");
1202 // Assign locations to all of the incoming arguments.
1203 SmallVector<CCValAssign, 16> ArgLocs;
1204 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1205 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(Op));
1207 SmallVector<SDValue, 8> ArgValues;
1208 unsigned LastVal = ~0U;
1209 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1210 CCValAssign &VA = ArgLocs[i];
1211 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1213 assert(VA.getValNo() != LastVal &&
1214 "Don't support value assigned to multiple locs yet");
1215 LastVal = VA.getValNo();
1217 if (VA.isRegLoc()) {
1218 MVT RegVT = VA.getLocVT();
1219 TargetRegisterClass *RC;
1220 if (RegVT == MVT::i32)
1221 RC = X86::GR32RegisterClass;
1222 else if (Is64Bit && RegVT == MVT::i64)
1223 RC = X86::GR64RegisterClass;
1224 else if (RegVT == MVT::f32)
1225 RC = X86::FR32RegisterClass;
1226 else if (RegVT == MVT::f64)
1227 RC = X86::FR64RegisterClass;
1228 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1229 RC = X86::VR128RegisterClass;
1230 else if (RegVT.isVector()) {
1231 assert(RegVT.getSizeInBits() == 64);
1233 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1235 // Darwin calling convention passes MMX values in either GPRs or
1236 // XMMs in x86-64. Other targets pass them in memory.
1237 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1238 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1241 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1246 assert(0 && "Unknown argument type!");
1249 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1250 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1252 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1253 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1255 if (VA.getLocInfo() == CCValAssign::SExt)
1256 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1257 DAG.getValueType(VA.getValVT()));
1258 else if (VA.getLocInfo() == CCValAssign::ZExt)
1259 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1260 DAG.getValueType(VA.getValVT()));
1262 if (VA.getLocInfo() != CCValAssign::Full)
1263 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1265 // Handle MMX values passed in GPRs.
1266 if (Is64Bit && RegVT != VA.getLocVT()) {
1267 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1268 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1269 else if (RC == X86::VR128RegisterClass) {
1270 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1271 DAG.getConstant(0, MVT::i64));
1272 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1276 ArgValues.push_back(ArgValue);
1278 assert(VA.isMemLoc());
1279 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1283 // The x86-64 ABI for returning structs by value requires that we copy
1284 // the sret argument into %rax for the return. Save the argument into
1285 // a virtual register so that we can access it from the return points.
1286 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1287 MachineFunction &MF = DAG.getMachineFunction();
1288 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1289 unsigned Reg = FuncInfo->getSRetReturnReg();
1291 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1292 FuncInfo->setSRetReturnReg(Reg);
1294 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1295 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1298 unsigned StackSize = CCInfo.getNextStackOffset();
1299 // align stack specially for tail calls
1300 if (PerformTailCallOpt && CC == CallingConv::Fast)
1301 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1303 // If the function takes variable number of arguments, make a frame index for
1304 // the start of the first vararg value... for expansion of llvm.va_start.
1306 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1307 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1310 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1312 // FIXME: We should really autogenerate these arrays
1313 static const unsigned GPR64ArgRegsWin64[] = {
1314 X86::RCX, X86::RDX, X86::R8, X86::R9
1316 static const unsigned XMMArgRegsWin64[] = {
1317 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1319 static const unsigned GPR64ArgRegs64Bit[] = {
1320 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1322 static const unsigned XMMArgRegs64Bit[] = {
1323 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1324 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1326 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1329 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1330 GPR64ArgRegs = GPR64ArgRegsWin64;
1331 XMMArgRegs = XMMArgRegsWin64;
1333 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1334 GPR64ArgRegs = GPR64ArgRegs64Bit;
1335 XMMArgRegs = XMMArgRegs64Bit;
1337 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1339 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1342 // For X86-64, if there are vararg parameters that are passed via
1343 // registers, then we must store them to their spots on the stack so they
1344 // may be loaded by deferencing the result of va_next.
1345 VarArgsGPOffset = NumIntRegs * 8;
1346 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1347 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1348 TotalNumXMMRegs * 16, 16);
1350 // Store the integer parameter registers.
1351 SmallVector<SDValue, 8> MemOps;
1352 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1353 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1354 DAG.getIntPtrConstant(VarArgsGPOffset));
1355 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1356 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1357 X86::GR64RegisterClass);
1358 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1360 DAG.getStore(Val.getValue(1), Val, FIN,
1361 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1362 MemOps.push_back(Store);
1363 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1364 DAG.getIntPtrConstant(8));
1367 // Now store the XMM (fp + vector) parameter registers.
1368 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1369 DAG.getIntPtrConstant(VarArgsFPOffset));
1370 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1371 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1372 X86::VR128RegisterClass);
1373 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1375 DAG.getStore(Val.getValue(1), Val, FIN,
1376 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1377 MemOps.push_back(Store);
1378 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1379 DAG.getIntPtrConstant(16));
1381 if (!MemOps.empty())
1382 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1383 &MemOps[0], MemOps.size());
1387 ArgValues.push_back(Root);
1389 // Some CCs need callee pop.
1390 if (IsCalleePop(Op)) {
1391 BytesToPopOnReturn = StackSize; // Callee pops everything.
1392 BytesCallerReserves = 0;
1394 BytesToPopOnReturn = 0; // Callee pops nothing.
1395 // If this is an sret function, the return should pop the hidden pointer.
1396 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1397 BytesToPopOnReturn = 4;
1398 BytesCallerReserves = StackSize;
1402 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1403 if (CC == CallingConv::X86_FastCall)
1404 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1407 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1409 // Return the new list of results.
1410 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1411 ArgValues.size()).getValue(Op.getResNo());
1415 X86TargetLowering::LowerMemOpCallTo(SDValue Op, SelectionDAG &DAG,
1416 const SDValue &StackPtr,
1417 const CCValAssign &VA,
1420 unsigned LocMemOffset = VA.getLocMemOffset();
1421 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1422 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1423 ISD::ArgFlagsTy Flags =
1424 cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->getArgFlags();
1425 if (Flags.isByVal()) {
1426 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1428 return DAG.getStore(Chain, Arg, PtrOff,
1429 PseudoSourceValue::getStack(), LocMemOffset);
1432 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1433 /// optimization is performed and it is required.
1435 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1436 SDValue &OutRetAddr,
1441 if (!IsTailCall || FPDiff==0) return Chain;
1443 // Adjust the Return address stack slot.
1444 MVT VT = getPointerTy();
1445 OutRetAddr = getReturnAddressFrameIndex(DAG);
1446 // Load the "old" Return address.
1447 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1448 return SDValue(OutRetAddr.getNode(), 1);
1451 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1452 /// optimization is performed and it is required (FPDiff!=0).
1454 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1455 SDValue Chain, SDValue RetAddrFrIdx,
1456 bool Is64Bit, int FPDiff) {
1457 // Store the return address to the appropriate stack slot.
1458 if (!FPDiff) return Chain;
1459 // Calculate the new stack slot for the return address.
1460 int SlotSize = Is64Bit ? 8 : 4;
1461 int NewReturnAddrFI =
1462 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1463 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1464 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1465 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1466 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1470 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1471 MachineFunction &MF = DAG.getMachineFunction();
1472 SDValue Chain = Op.getOperand(0);
1473 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1474 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1475 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1476 && CC == CallingConv::Fast && PerformTailCallOpt;
1477 SDValue Callee = Op.getOperand(4);
1478 bool Is64Bit = Subtarget->is64Bit();
1479 bool IsStructRet = CallIsStructReturn(Op);
1481 assert(!(isVarArg && CC == CallingConv::Fast) &&
1482 "Var args not supported with calling convention fastcc");
1484 // Analyze operands of the call, assigning locations to each operand.
1485 SmallVector<CCValAssign, 16> ArgLocs;
1486 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1487 CCInfo.AnalyzeCallOperands(Op.getNode(), CCAssignFnForNode(Op));
1489 // Get a count of how many bytes are to be pushed on the stack.
1490 unsigned NumBytes = CCInfo.getNextStackOffset();
1491 if (PerformTailCallOpt && CC == CallingConv::Fast)
1492 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1496 // Lower arguments at fp - stackoffset + fpdiff.
1497 unsigned NumBytesCallerPushed =
1498 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1499 FPDiff = NumBytesCallerPushed - NumBytes;
1501 // Set the delta of movement of the returnaddr stackslot.
1502 // But only set if delta is greater than previous delta.
1503 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1504 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1507 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1509 SDValue RetAddrFrIdx;
1510 // Load return adress for tail calls.
1511 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1514 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1515 SmallVector<SDValue, 8> MemOpChains;
1518 // Walk the register/memloc assignments, inserting copies/loads. In the case
1519 // of tail call optimization arguments are handle later.
1520 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1521 CCValAssign &VA = ArgLocs[i];
1522 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1523 bool isByVal = cast<ARG_FLAGSSDNode>(Op.getOperand(6+2*VA.getValNo()))->
1524 getArgFlags().isByVal();
1526 // Promote the value if needed.
1527 switch (VA.getLocInfo()) {
1528 default: assert(0 && "Unknown loc info!");
1529 case CCValAssign::Full: break;
1530 case CCValAssign::SExt:
1531 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1533 case CCValAssign::ZExt:
1534 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1536 case CCValAssign::AExt:
1537 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1541 if (VA.isRegLoc()) {
1543 MVT RegVT = VA.getLocVT();
1544 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1545 switch (VA.getLocReg()) {
1548 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1550 // Special case: passing MMX values in GPR registers.
1551 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1554 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1555 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1556 // Special case: passing MMX values in XMM registers.
1557 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1558 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1559 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1560 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1561 getMOVLMask(2, DAG));
1566 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1568 if (!IsTailCall || (IsTailCall && isByVal)) {
1569 assert(VA.isMemLoc());
1570 if (StackPtr.getNode() == 0)
1571 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1573 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1579 if (!MemOpChains.empty())
1580 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1581 &MemOpChains[0], MemOpChains.size());
1583 // Build a sequence of copy-to-reg nodes chained together with token chain
1584 // and flag operands which copy the outgoing args into registers.
1586 // Tail call byval lowering might overwrite argument registers so in case of
1587 // tail call optimization the copies to registers are lowered later.
1589 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1590 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1592 InFlag = Chain.getValue(1);
1595 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1597 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1598 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1599 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1601 InFlag = Chain.getValue(1);
1603 // If we are tail calling and generating PIC/GOT style code load the address
1604 // of the callee into ecx. The value in ecx is used as target of the tail
1605 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1606 // calls on PIC/GOT architectures. Normally we would just put the address of
1607 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1608 // restored (since ebx is callee saved) before jumping to the target@PLT.
1609 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1610 // Note: The actual moving to ecx is done further down.
1611 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1612 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1613 !G->getGlobal()->hasProtectedVisibility())
1614 Callee = LowerGlobalAddress(Callee, DAG);
1615 else if (isa<ExternalSymbolSDNode>(Callee))
1616 Callee = LowerExternalSymbol(Callee,DAG);
1619 if (Is64Bit && isVarArg) {
1620 // From AMD64 ABI document:
1621 // For calls that may call functions that use varargs or stdargs
1622 // (prototype-less calls or calls to functions containing ellipsis (...) in
1623 // the declaration) %al is used as hidden argument to specify the number
1624 // of SSE registers used. The contents of %al do not need to match exactly
1625 // the number of registers, but must be an ubound on the number of SSE
1626 // registers used and is in the range 0 - 8 inclusive.
1628 // FIXME: Verify this on Win64
1629 // Count the number of XMM registers allocated.
1630 static const unsigned XMMArgRegs[] = {
1631 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1632 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1634 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1636 Chain = DAG.getCopyToReg(Chain, X86::AL,
1637 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1638 InFlag = Chain.getValue(1);
1642 // For tail calls lower the arguments to the 'real' stack slot.
1644 SmallVector<SDValue, 8> MemOpChains2;
1647 // Do not flag preceeding copytoreg stuff together with the following stuff.
1649 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1650 CCValAssign &VA = ArgLocs[i];
1651 if (!VA.isRegLoc()) {
1652 assert(VA.isMemLoc());
1653 SDValue Arg = Op.getOperand(5+2*VA.getValNo());
1654 SDValue FlagsOp = Op.getOperand(6+2*VA.getValNo());
1655 ISD::ArgFlagsTy Flags =
1656 cast<ARG_FLAGSSDNode>(FlagsOp)->getArgFlags();
1657 // Create frame index.
1658 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1659 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1660 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1661 FIN = DAG.getFrameIndex(FI, getPointerTy());
1663 if (Flags.isByVal()) {
1664 // Copy relative to framepointer.
1665 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1666 if (StackPtr.getNode() == 0)
1667 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1668 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1670 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1673 // Store relative to framepointer.
1674 MemOpChains2.push_back(
1675 DAG.getStore(Chain, Arg, FIN,
1676 PseudoSourceValue::getFixedStack(FI), 0));
1681 if (!MemOpChains2.empty())
1682 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1683 &MemOpChains2[0], MemOpChains2.size());
1685 // Copy arguments to their registers.
1686 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1687 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1689 InFlag = Chain.getValue(1);
1693 // Store the return address to the appropriate stack slot.
1694 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1698 // If the callee is a GlobalAddress node (quite common, every direct call is)
1699 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1700 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1701 // We should use extra load for direct calls to dllimported functions in
1703 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1704 getTargetMachine(), true))
1705 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1706 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1707 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1708 } else if (IsTailCall) {
1709 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1711 Chain = DAG.getCopyToReg(Chain,
1712 DAG.getRegister(Opc, getPointerTy()),
1714 Callee = DAG.getRegister(Opc, getPointerTy());
1715 // Add register as live out.
1716 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1719 // Returns a chain & a flag for retval copy to use.
1720 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1721 SmallVector<SDValue, 8> Ops;
1724 Ops.push_back(Chain);
1725 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1726 Ops.push_back(DAG.getIntPtrConstant(0));
1727 if (InFlag.getNode())
1728 Ops.push_back(InFlag);
1729 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1730 InFlag = Chain.getValue(1);
1732 // Returns a chain & a flag for retval copy to use.
1733 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1737 Ops.push_back(Chain);
1738 Ops.push_back(Callee);
1741 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1743 // Add argument registers to the end of the list so that they are known live
1745 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1746 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1747 RegsToPass[i].second.getValueType()));
1749 // Add an implicit use GOT pointer in EBX.
1750 if (!IsTailCall && !Is64Bit &&
1751 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1752 Subtarget->isPICStyleGOT())
1753 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1755 // Add an implicit use of AL for x86 vararg functions.
1756 if (Is64Bit && isVarArg)
1757 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1759 if (InFlag.getNode())
1760 Ops.push_back(InFlag);
1763 assert(InFlag.getNode() &&
1764 "Flag must be set. Depend on flag being set in LowerRET");
1765 Chain = DAG.getNode(X86ISD::TAILCALL,
1766 Op.getNode()->getVTList(), &Ops[0], Ops.size());
1768 return SDValue(Chain.getNode(), Op.getResNo());
1771 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1772 InFlag = Chain.getValue(1);
1774 // Create the CALLSEQ_END node.
1775 unsigned NumBytesForCalleeToPush;
1776 if (IsCalleePop(Op))
1777 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1778 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1779 // If this is is a call to a struct-return function, the callee
1780 // pops the hidden struct pointer, so we have to push it back.
1781 // This is common for Darwin/X86, Linux & Mingw32 targets.
1782 NumBytesForCalleeToPush = 4;
1784 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1786 // Returns a flag for retval copy to use.
1787 Chain = DAG.getCALLSEQ_END(Chain,
1788 DAG.getIntPtrConstant(NumBytes),
1789 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1791 InFlag = Chain.getValue(1);
1793 // Handle result values, copying them out of physregs into vregs that we
1795 return SDValue(LowerCallResult(Chain, InFlag, Op.getNode(), CC, DAG),
1800 //===----------------------------------------------------------------------===//
1801 // Fast Calling Convention (tail call) implementation
1802 //===----------------------------------------------------------------------===//
1804 // Like std call, callee cleans arguments, convention except that ECX is
1805 // reserved for storing the tail called function address. Only 2 registers are
1806 // free for argument passing (inreg). Tail call optimization is performed
1808 // * tailcallopt is enabled
1809 // * caller/callee are fastcc
1810 // On X86_64 architecture with GOT-style position independent code only local
1811 // (within module) calls are supported at the moment.
1812 // To keep the stack aligned according to platform abi the function
1813 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1814 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1815 // If a tail called function callee has more arguments than the caller the
1816 // caller needs to make sure that there is room to move the RETADDR to. This is
1817 // achieved by reserving an area the size of the argument delta right after the
1818 // original REtADDR, but before the saved framepointer or the spilled registers
1819 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1831 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1832 /// for a 16 byte align requirement.
1833 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1834 SelectionDAG& DAG) {
1835 MachineFunction &MF = DAG.getMachineFunction();
1836 const TargetMachine &TM = MF.getTarget();
1837 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1838 unsigned StackAlignment = TFI.getStackAlignment();
1839 uint64_t AlignMask = StackAlignment - 1;
1840 int64_t Offset = StackSize;
1841 uint64_t SlotSize = TD->getPointerSize();
1842 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1843 // Number smaller than 12 so just add the difference.
1844 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1846 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1847 Offset = ((~AlignMask) & Offset) + StackAlignment +
1848 (StackAlignment-SlotSize);
1853 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1854 /// following the call is a return. A function is eligible if caller/callee
1855 /// calling conventions match, currently only fastcc supports tail calls, and
1856 /// the function CALL is immediatly followed by a RET.
1857 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1859 SelectionDAG& DAG) const {
1860 if (!PerformTailCallOpt)
1863 if (CheckTailCallReturnConstraints(Call, Ret)) {
1864 MachineFunction &MF = DAG.getMachineFunction();
1865 unsigned CallerCC = MF.getFunction()->getCallingConv();
1866 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1867 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1868 SDValue Callee = Call.getOperand(4);
1869 // On x86/32Bit PIC/GOT tail calls are supported.
1870 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1871 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1874 // Can only do local tail calls (in same module, hidden or protected) on
1875 // x86_64 PIC/GOT at the moment.
1876 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1877 return G->getGlobal()->hasHiddenVisibility()
1878 || G->getGlobal()->hasProtectedVisibility();
1886 X86TargetLowering::createFastISel(MachineFunction &mf,
1887 DenseMap<const Value *, unsigned> &vm,
1888 DenseMap<const BasicBlock *,
1889 MachineBasicBlock *> &bm,
1890 DenseMap<const AllocaInst *, int> &am) {
1892 return X86::createFastISel(mf, vm, bm, am);
1896 //===----------------------------------------------------------------------===//
1897 // Other Lowering Hooks
1898 //===----------------------------------------------------------------------===//
1901 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1904 int ReturnAddrIndex = FuncInfo->getRAIndex();
1905 uint64_t SlotSize = TD->getPointerSize();
1907 if (ReturnAddrIndex == 0) {
1908 // Set up a frame object for the return address.
1909 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1910 FuncInfo->setRAIndex(ReturnAddrIndex);
1913 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1917 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1918 /// specific condition code. It returns a false if it cannot do a direct
1919 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1921 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1922 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1923 SelectionDAG &DAG) {
1924 X86CC = X86::COND_INVALID;
1926 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1927 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1928 // X > -1 -> X == 0, jump !sign.
1929 RHS = DAG.getConstant(0, RHS.getValueType());
1930 X86CC = X86::COND_NS;
1932 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1933 // X < 0 -> X == 0, jump on sign.
1934 X86CC = X86::COND_S;
1936 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1938 RHS = DAG.getConstant(0, RHS.getValueType());
1939 X86CC = X86::COND_LE;
1944 switch (SetCCOpcode) {
1946 case ISD::SETEQ: X86CC = X86::COND_E; break;
1947 case ISD::SETGT: X86CC = X86::COND_G; break;
1948 case ISD::SETGE: X86CC = X86::COND_GE; break;
1949 case ISD::SETLT: X86CC = X86::COND_L; break;
1950 case ISD::SETLE: X86CC = X86::COND_LE; break;
1951 case ISD::SETNE: X86CC = X86::COND_NE; break;
1952 case ISD::SETULT: X86CC = X86::COND_B; break;
1953 case ISD::SETUGT: X86CC = X86::COND_A; break;
1954 case ISD::SETULE: X86CC = X86::COND_BE; break;
1955 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1958 // First determine if it requires or is profitable to flip the operands.
1960 switch (SetCCOpcode) {
1970 // If LHS is a foldable load, but RHS is not, flip the condition.
1972 (ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
1973 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
1974 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
1978 std::swap(LHS, RHS);
1980 // On a floating point condition, the flags are set as follows:
1982 // 0 | 0 | 0 | X > Y
1983 // 0 | 0 | 1 | X < Y
1984 // 1 | 0 | 0 | X == Y
1985 // 1 | 1 | 1 | unordered
1986 switch (SetCCOpcode) {
1990 X86CC = X86::COND_E;
1992 case ISD::SETOLT: // flipped
1995 X86CC = X86::COND_A;
1997 case ISD::SETOLE: // flipped
2000 X86CC = X86::COND_AE;
2002 case ISD::SETUGT: // flipped
2005 X86CC = X86::COND_B;
2007 case ISD::SETUGE: // flipped
2010 X86CC = X86::COND_BE;
2014 X86CC = X86::COND_NE;
2017 X86CC = X86::COND_P;
2020 X86CC = X86::COND_NP;
2025 return X86CC != X86::COND_INVALID;
2028 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2029 /// code. Current x86 isa includes the following FP cmov instructions:
2030 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2031 static bool hasFPCMov(unsigned X86CC) {
2047 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2048 /// true if Op is undef or if its value falls within the specified range (L, H].
2049 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2050 if (Op.getOpcode() == ISD::UNDEF)
2053 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2054 return (Val >= Low && Val < Hi);
2057 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2058 /// true if Op is undef or if its value equal to the specified value.
2059 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2060 if (Op.getOpcode() == ISD::UNDEF)
2062 return cast<ConstantSDNode>(Op)->getValue() == Val;
2065 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2066 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2067 bool X86::isPSHUFDMask(SDNode *N) {
2068 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2070 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2073 // Check if the value doesn't reference the second vector.
2074 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2075 SDValue Arg = N->getOperand(i);
2076 if (Arg.getOpcode() == ISD::UNDEF) continue;
2077 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2078 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2085 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2086 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2087 bool X86::isPSHUFHWMask(SDNode *N) {
2088 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2090 if (N->getNumOperands() != 8)
2093 // Lower quadword copied in order.
2094 for (unsigned i = 0; i != 4; ++i) {
2095 SDValue Arg = N->getOperand(i);
2096 if (Arg.getOpcode() == ISD::UNDEF) continue;
2097 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2098 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2102 // Upper quadword shuffled.
2103 for (unsigned i = 4; i != 8; ++i) {
2104 SDValue Arg = N->getOperand(i);
2105 if (Arg.getOpcode() == ISD::UNDEF) continue;
2106 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2107 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2108 if (Val < 4 || Val > 7)
2115 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2116 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2117 bool X86::isPSHUFLWMask(SDNode *N) {
2118 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2120 if (N->getNumOperands() != 8)
2123 // Upper quadword copied in order.
2124 for (unsigned i = 4; i != 8; ++i)
2125 if (!isUndefOrEqual(N->getOperand(i), i))
2128 // Lower quadword shuffled.
2129 for (unsigned i = 0; i != 4; ++i)
2130 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2136 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2137 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2138 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2139 if (NumElems != 2 && NumElems != 4) return false;
2141 unsigned Half = NumElems / 2;
2142 for (unsigned i = 0; i < Half; ++i)
2143 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2145 for (unsigned i = Half; i < NumElems; ++i)
2146 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2152 bool X86::isSHUFPMask(SDNode *N) {
2153 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2154 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2157 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2158 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2159 /// half elements to come from vector 1 (which would equal the dest.) and
2160 /// the upper half to come from vector 2.
2161 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2162 if (NumOps != 2 && NumOps != 4) return false;
2164 unsigned Half = NumOps / 2;
2165 for (unsigned i = 0; i < Half; ++i)
2166 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2168 for (unsigned i = Half; i < NumOps; ++i)
2169 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2174 static bool isCommutedSHUFP(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2176 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2179 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2180 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2181 bool X86::isMOVHLPSMask(SDNode *N) {
2182 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2184 if (N->getNumOperands() != 4)
2187 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2188 return isUndefOrEqual(N->getOperand(0), 6) &&
2189 isUndefOrEqual(N->getOperand(1), 7) &&
2190 isUndefOrEqual(N->getOperand(2), 2) &&
2191 isUndefOrEqual(N->getOperand(3), 3);
2194 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2195 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2197 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2198 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2200 if (N->getNumOperands() != 4)
2203 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2204 return isUndefOrEqual(N->getOperand(0), 2) &&
2205 isUndefOrEqual(N->getOperand(1), 3) &&
2206 isUndefOrEqual(N->getOperand(2), 2) &&
2207 isUndefOrEqual(N->getOperand(3), 3);
2210 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2211 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2212 bool X86::isMOVLPMask(SDNode *N) {
2213 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2215 unsigned NumElems = N->getNumOperands();
2216 if (NumElems != 2 && NumElems != 4)
2219 for (unsigned i = 0; i < NumElems/2; ++i)
2220 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2223 for (unsigned i = NumElems/2; i < NumElems; ++i)
2224 if (!isUndefOrEqual(N->getOperand(i), i))
2230 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2231 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2233 bool X86::isMOVHPMask(SDNode *N) {
2234 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2236 unsigned NumElems = N->getNumOperands();
2237 if (NumElems != 2 && NumElems != 4)
2240 for (unsigned i = 0; i < NumElems/2; ++i)
2241 if (!isUndefOrEqual(N->getOperand(i), i))
2244 for (unsigned i = 0; i < NumElems/2; ++i) {
2245 SDValue Arg = N->getOperand(i + NumElems/2);
2246 if (!isUndefOrEqual(Arg, i + NumElems))
2253 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2254 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2255 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2256 bool V2IsSplat = false) {
2257 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2260 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2261 SDValue BitI = Elts[i];
2262 SDValue BitI1 = Elts[i+1];
2263 if (!isUndefOrEqual(BitI, j))
2266 if (isUndefOrEqual(BitI1, NumElts))
2269 if (!isUndefOrEqual(BitI1, j + NumElts))
2277 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2278 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2279 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2282 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2283 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2284 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2285 bool V2IsSplat = false) {
2286 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2289 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2290 SDValue BitI = Elts[i];
2291 SDValue BitI1 = Elts[i+1];
2292 if (!isUndefOrEqual(BitI, j + NumElts/2))
2295 if (isUndefOrEqual(BitI1, NumElts))
2298 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2306 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2307 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2308 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2311 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2312 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2314 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2315 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2317 unsigned NumElems = N->getNumOperands();
2318 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2321 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2322 SDValue BitI = N->getOperand(i);
2323 SDValue BitI1 = N->getOperand(i+1);
2325 if (!isUndefOrEqual(BitI, j))
2327 if (!isUndefOrEqual(BitI1, j))
2334 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2335 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2337 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2338 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2340 unsigned NumElems = N->getNumOperands();
2341 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2344 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2345 SDValue BitI = N->getOperand(i);
2346 SDValue BitI1 = N->getOperand(i + 1);
2348 if (!isUndefOrEqual(BitI, j))
2350 if (!isUndefOrEqual(BitI1, j))
2357 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2358 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2359 /// MOVSD, and MOVD, i.e. setting the lowest element.
2360 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2361 if (NumElts != 2 && NumElts != 4)
2364 if (!isUndefOrEqual(Elts[0], NumElts))
2367 for (unsigned i = 1; i < NumElts; ++i) {
2368 if (!isUndefOrEqual(Elts[i], i))
2375 bool X86::isMOVLMask(SDNode *N) {
2376 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2377 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2380 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2381 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2382 /// element of vector 2 and the other elements to come from vector 1 in order.
2383 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2384 bool V2IsSplat = false,
2385 bool V2IsUndef = false) {
2386 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2389 if (!isUndefOrEqual(Ops[0], 0))
2392 for (unsigned i = 1; i < NumOps; ++i) {
2393 SDValue Arg = Ops[i];
2394 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2395 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2396 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2403 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2404 bool V2IsUndef = false) {
2405 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2406 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2407 V2IsSplat, V2IsUndef);
2410 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2411 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2412 bool X86::isMOVSHDUPMask(SDNode *N) {
2413 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2415 if (N->getNumOperands() != 4)
2418 // Expect 1, 1, 3, 3
2419 for (unsigned i = 0; i < 2; ++i) {
2420 SDValue Arg = N->getOperand(i);
2421 if (Arg.getOpcode() == ISD::UNDEF) continue;
2422 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2423 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2424 if (Val != 1) return false;
2428 for (unsigned i = 2; i < 4; ++i) {
2429 SDValue Arg = N->getOperand(i);
2430 if (Arg.getOpcode() == ISD::UNDEF) continue;
2431 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2432 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2433 if (Val != 3) return false;
2437 // Don't use movshdup if it can be done with a shufps.
2441 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2442 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2443 bool X86::isMOVSLDUPMask(SDNode *N) {
2444 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2446 if (N->getNumOperands() != 4)
2449 // Expect 0, 0, 2, 2
2450 for (unsigned i = 0; i < 2; ++i) {
2451 SDValue Arg = N->getOperand(i);
2452 if (Arg.getOpcode() == ISD::UNDEF) continue;
2453 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2454 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2455 if (Val != 0) return false;
2459 for (unsigned i = 2; i < 4; ++i) {
2460 SDValue Arg = N->getOperand(i);
2461 if (Arg.getOpcode() == ISD::UNDEF) continue;
2462 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2463 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2464 if (Val != 2) return false;
2468 // Don't use movshdup if it can be done with a shufps.
2472 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2473 /// specifies a identity operation on the LHS or RHS.
2474 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2475 unsigned NumElems = N->getNumOperands();
2476 for (unsigned i = 0; i < NumElems; ++i)
2477 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2482 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2483 /// a splat of a single element.
2484 static bool isSplatMask(SDNode *N) {
2485 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2487 // This is a splat operation if each element of the permute is the same, and
2488 // if the value doesn't reference the second vector.
2489 unsigned NumElems = N->getNumOperands();
2490 SDValue ElementBase;
2492 for (; i != NumElems; ++i) {
2493 SDValue Elt = N->getOperand(i);
2494 if (isa<ConstantSDNode>(Elt)) {
2500 if (!ElementBase.getNode())
2503 for (; i != NumElems; ++i) {
2504 SDValue Arg = N->getOperand(i);
2505 if (Arg.getOpcode() == ISD::UNDEF) continue;
2506 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2507 if (Arg != ElementBase) return false;
2510 // Make sure it is a splat of the first vector operand.
2511 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2514 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2515 /// a splat of a single element and it's a 2 or 4 element mask.
2516 bool X86::isSplatMask(SDNode *N) {
2517 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2519 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2520 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2522 return ::isSplatMask(N);
2525 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2526 /// specifies a splat of zero element.
2527 bool X86::isSplatLoMask(SDNode *N) {
2528 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2530 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2531 if (!isUndefOrEqual(N->getOperand(i), 0))
2536 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2537 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2539 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2540 unsigned NumOperands = N->getNumOperands();
2541 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2543 for (unsigned i = 0; i < NumOperands; ++i) {
2545 SDValue Arg = N->getOperand(NumOperands-i-1);
2546 if (Arg.getOpcode() != ISD::UNDEF)
2547 Val = cast<ConstantSDNode>(Arg)->getValue();
2548 if (Val >= NumOperands) Val -= NumOperands;
2550 if (i != NumOperands - 1)
2557 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2558 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2560 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2562 // 8 nodes, but we only care about the last 4.
2563 for (unsigned i = 7; i >= 4; --i) {
2565 SDValue Arg = N->getOperand(i);
2566 if (Arg.getOpcode() != ISD::UNDEF)
2567 Val = cast<ConstantSDNode>(Arg)->getValue();
2576 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2577 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2579 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2581 // 8 nodes, but we only care about the first 4.
2582 for (int i = 3; i >= 0; --i) {
2584 SDValue Arg = N->getOperand(i);
2585 if (Arg.getOpcode() != ISD::UNDEF)
2586 Val = cast<ConstantSDNode>(Arg)->getValue();
2595 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2596 /// specifies a 8 element shuffle that can be broken into a pair of
2597 /// PSHUFHW and PSHUFLW.
2598 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2599 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2601 if (N->getNumOperands() != 8)
2604 // Lower quadword shuffled.
2605 for (unsigned i = 0; i != 4; ++i) {
2606 SDValue Arg = N->getOperand(i);
2607 if (Arg.getOpcode() == ISD::UNDEF) continue;
2608 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2609 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2614 // Upper quadword shuffled.
2615 for (unsigned i = 4; i != 8; ++i) {
2616 SDValue Arg = N->getOperand(i);
2617 if (Arg.getOpcode() == ISD::UNDEF) continue;
2618 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2619 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2620 if (Val < 4 || Val > 7)
2627 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2628 /// values in ther permute mask.
2629 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2630 SDValue &V2, SDValue &Mask,
2631 SelectionDAG &DAG) {
2632 MVT VT = Op.getValueType();
2633 MVT MaskVT = Mask.getValueType();
2634 MVT EltVT = MaskVT.getVectorElementType();
2635 unsigned NumElems = Mask.getNumOperands();
2636 SmallVector<SDValue, 8> MaskVec;
2638 for (unsigned i = 0; i != NumElems; ++i) {
2639 SDValue Arg = Mask.getOperand(i);
2640 if (Arg.getOpcode() == ISD::UNDEF) {
2641 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2644 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2645 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2647 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2649 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2653 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2654 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2657 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2658 /// the two vector operands have swapped position.
2660 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2661 MVT MaskVT = Mask.getValueType();
2662 MVT EltVT = MaskVT.getVectorElementType();
2663 unsigned NumElems = Mask.getNumOperands();
2664 SmallVector<SDValue, 8> MaskVec;
2665 for (unsigned i = 0; i != NumElems; ++i) {
2666 SDValue Arg = Mask.getOperand(i);
2667 if (Arg.getOpcode() == ISD::UNDEF) {
2668 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2671 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2672 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2674 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2676 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2678 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2682 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2683 /// match movhlps. The lower half elements should come from upper half of
2684 /// V1 (and in order), and the upper half elements should come from the upper
2685 /// half of V2 (and in order).
2686 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2687 unsigned NumElems = Mask->getNumOperands();
2690 for (unsigned i = 0, e = 2; i != e; ++i)
2691 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2693 for (unsigned i = 2; i != 4; ++i)
2694 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2699 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2700 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2702 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2703 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2704 N = N->getOperand(0).getNode();
2705 if (ISD::isNON_EXTLoad(N)) {
2707 *LD = cast<LoadSDNode>(N);
2714 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2715 /// match movlp{s|d}. The lower half elements should come from lower half of
2716 /// V1 (and in order), and the upper half elements should come from the upper
2717 /// half of V2 (and in order). And since V1 will become the source of the
2718 /// MOVLP, it must be either a vector load or a scalar load to vector.
2719 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2720 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2722 // Is V2 is a vector load, don't do this transformation. We will try to use
2723 // load folding shufps op.
2724 if (ISD::isNON_EXTLoad(V2))
2727 unsigned NumElems = Mask->getNumOperands();
2728 if (NumElems != 2 && NumElems != 4)
2730 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2731 if (!isUndefOrEqual(Mask->getOperand(i), i))
2733 for (unsigned i = NumElems/2; i != NumElems; ++i)
2734 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2739 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2741 static bool isSplatVector(SDNode *N) {
2742 if (N->getOpcode() != ISD::BUILD_VECTOR)
2745 SDValue SplatValue = N->getOperand(0);
2746 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2747 if (N->getOperand(i) != SplatValue)
2752 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2754 static bool isUndefShuffle(SDNode *N) {
2755 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2758 SDValue V1 = N->getOperand(0);
2759 SDValue V2 = N->getOperand(1);
2760 SDValue Mask = N->getOperand(2);
2761 unsigned NumElems = Mask.getNumOperands();
2762 for (unsigned i = 0; i != NumElems; ++i) {
2763 SDValue Arg = Mask.getOperand(i);
2764 if (Arg.getOpcode() != ISD::UNDEF) {
2765 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2766 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2768 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2775 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2777 static inline bool isZeroNode(SDValue Elt) {
2778 return ((isa<ConstantSDNode>(Elt) &&
2779 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2780 (isa<ConstantFPSDNode>(Elt) &&
2781 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2784 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2785 /// to an zero vector.
2786 static bool isZeroShuffle(SDNode *N) {
2787 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2790 SDValue V1 = N->getOperand(0);
2791 SDValue V2 = N->getOperand(1);
2792 SDValue Mask = N->getOperand(2);
2793 unsigned NumElems = Mask.getNumOperands();
2794 for (unsigned i = 0; i != NumElems; ++i) {
2795 SDValue Arg = Mask.getOperand(i);
2796 if (Arg.getOpcode() == ISD::UNDEF)
2799 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2800 if (Idx < NumElems) {
2801 unsigned Opc = V1.getNode()->getOpcode();
2802 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2804 if (Opc != ISD::BUILD_VECTOR ||
2805 !isZeroNode(V1.getNode()->getOperand(Idx)))
2807 } else if (Idx >= NumElems) {
2808 unsigned Opc = V2.getNode()->getOpcode();
2809 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2811 if (Opc != ISD::BUILD_VECTOR ||
2812 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2819 /// getZeroVector - Returns a vector of specified type with all zero elements.
2821 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2822 assert(VT.isVector() && "Expected a vector type");
2824 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2825 // type. This ensures they get CSE'd.
2827 if (VT.getSizeInBits() == 64) { // MMX
2828 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2829 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2830 } else if (HasSSE2) { // SSE2
2831 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2832 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2834 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2835 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2837 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2840 /// getOnesVector - Returns a vector of specified type with all bits set.
2842 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2843 assert(VT.isVector() && "Expected a vector type");
2845 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2846 // type. This ensures they get CSE'd.
2847 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2849 if (VT.getSizeInBits() == 64) // MMX
2850 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2852 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2853 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2857 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2858 /// that point to V2 points to its first element.
2859 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2860 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2862 bool Changed = false;
2863 SmallVector<SDValue, 8> MaskVec;
2864 unsigned NumElems = Mask.getNumOperands();
2865 for (unsigned i = 0; i != NumElems; ++i) {
2866 SDValue Arg = Mask.getOperand(i);
2867 if (Arg.getOpcode() != ISD::UNDEF) {
2868 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2869 if (Val > NumElems) {
2870 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2874 MaskVec.push_back(Arg);
2878 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2879 &MaskVec[0], MaskVec.size());
2883 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2884 /// operation of specified width.
2885 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2886 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2887 MVT BaseVT = MaskVT.getVectorElementType();
2889 SmallVector<SDValue, 8> MaskVec;
2890 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2891 for (unsigned i = 1; i != NumElems; ++i)
2892 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2893 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2896 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2897 /// of specified width.
2898 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2899 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2900 MVT BaseVT = MaskVT.getVectorElementType();
2901 SmallVector<SDValue, 8> MaskVec;
2902 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2903 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2904 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2906 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2909 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2910 /// of specified width.
2911 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2912 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2913 MVT BaseVT = MaskVT.getVectorElementType();
2914 unsigned Half = NumElems/2;
2915 SmallVector<SDValue, 8> MaskVec;
2916 for (unsigned i = 0; i != Half; ++i) {
2917 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2918 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2920 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2923 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2924 /// element #0 of a vector with the specified index, leaving the rest of the
2925 /// elements in place.
2926 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2927 SelectionDAG &DAG) {
2928 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2929 MVT BaseVT = MaskVT.getVectorElementType();
2930 SmallVector<SDValue, 8> MaskVec;
2931 // Element #0 of the result gets the elt we are replacing.
2932 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2933 for (unsigned i = 1; i != NumElems; ++i)
2934 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2935 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2938 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2939 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2940 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2941 MVT VT = Op.getValueType();
2944 SDValue V1 = Op.getOperand(0);
2945 SDValue Mask = Op.getOperand(2);
2946 unsigned NumElems = Mask.getNumOperands();
2947 // Special handling of v4f32 -> v4i32.
2948 if (VT != MVT::v4f32) {
2949 Mask = getUnpacklMask(NumElems, DAG);
2950 while (NumElems > 4) {
2951 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2954 Mask = getZeroVector(MVT::v4i32, true, DAG);
2957 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2958 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2959 DAG.getNode(ISD::UNDEF, PVT), Mask);
2960 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2963 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2964 /// vector of zero or undef vector. This produces a shuffle where the low
2965 /// element of V2 is swizzled into the zero/undef vector, landing at element
2966 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2967 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2968 bool isZero, bool HasSSE2,
2969 SelectionDAG &DAG) {
2970 MVT VT = V2.getValueType();
2972 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
2973 unsigned NumElems = V2.getValueType().getVectorNumElements();
2974 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2975 MVT EVT = MaskVT.getVectorElementType();
2976 SmallVector<SDValue, 16> MaskVec;
2977 for (unsigned i = 0; i != NumElems; ++i)
2978 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2979 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2981 MaskVec.push_back(DAG.getConstant(i, EVT));
2982 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2983 &MaskVec[0], MaskVec.size());
2984 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2987 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2988 /// a shuffle that is zero.
2990 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
2991 unsigned NumElems, bool Low,
2992 SelectionDAG &DAG) {
2993 unsigned NumZeros = 0;
2994 for (unsigned i = 0; i < NumElems; ++i) {
2995 unsigned Index = Low ? i : NumElems-i-1;
2996 SDValue Idx = Mask.getOperand(Index);
2997 if (Idx.getOpcode() == ISD::UNDEF) {
3001 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3002 if (Elt.getNode() && isZeroNode(Elt))
3010 /// isVectorShift - Returns true if the shuffle can be implemented as a
3011 /// logical left or right shift of a vector.
3012 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3013 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3014 unsigned NumElems = Mask.getNumOperands();
3017 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3020 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3025 bool SeenV1 = false;
3026 bool SeenV2 = false;
3027 for (unsigned i = NumZeros; i < NumElems; ++i) {
3028 unsigned Val = isLeft ? (i - NumZeros) : i;
3029 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3030 if (Idx.getOpcode() == ISD::UNDEF)
3032 unsigned Index = cast<ConstantSDNode>(Idx)->getValue();
3033 if (Index < NumElems)
3042 if (SeenV1 && SeenV2)
3045 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3051 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3053 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3054 unsigned NumNonZero, unsigned NumZero,
3055 SelectionDAG &DAG, TargetLowering &TLI) {
3061 for (unsigned i = 0; i < 16; ++i) {
3062 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3063 if (ThisIsNonZero && First) {
3065 V = getZeroVector(MVT::v8i16, true, DAG);
3067 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3072 SDValue ThisElt(0, 0), LastElt(0, 0);
3073 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3074 if (LastIsNonZero) {
3075 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3077 if (ThisIsNonZero) {
3078 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3079 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3080 ThisElt, DAG.getConstant(8, MVT::i8));
3082 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3086 if (ThisElt.getNode())
3087 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3088 DAG.getIntPtrConstant(i/2));
3092 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3095 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3097 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3098 unsigned NumNonZero, unsigned NumZero,
3099 SelectionDAG &DAG, TargetLowering &TLI) {
3105 for (unsigned i = 0; i < 8; ++i) {
3106 bool isNonZero = (NonZeros & (1 << i)) != 0;
3110 V = getZeroVector(MVT::v8i16, true, DAG);
3112 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3115 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3116 DAG.getIntPtrConstant(i));
3123 /// getVShift - Return a vector logical shift node.
3125 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3126 unsigned NumBits, SelectionDAG &DAG,
3127 const TargetLowering &TLI) {
3128 bool isMMX = VT.getSizeInBits() == 64;
3129 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3130 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3131 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3132 return DAG.getNode(ISD::BIT_CONVERT, VT,
3133 DAG.getNode(Opc, ShVT, SrcOp,
3134 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3138 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3139 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3140 if (ISD::isBuildVectorAllZeros(Op.getNode())
3141 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3142 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3143 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3144 // eliminated on x86-32 hosts.
3145 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3148 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3149 return getOnesVector(Op.getValueType(), DAG);
3150 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3153 MVT VT = Op.getValueType();
3154 MVT EVT = VT.getVectorElementType();
3155 unsigned EVTBits = EVT.getSizeInBits();
3157 unsigned NumElems = Op.getNumOperands();
3158 unsigned NumZero = 0;
3159 unsigned NumNonZero = 0;
3160 unsigned NonZeros = 0;
3161 bool IsAllConstants = true;
3162 SmallSet<SDValue, 8> Values;
3163 for (unsigned i = 0; i < NumElems; ++i) {
3164 SDValue Elt = Op.getOperand(i);
3165 if (Elt.getOpcode() == ISD::UNDEF)
3168 if (Elt.getOpcode() != ISD::Constant &&
3169 Elt.getOpcode() != ISD::ConstantFP)
3170 IsAllConstants = false;
3171 if (isZeroNode(Elt))
3174 NonZeros |= (1 << i);
3179 if (NumNonZero == 0) {
3180 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3181 return DAG.getNode(ISD::UNDEF, VT);
3184 // Special case for single non-zero, non-undef, element.
3185 if (NumNonZero == 1 && NumElems <= 4) {
3186 unsigned Idx = CountTrailingZeros_32(NonZeros);
3187 SDValue Item = Op.getOperand(Idx);
3189 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3190 // the value are obviously zero, truncate the value to i32 and do the
3191 // insertion that way. Only do this if the value is non-constant or if the
3192 // value is a constant being inserted into element 0. It is cheaper to do
3193 // a constant pool load than it is to do a movd + shuffle.
3194 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3195 (!IsAllConstants || Idx == 0)) {
3196 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3197 // Handle MMX and SSE both.
3198 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3199 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3201 // Truncate the value (which may itself be a constant) to i32, and
3202 // convert it to a vector with movd (S2V+shuffle to zero extend).
3203 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3204 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3205 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3206 Subtarget->hasSSE2(), DAG);
3208 // Now we have our 32-bit value zero extended in the low element of
3209 // a vector. If Idx != 0, swizzle it into place.
3212 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3213 getSwapEltZeroMask(VecElts, Idx, DAG)
3215 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3217 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3221 // If we have a constant or non-constant insertion into the low element of
3222 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3223 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3224 // depending on what the source datatype is. Because we can only get here
3225 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3227 // Don't do this for i64 values on x86-32.
3228 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3229 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3230 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3231 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3232 Subtarget->hasSSE2(), DAG);
3235 // Is it a vector logical left shift?
3236 if (NumElems == 2 && Idx == 1 &&
3237 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3238 unsigned NumBits = VT.getSizeInBits();
3239 return getVShift(true, VT,
3240 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3241 NumBits/2, DAG, *this);
3244 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3247 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3248 // is a non-constant being inserted into an element other than the low one,
3249 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3250 // movd/movss) to move this into the low element, then shuffle it into
3252 if (EVTBits == 32) {
3253 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3255 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3256 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3257 Subtarget->hasSSE2(), DAG);
3258 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3259 MVT MaskEVT = MaskVT.getVectorElementType();
3260 SmallVector<SDValue, 8> MaskVec;
3261 for (unsigned i = 0; i < NumElems; i++)
3262 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3263 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3264 &MaskVec[0], MaskVec.size());
3265 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3266 DAG.getNode(ISD::UNDEF, VT), Mask);
3270 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3271 if (Values.size() == 1)
3274 // A vector full of immediates; various special cases are already
3275 // handled, so this is best done with a single constant-pool load.
3279 // Let legalizer expand 2-wide build_vectors.
3280 if (EVTBits == 64) {
3281 if (NumNonZero == 1) {
3282 // One half is zero or undef.
3283 unsigned Idx = CountTrailingZeros_32(NonZeros);
3284 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3285 Op.getOperand(Idx));
3286 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3287 Subtarget->hasSSE2(), DAG);
3292 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3293 if (EVTBits == 8 && NumElems == 16) {
3294 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3296 if (V.getNode()) return V;
3299 if (EVTBits == 16 && NumElems == 8) {
3300 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3302 if (V.getNode()) return V;
3305 // If element VT is == 32 bits, turn it into a number of shuffles.
3306 SmallVector<SDValue, 8> V;
3308 if (NumElems == 4 && NumZero > 0) {
3309 for (unsigned i = 0; i < 4; ++i) {
3310 bool isZero = !(NonZeros & (1 << i));
3312 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3314 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3317 for (unsigned i = 0; i < 2; ++i) {
3318 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3321 V[i] = V[i*2]; // Must be a zero vector.
3324 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3325 getMOVLMask(NumElems, DAG));
3328 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3329 getMOVLMask(NumElems, DAG));
3332 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3333 getUnpacklMask(NumElems, DAG));
3338 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3339 MVT EVT = MaskVT.getVectorElementType();
3340 SmallVector<SDValue, 8> MaskVec;
3341 bool Reverse = (NonZeros & 0x3) == 2;
3342 for (unsigned i = 0; i < 2; ++i)
3344 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3346 MaskVec.push_back(DAG.getConstant(i, EVT));
3347 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3348 for (unsigned i = 0; i < 2; ++i)
3350 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3352 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3353 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3354 &MaskVec[0], MaskVec.size());
3355 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3358 if (Values.size() > 2) {
3359 // Expand into a number of unpckl*.
3361 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3362 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3363 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3364 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3365 for (unsigned i = 0; i < NumElems; ++i)
3366 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3368 while (NumElems != 0) {
3369 for (unsigned i = 0; i < NumElems; ++i)
3370 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3381 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3382 SDValue PermMask, SelectionDAG &DAG,
3383 TargetLowering &TLI) {
3385 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3386 MVT MaskEVT = MaskVT.getVectorElementType();
3387 MVT PtrVT = TLI.getPointerTy();
3388 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3389 PermMask.getNode()->op_end());
3391 // First record which half of which vector the low elements come from.
3392 SmallVector<unsigned, 4> LowQuad(4);
3393 for (unsigned i = 0; i < 4; ++i) {
3394 SDValue Elt = MaskElts[i];
3395 if (Elt.getOpcode() == ISD::UNDEF)
3397 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3398 int QuadIdx = EltIdx / 4;
3402 int BestLowQuad = -1;
3403 unsigned MaxQuad = 1;
3404 for (unsigned i = 0; i < 4; ++i) {
3405 if (LowQuad[i] > MaxQuad) {
3407 MaxQuad = LowQuad[i];
3411 // Record which half of which vector the high elements come from.
3412 SmallVector<unsigned, 4> HighQuad(4);
3413 for (unsigned i = 4; i < 8; ++i) {
3414 SDValue Elt = MaskElts[i];
3415 if (Elt.getOpcode() == ISD::UNDEF)
3417 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3418 int QuadIdx = EltIdx / 4;
3419 ++HighQuad[QuadIdx];
3422 int BestHighQuad = -1;
3424 for (unsigned i = 0; i < 4; ++i) {
3425 if (HighQuad[i] > MaxQuad) {
3427 MaxQuad = HighQuad[i];
3431 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3432 if (BestLowQuad != -1 || BestHighQuad != -1) {
3433 // First sort the 4 chunks in order using shufpd.
3434 SmallVector<SDValue, 8> MaskVec;
3436 if (BestLowQuad != -1)
3437 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3439 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3441 if (BestHighQuad != -1)
3442 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3444 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3446 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3447 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3448 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3449 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3450 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3452 // Now sort high and low parts separately.
3453 BitVector InOrder(8);
3454 if (BestLowQuad != -1) {
3455 // Sort lower half in order using PSHUFLW.
3457 bool AnyOutOrder = false;
3459 for (unsigned i = 0; i != 4; ++i) {
3460 SDValue Elt = MaskElts[i];
3461 if (Elt.getOpcode() == ISD::UNDEF) {
3462 MaskVec.push_back(Elt);
3465 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3469 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3471 // If this element is in the right place after this shuffle, then
3473 if ((int)(EltIdx / 4) == BestLowQuad)
3478 for (unsigned i = 4; i != 8; ++i)
3479 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3480 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3481 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3485 if (BestHighQuad != -1) {
3486 // Sort high half in order using PSHUFHW if possible.
3489 for (unsigned i = 0; i != 4; ++i)
3490 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3492 bool AnyOutOrder = false;
3493 for (unsigned i = 4; i != 8; ++i) {
3494 SDValue Elt = MaskElts[i];
3495 if (Elt.getOpcode() == ISD::UNDEF) {
3496 MaskVec.push_back(Elt);
3499 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3503 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3505 // If this element is in the right place after this shuffle, then
3507 if ((int)(EltIdx / 4) == BestHighQuad)
3513 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3514 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3518 // The other elements are put in the right place using pextrw and pinsrw.
3519 for (unsigned i = 0; i != 8; ++i) {
3522 SDValue Elt = MaskElts[i];
3523 if (Elt.getOpcode() == ISD::UNDEF)
3525 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3526 SDValue ExtOp = (EltIdx < 8)
3527 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3528 DAG.getConstant(EltIdx, PtrVT))
3529 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3530 DAG.getConstant(EltIdx - 8, PtrVT));
3531 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3532 DAG.getConstant(i, PtrVT));
3538 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3539 // few as possible. First, let's find out how many elements are already in the
3541 unsigned V1InOrder = 0;
3542 unsigned V1FromV1 = 0;
3543 unsigned V2InOrder = 0;
3544 unsigned V2FromV2 = 0;
3545 SmallVector<SDValue, 8> V1Elts;
3546 SmallVector<SDValue, 8> V2Elts;
3547 for (unsigned i = 0; i < 8; ++i) {
3548 SDValue Elt = MaskElts[i];
3549 if (Elt.getOpcode() == ISD::UNDEF) {
3550 V1Elts.push_back(Elt);
3551 V2Elts.push_back(Elt);
3556 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3558 V1Elts.push_back(Elt);
3559 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3561 } else if (EltIdx == i+8) {
3562 V1Elts.push_back(Elt);
3563 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3565 } else if (EltIdx < 8) {
3566 V1Elts.push_back(Elt);
3569 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3574 if (V2InOrder > V1InOrder) {
3575 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3577 std::swap(V1Elts, V2Elts);
3578 std::swap(V1FromV1, V2FromV2);
3581 if ((V1FromV1 + V1InOrder) != 8) {
3582 // Some elements are from V2.
3584 // If there are elements that are from V1 but out of place,
3585 // then first sort them in place
3586 SmallVector<SDValue, 8> MaskVec;
3587 for (unsigned i = 0; i < 8; ++i) {
3588 SDValue Elt = V1Elts[i];
3589 if (Elt.getOpcode() == ISD::UNDEF) {
3590 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3593 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3595 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3597 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3599 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3600 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3604 for (unsigned i = 0; i < 8; ++i) {
3605 SDValue Elt = V1Elts[i];
3606 if (Elt.getOpcode() == ISD::UNDEF)
3608 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3611 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3612 DAG.getConstant(EltIdx - 8, PtrVT));
3613 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3614 DAG.getConstant(i, PtrVT));
3618 // All elements are from V1.
3620 for (unsigned i = 0; i < 8; ++i) {
3621 SDValue Elt = V1Elts[i];
3622 if (Elt.getOpcode() == ISD::UNDEF)
3624 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3625 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3626 DAG.getConstant(EltIdx, PtrVT));
3627 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3628 DAG.getConstant(i, PtrVT));
3634 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3635 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3636 /// done when every pair / quad of shuffle mask elements point to elements in
3637 /// the right sequence. e.g.
3638 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3640 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3642 SDValue PermMask, SelectionDAG &DAG,
3643 TargetLowering &TLI) {
3644 unsigned NumElems = PermMask.getNumOperands();
3645 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3646 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3647 MVT MaskEltVT = MaskVT.getVectorElementType();
3649 switch (VT.getSimpleVT()) {
3650 default: assert(false && "Unexpected!");
3651 case MVT::v4f32: NewVT = MVT::v2f64; break;
3652 case MVT::v4i32: NewVT = MVT::v2i64; break;
3653 case MVT::v8i16: NewVT = MVT::v4i32; break;
3654 case MVT::v16i8: NewVT = MVT::v4i32; break;
3657 if (NewWidth == 2) {
3663 unsigned Scale = NumElems / NewWidth;
3664 SmallVector<SDValue, 8> MaskVec;
3665 for (unsigned i = 0; i < NumElems; i += Scale) {
3666 unsigned StartIdx = ~0U;
3667 for (unsigned j = 0; j < Scale; ++j) {
3668 SDValue Elt = PermMask.getOperand(i+j);
3669 if (Elt.getOpcode() == ISD::UNDEF)
3671 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3672 if (StartIdx == ~0U)
3673 StartIdx = EltIdx - (EltIdx % Scale);
3674 if (EltIdx != StartIdx + j)
3677 if (StartIdx == ~0U)
3678 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3680 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3683 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3684 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3685 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3686 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3687 &MaskVec[0], MaskVec.size()));
3690 /// getVZextMovL - Return a zero-extending vector move low node.
3692 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3693 SDValue SrcOp, SelectionDAG &DAG,
3694 const X86Subtarget *Subtarget) {
3695 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3696 LoadSDNode *LD = NULL;
3697 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3698 LD = dyn_cast<LoadSDNode>(SrcOp);
3700 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3702 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3703 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3704 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3705 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3706 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3708 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3709 return DAG.getNode(ISD::BIT_CONVERT, VT,
3710 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3711 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3718 return DAG.getNode(ISD::BIT_CONVERT, VT,
3719 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3720 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3723 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3726 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3727 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3728 MVT MaskVT = PermMask.getValueType();
3729 MVT MaskEVT = MaskVT.getVectorElementType();
3730 SmallVector<std::pair<int, int>, 8> Locs;
3732 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3735 for (unsigned i = 0; i != 4; ++i) {
3736 SDValue Elt = PermMask.getOperand(i);
3737 if (Elt.getOpcode() == ISD::UNDEF) {
3738 Locs[i] = std::make_pair(-1, -1);
3740 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3741 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3743 Locs[i] = std::make_pair(0, NumLo);
3747 Locs[i] = std::make_pair(1, NumHi);
3749 Mask1[2+NumHi] = Elt;
3755 if (NumLo <= 2 && NumHi <= 2) {
3756 // If no more than two elements come from either vector. This can be
3757 // implemented with two shuffles. First shuffle gather the elements.
3758 // The second shuffle, which takes the first shuffle as both of its
3759 // vector operands, put the elements into the right order.
3760 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3761 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3762 &Mask1[0], Mask1.size()));
3764 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3765 for (unsigned i = 0; i != 4; ++i) {
3766 if (Locs[i].first == -1)
3769 unsigned Idx = (i < 2) ? 0 : 4;
3770 Idx += Locs[i].first * 2 + Locs[i].second;
3771 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3775 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3776 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3777 &Mask2[0], Mask2.size()));
3778 } else if (NumLo == 3 || NumHi == 3) {
3779 // Otherwise, we must have three elements from one vector, call it X, and
3780 // one element from the other, call it Y. First, use a shufps to build an
3781 // intermediate vector with the one element from Y and the element from X
3782 // that will be in the same half in the final destination (the indexes don't
3783 // matter). Then, use a shufps to build the final vector, taking the half
3784 // containing the element from Y from the intermediate, and the other half
3787 // Normalize it so the 3 elements come from V1.
3788 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3792 // Find the element from V2.
3794 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3795 SDValue Elt = PermMask.getOperand(HiIndex);
3796 if (Elt.getOpcode() == ISD::UNDEF)
3798 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3803 Mask1[0] = PermMask.getOperand(HiIndex);
3804 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3805 Mask1[2] = PermMask.getOperand(HiIndex^1);
3806 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3807 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3808 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3811 Mask1[0] = PermMask.getOperand(0);
3812 Mask1[1] = PermMask.getOperand(1);
3813 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3814 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3815 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3816 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3818 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3819 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3820 Mask1[2] = PermMask.getOperand(2);
3821 Mask1[3] = PermMask.getOperand(3);
3822 if (Mask1[2].getOpcode() != ISD::UNDEF)
3823 Mask1[2] = DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getValue()+4,
3825 if (Mask1[3].getOpcode() != ISD::UNDEF)
3826 Mask1[3] = DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getValue()+4,
3828 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3829 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3833 // Break it into (shuffle shuffle_hi, shuffle_lo).
3835 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3836 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3837 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3838 unsigned MaskIdx = 0;
3841 for (unsigned i = 0; i != 4; ++i) {
3848 SDValue Elt = PermMask.getOperand(i);
3849 if (Elt.getOpcode() == ISD::UNDEF) {
3850 Locs[i] = std::make_pair(-1, -1);
3851 } else if (cast<ConstantSDNode>(Elt)->getValue() < 4) {
3852 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3853 (*MaskPtr)[LoIdx] = Elt;
3856 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3857 (*MaskPtr)[HiIdx] = Elt;
3862 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3863 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3864 &LoMask[0], LoMask.size()));
3865 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3866 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3867 &HiMask[0], HiMask.size()));
3868 SmallVector<SDValue, 8> MaskOps;
3869 for (unsigned i = 0; i != 4; ++i) {
3870 if (Locs[i].first == -1) {
3871 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3873 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3874 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3877 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3878 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3879 &MaskOps[0], MaskOps.size()));
3883 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3884 SDValue V1 = Op.getOperand(0);
3885 SDValue V2 = Op.getOperand(1);
3886 SDValue PermMask = Op.getOperand(2);
3887 MVT VT = Op.getValueType();
3888 unsigned NumElems = PermMask.getNumOperands();
3889 bool isMMX = VT.getSizeInBits() == 64;
3890 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3891 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3892 bool V1IsSplat = false;
3893 bool V2IsSplat = false;
3895 if (isUndefShuffle(Op.getNode()))
3896 return DAG.getNode(ISD::UNDEF, VT);
3898 if (isZeroShuffle(Op.getNode()))
3899 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3901 if (isIdentityMask(PermMask.getNode()))
3903 else if (isIdentityMask(PermMask.getNode(), true))
3906 if (isSplatMask(PermMask.getNode())) {
3907 if (isMMX || NumElems < 4) return Op;
3908 // Promote it to a v4{if}32 splat.
3909 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3912 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3914 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3915 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3916 if (NewOp.getNode())
3917 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3918 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3919 // FIXME: Figure out a cleaner way to do this.
3920 // Try to make use of movq to zero out the top part.
3921 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3922 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3924 if (NewOp.getNode()) {
3925 SDValue NewV1 = NewOp.getOperand(0);
3926 SDValue NewV2 = NewOp.getOperand(1);
3927 SDValue NewMask = NewOp.getOperand(2);
3928 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3929 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3930 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3933 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3934 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3936 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3937 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3942 // Check if this can be converted into a logical shift.
3943 bool isLeft = false;
3946 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
3947 if (isShift && ShVal.hasOneUse()) {
3948 // If the shifted value has multiple uses, it may be cheaper to use
3949 // v_set0 + movlhps or movhlps, etc.
3950 MVT EVT = VT.getVectorElementType();
3951 ShAmt *= EVT.getSizeInBits();
3952 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3955 if (X86::isMOVLMask(PermMask.getNode())) {
3958 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3959 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
3964 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
3965 X86::isMOVSLDUPMask(PermMask.getNode()) ||
3966 X86::isMOVHLPSMask(PermMask.getNode()) ||
3967 X86::isMOVHPMask(PermMask.getNode()) ||
3968 X86::isMOVLPMask(PermMask.getNode())))
3971 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
3972 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
3973 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3976 // No better options. Use a vshl / vsrl.
3977 MVT EVT = VT.getVectorElementType();
3978 ShAmt *= EVT.getSizeInBits();
3979 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
3982 bool Commuted = false;
3983 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3984 // 1,1,1,1 -> v8i16 though.
3985 V1IsSplat = isSplatVector(V1.getNode());
3986 V2IsSplat = isSplatVector(V2.getNode());
3988 // Canonicalize the splat or undef, if present, to be on the RHS.
3989 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3990 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3991 std::swap(V1IsSplat, V2IsSplat);
3992 std::swap(V1IsUndef, V2IsUndef);
3996 // FIXME: Figure out a cleaner way to do this.
3997 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
3998 if (V2IsUndef) return V1;
3999 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4001 // V2 is a splat, so the mask may be malformed. That is, it may point
4002 // to any V2 element. The instruction selectior won't like this. Get
4003 // a corrected mask and commute to form a proper MOVS{S|D}.
4004 SDValue NewMask = getMOVLMask(NumElems, DAG);
4005 if (NewMask.getNode() != PermMask.getNode())
4006 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4011 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4012 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4013 X86::isUNPCKLMask(PermMask.getNode()) ||
4014 X86::isUNPCKHMask(PermMask.getNode()))
4018 // Normalize mask so all entries that point to V2 points to its first
4019 // element then try to match unpck{h|l} again. If match, return a
4020 // new vector_shuffle with the corrected mask.
4021 SDValue NewMask = NormalizeMask(PermMask, DAG);
4022 if (NewMask.getNode() != PermMask.getNode()) {
4023 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4024 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4025 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4026 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4027 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4028 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4033 // Normalize the node to match x86 shuffle ops if needed
4034 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4035 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4038 // Commute is back and try unpck* again.
4039 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4040 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4041 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4042 X86::isUNPCKLMask(PermMask.getNode()) ||
4043 X86::isUNPCKHMask(PermMask.getNode()))
4047 // Try PSHUF* first, then SHUFP*.
4048 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4049 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4050 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4051 if (V2.getOpcode() != ISD::UNDEF)
4052 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4053 DAG.getNode(ISD::UNDEF, VT), PermMask);
4058 if (Subtarget->hasSSE2() &&
4059 (X86::isPSHUFDMask(PermMask.getNode()) ||
4060 X86::isPSHUFHWMask(PermMask.getNode()) ||
4061 X86::isPSHUFLWMask(PermMask.getNode()))) {
4063 if (VT == MVT::v4f32) {
4065 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4066 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4067 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4068 } else if (V2.getOpcode() != ISD::UNDEF)
4069 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4070 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4072 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4076 // Binary or unary shufps.
4077 if (X86::isSHUFPMask(PermMask.getNode()) ||
4078 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4082 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4083 if (VT == MVT::v8i16) {
4084 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4085 if (NewOp.getNode())
4089 // Handle all 4 wide cases with a number of shuffles except for MMX.
4090 if (NumElems == 4 && !isMMX)
4091 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4097 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4098 SelectionDAG &DAG) {
4099 MVT VT = Op.getValueType();
4100 if (VT.getSizeInBits() == 8) {
4101 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4102 Op.getOperand(0), Op.getOperand(1));
4103 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4104 DAG.getValueType(VT));
4105 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4106 } else if (VT.getSizeInBits() == 16) {
4107 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4108 Op.getOperand(0), Op.getOperand(1));
4109 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4110 DAG.getValueType(VT));
4111 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4112 } else if (VT == MVT::f32) {
4113 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4114 // the result back to FR32 register. It's only worth matching if the
4115 // result has a single use which is a store or a bitcast to i32.
4116 if (!Op.hasOneUse())
4118 SDNode *User = *Op.getNode()->use_begin();
4119 if (User->getOpcode() != ISD::STORE &&
4120 (User->getOpcode() != ISD::BIT_CONVERT ||
4121 User->getValueType(0) != MVT::i32))
4123 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4124 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4126 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4133 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4134 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4137 if (Subtarget->hasSSE41()) {
4138 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4143 MVT VT = Op.getValueType();
4144 // TODO: handle v16i8.
4145 if (VT.getSizeInBits() == 16) {
4146 SDValue Vec = Op.getOperand(0);
4147 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4149 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4150 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4151 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4153 // Transform it so it match pextrw which produces a 32-bit result.
4154 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4155 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4156 Op.getOperand(0), Op.getOperand(1));
4157 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4158 DAG.getValueType(VT));
4159 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4160 } else if (VT.getSizeInBits() == 32) {
4161 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4164 // SHUFPS the element to the lowest double word, then movss.
4165 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4166 SmallVector<SDValue, 8> IdxVec;
4168 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4170 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4172 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4174 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4175 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4176 &IdxVec[0], IdxVec.size());
4177 SDValue Vec = Op.getOperand(0);
4178 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4179 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4180 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4181 DAG.getIntPtrConstant(0));
4182 } else if (VT.getSizeInBits() == 64) {
4183 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4184 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4185 // to match extract_elt for f64.
4186 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4190 // UNPCKHPD the element to the lowest double word, then movsd.
4191 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4192 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4193 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4194 SmallVector<SDValue, 8> IdxVec;
4195 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4197 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4198 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4199 &IdxVec[0], IdxVec.size());
4200 SDValue Vec = Op.getOperand(0);
4201 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4202 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4203 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4204 DAG.getIntPtrConstant(0));
4211 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4212 MVT VT = Op.getValueType();
4213 MVT EVT = VT.getVectorElementType();
4215 SDValue N0 = Op.getOperand(0);
4216 SDValue N1 = Op.getOperand(1);
4217 SDValue N2 = Op.getOperand(2);
4219 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4220 isa<ConstantSDNode>(N2)) {
4221 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4223 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4225 if (N1.getValueType() != MVT::i32)
4226 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4227 if (N2.getValueType() != MVT::i32)
4228 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4229 return DAG.getNode(Opc, VT, N0, N1, N2);
4230 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4231 // Bits [7:6] of the constant are the source select. This will always be
4232 // zero here. The DAG Combiner may combine an extract_elt index into these
4233 // bits. For example (insert (extract, 3), 2) could be matched by putting
4234 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4235 // Bits [5:4] of the constant are the destination select. This is the
4236 // value of the incoming immediate.
4237 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4238 // combine either bitwise AND or insert of float 0.0 to set these bits.
4239 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
4240 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4246 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4247 MVT VT = Op.getValueType();
4248 MVT EVT = VT.getVectorElementType();
4250 if (Subtarget->hasSSE41())
4251 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4256 SDValue N0 = Op.getOperand(0);
4257 SDValue N1 = Op.getOperand(1);
4258 SDValue N2 = Op.getOperand(2);
4260 if (EVT.getSizeInBits() == 16) {
4261 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4262 // as its second argument.
4263 if (N1.getValueType() != MVT::i32)
4264 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4265 if (N2.getValueType() != MVT::i32)
4266 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
4267 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4273 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4274 if (Op.getValueType() == MVT::v2f32)
4275 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4276 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4277 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4278 Op.getOperand(0))));
4280 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4281 MVT VT = MVT::v2i32;
4282 switch (Op.getValueType().getSimpleVT()) {
4289 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4290 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4293 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4294 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4295 // one of the above mentioned nodes. It has to be wrapped because otherwise
4296 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4297 // be used to form addressing mode. These wrapped nodes will be selected
4300 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4301 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4302 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4304 CP->getAlignment());
4305 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4306 // With PIC, the address is actually $g + Offset.
4307 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4308 !Subtarget->isPICStyleRIPRel()) {
4309 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4310 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4318 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4319 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4320 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4321 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4322 // With PIC, the address is actually $g + Offset.
4323 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4324 !Subtarget->isPICStyleRIPRel()) {
4325 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4326 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4330 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4331 // load the value at address GV, not the value of GV itself. This means that
4332 // the GlobalAddress must be in the base or index register of the address, not
4333 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4334 // The same applies for external symbols during PIC codegen
4335 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4336 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4337 PseudoSourceValue::getGOT(), 0);
4342 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4344 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4347 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4348 DAG.getNode(X86ISD::GlobalBaseReg,
4350 InFlag = Chain.getValue(1);
4352 // emit leal symbol@TLSGD(,%ebx,1), %eax
4353 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4354 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4355 GA->getValueType(0),
4357 SDValue Ops[] = { Chain, TGA, InFlag };
4358 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4359 InFlag = Result.getValue(2);
4360 Chain = Result.getValue(1);
4362 // call ___tls_get_addr. This function receives its argument in
4363 // the register EAX.
4364 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4365 InFlag = Chain.getValue(1);
4367 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4368 SDValue Ops1[] = { Chain,
4369 DAG.getTargetExternalSymbol("___tls_get_addr",
4371 DAG.getRegister(X86::EAX, PtrVT),
4372 DAG.getRegister(X86::EBX, PtrVT),
4374 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4375 InFlag = Chain.getValue(1);
4377 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4380 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4382 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4384 SDValue InFlag, Chain;
4386 // emit leaq symbol@TLSGD(%rip), %rdi
4387 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4388 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4389 GA->getValueType(0),
4391 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4392 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4393 Chain = Result.getValue(1);
4394 InFlag = Result.getValue(2);
4396 // call __tls_get_addr. This function receives its argument in
4397 // the register RDI.
4398 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4399 InFlag = Chain.getValue(1);
4401 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4402 SDValue Ops1[] = { Chain,
4403 DAG.getTargetExternalSymbol("__tls_get_addr",
4405 DAG.getRegister(X86::RDI, PtrVT),
4407 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4408 InFlag = Chain.getValue(1);
4410 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4413 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4414 // "local exec" model.
4415 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4417 // Get the Thread Pointer
4418 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4419 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4421 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4422 GA->getValueType(0),
4424 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4426 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4427 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4428 PseudoSourceValue::getGOT(), 0);
4430 // The address of the thread local variable is the add of the thread
4431 // pointer with the offset of the variable.
4432 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4436 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4437 // TODO: implement the "local dynamic" model
4438 // TODO: implement the "initial exec"model for pic executables
4439 assert(Subtarget->isTargetELF() &&
4440 "TLS not implemented for non-ELF targets");
4441 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4442 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4443 // otherwise use the "Local Exec"TLS Model
4444 if (Subtarget->is64Bit()) {
4445 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4447 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4448 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4450 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4455 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4456 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4457 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4458 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4459 // With PIC, the address is actually $g + Offset.
4460 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4461 !Subtarget->isPICStyleRIPRel()) {
4462 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4463 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4470 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4471 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4472 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4473 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4474 // With PIC, the address is actually $g + Offset.
4475 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4476 !Subtarget->isPICStyleRIPRel()) {
4477 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4478 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4485 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4486 /// take a 2 x i32 value to shift plus a shift amount.
4487 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4488 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4489 MVT VT = Op.getValueType();
4490 unsigned VTBits = VT.getSizeInBits();
4491 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4492 SDValue ShOpLo = Op.getOperand(0);
4493 SDValue ShOpHi = Op.getOperand(1);
4494 SDValue ShAmt = Op.getOperand(2);
4495 SDValue Tmp1 = isSRA ?
4496 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4497 DAG.getConstant(0, VT);
4500 if (Op.getOpcode() == ISD::SHL_PARTS) {
4501 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4502 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4504 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4505 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4508 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4509 DAG.getConstant(VTBits, MVT::i8));
4510 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4511 AndNode, DAG.getConstant(0, MVT::i8));
4514 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4515 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4516 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4518 if (Op.getOpcode() == ISD::SHL_PARTS) {
4519 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4520 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4522 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4523 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4526 SDValue Ops[2] = { Lo, Hi };
4527 return DAG.getMergeValues(Ops, 2);
4530 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4531 MVT SrcVT = Op.getOperand(0).getValueType();
4532 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4533 "Unknown SINT_TO_FP to lower!");
4535 // These are really Legal; caller falls through into that case.
4536 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4538 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4539 Subtarget->is64Bit())
4542 unsigned Size = SrcVT.getSizeInBits()/8;
4543 MachineFunction &MF = DAG.getMachineFunction();
4544 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4545 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4546 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4548 PseudoSourceValue::getFixedStack(SSFI), 0);
4552 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4554 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4556 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4557 SmallVector<SDValue, 8> Ops;
4558 Ops.push_back(Chain);
4559 Ops.push_back(StackSlot);
4560 Ops.push_back(DAG.getValueType(SrcVT));
4561 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4562 Tys, &Ops[0], Ops.size());
4565 Chain = Result.getValue(1);
4566 SDValue InFlag = Result.getValue(2);
4568 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4569 // shouldn't be necessary except that RFP cannot be live across
4570 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4571 MachineFunction &MF = DAG.getMachineFunction();
4572 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4573 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4574 Tys = DAG.getVTList(MVT::Other);
4575 SmallVector<SDValue, 8> Ops;
4576 Ops.push_back(Chain);
4577 Ops.push_back(Result);
4578 Ops.push_back(StackSlot);
4579 Ops.push_back(DAG.getValueType(Op.getValueType()));
4580 Ops.push_back(InFlag);
4581 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4582 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4583 PseudoSourceValue::getFixedStack(SSFI), 0);
4589 std::pair<SDValue,SDValue> X86TargetLowering::
4590 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4591 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4592 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4593 "Unknown FP_TO_SINT to lower!");
4595 // These are really Legal.
4596 if (Op.getValueType() == MVT::i32 &&
4597 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4598 return std::make_pair(SDValue(), SDValue());
4599 if (Subtarget->is64Bit() &&
4600 Op.getValueType() == MVT::i64 &&
4601 Op.getOperand(0).getValueType() != MVT::f80)
4602 return std::make_pair(SDValue(), SDValue());
4604 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4606 MachineFunction &MF = DAG.getMachineFunction();
4607 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4608 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4609 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4611 switch (Op.getValueType().getSimpleVT()) {
4612 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4613 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4614 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4615 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4618 SDValue Chain = DAG.getEntryNode();
4619 SDValue Value = Op.getOperand(0);
4620 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4621 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4622 Chain = DAG.getStore(Chain, Value, StackSlot,
4623 PseudoSourceValue::getFixedStack(SSFI), 0);
4624 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4626 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4628 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4629 Chain = Value.getValue(1);
4630 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4631 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4634 // Build the FP_TO_INT*_IN_MEM
4635 SDValue Ops[] = { Chain, Value, StackSlot };
4636 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4638 return std::make_pair(FIST, StackSlot);
4641 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4642 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4643 SDValue FIST = Vals.first, StackSlot = Vals.second;
4644 if (FIST.getNode() == 0) return SDValue();
4647 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4650 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4651 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4652 SDValue FIST = Vals.first, StackSlot = Vals.second;
4653 if (FIST.getNode() == 0) return 0;
4655 MVT VT = N->getValueType(0);
4657 // Return a load from the stack slot.
4658 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4660 // Use MERGE_VALUES to drop the chain result value and get a node with one
4661 // result. This requires turning off getMergeValues simplification, since
4662 // otherwise it will give us Res back.
4663 return DAG.getMergeValues(&Res, 1, false).getNode();
4666 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4667 MVT VT = Op.getValueType();
4670 EltVT = VT.getVectorElementType();
4671 std::vector<Constant*> CV;
4672 if (EltVT == MVT::f64) {
4673 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4677 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4683 Constant *C = ConstantVector::get(CV);
4684 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4685 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4686 PseudoSourceValue::getConstantPool(), 0,
4688 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4691 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4692 MVT VT = Op.getValueType();
4694 unsigned EltNum = 1;
4695 if (VT.isVector()) {
4696 EltVT = VT.getVectorElementType();
4697 EltNum = VT.getVectorNumElements();
4699 std::vector<Constant*> CV;
4700 if (EltVT == MVT::f64) {
4701 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4705 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4711 Constant *C = ConstantVector::get(CV);
4712 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4713 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4714 PseudoSourceValue::getConstantPool(), 0,
4716 if (VT.isVector()) {
4717 return DAG.getNode(ISD::BIT_CONVERT, VT,
4718 DAG.getNode(ISD::XOR, MVT::v2i64,
4719 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4720 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4722 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4726 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4727 SDValue Op0 = Op.getOperand(0);
4728 SDValue Op1 = Op.getOperand(1);
4729 MVT VT = Op.getValueType();
4730 MVT SrcVT = Op1.getValueType();
4732 // If second operand is smaller, extend it first.
4733 if (SrcVT.bitsLT(VT)) {
4734 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4737 // And if it is bigger, shrink it first.
4738 if (SrcVT.bitsGT(VT)) {
4739 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4743 // At this point the operands and the result should have the same
4744 // type, and that won't be f80 since that is not custom lowered.
4746 // First get the sign bit of second operand.
4747 std::vector<Constant*> CV;
4748 if (SrcVT == MVT::f64) {
4749 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4750 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4752 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4753 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4754 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4755 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4757 Constant *C = ConstantVector::get(CV);
4758 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4759 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4760 PseudoSourceValue::getConstantPool(), 0,
4762 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4764 // Shift sign bit right or left if the two operands have different types.
4765 if (SrcVT.bitsGT(VT)) {
4766 // Op0 is MVT::f32, Op1 is MVT::f64.
4767 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4768 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4769 DAG.getConstant(32, MVT::i32));
4770 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4771 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4772 DAG.getIntPtrConstant(0));
4775 // Clear first operand sign bit.
4777 if (VT == MVT::f64) {
4778 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4779 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4781 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4782 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4783 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4784 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4786 C = ConstantVector::get(CV);
4787 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4788 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4789 PseudoSourceValue::getConstantPool(), 0,
4791 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4793 // Or the value with the sign bit.
4794 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4797 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4798 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4800 SDValue Op0 = Op.getOperand(0);
4801 SDValue Op1 = Op.getOperand(1);
4802 SDValue CC = Op.getOperand(2);
4803 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4804 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4807 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4809 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4810 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4811 DAG.getConstant(X86CC, MVT::i8), Cond);
4814 assert(isFP && "Illegal integer SetCC!");
4816 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4817 switch (SetCCOpcode) {
4818 default: assert(false && "Illegal floating point SetCC!");
4819 case ISD::SETOEQ: { // !PF & ZF
4820 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4821 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4822 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4823 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4824 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4826 case ISD::SETUNE: { // PF | !ZF
4827 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4828 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4829 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4830 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4831 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4836 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4838 SDValue Op0 = Op.getOperand(0);
4839 SDValue Op1 = Op.getOperand(1);
4840 SDValue CC = Op.getOperand(2);
4841 MVT VT = Op.getValueType();
4842 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4843 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4847 MVT VT0 = Op0.getValueType();
4848 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4849 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4852 switch (SetCCOpcode) {
4855 case ISD::SETEQ: SSECC = 0; break;
4857 case ISD::SETGT: Swap = true; // Fallthrough
4859 case ISD::SETOLT: SSECC = 1; break;
4861 case ISD::SETGE: Swap = true; // Fallthrough
4863 case ISD::SETOLE: SSECC = 2; break;
4864 case ISD::SETUO: SSECC = 3; break;
4866 case ISD::SETNE: SSECC = 4; break;
4867 case ISD::SETULE: Swap = true;
4868 case ISD::SETUGE: SSECC = 5; break;
4869 case ISD::SETULT: Swap = true;
4870 case ISD::SETUGT: SSECC = 6; break;
4871 case ISD::SETO: SSECC = 7; break;
4874 std::swap(Op0, Op1);
4876 // In the two special cases we can't handle, emit two comparisons.
4878 if (SetCCOpcode == ISD::SETUEQ) {
4880 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4881 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4882 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4884 else if (SetCCOpcode == ISD::SETONE) {
4886 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4887 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4888 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4890 assert(0 && "Illegal FP comparison");
4892 // Handle all other FP comparisons here.
4893 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4896 // We are handling one of the integer comparisons here. Since SSE only has
4897 // GT and EQ comparisons for integer, swapping operands and multiple
4898 // operations may be required for some comparisons.
4899 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4900 bool Swap = false, Invert = false, FlipSigns = false;
4902 switch (VT.getSimpleVT()) {
4904 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4905 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4906 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4907 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4910 switch (SetCCOpcode) {
4912 case ISD::SETNE: Invert = true;
4913 case ISD::SETEQ: Opc = EQOpc; break;
4914 case ISD::SETLT: Swap = true;
4915 case ISD::SETGT: Opc = GTOpc; break;
4916 case ISD::SETGE: Swap = true;
4917 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4918 case ISD::SETULT: Swap = true;
4919 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4920 case ISD::SETUGE: Swap = true;
4921 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4924 std::swap(Op0, Op1);
4926 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4927 // bits of the inputs before performing those operations.
4929 MVT EltVT = VT.getVectorElementType();
4930 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4931 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4932 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4934 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
4935 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
4938 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
4940 // If the logical-not of the result is required, perform that now.
4942 MVT EltVT = VT.getVectorElementType();
4943 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
4944 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
4945 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
4947 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
4952 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
4953 bool addTest = true;
4954 SDValue Cond = Op.getOperand(0);
4957 if (Cond.getOpcode() == ISD::SETCC)
4958 Cond = LowerSETCC(Cond, DAG);
4960 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4961 // setting operand in place of the X86ISD::SETCC.
4962 if (Cond.getOpcode() == X86ISD::SETCC) {
4963 CC = Cond.getOperand(0);
4965 SDValue Cmp = Cond.getOperand(1);
4966 unsigned Opc = Cmp.getOpcode();
4967 MVT VT = Op.getValueType();
4969 bool IllegalFPCMov = false;
4970 if (VT.isFloatingPoint() && !VT.isVector() &&
4971 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4972 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4974 if ((Opc == X86ISD::CMP ||
4975 Opc == X86ISD::COMI ||
4976 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4983 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4984 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4987 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4989 SmallVector<SDValue, 4> Ops;
4990 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4991 // condition is true.
4992 Ops.push_back(Op.getOperand(2));
4993 Ops.push_back(Op.getOperand(1));
4995 Ops.push_back(Cond);
4996 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4999 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5000 bool addTest = true;
5001 SDValue Chain = Op.getOperand(0);
5002 SDValue Cond = Op.getOperand(1);
5003 SDValue Dest = Op.getOperand(2);
5006 if (Cond.getOpcode() == ISD::SETCC)
5007 Cond = LowerSETCC(Cond, DAG);
5009 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5010 // setting operand in place of the X86ISD::SETCC.
5011 if (Cond.getOpcode() == X86ISD::SETCC) {
5012 CC = Cond.getOperand(0);
5014 SDValue Cmp = Cond.getOperand(1);
5015 unsigned Opc = Cmp.getOpcode();
5016 if (Opc == X86ISD::CMP ||
5017 Opc == X86ISD::COMI ||
5018 Opc == X86ISD::UCOMI) {
5025 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5026 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5028 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5029 Chain, Op.getOperand(2), CC, Cond);
5033 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5034 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5035 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5036 // that the guard pages used by the OS virtual memory manager are allocated in
5037 // correct sequence.
5039 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5040 SelectionDAG &DAG) {
5041 assert(Subtarget->isTargetCygMing() &&
5042 "This should be used only on Cygwin/Mingw targets");
5045 SDValue Chain = Op.getOperand(0);
5046 SDValue Size = Op.getOperand(1);
5047 // FIXME: Ensure alignment here
5051 MVT IntPtr = getPointerTy();
5052 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5054 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0));
5056 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5057 Flag = Chain.getValue(1);
5059 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5060 SDValue Ops[] = { Chain,
5061 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5062 DAG.getRegister(X86::EAX, IntPtr),
5063 DAG.getRegister(X86StackPtr, SPTy),
5065 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5066 Flag = Chain.getValue(1);
5068 Chain = DAG.getCALLSEQ_END(Chain,
5069 DAG.getIntPtrConstant(0),
5070 DAG.getIntPtrConstant(0),
5073 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5075 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5076 return DAG.getMergeValues(Ops1, 2);
5080 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5082 SDValue Dst, SDValue Src,
5083 SDValue Size, unsigned Align,
5084 const Value *DstSV, uint64_t DstSVOff) {
5085 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5087 /// If not DWORD aligned or size is more than the threshold, call the library.
5088 /// The libc version is likely to be faster for these cases. It can use the
5089 /// address value and run time information about the CPU.
5090 if ((Align & 3) != 0 ||
5092 ConstantSize->getValue() > getSubtarget()->getMaxInlineSizeThreshold()) {
5093 SDValue InFlag(0, 0);
5095 // Check to see if there is a specialized entry-point for memory zeroing.
5096 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5097 if (const char *bzeroEntry =
5098 V && V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5099 MVT IntPtr = getPointerTy();
5100 const Type *IntPtrTy = TD->getIntPtrType();
5101 TargetLowering::ArgListTy Args;
5102 TargetLowering::ArgListEntry Entry;
5104 Entry.Ty = IntPtrTy;
5105 Args.push_back(Entry);
5107 Args.push_back(Entry);
5108 std::pair<SDValue,SDValue> CallResult =
5109 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
5110 false, DAG.getExternalSymbol(bzeroEntry, IntPtr),
5112 return CallResult.second;
5115 // Otherwise have the target-independent code call memset.
5119 uint64_t SizeVal = ConstantSize->getValue();
5120 SDValue InFlag(0, 0);
5123 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5124 unsigned BytesLeft = 0;
5125 bool TwoRepStos = false;
5128 uint64_t Val = ValC->getValue() & 255;
5130 // If the value is a constant, then we can potentially use larger sets.
5131 switch (Align & 3) {
5132 case 2: // WORD aligned
5135 Val = (Val << 8) | Val;
5137 case 0: // DWORD aligned
5140 Val = (Val << 8) | Val;
5141 Val = (Val << 16) | Val;
5142 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5145 Val = (Val << 32) | Val;
5148 default: // Byte aligned
5151 Count = DAG.getIntPtrConstant(SizeVal);
5155 if (AVT.bitsGT(MVT::i8)) {
5156 unsigned UBytes = AVT.getSizeInBits() / 8;
5157 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5158 BytesLeft = SizeVal % UBytes;
5161 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5163 InFlag = Chain.getValue(1);
5166 Count = DAG.getIntPtrConstant(SizeVal);
5167 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5168 InFlag = Chain.getValue(1);
5171 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5173 InFlag = Chain.getValue(1);
5174 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5176 InFlag = Chain.getValue(1);
5178 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5179 SmallVector<SDValue, 8> Ops;
5180 Ops.push_back(Chain);
5181 Ops.push_back(DAG.getValueType(AVT));
5182 Ops.push_back(InFlag);
5183 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5186 InFlag = Chain.getValue(1);
5188 MVT CVT = Count.getValueType();
5189 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5190 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5191 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5193 InFlag = Chain.getValue(1);
5194 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5196 Ops.push_back(Chain);
5197 Ops.push_back(DAG.getValueType(MVT::i8));
5198 Ops.push_back(InFlag);
5199 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5200 } else if (BytesLeft) {
5201 // Handle the last 1 - 7 bytes.
5202 unsigned Offset = SizeVal - BytesLeft;
5203 MVT AddrVT = Dst.getValueType();
5204 MVT SizeVT = Size.getValueType();
5206 Chain = DAG.getMemset(Chain,
5207 DAG.getNode(ISD::ADD, AddrVT, Dst,
5208 DAG.getConstant(Offset, AddrVT)),
5210 DAG.getConstant(BytesLeft, SizeVT),
5211 Align, DstSV, DstSVOff + Offset);
5214 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5219 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5220 SDValue Chain, SDValue Dst, SDValue Src,
5221 SDValue Size, unsigned Align,
5223 const Value *DstSV, uint64_t DstSVOff,
5224 const Value *SrcSV, uint64_t SrcSVOff) {
5225 // This requires the copy size to be a constant, preferrably
5226 // within a subtarget-specific limit.
5227 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5230 uint64_t SizeVal = ConstantSize->getValue();
5231 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5234 /// If not DWORD aligned, call the library.
5235 if ((Align & 3) != 0)
5240 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5243 unsigned UBytes = AVT.getSizeInBits() / 8;
5244 unsigned CountVal = SizeVal / UBytes;
5245 SDValue Count = DAG.getIntPtrConstant(CountVal);
5246 unsigned BytesLeft = SizeVal % UBytes;
5248 SDValue InFlag(0, 0);
5249 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5251 InFlag = Chain.getValue(1);
5252 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5254 InFlag = Chain.getValue(1);
5255 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5257 InFlag = Chain.getValue(1);
5259 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5260 SmallVector<SDValue, 8> Ops;
5261 Ops.push_back(Chain);
5262 Ops.push_back(DAG.getValueType(AVT));
5263 Ops.push_back(InFlag);
5264 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5266 SmallVector<SDValue, 4> Results;
5267 Results.push_back(RepMovs);
5269 // Handle the last 1 - 7 bytes.
5270 unsigned Offset = SizeVal - BytesLeft;
5271 MVT DstVT = Dst.getValueType();
5272 MVT SrcVT = Src.getValueType();
5273 MVT SizeVT = Size.getValueType();
5274 Results.push_back(DAG.getMemcpy(Chain,
5275 DAG.getNode(ISD::ADD, DstVT, Dst,
5276 DAG.getConstant(Offset, DstVT)),
5277 DAG.getNode(ISD::ADD, SrcVT, Src,
5278 DAG.getConstant(Offset, SrcVT)),
5279 DAG.getConstant(BytesLeft, SizeVT),
5280 Align, AlwaysInline,
5281 DstSV, DstSVOff + Offset,
5282 SrcSV, SrcSVOff + Offset));
5285 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5288 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5289 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5290 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5291 SDValue TheChain = N->getOperand(0);
5292 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5293 if (Subtarget->is64Bit()) {
5294 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5295 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5296 MVT::i64, rax.getValue(2));
5297 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5298 DAG.getConstant(32, MVT::i8));
5300 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5303 return DAG.getMergeValues(Ops, 2).getNode();
5306 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5307 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5308 MVT::i32, eax.getValue(2));
5309 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5310 SDValue Ops[] = { eax, edx };
5311 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5313 // Use a MERGE_VALUES to return the value and chain.
5314 Ops[1] = edx.getValue(1);
5315 return DAG.getMergeValues(Ops, 2).getNode();
5318 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5319 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5321 if (!Subtarget->is64Bit()) {
5322 // vastart just stores the address of the VarArgsFrameIndex slot into the
5323 // memory location argument.
5324 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5325 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5329 // gp_offset (0 - 6 * 8)
5330 // fp_offset (48 - 48 + 8 * 16)
5331 // overflow_arg_area (point to parameters coming in memory).
5333 SmallVector<SDValue, 8> MemOps;
5334 SDValue FIN = Op.getOperand(1);
5336 SDValue Store = DAG.getStore(Op.getOperand(0),
5337 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5339 MemOps.push_back(Store);
5342 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5343 Store = DAG.getStore(Op.getOperand(0),
5344 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5346 MemOps.push_back(Store);
5348 // Store ptr to overflow_arg_area
5349 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5350 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5351 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5352 MemOps.push_back(Store);
5354 // Store ptr to reg_save_area.
5355 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5356 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5357 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5358 MemOps.push_back(Store);
5359 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5362 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5363 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5364 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5365 SDValue Chain = Op.getOperand(0);
5366 SDValue SrcPtr = Op.getOperand(1);
5367 SDValue SrcSV = Op.getOperand(2);
5369 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5374 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5375 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5376 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5377 SDValue Chain = Op.getOperand(0);
5378 SDValue DstPtr = Op.getOperand(1);
5379 SDValue SrcPtr = Op.getOperand(2);
5380 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5381 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5383 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5384 DAG.getIntPtrConstant(24), 8, false,
5385 DstSV, 0, SrcSV, 0);
5389 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5390 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5392 default: return SDValue(); // Don't custom lower most intrinsics.
5393 // Comparison intrinsics.
5394 case Intrinsic::x86_sse_comieq_ss:
5395 case Intrinsic::x86_sse_comilt_ss:
5396 case Intrinsic::x86_sse_comile_ss:
5397 case Intrinsic::x86_sse_comigt_ss:
5398 case Intrinsic::x86_sse_comige_ss:
5399 case Intrinsic::x86_sse_comineq_ss:
5400 case Intrinsic::x86_sse_ucomieq_ss:
5401 case Intrinsic::x86_sse_ucomilt_ss:
5402 case Intrinsic::x86_sse_ucomile_ss:
5403 case Intrinsic::x86_sse_ucomigt_ss:
5404 case Intrinsic::x86_sse_ucomige_ss:
5405 case Intrinsic::x86_sse_ucomineq_ss:
5406 case Intrinsic::x86_sse2_comieq_sd:
5407 case Intrinsic::x86_sse2_comilt_sd:
5408 case Intrinsic::x86_sse2_comile_sd:
5409 case Intrinsic::x86_sse2_comigt_sd:
5410 case Intrinsic::x86_sse2_comige_sd:
5411 case Intrinsic::x86_sse2_comineq_sd:
5412 case Intrinsic::x86_sse2_ucomieq_sd:
5413 case Intrinsic::x86_sse2_ucomilt_sd:
5414 case Intrinsic::x86_sse2_ucomile_sd:
5415 case Intrinsic::x86_sse2_ucomigt_sd:
5416 case Intrinsic::x86_sse2_ucomige_sd:
5417 case Intrinsic::x86_sse2_ucomineq_sd: {
5419 ISD::CondCode CC = ISD::SETCC_INVALID;
5422 case Intrinsic::x86_sse_comieq_ss:
5423 case Intrinsic::x86_sse2_comieq_sd:
5427 case Intrinsic::x86_sse_comilt_ss:
5428 case Intrinsic::x86_sse2_comilt_sd:
5432 case Intrinsic::x86_sse_comile_ss:
5433 case Intrinsic::x86_sse2_comile_sd:
5437 case Intrinsic::x86_sse_comigt_ss:
5438 case Intrinsic::x86_sse2_comigt_sd:
5442 case Intrinsic::x86_sse_comige_ss:
5443 case Intrinsic::x86_sse2_comige_sd:
5447 case Intrinsic::x86_sse_comineq_ss:
5448 case Intrinsic::x86_sse2_comineq_sd:
5452 case Intrinsic::x86_sse_ucomieq_ss:
5453 case Intrinsic::x86_sse2_ucomieq_sd:
5454 Opc = X86ISD::UCOMI;
5457 case Intrinsic::x86_sse_ucomilt_ss:
5458 case Intrinsic::x86_sse2_ucomilt_sd:
5459 Opc = X86ISD::UCOMI;
5462 case Intrinsic::x86_sse_ucomile_ss:
5463 case Intrinsic::x86_sse2_ucomile_sd:
5464 Opc = X86ISD::UCOMI;
5467 case Intrinsic::x86_sse_ucomigt_ss:
5468 case Intrinsic::x86_sse2_ucomigt_sd:
5469 Opc = X86ISD::UCOMI;
5472 case Intrinsic::x86_sse_ucomige_ss:
5473 case Intrinsic::x86_sse2_ucomige_sd:
5474 Opc = X86ISD::UCOMI;
5477 case Intrinsic::x86_sse_ucomineq_ss:
5478 case Intrinsic::x86_sse2_ucomineq_sd:
5479 Opc = X86ISD::UCOMI;
5485 SDValue LHS = Op.getOperand(1);
5486 SDValue RHS = Op.getOperand(2);
5487 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5489 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5490 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5491 DAG.getConstant(X86CC, MVT::i8), Cond);
5492 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5495 // Fix vector shift instructions where the last operand is a non-immediate
5497 case Intrinsic::x86_sse2_pslli_w:
5498 case Intrinsic::x86_sse2_pslli_d:
5499 case Intrinsic::x86_sse2_pslli_q:
5500 case Intrinsic::x86_sse2_psrli_w:
5501 case Intrinsic::x86_sse2_psrli_d:
5502 case Intrinsic::x86_sse2_psrli_q:
5503 case Intrinsic::x86_sse2_psrai_w:
5504 case Intrinsic::x86_sse2_psrai_d:
5505 case Intrinsic::x86_mmx_pslli_w:
5506 case Intrinsic::x86_mmx_pslli_d:
5507 case Intrinsic::x86_mmx_pslli_q:
5508 case Intrinsic::x86_mmx_psrli_w:
5509 case Intrinsic::x86_mmx_psrli_d:
5510 case Intrinsic::x86_mmx_psrli_q:
5511 case Intrinsic::x86_mmx_psrai_w:
5512 case Intrinsic::x86_mmx_psrai_d: {
5513 SDValue ShAmt = Op.getOperand(2);
5514 if (isa<ConstantSDNode>(ShAmt))
5517 unsigned NewIntNo = 0;
5518 MVT ShAmtVT = MVT::v4i32;
5520 case Intrinsic::x86_sse2_pslli_w:
5521 NewIntNo = Intrinsic::x86_sse2_psll_w;
5523 case Intrinsic::x86_sse2_pslli_d:
5524 NewIntNo = Intrinsic::x86_sse2_psll_d;
5526 case Intrinsic::x86_sse2_pslli_q:
5527 NewIntNo = Intrinsic::x86_sse2_psll_q;
5529 case Intrinsic::x86_sse2_psrli_w:
5530 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5532 case Intrinsic::x86_sse2_psrli_d:
5533 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5535 case Intrinsic::x86_sse2_psrli_q:
5536 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5538 case Intrinsic::x86_sse2_psrai_w:
5539 NewIntNo = Intrinsic::x86_sse2_psra_w;
5541 case Intrinsic::x86_sse2_psrai_d:
5542 NewIntNo = Intrinsic::x86_sse2_psra_d;
5545 ShAmtVT = MVT::v2i32;
5547 case Intrinsic::x86_mmx_pslli_w:
5548 NewIntNo = Intrinsic::x86_mmx_psll_w;
5550 case Intrinsic::x86_mmx_pslli_d:
5551 NewIntNo = Intrinsic::x86_mmx_psll_d;
5553 case Intrinsic::x86_mmx_pslli_q:
5554 NewIntNo = Intrinsic::x86_mmx_psll_q;
5556 case Intrinsic::x86_mmx_psrli_w:
5557 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5559 case Intrinsic::x86_mmx_psrli_d:
5560 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5562 case Intrinsic::x86_mmx_psrli_q:
5563 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5565 case Intrinsic::x86_mmx_psrai_w:
5566 NewIntNo = Intrinsic::x86_mmx_psra_w;
5568 case Intrinsic::x86_mmx_psrai_d:
5569 NewIntNo = Intrinsic::x86_mmx_psra_d;
5571 default: abort(); // Can't reach here.
5576 MVT VT = Op.getValueType();
5577 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5578 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5579 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5580 DAG.getConstant(NewIntNo, MVT::i32),
5581 Op.getOperand(1), ShAmt);
5586 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5587 // Depths > 0 not supported yet!
5588 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5591 // Just load the return address
5592 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5593 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5596 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5597 // Depths > 0 not supported yet!
5598 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5601 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5602 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5603 DAG.getIntPtrConstant(TD->getPointerSize()));
5606 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5607 SelectionDAG &DAG) {
5608 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5611 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5613 MachineFunction &MF = DAG.getMachineFunction();
5614 SDValue Chain = Op.getOperand(0);
5615 SDValue Offset = Op.getOperand(1);
5616 SDValue Handler = Op.getOperand(2);
5618 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5620 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5622 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5623 DAG.getIntPtrConstant(-TD->getPointerSize()));
5624 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5625 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5626 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5627 MF.getRegInfo().addLiveOut(StoreAddrReg);
5629 return DAG.getNode(X86ISD::EH_RETURN,
5631 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5634 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5635 SelectionDAG &DAG) {
5636 SDValue Root = Op.getOperand(0);
5637 SDValue Trmp = Op.getOperand(1); // trampoline
5638 SDValue FPtr = Op.getOperand(2); // nested function
5639 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5641 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5643 const X86InstrInfo *TII =
5644 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5646 if (Subtarget->is64Bit()) {
5647 SDValue OutChains[6];
5649 // Large code-model.
5651 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5652 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5654 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5655 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5657 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5659 // Load the pointer to the nested function into R11.
5660 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5661 SDValue Addr = Trmp;
5662 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5665 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5666 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5668 // Load the 'nest' parameter value into R10.
5669 // R10 is specified in X86CallingConv.td
5670 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5671 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5672 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5675 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5676 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5678 // Jump to the nested function.
5679 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5680 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5681 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5684 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5685 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5686 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5690 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5691 return DAG.getMergeValues(Ops, 2);
5693 const Function *Func =
5694 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5695 unsigned CC = Func->getCallingConv();
5700 assert(0 && "Unsupported calling convention");
5701 case CallingConv::C:
5702 case CallingConv::X86_StdCall: {
5703 // Pass 'nest' parameter in ECX.
5704 // Must be kept in sync with X86CallingConv.td
5707 // Check that ECX wasn't needed by an 'inreg' parameter.
5708 const FunctionType *FTy = Func->getFunctionType();
5709 const PAListPtr &Attrs = Func->getParamAttrs();
5711 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5712 unsigned InRegCount = 0;
5715 for (FunctionType::param_iterator I = FTy->param_begin(),
5716 E = FTy->param_end(); I != E; ++I, ++Idx)
5717 if (Attrs.paramHasAttr(Idx, ParamAttr::InReg))
5718 // FIXME: should only count parameters that are lowered to integers.
5719 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
5721 if (InRegCount > 2) {
5722 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5728 case CallingConv::X86_FastCall:
5729 case CallingConv::Fast:
5730 // Pass 'nest' parameter in EAX.
5731 // Must be kept in sync with X86CallingConv.td
5736 SDValue OutChains[4];
5739 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5740 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5742 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5743 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5744 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5747 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5748 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5750 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5751 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5752 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5753 TrmpAddr, 5, false, 1);
5755 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5756 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5759 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5760 return DAG.getMergeValues(Ops, 2);
5764 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5766 The rounding mode is in bits 11:10 of FPSR, and has the following
5773 FLT_ROUNDS, on the other hand, expects the following:
5780 To perform the conversion, we do:
5781 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5784 MachineFunction &MF = DAG.getMachineFunction();
5785 const TargetMachine &TM = MF.getTarget();
5786 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5787 unsigned StackAlignment = TFI.getStackAlignment();
5788 MVT VT = Op.getValueType();
5790 // Save FP Control Word to stack slot
5791 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5792 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5794 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5795 DAG.getEntryNode(), StackSlot);
5797 // Load FP Control Word from stack slot
5798 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5800 // Transform as necessary
5802 DAG.getNode(ISD::SRL, MVT::i16,
5803 DAG.getNode(ISD::AND, MVT::i16,
5804 CWD, DAG.getConstant(0x800, MVT::i16)),
5805 DAG.getConstant(11, MVT::i8));
5807 DAG.getNode(ISD::SRL, MVT::i16,
5808 DAG.getNode(ISD::AND, MVT::i16,
5809 CWD, DAG.getConstant(0x400, MVT::i16)),
5810 DAG.getConstant(9, MVT::i8));
5813 DAG.getNode(ISD::AND, MVT::i16,
5814 DAG.getNode(ISD::ADD, MVT::i16,
5815 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5816 DAG.getConstant(1, MVT::i16)),
5817 DAG.getConstant(3, MVT::i16));
5820 return DAG.getNode((VT.getSizeInBits() < 16 ?
5821 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5824 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5825 MVT VT = Op.getValueType();
5827 unsigned NumBits = VT.getSizeInBits();
5829 Op = Op.getOperand(0);
5830 if (VT == MVT::i8) {
5831 // Zero extend to i32 since there is not an i8 bsr.
5833 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5836 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5837 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5838 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5840 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5841 SmallVector<SDValue, 4> Ops;
5843 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5844 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5845 Ops.push_back(Op.getValue(1));
5846 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5848 // Finally xor with NumBits-1.
5849 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5852 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5856 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5857 MVT VT = Op.getValueType();
5859 unsigned NumBits = VT.getSizeInBits();
5861 Op = Op.getOperand(0);
5862 if (VT == MVT::i8) {
5864 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5867 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5868 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5869 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5871 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5872 SmallVector<SDValue, 4> Ops;
5874 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5875 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5876 Ops.push_back(Op.getValue(1));
5877 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5880 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5884 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5885 MVT T = Op.getValueType();
5888 switch(T.getSimpleVT()) {
5890 assert(false && "Invalid value type!");
5891 case MVT::i8: Reg = X86::AL; size = 1; break;
5892 case MVT::i16: Reg = X86::AX; size = 2; break;
5893 case MVT::i32: Reg = X86::EAX; size = 4; break;
5895 if (Subtarget->is64Bit()) {
5896 Reg = X86::RAX; size = 8;
5897 } else //Should go away when LowerType stuff lands
5898 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5901 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5902 Op.getOperand(2), SDValue());
5903 SDValue Ops[] = { cpIn.getValue(0),
5906 DAG.getTargetConstant(size, MVT::i8),
5908 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5909 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5911 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5915 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5916 SelectionDAG &DAG) {
5917 MVT T = Op->getValueType(0);
5918 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5919 SDValue cpInL, cpInH;
5920 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5921 DAG.getConstant(0, MVT::i32));
5922 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5923 DAG.getConstant(1, MVT::i32));
5924 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5926 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5927 cpInH, cpInL.getValue(1));
5928 SDValue swapInL, swapInH;
5929 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5930 DAG.getConstant(0, MVT::i32));
5931 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
5932 DAG.getConstant(1, MVT::i32));
5933 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
5934 swapInL, cpInH.getValue(1));
5935 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
5936 swapInH, swapInL.getValue(1));
5937 SDValue Ops[] = { swapInH.getValue(0),
5939 swapInH.getValue(1)};
5940 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5941 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
5942 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
5943 Result.getValue(1));
5944 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
5945 cpOutL.getValue(2));
5946 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
5947 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
5948 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
5949 return DAG.getMergeValues(Vals, 2).getNode();
5952 SDNode* X86TargetLowering::ExpandATOMIC_LOAD_SUB(SDNode* Op,
5953 SelectionDAG &DAG) {
5954 MVT T = Op->getValueType(0);
5955 SDValue negOp = DAG.getNode(ISD::SUB, T,
5956 DAG.getConstant(0, T), Op->getOperand(2));
5957 return DAG.getAtomic((T==MVT::i8 ? ISD::ATOMIC_LOAD_ADD_8:
5958 T==MVT::i16 ? ISD::ATOMIC_LOAD_ADD_16:
5959 T==MVT::i32 ? ISD::ATOMIC_LOAD_ADD_32:
5960 T==MVT::i64 ? ISD::ATOMIC_LOAD_ADD_64: 0),
5961 Op->getOperand(0), Op->getOperand(1), negOp,
5962 cast<AtomicSDNode>(Op)->getSrcValue(),
5963 cast<AtomicSDNode>(Op)->getAlignment()).getNode();
5966 /// LowerOperation - Provide custom lowering hooks for some operations.
5968 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5969 switch (Op.getOpcode()) {
5970 default: assert(0 && "Should not custom lower this!");
5971 case ISD::ATOMIC_CMP_SWAP_8: return LowerCMP_SWAP(Op,DAG);
5972 case ISD::ATOMIC_CMP_SWAP_16: return LowerCMP_SWAP(Op,DAG);
5973 case ISD::ATOMIC_CMP_SWAP_32: return LowerCMP_SWAP(Op,DAG);
5974 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
5975 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5976 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5977 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5978 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5979 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5980 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5981 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5982 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5983 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5984 case ISD::SHL_PARTS:
5985 case ISD::SRA_PARTS:
5986 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5987 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5988 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5989 case ISD::FABS: return LowerFABS(Op, DAG);
5990 case ISD::FNEG: return LowerFNEG(Op, DAG);
5991 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5992 case ISD::SETCC: return LowerSETCC(Op, DAG);
5993 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
5994 case ISD::SELECT: return LowerSELECT(Op, DAG);
5995 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5996 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5997 case ISD::CALL: return LowerCALL(Op, DAG);
5998 case ISD::RET: return LowerRET(Op, DAG);
5999 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6000 case ISD::VASTART: return LowerVASTART(Op, DAG);
6001 case ISD::VAARG: return LowerVAARG(Op, DAG);
6002 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6003 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6004 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6005 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6006 case ISD::FRAME_TO_ARGS_OFFSET:
6007 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6008 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6009 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6010 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6011 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6012 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6013 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6015 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
6016 case ISD::READCYCLECOUNTER:
6017 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
6021 /// ReplaceNodeResults - Replace a node with an illegal result type
6022 /// with a new node built out of custom code.
6023 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
6024 switch (N->getOpcode()) {
6025 default: assert(0 && "Should not custom lower this!");
6026 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
6027 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
6028 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
6029 case ISD::ATOMIC_LOAD_SUB_8: return ExpandATOMIC_LOAD_SUB(N,DAG);
6030 case ISD::ATOMIC_LOAD_SUB_16: return ExpandATOMIC_LOAD_SUB(N,DAG);
6031 case ISD::ATOMIC_LOAD_SUB_32: return ExpandATOMIC_LOAD_SUB(N,DAG);
6032 case ISD::ATOMIC_LOAD_SUB_64: return ExpandATOMIC_LOAD_SUB(N,DAG);
6036 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6038 default: return NULL;
6039 case X86ISD::BSF: return "X86ISD::BSF";
6040 case X86ISD::BSR: return "X86ISD::BSR";
6041 case X86ISD::SHLD: return "X86ISD::SHLD";
6042 case X86ISD::SHRD: return "X86ISD::SHRD";
6043 case X86ISD::FAND: return "X86ISD::FAND";
6044 case X86ISD::FOR: return "X86ISD::FOR";
6045 case X86ISD::FXOR: return "X86ISD::FXOR";
6046 case X86ISD::FSRL: return "X86ISD::FSRL";
6047 case X86ISD::FILD: return "X86ISD::FILD";
6048 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6049 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6050 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6051 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6052 case X86ISD::FLD: return "X86ISD::FLD";
6053 case X86ISD::FST: return "X86ISD::FST";
6054 case X86ISD::CALL: return "X86ISD::CALL";
6055 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6056 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6057 case X86ISD::CMP: return "X86ISD::CMP";
6058 case X86ISD::COMI: return "X86ISD::COMI";
6059 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6060 case X86ISD::SETCC: return "X86ISD::SETCC";
6061 case X86ISD::CMOV: return "X86ISD::CMOV";
6062 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6063 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6064 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6065 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6066 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6067 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6068 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6069 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6070 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6071 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6072 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6073 case X86ISD::FMAX: return "X86ISD::FMAX";
6074 case X86ISD::FMIN: return "X86ISD::FMIN";
6075 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6076 case X86ISD::FRCP: return "X86ISD::FRCP";
6077 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6078 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6079 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6080 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6081 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6082 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6083 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6084 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6085 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6086 case X86ISD::VSHL: return "X86ISD::VSHL";
6087 case X86ISD::VSRL: return "X86ISD::VSRL";
6088 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6089 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6090 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6091 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6092 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6093 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6094 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6095 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6096 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6097 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6101 // isLegalAddressingMode - Return true if the addressing mode represented
6102 // by AM is legal for this target, for a load/store of the specified type.
6103 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6104 const Type *Ty) const {
6105 // X86 supports extremely general addressing modes.
6107 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6108 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6112 // We can only fold this if we don't need an extra load.
6113 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6116 // X86-64 only supports addr of globals in small code model.
6117 if (Subtarget->is64Bit()) {
6118 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6120 // If lower 4G is not available, then we must use rip-relative addressing.
6121 if (AM.BaseOffs || AM.Scale > 1)
6132 // These scales always work.
6137 // These scales are formed with basereg+scalereg. Only accept if there is
6142 default: // Other stuff never works.
6150 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6151 if (!Ty1->isInteger() || !Ty2->isInteger())
6153 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6154 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6155 if (NumBits1 <= NumBits2)
6157 return Subtarget->is64Bit() || NumBits1 < 64;
6160 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6161 if (!VT1.isInteger() || !VT2.isInteger())
6163 unsigned NumBits1 = VT1.getSizeInBits();
6164 unsigned NumBits2 = VT2.getSizeInBits();
6165 if (NumBits1 <= NumBits2)
6167 return Subtarget->is64Bit() || NumBits1 < 64;
6170 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6171 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6172 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6173 /// are assumed to be legal.
6175 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6176 // Only do shuffles on 128-bit vector types for now.
6177 if (VT.getSizeInBits() == 64) return false;
6178 return (Mask.getNode()->getNumOperands() <= 4 ||
6179 isIdentityMask(Mask.getNode()) ||
6180 isIdentityMask(Mask.getNode(), true) ||
6181 isSplatMask(Mask.getNode()) ||
6182 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6183 X86::isUNPCKLMask(Mask.getNode()) ||
6184 X86::isUNPCKHMask(Mask.getNode()) ||
6185 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6186 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6190 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6191 MVT EVT, SelectionDAG &DAG) const {
6192 unsigned NumElts = BVOps.size();
6193 // Only do shuffles on 128-bit vector types for now.
6194 if (EVT.getSizeInBits() * NumElts == 64) return false;
6195 if (NumElts == 2) return true;
6197 return (isMOVLMask(&BVOps[0], 4) ||
6198 isCommutedMOVL(&BVOps[0], 4, true) ||
6199 isSHUFPMask(&BVOps[0], 4) ||
6200 isCommutedSHUFP(&BVOps[0], 4));
6205 //===----------------------------------------------------------------------===//
6206 // X86 Scheduler Hooks
6207 //===----------------------------------------------------------------------===//
6209 // private utility function
6211 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6212 MachineBasicBlock *MBB,
6220 TargetRegisterClass *RC,
6222 // For the atomic bitwise operator, we generate
6225 // ld t1 = [bitinstr.addr]
6226 // op t2 = t1, [bitinstr.val]
6228 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6230 // fallthrough -->nextMBB
6231 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6232 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6233 MachineFunction::iterator MBBIter = MBB;
6236 /// First build the CFG
6237 MachineFunction *F = MBB->getParent();
6238 MachineBasicBlock *thisMBB = MBB;
6239 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6240 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6241 F->insert(MBBIter, newMBB);
6242 F->insert(MBBIter, nextMBB);
6244 // Move all successors to thisMBB to nextMBB
6245 nextMBB->transferSuccessors(thisMBB);
6247 // Update thisMBB to fall through to newMBB
6248 thisMBB->addSuccessor(newMBB);
6250 // newMBB jumps to itself and fall through to nextMBB
6251 newMBB->addSuccessor(nextMBB);
6252 newMBB->addSuccessor(newMBB);
6254 // Insert instructions into newMBB based on incoming instruction
6255 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6256 MachineOperand& destOper = bInstr->getOperand(0);
6257 MachineOperand* argOpers[6];
6258 int numArgs = bInstr->getNumOperands() - 1;
6259 for (int i=0; i < numArgs; ++i)
6260 argOpers[i] = &bInstr->getOperand(i+1);
6262 // x86 address has 4 operands: base, index, scale, and displacement
6263 int lastAddrIndx = 3; // [0,3]
6266 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6267 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6268 for (int i=0; i <= lastAddrIndx; ++i)
6269 (*MIB).addOperand(*argOpers[i]);
6271 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6273 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6278 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6279 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6280 && "invalid operand");
6281 if (argOpers[valArgIndx]->isReg())
6282 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6284 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6286 (*MIB).addOperand(*argOpers[valArgIndx]);
6288 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6291 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6292 for (int i=0; i <= lastAddrIndx; ++i)
6293 (*MIB).addOperand(*argOpers[i]);
6295 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6296 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6298 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6302 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6304 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6308 // private utility function
6310 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6311 MachineBasicBlock *MBB,
6313 // For the atomic min/max operator, we generate
6316 // ld t1 = [min/max.addr]
6317 // mov t2 = [min/max.val]
6319 // cmov[cond] t2 = t1
6321 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6323 // fallthrough -->nextMBB
6325 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6326 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6327 MachineFunction::iterator MBBIter = MBB;
6330 /// First build the CFG
6331 MachineFunction *F = MBB->getParent();
6332 MachineBasicBlock *thisMBB = MBB;
6333 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6334 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6335 F->insert(MBBIter, newMBB);
6336 F->insert(MBBIter, nextMBB);
6338 // Move all successors to thisMBB to nextMBB
6339 nextMBB->transferSuccessors(thisMBB);
6341 // Update thisMBB to fall through to newMBB
6342 thisMBB->addSuccessor(newMBB);
6344 // newMBB jumps to newMBB and fall through to nextMBB
6345 newMBB->addSuccessor(nextMBB);
6346 newMBB->addSuccessor(newMBB);
6348 // Insert instructions into newMBB based on incoming instruction
6349 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6350 MachineOperand& destOper = mInstr->getOperand(0);
6351 MachineOperand* argOpers[6];
6352 int numArgs = mInstr->getNumOperands() - 1;
6353 for (int i=0; i < numArgs; ++i)
6354 argOpers[i] = &mInstr->getOperand(i+1);
6356 // x86 address has 4 operands: base, index, scale, and displacement
6357 int lastAddrIndx = 3; // [0,3]
6360 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6361 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6362 for (int i=0; i <= lastAddrIndx; ++i)
6363 (*MIB).addOperand(*argOpers[i]);
6365 // We only support register and immediate values
6366 assert( (argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm())
6367 && "invalid operand");
6369 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6370 if (argOpers[valArgIndx]->isReg())
6371 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6373 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6374 (*MIB).addOperand(*argOpers[valArgIndx]);
6376 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6379 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6384 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6385 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6389 // Cmp and exchange if none has modified the memory location
6390 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6391 for (int i=0; i <= lastAddrIndx; ++i)
6392 (*MIB).addOperand(*argOpers[i]);
6394 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6395 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6397 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6398 MIB.addReg(X86::EAX);
6401 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6403 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6409 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6410 MachineBasicBlock *BB) {
6411 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6412 switch (MI->getOpcode()) {
6413 default: assert(false && "Unexpected instr type to insert");
6414 case X86::CMOV_FR32:
6415 case X86::CMOV_FR64:
6416 case X86::CMOV_V4F32:
6417 case X86::CMOV_V2F64:
6418 case X86::CMOV_V2I64: {
6419 // To "insert" a SELECT_CC instruction, we actually have to insert the
6420 // diamond control-flow pattern. The incoming instruction knows the
6421 // destination vreg to set, the condition code register to branch on, the
6422 // true/false values to select between, and a branch opcode to use.
6423 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6424 MachineFunction::iterator It = BB;
6430 // cmpTY ccX, r1, r2
6432 // fallthrough --> copy0MBB
6433 MachineBasicBlock *thisMBB = BB;
6434 MachineFunction *F = BB->getParent();
6435 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6436 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6438 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6439 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6440 F->insert(It, copy0MBB);
6441 F->insert(It, sinkMBB);
6442 // Update machine-CFG edges by transferring all successors of the current
6443 // block to the new block which will contain the Phi node for the select.
6444 sinkMBB->transferSuccessors(BB);
6446 // Add the true and fallthrough blocks as its successors.
6447 BB->addSuccessor(copy0MBB);
6448 BB->addSuccessor(sinkMBB);
6451 // %FalseValue = ...
6452 // # fallthrough to sinkMBB
6455 // Update machine-CFG edges
6456 BB->addSuccessor(sinkMBB);
6459 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6462 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6463 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6464 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6466 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6470 case X86::FP32_TO_INT16_IN_MEM:
6471 case X86::FP32_TO_INT32_IN_MEM:
6472 case X86::FP32_TO_INT64_IN_MEM:
6473 case X86::FP64_TO_INT16_IN_MEM:
6474 case X86::FP64_TO_INT32_IN_MEM:
6475 case X86::FP64_TO_INT64_IN_MEM:
6476 case X86::FP80_TO_INT16_IN_MEM:
6477 case X86::FP80_TO_INT32_IN_MEM:
6478 case X86::FP80_TO_INT64_IN_MEM: {
6479 // Change the floating point control register to use "round towards zero"
6480 // mode when truncating to an integer value.
6481 MachineFunction *F = BB->getParent();
6482 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6483 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6485 // Load the old value of the high byte of the control word...
6487 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6488 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6490 // Set the high part to be round to zero...
6491 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6494 // Reload the modified control word now...
6495 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6497 // Restore the memory image of control word to original value
6498 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6501 // Get the X86 opcode to use.
6503 switch (MI->getOpcode()) {
6504 default: assert(0 && "illegal opcode!");
6505 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6506 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6507 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6508 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6509 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6510 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6511 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6512 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6513 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6517 MachineOperand &Op = MI->getOperand(0);
6518 if (Op.isRegister()) {
6519 AM.BaseType = X86AddressMode::RegBase;
6520 AM.Base.Reg = Op.getReg();
6522 AM.BaseType = X86AddressMode::FrameIndexBase;
6523 AM.Base.FrameIndex = Op.getIndex();
6525 Op = MI->getOperand(1);
6526 if (Op.isImmediate())
6527 AM.Scale = Op.getImm();
6528 Op = MI->getOperand(2);
6529 if (Op.isImmediate())
6530 AM.IndexReg = Op.getImm();
6531 Op = MI->getOperand(3);
6532 if (Op.isGlobalAddress()) {
6533 AM.GV = Op.getGlobal();
6535 AM.Disp = Op.getImm();
6537 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6538 .addReg(MI->getOperand(4).getReg());
6540 // Reload the original control word now.
6541 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6543 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6546 case X86::ATOMAND32:
6547 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6548 X86::AND32ri, X86::MOV32rm,
6549 X86::LCMPXCHG32, X86::MOV32rr,
6550 X86::NOT32r, X86::EAX,
6551 X86::GR32RegisterClass);
6553 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6554 X86::OR32ri, X86::MOV32rm,
6555 X86::LCMPXCHG32, X86::MOV32rr,
6556 X86::NOT32r, X86::EAX,
6557 X86::GR32RegisterClass);
6558 case X86::ATOMXOR32:
6559 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6560 X86::XOR32ri, X86::MOV32rm,
6561 X86::LCMPXCHG32, X86::MOV32rr,
6562 X86::NOT32r, X86::EAX,
6563 X86::GR32RegisterClass);
6564 case X86::ATOMNAND32:
6565 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6566 X86::AND32ri, X86::MOV32rm,
6567 X86::LCMPXCHG32, X86::MOV32rr,
6568 X86::NOT32r, X86::EAX,
6569 X86::GR32RegisterClass, true);
6570 case X86::ATOMMIN32:
6571 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6572 case X86::ATOMMAX32:
6573 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6574 case X86::ATOMUMIN32:
6575 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6576 case X86::ATOMUMAX32:
6577 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6579 case X86::ATOMAND16:
6580 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6581 X86::AND16ri, X86::MOV16rm,
6582 X86::LCMPXCHG16, X86::MOV16rr,
6583 X86::NOT16r, X86::AX,
6584 X86::GR16RegisterClass);
6586 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6587 X86::OR16ri, X86::MOV16rm,
6588 X86::LCMPXCHG16, X86::MOV16rr,
6589 X86::NOT16r, X86::AX,
6590 X86::GR16RegisterClass);
6591 case X86::ATOMXOR16:
6592 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6593 X86::XOR16ri, X86::MOV16rm,
6594 X86::LCMPXCHG16, X86::MOV16rr,
6595 X86::NOT16r, X86::AX,
6596 X86::GR16RegisterClass);
6597 case X86::ATOMNAND16:
6598 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6599 X86::AND16ri, X86::MOV16rm,
6600 X86::LCMPXCHG16, X86::MOV16rr,
6601 X86::NOT16r, X86::AX,
6602 X86::GR16RegisterClass, true);
6603 case X86::ATOMMIN16:
6604 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6605 case X86::ATOMMAX16:
6606 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6607 case X86::ATOMUMIN16:
6608 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6609 case X86::ATOMUMAX16:
6610 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6613 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6614 X86::AND8ri, X86::MOV8rm,
6615 X86::LCMPXCHG8, X86::MOV8rr,
6616 X86::NOT8r, X86::AL,
6617 X86::GR8RegisterClass);
6619 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6620 X86::OR8ri, X86::MOV8rm,
6621 X86::LCMPXCHG8, X86::MOV8rr,
6622 X86::NOT8r, X86::AL,
6623 X86::GR8RegisterClass);
6625 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6626 X86::XOR8ri, X86::MOV8rm,
6627 X86::LCMPXCHG8, X86::MOV8rr,
6628 X86::NOT8r, X86::AL,
6629 X86::GR8RegisterClass);
6630 case X86::ATOMNAND8:
6631 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6632 X86::AND8ri, X86::MOV8rm,
6633 X86::LCMPXCHG8, X86::MOV8rr,
6634 X86::NOT8r, X86::AL,
6635 X86::GR8RegisterClass, true);
6636 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6637 case X86::ATOMAND64:
6638 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6639 X86::AND64ri32, X86::MOV64rm,
6640 X86::LCMPXCHG64, X86::MOV64rr,
6641 X86::NOT64r, X86::RAX,
6642 X86::GR64RegisterClass);
6644 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6645 X86::OR64ri32, X86::MOV64rm,
6646 X86::LCMPXCHG64, X86::MOV64rr,
6647 X86::NOT64r, X86::RAX,
6648 X86::GR64RegisterClass);
6649 case X86::ATOMXOR64:
6650 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6651 X86::XOR64ri32, X86::MOV64rm,
6652 X86::LCMPXCHG64, X86::MOV64rr,
6653 X86::NOT64r, X86::RAX,
6654 X86::GR64RegisterClass);
6655 case X86::ATOMNAND64:
6656 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6657 X86::AND64ri32, X86::MOV64rm,
6658 X86::LCMPXCHG64, X86::MOV64rr,
6659 X86::NOT64r, X86::RAX,
6660 X86::GR64RegisterClass, true);
6661 case X86::ATOMMIN64:
6662 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6663 case X86::ATOMMAX64:
6664 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6665 case X86::ATOMUMIN64:
6666 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6667 case X86::ATOMUMAX64:
6668 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6672 //===----------------------------------------------------------------------===//
6673 // X86 Optimization Hooks
6674 //===----------------------------------------------------------------------===//
6676 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6680 const SelectionDAG &DAG,
6681 unsigned Depth) const {
6682 unsigned Opc = Op.getOpcode();
6683 assert((Opc >= ISD::BUILTIN_OP_END ||
6684 Opc == ISD::INTRINSIC_WO_CHAIN ||
6685 Opc == ISD::INTRINSIC_W_CHAIN ||
6686 Opc == ISD::INTRINSIC_VOID) &&
6687 "Should use MaskedValueIsZero if you don't know whether Op"
6688 " is a target node!");
6690 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
6694 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
6695 Mask.getBitWidth() - 1);
6700 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
6701 /// node is a GlobalAddress + offset.
6702 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
6703 GlobalValue* &GA, int64_t &Offset) const{
6704 if (N->getOpcode() == X86ISD::Wrapper) {
6705 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
6706 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
6710 return TargetLowering::isGAPlusOffset(N, GA, Offset);
6713 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
6714 const TargetLowering &TLI) {
6717 if (TLI.isGAPlusOffset(Base, GV, Offset))
6718 return (GV->getAlignment() >= N && (Offset % N) == 0);
6719 // DAG combine handles the stack object case.
6723 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
6724 unsigned NumElems, MVT EVT,
6726 SelectionDAG &DAG, MachineFrameInfo *MFI,
6727 const TargetLowering &TLI) {
6729 for (unsigned i = 0; i < NumElems; ++i) {
6730 SDValue Idx = PermMask.getOperand(i);
6731 if (Idx.getOpcode() == ISD::UNDEF) {
6737 SDValue Elt = DAG.getShuffleScalarElt(N, i);
6738 if (!Elt.getNode() ||
6739 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6742 Base = Elt.getNode();
6743 if (Base->getOpcode() == ISD::UNDEF)
6747 if (Elt.getOpcode() == ISD::UNDEF)
6750 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
6751 EVT.getSizeInBits()/8, i, MFI))
6757 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
6758 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
6759 /// if the load addresses are consecutive, non-overlapping, and in the right
6761 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
6762 const TargetLowering &TLI) {
6763 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6764 MVT VT = N->getValueType(0);
6765 MVT EVT = VT.getVectorElementType();
6766 SDValue PermMask = N->getOperand(2);
6767 unsigned NumElems = PermMask.getNumOperands();
6768 SDNode *Base = NULL;
6769 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
6773 LoadSDNode *LD = cast<LoadSDNode>(Base);
6774 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
6775 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6776 LD->getSrcValueOffset(), LD->isVolatile());
6777 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6778 LD->getSrcValueOffset(), LD->isVolatile(),
6779 LD->getAlignment());
6782 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
6783 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
6784 const X86Subtarget *Subtarget,
6785 const TargetLowering &TLI) {
6786 unsigned NumOps = N->getNumOperands();
6788 // Ignore single operand BUILD_VECTOR.
6792 MVT VT = N->getValueType(0);
6793 MVT EVT = VT.getVectorElementType();
6794 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
6795 // We are looking for load i64 and zero extend. We want to transform
6796 // it before legalizer has a chance to expand it. Also look for i64
6797 // BUILD_PAIR bit casted to f64.
6799 // This must be an insertion into a zero vector.
6800 SDValue HighElt = N->getOperand(1);
6801 if (!isZeroNode(HighElt))
6804 // Value must be a load.
6805 SDNode *Base = N->getOperand(0).getNode();
6806 if (!isa<LoadSDNode>(Base)) {
6807 if (Base->getOpcode() != ISD::BIT_CONVERT)
6809 Base = Base->getOperand(0).getNode();
6810 if (!isa<LoadSDNode>(Base))
6814 // Transform it into VZEXT_LOAD addr.
6815 LoadSDNode *LD = cast<LoadSDNode>(Base);
6817 // Load must not be an extload.
6818 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
6821 return DAG.getNode(X86ISD::VZEXT_LOAD, VT, LD->getChain(), LD->getBasePtr());
6824 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6825 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6826 const X86Subtarget *Subtarget) {
6827 SDValue Cond = N->getOperand(0);
6829 // If we have SSE[12] support, try to form min/max nodes.
6830 if (Subtarget->hasSSE2() &&
6831 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6832 if (Cond.getOpcode() == ISD::SETCC) {
6833 // Get the LHS/RHS of the select.
6834 SDValue LHS = N->getOperand(1);
6835 SDValue RHS = N->getOperand(2);
6836 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6838 unsigned Opcode = 0;
6839 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6842 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6845 if (!UnsafeFPMath) break;
6847 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6849 Opcode = X86ISD::FMIN;
6852 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6855 if (!UnsafeFPMath) break;
6857 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6859 Opcode = X86ISD::FMAX;
6862 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6865 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6868 if (!UnsafeFPMath) break;
6870 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6872 Opcode = X86ISD::FMIN;
6875 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6878 if (!UnsafeFPMath) break;
6880 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6882 Opcode = X86ISD::FMAX;
6888 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6896 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
6897 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
6898 const X86Subtarget *Subtarget) {
6899 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
6900 // the FP state in cases where an emms may be missing.
6901 // A preferable solution to the general problem is to figure out the right
6902 // places to insert EMMS. This qualifies as a quick hack.
6903 StoreSDNode *St = cast<StoreSDNode>(N);
6904 if (St->getValue().getValueType().isVector() &&
6905 St->getValue().getValueType().getSizeInBits() == 64 &&
6906 isa<LoadSDNode>(St->getValue()) &&
6907 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
6908 St->getChain().hasOneUse() && !St->isVolatile()) {
6909 SDNode* LdVal = St->getValue().getNode();
6911 int TokenFactorIndex = -1;
6912 SmallVector<SDValue, 8> Ops;
6913 SDNode* ChainVal = St->getChain().getNode();
6914 // Must be a store of a load. We currently handle two cases: the load
6915 // is a direct child, and it's under an intervening TokenFactor. It is
6916 // possible to dig deeper under nested TokenFactors.
6917 if (ChainVal == LdVal)
6918 Ld = cast<LoadSDNode>(St->getChain());
6919 else if (St->getValue().hasOneUse() &&
6920 ChainVal->getOpcode() == ISD::TokenFactor) {
6921 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
6922 if (ChainVal->getOperand(i).getNode() == LdVal) {
6923 TokenFactorIndex = i;
6924 Ld = cast<LoadSDNode>(St->getValue());
6926 Ops.push_back(ChainVal->getOperand(i));
6930 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
6931 if (Subtarget->is64Bit()) {
6932 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
6933 Ld->getBasePtr(), Ld->getSrcValue(),
6934 Ld->getSrcValueOffset(), Ld->isVolatile(),
6935 Ld->getAlignment());
6936 SDValue NewChain = NewLd.getValue(1);
6937 if (TokenFactorIndex != -1) {
6938 Ops.push_back(NewChain);
6939 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6942 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
6943 St->getSrcValue(), St->getSrcValueOffset(),
6944 St->isVolatile(), St->getAlignment());
6947 // Otherwise, lower to two 32-bit copies.
6948 SDValue LoAddr = Ld->getBasePtr();
6949 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6950 DAG.getConstant(4, MVT::i32));
6952 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
6953 Ld->getSrcValue(), Ld->getSrcValueOffset(),
6954 Ld->isVolatile(), Ld->getAlignment());
6955 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
6956 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
6958 MinAlign(Ld->getAlignment(), 4));
6960 SDValue NewChain = LoLd.getValue(1);
6961 if (TokenFactorIndex != -1) {
6962 Ops.push_back(LoLd);
6963 Ops.push_back(HiLd);
6964 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
6968 LoAddr = St->getBasePtr();
6969 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
6970 DAG.getConstant(4, MVT::i32));
6972 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
6973 St->getSrcValue(), St->getSrcValueOffset(),
6974 St->isVolatile(), St->getAlignment());
6975 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
6977 St->getSrcValueOffset() + 4,
6979 MinAlign(St->getAlignment(), 4));
6980 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
6986 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
6987 /// X86ISD::FXOR nodes.
6988 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
6989 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
6990 // F[X]OR(0.0, x) -> x
6991 // F[X]OR(x, 0.0) -> x
6992 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
6993 if (C->getValueAPF().isPosZero())
6994 return N->getOperand(1);
6995 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
6996 if (C->getValueAPF().isPosZero())
6997 return N->getOperand(0);
7001 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
7002 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
7003 // FAND(0.0, x) -> 0.0
7004 // FAND(x, 0.0) -> 0.0
7005 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7006 if (C->getValueAPF().isPosZero())
7007 return N->getOperand(0);
7008 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7009 if (C->getValueAPF().isPosZero())
7010 return N->getOperand(1);
7015 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7016 DAGCombinerInfo &DCI) const {
7017 SelectionDAG &DAG = DCI.DAG;
7018 switch (N->getOpcode()) {
7020 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7021 case ISD::BUILD_VECTOR:
7022 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7023 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7024 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7026 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7027 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7033 //===----------------------------------------------------------------------===//
7034 // X86 Inline Assembly Support
7035 //===----------------------------------------------------------------------===//
7037 /// getConstraintType - Given a constraint letter, return the type of
7038 /// constraint it is for this target.
7039 X86TargetLowering::ConstraintType
7040 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7041 if (Constraint.size() == 1) {
7042 switch (Constraint[0]) {
7053 return C_RegisterClass;
7058 return TargetLowering::getConstraintType(Constraint);
7061 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7062 /// with another that has more specific requirements based on the type of the
7063 /// corresponding operand.
7064 const char *X86TargetLowering::
7065 LowerXConstraint(MVT ConstraintVT) const {
7066 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7067 // 'f' like normal targets.
7068 if (ConstraintVT.isFloatingPoint()) {
7069 if (Subtarget->hasSSE2())
7071 if (Subtarget->hasSSE1())
7075 return TargetLowering::LowerXConstraint(ConstraintVT);
7078 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7079 /// vector. If it is invalid, don't add anything to Ops.
7080 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7082 std::vector<SDValue>&Ops,
7083 SelectionDAG &DAG) const {
7084 SDValue Result(0, 0);
7086 switch (Constraint) {
7089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7090 if (C->getValue() <= 31) {
7091 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7097 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7098 if (C->getValue() <= 255) {
7099 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
7105 // Literal immediates are always ok.
7106 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7107 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
7111 // If we are in non-pic codegen mode, we allow the address of a global (with
7112 // an optional displacement) to be used with 'i'.
7113 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7116 // Match either (GA) or (GA+C)
7118 Offset = GA->getOffset();
7119 } else if (Op.getOpcode() == ISD::ADD) {
7120 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7121 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7123 Offset = GA->getOffset()+C->getValue();
7125 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7126 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7128 Offset = GA->getOffset()+C->getValue();
7135 // If addressing this global requires a load (e.g. in PIC mode), we can't
7137 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
7141 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7147 // Otherwise, not valid for this mode.
7152 if (Result.getNode()) {
7153 Ops.push_back(Result);
7156 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7159 std::vector<unsigned> X86TargetLowering::
7160 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7162 if (Constraint.size() == 1) {
7163 // FIXME: not handling fp-stack yet!
7164 switch (Constraint[0]) { // GCC X86 Constraint Letters
7165 default: break; // Unknown constraint letter
7166 case 'A': // EAX/EDX
7167 if (VT == MVT::i32 || VT == MVT::i64)
7168 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7170 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7173 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7174 else if (VT == MVT::i16)
7175 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7176 else if (VT == MVT::i8)
7177 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7178 else if (VT == MVT::i64)
7179 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7184 return std::vector<unsigned>();
7187 std::pair<unsigned, const TargetRegisterClass*>
7188 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7190 // First, see if this is a constraint that directly corresponds to an LLVM
7192 if (Constraint.size() == 1) {
7193 // GCC Constraint Letters
7194 switch (Constraint[0]) {
7196 case 'r': // GENERAL_REGS
7197 case 'R': // LEGACY_REGS
7198 case 'l': // INDEX_REGS
7199 if (VT == MVT::i64 && Subtarget->is64Bit())
7200 return std::make_pair(0U, X86::GR64RegisterClass);
7202 return std::make_pair(0U, X86::GR32RegisterClass);
7203 else if (VT == MVT::i16)
7204 return std::make_pair(0U, X86::GR16RegisterClass);
7205 else if (VT == MVT::i8)
7206 return std::make_pair(0U, X86::GR8RegisterClass);
7208 case 'f': // FP Stack registers.
7209 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7210 // value to the correct fpstack register class.
7211 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7212 return std::make_pair(0U, X86::RFP32RegisterClass);
7213 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7214 return std::make_pair(0U, X86::RFP64RegisterClass);
7215 return std::make_pair(0U, X86::RFP80RegisterClass);
7216 case 'y': // MMX_REGS if MMX allowed.
7217 if (!Subtarget->hasMMX()) break;
7218 return std::make_pair(0U, X86::VR64RegisterClass);
7220 case 'Y': // SSE_REGS if SSE2 allowed
7221 if (!Subtarget->hasSSE2()) break;
7223 case 'x': // SSE_REGS if SSE1 allowed
7224 if (!Subtarget->hasSSE1()) break;
7226 switch (VT.getSimpleVT()) {
7228 // Scalar SSE types.
7231 return std::make_pair(0U, X86::FR32RegisterClass);
7234 return std::make_pair(0U, X86::FR64RegisterClass);
7242 return std::make_pair(0U, X86::VR128RegisterClass);
7248 // Use the default implementation in TargetLowering to convert the register
7249 // constraint into a member of a register class.
7250 std::pair<unsigned, const TargetRegisterClass*> Res;
7251 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7253 // Not found as a standard register?
7254 if (Res.second == 0) {
7255 // GCC calls "st(0)" just plain "st".
7256 if (StringsEqualNoCase("{st}", Constraint)) {
7257 Res.first = X86::ST0;
7258 Res.second = X86::RFP80RegisterClass;
7264 // Otherwise, check to see if this is a register class of the wrong value
7265 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7266 // turn into {ax},{dx}.
7267 if (Res.second->hasType(VT))
7268 return Res; // Correct type already, nothing to do.
7270 // All of the single-register GCC register classes map their values onto
7271 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7272 // really want an 8-bit or 32-bit register, map to the appropriate register
7273 // class and return the appropriate register.
7274 if (Res.second == X86::GR16RegisterClass) {
7275 if (VT == MVT::i8) {
7276 unsigned DestReg = 0;
7277 switch (Res.first) {
7279 case X86::AX: DestReg = X86::AL; break;
7280 case X86::DX: DestReg = X86::DL; break;
7281 case X86::CX: DestReg = X86::CL; break;
7282 case X86::BX: DestReg = X86::BL; break;
7285 Res.first = DestReg;
7286 Res.second = Res.second = X86::GR8RegisterClass;
7288 } else if (VT == MVT::i32) {
7289 unsigned DestReg = 0;
7290 switch (Res.first) {
7292 case X86::AX: DestReg = X86::EAX; break;
7293 case X86::DX: DestReg = X86::EDX; break;
7294 case X86::CX: DestReg = X86::ECX; break;
7295 case X86::BX: DestReg = X86::EBX; break;
7296 case X86::SI: DestReg = X86::ESI; break;
7297 case X86::DI: DestReg = X86::EDI; break;
7298 case X86::BP: DestReg = X86::EBP; break;
7299 case X86::SP: DestReg = X86::ESP; break;
7302 Res.first = DestReg;
7303 Res.second = Res.second = X86::GR32RegisterClass;
7305 } else if (VT == MVT::i64) {
7306 unsigned DestReg = 0;
7307 switch (Res.first) {
7309 case X86::AX: DestReg = X86::RAX; break;
7310 case X86::DX: DestReg = X86::RDX; break;
7311 case X86::CX: DestReg = X86::RCX; break;
7312 case X86::BX: DestReg = X86::RBX; break;
7313 case X86::SI: DestReg = X86::RSI; break;
7314 case X86::DI: DestReg = X86::RDI; break;
7315 case X86::BP: DestReg = X86::RBP; break;
7316 case X86::SP: DestReg = X86::RSP; break;
7319 Res.first = DestReg;
7320 Res.second = Res.second = X86::GR64RegisterClass;
7323 } else if (Res.second == X86::FR32RegisterClass ||
7324 Res.second == X86::FR64RegisterClass ||
7325 Res.second == X86::VR128RegisterClass) {
7326 // Handle references to XMM physical registers that got mapped into the
7327 // wrong class. This can happen with constraints like {xmm0} where the
7328 // target independent register mapper will just pick the first match it can
7329 // find, ignoring the required type.
7331 Res.second = X86::FR32RegisterClass;
7332 else if (VT == MVT::f64)
7333 Res.second = X86::FR64RegisterClass;
7334 else if (X86::VR128RegisterClass->hasType(VT))
7335 Res.second = X86::VR128RegisterClass;