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 setLoadExtAction(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, Custom);
301 setOperationAction(ISD::ATOMIC_LOAD_SUB_16, MVT::i16, Custom);
302 setOperationAction(ISD::ATOMIC_LOAD_SUB_32, MVT::i32, Custom);
303 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Custom);
305 if (!Subtarget->is64Bit()) {
306 setOperationAction(ISD::ATOMIC_LOAD_ADD_64, MVT::i64, Custom);
307 setOperationAction(ISD::ATOMIC_LOAD_SUB_64, MVT::i64, Custom);
308 setOperationAction(ISD::ATOMIC_LOAD_AND_64, MVT::i64, Custom);
309 setOperationAction(ISD::ATOMIC_LOAD_OR_64, MVT::i64, Custom);
310 setOperationAction(ISD::ATOMIC_LOAD_XOR_64, MVT::i64, Custom);
311 setOperationAction(ISD::ATOMIC_LOAD_NAND_64, MVT::i64, Custom);
312 setOperationAction(ISD::ATOMIC_SWAP_64, MVT::i64, Custom);
315 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
316 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
317 // FIXME - use subtarget debug flags
318 if (!Subtarget->isTargetDarwin() &&
319 !Subtarget->isTargetELF() &&
320 !Subtarget->isTargetCygMing()) {
321 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
322 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
325 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
326 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
327 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
328 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
329 if (Subtarget->is64Bit()) {
330 setExceptionPointerRegister(X86::RAX);
331 setExceptionSelectorRegister(X86::RDX);
333 setExceptionPointerRegister(X86::EAX);
334 setExceptionSelectorRegister(X86::EDX);
336 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
337 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
339 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
341 setOperationAction(ISD::TRAP, MVT::Other, Legal);
343 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
344 setOperationAction(ISD::VASTART , MVT::Other, Custom);
345 setOperationAction(ISD::VAEND , MVT::Other, Expand);
346 if (Subtarget->is64Bit()) {
347 setOperationAction(ISD::VAARG , MVT::Other, Custom);
348 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
350 setOperationAction(ISD::VAARG , MVT::Other, Expand);
351 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
354 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
355 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
356 if (Subtarget->is64Bit())
357 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
358 if (Subtarget->isTargetCygMing())
359 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
361 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
363 if (X86ScalarSSEf64) {
364 // f32 and f64 use SSE.
365 // Set up the FP register classes.
366 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
367 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
369 // Use ANDPD to simulate FABS.
370 setOperationAction(ISD::FABS , MVT::f64, Custom);
371 setOperationAction(ISD::FABS , MVT::f32, Custom);
373 // Use XORP to simulate FNEG.
374 setOperationAction(ISD::FNEG , MVT::f64, Custom);
375 setOperationAction(ISD::FNEG , MVT::f32, Custom);
377 // Use ANDPD and ORPD to simulate FCOPYSIGN.
378 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
379 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
381 // We don't support sin/cos/fmod
382 setOperationAction(ISD::FSIN , MVT::f64, Expand);
383 setOperationAction(ISD::FCOS , MVT::f64, Expand);
384 setOperationAction(ISD::FSIN , MVT::f32, Expand);
385 setOperationAction(ISD::FCOS , MVT::f32, Expand);
387 // Expand FP immediates into loads from the stack, except for the special
389 addLegalFPImmediate(APFloat(+0.0)); // xorpd
390 addLegalFPImmediate(APFloat(+0.0f)); // xorps
392 // Floating truncations from f80 and extensions to f80 go through memory.
393 // If optimizing, we lie about this though and handle it in
394 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
396 setConvertAction(MVT::f32, MVT::f80, Expand);
397 setConvertAction(MVT::f64, MVT::f80, Expand);
398 setConvertAction(MVT::f80, MVT::f32, Expand);
399 setConvertAction(MVT::f80, MVT::f64, Expand);
401 } else if (X86ScalarSSEf32) {
402 // Use SSE for f32, x87 for f64.
403 // Set up the FP register classes.
404 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
405 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
407 // Use ANDPS to simulate FABS.
408 setOperationAction(ISD::FABS , MVT::f32, Custom);
410 // Use XORP to simulate FNEG.
411 setOperationAction(ISD::FNEG , MVT::f32, Custom);
413 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
415 // Use ANDPS and ORPS to simulate FCOPYSIGN.
416 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
417 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
419 // We don't support sin/cos/fmod
420 setOperationAction(ISD::FSIN , MVT::f32, Expand);
421 setOperationAction(ISD::FCOS , MVT::f32, Expand);
423 // Special cases we handle for FP constants.
424 addLegalFPImmediate(APFloat(+0.0f)); // xorps
425 addLegalFPImmediate(APFloat(+0.0)); // FLD0
426 addLegalFPImmediate(APFloat(+1.0)); // FLD1
427 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
428 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
430 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
431 // this though and handle it in InstructionSelectPreprocess so that
432 // dagcombine2 can hack on these.
434 setConvertAction(MVT::f32, MVT::f64, Expand);
435 setConvertAction(MVT::f32, MVT::f80, Expand);
436 setConvertAction(MVT::f80, MVT::f32, Expand);
437 setConvertAction(MVT::f64, MVT::f32, Expand);
438 // And x87->x87 truncations also.
439 setConvertAction(MVT::f80, MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
444 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
447 // f32 and f64 in x87.
448 // Set up the FP register classes.
449 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
450 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
452 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
453 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
454 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
455 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
457 // Floating truncations go through memory. If optimizing, we lie about
458 // this though and handle it in InstructionSelectPreprocess so that
459 // dagcombine2 can hack on these.
461 setConvertAction(MVT::f80, MVT::f32, Expand);
462 setConvertAction(MVT::f64, MVT::f32, Expand);
463 setConvertAction(MVT::f80, MVT::f64, Expand);
467 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
468 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
470 addLegalFPImmediate(APFloat(+0.0)); // FLD0
471 addLegalFPImmediate(APFloat(+1.0)); // FLD1
472 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
473 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
474 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
480 // Long double always uses X87.
481 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
482 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
483 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
486 APFloat TmpFlt(+0.0);
487 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
489 addLegalFPImmediate(TmpFlt); // FLD0
491 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
492 APFloat TmpFlt2(+1.0);
493 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
495 addLegalFPImmediate(TmpFlt2); // FLD1
496 TmpFlt2.changeSign();
497 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
501 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
502 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
505 // Always use a library call for pow.
506 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
507 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
508 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
510 setOperationAction(ISD::FLOG, MVT::f80, Expand);
511 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
512 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
513 setOperationAction(ISD::FEXP, MVT::f80, Expand);
514 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
516 // First set operation action for all vector types to expand. Then we
517 // will selectively turn on ones that can be effectively codegen'd.
518 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
519 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
520 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
535 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
565 if (Subtarget->hasMMX()) {
566 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
567 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
568 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
569 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
570 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
572 // FIXME: add MMX packed arithmetics
574 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
575 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
576 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
577 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
579 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
580 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
581 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
582 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
584 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
585 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
587 setOperationAction(ISD::AND, MVT::v8i8, Promote);
588 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
589 setOperationAction(ISD::AND, MVT::v4i16, Promote);
590 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
591 setOperationAction(ISD::AND, MVT::v2i32, Promote);
592 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
593 setOperationAction(ISD::AND, MVT::v1i64, Legal);
595 setOperationAction(ISD::OR, MVT::v8i8, Promote);
596 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
597 setOperationAction(ISD::OR, MVT::v4i16, Promote);
598 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
599 setOperationAction(ISD::OR, MVT::v2i32, Promote);
600 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
601 setOperationAction(ISD::OR, MVT::v1i64, Legal);
603 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
604 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
605 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
606 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
607 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
608 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
609 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
611 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
612 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
613 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
614 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
615 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
616 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
617 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
618 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
619 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
621 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
622 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
623 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
624 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
625 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
628 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
629 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
630 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
632 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
633 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
634 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
635 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
637 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
640 if (Subtarget->hasSSE1()) {
641 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
643 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
644 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
645 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
646 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
647 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
648 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
649 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
651 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
652 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
653 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
654 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
657 if (Subtarget->hasSSE2()) {
658 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
659 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
660 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
661 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
662 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
664 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
665 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
666 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
667 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
668 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
669 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
670 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
671 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
672 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
673 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
674 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
675 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
676 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
677 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
678 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
680 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
681 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
682 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
683 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
685 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
686 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
687 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
688 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
689 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
691 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
692 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
693 MVT VT = (MVT::SimpleValueType)i;
694 // Do not attempt to custom lower non-power-of-2 vectors
695 if (!isPowerOf2_32(VT.getVectorNumElements()))
697 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
698 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
699 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
701 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
702 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
703 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
704 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
705 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
706 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
707 if (Subtarget->is64Bit()) {
708 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
709 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
712 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
713 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
714 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
715 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
716 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
717 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
718 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
719 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
720 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
721 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
722 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
723 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
726 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
728 // Custom lower v2i64 and v2f64 selects.
729 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
730 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
731 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
732 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
736 if (Subtarget->hasSSE41()) {
737 // FIXME: Do we need to handle scalar-to-vector here?
738 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
739 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
741 // i8 and i16 vectors are custom , because the source register and source
742 // source memory operand types are not the same width. f32 vectors are
743 // custom since the immediate controlling the insert encodes additional
745 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
746 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
748 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
750 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
751 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
752 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
753 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
755 if (Subtarget->is64Bit()) {
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
757 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
761 if (Subtarget->hasSSE42()) {
762 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
765 // We want to custom lower some of our intrinsics.
766 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
768 // We have target-specific dag combine patterns for the following nodes:
769 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
770 setTargetDAGCombine(ISD::BUILD_VECTOR);
771 setTargetDAGCombine(ISD::SELECT);
772 setTargetDAGCombine(ISD::STORE);
774 computeRegisterProperties();
776 // FIXME: These should be based on subtarget info. Plus, the values should
777 // be smaller when we are in optimizing for size mode.
778 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
779 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
780 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
781 allowUnalignedMemoryAccesses = true; // x86 supports it!
782 setPrefLoopAlignment(16);
786 MVT X86TargetLowering::getSetCCResultType(const SDValue &) const {
791 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
792 /// the desired ByVal argument alignment.
793 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
796 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
797 if (VTy->getBitWidth() == 128)
799 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
800 unsigned EltAlign = 0;
801 getMaxByValAlign(ATy->getElementType(), EltAlign);
802 if (EltAlign > MaxAlign)
804 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
805 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
806 unsigned EltAlign = 0;
807 getMaxByValAlign(STy->getElementType(i), EltAlign);
808 if (EltAlign > MaxAlign)
817 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
818 /// function arguments in the caller parameter area. For X86, aggregates
819 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
820 /// are at 4-byte boundaries.
821 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
822 if (Subtarget->is64Bit()) {
823 // Max of 8 and alignment of type.
824 unsigned TyAlign = TD->getABITypeAlignment(Ty);
831 if (Subtarget->hasSSE1())
832 getMaxByValAlign(Ty, Align);
836 /// getOptimalMemOpType - Returns the target specific optimal type for load
837 /// and store operations as a result of memset, memcpy, and memmove
838 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
841 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
842 bool isSrcConst, bool isSrcStr) const {
843 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
845 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
847 if (Subtarget->is64Bit() && Size >= 8)
853 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
855 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
856 SelectionDAG &DAG) const {
857 if (usesGlobalOffsetTable())
858 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
859 if (!Subtarget->isPICStyleRIPRel())
860 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
864 //===----------------------------------------------------------------------===//
865 // Return Value Calling Convention Implementation
866 //===----------------------------------------------------------------------===//
868 #include "X86GenCallingConv.inc"
870 /// LowerRET - Lower an ISD::RET node.
871 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
872 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
874 SmallVector<CCValAssign, 16> RVLocs;
875 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
876 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
877 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
878 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
880 // If this is the first return lowered for this function, add the regs to the
881 // liveout set for the function.
882 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
883 for (unsigned i = 0; i != RVLocs.size(); ++i)
884 if (RVLocs[i].isRegLoc())
885 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
887 SDValue Chain = Op.getOperand(0);
889 // Handle tail call return.
890 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
891 if (Chain.getOpcode() == X86ISD::TAILCALL) {
892 SDValue TailCall = Chain;
893 SDValue TargetAddress = TailCall.getOperand(1);
894 SDValue StackAdjustment = TailCall.getOperand(2);
895 assert(((TargetAddress.getOpcode() == ISD::Register &&
896 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
897 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
898 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
899 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
900 "Expecting an global address, external symbol, or register");
901 assert(StackAdjustment.getOpcode() == ISD::Constant &&
902 "Expecting a const value");
904 SmallVector<SDValue,8> Operands;
905 Operands.push_back(Chain.getOperand(0));
906 Operands.push_back(TargetAddress);
907 Operands.push_back(StackAdjustment);
908 // Copy registers used by the call. Last operand is a flag so it is not
910 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
911 Operands.push_back(Chain.getOperand(i));
913 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
920 SmallVector<SDValue, 6> RetOps;
921 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
922 // Operand #1 = Bytes To Pop
923 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
925 // Copy the result values into the output registers.
926 for (unsigned i = 0; i != RVLocs.size(); ++i) {
927 CCValAssign &VA = RVLocs[i];
928 assert(VA.isRegLoc() && "Can only return in registers!");
929 SDValue ValToCopy = Op.getOperand(i*2+1);
931 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
932 // the RET instruction and handled by the FP Stackifier.
933 if (RVLocs[i].getLocReg() == X86::ST0 ||
934 RVLocs[i].getLocReg() == X86::ST1) {
935 // If this is a copy from an xmm register to ST(0), use an FPExtend to
936 // change the value to the FP stack register class.
937 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
938 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
939 RetOps.push_back(ValToCopy);
940 // Don't emit a copytoreg.
944 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
945 Flag = Chain.getValue(1);
948 // The x86-64 ABI for returning structs by value requires that we copy
949 // the sret argument into %rax for the return. We saved the argument into
950 // a virtual register in the entry block, so now we copy the value out
952 if (Subtarget->is64Bit() &&
953 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
954 MachineFunction &MF = DAG.getMachineFunction();
955 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
956 unsigned Reg = FuncInfo->getSRetReturnReg();
958 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
959 FuncInfo->setSRetReturnReg(Reg);
961 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
963 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
964 Flag = Chain.getValue(1);
967 RetOps[0] = Chain; // Update chain.
969 // Add the flag if we have it.
971 RetOps.push_back(Flag);
973 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
977 /// LowerCallResult - Lower the result values of an ISD::CALL into the
978 /// appropriate copies out of appropriate physical registers. This assumes that
979 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
980 /// being lowered. The returns a SDNode with the same number of values as the
982 SDNode *X86TargetLowering::
983 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
984 unsigned CallingConv, SelectionDAG &DAG) {
986 // Assign locations to each value returned by this call.
987 SmallVector<CCValAssign, 16> RVLocs;
988 bool isVarArg = TheCall->isVarArg();
989 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
990 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
992 SmallVector<SDValue, 8> ResultVals;
994 // Copy all of the result registers out of their specified physreg.
995 for (unsigned i = 0; i != RVLocs.size(); ++i) {
996 MVT CopyVT = RVLocs[i].getValVT();
998 // If this is a call to a function that returns an fp value on the floating
999 // point stack, but where we prefer to use the value in xmm registers, copy
1000 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1001 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1002 RVLocs[i].getLocReg() == X86::ST1) &&
1003 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
1007 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1008 CopyVT, InFlag).getValue(1);
1009 SDValue Val = Chain.getValue(0);
1010 InFlag = Chain.getValue(2);
1012 if (CopyVT != RVLocs[i].getValVT()) {
1013 // Round the F80 the right size, which also moves to the appropriate xmm
1015 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1016 // This truncation won't change the value.
1017 DAG.getIntPtrConstant(1));
1020 ResultVals.push_back(Val);
1023 // Merge everything together with a MERGE_VALUES node.
1024 ResultVals.push_back(Chain);
1025 return DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
1026 ResultVals.size()).getNode();
1030 //===----------------------------------------------------------------------===//
1031 // C & StdCall & Fast Calling Convention implementation
1032 //===----------------------------------------------------------------------===//
1033 // StdCall calling convention seems to be standard for many Windows' API
1034 // routines and around. It differs from C calling convention just a little:
1035 // callee should clean up the stack, not caller. Symbols should be also
1036 // decorated in some fancy way :) It doesn't support any vector arguments.
1037 // For info on fast calling convention see Fast Calling Convention (tail call)
1038 // implementation LowerX86_32FastCCCallTo.
1040 /// AddLiveIn - This helper function adds the specified physical register to the
1041 /// MachineFunction as a live in value. It also creates a corresponding virtual
1042 /// register for it.
1043 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1044 const TargetRegisterClass *RC) {
1045 assert(RC->contains(PReg) && "Not the correct regclass!");
1046 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1047 MF.getRegInfo().addLiveIn(PReg, VReg);
1051 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1053 static bool CallIsStructReturn(CallSDNode *TheCall) {
1054 unsigned NumOps = TheCall->getNumArgs();
1058 return TheCall->getArgFlags(0).isSRet();
1061 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1062 /// return semantics.
1063 static bool ArgsAreStructReturn(SDValue Op) {
1064 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1068 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1071 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1072 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1074 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1078 switch (CallingConv) {
1081 case CallingConv::X86_StdCall:
1082 return !Subtarget->is64Bit();
1083 case CallingConv::X86_FastCall:
1084 return !Subtarget->is64Bit();
1085 case CallingConv::Fast:
1086 return PerformTailCallOpt;
1090 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1091 /// given CallingConvention value.
1092 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
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)
1105 return CC_X86_32_FastCC;
1110 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1111 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1113 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1114 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1115 if (CC == CallingConv::X86_FastCall)
1117 else if (CC == CallingConv::X86_StdCall)
1123 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1124 /// in a register before calling.
1125 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1126 return !IsTailCall && !Is64Bit &&
1127 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1128 Subtarget->isPICStyleGOT();
1131 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1132 /// address to be loaded in a register.
1134 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1135 return !Is64Bit && IsTailCall &&
1136 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1137 Subtarget->isPICStyleGOT();
1140 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1141 /// by "Src" to address "Dst" with size and alignment information specified by
1142 /// the specific parameter attribute. The copy will be passed as a byval
1143 /// function parameter.
1145 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1146 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1147 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1148 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1149 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1152 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1153 const CCValAssign &VA,
1154 MachineFrameInfo *MFI,
1156 SDValue Root, unsigned i) {
1157 // Create the nodes corresponding to a load from this parameter slot.
1158 ISD::ArgFlagsTy Flags =
1159 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1160 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1161 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1163 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1164 // changed with more analysis.
1165 // In case of tail call optimization mark all arguments mutable. Since they
1166 // could be overwritten by lowering of arguments in case of a tail call.
1167 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1168 VA.getLocMemOffset(), isImmutable);
1169 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1170 if (Flags.isByVal())
1172 return DAG.getLoad(VA.getValVT(), Root, FIN,
1173 PseudoSourceValue::getFixedStack(FI), 0);
1177 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1178 MachineFunction &MF = DAG.getMachineFunction();
1179 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1181 const Function* Fn = MF.getFunction();
1182 if (Fn->hasExternalLinkage() &&
1183 Subtarget->isTargetCygMing() &&
1184 Fn->getName() == "main")
1185 FuncInfo->setForceFramePointer(true);
1187 // Decorate the function name.
1188 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1190 MachineFrameInfo *MFI = MF.getFrameInfo();
1191 SDValue Root = Op.getOperand(0);
1192 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1193 unsigned CC = MF.getFunction()->getCallingConv();
1194 bool Is64Bit = Subtarget->is64Bit();
1195 bool IsWin64 = Subtarget->isTargetWin64();
1197 assert(!(isVarArg && CC == CallingConv::Fast) &&
1198 "Var args not supported with calling convention fastcc");
1200 // Assign locations to all of the incoming arguments.
1201 SmallVector<CCValAssign, 16> ArgLocs;
1202 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1203 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1205 SmallVector<SDValue, 8> ArgValues;
1206 unsigned LastVal = ~0U;
1207 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1208 CCValAssign &VA = ArgLocs[i];
1209 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1211 assert(VA.getValNo() != LastVal &&
1212 "Don't support value assigned to multiple locs yet");
1213 LastVal = VA.getValNo();
1215 if (VA.isRegLoc()) {
1216 MVT RegVT = VA.getLocVT();
1217 TargetRegisterClass *RC;
1218 if (RegVT == MVT::i32)
1219 RC = X86::GR32RegisterClass;
1220 else if (Is64Bit && RegVT == MVT::i64)
1221 RC = X86::GR64RegisterClass;
1222 else if (RegVT == MVT::f32)
1223 RC = X86::FR32RegisterClass;
1224 else if (RegVT == MVT::f64)
1225 RC = X86::FR64RegisterClass;
1226 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1227 RC = X86::VR128RegisterClass;
1228 else if (RegVT.isVector()) {
1229 assert(RegVT.getSizeInBits() == 64);
1231 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1233 // Darwin calling convention passes MMX values in either GPRs or
1234 // XMMs in x86-64. Other targets pass them in memory.
1235 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1236 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1239 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1244 assert(0 && "Unknown argument type!");
1247 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1248 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1250 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1251 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1253 if (VA.getLocInfo() == CCValAssign::SExt)
1254 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1255 DAG.getValueType(VA.getValVT()));
1256 else if (VA.getLocInfo() == CCValAssign::ZExt)
1257 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1258 DAG.getValueType(VA.getValVT()));
1260 if (VA.getLocInfo() != CCValAssign::Full)
1261 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1263 // Handle MMX values passed in GPRs.
1264 if (Is64Bit && RegVT != VA.getLocVT()) {
1265 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1266 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1267 else if (RC == X86::VR128RegisterClass) {
1268 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1269 DAG.getConstant(0, MVT::i64));
1270 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1274 ArgValues.push_back(ArgValue);
1276 assert(VA.isMemLoc());
1277 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1281 // The x86-64 ABI for returning structs by value requires that we copy
1282 // the sret argument into %rax for the return. Save the argument into
1283 // a virtual register so that we can access it from the return points.
1284 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1285 MachineFunction &MF = DAG.getMachineFunction();
1286 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1287 unsigned Reg = FuncInfo->getSRetReturnReg();
1289 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1290 FuncInfo->setSRetReturnReg(Reg);
1292 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1293 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1296 unsigned StackSize = CCInfo.getNextStackOffset();
1297 // align stack specially for tail calls
1298 if (PerformTailCallOpt && CC == CallingConv::Fast)
1299 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1301 // If the function takes variable number of arguments, make a frame index for
1302 // the start of the first vararg value... for expansion of llvm.va_start.
1304 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1305 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1308 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1310 // FIXME: We should really autogenerate these arrays
1311 static const unsigned GPR64ArgRegsWin64[] = {
1312 X86::RCX, X86::RDX, X86::R8, X86::R9
1314 static const unsigned XMMArgRegsWin64[] = {
1315 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1317 static const unsigned GPR64ArgRegs64Bit[] = {
1318 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1320 static const unsigned XMMArgRegs64Bit[] = {
1321 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1322 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1324 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1327 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1328 GPR64ArgRegs = GPR64ArgRegsWin64;
1329 XMMArgRegs = XMMArgRegsWin64;
1331 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1332 GPR64ArgRegs = GPR64ArgRegs64Bit;
1333 XMMArgRegs = XMMArgRegs64Bit;
1335 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1337 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1340 // For X86-64, if there are vararg parameters that are passed via
1341 // registers, then we must store them to their spots on the stack so they
1342 // may be loaded by deferencing the result of va_next.
1343 VarArgsGPOffset = NumIntRegs * 8;
1344 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1345 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1346 TotalNumXMMRegs * 16, 16);
1348 // Store the integer parameter registers.
1349 SmallVector<SDValue, 8> MemOps;
1350 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1351 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1352 DAG.getIntPtrConstant(VarArgsGPOffset));
1353 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1354 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1355 X86::GR64RegisterClass);
1356 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1358 DAG.getStore(Val.getValue(1), Val, FIN,
1359 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1360 MemOps.push_back(Store);
1361 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1362 DAG.getIntPtrConstant(8));
1365 // Now store the XMM (fp + vector) parameter registers.
1366 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1367 DAG.getIntPtrConstant(VarArgsFPOffset));
1368 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1369 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1370 X86::VR128RegisterClass);
1371 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1373 DAG.getStore(Val.getValue(1), Val, FIN,
1374 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1375 MemOps.push_back(Store);
1376 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1377 DAG.getIntPtrConstant(16));
1379 if (!MemOps.empty())
1380 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1381 &MemOps[0], MemOps.size());
1385 ArgValues.push_back(Root);
1387 // Some CCs need callee pop.
1388 if (IsCalleePop(isVarArg, CC)) {
1389 BytesToPopOnReturn = StackSize; // Callee pops everything.
1390 BytesCallerReserves = 0;
1392 BytesToPopOnReturn = 0; // Callee pops nothing.
1393 // If this is an sret function, the return should pop the hidden pointer.
1394 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1395 BytesToPopOnReturn = 4;
1396 BytesCallerReserves = StackSize;
1400 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1401 if (CC == CallingConv::X86_FastCall)
1402 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1405 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1407 // Return the new list of results.
1408 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1409 ArgValues.size()).getValue(Op.getResNo());
1413 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1414 const SDValue &StackPtr,
1415 const CCValAssign &VA,
1417 SDValue Arg, ISD::ArgFlagsTy Flags) {
1418 unsigned LocMemOffset = VA.getLocMemOffset();
1419 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1420 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1421 if (Flags.isByVal()) {
1422 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1424 return DAG.getStore(Chain, Arg, PtrOff,
1425 PseudoSourceValue::getStack(), LocMemOffset);
1428 /// EmitTailCallLoadRetAddr - Emit a load of return adress if tail call
1429 /// optimization is performed and it is required.
1431 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1432 SDValue &OutRetAddr,
1437 if (!IsTailCall || FPDiff==0) return Chain;
1439 // Adjust the Return address stack slot.
1440 MVT VT = getPointerTy();
1441 OutRetAddr = getReturnAddressFrameIndex(DAG);
1442 // Load the "old" Return address.
1443 OutRetAddr = DAG.getLoad(VT, Chain,OutRetAddr, NULL, 0);
1444 return SDValue(OutRetAddr.getNode(), 1);
1447 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1448 /// optimization is performed and it is required (FPDiff!=0).
1450 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1451 SDValue Chain, SDValue RetAddrFrIdx,
1452 bool Is64Bit, int FPDiff) {
1453 // Store the return address to the appropriate stack slot.
1454 if (!FPDiff) return Chain;
1455 // Calculate the new stack slot for the return address.
1456 int SlotSize = Is64Bit ? 8 : 4;
1457 int NewReturnAddrFI =
1458 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1459 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1460 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1461 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1462 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1466 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1467 MachineFunction &MF = DAG.getMachineFunction();
1468 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1469 SDValue Chain = TheCall->getChain();
1470 unsigned CC = TheCall->getCallingConv();
1471 bool isVarArg = TheCall->isVarArg();
1472 bool IsTailCall = TheCall->isTailCall() &&
1473 CC == CallingConv::Fast && PerformTailCallOpt;
1474 SDValue Callee = TheCall->getCallee();
1475 bool Is64Bit = Subtarget->is64Bit();
1476 bool IsStructRet = CallIsStructReturn(TheCall);
1478 assert(!(isVarArg && CC == CallingConv::Fast) &&
1479 "Var args not supported with calling convention fastcc");
1481 // Analyze operands of the call, assigning locations to each operand.
1482 SmallVector<CCValAssign, 16> ArgLocs;
1483 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1484 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1486 // Get a count of how many bytes are to be pushed on the stack.
1487 unsigned NumBytes = CCInfo.getNextStackOffset();
1488 if (PerformTailCallOpt && CC == CallingConv::Fast)
1489 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1493 // Lower arguments at fp - stackoffset + fpdiff.
1494 unsigned NumBytesCallerPushed =
1495 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1496 FPDiff = NumBytesCallerPushed - NumBytes;
1498 // Set the delta of movement of the returnaddr stackslot.
1499 // But only set if delta is greater than previous delta.
1500 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1501 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1504 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1506 SDValue RetAddrFrIdx;
1507 // Load return adress for tail calls.
1508 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1511 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1512 SmallVector<SDValue, 8> MemOpChains;
1515 // Walk the register/memloc assignments, inserting copies/loads. In the case
1516 // of tail call optimization arguments are handle later.
1517 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1518 CCValAssign &VA = ArgLocs[i];
1519 SDValue Arg = TheCall->getArg(i);
1520 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1521 bool isByVal = Flags.isByVal();
1523 // Promote the value if needed.
1524 switch (VA.getLocInfo()) {
1525 default: assert(0 && "Unknown loc info!");
1526 case CCValAssign::Full: break;
1527 case CCValAssign::SExt:
1528 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1530 case CCValAssign::ZExt:
1531 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1533 case CCValAssign::AExt:
1534 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1538 if (VA.isRegLoc()) {
1540 MVT RegVT = VA.getLocVT();
1541 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1542 switch (VA.getLocReg()) {
1545 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1547 // Special case: passing MMX values in GPR registers.
1548 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1551 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1552 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1553 // Special case: passing MMX values in XMM registers.
1554 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1555 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1556 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1557 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1558 getMOVLMask(2, DAG));
1563 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1565 if (!IsTailCall || (IsTailCall && isByVal)) {
1566 assert(VA.isMemLoc());
1567 if (StackPtr.getNode() == 0)
1568 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1570 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1571 Chain, Arg, Flags));
1576 if (!MemOpChains.empty())
1577 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1578 &MemOpChains[0], MemOpChains.size());
1580 // Build a sequence of copy-to-reg nodes chained together with token chain
1581 // and flag operands which copy the outgoing args into registers.
1583 // Tail call byval lowering might overwrite argument registers so in case of
1584 // tail call optimization the copies to registers are lowered later.
1586 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1587 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1589 InFlag = Chain.getValue(1);
1592 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1594 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1595 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1596 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1598 InFlag = Chain.getValue(1);
1600 // If we are tail calling and generating PIC/GOT style code load the address
1601 // of the callee into ecx. The value in ecx is used as target of the tail
1602 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1603 // calls on PIC/GOT architectures. Normally we would just put the address of
1604 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1605 // restored (since ebx is callee saved) before jumping to the target@PLT.
1606 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1607 // Note: The actual moving to ecx is done further down.
1608 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1609 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1610 !G->getGlobal()->hasProtectedVisibility())
1611 Callee = LowerGlobalAddress(Callee, DAG);
1612 else if (isa<ExternalSymbolSDNode>(Callee))
1613 Callee = LowerExternalSymbol(Callee,DAG);
1616 if (Is64Bit && isVarArg) {
1617 // From AMD64 ABI document:
1618 // For calls that may call functions that use varargs or stdargs
1619 // (prototype-less calls or calls to functions containing ellipsis (...) in
1620 // the declaration) %al is used as hidden argument to specify the number
1621 // of SSE registers used. The contents of %al do not need to match exactly
1622 // the number of registers, but must be an ubound on the number of SSE
1623 // registers used and is in the range 0 - 8 inclusive.
1625 // FIXME: Verify this on Win64
1626 // Count the number of XMM registers allocated.
1627 static const unsigned XMMArgRegs[] = {
1628 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1629 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1631 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1633 Chain = DAG.getCopyToReg(Chain, X86::AL,
1634 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1635 InFlag = Chain.getValue(1);
1639 // For tail calls lower the arguments to the 'real' stack slot.
1641 SmallVector<SDValue, 8> MemOpChains2;
1644 // Do not flag preceeding copytoreg stuff together with the following stuff.
1646 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1647 CCValAssign &VA = ArgLocs[i];
1648 if (!VA.isRegLoc()) {
1649 assert(VA.isMemLoc());
1650 SDValue Arg = TheCall->getArg(i);
1651 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1652 // Create frame index.
1653 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1654 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1655 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1656 FIN = DAG.getFrameIndex(FI, getPointerTy());
1658 if (Flags.isByVal()) {
1659 // Copy relative to framepointer.
1660 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1661 if (StackPtr.getNode() == 0)
1662 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1663 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1665 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1668 // Store relative to framepointer.
1669 MemOpChains2.push_back(
1670 DAG.getStore(Chain, Arg, FIN,
1671 PseudoSourceValue::getFixedStack(FI), 0));
1676 if (!MemOpChains2.empty())
1677 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1678 &MemOpChains2[0], MemOpChains2.size());
1680 // Copy arguments to their registers.
1681 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1682 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1684 InFlag = Chain.getValue(1);
1688 // Store the return address to the appropriate stack slot.
1689 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1693 // If the callee is a GlobalAddress node (quite common, every direct call is)
1694 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1695 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1696 // We should use extra load for direct calls to dllimported functions in
1698 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1699 getTargetMachine(), true))
1700 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1701 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1702 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1703 } else if (IsTailCall) {
1704 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1706 Chain = DAG.getCopyToReg(Chain,
1707 DAG.getRegister(Opc, getPointerTy()),
1709 Callee = DAG.getRegister(Opc, getPointerTy());
1710 // Add register as live out.
1711 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1714 // Returns a chain & a flag for retval copy to use.
1715 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1716 SmallVector<SDValue, 8> Ops;
1719 Ops.push_back(Chain);
1720 Ops.push_back(DAG.getIntPtrConstant(NumBytes, true));
1721 Ops.push_back(DAG.getIntPtrConstant(0, true));
1722 if (InFlag.getNode())
1723 Ops.push_back(InFlag);
1724 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1725 InFlag = Chain.getValue(1);
1727 // Returns a chain & a flag for retval copy to use.
1728 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1732 Ops.push_back(Chain);
1733 Ops.push_back(Callee);
1736 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1738 // Add argument registers to the end of the list so that they are known live
1740 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1741 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1742 RegsToPass[i].second.getValueType()));
1744 // Add an implicit use GOT pointer in EBX.
1745 if (!IsTailCall && !Is64Bit &&
1746 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1747 Subtarget->isPICStyleGOT())
1748 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1750 // Add an implicit use of AL for x86 vararg functions.
1751 if (Is64Bit && isVarArg)
1752 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1754 if (InFlag.getNode())
1755 Ops.push_back(InFlag);
1758 assert(InFlag.getNode() &&
1759 "Flag must be set. Depend on flag being set in LowerRET");
1760 Chain = DAG.getNode(X86ISD::TAILCALL,
1761 TheCall->getVTList(), &Ops[0], Ops.size());
1763 return SDValue(Chain.getNode(), Op.getResNo());
1766 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1767 InFlag = Chain.getValue(1);
1769 // Create the CALLSEQ_END node.
1770 unsigned NumBytesForCalleeToPush;
1771 if (IsCalleePop(isVarArg, CC))
1772 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1773 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1774 // If this is is a call to a struct-return function, the callee
1775 // pops the hidden struct pointer, so we have to push it back.
1776 // This is common for Darwin/X86, Linux & Mingw32 targets.
1777 NumBytesForCalleeToPush = 4;
1779 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1781 // Returns a flag for retval copy to use.
1782 Chain = DAG.getCALLSEQ_END(Chain,
1783 DAG.getIntPtrConstant(NumBytes, true),
1784 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1787 InFlag = Chain.getValue(1);
1789 // Handle result values, copying them out of physregs into vregs that we
1791 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1796 //===----------------------------------------------------------------------===//
1797 // Fast Calling Convention (tail call) implementation
1798 //===----------------------------------------------------------------------===//
1800 // Like std call, callee cleans arguments, convention except that ECX is
1801 // reserved for storing the tail called function address. Only 2 registers are
1802 // free for argument passing (inreg). Tail call optimization is performed
1804 // * tailcallopt is enabled
1805 // * caller/callee are fastcc
1806 // On X86_64 architecture with GOT-style position independent code only local
1807 // (within module) calls are supported at the moment.
1808 // To keep the stack aligned according to platform abi the function
1809 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1810 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1811 // If a tail called function callee has more arguments than the caller the
1812 // caller needs to make sure that there is room to move the RETADDR to. This is
1813 // achieved by reserving an area the size of the argument delta right after the
1814 // original REtADDR, but before the saved framepointer or the spilled registers
1815 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1827 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1828 /// for a 16 byte align requirement.
1829 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1830 SelectionDAG& DAG) {
1831 MachineFunction &MF = DAG.getMachineFunction();
1832 const TargetMachine &TM = MF.getTarget();
1833 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1834 unsigned StackAlignment = TFI.getStackAlignment();
1835 uint64_t AlignMask = StackAlignment - 1;
1836 int64_t Offset = StackSize;
1837 uint64_t SlotSize = TD->getPointerSize();
1838 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1839 // Number smaller than 12 so just add the difference.
1840 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1842 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1843 Offset = ((~AlignMask) & Offset) + StackAlignment +
1844 (StackAlignment-SlotSize);
1849 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1850 /// following the call is a return. A function is eligible if caller/callee
1851 /// calling conventions match, currently only fastcc supports tail calls, and
1852 /// the function CALL is immediatly followed by a RET.
1853 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1855 SelectionDAG& DAG) const {
1856 if (!PerformTailCallOpt)
1859 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1860 MachineFunction &MF = DAG.getMachineFunction();
1861 unsigned CallerCC = MF.getFunction()->getCallingConv();
1862 unsigned CalleeCC= TheCall->getCallingConv();
1863 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1864 SDValue Callee = TheCall->getCallee();
1865 // On x86/32Bit PIC/GOT tail calls are supported.
1866 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1867 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1870 // Can only do local tail calls (in same module, hidden or protected) on
1871 // x86_64 PIC/GOT at the moment.
1872 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1873 return G->getGlobal()->hasHiddenVisibility()
1874 || G->getGlobal()->hasProtectedVisibility();
1882 X86TargetLowering::createFastISel(MachineFunction &mf,
1883 MachineModuleInfo *mmo,
1884 DenseMap<const Value *, unsigned> &vm,
1885 DenseMap<const BasicBlock *,
1886 MachineBasicBlock *> &bm,
1887 DenseMap<const AllocaInst *, int> &am) {
1889 return X86::createFastISel(mf, mmo, vm, bm, am);
1893 //===----------------------------------------------------------------------===//
1894 // Other Lowering Hooks
1895 //===----------------------------------------------------------------------===//
1898 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1899 MachineFunction &MF = DAG.getMachineFunction();
1900 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1901 int ReturnAddrIndex = FuncInfo->getRAIndex();
1902 uint64_t SlotSize = TD->getPointerSize();
1904 if (ReturnAddrIndex == 0) {
1905 // Set up a frame object for the return address.
1906 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1907 FuncInfo->setRAIndex(ReturnAddrIndex);
1910 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1914 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1915 /// specific condition code. It returns a false if it cannot do a direct
1916 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1918 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1919 unsigned &X86CC, SDValue &LHS, SDValue &RHS,
1920 SelectionDAG &DAG) {
1921 X86CC = X86::COND_INVALID;
1923 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1924 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1925 // X > -1 -> X == 0, jump !sign.
1926 RHS = DAG.getConstant(0, RHS.getValueType());
1927 X86CC = X86::COND_NS;
1929 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1930 // X < 0 -> X == 0, jump on sign.
1931 X86CC = X86::COND_S;
1933 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
1935 RHS = DAG.getConstant(0, RHS.getValueType());
1936 X86CC = X86::COND_LE;
1941 switch (SetCCOpcode) {
1943 case ISD::SETEQ: X86CC = X86::COND_E; break;
1944 case ISD::SETGT: X86CC = X86::COND_G; break;
1945 case ISD::SETGE: X86CC = X86::COND_GE; break;
1946 case ISD::SETLT: X86CC = X86::COND_L; break;
1947 case ISD::SETLE: X86CC = X86::COND_LE; break;
1948 case ISD::SETNE: X86CC = X86::COND_NE; break;
1949 case ISD::SETULT: X86CC = X86::COND_B; break;
1950 case ISD::SETUGT: X86CC = X86::COND_A; break;
1951 case ISD::SETULE: X86CC = X86::COND_BE; break;
1952 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1955 // First determine if it requires or is profitable to flip the operands.
1957 switch (SetCCOpcode) {
1967 // If LHS is a foldable load, but RHS is not, flip the condition.
1969 (ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
1970 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
1971 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
1975 std::swap(LHS, RHS);
1977 // On a floating point condition, the flags are set as follows:
1979 // 0 | 0 | 0 | X > Y
1980 // 0 | 0 | 1 | X < Y
1981 // 1 | 0 | 0 | X == Y
1982 // 1 | 1 | 1 | unordered
1983 switch (SetCCOpcode) {
1987 X86CC = X86::COND_E;
1989 case ISD::SETOLT: // flipped
1992 X86CC = X86::COND_A;
1994 case ISD::SETOLE: // flipped
1997 X86CC = X86::COND_AE;
1999 case ISD::SETUGT: // flipped
2002 X86CC = X86::COND_B;
2004 case ISD::SETUGE: // flipped
2007 X86CC = X86::COND_BE;
2011 X86CC = X86::COND_NE;
2014 X86CC = X86::COND_P;
2017 X86CC = X86::COND_NP;
2022 return X86CC != X86::COND_INVALID;
2025 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2026 /// code. Current x86 isa includes the following FP cmov instructions:
2027 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2028 static bool hasFPCMov(unsigned X86CC) {
2044 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2045 /// true if Op is undef or if its value falls within the specified range (L, H].
2046 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2047 if (Op.getOpcode() == ISD::UNDEF)
2050 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2051 return (Val >= Low && Val < Hi);
2054 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2055 /// true if Op is undef or if its value equal to the specified value.
2056 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2057 if (Op.getOpcode() == ISD::UNDEF)
2059 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2062 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2063 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2064 bool X86::isPSHUFDMask(SDNode *N) {
2065 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2067 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2070 // Check if the value doesn't reference the second vector.
2071 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2072 SDValue Arg = N->getOperand(i);
2073 if (Arg.getOpcode() == ISD::UNDEF) continue;
2074 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2075 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2082 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2083 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2084 bool X86::isPSHUFHWMask(SDNode *N) {
2085 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2087 if (N->getNumOperands() != 8)
2090 // Lower quadword copied in order.
2091 for (unsigned i = 0; i != 4; ++i) {
2092 SDValue Arg = N->getOperand(i);
2093 if (Arg.getOpcode() == ISD::UNDEF) continue;
2094 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2095 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2099 // Upper quadword shuffled.
2100 for (unsigned i = 4; i != 8; ++i) {
2101 SDValue Arg = N->getOperand(i);
2102 if (Arg.getOpcode() == ISD::UNDEF) continue;
2103 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2104 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2105 if (Val < 4 || Val > 7)
2112 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2113 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2114 bool X86::isPSHUFLWMask(SDNode *N) {
2115 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2117 if (N->getNumOperands() != 8)
2120 // Upper quadword copied in order.
2121 for (unsigned i = 4; i != 8; ++i)
2122 if (!isUndefOrEqual(N->getOperand(i), i))
2125 // Lower quadword shuffled.
2126 for (unsigned i = 0; i != 4; ++i)
2127 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2133 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2134 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2135 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2136 if (NumElems != 2 && NumElems != 4) return false;
2138 unsigned Half = NumElems / 2;
2139 for (unsigned i = 0; i < Half; ++i)
2140 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2142 for (unsigned i = Half; i < NumElems; ++i)
2143 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2149 bool X86::isSHUFPMask(SDNode *N) {
2150 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2151 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2154 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2155 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2156 /// half elements to come from vector 1 (which would equal the dest.) and
2157 /// the upper half to come from vector 2.
2158 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2159 if (NumOps != 2 && NumOps != 4) return false;
2161 unsigned Half = NumOps / 2;
2162 for (unsigned i = 0; i < Half; ++i)
2163 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2165 for (unsigned i = Half; i < NumOps; ++i)
2166 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2171 static bool isCommutedSHUFP(SDNode *N) {
2172 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2173 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2176 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2177 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2178 bool X86::isMOVHLPSMask(SDNode *N) {
2179 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2181 if (N->getNumOperands() != 4)
2184 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2185 return isUndefOrEqual(N->getOperand(0), 6) &&
2186 isUndefOrEqual(N->getOperand(1), 7) &&
2187 isUndefOrEqual(N->getOperand(2), 2) &&
2188 isUndefOrEqual(N->getOperand(3), 3);
2191 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2192 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2194 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2195 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2197 if (N->getNumOperands() != 4)
2200 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2201 return isUndefOrEqual(N->getOperand(0), 2) &&
2202 isUndefOrEqual(N->getOperand(1), 3) &&
2203 isUndefOrEqual(N->getOperand(2), 2) &&
2204 isUndefOrEqual(N->getOperand(3), 3);
2207 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2208 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2209 bool X86::isMOVLPMask(SDNode *N) {
2210 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2212 unsigned NumElems = N->getNumOperands();
2213 if (NumElems != 2 && NumElems != 4)
2216 for (unsigned i = 0; i < NumElems/2; ++i)
2217 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2220 for (unsigned i = NumElems/2; i < NumElems; ++i)
2221 if (!isUndefOrEqual(N->getOperand(i), i))
2227 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2228 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2230 bool X86::isMOVHPMask(SDNode *N) {
2231 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2233 unsigned NumElems = N->getNumOperands();
2234 if (NumElems != 2 && NumElems != 4)
2237 for (unsigned i = 0; i < NumElems/2; ++i)
2238 if (!isUndefOrEqual(N->getOperand(i), i))
2241 for (unsigned i = 0; i < NumElems/2; ++i) {
2242 SDValue Arg = N->getOperand(i + NumElems/2);
2243 if (!isUndefOrEqual(Arg, i + NumElems))
2250 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2251 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2252 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2253 bool V2IsSplat = false) {
2254 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2257 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2258 SDValue BitI = Elts[i];
2259 SDValue BitI1 = Elts[i+1];
2260 if (!isUndefOrEqual(BitI, j))
2263 if (isUndefOrEqual(BitI1, NumElts))
2266 if (!isUndefOrEqual(BitI1, j + NumElts))
2274 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2275 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2276 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2279 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2280 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2281 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2282 bool V2IsSplat = false) {
2283 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2286 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2287 SDValue BitI = Elts[i];
2288 SDValue BitI1 = Elts[i+1];
2289 if (!isUndefOrEqual(BitI, j + NumElts/2))
2292 if (isUndefOrEqual(BitI1, NumElts))
2295 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2303 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2304 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2305 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2308 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2309 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2311 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2312 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2314 unsigned NumElems = N->getNumOperands();
2315 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2318 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2319 SDValue BitI = N->getOperand(i);
2320 SDValue BitI1 = N->getOperand(i+1);
2322 if (!isUndefOrEqual(BitI, j))
2324 if (!isUndefOrEqual(BitI1, j))
2331 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2332 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2334 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2335 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2337 unsigned NumElems = N->getNumOperands();
2338 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2341 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2342 SDValue BitI = N->getOperand(i);
2343 SDValue BitI1 = N->getOperand(i + 1);
2345 if (!isUndefOrEqual(BitI, j))
2347 if (!isUndefOrEqual(BitI1, j))
2354 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2355 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2356 /// MOVSD, and MOVD, i.e. setting the lowest element.
2357 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2358 if (NumElts != 2 && NumElts != 4)
2361 if (!isUndefOrEqual(Elts[0], NumElts))
2364 for (unsigned i = 1; i < NumElts; ++i) {
2365 if (!isUndefOrEqual(Elts[i], i))
2372 bool X86::isMOVLMask(SDNode *N) {
2373 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2374 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2377 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2378 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2379 /// element of vector 2 and the other elements to come from vector 1 in order.
2380 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2381 bool V2IsSplat = false,
2382 bool V2IsUndef = false) {
2383 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2386 if (!isUndefOrEqual(Ops[0], 0))
2389 for (unsigned i = 1; i < NumOps; ++i) {
2390 SDValue Arg = Ops[i];
2391 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2392 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2393 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2400 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2401 bool V2IsUndef = false) {
2402 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2403 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2404 V2IsSplat, V2IsUndef);
2407 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2408 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2409 bool X86::isMOVSHDUPMask(SDNode *N) {
2410 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2412 if (N->getNumOperands() != 4)
2415 // Expect 1, 1, 3, 3
2416 for (unsigned i = 0; i < 2; ++i) {
2417 SDValue Arg = N->getOperand(i);
2418 if (Arg.getOpcode() == ISD::UNDEF) continue;
2419 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2420 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2421 if (Val != 1) return false;
2425 for (unsigned i = 2; i < 4; ++i) {
2426 SDValue Arg = N->getOperand(i);
2427 if (Arg.getOpcode() == ISD::UNDEF) continue;
2428 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2429 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2430 if (Val != 3) return false;
2434 // Don't use movshdup if it can be done with a shufps.
2438 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2439 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2440 bool X86::isMOVSLDUPMask(SDNode *N) {
2441 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2443 if (N->getNumOperands() != 4)
2446 // Expect 0, 0, 2, 2
2447 for (unsigned i = 0; i < 2; ++i) {
2448 SDValue Arg = N->getOperand(i);
2449 if (Arg.getOpcode() == ISD::UNDEF) continue;
2450 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2451 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2452 if (Val != 0) return false;
2456 for (unsigned i = 2; i < 4; ++i) {
2457 SDValue Arg = N->getOperand(i);
2458 if (Arg.getOpcode() == ISD::UNDEF) continue;
2459 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2460 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2461 if (Val != 2) return false;
2465 // Don't use movshdup if it can be done with a shufps.
2469 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2470 /// specifies a identity operation on the LHS or RHS.
2471 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2472 unsigned NumElems = N->getNumOperands();
2473 for (unsigned i = 0; i < NumElems; ++i)
2474 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2479 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2480 /// a splat of a single element.
2481 static bool isSplatMask(SDNode *N) {
2482 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2484 // This is a splat operation if each element of the permute is the same, and
2485 // if the value doesn't reference the second vector.
2486 unsigned NumElems = N->getNumOperands();
2487 SDValue ElementBase;
2489 for (; i != NumElems; ++i) {
2490 SDValue Elt = N->getOperand(i);
2491 if (isa<ConstantSDNode>(Elt)) {
2497 if (!ElementBase.getNode())
2500 for (; i != NumElems; ++i) {
2501 SDValue Arg = N->getOperand(i);
2502 if (Arg.getOpcode() == ISD::UNDEF) continue;
2503 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2504 if (Arg != ElementBase) return false;
2507 // Make sure it is a splat of the first vector operand.
2508 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2511 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2512 /// a splat of a single element and it's a 2 or 4 element mask.
2513 bool X86::isSplatMask(SDNode *N) {
2514 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2516 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2517 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2519 return ::isSplatMask(N);
2522 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2523 /// specifies a splat of zero element.
2524 bool X86::isSplatLoMask(SDNode *N) {
2525 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2527 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2528 if (!isUndefOrEqual(N->getOperand(i), 0))
2533 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2534 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2535 bool X86::isMOVDDUPMask(SDNode *N) {
2536 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2538 unsigned e = N->getNumOperands() / 2;
2539 for (unsigned i = 0; i < e; ++i)
2540 if (!isUndefOrEqual(N->getOperand(i), i))
2542 for (unsigned i = 0; i < e; ++i)
2543 if (!isUndefOrEqual(N->getOperand(e+i), i))
2548 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2549 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2551 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2552 unsigned NumOperands = N->getNumOperands();
2553 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2555 for (unsigned i = 0; i < NumOperands; ++i) {
2557 SDValue Arg = N->getOperand(NumOperands-i-1);
2558 if (Arg.getOpcode() != ISD::UNDEF)
2559 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2560 if (Val >= NumOperands) Val -= NumOperands;
2562 if (i != NumOperands - 1)
2569 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2570 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2572 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2574 // 8 nodes, but we only care about the last 4.
2575 for (unsigned i = 7; i >= 4; --i) {
2577 SDValue Arg = N->getOperand(i);
2578 if (Arg.getOpcode() != ISD::UNDEF)
2579 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2588 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2589 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2591 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2593 // 8 nodes, but we only care about the first 4.
2594 for (int i = 3; i >= 0; --i) {
2596 SDValue Arg = N->getOperand(i);
2597 if (Arg.getOpcode() != ISD::UNDEF)
2598 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2607 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2608 /// specifies a 8 element shuffle that can be broken into a pair of
2609 /// PSHUFHW and PSHUFLW.
2610 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2611 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2613 if (N->getNumOperands() != 8)
2616 // Lower quadword shuffled.
2617 for (unsigned i = 0; i != 4; ++i) {
2618 SDValue Arg = N->getOperand(i);
2619 if (Arg.getOpcode() == ISD::UNDEF) continue;
2620 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2621 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2626 // Upper quadword shuffled.
2627 for (unsigned i = 4; i != 8; ++i) {
2628 SDValue Arg = N->getOperand(i);
2629 if (Arg.getOpcode() == ISD::UNDEF) continue;
2630 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2631 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2632 if (Val < 4 || Val > 7)
2639 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2640 /// values in ther permute mask.
2641 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2642 SDValue &V2, SDValue &Mask,
2643 SelectionDAG &DAG) {
2644 MVT VT = Op.getValueType();
2645 MVT MaskVT = Mask.getValueType();
2646 MVT EltVT = MaskVT.getVectorElementType();
2647 unsigned NumElems = Mask.getNumOperands();
2648 SmallVector<SDValue, 8> MaskVec;
2650 for (unsigned i = 0; i != NumElems; ++i) {
2651 SDValue Arg = Mask.getOperand(i);
2652 if (Arg.getOpcode() == ISD::UNDEF) {
2653 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2656 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2657 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2659 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2661 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2665 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2666 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2669 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2670 /// the two vector operands have swapped position.
2672 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2673 MVT MaskVT = Mask.getValueType();
2674 MVT EltVT = MaskVT.getVectorElementType();
2675 unsigned NumElems = Mask.getNumOperands();
2676 SmallVector<SDValue, 8> MaskVec;
2677 for (unsigned i = 0; i != NumElems; ++i) {
2678 SDValue Arg = Mask.getOperand(i);
2679 if (Arg.getOpcode() == ISD::UNDEF) {
2680 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2683 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2684 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2686 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2688 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2690 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2694 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2695 /// match movhlps. The lower half elements should come from upper half of
2696 /// V1 (and in order), and the upper half elements should come from the upper
2697 /// half of V2 (and in order).
2698 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2699 unsigned NumElems = Mask->getNumOperands();
2702 for (unsigned i = 0, e = 2; i != e; ++i)
2703 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2705 for (unsigned i = 2; i != 4; ++i)
2706 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2711 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2712 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2714 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2715 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2717 N = N->getOperand(0).getNode();
2718 if (!ISD::isNON_EXTLoad(N))
2721 *LD = cast<LoadSDNode>(N);
2725 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2726 /// match movlp{s|d}. The lower half elements should come from lower half of
2727 /// V1 (and in order), and the upper half elements should come from the upper
2728 /// half of V2 (and in order). And since V1 will become the source of the
2729 /// MOVLP, it must be either a vector load or a scalar load to vector.
2730 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2731 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2733 // Is V2 is a vector load, don't do this transformation. We will try to use
2734 // load folding shufps op.
2735 if (ISD::isNON_EXTLoad(V2))
2738 unsigned NumElems = Mask->getNumOperands();
2739 if (NumElems != 2 && NumElems != 4)
2741 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2742 if (!isUndefOrEqual(Mask->getOperand(i), i))
2744 for (unsigned i = NumElems/2; i != NumElems; ++i)
2745 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2750 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2752 static bool isSplatVector(SDNode *N) {
2753 if (N->getOpcode() != ISD::BUILD_VECTOR)
2756 SDValue SplatValue = N->getOperand(0);
2757 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2758 if (N->getOperand(i) != SplatValue)
2763 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2765 static bool isUndefShuffle(SDNode *N) {
2766 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2769 SDValue V1 = N->getOperand(0);
2770 SDValue V2 = N->getOperand(1);
2771 SDValue Mask = N->getOperand(2);
2772 unsigned NumElems = Mask.getNumOperands();
2773 for (unsigned i = 0; i != NumElems; ++i) {
2774 SDValue Arg = Mask.getOperand(i);
2775 if (Arg.getOpcode() != ISD::UNDEF) {
2776 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2777 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2779 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2786 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2788 static inline bool isZeroNode(SDValue Elt) {
2789 return ((isa<ConstantSDNode>(Elt) &&
2790 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2791 (isa<ConstantFPSDNode>(Elt) &&
2792 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2795 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2796 /// to an zero vector.
2797 static bool isZeroShuffle(SDNode *N) {
2798 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2801 SDValue V1 = N->getOperand(0);
2802 SDValue V2 = N->getOperand(1);
2803 SDValue Mask = N->getOperand(2);
2804 unsigned NumElems = Mask.getNumOperands();
2805 for (unsigned i = 0; i != NumElems; ++i) {
2806 SDValue Arg = Mask.getOperand(i);
2807 if (Arg.getOpcode() == ISD::UNDEF)
2810 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2811 if (Idx < NumElems) {
2812 unsigned Opc = V1.getNode()->getOpcode();
2813 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2815 if (Opc != ISD::BUILD_VECTOR ||
2816 !isZeroNode(V1.getNode()->getOperand(Idx)))
2818 } else if (Idx >= NumElems) {
2819 unsigned Opc = V2.getNode()->getOpcode();
2820 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2822 if (Opc != ISD::BUILD_VECTOR ||
2823 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2830 /// getZeroVector - Returns a vector of specified type with all zero elements.
2832 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2833 assert(VT.isVector() && "Expected a vector type");
2835 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2836 // type. This ensures they get CSE'd.
2838 if (VT.getSizeInBits() == 64) { // MMX
2839 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2840 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2841 } else if (HasSSE2) { // SSE2
2842 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2843 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2845 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2846 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2848 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2851 /// getOnesVector - Returns a vector of specified type with all bits set.
2853 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2854 assert(VT.isVector() && "Expected a vector type");
2856 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2857 // type. This ensures they get CSE'd.
2858 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2860 if (VT.getSizeInBits() == 64) // MMX
2861 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2863 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2864 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2868 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2869 /// that point to V2 points to its first element.
2870 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2871 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2873 bool Changed = false;
2874 SmallVector<SDValue, 8> MaskVec;
2875 unsigned NumElems = Mask.getNumOperands();
2876 for (unsigned i = 0; i != NumElems; ++i) {
2877 SDValue Arg = Mask.getOperand(i);
2878 if (Arg.getOpcode() != ISD::UNDEF) {
2879 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2880 if (Val > NumElems) {
2881 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2885 MaskVec.push_back(Arg);
2889 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2890 &MaskVec[0], MaskVec.size());
2894 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2895 /// operation of specified width.
2896 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2897 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2898 MVT BaseVT = MaskVT.getVectorElementType();
2900 SmallVector<SDValue, 8> MaskVec;
2901 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2902 for (unsigned i = 1; i != NumElems; ++i)
2903 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2904 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2907 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2908 /// of specified width.
2909 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2910 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2911 MVT BaseVT = MaskVT.getVectorElementType();
2912 SmallVector<SDValue, 8> MaskVec;
2913 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2914 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2915 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2917 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2920 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2921 /// of specified width.
2922 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2923 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2924 MVT BaseVT = MaskVT.getVectorElementType();
2925 unsigned Half = NumElems/2;
2926 SmallVector<SDValue, 8> MaskVec;
2927 for (unsigned i = 0; i != Half; ++i) {
2928 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2929 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2931 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2934 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2935 /// element #0 of a vector with the specified index, leaving the rest of the
2936 /// elements in place.
2937 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2938 SelectionDAG &DAG) {
2939 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2940 MVT BaseVT = MaskVT.getVectorElementType();
2941 SmallVector<SDValue, 8> MaskVec;
2942 // Element #0 of the result gets the elt we are replacing.
2943 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2944 for (unsigned i = 1; i != NumElems; ++i)
2945 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2946 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2949 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2950 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2951 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2952 MVT VT = Op.getValueType();
2955 SDValue V1 = Op.getOperand(0);
2956 SDValue Mask = Op.getOperand(2);
2957 unsigned NumElems = Mask.getNumOperands();
2958 // Special handling of v4f32 -> v4i32.
2959 if (VT != MVT::v4f32) {
2960 Mask = getUnpacklMask(NumElems, DAG);
2961 while (NumElems > 4) {
2962 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2965 Mask = getZeroVector(MVT::v4i32, true, DAG);
2968 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
2969 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
2970 DAG.getNode(ISD::UNDEF, PVT), Mask);
2971 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2974 /// isVectorLoad - Returns true if the node is a vector load, a scalar
2975 /// load that's promoted to vector, or a load bitcasted.
2976 static bool isVectorLoad(SDValue Op) {
2977 assert(Op.getValueType().isVector() && "Expected a vector type");
2978 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
2979 Op.getOpcode() == ISD::BIT_CONVERT) {
2980 return isa<LoadSDNode>(Op.getOperand(0));
2982 return isa<LoadSDNode>(Op);
2986 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
2988 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
2989 SelectionDAG &DAG, bool HasSSE3) {
2990 // If we have sse3 and shuffle has more than one use or input is a load, then
2991 // use movddup. Otherwise, use movlhps.
2992 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
2993 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
2994 MVT VT = Op.getValueType();
2997 unsigned NumElems = PVT.getVectorNumElements();
2998 if (NumElems == 2) {
2999 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3000 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
3002 assert(NumElems == 4);
3003 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3004 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3005 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst0, Cst1, Cst0, Cst1);
3008 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3009 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3010 DAG.getNode(ISD::UNDEF, PVT), Mask);
3011 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3014 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3015 /// vector of zero or undef vector. This produces a shuffle where the low
3016 /// element of V2 is swizzled into the zero/undef vector, landing at element
3017 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3018 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3019 bool isZero, bool HasSSE2,
3020 SelectionDAG &DAG) {
3021 MVT VT = V2.getValueType();
3023 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
3024 unsigned NumElems = V2.getValueType().getVectorNumElements();
3025 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3026 MVT EVT = MaskVT.getVectorElementType();
3027 SmallVector<SDValue, 16> MaskVec;
3028 for (unsigned i = 0; i != NumElems; ++i)
3029 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3030 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3032 MaskVec.push_back(DAG.getConstant(i, EVT));
3033 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3034 &MaskVec[0], MaskVec.size());
3035 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3038 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3039 /// a shuffle that is zero.
3041 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3042 unsigned NumElems, bool Low,
3043 SelectionDAG &DAG) {
3044 unsigned NumZeros = 0;
3045 for (unsigned i = 0; i < NumElems; ++i) {
3046 unsigned Index = Low ? i : NumElems-i-1;
3047 SDValue Idx = Mask.getOperand(Index);
3048 if (Idx.getOpcode() == ISD::UNDEF) {
3052 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3053 if (Elt.getNode() && isZeroNode(Elt))
3061 /// isVectorShift - Returns true if the shuffle can be implemented as a
3062 /// logical left or right shift of a vector.
3063 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3064 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3065 unsigned NumElems = Mask.getNumOperands();
3068 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3071 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3076 bool SeenV1 = false;
3077 bool SeenV2 = false;
3078 for (unsigned i = NumZeros; i < NumElems; ++i) {
3079 unsigned Val = isLeft ? (i - NumZeros) : i;
3080 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3081 if (Idx.getOpcode() == ISD::UNDEF)
3083 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3084 if (Index < NumElems)
3093 if (SeenV1 && SeenV2)
3096 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3102 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3104 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3105 unsigned NumNonZero, unsigned NumZero,
3106 SelectionDAG &DAG, TargetLowering &TLI) {
3112 for (unsigned i = 0; i < 16; ++i) {
3113 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3114 if (ThisIsNonZero && First) {
3116 V = getZeroVector(MVT::v8i16, true, DAG);
3118 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3123 SDValue ThisElt(0, 0), LastElt(0, 0);
3124 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3125 if (LastIsNonZero) {
3126 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3128 if (ThisIsNonZero) {
3129 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3130 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3131 ThisElt, DAG.getConstant(8, MVT::i8));
3133 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3137 if (ThisElt.getNode())
3138 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3139 DAG.getIntPtrConstant(i/2));
3143 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3146 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3148 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3149 unsigned NumNonZero, unsigned NumZero,
3150 SelectionDAG &DAG, TargetLowering &TLI) {
3156 for (unsigned i = 0; i < 8; ++i) {
3157 bool isNonZero = (NonZeros & (1 << i)) != 0;
3161 V = getZeroVector(MVT::v8i16, true, DAG);
3163 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3166 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3167 DAG.getIntPtrConstant(i));
3174 /// getVShift - Return a vector logical shift node.
3176 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3177 unsigned NumBits, SelectionDAG &DAG,
3178 const TargetLowering &TLI) {
3179 bool isMMX = VT.getSizeInBits() == 64;
3180 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3181 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3182 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3183 return DAG.getNode(ISD::BIT_CONVERT, VT,
3184 DAG.getNode(Opc, ShVT, SrcOp,
3185 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3189 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3190 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3191 if (ISD::isBuildVectorAllZeros(Op.getNode())
3192 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3193 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3194 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3195 // eliminated on x86-32 hosts.
3196 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3199 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3200 return getOnesVector(Op.getValueType(), DAG);
3201 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3204 MVT VT = Op.getValueType();
3205 MVT EVT = VT.getVectorElementType();
3206 unsigned EVTBits = EVT.getSizeInBits();
3208 unsigned NumElems = Op.getNumOperands();
3209 unsigned NumZero = 0;
3210 unsigned NumNonZero = 0;
3211 unsigned NonZeros = 0;
3212 bool IsAllConstants = true;
3213 SmallSet<SDValue, 8> Values;
3214 for (unsigned i = 0; i < NumElems; ++i) {
3215 SDValue Elt = Op.getOperand(i);
3216 if (Elt.getOpcode() == ISD::UNDEF)
3219 if (Elt.getOpcode() != ISD::Constant &&
3220 Elt.getOpcode() != ISD::ConstantFP)
3221 IsAllConstants = false;
3222 if (isZeroNode(Elt))
3225 NonZeros |= (1 << i);
3230 if (NumNonZero == 0) {
3231 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3232 return DAG.getNode(ISD::UNDEF, VT);
3235 // Special case for single non-zero, non-undef, element.
3236 if (NumNonZero == 1 && NumElems <= 4) {
3237 unsigned Idx = CountTrailingZeros_32(NonZeros);
3238 SDValue Item = Op.getOperand(Idx);
3240 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3241 // the value are obviously zero, truncate the value to i32 and do the
3242 // insertion that way. Only do this if the value is non-constant or if the
3243 // value is a constant being inserted into element 0. It is cheaper to do
3244 // a constant pool load than it is to do a movd + shuffle.
3245 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3246 (!IsAllConstants || Idx == 0)) {
3247 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3248 // Handle MMX and SSE both.
3249 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3250 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3252 // Truncate the value (which may itself be a constant) to i32, and
3253 // convert it to a vector with movd (S2V+shuffle to zero extend).
3254 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3255 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3256 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3257 Subtarget->hasSSE2(), DAG);
3259 // Now we have our 32-bit value zero extended in the low element of
3260 // a vector. If Idx != 0, swizzle it into place.
3263 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3264 getSwapEltZeroMask(VecElts, Idx, DAG)
3266 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3268 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3272 // If we have a constant or non-constant insertion into the low element of
3273 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3274 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3275 // depending on what the source datatype is. Because we can only get here
3276 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3278 // Don't do this for i64 values on x86-32.
3279 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3280 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3281 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3282 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3283 Subtarget->hasSSE2(), DAG);
3286 // Is it a vector logical left shift?
3287 if (NumElems == 2 && Idx == 1 &&
3288 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3289 unsigned NumBits = VT.getSizeInBits();
3290 return getVShift(true, VT,
3291 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3292 NumBits/2, DAG, *this);
3295 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3298 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3299 // is a non-constant being inserted into an element other than the low one,
3300 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3301 // movd/movss) to move this into the low element, then shuffle it into
3303 if (EVTBits == 32) {
3304 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3306 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3307 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3308 Subtarget->hasSSE2(), DAG);
3309 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3310 MVT MaskEVT = MaskVT.getVectorElementType();
3311 SmallVector<SDValue, 8> MaskVec;
3312 for (unsigned i = 0; i < NumElems; i++)
3313 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3314 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3315 &MaskVec[0], MaskVec.size());
3316 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3317 DAG.getNode(ISD::UNDEF, VT), Mask);
3321 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3322 if (Values.size() == 1)
3325 // A vector full of immediates; various special cases are already
3326 // handled, so this is best done with a single constant-pool load.
3330 // Let legalizer expand 2-wide build_vectors.
3331 if (EVTBits == 64) {
3332 if (NumNonZero == 1) {
3333 // One half is zero or undef.
3334 unsigned Idx = CountTrailingZeros_32(NonZeros);
3335 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3336 Op.getOperand(Idx));
3337 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3338 Subtarget->hasSSE2(), DAG);
3343 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3344 if (EVTBits == 8 && NumElems == 16) {
3345 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3347 if (V.getNode()) return V;
3350 if (EVTBits == 16 && NumElems == 8) {
3351 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3353 if (V.getNode()) return V;
3356 // If element VT is == 32 bits, turn it into a number of shuffles.
3357 SmallVector<SDValue, 8> V;
3359 if (NumElems == 4 && NumZero > 0) {
3360 for (unsigned i = 0; i < 4; ++i) {
3361 bool isZero = !(NonZeros & (1 << i));
3363 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3365 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3368 for (unsigned i = 0; i < 2; ++i) {
3369 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3372 V[i] = V[i*2]; // Must be a zero vector.
3375 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3376 getMOVLMask(NumElems, DAG));
3379 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3380 getMOVLMask(NumElems, DAG));
3383 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3384 getUnpacklMask(NumElems, DAG));
3389 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3390 MVT EVT = MaskVT.getVectorElementType();
3391 SmallVector<SDValue, 8> MaskVec;
3392 bool Reverse = (NonZeros & 0x3) == 2;
3393 for (unsigned i = 0; i < 2; ++i)
3395 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3397 MaskVec.push_back(DAG.getConstant(i, EVT));
3398 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3399 for (unsigned i = 0; i < 2; ++i)
3401 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3403 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3404 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3405 &MaskVec[0], MaskVec.size());
3406 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3409 if (Values.size() > 2) {
3410 // Expand into a number of unpckl*.
3412 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3413 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3414 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3415 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3416 for (unsigned i = 0; i < NumElems; ++i)
3417 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3419 while (NumElems != 0) {
3420 for (unsigned i = 0; i < NumElems; ++i)
3421 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3432 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3433 SDValue PermMask, SelectionDAG &DAG,
3434 TargetLowering &TLI) {
3436 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3437 MVT MaskEVT = MaskVT.getVectorElementType();
3438 MVT PtrVT = TLI.getPointerTy();
3439 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3440 PermMask.getNode()->op_end());
3442 // First record which half of which vector the low elements come from.
3443 SmallVector<unsigned, 4> LowQuad(4);
3444 for (unsigned i = 0; i < 4; ++i) {
3445 SDValue Elt = MaskElts[i];
3446 if (Elt.getOpcode() == ISD::UNDEF)
3448 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3449 int QuadIdx = EltIdx / 4;
3453 int BestLowQuad = -1;
3454 unsigned MaxQuad = 1;
3455 for (unsigned i = 0; i < 4; ++i) {
3456 if (LowQuad[i] > MaxQuad) {
3458 MaxQuad = LowQuad[i];
3462 // Record which half of which vector the high elements come from.
3463 SmallVector<unsigned, 4> HighQuad(4);
3464 for (unsigned i = 4; i < 8; ++i) {
3465 SDValue Elt = MaskElts[i];
3466 if (Elt.getOpcode() == ISD::UNDEF)
3468 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3469 int QuadIdx = EltIdx / 4;
3470 ++HighQuad[QuadIdx];
3473 int BestHighQuad = -1;
3475 for (unsigned i = 0; i < 4; ++i) {
3476 if (HighQuad[i] > MaxQuad) {
3478 MaxQuad = HighQuad[i];
3482 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3483 if (BestLowQuad != -1 || BestHighQuad != -1) {
3484 // First sort the 4 chunks in order using shufpd.
3485 SmallVector<SDValue, 8> MaskVec;
3487 if (BestLowQuad != -1)
3488 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3490 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3492 if (BestHighQuad != -1)
3493 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3495 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3497 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3498 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3499 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3500 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3501 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3503 // Now sort high and low parts separately.
3504 BitVector InOrder(8);
3505 if (BestLowQuad != -1) {
3506 // Sort lower half in order using PSHUFLW.
3508 bool AnyOutOrder = false;
3510 for (unsigned i = 0; i != 4; ++i) {
3511 SDValue Elt = MaskElts[i];
3512 if (Elt.getOpcode() == ISD::UNDEF) {
3513 MaskVec.push_back(Elt);
3516 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3520 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3522 // If this element is in the right place after this shuffle, then
3524 if ((int)(EltIdx / 4) == BestLowQuad)
3529 for (unsigned i = 4; i != 8; ++i)
3530 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3531 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3532 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3536 if (BestHighQuad != -1) {
3537 // Sort high half in order using PSHUFHW if possible.
3540 for (unsigned i = 0; i != 4; ++i)
3541 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3543 bool AnyOutOrder = false;
3544 for (unsigned i = 4; i != 8; ++i) {
3545 SDValue Elt = MaskElts[i];
3546 if (Elt.getOpcode() == ISD::UNDEF) {
3547 MaskVec.push_back(Elt);
3550 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3554 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3556 // If this element is in the right place after this shuffle, then
3558 if ((int)(EltIdx / 4) == BestHighQuad)
3564 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3565 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3569 // The other elements are put in the right place using pextrw and pinsrw.
3570 for (unsigned i = 0; i != 8; ++i) {
3573 SDValue Elt = MaskElts[i];
3574 if (Elt.getOpcode() == ISD::UNDEF)
3576 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3577 SDValue ExtOp = (EltIdx < 8)
3578 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3579 DAG.getConstant(EltIdx, PtrVT))
3580 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3581 DAG.getConstant(EltIdx - 8, PtrVT));
3582 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3583 DAG.getConstant(i, PtrVT));
3589 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3590 // few as possible. First, let's find out how many elements are already in the
3592 unsigned V1InOrder = 0;
3593 unsigned V1FromV1 = 0;
3594 unsigned V2InOrder = 0;
3595 unsigned V2FromV2 = 0;
3596 SmallVector<SDValue, 8> V1Elts;
3597 SmallVector<SDValue, 8> V2Elts;
3598 for (unsigned i = 0; i < 8; ++i) {
3599 SDValue Elt = MaskElts[i];
3600 if (Elt.getOpcode() == ISD::UNDEF) {
3601 V1Elts.push_back(Elt);
3602 V2Elts.push_back(Elt);
3607 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3609 V1Elts.push_back(Elt);
3610 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3612 } else if (EltIdx == i+8) {
3613 V1Elts.push_back(Elt);
3614 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3616 } else if (EltIdx < 8) {
3617 V1Elts.push_back(Elt);
3620 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3625 if (V2InOrder > V1InOrder) {
3626 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3628 std::swap(V1Elts, V2Elts);
3629 std::swap(V1FromV1, V2FromV2);
3632 if ((V1FromV1 + V1InOrder) != 8) {
3633 // Some elements are from V2.
3635 // If there are elements that are from V1 but out of place,
3636 // then first sort them in place
3637 SmallVector<SDValue, 8> MaskVec;
3638 for (unsigned i = 0; i < 8; ++i) {
3639 SDValue Elt = V1Elts[i];
3640 if (Elt.getOpcode() == ISD::UNDEF) {
3641 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3644 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3646 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3648 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3650 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3651 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3655 for (unsigned i = 0; i < 8; ++i) {
3656 SDValue Elt = V1Elts[i];
3657 if (Elt.getOpcode() == ISD::UNDEF)
3659 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3662 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3663 DAG.getConstant(EltIdx - 8, PtrVT));
3664 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3665 DAG.getConstant(i, PtrVT));
3669 // All elements are from V1.
3671 for (unsigned i = 0; i < 8; ++i) {
3672 SDValue Elt = V1Elts[i];
3673 if (Elt.getOpcode() == ISD::UNDEF)
3675 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3676 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3677 DAG.getConstant(EltIdx, PtrVT));
3678 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3679 DAG.getConstant(i, PtrVT));
3685 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3686 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3687 /// done when every pair / quad of shuffle mask elements point to elements in
3688 /// the right sequence. e.g.
3689 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3691 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3693 SDValue PermMask, SelectionDAG &DAG,
3694 TargetLowering &TLI) {
3695 unsigned NumElems = PermMask.getNumOperands();
3696 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3697 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3698 MVT MaskEltVT = MaskVT.getVectorElementType();
3700 switch (VT.getSimpleVT()) {
3701 default: assert(false && "Unexpected!");
3702 case MVT::v4f32: NewVT = MVT::v2f64; break;
3703 case MVT::v4i32: NewVT = MVT::v2i64; break;
3704 case MVT::v8i16: NewVT = MVT::v4i32; break;
3705 case MVT::v16i8: NewVT = MVT::v4i32; break;
3708 if (NewWidth == 2) {
3714 unsigned Scale = NumElems / NewWidth;
3715 SmallVector<SDValue, 8> MaskVec;
3716 for (unsigned i = 0; i < NumElems; i += Scale) {
3717 unsigned StartIdx = ~0U;
3718 for (unsigned j = 0; j < Scale; ++j) {
3719 SDValue Elt = PermMask.getOperand(i+j);
3720 if (Elt.getOpcode() == ISD::UNDEF)
3722 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3723 if (StartIdx == ~0U)
3724 StartIdx = EltIdx - (EltIdx % Scale);
3725 if (EltIdx != StartIdx + j)
3728 if (StartIdx == ~0U)
3729 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3731 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3734 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3735 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3736 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3737 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3738 &MaskVec[0], MaskVec.size()));
3741 /// getVZextMovL - Return a zero-extending vector move low node.
3743 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3744 SDValue SrcOp, SelectionDAG &DAG,
3745 const X86Subtarget *Subtarget) {
3746 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3747 LoadSDNode *LD = NULL;
3748 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3749 LD = dyn_cast<LoadSDNode>(SrcOp);
3751 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3753 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3754 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3755 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3756 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3757 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3759 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3760 return DAG.getNode(ISD::BIT_CONVERT, VT,
3761 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3762 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3769 return DAG.getNode(ISD::BIT_CONVERT, VT,
3770 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3771 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3774 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3777 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3778 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3779 MVT MaskVT = PermMask.getValueType();
3780 MVT MaskEVT = MaskVT.getVectorElementType();
3781 SmallVector<std::pair<int, int>, 8> Locs;
3783 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3786 for (unsigned i = 0; i != 4; ++i) {
3787 SDValue Elt = PermMask.getOperand(i);
3788 if (Elt.getOpcode() == ISD::UNDEF) {
3789 Locs[i] = std::make_pair(-1, -1);
3791 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3792 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3794 Locs[i] = std::make_pair(0, NumLo);
3798 Locs[i] = std::make_pair(1, NumHi);
3800 Mask1[2+NumHi] = Elt;
3806 if (NumLo <= 2 && NumHi <= 2) {
3807 // If no more than two elements come from either vector. This can be
3808 // implemented with two shuffles. First shuffle gather the elements.
3809 // The second shuffle, which takes the first shuffle as both of its
3810 // vector operands, put the elements into the right order.
3811 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3812 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3813 &Mask1[0], Mask1.size()));
3815 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3816 for (unsigned i = 0; i != 4; ++i) {
3817 if (Locs[i].first == -1)
3820 unsigned Idx = (i < 2) ? 0 : 4;
3821 Idx += Locs[i].first * 2 + Locs[i].second;
3822 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3826 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3827 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3828 &Mask2[0], Mask2.size()));
3829 } else if (NumLo == 3 || NumHi == 3) {
3830 // Otherwise, we must have three elements from one vector, call it X, and
3831 // one element from the other, call it Y. First, use a shufps to build an
3832 // intermediate vector with the one element from Y and the element from X
3833 // that will be in the same half in the final destination (the indexes don't
3834 // matter). Then, use a shufps to build the final vector, taking the half
3835 // containing the element from Y from the intermediate, and the other half
3838 // Normalize it so the 3 elements come from V1.
3839 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3843 // Find the element from V2.
3845 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3846 SDValue Elt = PermMask.getOperand(HiIndex);
3847 if (Elt.getOpcode() == ISD::UNDEF)
3849 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3854 Mask1[0] = PermMask.getOperand(HiIndex);
3855 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3856 Mask1[2] = PermMask.getOperand(HiIndex^1);
3857 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3858 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3859 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3862 Mask1[0] = PermMask.getOperand(0);
3863 Mask1[1] = PermMask.getOperand(1);
3864 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3865 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3866 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3867 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3869 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3870 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3871 Mask1[2] = PermMask.getOperand(2);
3872 Mask1[3] = PermMask.getOperand(3);
3873 if (Mask1[2].getOpcode() != ISD::UNDEF)
3875 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
3877 if (Mask1[3].getOpcode() != ISD::UNDEF)
3879 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
3881 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3882 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3886 // Break it into (shuffle shuffle_hi, shuffle_lo).
3888 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3889 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3890 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3891 unsigned MaskIdx = 0;
3894 for (unsigned i = 0; i != 4; ++i) {
3901 SDValue Elt = PermMask.getOperand(i);
3902 if (Elt.getOpcode() == ISD::UNDEF) {
3903 Locs[i] = std::make_pair(-1, -1);
3904 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
3905 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3906 (*MaskPtr)[LoIdx] = Elt;
3909 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3910 (*MaskPtr)[HiIdx] = Elt;
3915 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3916 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3917 &LoMask[0], LoMask.size()));
3918 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3919 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3920 &HiMask[0], HiMask.size()));
3921 SmallVector<SDValue, 8> MaskOps;
3922 for (unsigned i = 0; i != 4; ++i) {
3923 if (Locs[i].first == -1) {
3924 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3926 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3927 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3930 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3931 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3932 &MaskOps[0], MaskOps.size()));
3936 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3937 SDValue V1 = Op.getOperand(0);
3938 SDValue V2 = Op.getOperand(1);
3939 SDValue PermMask = Op.getOperand(2);
3940 MVT VT = Op.getValueType();
3941 unsigned NumElems = PermMask.getNumOperands();
3942 bool isMMX = VT.getSizeInBits() == 64;
3943 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3944 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3945 bool V1IsSplat = false;
3946 bool V2IsSplat = false;
3948 if (isUndefShuffle(Op.getNode()))
3949 return DAG.getNode(ISD::UNDEF, VT);
3951 if (isZeroShuffle(Op.getNode()))
3952 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3954 if (isIdentityMask(PermMask.getNode()))
3956 else if (isIdentityMask(PermMask.getNode(), true))
3959 // Canonicalize movddup shuffles.
3960 if (V2IsUndef && Subtarget->hasSSE2() &&
3961 VT.getSizeInBits() == 128 &&
3962 X86::isMOVDDUPMask(PermMask.getNode()))
3963 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
3965 if (isSplatMask(PermMask.getNode())) {
3966 if (isMMX || NumElems < 4) return Op;
3967 // Promote it to a v4{if}32 splat.
3968 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
3971 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3973 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3974 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3975 if (NewOp.getNode())
3976 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3977 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3978 // FIXME: Figure out a cleaner way to do this.
3979 // Try to make use of movq to zero out the top part.
3980 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3981 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3983 if (NewOp.getNode()) {
3984 SDValue NewV1 = NewOp.getOperand(0);
3985 SDValue NewV2 = NewOp.getOperand(1);
3986 SDValue NewMask = NewOp.getOperand(2);
3987 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
3988 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3989 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
3992 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3993 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
3995 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
3996 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4001 // Check if this can be converted into a logical shift.
4002 bool isLeft = false;
4005 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4006 if (isShift && ShVal.hasOneUse()) {
4007 // If the shifted value has multiple uses, it may be cheaper to use
4008 // v_set0 + movlhps or movhlps, etc.
4009 MVT EVT = VT.getVectorElementType();
4010 ShAmt *= EVT.getSizeInBits();
4011 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4014 if (X86::isMOVLMask(PermMask.getNode())) {
4017 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4018 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
4023 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4024 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4025 X86::isMOVHLPSMask(PermMask.getNode()) ||
4026 X86::isMOVHPMask(PermMask.getNode()) ||
4027 X86::isMOVLPMask(PermMask.getNode())))
4030 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4031 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4032 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4035 // No better options. Use a vshl / vsrl.
4036 MVT EVT = VT.getVectorElementType();
4037 ShAmt *= EVT.getSizeInBits();
4038 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4041 bool Commuted = false;
4042 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4043 // 1,1,1,1 -> v8i16 though.
4044 V1IsSplat = isSplatVector(V1.getNode());
4045 V2IsSplat = isSplatVector(V2.getNode());
4047 // Canonicalize the splat or undef, if present, to be on the RHS.
4048 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4049 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4050 std::swap(V1IsSplat, V2IsSplat);
4051 std::swap(V1IsUndef, V2IsUndef);
4055 // FIXME: Figure out a cleaner way to do this.
4056 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4057 if (V2IsUndef) return V1;
4058 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4060 // V2 is a splat, so the mask may be malformed. That is, it may point
4061 // to any V2 element. The instruction selectior won't like this. Get
4062 // a corrected mask and commute to form a proper MOVS{S|D}.
4063 SDValue NewMask = getMOVLMask(NumElems, DAG);
4064 if (NewMask.getNode() != PermMask.getNode())
4065 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4070 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4071 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4072 X86::isUNPCKLMask(PermMask.getNode()) ||
4073 X86::isUNPCKHMask(PermMask.getNode()))
4077 // Normalize mask so all entries that point to V2 points to its first
4078 // element then try to match unpck{h|l} again. If match, return a
4079 // new vector_shuffle with the corrected mask.
4080 SDValue NewMask = NormalizeMask(PermMask, DAG);
4081 if (NewMask.getNode() != PermMask.getNode()) {
4082 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4083 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4084 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4085 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4086 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4087 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4092 // Normalize the node to match x86 shuffle ops if needed
4093 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4094 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4097 // Commute is back and try unpck* again.
4098 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4099 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4100 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4101 X86::isUNPCKLMask(PermMask.getNode()) ||
4102 X86::isUNPCKHMask(PermMask.getNode()))
4106 // Try PSHUF* first, then SHUFP*.
4107 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4108 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4109 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4110 if (V2.getOpcode() != ISD::UNDEF)
4111 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4112 DAG.getNode(ISD::UNDEF, VT), PermMask);
4117 if (Subtarget->hasSSE2() &&
4118 (X86::isPSHUFDMask(PermMask.getNode()) ||
4119 X86::isPSHUFHWMask(PermMask.getNode()) ||
4120 X86::isPSHUFLWMask(PermMask.getNode()))) {
4122 if (VT == MVT::v4f32) {
4124 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4125 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4126 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4127 } else if (V2.getOpcode() != ISD::UNDEF)
4128 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4129 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4131 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4135 // Binary or unary shufps.
4136 if (X86::isSHUFPMask(PermMask.getNode()) ||
4137 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4141 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4142 if (VT == MVT::v8i16) {
4143 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4144 if (NewOp.getNode())
4148 // Handle all 4 wide cases with a number of shuffles except for MMX.
4149 if (NumElems == 4 && !isMMX)
4150 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4156 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4157 SelectionDAG &DAG) {
4158 MVT VT = Op.getValueType();
4159 if (VT.getSizeInBits() == 8) {
4160 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4161 Op.getOperand(0), Op.getOperand(1));
4162 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4163 DAG.getValueType(VT));
4164 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4165 } else if (VT.getSizeInBits() == 16) {
4166 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4167 Op.getOperand(0), Op.getOperand(1));
4168 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4169 DAG.getValueType(VT));
4170 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4171 } else if (VT == MVT::f32) {
4172 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4173 // the result back to FR32 register. It's only worth matching if the
4174 // result has a single use which is a store or a bitcast to i32.
4175 if (!Op.hasOneUse())
4177 SDNode *User = *Op.getNode()->use_begin();
4178 if (User->getOpcode() != ISD::STORE &&
4179 (User->getOpcode() != ISD::BIT_CONVERT ||
4180 User->getValueType(0) != MVT::i32))
4182 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4183 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4185 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4192 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4193 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4196 if (Subtarget->hasSSE41()) {
4197 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4202 MVT VT = Op.getValueType();
4203 // TODO: handle v16i8.
4204 if (VT.getSizeInBits() == 16) {
4205 SDValue Vec = Op.getOperand(0);
4206 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4208 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4209 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4210 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4212 // Transform it so it match pextrw which produces a 32-bit result.
4213 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4214 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4215 Op.getOperand(0), Op.getOperand(1));
4216 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4217 DAG.getValueType(VT));
4218 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4219 } else if (VT.getSizeInBits() == 32) {
4220 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4223 // SHUFPS the element to the lowest double word, then movss.
4224 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4225 SmallVector<SDValue, 8> IdxVec;
4227 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4229 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4231 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4233 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4234 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4235 &IdxVec[0], IdxVec.size());
4236 SDValue Vec = Op.getOperand(0);
4237 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4238 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4239 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4240 DAG.getIntPtrConstant(0));
4241 } else if (VT.getSizeInBits() == 64) {
4242 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4243 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4244 // to match extract_elt for f64.
4245 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4249 // UNPCKHPD the element to the lowest double word, then movsd.
4250 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4251 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4252 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4253 SmallVector<SDValue, 8> IdxVec;
4254 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4256 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4257 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4258 &IdxVec[0], IdxVec.size());
4259 SDValue Vec = Op.getOperand(0);
4260 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4261 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4262 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4263 DAG.getIntPtrConstant(0));
4270 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4271 MVT VT = Op.getValueType();
4272 MVT EVT = VT.getVectorElementType();
4274 SDValue N0 = Op.getOperand(0);
4275 SDValue N1 = Op.getOperand(1);
4276 SDValue N2 = Op.getOperand(2);
4278 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4279 isa<ConstantSDNode>(N2)) {
4280 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4282 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4284 if (N1.getValueType() != MVT::i32)
4285 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4286 if (N2.getValueType() != MVT::i32)
4287 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4288 return DAG.getNode(Opc, VT, N0, N1, N2);
4289 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4290 // Bits [7:6] of the constant are the source select. This will always be
4291 // zero here. The DAG Combiner may combine an extract_elt index into these
4292 // bits. For example (insert (extract, 3), 2) could be matched by putting
4293 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4294 // Bits [5:4] of the constant are the destination select. This is the
4295 // value of the incoming immediate.
4296 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4297 // combine either bitwise AND or insert of float 0.0 to set these bits.
4298 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4299 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4305 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4306 MVT VT = Op.getValueType();
4307 MVT EVT = VT.getVectorElementType();
4309 if (Subtarget->hasSSE41())
4310 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4315 SDValue N0 = Op.getOperand(0);
4316 SDValue N1 = Op.getOperand(1);
4317 SDValue N2 = Op.getOperand(2);
4319 if (EVT.getSizeInBits() == 16) {
4320 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4321 // as its second argument.
4322 if (N1.getValueType() != MVT::i32)
4323 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4324 if (N2.getValueType() != MVT::i32)
4325 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4326 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4332 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4333 if (Op.getValueType() == MVT::v2f32)
4334 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4335 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4336 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4337 Op.getOperand(0))));
4339 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4340 MVT VT = MVT::v2i32;
4341 switch (Op.getValueType().getSimpleVT()) {
4348 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4349 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4352 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4353 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4354 // one of the above mentioned nodes. It has to be wrapped because otherwise
4355 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4356 // be used to form addressing mode. These wrapped nodes will be selected
4359 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4360 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4361 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4363 CP->getAlignment());
4364 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4365 // With PIC, the address is actually $g + Offset.
4366 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4367 !Subtarget->isPICStyleRIPRel()) {
4368 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4369 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4377 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV,
4378 SelectionDAG &DAG) const {
4379 SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4380 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4381 // With PIC, the address is actually $g + Offset.
4382 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4383 !Subtarget->isPICStyleRIPRel()) {
4384 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4385 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4389 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4390 // load the value at address GV, not the value of GV itself. This means that
4391 // the GlobalAddress must be in the base or index register of the address, not
4392 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4393 // The same applies for external symbols during PIC codegen
4394 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4395 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4396 PseudoSourceValue::getGOT(), 0);
4402 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4403 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4404 return LowerGlobalAddress(GV, DAG);
4407 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4409 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4412 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4413 DAG.getNode(X86ISD::GlobalBaseReg,
4415 InFlag = Chain.getValue(1);
4417 // emit leal symbol@TLSGD(,%ebx,1), %eax
4418 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4419 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4420 GA->getValueType(0),
4422 SDValue Ops[] = { Chain, TGA, InFlag };
4423 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4424 InFlag = Result.getValue(2);
4425 Chain = Result.getValue(1);
4427 // call ___tls_get_addr. This function receives its argument in
4428 // the register EAX.
4429 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4430 InFlag = Chain.getValue(1);
4432 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4433 SDValue Ops1[] = { Chain,
4434 DAG.getTargetExternalSymbol("___tls_get_addr",
4436 DAG.getRegister(X86::EAX, PtrVT),
4437 DAG.getRegister(X86::EBX, PtrVT),
4439 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4440 InFlag = Chain.getValue(1);
4442 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4445 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4447 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4449 SDValue InFlag, Chain;
4451 // emit leaq symbol@TLSGD(%rip), %rdi
4452 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4453 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4454 GA->getValueType(0),
4456 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4457 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4458 Chain = Result.getValue(1);
4459 InFlag = Result.getValue(2);
4461 // call __tls_get_addr. This function receives its argument in
4462 // the register RDI.
4463 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4464 InFlag = Chain.getValue(1);
4466 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4467 SDValue Ops1[] = { Chain,
4468 DAG.getTargetExternalSymbol("__tls_get_addr",
4470 DAG.getRegister(X86::RDI, PtrVT),
4472 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4473 InFlag = Chain.getValue(1);
4475 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4478 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4479 // "local exec" model.
4480 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4482 // Get the Thread Pointer
4483 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4484 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4486 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4487 GA->getValueType(0),
4489 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4491 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4492 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4493 PseudoSourceValue::getGOT(), 0);
4495 // The address of the thread local variable is the add of the thread
4496 // pointer with the offset of the variable.
4497 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4501 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4502 // TODO: implement the "local dynamic" model
4503 // TODO: implement the "initial exec"model for pic executables
4504 assert(Subtarget->isTargetELF() &&
4505 "TLS not implemented for non-ELF targets");
4506 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4507 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4508 // otherwise use the "Local Exec"TLS Model
4509 if (Subtarget->is64Bit()) {
4510 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4512 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4513 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4515 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4520 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4521 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4522 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4523 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4524 // With PIC, the address is actually $g + Offset.
4525 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4526 !Subtarget->isPICStyleRIPRel()) {
4527 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4528 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4535 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4536 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4537 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4538 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4539 // With PIC, the address is actually $g + Offset.
4540 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4541 !Subtarget->isPICStyleRIPRel()) {
4542 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4543 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4550 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4551 /// take a 2 x i32 value to shift plus a shift amount.
4552 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4553 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4554 MVT VT = Op.getValueType();
4555 unsigned VTBits = VT.getSizeInBits();
4556 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4557 SDValue ShOpLo = Op.getOperand(0);
4558 SDValue ShOpHi = Op.getOperand(1);
4559 SDValue ShAmt = Op.getOperand(2);
4560 SDValue Tmp1 = isSRA ?
4561 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4562 DAG.getConstant(0, VT);
4565 if (Op.getOpcode() == ISD::SHL_PARTS) {
4566 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4567 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4569 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4570 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4573 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4574 DAG.getConstant(VTBits, MVT::i8));
4575 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4576 AndNode, DAG.getConstant(0, MVT::i8));
4579 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4580 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4581 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4583 if (Op.getOpcode() == ISD::SHL_PARTS) {
4584 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4585 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4587 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4588 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4591 SDValue Ops[2] = { Lo, Hi };
4592 return DAG.getMergeValues(Ops, 2);
4595 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4596 MVT SrcVT = Op.getOperand(0).getValueType();
4597 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4598 "Unknown SINT_TO_FP to lower!");
4600 // These are really Legal; caller falls through into that case.
4601 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4603 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4604 Subtarget->is64Bit())
4607 unsigned Size = SrcVT.getSizeInBits()/8;
4608 MachineFunction &MF = DAG.getMachineFunction();
4609 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4610 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4611 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4613 PseudoSourceValue::getFixedStack(SSFI), 0);
4617 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4619 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4621 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4622 SmallVector<SDValue, 8> Ops;
4623 Ops.push_back(Chain);
4624 Ops.push_back(StackSlot);
4625 Ops.push_back(DAG.getValueType(SrcVT));
4626 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4627 Tys, &Ops[0], Ops.size());
4630 Chain = Result.getValue(1);
4631 SDValue InFlag = Result.getValue(2);
4633 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4634 // shouldn't be necessary except that RFP cannot be live across
4635 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4636 MachineFunction &MF = DAG.getMachineFunction();
4637 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4638 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4639 Tys = DAG.getVTList(MVT::Other);
4640 SmallVector<SDValue, 8> Ops;
4641 Ops.push_back(Chain);
4642 Ops.push_back(Result);
4643 Ops.push_back(StackSlot);
4644 Ops.push_back(DAG.getValueType(Op.getValueType()));
4645 Ops.push_back(InFlag);
4646 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4647 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4648 PseudoSourceValue::getFixedStack(SSFI), 0);
4654 std::pair<SDValue,SDValue> X86TargetLowering::
4655 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4656 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4657 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4658 "Unknown FP_TO_SINT to lower!");
4660 // These are really Legal.
4661 if (Op.getValueType() == MVT::i32 &&
4662 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4663 return std::make_pair(SDValue(), SDValue());
4664 if (Subtarget->is64Bit() &&
4665 Op.getValueType() == MVT::i64 &&
4666 Op.getOperand(0).getValueType() != MVT::f80)
4667 return std::make_pair(SDValue(), SDValue());
4669 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4671 MachineFunction &MF = DAG.getMachineFunction();
4672 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4673 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4674 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4676 switch (Op.getValueType().getSimpleVT()) {
4677 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4678 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4679 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4680 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4683 SDValue Chain = DAG.getEntryNode();
4684 SDValue Value = Op.getOperand(0);
4685 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4686 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4687 Chain = DAG.getStore(Chain, Value, StackSlot,
4688 PseudoSourceValue::getFixedStack(SSFI), 0);
4689 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4691 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4693 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4694 Chain = Value.getValue(1);
4695 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4696 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4699 // Build the FP_TO_INT*_IN_MEM
4700 SDValue Ops[] = { Chain, Value, StackSlot };
4701 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4703 return std::make_pair(FIST, StackSlot);
4706 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4707 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4708 SDValue FIST = Vals.first, StackSlot = Vals.second;
4709 if (FIST.getNode() == 0) return SDValue();
4712 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4715 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4716 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
4717 SDValue FIST = Vals.first, StackSlot = Vals.second;
4718 if (FIST.getNode() == 0) return 0;
4720 MVT VT = N->getValueType(0);
4722 // Return a load from the stack slot.
4723 SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);
4725 // Use MERGE_VALUES to drop the chain result value and get a node with one
4726 // result. This requires turning off getMergeValues simplification, since
4727 // otherwise it will give us Res back.
4728 return DAG.getMergeValues(&Res, 1, false).getNode();
4731 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4732 MVT VT = Op.getValueType();
4735 EltVT = VT.getVectorElementType();
4736 std::vector<Constant*> CV;
4737 if (EltVT == MVT::f64) {
4738 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4742 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4748 Constant *C = ConstantVector::get(CV);
4749 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4750 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4751 PseudoSourceValue::getConstantPool(), 0,
4753 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4756 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4757 MVT VT = Op.getValueType();
4759 unsigned EltNum = 1;
4760 if (VT.isVector()) {
4761 EltVT = VT.getVectorElementType();
4762 EltNum = VT.getVectorNumElements();
4764 std::vector<Constant*> CV;
4765 if (EltVT == MVT::f64) {
4766 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4770 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4776 Constant *C = ConstantVector::get(CV);
4777 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4778 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4779 PseudoSourceValue::getConstantPool(), 0,
4781 if (VT.isVector()) {
4782 return DAG.getNode(ISD::BIT_CONVERT, VT,
4783 DAG.getNode(ISD::XOR, MVT::v2i64,
4784 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4785 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4787 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4791 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4792 SDValue Op0 = Op.getOperand(0);
4793 SDValue Op1 = Op.getOperand(1);
4794 MVT VT = Op.getValueType();
4795 MVT SrcVT = Op1.getValueType();
4797 // If second operand is smaller, extend it first.
4798 if (SrcVT.bitsLT(VT)) {
4799 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4802 // And if it is bigger, shrink it first.
4803 if (SrcVT.bitsGT(VT)) {
4804 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4808 // At this point the operands and the result should have the same
4809 // type, and that won't be f80 since that is not custom lowered.
4811 // First get the sign bit of second operand.
4812 std::vector<Constant*> CV;
4813 if (SrcVT == MVT::f64) {
4814 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4815 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4817 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4818 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4819 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4820 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4822 Constant *C = ConstantVector::get(CV);
4823 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4824 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
4825 PseudoSourceValue::getConstantPool(), 0,
4827 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4829 // Shift sign bit right or left if the two operands have different types.
4830 if (SrcVT.bitsGT(VT)) {
4831 // Op0 is MVT::f32, Op1 is MVT::f64.
4832 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4833 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4834 DAG.getConstant(32, MVT::i32));
4835 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4836 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4837 DAG.getIntPtrConstant(0));
4840 // Clear first operand sign bit.
4842 if (VT == MVT::f64) {
4843 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
4844 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4846 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
4847 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4848 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4849 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
4851 C = ConstantVector::get(CV);
4852 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4853 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4854 PseudoSourceValue::getConstantPool(), 0,
4856 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4858 // Or the value with the sign bit.
4859 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4862 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
4863 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4865 SDValue Op0 = Op.getOperand(0);
4866 SDValue Op1 = Op.getOperand(1);
4867 SDValue CC = Op.getOperand(2);
4868 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4869 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4872 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4874 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4875 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4876 DAG.getConstant(X86CC, MVT::i8), Cond);
4879 assert(isFP && "Illegal integer SetCC!");
4881 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4882 switch (SetCCOpcode) {
4883 default: assert(false && "Illegal floating point SetCC!");
4884 case ISD::SETOEQ: { // !PF & ZF
4885 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4886 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4887 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4888 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4889 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4891 case ISD::SETUNE: { // PF | !ZF
4892 SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4893 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4894 SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4895 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4896 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4901 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4903 SDValue Op0 = Op.getOperand(0);
4904 SDValue Op1 = Op.getOperand(1);
4905 SDValue CC = Op.getOperand(2);
4906 MVT VT = Op.getValueType();
4907 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4908 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
4912 MVT VT0 = Op0.getValueType();
4913 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
4914 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
4917 switch (SetCCOpcode) {
4920 case ISD::SETEQ: SSECC = 0; break;
4922 case ISD::SETGT: Swap = true; // Fallthrough
4924 case ISD::SETOLT: SSECC = 1; break;
4926 case ISD::SETGE: Swap = true; // Fallthrough
4928 case ISD::SETOLE: SSECC = 2; break;
4929 case ISD::SETUO: SSECC = 3; break;
4931 case ISD::SETNE: SSECC = 4; break;
4932 case ISD::SETULE: Swap = true;
4933 case ISD::SETUGE: SSECC = 5; break;
4934 case ISD::SETULT: Swap = true;
4935 case ISD::SETUGT: SSECC = 6; break;
4936 case ISD::SETO: SSECC = 7; break;
4939 std::swap(Op0, Op1);
4941 // In the two special cases we can't handle, emit two comparisons.
4943 if (SetCCOpcode == ISD::SETUEQ) {
4945 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
4946 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
4947 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
4949 else if (SetCCOpcode == ISD::SETONE) {
4951 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
4952 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
4953 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
4955 assert(0 && "Illegal FP comparison");
4957 // Handle all other FP comparisons here.
4958 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
4961 // We are handling one of the integer comparisons here. Since SSE only has
4962 // GT and EQ comparisons for integer, swapping operands and multiple
4963 // operations may be required for some comparisons.
4964 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
4965 bool Swap = false, Invert = false, FlipSigns = false;
4967 switch (VT.getSimpleVT()) {
4969 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
4970 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
4971 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
4972 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
4975 switch (SetCCOpcode) {
4977 case ISD::SETNE: Invert = true;
4978 case ISD::SETEQ: Opc = EQOpc; break;
4979 case ISD::SETLT: Swap = true;
4980 case ISD::SETGT: Opc = GTOpc; break;
4981 case ISD::SETGE: Swap = true;
4982 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
4983 case ISD::SETULT: Swap = true;
4984 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
4985 case ISD::SETUGE: Swap = true;
4986 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
4989 std::swap(Op0, Op1);
4991 // Since SSE has no unsigned integer comparisons, we need to flip the sign
4992 // bits of the inputs before performing those operations.
4994 MVT EltVT = VT.getVectorElementType();
4995 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
4996 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
4997 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
4999 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
5000 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
5003 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
5005 // If the logical-not of the result is required, perform that now.
5007 MVT EltVT = VT.getVectorElementType();
5008 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
5009 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
5010 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
5012 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
5017 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5018 bool addTest = true;
5019 SDValue Cond = Op.getOperand(0);
5022 if (Cond.getOpcode() == ISD::SETCC)
5023 Cond = LowerSETCC(Cond, DAG);
5025 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5026 // setting operand in place of the X86ISD::SETCC.
5027 if (Cond.getOpcode() == X86ISD::SETCC) {
5028 CC = Cond.getOperand(0);
5030 SDValue Cmp = Cond.getOperand(1);
5031 unsigned Opc = Cmp.getOpcode();
5032 MVT VT = Op.getValueType();
5034 bool IllegalFPCMov = false;
5035 if (VT.isFloatingPoint() && !VT.isVector() &&
5036 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5037 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5039 if ((Opc == X86ISD::CMP ||
5040 Opc == X86ISD::COMI ||
5041 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
5048 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5049 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5052 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5054 SmallVector<SDValue, 4> Ops;
5055 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5056 // condition is true.
5057 Ops.push_back(Op.getOperand(2));
5058 Ops.push_back(Op.getOperand(1));
5060 Ops.push_back(Cond);
5061 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
5064 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5065 bool addTest = true;
5066 SDValue Chain = Op.getOperand(0);
5067 SDValue Cond = Op.getOperand(1);
5068 SDValue Dest = Op.getOperand(2);
5071 if (Cond.getOpcode() == ISD::SETCC)
5072 Cond = LowerSETCC(Cond, DAG);
5074 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5075 // setting operand in place of the X86ISD::SETCC.
5076 if (Cond.getOpcode() == X86ISD::SETCC) {
5077 CC = Cond.getOperand(0);
5079 SDValue Cmp = Cond.getOperand(1);
5080 unsigned Opc = Cmp.getOpcode();
5081 if (Opc == X86ISD::CMP ||
5082 Opc == X86ISD::COMI ||
5083 Opc == X86ISD::UCOMI) {
5090 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5091 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5093 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5094 Chain, Op.getOperand(2), CC, Cond);
5098 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5099 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5100 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5101 // that the guard pages used by the OS virtual memory manager are allocated in
5102 // correct sequence.
5104 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5105 SelectionDAG &DAG) {
5106 assert(Subtarget->isTargetCygMing() &&
5107 "This should be used only on Cygwin/Mingw targets");
5110 SDValue Chain = Op.getOperand(0);
5111 SDValue Size = Op.getOperand(1);
5112 // FIXME: Ensure alignment here
5116 MVT IntPtr = getPointerTy();
5117 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5119 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5121 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5122 Flag = Chain.getValue(1);
5124 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5125 SDValue Ops[] = { Chain,
5126 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5127 DAG.getRegister(X86::EAX, IntPtr),
5128 DAG.getRegister(X86StackPtr, SPTy),
5130 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5131 Flag = Chain.getValue(1);
5133 Chain = DAG.getCALLSEQ_END(Chain,
5134 DAG.getIntPtrConstant(0, true),
5135 DAG.getIntPtrConstant(0, true),
5138 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5140 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5141 return DAG.getMergeValues(Ops1, 2);
5145 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5147 SDValue Dst, SDValue Src,
5148 SDValue Size, unsigned Align,
5150 uint64_t DstSVOff) {
5151 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5153 // If not DWORD aligned or size is more than the threshold, call the library.
5154 // The libc version is likely to be faster for these cases. It can use the
5155 // address value and run time information about the CPU.
5156 if ((Align & 3) != 0 ||
5158 ConstantSize->getZExtValue() >
5159 getSubtarget()->getMaxInlineSizeThreshold()) {
5160 SDValue InFlag(0, 0);
5162 // Check to see if there is a specialized entry-point for memory zeroing.
5163 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5165 if (const char *bzeroEntry = V &&
5166 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5167 MVT IntPtr = getPointerTy();
5168 const Type *IntPtrTy = TD->getIntPtrType();
5169 TargetLowering::ArgListTy Args;
5170 TargetLowering::ArgListEntry Entry;
5172 Entry.Ty = IntPtrTy;
5173 Args.push_back(Entry);
5175 Args.push_back(Entry);
5176 std::pair<SDValue,SDValue> CallResult =
5177 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5178 CallingConv::C, false,
5179 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG);
5180 return CallResult.second;
5183 // Otherwise have the target-independent code call memset.
5187 uint64_t SizeVal = ConstantSize->getZExtValue();
5188 SDValue InFlag(0, 0);
5191 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5192 unsigned BytesLeft = 0;
5193 bool TwoRepStos = false;
5196 uint64_t Val = ValC->getZExtValue() & 255;
5198 // If the value is a constant, then we can potentially use larger sets.
5199 switch (Align & 3) {
5200 case 2: // WORD aligned
5203 Val = (Val << 8) | Val;
5205 case 0: // DWORD aligned
5208 Val = (Val << 8) | Val;
5209 Val = (Val << 16) | Val;
5210 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5213 Val = (Val << 32) | Val;
5216 default: // Byte aligned
5219 Count = DAG.getIntPtrConstant(SizeVal);
5223 if (AVT.bitsGT(MVT::i8)) {
5224 unsigned UBytes = AVT.getSizeInBits() / 8;
5225 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5226 BytesLeft = SizeVal % UBytes;
5229 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5231 InFlag = Chain.getValue(1);
5234 Count = DAG.getIntPtrConstant(SizeVal);
5235 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5236 InFlag = Chain.getValue(1);
5239 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5241 InFlag = Chain.getValue(1);
5242 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5244 InFlag = Chain.getValue(1);
5246 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5247 SmallVector<SDValue, 8> Ops;
5248 Ops.push_back(Chain);
5249 Ops.push_back(DAG.getValueType(AVT));
5250 Ops.push_back(InFlag);
5251 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5254 InFlag = Chain.getValue(1);
5256 MVT CVT = Count.getValueType();
5257 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5258 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5259 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5261 InFlag = Chain.getValue(1);
5262 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5264 Ops.push_back(Chain);
5265 Ops.push_back(DAG.getValueType(MVT::i8));
5266 Ops.push_back(InFlag);
5267 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5268 } else if (BytesLeft) {
5269 // Handle the last 1 - 7 bytes.
5270 unsigned Offset = SizeVal - BytesLeft;
5271 MVT AddrVT = Dst.getValueType();
5272 MVT SizeVT = Size.getValueType();
5274 Chain = DAG.getMemset(Chain,
5275 DAG.getNode(ISD::ADD, AddrVT, Dst,
5276 DAG.getConstant(Offset, AddrVT)),
5278 DAG.getConstant(BytesLeft, SizeVT),
5279 Align, DstSV, DstSVOff + Offset);
5282 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5287 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5288 SDValue Chain, SDValue Dst, SDValue Src,
5289 SDValue Size, unsigned Align,
5291 const Value *DstSV, uint64_t DstSVOff,
5292 const Value *SrcSV, uint64_t SrcSVOff) {
5293 // This requires the copy size to be a constant, preferrably
5294 // within a subtarget-specific limit.
5295 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5298 uint64_t SizeVal = ConstantSize->getZExtValue();
5299 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5302 /// If not DWORD aligned, call the library.
5303 if ((Align & 3) != 0)
5308 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5311 unsigned UBytes = AVT.getSizeInBits() / 8;
5312 unsigned CountVal = SizeVal / UBytes;
5313 SDValue Count = DAG.getIntPtrConstant(CountVal);
5314 unsigned BytesLeft = SizeVal % UBytes;
5316 SDValue InFlag(0, 0);
5317 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5319 InFlag = Chain.getValue(1);
5320 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5322 InFlag = Chain.getValue(1);
5323 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5325 InFlag = Chain.getValue(1);
5327 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5328 SmallVector<SDValue, 8> Ops;
5329 Ops.push_back(Chain);
5330 Ops.push_back(DAG.getValueType(AVT));
5331 Ops.push_back(InFlag);
5332 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5334 SmallVector<SDValue, 4> Results;
5335 Results.push_back(RepMovs);
5337 // Handle the last 1 - 7 bytes.
5338 unsigned Offset = SizeVal - BytesLeft;
5339 MVT DstVT = Dst.getValueType();
5340 MVT SrcVT = Src.getValueType();
5341 MVT SizeVT = Size.getValueType();
5342 Results.push_back(DAG.getMemcpy(Chain,
5343 DAG.getNode(ISD::ADD, DstVT, Dst,
5344 DAG.getConstant(Offset, DstVT)),
5345 DAG.getNode(ISD::ADD, SrcVT, Src,
5346 DAG.getConstant(Offset, SrcVT)),
5347 DAG.getConstant(BytesLeft, SizeVT),
5348 Align, AlwaysInline,
5349 DstSV, DstSVOff + Offset,
5350 SrcSV, SrcSVOff + Offset));
5353 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5356 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5357 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5358 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5359 SDValue TheChain = N->getOperand(0);
5360 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5361 if (Subtarget->is64Bit()) {
5362 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5363 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5364 MVT::i64, rax.getValue(2));
5365 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5366 DAG.getConstant(32, MVT::i8));
5368 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5371 return DAG.getMergeValues(Ops, 2).getNode();
5374 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5375 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5376 MVT::i32, eax.getValue(2));
5377 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5378 SDValue Ops[] = { eax, edx };
5379 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5381 // Use a MERGE_VALUES to return the value and chain.
5382 Ops[1] = edx.getValue(1);
5383 return DAG.getMergeValues(Ops, 2).getNode();
5386 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5387 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5389 if (!Subtarget->is64Bit()) {
5390 // vastart just stores the address of the VarArgsFrameIndex slot into the
5391 // memory location argument.
5392 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5393 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5397 // gp_offset (0 - 6 * 8)
5398 // fp_offset (48 - 48 + 8 * 16)
5399 // overflow_arg_area (point to parameters coming in memory).
5401 SmallVector<SDValue, 8> MemOps;
5402 SDValue FIN = Op.getOperand(1);
5404 SDValue Store = DAG.getStore(Op.getOperand(0),
5405 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5407 MemOps.push_back(Store);
5410 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5411 Store = DAG.getStore(Op.getOperand(0),
5412 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5414 MemOps.push_back(Store);
5416 // Store ptr to overflow_arg_area
5417 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5418 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5419 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5420 MemOps.push_back(Store);
5422 // Store ptr to reg_save_area.
5423 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5424 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5425 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5426 MemOps.push_back(Store);
5427 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5430 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5431 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5432 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5433 SDValue Chain = Op.getOperand(0);
5434 SDValue SrcPtr = Op.getOperand(1);
5435 SDValue SrcSV = Op.getOperand(2);
5437 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5442 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5443 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5444 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5445 SDValue Chain = Op.getOperand(0);
5446 SDValue DstPtr = Op.getOperand(1);
5447 SDValue SrcPtr = Op.getOperand(2);
5448 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5449 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5451 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5452 DAG.getIntPtrConstant(24), 8, false,
5453 DstSV, 0, SrcSV, 0);
5457 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5458 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5460 default: return SDValue(); // Don't custom lower most intrinsics.
5461 // Comparison intrinsics.
5462 case Intrinsic::x86_sse_comieq_ss:
5463 case Intrinsic::x86_sse_comilt_ss:
5464 case Intrinsic::x86_sse_comile_ss:
5465 case Intrinsic::x86_sse_comigt_ss:
5466 case Intrinsic::x86_sse_comige_ss:
5467 case Intrinsic::x86_sse_comineq_ss:
5468 case Intrinsic::x86_sse_ucomieq_ss:
5469 case Intrinsic::x86_sse_ucomilt_ss:
5470 case Intrinsic::x86_sse_ucomile_ss:
5471 case Intrinsic::x86_sse_ucomigt_ss:
5472 case Intrinsic::x86_sse_ucomige_ss:
5473 case Intrinsic::x86_sse_ucomineq_ss:
5474 case Intrinsic::x86_sse2_comieq_sd:
5475 case Intrinsic::x86_sse2_comilt_sd:
5476 case Intrinsic::x86_sse2_comile_sd:
5477 case Intrinsic::x86_sse2_comigt_sd:
5478 case Intrinsic::x86_sse2_comige_sd:
5479 case Intrinsic::x86_sse2_comineq_sd:
5480 case Intrinsic::x86_sse2_ucomieq_sd:
5481 case Intrinsic::x86_sse2_ucomilt_sd:
5482 case Intrinsic::x86_sse2_ucomile_sd:
5483 case Intrinsic::x86_sse2_ucomigt_sd:
5484 case Intrinsic::x86_sse2_ucomige_sd:
5485 case Intrinsic::x86_sse2_ucomineq_sd: {
5487 ISD::CondCode CC = ISD::SETCC_INVALID;
5490 case Intrinsic::x86_sse_comieq_ss:
5491 case Intrinsic::x86_sse2_comieq_sd:
5495 case Intrinsic::x86_sse_comilt_ss:
5496 case Intrinsic::x86_sse2_comilt_sd:
5500 case Intrinsic::x86_sse_comile_ss:
5501 case Intrinsic::x86_sse2_comile_sd:
5505 case Intrinsic::x86_sse_comigt_ss:
5506 case Intrinsic::x86_sse2_comigt_sd:
5510 case Intrinsic::x86_sse_comige_ss:
5511 case Intrinsic::x86_sse2_comige_sd:
5515 case Intrinsic::x86_sse_comineq_ss:
5516 case Intrinsic::x86_sse2_comineq_sd:
5520 case Intrinsic::x86_sse_ucomieq_ss:
5521 case Intrinsic::x86_sse2_ucomieq_sd:
5522 Opc = X86ISD::UCOMI;
5525 case Intrinsic::x86_sse_ucomilt_ss:
5526 case Intrinsic::x86_sse2_ucomilt_sd:
5527 Opc = X86ISD::UCOMI;
5530 case Intrinsic::x86_sse_ucomile_ss:
5531 case Intrinsic::x86_sse2_ucomile_sd:
5532 Opc = X86ISD::UCOMI;
5535 case Intrinsic::x86_sse_ucomigt_ss:
5536 case Intrinsic::x86_sse2_ucomigt_sd:
5537 Opc = X86ISD::UCOMI;
5540 case Intrinsic::x86_sse_ucomige_ss:
5541 case Intrinsic::x86_sse2_ucomige_sd:
5542 Opc = X86ISD::UCOMI;
5545 case Intrinsic::x86_sse_ucomineq_ss:
5546 case Intrinsic::x86_sse2_ucomineq_sd:
5547 Opc = X86ISD::UCOMI;
5553 SDValue LHS = Op.getOperand(1);
5554 SDValue RHS = Op.getOperand(2);
5555 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5557 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5558 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5559 DAG.getConstant(X86CC, MVT::i8), Cond);
5560 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5563 // Fix vector shift instructions where the last operand is a non-immediate
5565 case Intrinsic::x86_sse2_pslli_w:
5566 case Intrinsic::x86_sse2_pslli_d:
5567 case Intrinsic::x86_sse2_pslli_q:
5568 case Intrinsic::x86_sse2_psrli_w:
5569 case Intrinsic::x86_sse2_psrli_d:
5570 case Intrinsic::x86_sse2_psrli_q:
5571 case Intrinsic::x86_sse2_psrai_w:
5572 case Intrinsic::x86_sse2_psrai_d:
5573 case Intrinsic::x86_mmx_pslli_w:
5574 case Intrinsic::x86_mmx_pslli_d:
5575 case Intrinsic::x86_mmx_pslli_q:
5576 case Intrinsic::x86_mmx_psrli_w:
5577 case Intrinsic::x86_mmx_psrli_d:
5578 case Intrinsic::x86_mmx_psrli_q:
5579 case Intrinsic::x86_mmx_psrai_w:
5580 case Intrinsic::x86_mmx_psrai_d: {
5581 SDValue ShAmt = Op.getOperand(2);
5582 if (isa<ConstantSDNode>(ShAmt))
5585 unsigned NewIntNo = 0;
5586 MVT ShAmtVT = MVT::v4i32;
5588 case Intrinsic::x86_sse2_pslli_w:
5589 NewIntNo = Intrinsic::x86_sse2_psll_w;
5591 case Intrinsic::x86_sse2_pslli_d:
5592 NewIntNo = Intrinsic::x86_sse2_psll_d;
5594 case Intrinsic::x86_sse2_pslli_q:
5595 NewIntNo = Intrinsic::x86_sse2_psll_q;
5597 case Intrinsic::x86_sse2_psrli_w:
5598 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5600 case Intrinsic::x86_sse2_psrli_d:
5601 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5603 case Intrinsic::x86_sse2_psrli_q:
5604 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5606 case Intrinsic::x86_sse2_psrai_w:
5607 NewIntNo = Intrinsic::x86_sse2_psra_w;
5609 case Intrinsic::x86_sse2_psrai_d:
5610 NewIntNo = Intrinsic::x86_sse2_psra_d;
5613 ShAmtVT = MVT::v2i32;
5615 case Intrinsic::x86_mmx_pslli_w:
5616 NewIntNo = Intrinsic::x86_mmx_psll_w;
5618 case Intrinsic::x86_mmx_pslli_d:
5619 NewIntNo = Intrinsic::x86_mmx_psll_d;
5621 case Intrinsic::x86_mmx_pslli_q:
5622 NewIntNo = Intrinsic::x86_mmx_psll_q;
5624 case Intrinsic::x86_mmx_psrli_w:
5625 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5627 case Intrinsic::x86_mmx_psrli_d:
5628 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5630 case Intrinsic::x86_mmx_psrli_q:
5631 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5633 case Intrinsic::x86_mmx_psrai_w:
5634 NewIntNo = Intrinsic::x86_mmx_psra_w;
5636 case Intrinsic::x86_mmx_psrai_d:
5637 NewIntNo = Intrinsic::x86_mmx_psra_d;
5639 default: abort(); // Can't reach here.
5644 MVT VT = Op.getValueType();
5645 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5646 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5647 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5648 DAG.getConstant(NewIntNo, MVT::i32),
5649 Op.getOperand(1), ShAmt);
5654 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5655 // Depths > 0 not supported yet!
5656 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5659 // Just load the return address
5660 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5661 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5664 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5665 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5666 MFI->setFrameAddressIsTaken(true);
5667 MVT VT = Op.getValueType();
5668 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5669 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
5670 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), FrameReg, VT);
5672 FrameAddr = DAG.getLoad(VT, DAG.getEntryNode(), FrameAddr, NULL, 0);
5676 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5677 SelectionDAG &DAG) {
5678 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5681 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5683 MachineFunction &MF = DAG.getMachineFunction();
5684 SDValue Chain = Op.getOperand(0);
5685 SDValue Offset = Op.getOperand(1);
5686 SDValue Handler = Op.getOperand(2);
5688 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
5690 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
5692 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5693 DAG.getIntPtrConstant(-TD->getPointerSize()));
5694 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5695 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5696 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
5697 MF.getRegInfo().addLiveOut(StoreAddrReg);
5699 return DAG.getNode(X86ISD::EH_RETURN,
5701 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
5704 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
5705 SelectionDAG &DAG) {
5706 SDValue Root = Op.getOperand(0);
5707 SDValue Trmp = Op.getOperand(1); // trampoline
5708 SDValue FPtr = Op.getOperand(2); // nested function
5709 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
5711 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5713 const X86InstrInfo *TII =
5714 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5716 if (Subtarget->is64Bit()) {
5717 SDValue OutChains[6];
5719 // Large code-model.
5721 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
5722 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
5724 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
5725 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
5727 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
5729 // Load the pointer to the nested function into R11.
5730 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
5731 SDValue Addr = Trmp;
5732 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5735 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
5736 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
5738 // Load the 'nest' parameter value into R10.
5739 // R10 is specified in X86CallingConv.td
5740 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
5741 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
5742 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5745 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
5746 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
5748 // Jump to the nested function.
5749 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
5750 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
5751 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
5754 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
5755 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
5756 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
5760 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
5761 return DAG.getMergeValues(Ops, 2);
5763 const Function *Func =
5764 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5765 unsigned CC = Func->getCallingConv();
5770 assert(0 && "Unsupported calling convention");
5771 case CallingConv::C:
5772 case CallingConv::X86_StdCall: {
5773 // Pass 'nest' parameter in ECX.
5774 // Must be kept in sync with X86CallingConv.td
5777 // Check that ECX wasn't needed by an 'inreg' parameter.
5778 const FunctionType *FTy = Func->getFunctionType();
5779 const AttrListPtr &Attrs = Func->getAttributes();
5781 if (!Attrs.isEmpty() && !Func->isVarArg()) {
5782 unsigned InRegCount = 0;
5785 for (FunctionType::param_iterator I = FTy->param_begin(),
5786 E = FTy->param_end(); I != E; ++I, ++Idx)
5787 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
5788 // FIXME: should only count parameters that are lowered to integers.
5789 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
5791 if (InRegCount > 2) {
5792 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5798 case CallingConv::X86_FastCall:
5799 case CallingConv::Fast:
5800 // Pass 'nest' parameter in EAX.
5801 // Must be kept in sync with X86CallingConv.td
5806 SDValue OutChains[4];
5809 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5810 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5812 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5813 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
5814 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5817 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5818 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
5820 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5821 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5822 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5823 TrmpAddr, 5, false, 1);
5825 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5826 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
5829 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5830 return DAG.getMergeValues(Ops, 2);
5834 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
5836 The rounding mode is in bits 11:10 of FPSR, and has the following
5843 FLT_ROUNDS, on the other hand, expects the following:
5850 To perform the conversion, we do:
5851 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5854 MachineFunction &MF = DAG.getMachineFunction();
5855 const TargetMachine &TM = MF.getTarget();
5856 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5857 unsigned StackAlignment = TFI.getStackAlignment();
5858 MVT VT = Op.getValueType();
5860 // Save FP Control Word to stack slot
5861 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5862 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5864 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5865 DAG.getEntryNode(), StackSlot);
5867 // Load FP Control Word from stack slot
5868 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5870 // Transform as necessary
5872 DAG.getNode(ISD::SRL, MVT::i16,
5873 DAG.getNode(ISD::AND, MVT::i16,
5874 CWD, DAG.getConstant(0x800, MVT::i16)),
5875 DAG.getConstant(11, MVT::i8));
5877 DAG.getNode(ISD::SRL, MVT::i16,
5878 DAG.getNode(ISD::AND, MVT::i16,
5879 CWD, DAG.getConstant(0x400, MVT::i16)),
5880 DAG.getConstant(9, MVT::i8));
5883 DAG.getNode(ISD::AND, MVT::i16,
5884 DAG.getNode(ISD::ADD, MVT::i16,
5885 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5886 DAG.getConstant(1, MVT::i16)),
5887 DAG.getConstant(3, MVT::i16));
5890 return DAG.getNode((VT.getSizeInBits() < 16 ?
5891 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5894 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
5895 MVT VT = Op.getValueType();
5897 unsigned NumBits = VT.getSizeInBits();
5899 Op = Op.getOperand(0);
5900 if (VT == MVT::i8) {
5901 // Zero extend to i32 since there is not an i8 bsr.
5903 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5906 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5907 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5908 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5910 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5911 SmallVector<SDValue, 4> Ops;
5913 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5914 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5915 Ops.push_back(Op.getValue(1));
5916 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5918 // Finally xor with NumBits-1.
5919 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5922 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5926 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
5927 MVT VT = Op.getValueType();
5929 unsigned NumBits = VT.getSizeInBits();
5931 Op = Op.getOperand(0);
5932 if (VT == MVT::i8) {
5934 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5937 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5938 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5939 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5941 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5942 SmallVector<SDValue, 4> Ops;
5944 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5945 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5946 Ops.push_back(Op.getValue(1));
5947 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5950 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5954 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
5955 MVT T = Op.getValueType();
5958 switch(T.getSimpleVT()) {
5960 assert(false && "Invalid value type!");
5961 case MVT::i8: Reg = X86::AL; size = 1; break;
5962 case MVT::i16: Reg = X86::AX; size = 2; break;
5963 case MVT::i32: Reg = X86::EAX; size = 4; break;
5965 if (Subtarget->is64Bit()) {
5966 Reg = X86::RAX; size = 8;
5967 } else //Should go away when LowerType stuff lands
5968 return SDValue(ExpandATOMIC_CMP_SWAP(Op.getNode(), DAG), 0);
5971 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
5972 Op.getOperand(2), SDValue());
5973 SDValue Ops[] = { cpIn.getValue(0),
5976 DAG.getTargetConstant(size, MVT::i8),
5978 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5979 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
5981 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
5985 SDNode* X86TargetLowering::ExpandATOMIC_CMP_SWAP(SDNode* Op,
5986 SelectionDAG &DAG) {
5987 MVT T = Op->getValueType(0);
5988 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
5989 SDValue cpInL, cpInH;
5990 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5991 DAG.getConstant(0, MVT::i32));
5992 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(2),
5993 DAG.getConstant(1, MVT::i32));
5994 cpInL = DAG.getCopyToReg(Op->getOperand(0), X86::EAX,
5996 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX,
5997 cpInH, cpInL.getValue(1));
5998 SDValue swapInL, swapInH;
5999 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
6000 DAG.getConstant(0, MVT::i32));
6001 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op->getOperand(3),
6002 DAG.getConstant(1, MVT::i32));
6003 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX,
6004 swapInL, cpInH.getValue(1));
6005 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX,
6006 swapInH, swapInL.getValue(1));
6007 SDValue Ops[] = { swapInH.getValue(0),
6009 swapInH.getValue(1) };
6010 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6011 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
6012 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
6013 Result.getValue(1));
6014 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
6015 cpOutL.getValue(2));
6016 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6017 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
6018 SDValue Vals[2] = { ResultVal, cpOutH.getValue(1) };
6019 return DAG.getMergeValues(Vals, 2).getNode();
6022 SDValue X86TargetLowering::LowerATOMIC_BINARY_64(SDValue Op,
6025 SDNode *Node = Op.getNode();
6026 MVT T = Node->getValueType(0);
6027 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6029 SDValue Chain = Node->getOperand(0);
6030 SDValue In1 = Node->getOperand(1);
6031 assert(Node->getOperand(2).getNode()->getOpcode()==ISD::BUILD_PAIR);
6032 SDValue In2L = Node->getOperand(2).getNode()->getOperand(0);
6033 SDValue In2H = Node->getOperand(2).getNode()->getOperand(1);
6034 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6035 // have a MemOperand. Pass the info through as a normal operand.
6036 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6037 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6038 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6039 SDValue Result = DAG.getNode(NewOp, Tys, Ops, 5);
6040 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6041 SDValue ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2);
6042 SDValue Vals[2] = { ResultVal, Result.getValue(2) };
6043 return SDValue(DAG.getMergeValues(Vals, 2).getNode(), 0);
6046 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6047 SDNode *Node = Op.getNode();
6048 MVT T = Node->getValueType(0);
6049 SDValue negOp = DAG.getNode(ISD::SUB, T,
6050 DAG.getConstant(0, T), Node->getOperand(2));
6051 return DAG.getAtomic((Op.getOpcode()==ISD::ATOMIC_LOAD_SUB_8 ?
6052 ISD::ATOMIC_LOAD_ADD_8 :
6053 Op.getOpcode()==ISD::ATOMIC_LOAD_SUB_16 ?
6054 ISD::ATOMIC_LOAD_ADD_16 :
6055 Op.getOpcode()==ISD::ATOMIC_LOAD_SUB_32 ?
6056 ISD::ATOMIC_LOAD_ADD_32 :
6057 ISD::ATOMIC_LOAD_ADD_64),
6058 Node->getOperand(0),
6059 Node->getOperand(1), negOp,
6060 cast<AtomicSDNode>(Node)->getSrcValue(),
6061 cast<AtomicSDNode>(Node)->getAlignment());
6064 /// LowerOperation - Provide custom lowering hooks for some operations.
6066 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6067 switch (Op.getOpcode()) {
6068 default: assert(0 && "Should not custom lower this!");
6069 case ISD::ATOMIC_CMP_SWAP_8:
6070 case ISD::ATOMIC_CMP_SWAP_16:
6071 case ISD::ATOMIC_CMP_SWAP_32:
6072 case ISD::ATOMIC_CMP_SWAP_64: return LowerCMP_SWAP(Op,DAG);
6073 case ISD::ATOMIC_LOAD_SUB_8:
6074 case ISD::ATOMIC_LOAD_SUB_16:
6075 case ISD::ATOMIC_LOAD_SUB_32: return LowerLOAD_SUB(Op,DAG);
6076 case ISD::ATOMIC_LOAD_SUB_64: return (Subtarget->is64Bit()) ?
6077 LowerLOAD_SUB(Op,DAG) :
6078 LowerATOMIC_BINARY_64(Op,DAG,
6079 X86ISD::ATOMSUB64_DAG);
6080 case ISD::ATOMIC_LOAD_AND_64: return LowerATOMIC_BINARY_64(Op,DAG,
6081 X86ISD::ATOMAND64_DAG);
6082 case ISD::ATOMIC_LOAD_OR_64: return LowerATOMIC_BINARY_64(Op, DAG,
6083 X86ISD::ATOMOR64_DAG);
6084 case ISD::ATOMIC_LOAD_XOR_64: return LowerATOMIC_BINARY_64(Op,DAG,
6085 X86ISD::ATOMXOR64_DAG);
6086 case ISD::ATOMIC_LOAD_NAND_64:return LowerATOMIC_BINARY_64(Op,DAG,
6087 X86ISD::ATOMNAND64_DAG);
6088 case ISD::ATOMIC_LOAD_ADD_64: return LowerATOMIC_BINARY_64(Op,DAG,
6089 X86ISD::ATOMADD64_DAG);
6090 case ISD::ATOMIC_SWAP_64: return LowerATOMIC_BINARY_64(Op,DAG,
6091 X86ISD::ATOMSWAP64_DAG);
6092 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6093 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6094 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6095 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6096 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6097 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6098 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6099 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6100 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6101 case ISD::SHL_PARTS:
6102 case ISD::SRA_PARTS:
6103 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6104 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6105 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6106 case ISD::FABS: return LowerFABS(Op, DAG);
6107 case ISD::FNEG: return LowerFNEG(Op, DAG);
6108 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6109 case ISD::SETCC: return LowerSETCC(Op, DAG);
6110 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6111 case ISD::SELECT: return LowerSELECT(Op, DAG);
6112 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6113 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6114 case ISD::CALL: return LowerCALL(Op, DAG);
6115 case ISD::RET: return LowerRET(Op, DAG);
6116 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6117 case ISD::VASTART: return LowerVASTART(Op, DAG);
6118 case ISD::VAARG: return LowerVAARG(Op, DAG);
6119 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6120 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6121 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6122 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6123 case ISD::FRAME_TO_ARGS_OFFSET:
6124 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6125 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6126 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6127 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6128 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6129 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6130 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6132 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
6133 case ISD::READCYCLECOUNTER:
6134 return SDValue(ExpandREADCYCLECOUNTER(Op.getNode(), DAG), 0);
6138 /// ReplaceNodeResults - Replace a node with an illegal result type
6139 /// with a new node built out of custom code.
6140 SDNode *X86TargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
6141 switch (N->getOpcode()) {
6142 default: assert(0 && "Should not custom lower this!");
6143 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
6144 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
6145 case ISD::ATOMIC_CMP_SWAP_64: return ExpandATOMIC_CMP_SWAP(N, DAG);
6149 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6151 default: return NULL;
6152 case X86ISD::BSF: return "X86ISD::BSF";
6153 case X86ISD::BSR: return "X86ISD::BSR";
6154 case X86ISD::SHLD: return "X86ISD::SHLD";
6155 case X86ISD::SHRD: return "X86ISD::SHRD";
6156 case X86ISD::FAND: return "X86ISD::FAND";
6157 case X86ISD::FOR: return "X86ISD::FOR";
6158 case X86ISD::FXOR: return "X86ISD::FXOR";
6159 case X86ISD::FSRL: return "X86ISD::FSRL";
6160 case X86ISD::FILD: return "X86ISD::FILD";
6161 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6162 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6163 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6164 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6165 case X86ISD::FLD: return "X86ISD::FLD";
6166 case X86ISD::FST: return "X86ISD::FST";
6167 case X86ISD::CALL: return "X86ISD::CALL";
6168 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6169 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6170 case X86ISD::CMP: return "X86ISD::CMP";
6171 case X86ISD::COMI: return "X86ISD::COMI";
6172 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6173 case X86ISD::SETCC: return "X86ISD::SETCC";
6174 case X86ISD::CMOV: return "X86ISD::CMOV";
6175 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6176 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6177 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6178 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6179 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6180 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6181 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6182 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6183 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6184 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6185 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6186 case X86ISD::FMAX: return "X86ISD::FMAX";
6187 case X86ISD::FMIN: return "X86ISD::FMIN";
6188 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6189 case X86ISD::FRCP: return "X86ISD::FRCP";
6190 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6191 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6192 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6193 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6194 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6195 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6196 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6197 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6198 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6199 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6200 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6201 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6202 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6203 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6204 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6205 case X86ISD::VSHL: return "X86ISD::VSHL";
6206 case X86ISD::VSRL: return "X86ISD::VSRL";
6207 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6208 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6209 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6210 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6211 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6212 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6213 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6214 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6215 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6216 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6220 // isLegalAddressingMode - Return true if the addressing mode represented
6221 // by AM is legal for this target, for a load/store of the specified type.
6222 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6223 const Type *Ty) const {
6224 // X86 supports extremely general addressing modes.
6226 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6227 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6231 // We can only fold this if we don't need an extra load.
6232 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6235 // X86-64 only supports addr of globals in small code model.
6236 if (Subtarget->is64Bit()) {
6237 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6239 // If lower 4G is not available, then we must use rip-relative addressing.
6240 if (AM.BaseOffs || AM.Scale > 1)
6251 // These scales always work.
6256 // These scales are formed with basereg+scalereg. Only accept if there is
6261 default: // Other stuff never works.
6269 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6270 if (!Ty1->isInteger() || !Ty2->isInteger())
6272 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6273 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6274 if (NumBits1 <= NumBits2)
6276 return Subtarget->is64Bit() || NumBits1 < 64;
6279 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6280 if (!VT1.isInteger() || !VT2.isInteger())
6282 unsigned NumBits1 = VT1.getSizeInBits();
6283 unsigned NumBits2 = VT2.getSizeInBits();
6284 if (NumBits1 <= NumBits2)
6286 return Subtarget->is64Bit() || NumBits1 < 64;
6289 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6290 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6291 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6292 /// are assumed to be legal.
6294 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6295 // Only do shuffles on 128-bit vector types for now.
6296 if (VT.getSizeInBits() == 64) return false;
6297 return (Mask.getNode()->getNumOperands() <= 4 ||
6298 isIdentityMask(Mask.getNode()) ||
6299 isIdentityMask(Mask.getNode(), true) ||
6300 isSplatMask(Mask.getNode()) ||
6301 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6302 X86::isUNPCKLMask(Mask.getNode()) ||
6303 X86::isUNPCKHMask(Mask.getNode()) ||
6304 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6305 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6309 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6310 MVT EVT, SelectionDAG &DAG) const {
6311 unsigned NumElts = BVOps.size();
6312 // Only do shuffles on 128-bit vector types for now.
6313 if (EVT.getSizeInBits() * NumElts == 64) return false;
6314 if (NumElts == 2) return true;
6316 return (isMOVLMask(&BVOps[0], 4) ||
6317 isCommutedMOVL(&BVOps[0], 4, true) ||
6318 isSHUFPMask(&BVOps[0], 4) ||
6319 isCommutedSHUFP(&BVOps[0], 4));
6324 //===----------------------------------------------------------------------===//
6325 // X86 Scheduler Hooks
6326 //===----------------------------------------------------------------------===//
6328 // private utility function
6330 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6331 MachineBasicBlock *MBB,
6339 TargetRegisterClass *RC,
6341 // For the atomic bitwise operator, we generate
6344 // ld t1 = [bitinstr.addr]
6345 // op t2 = t1, [bitinstr.val]
6347 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6349 // fallthrough -->nextMBB
6350 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6351 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6352 MachineFunction::iterator MBBIter = MBB;
6355 /// First build the CFG
6356 MachineFunction *F = MBB->getParent();
6357 MachineBasicBlock *thisMBB = MBB;
6358 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6359 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6360 F->insert(MBBIter, newMBB);
6361 F->insert(MBBIter, nextMBB);
6363 // Move all successors to thisMBB to nextMBB
6364 nextMBB->transferSuccessors(thisMBB);
6366 // Update thisMBB to fall through to newMBB
6367 thisMBB->addSuccessor(newMBB);
6369 // newMBB jumps to itself and fall through to nextMBB
6370 newMBB->addSuccessor(nextMBB);
6371 newMBB->addSuccessor(newMBB);
6373 // Insert instructions into newMBB based on incoming instruction
6374 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6375 MachineOperand& destOper = bInstr->getOperand(0);
6376 MachineOperand* argOpers[6];
6377 int numArgs = bInstr->getNumOperands() - 1;
6378 for (int i=0; i < numArgs; ++i)
6379 argOpers[i] = &bInstr->getOperand(i+1);
6381 // x86 address has 4 operands: base, index, scale, and displacement
6382 int lastAddrIndx = 3; // [0,3]
6385 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6386 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6387 for (int i=0; i <= lastAddrIndx; ++i)
6388 (*MIB).addOperand(*argOpers[i]);
6390 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6392 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6397 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6398 assert((argOpers[valArgIndx]->isReg() ||
6399 argOpers[valArgIndx]->isImm()) &&
6401 if (argOpers[valArgIndx]->isReg())
6402 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6404 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6406 (*MIB).addOperand(*argOpers[valArgIndx]);
6408 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6411 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6412 for (int i=0; i <= lastAddrIndx; ++i)
6413 (*MIB).addOperand(*argOpers[i]);
6415 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6416 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6418 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6422 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6424 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6428 // private utility function: 64 bit atomics on 32 bit host.
6430 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
6431 MachineBasicBlock *MBB,
6437 // For the atomic bitwise operator, we generate
6438 // thisMBB (instructions are in pairs, except cmpxchg8b)
6439 // ld t1,t2 = [bitinstr.addr]
6441 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
6442 // op t5, t6 <- out1, out2, [bitinstr.val]
6443 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
6444 // mov ECX, EBX <- t5, t6
6445 // mov EAX, EDX <- t1, t2
6446 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
6447 // mov t3, t4 <- EAX, EDX
6449 // result in out1, out2
6450 // fallthrough -->nextMBB
6452 const TargetRegisterClass *RC = X86::GR32RegisterClass;
6453 const unsigned LoadOpc = X86::MOV32rm;
6454 const unsigned copyOpc = X86::MOV32rr;
6455 const unsigned NotOpc = X86::NOT32r;
6456 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6457 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6458 MachineFunction::iterator MBBIter = MBB;
6461 /// First build the CFG
6462 MachineFunction *F = MBB->getParent();
6463 MachineBasicBlock *thisMBB = MBB;
6464 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6465 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6466 F->insert(MBBIter, newMBB);
6467 F->insert(MBBIter, nextMBB);
6469 // Move all successors to thisMBB to nextMBB
6470 nextMBB->transferSuccessors(thisMBB);
6472 // Update thisMBB to fall through to newMBB
6473 thisMBB->addSuccessor(newMBB);
6475 // newMBB jumps to itself and fall through to nextMBB
6476 newMBB->addSuccessor(nextMBB);
6477 newMBB->addSuccessor(newMBB);
6479 // Insert instructions into newMBB based on incoming instruction
6480 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
6481 assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
6482 MachineOperand& dest1Oper = bInstr->getOperand(0);
6483 MachineOperand& dest2Oper = bInstr->getOperand(1);
6484 MachineOperand* argOpers[6];
6485 for (int i=0; i < 6; ++i)
6486 argOpers[i] = &bInstr->getOperand(i+2);
6488 // x86 address has 4 operands: base, index, scale, and displacement
6489 int lastAddrIndx = 3; // [0,3]
6491 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6492 MachineInstrBuilder MIB = BuildMI(thisMBB, TII->get(LoadOpc), t1);
6493 for (int i=0; i <= lastAddrIndx; ++i)
6494 (*MIB).addOperand(*argOpers[i]);
6495 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6496 MIB = BuildMI(thisMBB, TII->get(LoadOpc), t2);
6497 // add 4 to displacement.
6498 for (int i=0; i <= lastAddrIndx-1; ++i)
6499 (*MIB).addOperand(*argOpers[i]);
6500 MachineOperand newOp3 = *(argOpers[3]);
6502 newOp3.setImm(newOp3.getImm()+4);
6504 newOp3.setOffset(newOp3.getOffset()+4);
6505 (*MIB).addOperand(newOp3);
6507 // t3/4 are defined later, at the bottom of the loop
6508 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
6509 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
6510 BuildMI(newMBB, TII->get(X86::PHI), dest1Oper.getReg())
6511 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
6512 BuildMI(newMBB, TII->get(X86::PHI), dest2Oper.getReg())
6513 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
6515 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
6516 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
6518 MIB = BuildMI(newMBB, TII->get(NotOpc), tt1).addReg(t1);
6519 MIB = BuildMI(newMBB, TII->get(NotOpc), tt2).addReg(t2);
6525 assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
6527 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
6528 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
6529 if (argOpers[4]->isReg())
6530 MIB = BuildMI(newMBB, TII->get(regOpcL), t5);
6532 MIB = BuildMI(newMBB, TII->get(immOpcL), t5);
6533 if (regOpcL != X86::MOV32rr)
6535 (*MIB).addOperand(*argOpers[4]);
6536 assert(argOpers[5]->isReg() == argOpers[4]->isReg());
6537 assert(argOpers[5]->isImm() == argOpers[4]->isImm());
6538 if (argOpers[5]->isReg())
6539 MIB = BuildMI(newMBB, TII->get(regOpcH), t6);
6541 MIB = BuildMI(newMBB, TII->get(immOpcH), t6);
6542 if (regOpcH != X86::MOV32rr)
6544 (*MIB).addOperand(*argOpers[5]);
6546 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EAX);
6548 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EDX);
6551 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EBX);
6553 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::ECX);
6556 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG8B));
6557 for (int i=0; i <= lastAddrIndx; ++i)
6558 (*MIB).addOperand(*argOpers[i]);
6560 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6561 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6563 MIB = BuildMI(newMBB, TII->get(copyOpc), t3);
6564 MIB.addReg(X86::EAX);
6565 MIB = BuildMI(newMBB, TII->get(copyOpc), t4);
6566 MIB.addReg(X86::EDX);
6569 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6571 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6575 // private utility function
6577 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
6578 MachineBasicBlock *MBB,
6580 // For the atomic min/max operator, we generate
6583 // ld t1 = [min/max.addr]
6584 // mov t2 = [min/max.val]
6586 // cmov[cond] t2 = t1
6588 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6590 // fallthrough -->nextMBB
6592 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6593 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6594 MachineFunction::iterator MBBIter = MBB;
6597 /// First build the CFG
6598 MachineFunction *F = MBB->getParent();
6599 MachineBasicBlock *thisMBB = MBB;
6600 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6601 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6602 F->insert(MBBIter, newMBB);
6603 F->insert(MBBIter, nextMBB);
6605 // Move all successors to thisMBB to nextMBB
6606 nextMBB->transferSuccessors(thisMBB);
6608 // Update thisMBB to fall through to newMBB
6609 thisMBB->addSuccessor(newMBB);
6611 // newMBB jumps to newMBB and fall through to nextMBB
6612 newMBB->addSuccessor(nextMBB);
6613 newMBB->addSuccessor(newMBB);
6615 // Insert instructions into newMBB based on incoming instruction
6616 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
6617 MachineOperand& destOper = mInstr->getOperand(0);
6618 MachineOperand* argOpers[6];
6619 int numArgs = mInstr->getNumOperands() - 1;
6620 for (int i=0; i < numArgs; ++i)
6621 argOpers[i] = &mInstr->getOperand(i+1);
6623 // x86 address has 4 operands: base, index, scale, and displacement
6624 int lastAddrIndx = 3; // [0,3]
6627 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6628 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
6629 for (int i=0; i <= lastAddrIndx; ++i)
6630 (*MIB).addOperand(*argOpers[i]);
6632 // We only support register and immediate values
6633 assert((argOpers[valArgIndx]->isReg() ||
6634 argOpers[valArgIndx]->isImm()) &&
6637 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6638 if (argOpers[valArgIndx]->isReg())
6639 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6641 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
6642 (*MIB).addOperand(*argOpers[valArgIndx]);
6644 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
6647 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
6652 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
6653 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
6657 // Cmp and exchange if none has modified the memory location
6658 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
6659 for (int i=0; i <= lastAddrIndx; ++i)
6660 (*MIB).addOperand(*argOpers[i]);
6662 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6663 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
6665 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
6666 MIB.addReg(X86::EAX);
6669 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6671 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
6677 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6678 MachineBasicBlock *BB) {
6679 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6680 switch (MI->getOpcode()) {
6681 default: assert(false && "Unexpected instr type to insert");
6682 case X86::CMOV_FR32:
6683 case X86::CMOV_FR64:
6684 case X86::CMOV_V4F32:
6685 case X86::CMOV_V2F64:
6686 case X86::CMOV_V2I64: {
6687 // To "insert" a SELECT_CC instruction, we actually have to insert the
6688 // diamond control-flow pattern. The incoming instruction knows the
6689 // destination vreg to set, the condition code register to branch on, the
6690 // true/false values to select between, and a branch opcode to use.
6691 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6692 MachineFunction::iterator It = BB;
6698 // cmpTY ccX, r1, r2
6700 // fallthrough --> copy0MBB
6701 MachineBasicBlock *thisMBB = BB;
6702 MachineFunction *F = BB->getParent();
6703 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6704 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6706 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
6707 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
6708 F->insert(It, copy0MBB);
6709 F->insert(It, sinkMBB);
6710 // Update machine-CFG edges by transferring all successors of the current
6711 // block to the new block which will contain the Phi node for the select.
6712 sinkMBB->transferSuccessors(BB);
6714 // Add the true and fallthrough blocks as its successors.
6715 BB->addSuccessor(copy0MBB);
6716 BB->addSuccessor(sinkMBB);
6719 // %FalseValue = ...
6720 // # fallthrough to sinkMBB
6723 // Update machine-CFG edges
6724 BB->addSuccessor(sinkMBB);
6727 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6730 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
6731 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6732 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6734 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6738 case X86::FP32_TO_INT16_IN_MEM:
6739 case X86::FP32_TO_INT32_IN_MEM:
6740 case X86::FP32_TO_INT64_IN_MEM:
6741 case X86::FP64_TO_INT16_IN_MEM:
6742 case X86::FP64_TO_INT32_IN_MEM:
6743 case X86::FP64_TO_INT64_IN_MEM:
6744 case X86::FP80_TO_INT16_IN_MEM:
6745 case X86::FP80_TO_INT32_IN_MEM:
6746 case X86::FP80_TO_INT64_IN_MEM: {
6747 // Change the floating point control register to use "round towards zero"
6748 // mode when truncating to an integer value.
6749 MachineFunction *F = BB->getParent();
6750 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
6751 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
6753 // Load the old value of the high byte of the control word...
6755 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
6756 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
6758 // Set the high part to be round to zero...
6759 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
6762 // Reload the modified control word now...
6763 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6765 // Restore the memory image of control word to original value
6766 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
6769 // Get the X86 opcode to use.
6771 switch (MI->getOpcode()) {
6772 default: assert(0 && "illegal opcode!");
6773 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
6774 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
6775 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
6776 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
6777 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
6778 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
6779 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
6780 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
6781 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
6785 MachineOperand &Op = MI->getOperand(0);
6787 AM.BaseType = X86AddressMode::RegBase;
6788 AM.Base.Reg = Op.getReg();
6790 AM.BaseType = X86AddressMode::FrameIndexBase;
6791 AM.Base.FrameIndex = Op.getIndex();
6793 Op = MI->getOperand(1);
6795 AM.Scale = Op.getImm();
6796 Op = MI->getOperand(2);
6798 AM.IndexReg = Op.getImm();
6799 Op = MI->getOperand(3);
6800 if (Op.isGlobal()) {
6801 AM.GV = Op.getGlobal();
6803 AM.Disp = Op.getImm();
6805 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
6806 .addReg(MI->getOperand(4).getReg());
6808 // Reload the original control word now.
6809 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
6811 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
6814 case X86::ATOMAND32:
6815 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6816 X86::AND32ri, X86::MOV32rm,
6817 X86::LCMPXCHG32, X86::MOV32rr,
6818 X86::NOT32r, X86::EAX,
6819 X86::GR32RegisterClass);
6821 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
6822 X86::OR32ri, X86::MOV32rm,
6823 X86::LCMPXCHG32, X86::MOV32rr,
6824 X86::NOT32r, X86::EAX,
6825 X86::GR32RegisterClass);
6826 case X86::ATOMXOR32:
6827 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
6828 X86::XOR32ri, X86::MOV32rm,
6829 X86::LCMPXCHG32, X86::MOV32rr,
6830 X86::NOT32r, X86::EAX,
6831 X86::GR32RegisterClass);
6832 case X86::ATOMNAND32:
6833 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
6834 X86::AND32ri, X86::MOV32rm,
6835 X86::LCMPXCHG32, X86::MOV32rr,
6836 X86::NOT32r, X86::EAX,
6837 X86::GR32RegisterClass, true);
6838 case X86::ATOMMIN32:
6839 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
6840 case X86::ATOMMAX32:
6841 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
6842 case X86::ATOMUMIN32:
6843 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
6844 case X86::ATOMUMAX32:
6845 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
6847 case X86::ATOMAND16:
6848 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6849 X86::AND16ri, X86::MOV16rm,
6850 X86::LCMPXCHG16, X86::MOV16rr,
6851 X86::NOT16r, X86::AX,
6852 X86::GR16RegisterClass);
6854 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
6855 X86::OR16ri, X86::MOV16rm,
6856 X86::LCMPXCHG16, X86::MOV16rr,
6857 X86::NOT16r, X86::AX,
6858 X86::GR16RegisterClass);
6859 case X86::ATOMXOR16:
6860 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
6861 X86::XOR16ri, X86::MOV16rm,
6862 X86::LCMPXCHG16, X86::MOV16rr,
6863 X86::NOT16r, X86::AX,
6864 X86::GR16RegisterClass);
6865 case X86::ATOMNAND16:
6866 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
6867 X86::AND16ri, X86::MOV16rm,
6868 X86::LCMPXCHG16, X86::MOV16rr,
6869 X86::NOT16r, X86::AX,
6870 X86::GR16RegisterClass, true);
6871 case X86::ATOMMIN16:
6872 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
6873 case X86::ATOMMAX16:
6874 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
6875 case X86::ATOMUMIN16:
6876 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
6877 case X86::ATOMUMAX16:
6878 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
6881 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6882 X86::AND8ri, X86::MOV8rm,
6883 X86::LCMPXCHG8, X86::MOV8rr,
6884 X86::NOT8r, X86::AL,
6885 X86::GR8RegisterClass);
6887 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
6888 X86::OR8ri, X86::MOV8rm,
6889 X86::LCMPXCHG8, X86::MOV8rr,
6890 X86::NOT8r, X86::AL,
6891 X86::GR8RegisterClass);
6893 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
6894 X86::XOR8ri, X86::MOV8rm,
6895 X86::LCMPXCHG8, X86::MOV8rr,
6896 X86::NOT8r, X86::AL,
6897 X86::GR8RegisterClass);
6898 case X86::ATOMNAND8:
6899 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
6900 X86::AND8ri, X86::MOV8rm,
6901 X86::LCMPXCHG8, X86::MOV8rr,
6902 X86::NOT8r, X86::AL,
6903 X86::GR8RegisterClass, true);
6904 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
6905 // This group is for 64-bit host.
6906 case X86::ATOMAND64:
6907 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6908 X86::AND64ri32, X86::MOV64rm,
6909 X86::LCMPXCHG64, X86::MOV64rr,
6910 X86::NOT64r, X86::RAX,
6911 X86::GR64RegisterClass);
6913 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
6914 X86::OR64ri32, X86::MOV64rm,
6915 X86::LCMPXCHG64, X86::MOV64rr,
6916 X86::NOT64r, X86::RAX,
6917 X86::GR64RegisterClass);
6918 case X86::ATOMXOR64:
6919 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
6920 X86::XOR64ri32, X86::MOV64rm,
6921 X86::LCMPXCHG64, X86::MOV64rr,
6922 X86::NOT64r, X86::RAX,
6923 X86::GR64RegisterClass);
6924 case X86::ATOMNAND64:
6925 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
6926 X86::AND64ri32, X86::MOV64rm,
6927 X86::LCMPXCHG64, X86::MOV64rr,
6928 X86::NOT64r, X86::RAX,
6929 X86::GR64RegisterClass, true);
6930 case X86::ATOMMIN64:
6931 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
6932 case X86::ATOMMAX64:
6933 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
6934 case X86::ATOMUMIN64:
6935 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
6936 case X86::ATOMUMAX64:
6937 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
6939 // This group does 64-bit operations on a 32-bit host.
6940 case X86::ATOMAND6432:
6941 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6942 X86::AND32rr, X86::AND32rr,
6943 X86::AND32ri, X86::AND32ri,
6945 case X86::ATOMOR6432:
6946 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6947 X86::OR32rr, X86::OR32rr,
6948 X86::OR32ri, X86::OR32ri,
6950 case X86::ATOMXOR6432:
6951 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6952 X86::XOR32rr, X86::XOR32rr,
6953 X86::XOR32ri, X86::XOR32ri,
6955 case X86::ATOMNAND6432:
6956 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6957 X86::AND32rr, X86::AND32rr,
6958 X86::AND32ri, X86::AND32ri,
6960 case X86::ATOMADD6432:
6961 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6962 X86::ADD32rr, X86::ADC32rr,
6963 X86::ADD32ri, X86::ADC32ri,
6965 case X86::ATOMSUB6432:
6966 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6967 X86::SUB32rr, X86::SBB32rr,
6968 X86::SUB32ri, X86::SBB32ri,
6970 case X86::ATOMSWAP6432:
6971 return EmitAtomicBit6432WithCustomInserter(MI, BB,
6972 X86::MOV32rr, X86::MOV32rr,
6973 X86::MOV32ri, X86::MOV32ri,
6978 //===----------------------------------------------------------------------===//
6979 // X86 Optimization Hooks
6980 //===----------------------------------------------------------------------===//
6982 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6986 const SelectionDAG &DAG,
6987 unsigned Depth) const {
6988 unsigned Opc = Op.getOpcode();
6989 assert((Opc >= ISD::BUILTIN_OP_END ||
6990 Opc == ISD::INTRINSIC_WO_CHAIN ||
6991 Opc == ISD::INTRINSIC_W_CHAIN ||
6992 Opc == ISD::INTRINSIC_VOID) &&
6993 "Should use MaskedValueIsZero if you don't know whether Op"
6994 " is a target node!");
6996 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7000 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7001 Mask.getBitWidth() - 1);
7006 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7007 /// node is a GlobalAddress + offset.
7008 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7009 GlobalValue* &GA, int64_t &Offset) const{
7010 if (N->getOpcode() == X86ISD::Wrapper) {
7011 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7012 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7016 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7019 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7020 const TargetLowering &TLI) {
7023 if (TLI.isGAPlusOffset(Base, GV, Offset))
7024 return (GV->getAlignment() >= N && (Offset % N) == 0);
7025 // DAG combine handles the stack object case.
7029 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
7030 unsigned NumElems, MVT EVT,
7032 SelectionDAG &DAG, MachineFrameInfo *MFI,
7033 const TargetLowering &TLI) {
7035 for (unsigned i = 0; i < NumElems; ++i) {
7036 SDValue Idx = PermMask.getOperand(i);
7037 if (Idx.getOpcode() == ISD::UNDEF) {
7043 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7044 if (!Elt.getNode() ||
7045 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7048 Base = Elt.getNode();
7049 if (Base->getOpcode() == ISD::UNDEF)
7053 if (Elt.getOpcode() == ISD::UNDEF)
7056 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7057 EVT.getSizeInBits()/8, i, MFI))
7063 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7064 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7065 /// if the load addresses are consecutive, non-overlapping, and in the right
7067 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7068 const TargetLowering &TLI) {
7069 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7070 MVT VT = N->getValueType(0);
7071 MVT EVT = VT.getVectorElementType();
7072 SDValue PermMask = N->getOperand(2);
7073 unsigned NumElems = PermMask.getNumOperands();
7074 SDNode *Base = NULL;
7075 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
7079 LoadSDNode *LD = cast<LoadSDNode>(Base);
7080 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7081 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7082 LD->getSrcValueOffset(), LD->isVolatile());
7083 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7084 LD->getSrcValueOffset(), LD->isVolatile(),
7085 LD->getAlignment());
7088 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7089 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7090 const X86Subtarget *Subtarget,
7091 const TargetLowering &TLI) {
7092 unsigned NumOps = N->getNumOperands();
7094 // Ignore single operand BUILD_VECTOR.
7098 MVT VT = N->getValueType(0);
7099 MVT EVT = VT.getVectorElementType();
7100 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7101 // We are looking for load i64 and zero extend. We want to transform
7102 // it before legalizer has a chance to expand it. Also look for i64
7103 // BUILD_PAIR bit casted to f64.
7105 // This must be an insertion into a zero vector.
7106 SDValue HighElt = N->getOperand(1);
7107 if (!isZeroNode(HighElt))
7110 // Value must be a load.
7111 SDNode *Base = N->getOperand(0).getNode();
7112 if (!isa<LoadSDNode>(Base)) {
7113 if (Base->getOpcode() != ISD::BIT_CONVERT)
7115 Base = Base->getOperand(0).getNode();
7116 if (!isa<LoadSDNode>(Base))
7120 // Transform it into VZEXT_LOAD addr.
7121 LoadSDNode *LD = cast<LoadSDNode>(Base);
7123 // Load must not be an extload.
7124 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7127 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7128 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7129 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, Tys, Ops, 2);
7130 DAG.ReplaceAllUsesOfValueWith(SDValue(Base, 1), ResNode.getValue(1));
7134 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7135 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7136 const X86Subtarget *Subtarget) {
7137 SDValue Cond = N->getOperand(0);
7139 // If we have SSE[12] support, try to form min/max nodes.
7140 if (Subtarget->hasSSE2() &&
7141 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
7142 if (Cond.getOpcode() == ISD::SETCC) {
7143 // Get the LHS/RHS of the select.
7144 SDValue LHS = N->getOperand(1);
7145 SDValue RHS = N->getOperand(2);
7146 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7148 unsigned Opcode = 0;
7149 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7152 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7155 if (!UnsafeFPMath) break;
7157 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7159 Opcode = X86ISD::FMIN;
7162 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7165 if (!UnsafeFPMath) break;
7167 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7169 Opcode = X86ISD::FMAX;
7172 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7175 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7178 if (!UnsafeFPMath) break;
7180 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7182 Opcode = X86ISD::FMIN;
7185 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7188 if (!UnsafeFPMath) break;
7190 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7192 Opcode = X86ISD::FMAX;
7198 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
7206 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
7207 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
7208 const X86Subtarget *Subtarget) {
7209 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
7210 // the FP state in cases where an emms may be missing.
7211 // A preferable solution to the general problem is to figure out the right
7212 // places to insert EMMS. This qualifies as a quick hack.
7213 StoreSDNode *St = cast<StoreSDNode>(N);
7214 if (St->getValue().getValueType().isVector() &&
7215 St->getValue().getValueType().getSizeInBits() == 64 &&
7216 isa<LoadSDNode>(St->getValue()) &&
7217 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
7218 St->getChain().hasOneUse() && !St->isVolatile()) {
7219 SDNode* LdVal = St->getValue().getNode();
7221 int TokenFactorIndex = -1;
7222 SmallVector<SDValue, 8> Ops;
7223 SDNode* ChainVal = St->getChain().getNode();
7224 // Must be a store of a load. We currently handle two cases: the load
7225 // is a direct child, and it's under an intervening TokenFactor. It is
7226 // possible to dig deeper under nested TokenFactors.
7227 if (ChainVal == LdVal)
7228 Ld = cast<LoadSDNode>(St->getChain());
7229 else if (St->getValue().hasOneUse() &&
7230 ChainVal->getOpcode() == ISD::TokenFactor) {
7231 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
7232 if (ChainVal->getOperand(i).getNode() == LdVal) {
7233 TokenFactorIndex = i;
7234 Ld = cast<LoadSDNode>(St->getValue());
7236 Ops.push_back(ChainVal->getOperand(i));
7240 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
7241 if (Subtarget->is64Bit()) {
7242 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
7243 Ld->getBasePtr(), Ld->getSrcValue(),
7244 Ld->getSrcValueOffset(), Ld->isVolatile(),
7245 Ld->getAlignment());
7246 SDValue NewChain = NewLd.getValue(1);
7247 if (TokenFactorIndex != -1) {
7248 Ops.push_back(NewChain);
7249 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7252 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
7253 St->getSrcValue(), St->getSrcValueOffset(),
7254 St->isVolatile(), St->getAlignment());
7257 // Otherwise, lower to two 32-bit copies.
7258 SDValue LoAddr = Ld->getBasePtr();
7259 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7260 DAG.getConstant(4, MVT::i32));
7262 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
7263 Ld->getSrcValue(), Ld->getSrcValueOffset(),
7264 Ld->isVolatile(), Ld->getAlignment());
7265 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
7266 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
7268 MinAlign(Ld->getAlignment(), 4));
7270 SDValue NewChain = LoLd.getValue(1);
7271 if (TokenFactorIndex != -1) {
7272 Ops.push_back(LoLd);
7273 Ops.push_back(HiLd);
7274 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7278 LoAddr = St->getBasePtr();
7279 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7280 DAG.getConstant(4, MVT::i32));
7282 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
7283 St->getSrcValue(), St->getSrcValueOffset(),
7284 St->isVolatile(), St->getAlignment());
7285 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
7287 St->getSrcValueOffset() + 4,
7289 MinAlign(St->getAlignment(), 4));
7290 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
7296 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
7297 /// X86ISD::FXOR nodes.
7298 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
7299 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
7300 // F[X]OR(0.0, x) -> x
7301 // F[X]OR(x, 0.0) -> x
7302 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7303 if (C->getValueAPF().isPosZero())
7304 return N->getOperand(1);
7305 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7306 if (C->getValueAPF().isPosZero())
7307 return N->getOperand(0);
7311 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
7312 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
7313 // FAND(0.0, x) -> 0.0
7314 // FAND(x, 0.0) -> 0.0
7315 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7316 if (C->getValueAPF().isPosZero())
7317 return N->getOperand(0);
7318 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7319 if (C->getValueAPF().isPosZero())
7320 return N->getOperand(1);
7325 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7326 DAGCombinerInfo &DCI) const {
7327 SelectionDAG &DAG = DCI.DAG;
7328 switch (N->getOpcode()) {
7330 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7331 case ISD::BUILD_VECTOR:
7332 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7333 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7334 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7336 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7337 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7343 //===----------------------------------------------------------------------===//
7344 // X86 Inline Assembly Support
7345 //===----------------------------------------------------------------------===//
7347 /// getConstraintType - Given a constraint letter, return the type of
7348 /// constraint it is for this target.
7349 X86TargetLowering::ConstraintType
7350 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7351 if (Constraint.size() == 1) {
7352 switch (Constraint[0]) {
7363 return C_RegisterClass;
7368 return TargetLowering::getConstraintType(Constraint);
7371 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7372 /// with another that has more specific requirements based on the type of the
7373 /// corresponding operand.
7374 const char *X86TargetLowering::
7375 LowerXConstraint(MVT ConstraintVT) const {
7376 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7377 // 'f' like normal targets.
7378 if (ConstraintVT.isFloatingPoint()) {
7379 if (Subtarget->hasSSE2())
7381 if (Subtarget->hasSSE1())
7385 return TargetLowering::LowerXConstraint(ConstraintVT);
7388 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7389 /// vector. If it is invalid, don't add anything to Ops.
7390 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7393 std::vector<SDValue>&Ops,
7394 SelectionDAG &DAG) const {
7395 SDValue Result(0, 0);
7397 switch (Constraint) {
7400 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7401 if (C->getZExtValue() <= 31) {
7402 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7408 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7409 if (C->getZExtValue() <= 63) {
7410 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7416 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7417 if (C->getZExtValue() <= 255) {
7418 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7424 // Literal immediates are always ok.
7425 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7426 Result = DAG.getTargetConstant(CST->getZExtValue(), Op.getValueType());
7430 // If we are in non-pic codegen mode, we allow the address of a global (with
7431 // an optional displacement) to be used with 'i'.
7432 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7435 // Match either (GA) or (GA+C)
7437 Offset = GA->getOffset();
7438 } else if (Op.getOpcode() == ISD::ADD) {
7439 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7440 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7442 Offset = GA->getOffset()+C->getZExtValue();
7444 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7445 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7447 Offset = GA->getOffset()+C->getZExtValue();
7455 Op = LowerGlobalAddress(GA->getGlobal(), DAG);
7457 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7463 // Otherwise, not valid for this mode.
7468 if (Result.getNode()) {
7469 Ops.push_back(Result);
7472 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
7476 std::vector<unsigned> X86TargetLowering::
7477 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7479 if (Constraint.size() == 1) {
7480 // FIXME: not handling fp-stack yet!
7481 switch (Constraint[0]) { // GCC X86 Constraint Letters
7482 default: break; // Unknown constraint letter
7483 case 'A': // EAX/EDX
7484 if (VT == MVT::i32 || VT == MVT::i64)
7485 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
7487 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7490 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7491 else if (VT == MVT::i16)
7492 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7493 else if (VT == MVT::i8)
7494 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7495 else if (VT == MVT::i64)
7496 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7501 return std::vector<unsigned>();
7504 std::pair<unsigned, const TargetRegisterClass*>
7505 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7507 // First, see if this is a constraint that directly corresponds to an LLVM
7509 if (Constraint.size() == 1) {
7510 // GCC Constraint Letters
7511 switch (Constraint[0]) {
7513 case 'r': // GENERAL_REGS
7514 case 'R': // LEGACY_REGS
7515 case 'l': // INDEX_REGS
7516 if (VT == MVT::i64 && Subtarget->is64Bit())
7517 return std::make_pair(0U, X86::GR64RegisterClass);
7519 return std::make_pair(0U, X86::GR32RegisterClass);
7520 else if (VT == MVT::i16)
7521 return std::make_pair(0U, X86::GR16RegisterClass);
7522 else if (VT == MVT::i8)
7523 return std::make_pair(0U, X86::GR8RegisterClass);
7525 case 'f': // FP Stack registers.
7526 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7527 // value to the correct fpstack register class.
7528 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7529 return std::make_pair(0U, X86::RFP32RegisterClass);
7530 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7531 return std::make_pair(0U, X86::RFP64RegisterClass);
7532 return std::make_pair(0U, X86::RFP80RegisterClass);
7533 case 'y': // MMX_REGS if MMX allowed.
7534 if (!Subtarget->hasMMX()) break;
7535 return std::make_pair(0U, X86::VR64RegisterClass);
7537 case 'Y': // SSE_REGS if SSE2 allowed
7538 if (!Subtarget->hasSSE2()) break;
7540 case 'x': // SSE_REGS if SSE1 allowed
7541 if (!Subtarget->hasSSE1()) break;
7543 switch (VT.getSimpleVT()) {
7545 // Scalar SSE types.
7548 return std::make_pair(0U, X86::FR32RegisterClass);
7551 return std::make_pair(0U, X86::FR64RegisterClass);
7559 return std::make_pair(0U, X86::VR128RegisterClass);
7565 // Use the default implementation in TargetLowering to convert the register
7566 // constraint into a member of a register class.
7567 std::pair<unsigned, const TargetRegisterClass*> Res;
7568 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7570 // Not found as a standard register?
7571 if (Res.second == 0) {
7572 // GCC calls "st(0)" just plain "st".
7573 if (StringsEqualNoCase("{st}", Constraint)) {
7574 Res.first = X86::ST0;
7575 Res.second = X86::RFP80RegisterClass;
7581 // Otherwise, check to see if this is a register class of the wrong value
7582 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
7583 // turn into {ax},{dx}.
7584 if (Res.second->hasType(VT))
7585 return Res; // Correct type already, nothing to do.
7587 // All of the single-register GCC register classes map their values onto
7588 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
7589 // really want an 8-bit or 32-bit register, map to the appropriate register
7590 // class and return the appropriate register.
7591 if (Res.second == X86::GR16RegisterClass) {
7592 if (VT == MVT::i8) {
7593 unsigned DestReg = 0;
7594 switch (Res.first) {
7596 case X86::AX: DestReg = X86::AL; break;
7597 case X86::DX: DestReg = X86::DL; break;
7598 case X86::CX: DestReg = X86::CL; break;
7599 case X86::BX: DestReg = X86::BL; break;
7602 Res.first = DestReg;
7603 Res.second = Res.second = X86::GR8RegisterClass;
7605 } else if (VT == MVT::i32) {
7606 unsigned DestReg = 0;
7607 switch (Res.first) {
7609 case X86::AX: DestReg = X86::EAX; break;
7610 case X86::DX: DestReg = X86::EDX; break;
7611 case X86::CX: DestReg = X86::ECX; break;
7612 case X86::BX: DestReg = X86::EBX; break;
7613 case X86::SI: DestReg = X86::ESI; break;
7614 case X86::DI: DestReg = X86::EDI; break;
7615 case X86::BP: DestReg = X86::EBP; break;
7616 case X86::SP: DestReg = X86::ESP; break;
7619 Res.first = DestReg;
7620 Res.second = Res.second = X86::GR32RegisterClass;
7622 } else if (VT == MVT::i64) {
7623 unsigned DestReg = 0;
7624 switch (Res.first) {
7626 case X86::AX: DestReg = X86::RAX; break;
7627 case X86::DX: DestReg = X86::RDX; break;
7628 case X86::CX: DestReg = X86::RCX; break;
7629 case X86::BX: DestReg = X86::RBX; break;
7630 case X86::SI: DestReg = X86::RSI; break;
7631 case X86::DI: DestReg = X86::RDI; break;
7632 case X86::BP: DestReg = X86::RBP; break;
7633 case X86::SP: DestReg = X86::RSP; break;
7636 Res.first = DestReg;
7637 Res.second = Res.second = X86::GR64RegisterClass;
7640 } else if (Res.second == X86::FR32RegisterClass ||
7641 Res.second == X86::FR64RegisterClass ||
7642 Res.second == X86::VR128RegisterClass) {
7643 // Handle references to XMM physical registers that got mapped into the
7644 // wrong class. This can happen with constraints like {xmm0} where the
7645 // target independent register mapper will just pick the first match it can
7646 // find, ignoring the required type.
7648 Res.second = X86::FR32RegisterClass;
7649 else if (VT == MVT::f64)
7650 Res.second = X86::FR64RegisterClass;
7651 else if (X86::VR128RegisterClass->hasType(VT))
7652 Res.second = X86::VR128RegisterClass;