1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/StringExtras.h"
41 #include "llvm/Support/CommandLine.h"
45 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
47 // Forward declarations.
48 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
50 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
51 : TargetLowering(TM) {
52 Subtarget = &TM.getSubtarget<X86Subtarget>();
53 X86ScalarSSEf64 = Subtarget->hasSSE2();
54 X86ScalarSSEf32 = Subtarget->hasSSE1();
55 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
59 RegInfo = TM.getRegisterInfo();
62 // Set up the TargetLowering object.
64 // X86 is weird, it always uses i8 for shift amounts and setcc results.
65 setShiftAmountType(MVT::i8);
66 setBooleanContents(ZeroOrOneBooleanContent);
67 setSchedulingPreference(SchedulingForRegPressure);
68 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
69 setStackPointerRegisterToSaveRestore(X86StackPtr);
71 if (Subtarget->isTargetDarwin()) {
72 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(false);
74 setUseUnderscoreLongJmp(false);
75 } else if (Subtarget->isTargetMingw()) {
76 // MS runtime is weird: it exports _setjmp, but longjmp!
77 setUseUnderscoreSetJmp(true);
78 setUseUnderscoreLongJmp(false);
80 setUseUnderscoreSetJmp(true);
81 setUseUnderscoreLongJmp(true);
84 // Set up the register classes.
85 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
86 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
87 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
88 if (Subtarget->is64Bit())
89 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
91 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
93 // We don't accept any truncstore of integer registers.
94 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
95 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
96 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
97 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
98 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
99 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
101 // SETOEQ and SETUNE require checking two conditions.
102 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
103 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
104 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
105 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
106 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
107 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
109 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
111 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
112 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
113 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
115 if (Subtarget->is64Bit()) {
116 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
117 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
119 if (X86ScalarSSEf64) {
120 // We have an impenetrably clever algorithm for ui64->double only.
121 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
123 // We have faster algorithm for ui32->single only.
124 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
126 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
129 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
131 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
132 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
133 // SSE has no i16 to fp conversion, only i32
134 if (X86ScalarSSEf32) {
135 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
136 // f32 and f64 cases are Legal, f80 case is not
137 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
139 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
140 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
143 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
144 // are Legal, f80 is custom lowered.
145 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
146 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
148 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
150 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
151 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
153 if (X86ScalarSSEf32) {
154 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
155 // f32 and f64 cases are Legal, f80 case is not
156 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
158 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
159 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
162 // Handle FP_TO_UINT by promoting the destination to a larger signed
164 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
165 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
166 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
168 if (Subtarget->is64Bit()) {
169 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
170 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
172 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
173 // Expand FP_TO_UINT into a select.
174 // FIXME: We would like to use a Custom expander here eventually to do
175 // the optimal thing for SSE vs. the default expansion in the legalizer.
176 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
178 // With SSE3 we can use fisttpll to convert to a signed i64.
179 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
182 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
183 if (!X86ScalarSSEf64) {
184 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
185 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
188 // Scalar integer divide and remainder are lowered to use operations that
189 // produce two results, to match the available instructions. This exposes
190 // the two-result form to trivial CSE, which is able to combine x/y and x%y
191 // into a single instruction.
193 // Scalar integer multiply-high is also lowered to use two-result
194 // operations, to match the available instructions. However, plain multiply
195 // (low) operations are left as Legal, as there are single-result
196 // instructions for this in x86. Using the two-result multiply instructions
197 // when both high and low results are needed must be arranged by dagcombine.
198 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
199 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
200 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
201 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
202 setOperationAction(ISD::SREM , MVT::i8 , Expand);
203 setOperationAction(ISD::UREM , MVT::i8 , Expand);
204 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
205 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
206 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
207 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
208 setOperationAction(ISD::SREM , MVT::i16 , Expand);
209 setOperationAction(ISD::UREM , MVT::i16 , Expand);
210 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
211 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
212 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
213 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
214 setOperationAction(ISD::SREM , MVT::i32 , Expand);
215 setOperationAction(ISD::UREM , MVT::i32 , Expand);
216 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
217 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
218 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
219 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
220 setOperationAction(ISD::SREM , MVT::i64 , Expand);
221 setOperationAction(ISD::UREM , MVT::i64 , Expand);
223 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
224 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
225 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
226 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
227 if (Subtarget->is64Bit())
228 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
229 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
230 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
231 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
232 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
233 setOperationAction(ISD::FREM , MVT::f32 , Expand);
234 setOperationAction(ISD::FREM , MVT::f64 , Expand);
235 setOperationAction(ISD::FREM , MVT::f80 , Expand);
236 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
238 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
239 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
240 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
241 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
242 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
243 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
244 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
245 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
246 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
247 if (Subtarget->is64Bit()) {
248 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
249 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
250 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
253 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
254 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
256 // These should be promoted to a larger select which is supported.
257 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
258 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
259 // X86 wants to expand cmov itself.
260 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
261 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
262 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
263 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
264 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
265 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
266 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
267 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
268 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
269 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
270 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
271 if (Subtarget->is64Bit()) {
272 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
275 // X86 ret instruction may pop stack.
276 setOperationAction(ISD::RET , MVT::Other, Custom);
277 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
280 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
281 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
282 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
283 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
284 if (Subtarget->is64Bit())
285 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
286 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
289 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
290 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
291 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
293 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
294 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
295 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
296 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
297 if (Subtarget->is64Bit()) {
298 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
299 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
300 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
303 if (Subtarget->hasSSE1())
304 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
306 if (!Subtarget->hasSSE2())
307 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
309 // Expand certain atomics
310 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
311 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
312 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
313 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
315 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
316 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
317 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
318 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
320 if (!Subtarget->is64Bit()) {
321 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
327 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
330 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
331 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
332 // FIXME - use subtarget debug flags
333 if (!Subtarget->isTargetDarwin() &&
334 !Subtarget->isTargetELF() &&
335 !Subtarget->isTargetCygMing()) {
336 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
337 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
340 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
341 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
342 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
343 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
344 if (Subtarget->is64Bit()) {
345 setExceptionPointerRegister(X86::RAX);
346 setExceptionSelectorRegister(X86::RDX);
348 setExceptionPointerRegister(X86::EAX);
349 setExceptionSelectorRegister(X86::EDX);
351 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
352 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
354 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
356 setOperationAction(ISD::TRAP, MVT::Other, Legal);
358 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
359 setOperationAction(ISD::VASTART , MVT::Other, Custom);
360 setOperationAction(ISD::VAEND , MVT::Other, Expand);
361 if (Subtarget->is64Bit()) {
362 setOperationAction(ISD::VAARG , MVT::Other, Custom);
363 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
365 setOperationAction(ISD::VAARG , MVT::Other, Expand);
366 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
369 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
370 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
371 if (Subtarget->is64Bit())
372 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
373 if (Subtarget->isTargetCygMing())
374 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
376 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
378 if (X86ScalarSSEf64) {
379 // f32 and f64 use SSE.
380 // Set up the FP register classes.
381 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
382 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
384 // Use ANDPD to simulate FABS.
385 setOperationAction(ISD::FABS , MVT::f64, Custom);
386 setOperationAction(ISD::FABS , MVT::f32, Custom);
388 // Use XORP to simulate FNEG.
389 setOperationAction(ISD::FNEG , MVT::f64, Custom);
390 setOperationAction(ISD::FNEG , MVT::f32, Custom);
392 // Use ANDPD and ORPD to simulate FCOPYSIGN.
393 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
394 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
396 // We don't support sin/cos/fmod
397 setOperationAction(ISD::FSIN , MVT::f64, Expand);
398 setOperationAction(ISD::FCOS , MVT::f64, Expand);
399 setOperationAction(ISD::FSIN , MVT::f32, Expand);
400 setOperationAction(ISD::FCOS , MVT::f32, Expand);
402 // Expand FP immediates into loads from the stack, except for the special
404 addLegalFPImmediate(APFloat(+0.0)); // xorpd
405 addLegalFPImmediate(APFloat(+0.0f)); // xorps
407 // Floating truncations from f80 and extensions to f80 go through memory.
408 // If optimizing, we lie about this though and handle it in
409 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
411 setConvertAction(MVT::f32, MVT::f80, Expand);
412 setConvertAction(MVT::f64, MVT::f80, Expand);
413 setConvertAction(MVT::f80, MVT::f32, Expand);
414 setConvertAction(MVT::f80, MVT::f64, Expand);
416 } else if (X86ScalarSSEf32) {
417 // Use SSE for f32, x87 for f64.
418 // Set up the FP register classes.
419 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
420 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
422 // Use ANDPS to simulate FABS.
423 setOperationAction(ISD::FABS , MVT::f32, Custom);
425 // Use XORP to simulate FNEG.
426 setOperationAction(ISD::FNEG , MVT::f32, Custom);
428 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
430 // Use ANDPS and ORPS to simulate FCOPYSIGN.
431 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
432 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
434 // We don't support sin/cos/fmod
435 setOperationAction(ISD::FSIN , MVT::f32, Expand);
436 setOperationAction(ISD::FCOS , MVT::f32, Expand);
438 // Special cases we handle for FP constants.
439 addLegalFPImmediate(APFloat(+0.0f)); // xorps
440 addLegalFPImmediate(APFloat(+0.0)); // FLD0
441 addLegalFPImmediate(APFloat(+1.0)); // FLD1
442 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
443 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
445 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
446 // this though and handle it in InstructionSelectPreprocess so that
447 // dagcombine2 can hack on these.
449 setConvertAction(MVT::f32, MVT::f64, Expand);
450 setConvertAction(MVT::f32, MVT::f80, Expand);
451 setConvertAction(MVT::f80, MVT::f32, Expand);
452 setConvertAction(MVT::f64, MVT::f32, Expand);
453 // And x87->x87 truncations also.
454 setConvertAction(MVT::f80, MVT::f64, Expand);
458 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
459 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
462 // f32 and f64 in x87.
463 // Set up the FP register classes.
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
465 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
467 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
468 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
472 // Floating truncations go through memory. If optimizing, we lie about
473 // this though and handle it in InstructionSelectPreprocess so that
474 // dagcombine2 can hack on these.
476 setConvertAction(MVT::f80, MVT::f32, Expand);
477 setConvertAction(MVT::f64, MVT::f32, Expand);
478 setConvertAction(MVT::f80, MVT::f64, Expand);
482 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
483 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
485 addLegalFPImmediate(APFloat(+0.0)); // FLD0
486 addLegalFPImmediate(APFloat(+1.0)); // FLD1
487 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
488 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
489 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
490 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
491 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
492 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
495 // Long double always uses X87.
496 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
497 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
501 APFloat TmpFlt(+0.0);
502 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
504 addLegalFPImmediate(TmpFlt); // FLD0
506 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
507 APFloat TmpFlt2(+1.0);
508 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
510 addLegalFPImmediate(TmpFlt2); // FLD1
511 TmpFlt2.changeSign();
512 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
516 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
517 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
520 // Always use a library call for pow.
521 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
522 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
523 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
525 setOperationAction(ISD::FLOG, MVT::f80, Expand);
526 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
527 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
528 setOperationAction(ISD::FEXP, MVT::f80, Expand);
529 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
531 // First set operation action for all vector types to either promote
532 // (for widening) or expand (for scalarization). Then we will selectively
533 // turn on ones that can be effectively codegen'd.
534 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
535 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
536 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
551 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
581 if (!DisableMMX && Subtarget->hasMMX()) {
582 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
583 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
584 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
585 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
586 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
588 // FIXME: add MMX packed arithmetics
590 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
591 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
592 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
593 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
595 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
596 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
597 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
598 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
600 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
601 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
603 setOperationAction(ISD::AND, MVT::v8i8, Promote);
604 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
605 setOperationAction(ISD::AND, MVT::v4i16, Promote);
606 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
607 setOperationAction(ISD::AND, MVT::v2i32, Promote);
608 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
609 setOperationAction(ISD::AND, MVT::v1i64, Legal);
611 setOperationAction(ISD::OR, MVT::v8i8, Promote);
612 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
613 setOperationAction(ISD::OR, MVT::v4i16, Promote);
614 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
615 setOperationAction(ISD::OR, MVT::v2i32, Promote);
616 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
617 setOperationAction(ISD::OR, MVT::v1i64, Legal);
619 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
620 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
621 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
622 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
623 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
624 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
625 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
627 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
628 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
629 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
630 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
631 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
632 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
633 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
634 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
635 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
637 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
638 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
639 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
640 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
641 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
645 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
646 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
648 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
649 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
650 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
651 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
653 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
655 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
656 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
657 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
658 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
659 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
660 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
663 if (Subtarget->hasSSE1()) {
664 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
666 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
667 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
668 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
669 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
670 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
671 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
672 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
673 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
675 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
676 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
677 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
680 if (Subtarget->hasSSE2()) {
681 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
682 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
683 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
684 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
685 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
687 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
688 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
689 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
690 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
691 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
692 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
693 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
694 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
695 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
696 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
697 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
698 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
699 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
700 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
701 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
702 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
704 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
705 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
706 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
707 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
709 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
710 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
711 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
712 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
713 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
715 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
716 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
717 MVT VT = (MVT::SimpleValueType)i;
718 // Do not attempt to custom lower non-power-of-2 vectors
719 if (!isPowerOf2_32(VT.getVectorNumElements()))
721 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
722 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
723 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
725 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
726 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
727 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
728 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
729 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
731 if (Subtarget->is64Bit()) {
732 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
733 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
736 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
737 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
738 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
739 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
740 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
741 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
742 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
743 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
744 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
745 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
746 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
747 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
750 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
752 // Custom lower v2i64 and v2f64 selects.
753 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
754 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
755 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
756 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
760 if (Subtarget->hasSSE41()) {
761 // FIXME: Do we need to handle scalar-to-vector here?
762 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
764 // i8 and i16 vectors are custom , because the source register and source
765 // source memory operand types are not the same width. f32 vectors are
766 // custom since the immediate controlling the insert encodes additional
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
770 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
771 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
774 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
775 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
776 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
778 if (Subtarget->is64Bit()) {
779 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
780 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
784 if (Subtarget->hasSSE42()) {
785 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
788 // We want to custom lower some of our intrinsics.
789 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
791 // Add/Sub/Mul with overflow operations are custom lowered.
792 setOperationAction(ISD::SADDO, MVT::i32, Custom);
793 setOperationAction(ISD::SADDO, MVT::i64, Custom);
794 setOperationAction(ISD::UADDO, MVT::i32, Custom);
795 setOperationAction(ISD::UADDO, MVT::i64, Custom);
796 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
797 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
798 setOperationAction(ISD::USUBO, MVT::i32, Custom);
799 setOperationAction(ISD::USUBO, MVT::i64, Custom);
800 setOperationAction(ISD::SMULO, MVT::i32, Custom);
801 setOperationAction(ISD::SMULO, MVT::i64, Custom);
802 setOperationAction(ISD::UMULO, MVT::i32, Custom);
803 setOperationAction(ISD::UMULO, MVT::i64, Custom);
805 // We have target-specific dag combine patterns for the following nodes:
806 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
807 setTargetDAGCombine(ISD::BUILD_VECTOR);
808 setTargetDAGCombine(ISD::SELECT);
809 setTargetDAGCombine(ISD::STORE);
811 computeRegisterProperties();
813 // FIXME: These should be based on subtarget info. Plus, the values should
814 // be smaller when we are in optimizing for size mode.
815 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
816 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
817 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
818 allowUnalignedMemoryAccesses = true; // x86 supports it!
819 setPrefLoopAlignment(16);
823 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
828 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
829 /// the desired ByVal argument alignment.
830 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
833 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
834 if (VTy->getBitWidth() == 128)
836 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
837 unsigned EltAlign = 0;
838 getMaxByValAlign(ATy->getElementType(), EltAlign);
839 if (EltAlign > MaxAlign)
841 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
842 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
843 unsigned EltAlign = 0;
844 getMaxByValAlign(STy->getElementType(i), EltAlign);
845 if (EltAlign > MaxAlign)
854 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
855 /// function arguments in the caller parameter area. For X86, aggregates
856 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
857 /// are at 4-byte boundaries.
858 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
859 if (Subtarget->is64Bit()) {
860 // Max of 8 and alignment of type.
861 unsigned TyAlign = TD->getABITypeAlignment(Ty);
868 if (Subtarget->hasSSE1())
869 getMaxByValAlign(Ty, Align);
873 /// getOptimalMemOpType - Returns the target specific optimal type for load
874 /// and store operations as a result of memset, memcpy, and memmove
875 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
878 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
879 bool isSrcConst, bool isSrcStr) const {
880 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
881 // linux. This is because the stack realignment code can't handle certain
882 // cases like PR2962. This should be removed when PR2962 is fixed.
883 if (Subtarget->getStackAlignment() >= 16) {
884 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
886 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
889 if (Subtarget->is64Bit() && Size >= 8)
895 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
897 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
898 SelectionDAG &DAG) const {
899 if (usesGlobalOffsetTable())
900 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
901 if (!Subtarget->isPICStyleRIPRel())
902 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
906 //===----------------------------------------------------------------------===//
907 // Return Value Calling Convention Implementation
908 //===----------------------------------------------------------------------===//
910 #include "X86GenCallingConv.inc"
912 /// LowerRET - Lower an ISD::RET node.
913 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
914 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
916 SmallVector<CCValAssign, 16> RVLocs;
917 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
918 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
919 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
920 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
922 // If this is the first return lowered for this function, add the regs to the
923 // liveout set for the function.
924 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
925 for (unsigned i = 0; i != RVLocs.size(); ++i)
926 if (RVLocs[i].isRegLoc())
927 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
929 SDValue Chain = Op.getOperand(0);
931 // Handle tail call return.
932 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
933 if (Chain.getOpcode() == X86ISD::TAILCALL) {
934 SDValue TailCall = Chain;
935 SDValue TargetAddress = TailCall.getOperand(1);
936 SDValue StackAdjustment = TailCall.getOperand(2);
937 assert(((TargetAddress.getOpcode() == ISD::Register &&
938 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
939 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
940 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
941 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
942 "Expecting an global address, external symbol, or register");
943 assert(StackAdjustment.getOpcode() == ISD::Constant &&
944 "Expecting a const value");
946 SmallVector<SDValue,8> Operands;
947 Operands.push_back(Chain.getOperand(0));
948 Operands.push_back(TargetAddress);
949 Operands.push_back(StackAdjustment);
950 // Copy registers used by the call. Last operand is a flag so it is not
952 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
953 Operands.push_back(Chain.getOperand(i));
955 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
962 SmallVector<SDValue, 6> RetOps;
963 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
964 // Operand #1 = Bytes To Pop
965 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
967 // Copy the result values into the output registers.
968 for (unsigned i = 0; i != RVLocs.size(); ++i) {
969 CCValAssign &VA = RVLocs[i];
970 assert(VA.isRegLoc() && "Can only return in registers!");
971 SDValue ValToCopy = Op.getOperand(i*2+1);
973 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
974 // the RET instruction and handled by the FP Stackifier.
975 if (RVLocs[i].getLocReg() == X86::ST0 ||
976 RVLocs[i].getLocReg() == X86::ST1) {
977 // If this is a copy from an xmm register to ST(0), use an FPExtend to
978 // change the value to the FP stack register class.
979 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT()))
980 ValToCopy = DAG.getNode(ISD::FP_EXTEND, MVT::f80, ValToCopy);
981 RetOps.push_back(ValToCopy);
982 // Don't emit a copytoreg.
986 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), ValToCopy, Flag);
987 Flag = Chain.getValue(1);
990 // The x86-64 ABI for returning structs by value requires that we copy
991 // the sret argument into %rax for the return. We saved the argument into
992 // a virtual register in the entry block, so now we copy the value out
994 if (Subtarget->is64Bit() &&
995 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
996 MachineFunction &MF = DAG.getMachineFunction();
997 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
998 unsigned Reg = FuncInfo->getSRetReturnReg();
1000 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1001 FuncInfo->setSRetReturnReg(Reg);
1003 SDValue Val = DAG.getCopyFromReg(Chain, Reg, getPointerTy());
1005 Chain = DAG.getCopyToReg(Chain, X86::RAX, Val, Flag);
1006 Flag = Chain.getValue(1);
1009 RetOps[0] = Chain; // Update chain.
1011 // Add the flag if we have it.
1013 RetOps.push_back(Flag);
1015 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, &RetOps[0], RetOps.size());
1019 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1020 /// appropriate copies out of appropriate physical registers. This assumes that
1021 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1022 /// being lowered. The returns a SDNode with the same number of values as the
1024 SDNode *X86TargetLowering::
1025 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1026 unsigned CallingConv, SelectionDAG &DAG) {
1028 // Assign locations to each value returned by this call.
1029 SmallVector<CCValAssign, 16> RVLocs;
1030 bool isVarArg = TheCall->isVarArg();
1031 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1032 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1034 SmallVector<SDValue, 8> ResultVals;
1036 // Copy all of the result registers out of their specified physreg.
1037 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1038 MVT CopyVT = RVLocs[i].getValVT();
1040 // If this is a call to a function that returns an fp value on the floating
1041 // point stack, but where we prefer to use the value in xmm registers, copy
1042 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1043 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1044 RVLocs[i].getLocReg() == X86::ST1) &&
1045 isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
1049 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
1050 CopyVT, InFlag).getValue(1);
1051 SDValue Val = Chain.getValue(0);
1052 InFlag = Chain.getValue(2);
1054 if (CopyVT != RVLocs[i].getValVT()) {
1055 // Round the F80 the right size, which also moves to the appropriate xmm
1057 Val = DAG.getNode(ISD::FP_ROUND, RVLocs[i].getValVT(), Val,
1058 // This truncation won't change the value.
1059 DAG.getIntPtrConstant(1));
1062 ResultVals.push_back(Val);
1065 // Merge everything together with a MERGE_VALUES node.
1066 ResultVals.push_back(Chain);
1067 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(), &ResultVals[0],
1068 ResultVals.size()).getNode();
1072 //===----------------------------------------------------------------------===//
1073 // C & StdCall & Fast Calling Convention implementation
1074 //===----------------------------------------------------------------------===//
1075 // StdCall calling convention seems to be standard for many Windows' API
1076 // routines and around. It differs from C calling convention just a little:
1077 // callee should clean up the stack, not caller. Symbols should be also
1078 // decorated in some fancy way :) It doesn't support any vector arguments.
1079 // For info on fast calling convention see Fast Calling Convention (tail call)
1080 // implementation LowerX86_32FastCCCallTo.
1082 /// AddLiveIn - This helper function adds the specified physical register to the
1083 /// MachineFunction as a live in value. It also creates a corresponding virtual
1084 /// register for it.
1085 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1086 const TargetRegisterClass *RC) {
1087 assert(RC->contains(PReg) && "Not the correct regclass!");
1088 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1089 MF.getRegInfo().addLiveIn(PReg, VReg);
1093 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1095 static bool CallIsStructReturn(CallSDNode *TheCall) {
1096 unsigned NumOps = TheCall->getNumArgs();
1100 return TheCall->getArgFlags(0).isSRet();
1103 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1104 /// return semantics.
1105 static bool ArgsAreStructReturn(SDValue Op) {
1106 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1110 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1113 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1114 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1116 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1120 switch (CallingConv) {
1123 case CallingConv::X86_StdCall:
1124 return !Subtarget->is64Bit();
1125 case CallingConv::X86_FastCall:
1126 return !Subtarget->is64Bit();
1127 case CallingConv::Fast:
1128 return PerformTailCallOpt;
1132 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1133 /// given CallingConvention value.
1134 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1135 if (Subtarget->is64Bit()) {
1136 if (Subtarget->isTargetWin64())
1137 return CC_X86_Win64_C;
1138 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1139 return CC_X86_64_TailCall;
1144 if (CC == CallingConv::X86_FastCall)
1145 return CC_X86_32_FastCall;
1146 else if (CC == CallingConv::Fast)
1147 return CC_X86_32_FastCC;
1152 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1153 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1155 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1156 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1157 if (CC == CallingConv::X86_FastCall)
1159 else if (CC == CallingConv::X86_StdCall)
1165 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1166 /// in a register before calling.
1167 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1168 return !IsTailCall && !Is64Bit &&
1169 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1170 Subtarget->isPICStyleGOT();
1173 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1174 /// address to be loaded in a register.
1176 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1177 return !Is64Bit && IsTailCall &&
1178 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1179 Subtarget->isPICStyleGOT();
1182 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1183 /// by "Src" to address "Dst" with size and alignment information specified by
1184 /// the specific parameter attribute. The copy will be passed as a byval
1185 /// function parameter.
1187 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1188 ISD::ArgFlagsTy Flags, SelectionDAG &DAG) {
1189 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1190 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(),
1191 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1194 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1195 const CCValAssign &VA,
1196 MachineFrameInfo *MFI,
1198 SDValue Root, unsigned i) {
1199 // Create the nodes corresponding to a load from this parameter slot.
1200 ISD::ArgFlagsTy Flags =
1201 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1202 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1203 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1205 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1206 // changed with more analysis.
1207 // In case of tail call optimization mark all arguments mutable. Since they
1208 // could be overwritten by lowering of arguments in case of a tail call.
1209 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1210 VA.getLocMemOffset(), isImmutable);
1211 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1212 if (Flags.isByVal())
1214 return DAG.getLoad(VA.getValVT(), Root, FIN,
1215 PseudoSourceValue::getFixedStack(FI), 0);
1219 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1220 MachineFunction &MF = DAG.getMachineFunction();
1221 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1223 const Function* Fn = MF.getFunction();
1224 if (Fn->hasExternalLinkage() &&
1225 Subtarget->isTargetCygMing() &&
1226 Fn->getName() == "main")
1227 FuncInfo->setForceFramePointer(true);
1229 // Decorate the function name.
1230 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1232 MachineFrameInfo *MFI = MF.getFrameInfo();
1233 SDValue Root = Op.getOperand(0);
1234 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1235 unsigned CC = MF.getFunction()->getCallingConv();
1236 bool Is64Bit = Subtarget->is64Bit();
1237 bool IsWin64 = Subtarget->isTargetWin64();
1239 assert(!(isVarArg && CC == CallingConv::Fast) &&
1240 "Var args not supported with calling convention fastcc");
1242 // Assign locations to all of the incoming arguments.
1243 SmallVector<CCValAssign, 16> ArgLocs;
1244 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1245 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1247 SmallVector<SDValue, 8> ArgValues;
1248 unsigned LastVal = ~0U;
1249 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1250 CCValAssign &VA = ArgLocs[i];
1251 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1253 assert(VA.getValNo() != LastVal &&
1254 "Don't support value assigned to multiple locs yet");
1255 LastVal = VA.getValNo();
1257 if (VA.isRegLoc()) {
1258 MVT RegVT = VA.getLocVT();
1259 TargetRegisterClass *RC = NULL;
1260 if (RegVT == MVT::i32)
1261 RC = X86::GR32RegisterClass;
1262 else if (Is64Bit && RegVT == MVT::i64)
1263 RC = X86::GR64RegisterClass;
1264 else if (RegVT == MVT::f32)
1265 RC = X86::FR32RegisterClass;
1266 else if (RegVT == MVT::f64)
1267 RC = X86::FR64RegisterClass;
1268 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1269 RC = X86::VR128RegisterClass;
1270 else if (RegVT.isVector()) {
1271 assert(RegVT.getSizeInBits() == 64);
1273 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1275 // Darwin calling convention passes MMX values in either GPRs or
1276 // XMMs in x86-64. Other targets pass them in memory.
1277 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1278 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1281 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1286 assert(0 && "Unknown argument type!");
1289 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1290 SDValue ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1292 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1293 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1295 if (VA.getLocInfo() == CCValAssign::SExt)
1296 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1297 DAG.getValueType(VA.getValVT()));
1298 else if (VA.getLocInfo() == CCValAssign::ZExt)
1299 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1300 DAG.getValueType(VA.getValVT()));
1302 if (VA.getLocInfo() != CCValAssign::Full)
1303 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1305 // Handle MMX values passed in GPRs.
1306 if (Is64Bit && RegVT != VA.getLocVT()) {
1307 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1308 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1309 else if (RC == X86::VR128RegisterClass) {
1310 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i64, ArgValue,
1311 DAG.getConstant(0, MVT::i64));
1312 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1316 ArgValues.push_back(ArgValue);
1318 assert(VA.isMemLoc());
1319 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1323 // The x86-64 ABI for returning structs by value requires that we copy
1324 // the sret argument into %rax for the return. Save the argument into
1325 // a virtual register so that we can access it from the return points.
1326 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1327 MachineFunction &MF = DAG.getMachineFunction();
1328 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1329 unsigned Reg = FuncInfo->getSRetReturnReg();
1331 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1332 FuncInfo->setSRetReturnReg(Reg);
1334 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), Reg, ArgValues[0]);
1335 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, Copy, Root);
1338 unsigned StackSize = CCInfo.getNextStackOffset();
1339 // align stack specially for tail calls
1340 if (PerformTailCallOpt && CC == CallingConv::Fast)
1341 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1343 // If the function takes variable number of arguments, make a frame index for
1344 // the start of the first vararg value... for expansion of llvm.va_start.
1346 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1347 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1350 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1352 // FIXME: We should really autogenerate these arrays
1353 static const unsigned GPR64ArgRegsWin64[] = {
1354 X86::RCX, X86::RDX, X86::R8, X86::R9
1356 static const unsigned XMMArgRegsWin64[] = {
1357 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1359 static const unsigned GPR64ArgRegs64Bit[] = {
1360 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1362 static const unsigned XMMArgRegs64Bit[] = {
1363 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1364 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1366 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1369 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1370 GPR64ArgRegs = GPR64ArgRegsWin64;
1371 XMMArgRegs = XMMArgRegsWin64;
1373 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1374 GPR64ArgRegs = GPR64ArgRegs64Bit;
1375 XMMArgRegs = XMMArgRegs64Bit;
1377 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1379 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1382 // For X86-64, if there are vararg parameters that are passed via
1383 // registers, then we must store them to their spots on the stack so they
1384 // may be loaded by deferencing the result of va_next.
1385 VarArgsGPOffset = NumIntRegs * 8;
1386 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1387 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1388 TotalNumXMMRegs * 16, 16);
1390 // Store the integer parameter registers.
1391 SmallVector<SDValue, 8> MemOps;
1392 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1393 SDValue FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1394 DAG.getIntPtrConstant(VarArgsGPOffset));
1395 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1396 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1397 X86::GR64RegisterClass);
1398 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1400 DAG.getStore(Val.getValue(1), Val, FIN,
1401 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1402 MemOps.push_back(Store);
1403 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1404 DAG.getIntPtrConstant(8));
1407 // Now store the XMM (fp + vector) parameter registers.
1408 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1409 DAG.getIntPtrConstant(VarArgsFPOffset));
1410 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1411 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1412 X86::VR128RegisterClass);
1413 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1415 DAG.getStore(Val.getValue(1), Val, FIN,
1416 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1417 MemOps.push_back(Store);
1418 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1419 DAG.getIntPtrConstant(16));
1421 if (!MemOps.empty())
1422 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1423 &MemOps[0], MemOps.size());
1427 ArgValues.push_back(Root);
1429 // Some CCs need callee pop.
1430 if (IsCalleePop(isVarArg, CC)) {
1431 BytesToPopOnReturn = StackSize; // Callee pops everything.
1432 BytesCallerReserves = 0;
1434 BytesToPopOnReturn = 0; // Callee pops nothing.
1435 // If this is an sret function, the return should pop the hidden pointer.
1436 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1437 BytesToPopOnReturn = 4;
1438 BytesCallerReserves = StackSize;
1442 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1443 if (CC == CallingConv::X86_FastCall)
1444 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1447 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1449 // Return the new list of results.
1450 return DAG.getNode(ISD::MERGE_VALUES, Op.getNode()->getVTList(),
1451 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1455 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1456 const SDValue &StackPtr,
1457 const CCValAssign &VA,
1459 SDValue Arg, ISD::ArgFlagsTy Flags) {
1460 unsigned LocMemOffset = VA.getLocMemOffset();
1461 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1462 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1463 if (Flags.isByVal()) {
1464 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1466 return DAG.getStore(Chain, Arg, PtrOff,
1467 PseudoSourceValue::getStack(), LocMemOffset);
1470 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1471 /// optimization is performed and it is required.
1473 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1474 SDValue &OutRetAddr,
1479 if (!IsTailCall || FPDiff==0) return Chain;
1481 // Adjust the Return address stack slot.
1482 MVT VT = getPointerTy();
1483 OutRetAddr = getReturnAddressFrameIndex(DAG);
1485 // Load the "old" Return address.
1486 OutRetAddr = DAG.getLoad(VT, Chain, OutRetAddr, NULL, 0);
1487 return SDValue(OutRetAddr.getNode(), 1);
1490 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1491 /// optimization is performed and it is required (FPDiff!=0).
1493 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1494 SDValue Chain, SDValue RetAddrFrIdx,
1495 bool Is64Bit, int FPDiff) {
1496 // Store the return address to the appropriate stack slot.
1497 if (!FPDiff) return Chain;
1498 // Calculate the new stack slot for the return address.
1499 int SlotSize = Is64Bit ? 8 : 4;
1500 int NewReturnAddrFI =
1501 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1502 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1503 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1504 Chain = DAG.getStore(Chain, RetAddrFrIdx, NewRetAddrFrIdx,
1505 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1509 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1510 MachineFunction &MF = DAG.getMachineFunction();
1511 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1512 SDValue Chain = TheCall->getChain();
1513 unsigned CC = TheCall->getCallingConv();
1514 bool isVarArg = TheCall->isVarArg();
1515 bool IsTailCall = TheCall->isTailCall() &&
1516 CC == CallingConv::Fast && PerformTailCallOpt;
1517 SDValue Callee = TheCall->getCallee();
1518 bool Is64Bit = Subtarget->is64Bit();
1519 bool IsStructRet = CallIsStructReturn(TheCall);
1521 assert(!(isVarArg && CC == CallingConv::Fast) &&
1522 "Var args not supported with calling convention fastcc");
1524 // Analyze operands of the call, assigning locations to each operand.
1525 SmallVector<CCValAssign, 16> ArgLocs;
1526 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1527 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1529 // Get a count of how many bytes are to be pushed on the stack.
1530 unsigned NumBytes = CCInfo.getNextStackOffset();
1531 if (PerformTailCallOpt && CC == CallingConv::Fast)
1532 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1536 // Lower arguments at fp - stackoffset + fpdiff.
1537 unsigned NumBytesCallerPushed =
1538 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1539 FPDiff = NumBytesCallerPushed - NumBytes;
1541 // Set the delta of movement of the returnaddr stackslot.
1542 // But only set if delta is greater than previous delta.
1543 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1544 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1547 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1549 SDValue RetAddrFrIdx;
1550 // Load return adress for tail calls.
1551 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1554 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1555 SmallVector<SDValue, 8> MemOpChains;
1558 // Walk the register/memloc assignments, inserting copies/loads. In the case
1559 // of tail call optimization arguments are handle later.
1560 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1561 CCValAssign &VA = ArgLocs[i];
1562 SDValue Arg = TheCall->getArg(i);
1563 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1564 bool isByVal = Flags.isByVal();
1566 // Promote the value if needed.
1567 switch (VA.getLocInfo()) {
1568 default: assert(0 && "Unknown loc info!");
1569 case CCValAssign::Full: break;
1570 case CCValAssign::SExt:
1571 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1573 case CCValAssign::ZExt:
1574 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1576 case CCValAssign::AExt:
1577 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1581 if (VA.isRegLoc()) {
1583 MVT RegVT = VA.getLocVT();
1584 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1585 switch (VA.getLocReg()) {
1588 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1590 // Special case: passing MMX values in GPR registers.
1591 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1594 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1595 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1596 // Special case: passing MMX values in XMM registers.
1597 Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
1598 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
1599 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
1600 DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
1601 getMOVLMask(2, DAG));
1606 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1608 if (!IsTailCall || (IsTailCall && isByVal)) {
1609 assert(VA.isMemLoc());
1610 if (StackPtr.getNode() == 0)
1611 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1613 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1614 Chain, Arg, Flags));
1619 if (!MemOpChains.empty())
1620 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1621 &MemOpChains[0], MemOpChains.size());
1623 // Build a sequence of copy-to-reg nodes chained together with token chain
1624 // and flag operands which copy the outgoing args into registers.
1626 // Tail call byval lowering might overwrite argument registers so in case of
1627 // tail call optimization the copies to registers are lowered later.
1629 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1630 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1632 InFlag = Chain.getValue(1);
1635 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1637 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1638 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1639 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1641 InFlag = Chain.getValue(1);
1643 // If we are tail calling and generating PIC/GOT style code load the address
1644 // of the callee into ecx. The value in ecx is used as target of the tail
1645 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1646 // calls on PIC/GOT architectures. Normally we would just put the address of
1647 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1648 // restored (since ebx is callee saved) before jumping to the target@PLT.
1649 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1650 // Note: The actual moving to ecx is done further down.
1651 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1652 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1653 !G->getGlobal()->hasProtectedVisibility())
1654 Callee = LowerGlobalAddress(Callee, DAG);
1655 else if (isa<ExternalSymbolSDNode>(Callee))
1656 Callee = LowerExternalSymbol(Callee,DAG);
1659 if (Is64Bit && isVarArg) {
1660 // From AMD64 ABI document:
1661 // For calls that may call functions that use varargs or stdargs
1662 // (prototype-less calls or calls to functions containing ellipsis (...) in
1663 // the declaration) %al is used as hidden argument to specify the number
1664 // of SSE registers used. The contents of %al do not need to match exactly
1665 // the number of registers, but must be an ubound on the number of SSE
1666 // registers used and is in the range 0 - 8 inclusive.
1668 // FIXME: Verify this on Win64
1669 // Count the number of XMM registers allocated.
1670 static const unsigned XMMArgRegs[] = {
1671 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1672 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1674 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1676 Chain = DAG.getCopyToReg(Chain, X86::AL,
1677 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1678 InFlag = Chain.getValue(1);
1682 // For tail calls lower the arguments to the 'real' stack slot.
1684 SmallVector<SDValue, 8> MemOpChains2;
1687 // Do not flag preceeding copytoreg stuff together with the following stuff.
1689 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1690 CCValAssign &VA = ArgLocs[i];
1691 if (!VA.isRegLoc()) {
1692 assert(VA.isMemLoc());
1693 SDValue Arg = TheCall->getArg(i);
1694 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1695 // Create frame index.
1696 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1697 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1698 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1699 FIN = DAG.getFrameIndex(FI, getPointerTy());
1701 if (Flags.isByVal()) {
1702 // Copy relative to framepointer.
1703 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1704 if (StackPtr.getNode() == 0)
1705 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1706 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1708 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1711 // Store relative to framepointer.
1712 MemOpChains2.push_back(
1713 DAG.getStore(Chain, Arg, FIN,
1714 PseudoSourceValue::getFixedStack(FI), 0));
1719 if (!MemOpChains2.empty())
1720 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1721 &MemOpChains2[0], MemOpChains2.size());
1723 // Copy arguments to their registers.
1724 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1725 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1727 InFlag = Chain.getValue(1);
1731 // Store the return address to the appropriate stack slot.
1732 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1736 // If the callee is a GlobalAddress node (quite common, every direct call is)
1737 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1738 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1739 // We should use extra load for direct calls to dllimported functions in
1741 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1742 getTargetMachine(), true))
1743 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1745 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1746 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1747 } else if (IsTailCall) {
1748 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1750 Chain = DAG.getCopyToReg(Chain,
1751 DAG.getRegister(Opc, getPointerTy()),
1753 Callee = DAG.getRegister(Opc, getPointerTy());
1754 // Add register as live out.
1755 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1758 // Returns a chain & a flag for retval copy to use.
1759 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1760 SmallVector<SDValue, 8> Ops;
1763 Ops.push_back(Chain);
1764 Ops.push_back(DAG.getIntPtrConstant(NumBytes, true));
1765 Ops.push_back(DAG.getIntPtrConstant(0, true));
1766 if (InFlag.getNode())
1767 Ops.push_back(InFlag);
1768 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1769 InFlag = Chain.getValue(1);
1771 // Returns a chain & a flag for retval copy to use.
1772 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1776 Ops.push_back(Chain);
1777 Ops.push_back(Callee);
1780 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1782 // Add argument registers to the end of the list so that they are known live
1784 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1785 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1786 RegsToPass[i].second.getValueType()));
1788 // Add an implicit use GOT pointer in EBX.
1789 if (!IsTailCall && !Is64Bit &&
1790 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1791 Subtarget->isPICStyleGOT())
1792 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1794 // Add an implicit use of AL for x86 vararg functions.
1795 if (Is64Bit && isVarArg)
1796 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1798 if (InFlag.getNode())
1799 Ops.push_back(InFlag);
1802 assert(InFlag.getNode() &&
1803 "Flag must be set. Depend on flag being set in LowerRET");
1804 Chain = DAG.getNode(X86ISD::TAILCALL,
1805 TheCall->getVTList(), &Ops[0], Ops.size());
1807 return SDValue(Chain.getNode(), Op.getResNo());
1810 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1811 InFlag = Chain.getValue(1);
1813 // Create the CALLSEQ_END node.
1814 unsigned NumBytesForCalleeToPush;
1815 if (IsCalleePop(isVarArg, CC))
1816 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1817 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1818 // If this is is a call to a struct-return function, the callee
1819 // pops the hidden struct pointer, so we have to push it back.
1820 // This is common for Darwin/X86, Linux & Mingw32 targets.
1821 NumBytesForCalleeToPush = 4;
1823 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1825 // Returns a flag for retval copy to use.
1826 Chain = DAG.getCALLSEQ_END(Chain,
1827 DAG.getIntPtrConstant(NumBytes, true),
1828 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1831 InFlag = Chain.getValue(1);
1833 // Handle result values, copying them out of physregs into vregs that we
1835 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1840 //===----------------------------------------------------------------------===//
1841 // Fast Calling Convention (tail call) implementation
1842 //===----------------------------------------------------------------------===//
1844 // Like std call, callee cleans arguments, convention except that ECX is
1845 // reserved for storing the tail called function address. Only 2 registers are
1846 // free for argument passing (inreg). Tail call optimization is performed
1848 // * tailcallopt is enabled
1849 // * caller/callee are fastcc
1850 // On X86_64 architecture with GOT-style position independent code only local
1851 // (within module) calls are supported at the moment.
1852 // To keep the stack aligned according to platform abi the function
1853 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1854 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1855 // If a tail called function callee has more arguments than the caller the
1856 // caller needs to make sure that there is room to move the RETADDR to. This is
1857 // achieved by reserving an area the size of the argument delta right after the
1858 // original REtADDR, but before the saved framepointer or the spilled registers
1859 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1871 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1872 /// for a 16 byte align requirement.
1873 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1874 SelectionDAG& DAG) {
1875 MachineFunction &MF = DAG.getMachineFunction();
1876 const TargetMachine &TM = MF.getTarget();
1877 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1878 unsigned StackAlignment = TFI.getStackAlignment();
1879 uint64_t AlignMask = StackAlignment - 1;
1880 int64_t Offset = StackSize;
1881 uint64_t SlotSize = TD->getPointerSize();
1882 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1883 // Number smaller than 12 so just add the difference.
1884 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1886 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1887 Offset = ((~AlignMask) & Offset) + StackAlignment +
1888 (StackAlignment-SlotSize);
1893 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1894 /// following the call is a return. A function is eligible if caller/callee
1895 /// calling conventions match, currently only fastcc supports tail calls, and
1896 /// the function CALL is immediatly followed by a RET.
1897 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1899 SelectionDAG& DAG) const {
1900 if (!PerformTailCallOpt)
1903 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1904 MachineFunction &MF = DAG.getMachineFunction();
1905 unsigned CallerCC = MF.getFunction()->getCallingConv();
1906 unsigned CalleeCC= TheCall->getCallingConv();
1907 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1908 SDValue Callee = TheCall->getCallee();
1909 // On x86/32Bit PIC/GOT tail calls are supported.
1910 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1911 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1914 // Can only do local tail calls (in same module, hidden or protected) on
1915 // x86_64 PIC/GOT at the moment.
1916 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1917 return G->getGlobal()->hasHiddenVisibility()
1918 || G->getGlobal()->hasProtectedVisibility();
1926 X86TargetLowering::createFastISel(MachineFunction &mf,
1927 MachineModuleInfo *mmo,
1929 DenseMap<const Value *, unsigned> &vm,
1930 DenseMap<const BasicBlock *,
1931 MachineBasicBlock *> &bm,
1932 DenseMap<const AllocaInst *, int> &am
1934 , SmallSet<Instruction*, 8> &cil
1937 return X86::createFastISel(mf, mmo, dw, vm, bm, am
1945 //===----------------------------------------------------------------------===//
1946 // Other Lowering Hooks
1947 //===----------------------------------------------------------------------===//
1950 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1951 MachineFunction &MF = DAG.getMachineFunction();
1952 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1953 int ReturnAddrIndex = FuncInfo->getRAIndex();
1955 if (ReturnAddrIndex == 0) {
1956 // Set up a frame object for the return address.
1957 uint64_t SlotSize = TD->getPointerSize();
1958 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
1959 FuncInfo->setRAIndex(ReturnAddrIndex);
1962 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1966 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
1967 /// specific condition code, returning the condition code and the LHS/RHS of the
1968 /// comparison to make.
1969 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1970 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
1972 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1973 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1974 // X > -1 -> X == 0, jump !sign.
1975 RHS = DAG.getConstant(0, RHS.getValueType());
1976 return X86::COND_NS;
1977 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1978 // X < 0 -> X == 0, jump on sign.
1980 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
1982 RHS = DAG.getConstant(0, RHS.getValueType());
1983 return X86::COND_LE;
1987 switch (SetCCOpcode) {
1988 default: assert(0 && "Invalid integer condition!");
1989 case ISD::SETEQ: return X86::COND_E;
1990 case ISD::SETGT: return X86::COND_G;
1991 case ISD::SETGE: return X86::COND_GE;
1992 case ISD::SETLT: return X86::COND_L;
1993 case ISD::SETLE: return X86::COND_LE;
1994 case ISD::SETNE: return X86::COND_NE;
1995 case ISD::SETULT: return X86::COND_B;
1996 case ISD::SETUGT: return X86::COND_A;
1997 case ISD::SETULE: return X86::COND_BE;
1998 case ISD::SETUGE: return X86::COND_AE;
2002 // First determine if it is required or is profitable to flip the operands.
2004 // If LHS is a foldable load, but RHS is not, flip the condition.
2005 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2006 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2007 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2008 std::swap(LHS, RHS);
2011 switch (SetCCOpcode) {
2017 std::swap(LHS, RHS);
2021 // On a floating point condition, the flags are set as follows:
2023 // 0 | 0 | 0 | X > Y
2024 // 0 | 0 | 1 | X < Y
2025 // 1 | 0 | 0 | X == Y
2026 // 1 | 1 | 1 | unordered
2027 switch (SetCCOpcode) {
2028 default: assert(0 && "Condcode should be pre-legalized away");
2030 case ISD::SETEQ: return X86::COND_E;
2031 case ISD::SETOLT: // flipped
2033 case ISD::SETGT: return X86::COND_A;
2034 case ISD::SETOLE: // flipped
2036 case ISD::SETGE: return X86::COND_AE;
2037 case ISD::SETUGT: // flipped
2039 case ISD::SETLT: return X86::COND_B;
2040 case ISD::SETUGE: // flipped
2042 case ISD::SETLE: return X86::COND_BE;
2044 case ISD::SETNE: return X86::COND_NE;
2045 case ISD::SETUO: return X86::COND_P;
2046 case ISD::SETO: return X86::COND_NP;
2050 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2051 /// code. Current x86 isa includes the following FP cmov instructions:
2052 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2053 static bool hasFPCMov(unsigned X86CC) {
2069 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2070 /// true if Op is undef or if its value falls within the specified range (L, H].
2071 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2072 if (Op.getOpcode() == ISD::UNDEF)
2075 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2076 return (Val >= Low && Val < Hi);
2079 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2080 /// true if Op is undef or if its value equal to the specified value.
2081 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2082 if (Op.getOpcode() == ISD::UNDEF)
2084 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2087 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2088 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2089 bool X86::isPSHUFDMask(SDNode *N) {
2090 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2092 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2095 // Check if the value doesn't reference the second vector.
2096 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2097 SDValue Arg = N->getOperand(i);
2098 if (Arg.getOpcode() == ISD::UNDEF) continue;
2099 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2100 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2107 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2108 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2109 bool X86::isPSHUFHWMask(SDNode *N) {
2110 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2112 if (N->getNumOperands() != 8)
2115 // Lower quadword copied in order.
2116 for (unsigned i = 0; i != 4; ++i) {
2117 SDValue Arg = N->getOperand(i);
2118 if (Arg.getOpcode() == ISD::UNDEF) continue;
2119 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2120 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2124 // Upper quadword shuffled.
2125 for (unsigned i = 4; i != 8; ++i) {
2126 SDValue Arg = N->getOperand(i);
2127 if (Arg.getOpcode() == ISD::UNDEF) continue;
2128 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2129 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2130 if (Val < 4 || Val > 7)
2137 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2138 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2139 bool X86::isPSHUFLWMask(SDNode *N) {
2140 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2142 if (N->getNumOperands() != 8)
2145 // Upper quadword copied in order.
2146 for (unsigned i = 4; i != 8; ++i)
2147 if (!isUndefOrEqual(N->getOperand(i), i))
2150 // Lower quadword shuffled.
2151 for (unsigned i = 0; i != 4; ++i)
2152 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2158 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2159 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2160 static bool isSHUFPMask(SDOperandPtr Elems, unsigned NumElems) {
2161 if (NumElems != 2 && NumElems != 4) return false;
2163 unsigned Half = NumElems / 2;
2164 for (unsigned i = 0; i < Half; ++i)
2165 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2167 for (unsigned i = Half; i < NumElems; ++i)
2168 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2174 bool X86::isSHUFPMask(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2176 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2179 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2180 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2181 /// half elements to come from vector 1 (which would equal the dest.) and
2182 /// the upper half to come from vector 2.
2183 static bool isCommutedSHUFP(SDOperandPtr Ops, unsigned NumOps) {
2184 if (NumOps != 2 && NumOps != 4) return false;
2186 unsigned Half = NumOps / 2;
2187 for (unsigned i = 0; i < Half; ++i)
2188 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2190 for (unsigned i = Half; i < NumOps; ++i)
2191 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2196 static bool isCommutedSHUFP(SDNode *N) {
2197 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2198 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2201 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2202 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2203 bool X86::isMOVHLPSMask(SDNode *N) {
2204 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2206 if (N->getNumOperands() != 4)
2209 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2210 return isUndefOrEqual(N->getOperand(0), 6) &&
2211 isUndefOrEqual(N->getOperand(1), 7) &&
2212 isUndefOrEqual(N->getOperand(2), 2) &&
2213 isUndefOrEqual(N->getOperand(3), 3);
2216 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2217 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2219 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2220 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2222 if (N->getNumOperands() != 4)
2225 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2226 return isUndefOrEqual(N->getOperand(0), 2) &&
2227 isUndefOrEqual(N->getOperand(1), 3) &&
2228 isUndefOrEqual(N->getOperand(2), 2) &&
2229 isUndefOrEqual(N->getOperand(3), 3);
2232 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2233 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2234 bool X86::isMOVLPMask(SDNode *N) {
2235 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2237 unsigned NumElems = N->getNumOperands();
2238 if (NumElems != 2 && NumElems != 4)
2241 for (unsigned i = 0; i < NumElems/2; ++i)
2242 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2245 for (unsigned i = NumElems/2; i < NumElems; ++i)
2246 if (!isUndefOrEqual(N->getOperand(i), i))
2252 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2253 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2255 bool X86::isMOVHPMask(SDNode *N) {
2256 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2258 unsigned NumElems = N->getNumOperands();
2259 if (NumElems != 2 && NumElems != 4)
2262 for (unsigned i = 0; i < NumElems/2; ++i)
2263 if (!isUndefOrEqual(N->getOperand(i), i))
2266 for (unsigned i = 0; i < NumElems/2; ++i) {
2267 SDValue Arg = N->getOperand(i + NumElems/2);
2268 if (!isUndefOrEqual(Arg, i + NumElems))
2275 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2276 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2277 bool static isUNPCKLMask(SDOperandPtr Elts, unsigned NumElts,
2278 bool V2IsSplat = false) {
2279 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2282 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2283 SDValue BitI = Elts[i];
2284 SDValue BitI1 = Elts[i+1];
2285 if (!isUndefOrEqual(BitI, j))
2288 if (isUndefOrEqual(BitI1, NumElts))
2291 if (!isUndefOrEqual(BitI1, j + NumElts))
2299 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2300 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2301 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2304 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2305 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2306 bool static isUNPCKHMask(SDOperandPtr Elts, unsigned NumElts,
2307 bool V2IsSplat = false) {
2308 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2311 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2312 SDValue BitI = Elts[i];
2313 SDValue BitI1 = Elts[i+1];
2314 if (!isUndefOrEqual(BitI, j + NumElts/2))
2317 if (isUndefOrEqual(BitI1, NumElts))
2320 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2328 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2329 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2330 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2333 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2334 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2336 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2337 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2339 unsigned NumElems = N->getNumOperands();
2340 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2343 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2344 SDValue BitI = N->getOperand(i);
2345 SDValue BitI1 = N->getOperand(i+1);
2347 if (!isUndefOrEqual(BitI, j))
2349 if (!isUndefOrEqual(BitI1, j))
2356 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2357 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2359 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2360 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2362 unsigned NumElems = N->getNumOperands();
2363 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2366 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2367 SDValue BitI = N->getOperand(i);
2368 SDValue BitI1 = N->getOperand(i + 1);
2370 if (!isUndefOrEqual(BitI, j))
2372 if (!isUndefOrEqual(BitI1, j))
2379 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2380 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2381 /// MOVSD, and MOVD, i.e. setting the lowest element.
2382 static bool isMOVLMask(SDOperandPtr Elts, unsigned NumElts) {
2383 if (NumElts != 2 && NumElts != 4)
2386 if (!isUndefOrEqual(Elts[0], NumElts))
2389 for (unsigned i = 1; i < NumElts; ++i) {
2390 if (!isUndefOrEqual(Elts[i], i))
2397 bool X86::isMOVLMask(SDNode *N) {
2398 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2399 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2402 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2403 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2404 /// element of vector 2 and the other elements to come from vector 1 in order.
2405 static bool isCommutedMOVL(SDOperandPtr Ops, unsigned NumOps,
2406 bool V2IsSplat = false,
2407 bool V2IsUndef = false) {
2408 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2411 if (!isUndefOrEqual(Ops[0], 0))
2414 for (unsigned i = 1; i < NumOps; ++i) {
2415 SDValue Arg = Ops[i];
2416 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2417 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2418 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2425 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2426 bool V2IsUndef = false) {
2427 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2428 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2429 V2IsSplat, V2IsUndef);
2432 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2433 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2434 bool X86::isMOVSHDUPMask(SDNode *N) {
2435 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2437 if (N->getNumOperands() != 4)
2440 // Expect 1, 1, 3, 3
2441 for (unsigned i = 0; i < 2; ++i) {
2442 SDValue Arg = N->getOperand(i);
2443 if (Arg.getOpcode() == ISD::UNDEF) continue;
2444 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2445 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2446 if (Val != 1) return false;
2450 for (unsigned i = 2; i < 4; ++i) {
2451 SDValue Arg = N->getOperand(i);
2452 if (Arg.getOpcode() == ISD::UNDEF) continue;
2453 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2454 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2455 if (Val != 3) return false;
2459 // Don't use movshdup if it can be done with a shufps.
2463 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2464 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2465 bool X86::isMOVSLDUPMask(SDNode *N) {
2466 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2468 if (N->getNumOperands() != 4)
2471 // Expect 0, 0, 2, 2
2472 for (unsigned i = 0; i < 2; ++i) {
2473 SDValue Arg = N->getOperand(i);
2474 if (Arg.getOpcode() == ISD::UNDEF) continue;
2475 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2476 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2477 if (Val != 0) return false;
2481 for (unsigned i = 2; i < 4; ++i) {
2482 SDValue Arg = N->getOperand(i);
2483 if (Arg.getOpcode() == ISD::UNDEF) continue;
2484 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2485 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2486 if (Val != 2) return false;
2490 // Don't use movshdup if it can be done with a shufps.
2494 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2495 /// specifies a identity operation on the LHS or RHS.
2496 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2497 unsigned NumElems = N->getNumOperands();
2498 for (unsigned i = 0; i < NumElems; ++i)
2499 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2504 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2505 /// a splat of a single element.
2506 static bool isSplatMask(SDNode *N) {
2507 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2509 // This is a splat operation if each element of the permute is the same, and
2510 // if the value doesn't reference the second vector.
2511 unsigned NumElems = N->getNumOperands();
2512 SDValue ElementBase;
2514 for (; i != NumElems; ++i) {
2515 SDValue Elt = N->getOperand(i);
2516 if (isa<ConstantSDNode>(Elt)) {
2522 if (!ElementBase.getNode())
2525 for (; i != NumElems; ++i) {
2526 SDValue Arg = N->getOperand(i);
2527 if (Arg.getOpcode() == ISD::UNDEF) continue;
2528 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2529 if (Arg != ElementBase) return false;
2532 // Make sure it is a splat of the first vector operand.
2533 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2536 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2537 /// we want to splat.
2538 static SDValue getSplatMaskEltNo(SDNode *N) {
2539 assert(isSplatMask(N) && "Not a splat mask");
2540 unsigned NumElems = N->getNumOperands();
2541 SDValue ElementBase;
2543 for (; i != NumElems; ++i) {
2544 SDValue Elt = N->getOperand(i);
2545 if (isa<ConstantSDNode>(Elt))
2548 assert(0 && " No splat value found!");
2553 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2554 /// a splat of a single element and it's a 2 or 4 element mask.
2555 bool X86::isSplatMask(SDNode *N) {
2556 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2558 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2559 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2561 return ::isSplatMask(N);
2564 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2565 /// specifies a splat of zero element.
2566 bool X86::isSplatLoMask(SDNode *N) {
2567 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2569 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2570 if (!isUndefOrEqual(N->getOperand(i), 0))
2575 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2576 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2577 bool X86::isMOVDDUPMask(SDNode *N) {
2578 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2580 unsigned e = N->getNumOperands() / 2;
2581 for (unsigned i = 0; i < e; ++i)
2582 if (!isUndefOrEqual(N->getOperand(i), i))
2584 for (unsigned i = 0; i < e; ++i)
2585 if (!isUndefOrEqual(N->getOperand(e+i), i))
2590 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2591 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2593 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2594 unsigned NumOperands = N->getNumOperands();
2595 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2597 for (unsigned i = 0; i < NumOperands; ++i) {
2599 SDValue Arg = N->getOperand(NumOperands-i-1);
2600 if (Arg.getOpcode() != ISD::UNDEF)
2601 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2602 if (Val >= NumOperands) Val -= NumOperands;
2604 if (i != NumOperands - 1)
2611 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2612 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2614 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2616 // 8 nodes, but we only care about the last 4.
2617 for (unsigned i = 7; i >= 4; --i) {
2619 SDValue Arg = N->getOperand(i);
2620 if (Arg.getOpcode() != ISD::UNDEF)
2621 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2630 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2631 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2633 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2635 // 8 nodes, but we only care about the first 4.
2636 for (int i = 3; i >= 0; --i) {
2638 SDValue Arg = N->getOperand(i);
2639 if (Arg.getOpcode() != ISD::UNDEF)
2640 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2649 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2650 /// specifies a 8 element shuffle that can be broken into a pair of
2651 /// PSHUFHW and PSHUFLW.
2652 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2653 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2655 if (N->getNumOperands() != 8)
2658 // Lower quadword shuffled.
2659 for (unsigned i = 0; i != 4; ++i) {
2660 SDValue Arg = N->getOperand(i);
2661 if (Arg.getOpcode() == ISD::UNDEF) continue;
2662 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2663 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2668 // Upper quadword shuffled.
2669 for (unsigned i = 4; i != 8; ++i) {
2670 SDValue Arg = N->getOperand(i);
2671 if (Arg.getOpcode() == ISD::UNDEF) continue;
2672 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2673 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2674 if (Val < 4 || Val > 7)
2681 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2682 /// values in ther permute mask.
2683 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2684 SDValue &V2, SDValue &Mask,
2685 SelectionDAG &DAG) {
2686 MVT VT = Op.getValueType();
2687 MVT MaskVT = Mask.getValueType();
2688 MVT EltVT = MaskVT.getVectorElementType();
2689 unsigned NumElems = Mask.getNumOperands();
2690 SmallVector<SDValue, 8> MaskVec;
2692 for (unsigned i = 0; i != NumElems; ++i) {
2693 SDValue Arg = Mask.getOperand(i);
2694 if (Arg.getOpcode() == ISD::UNDEF) {
2695 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2698 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2699 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2701 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2703 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2707 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2708 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2711 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2712 /// the two vector operands have swapped position.
2714 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG) {
2715 MVT MaskVT = Mask.getValueType();
2716 MVT EltVT = MaskVT.getVectorElementType();
2717 unsigned NumElems = Mask.getNumOperands();
2718 SmallVector<SDValue, 8> MaskVec;
2719 for (unsigned i = 0; i != NumElems; ++i) {
2720 SDValue Arg = Mask.getOperand(i);
2721 if (Arg.getOpcode() == ISD::UNDEF) {
2722 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2725 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2726 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2728 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2730 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2732 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2736 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2737 /// match movhlps. The lower half elements should come from upper half of
2738 /// V1 (and in order), and the upper half elements should come from the upper
2739 /// half of V2 (and in order).
2740 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2741 unsigned NumElems = Mask->getNumOperands();
2744 for (unsigned i = 0, e = 2; i != e; ++i)
2745 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2747 for (unsigned i = 2; i != 4; ++i)
2748 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2753 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2754 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2756 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2757 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2759 N = N->getOperand(0).getNode();
2760 if (!ISD::isNON_EXTLoad(N))
2763 *LD = cast<LoadSDNode>(N);
2767 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2768 /// match movlp{s|d}. The lower half elements should come from lower half of
2769 /// V1 (and in order), and the upper half elements should come from the upper
2770 /// half of V2 (and in order). And since V1 will become the source of the
2771 /// MOVLP, it must be either a vector load or a scalar load to vector.
2772 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2773 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2775 // Is V2 is a vector load, don't do this transformation. We will try to use
2776 // load folding shufps op.
2777 if (ISD::isNON_EXTLoad(V2))
2780 unsigned NumElems = Mask->getNumOperands();
2781 if (NumElems != 2 && NumElems != 4)
2783 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2784 if (!isUndefOrEqual(Mask->getOperand(i), i))
2786 for (unsigned i = NumElems/2; i != NumElems; ++i)
2787 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2792 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2794 static bool isSplatVector(SDNode *N) {
2795 if (N->getOpcode() != ISD::BUILD_VECTOR)
2798 SDValue SplatValue = N->getOperand(0);
2799 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2800 if (N->getOperand(i) != SplatValue)
2805 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2807 static bool isUndefShuffle(SDNode *N) {
2808 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2811 SDValue V1 = N->getOperand(0);
2812 SDValue V2 = N->getOperand(1);
2813 SDValue Mask = N->getOperand(2);
2814 unsigned NumElems = Mask.getNumOperands();
2815 for (unsigned i = 0; i != NumElems; ++i) {
2816 SDValue Arg = Mask.getOperand(i);
2817 if (Arg.getOpcode() != ISD::UNDEF) {
2818 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2819 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2821 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2828 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2830 static inline bool isZeroNode(SDValue Elt) {
2831 return ((isa<ConstantSDNode>(Elt) &&
2832 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2833 (isa<ConstantFPSDNode>(Elt) &&
2834 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2837 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2838 /// to an zero vector.
2839 static bool isZeroShuffle(SDNode *N) {
2840 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2843 SDValue V1 = N->getOperand(0);
2844 SDValue V2 = N->getOperand(1);
2845 SDValue Mask = N->getOperand(2);
2846 unsigned NumElems = Mask.getNumOperands();
2847 for (unsigned i = 0; i != NumElems; ++i) {
2848 SDValue Arg = Mask.getOperand(i);
2849 if (Arg.getOpcode() == ISD::UNDEF)
2852 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2853 if (Idx < NumElems) {
2854 unsigned Opc = V1.getNode()->getOpcode();
2855 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2857 if (Opc != ISD::BUILD_VECTOR ||
2858 !isZeroNode(V1.getNode()->getOperand(Idx)))
2860 } else if (Idx >= NumElems) {
2861 unsigned Opc = V2.getNode()->getOpcode();
2862 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2864 if (Opc != ISD::BUILD_VECTOR ||
2865 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2872 /// getZeroVector - Returns a vector of specified type with all zero elements.
2874 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG) {
2875 assert(VT.isVector() && "Expected a vector type");
2877 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2878 // type. This ensures they get CSE'd.
2880 if (VT.getSizeInBits() == 64) { // MMX
2881 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2882 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2883 } else if (HasSSE2) { // SSE2
2884 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2885 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2887 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2888 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4f32, Cst, Cst, Cst, Cst);
2890 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2893 /// getOnesVector - Returns a vector of specified type with all bits set.
2895 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG) {
2896 assert(VT.isVector() && "Expected a vector type");
2898 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2899 // type. This ensures they get CSE'd.
2900 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2902 if (VT.getSizeInBits() == 64) // MMX
2903 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2905 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2906 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2910 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2911 /// that point to V2 points to its first element.
2912 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2913 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2915 bool Changed = false;
2916 SmallVector<SDValue, 8> MaskVec;
2917 unsigned NumElems = Mask.getNumOperands();
2918 for (unsigned i = 0; i != NumElems; ++i) {
2919 SDValue Arg = Mask.getOperand(i);
2920 if (Arg.getOpcode() != ISD::UNDEF) {
2921 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2922 if (Val > NumElems) {
2923 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2927 MaskVec.push_back(Arg);
2931 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2932 &MaskVec[0], MaskVec.size());
2936 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2937 /// operation of specified width.
2938 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2939 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2940 MVT BaseVT = MaskVT.getVectorElementType();
2942 SmallVector<SDValue, 8> MaskVec;
2943 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2944 for (unsigned i = 1; i != NumElems; ++i)
2945 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2946 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2949 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2950 /// of specified width.
2951 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2952 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2953 MVT BaseVT = MaskVT.getVectorElementType();
2954 SmallVector<SDValue, 8> MaskVec;
2955 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2956 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2957 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2959 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2962 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2963 /// of specified width.
2964 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2965 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2966 MVT BaseVT = MaskVT.getVectorElementType();
2967 unsigned Half = NumElems/2;
2968 SmallVector<SDValue, 8> MaskVec;
2969 for (unsigned i = 0; i != Half; ++i) {
2970 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2971 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2973 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2976 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
2977 /// element #0 of a vector with the specified index, leaving the rest of the
2978 /// elements in place.
2979 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
2980 SelectionDAG &DAG) {
2981 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2982 MVT BaseVT = MaskVT.getVectorElementType();
2983 SmallVector<SDValue, 8> MaskVec;
2984 // Element #0 of the result gets the elt we are replacing.
2985 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
2986 for (unsigned i = 1; i != NumElems; ++i)
2987 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
2988 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2991 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2992 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
2993 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
2994 MVT VT = Op.getValueType();
2997 SDValue V1 = Op.getOperand(0);
2998 SDValue Mask = Op.getOperand(2);
2999 unsigned MaskNumElems = Mask.getNumOperands();
3000 unsigned NumElems = MaskNumElems;
3001 // Special handling of v4f32 -> v4i32.
3002 if (VT != MVT::v4f32) {
3003 // Find which element we want to splat.
3004 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3005 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3006 // unpack elements to the correct location
3007 while (NumElems > 4) {
3008 if (EltNo < NumElems/2) {
3009 Mask = getUnpacklMask(MaskNumElems, DAG);
3011 Mask = getUnpackhMask(MaskNumElems, DAG);
3012 EltNo -= NumElems/2;
3014 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
3017 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3018 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
3021 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3022 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3023 DAG.getNode(ISD::UNDEF, PVT), Mask);
3024 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3027 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3028 /// load that's promoted to vector, or a load bitcasted.
3029 static bool isVectorLoad(SDValue Op) {
3030 assert(Op.getValueType().isVector() && "Expected a vector type");
3031 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3032 Op.getOpcode() == ISD::BIT_CONVERT) {
3033 return isa<LoadSDNode>(Op.getOperand(0));
3035 return isa<LoadSDNode>(Op);
3039 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3041 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3042 SelectionDAG &DAG, bool HasSSE3) {
3043 // If we have sse3 and shuffle has more than one use or input is a load, then
3044 // use movddup. Otherwise, use movlhps.
3045 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3046 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3047 MVT VT = Op.getValueType();
3050 unsigned NumElems = PVT.getVectorNumElements();
3051 if (NumElems == 2) {
3052 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3053 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
3055 assert(NumElems == 4);
3056 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3057 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3058 Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst0, Cst1, Cst0, Cst1);
3061 V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
3062 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
3063 DAG.getNode(ISD::UNDEF, PVT), Mask);
3064 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3067 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3068 /// vector of zero or undef vector. This produces a shuffle where the low
3069 /// element of V2 is swizzled into the zero/undef vector, landing at element
3070 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3071 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3072 bool isZero, bool HasSSE2,
3073 SelectionDAG &DAG) {
3074 MVT VT = V2.getValueType();
3076 ? getZeroVector(VT, HasSSE2, DAG) : DAG.getNode(ISD::UNDEF, VT);
3077 unsigned NumElems = V2.getValueType().getVectorNumElements();
3078 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3079 MVT EVT = MaskVT.getVectorElementType();
3080 SmallVector<SDValue, 16> MaskVec;
3081 for (unsigned i = 0; i != NumElems; ++i)
3082 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3083 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3085 MaskVec.push_back(DAG.getConstant(i, EVT));
3086 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3087 &MaskVec[0], MaskVec.size());
3088 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3091 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3092 /// a shuffle that is zero.
3094 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3095 unsigned NumElems, bool Low,
3096 SelectionDAG &DAG) {
3097 unsigned NumZeros = 0;
3098 for (unsigned i = 0; i < NumElems; ++i) {
3099 unsigned Index = Low ? i : NumElems-i-1;
3100 SDValue Idx = Mask.getOperand(Index);
3101 if (Idx.getOpcode() == ISD::UNDEF) {
3105 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3106 if (Elt.getNode() && isZeroNode(Elt))
3114 /// isVectorShift - Returns true if the shuffle can be implemented as a
3115 /// logical left or right shift of a vector.
3116 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3117 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3118 unsigned NumElems = Mask.getNumOperands();
3121 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3124 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3129 bool SeenV1 = false;
3130 bool SeenV2 = false;
3131 for (unsigned i = NumZeros; i < NumElems; ++i) {
3132 unsigned Val = isLeft ? (i - NumZeros) : i;
3133 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3134 if (Idx.getOpcode() == ISD::UNDEF)
3136 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3137 if (Index < NumElems)
3146 if (SeenV1 && SeenV2)
3149 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3155 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3157 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3158 unsigned NumNonZero, unsigned NumZero,
3159 SelectionDAG &DAG, TargetLowering &TLI) {
3165 for (unsigned i = 0; i < 16; ++i) {
3166 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3167 if (ThisIsNonZero && First) {
3169 V = getZeroVector(MVT::v8i16, true, DAG);
3171 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3176 SDValue ThisElt(0, 0), LastElt(0, 0);
3177 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3178 if (LastIsNonZero) {
3179 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3181 if (ThisIsNonZero) {
3182 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3183 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3184 ThisElt, DAG.getConstant(8, MVT::i8));
3186 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3190 if (ThisElt.getNode())
3191 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3192 DAG.getIntPtrConstant(i/2));
3196 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3199 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3201 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3202 unsigned NumNonZero, unsigned NumZero,
3203 SelectionDAG &DAG, TargetLowering &TLI) {
3209 for (unsigned i = 0; i < 8; ++i) {
3210 bool isNonZero = (NonZeros & (1 << i)) != 0;
3214 V = getZeroVector(MVT::v8i16, true, DAG);
3216 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3219 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3220 DAG.getIntPtrConstant(i));
3227 /// getVShift - Return a vector logical shift node.
3229 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3230 unsigned NumBits, SelectionDAG &DAG,
3231 const TargetLowering &TLI) {
3232 bool isMMX = VT.getSizeInBits() == 64;
3233 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3234 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3235 SrcOp = DAG.getNode(ISD::BIT_CONVERT, ShVT, SrcOp);
3236 return DAG.getNode(ISD::BIT_CONVERT, VT,
3237 DAG.getNode(Opc, ShVT, SrcOp,
3238 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3242 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3243 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3244 if (ISD::isBuildVectorAllZeros(Op.getNode())
3245 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3246 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3247 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3248 // eliminated on x86-32 hosts.
3249 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3252 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3253 return getOnesVector(Op.getValueType(), DAG);
3254 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG);
3257 MVT VT = Op.getValueType();
3258 MVT EVT = VT.getVectorElementType();
3259 unsigned EVTBits = EVT.getSizeInBits();
3261 unsigned NumElems = Op.getNumOperands();
3262 unsigned NumZero = 0;
3263 unsigned NumNonZero = 0;
3264 unsigned NonZeros = 0;
3265 bool IsAllConstants = true;
3266 SmallSet<SDValue, 8> Values;
3267 for (unsigned i = 0; i < NumElems; ++i) {
3268 SDValue Elt = Op.getOperand(i);
3269 if (Elt.getOpcode() == ISD::UNDEF)
3272 if (Elt.getOpcode() != ISD::Constant &&
3273 Elt.getOpcode() != ISD::ConstantFP)
3274 IsAllConstants = false;
3275 if (isZeroNode(Elt))
3278 NonZeros |= (1 << i);
3283 if (NumNonZero == 0) {
3284 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3285 return DAG.getNode(ISD::UNDEF, VT);
3288 // Special case for single non-zero, non-undef, element.
3289 if (NumNonZero == 1 && NumElems <= 4) {
3290 unsigned Idx = CountTrailingZeros_32(NonZeros);
3291 SDValue Item = Op.getOperand(Idx);
3293 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3294 // the value are obviously zero, truncate the value to i32 and do the
3295 // insertion that way. Only do this if the value is non-constant or if the
3296 // value is a constant being inserted into element 0. It is cheaper to do
3297 // a constant pool load than it is to do a movd + shuffle.
3298 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3299 (!IsAllConstants || Idx == 0)) {
3300 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3301 // Handle MMX and SSE both.
3302 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3303 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3305 // Truncate the value (which may itself be a constant) to i32, and
3306 // convert it to a vector with movd (S2V+shuffle to zero extend).
3307 Item = DAG.getNode(ISD::TRUNCATE, MVT::i32, Item);
3308 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VecVT, Item);
3309 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3310 Subtarget->hasSSE2(), DAG);
3312 // Now we have our 32-bit value zero extended in the low element of
3313 // a vector. If Idx != 0, swizzle it into place.
3316 Item, DAG.getNode(ISD::UNDEF, Item.getValueType()),
3317 getSwapEltZeroMask(VecElts, Idx, DAG)
3319 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, VecVT, Ops, 3);
3321 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Item);
3325 // If we have a constant or non-constant insertion into the low element of
3326 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3327 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3328 // depending on what the source datatype is. Because we can only get here
3329 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3331 // Don't do this for i64 values on x86-32.
3332 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3333 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3334 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3335 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3336 Subtarget->hasSSE2(), DAG);
3339 // Is it a vector logical left shift?
3340 if (NumElems == 2 && Idx == 1 &&
3341 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3342 unsigned NumBits = VT.getSizeInBits();
3343 return getVShift(true, VT,
3344 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(1)),
3345 NumBits/2, DAG, *this);
3348 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3351 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3352 // is a non-constant being inserted into an element other than the low one,
3353 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3354 // movd/movss) to move this into the low element, then shuffle it into
3356 if (EVTBits == 32) {
3357 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3359 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3360 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3361 Subtarget->hasSSE2(), DAG);
3362 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3363 MVT MaskEVT = MaskVT.getVectorElementType();
3364 SmallVector<SDValue, 8> MaskVec;
3365 for (unsigned i = 0; i < NumElems; i++)
3366 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3367 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3368 &MaskVec[0], MaskVec.size());
3369 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3370 DAG.getNode(ISD::UNDEF, VT), Mask);
3374 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3375 if (Values.size() == 1)
3378 // A vector full of immediates; various special cases are already
3379 // handled, so this is best done with a single constant-pool load.
3383 // Let legalizer expand 2-wide build_vectors.
3384 if (EVTBits == 64) {
3385 if (NumNonZero == 1) {
3386 // One half is zero or undef.
3387 unsigned Idx = CountTrailingZeros_32(NonZeros);
3388 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT,
3389 Op.getOperand(Idx));
3390 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3391 Subtarget->hasSSE2(), DAG);
3396 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3397 if (EVTBits == 8 && NumElems == 16) {
3398 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3400 if (V.getNode()) return V;
3403 if (EVTBits == 16 && NumElems == 8) {
3404 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3406 if (V.getNode()) return V;
3409 // If element VT is == 32 bits, turn it into a number of shuffles.
3410 SmallVector<SDValue, 8> V;
3412 if (NumElems == 4 && NumZero > 0) {
3413 for (unsigned i = 0; i < 4; ++i) {
3414 bool isZero = !(NonZeros & (1 << i));
3416 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG);
3418 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3421 for (unsigned i = 0; i < 2; ++i) {
3422 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3425 V[i] = V[i*2]; // Must be a zero vector.
3428 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3429 getMOVLMask(NumElems, DAG));
3432 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3433 getMOVLMask(NumElems, DAG));
3436 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3437 getUnpacklMask(NumElems, DAG));
3442 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3443 MVT EVT = MaskVT.getVectorElementType();
3444 SmallVector<SDValue, 8> MaskVec;
3445 bool Reverse = (NonZeros & 0x3) == 2;
3446 for (unsigned i = 0; i < 2; ++i)
3448 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3450 MaskVec.push_back(DAG.getConstant(i, EVT));
3451 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3452 for (unsigned i = 0; i < 2; ++i)
3454 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3456 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3457 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3458 &MaskVec[0], MaskVec.size());
3459 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3462 if (Values.size() > 2) {
3463 // Expand into a number of unpckl*.
3465 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3466 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3467 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3468 SDValue UnpckMask = getUnpacklMask(NumElems, DAG);
3469 for (unsigned i = 0; i < NumElems; ++i)
3470 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3472 while (NumElems != 0) {
3473 for (unsigned i = 0; i < NumElems; ++i)
3474 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3485 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3486 SDValue PermMask, SelectionDAG &DAG,
3487 TargetLowering &TLI) {
3489 MVT MaskVT = MVT::getIntVectorWithNumElements(8);
3490 MVT MaskEVT = MaskVT.getVectorElementType();
3491 MVT PtrVT = TLI.getPointerTy();
3492 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3493 PermMask.getNode()->op_end());
3495 // First record which half of which vector the low elements come from.
3496 SmallVector<unsigned, 4> LowQuad(4);
3497 for (unsigned i = 0; i < 4; ++i) {
3498 SDValue Elt = MaskElts[i];
3499 if (Elt.getOpcode() == ISD::UNDEF)
3501 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3502 int QuadIdx = EltIdx / 4;
3506 int BestLowQuad = -1;
3507 unsigned MaxQuad = 1;
3508 for (unsigned i = 0; i < 4; ++i) {
3509 if (LowQuad[i] > MaxQuad) {
3511 MaxQuad = LowQuad[i];
3515 // Record which half of which vector the high elements come from.
3516 SmallVector<unsigned, 4> HighQuad(4);
3517 for (unsigned i = 4; i < 8; ++i) {
3518 SDValue Elt = MaskElts[i];
3519 if (Elt.getOpcode() == ISD::UNDEF)
3521 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3522 int QuadIdx = EltIdx / 4;
3523 ++HighQuad[QuadIdx];
3526 int BestHighQuad = -1;
3528 for (unsigned i = 0; i < 4; ++i) {
3529 if (HighQuad[i] > MaxQuad) {
3531 MaxQuad = HighQuad[i];
3535 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3536 if (BestLowQuad != -1 || BestHighQuad != -1) {
3537 // First sort the 4 chunks in order using shufpd.
3538 SmallVector<SDValue, 8> MaskVec;
3540 if (BestLowQuad != -1)
3541 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3543 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3545 if (BestHighQuad != -1)
3546 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3548 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3550 SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3551 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3552 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3553 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3554 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3556 // Now sort high and low parts separately.
3557 BitVector InOrder(8);
3558 if (BestLowQuad != -1) {
3559 // Sort lower half in order using PSHUFLW.
3561 bool AnyOutOrder = false;
3563 for (unsigned i = 0; i != 4; ++i) {
3564 SDValue Elt = MaskElts[i];
3565 if (Elt.getOpcode() == ISD::UNDEF) {
3566 MaskVec.push_back(Elt);
3569 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3573 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3575 // If this element is in the right place after this shuffle, then
3577 if ((int)(EltIdx / 4) == BestLowQuad)
3582 for (unsigned i = 4; i != 8; ++i)
3583 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3584 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3585 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3589 if (BestHighQuad != -1) {
3590 // Sort high half in order using PSHUFHW if possible.
3593 for (unsigned i = 0; i != 4; ++i)
3594 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3596 bool AnyOutOrder = false;
3597 for (unsigned i = 4; i != 8; ++i) {
3598 SDValue Elt = MaskElts[i];
3599 if (Elt.getOpcode() == ISD::UNDEF) {
3600 MaskVec.push_back(Elt);
3603 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3607 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3609 // If this element is in the right place after this shuffle, then
3611 if ((int)(EltIdx / 4) == BestHighQuad)
3617 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3618 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3622 // The other elements are put in the right place using pextrw and pinsrw.
3623 for (unsigned i = 0; i != 8; ++i) {
3626 SDValue Elt = MaskElts[i];
3627 if (Elt.getOpcode() == ISD::UNDEF)
3629 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3630 SDValue ExtOp = (EltIdx < 8)
3631 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3632 DAG.getConstant(EltIdx, PtrVT))
3633 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3634 DAG.getConstant(EltIdx - 8, PtrVT));
3635 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3636 DAG.getConstant(i, PtrVT));
3642 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
3643 // few as possible. First, let's find out how many elements are already in the
3645 unsigned V1InOrder = 0;
3646 unsigned V1FromV1 = 0;
3647 unsigned V2InOrder = 0;
3648 unsigned V2FromV2 = 0;
3649 SmallVector<SDValue, 8> V1Elts;
3650 SmallVector<SDValue, 8> V2Elts;
3651 for (unsigned i = 0; i < 8; ++i) {
3652 SDValue Elt = MaskElts[i];
3653 if (Elt.getOpcode() == ISD::UNDEF) {
3654 V1Elts.push_back(Elt);
3655 V2Elts.push_back(Elt);
3660 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3662 V1Elts.push_back(Elt);
3663 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3665 } else if (EltIdx == i+8) {
3666 V1Elts.push_back(Elt);
3667 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3669 } else if (EltIdx < 8) {
3670 V1Elts.push_back(Elt);
3671 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3674 V1Elts.push_back(Elt);
3675 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3680 if (V2InOrder > V1InOrder) {
3681 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3683 std::swap(V1Elts, V2Elts);
3684 std::swap(V1FromV1, V2FromV2);
3687 if ((V1FromV1 + V1InOrder) != 8) {
3688 // Some elements are from V2.
3690 // If there are elements that are from V1 but out of place,
3691 // then first sort them in place
3692 SmallVector<SDValue, 8> MaskVec;
3693 for (unsigned i = 0; i < 8; ++i) {
3694 SDValue Elt = V1Elts[i];
3695 if (Elt.getOpcode() == ISD::UNDEF) {
3696 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3699 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3701 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3703 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3705 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3706 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3710 for (unsigned i = 0; i < 8; ++i) {
3711 SDValue Elt = V1Elts[i];
3712 if (Elt.getOpcode() == ISD::UNDEF)
3714 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3717 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3718 DAG.getConstant(EltIdx - 8, PtrVT));
3719 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3720 DAG.getConstant(i, PtrVT));
3724 // All elements are from V1.
3726 for (unsigned i = 0; i < 8; ++i) {
3727 SDValue Elt = V1Elts[i];
3728 if (Elt.getOpcode() == ISD::UNDEF)
3730 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3731 SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3732 DAG.getConstant(EltIdx, PtrVT));
3733 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3734 DAG.getConstant(i, PtrVT));
3740 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3741 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3742 /// done when every pair / quad of shuffle mask elements point to elements in
3743 /// the right sequence. e.g.
3744 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3746 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3748 SDValue PermMask, SelectionDAG &DAG,
3749 TargetLowering &TLI) {
3750 unsigned NumElems = PermMask.getNumOperands();
3751 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3752 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3753 MVT MaskEltVT = MaskVT.getVectorElementType();
3755 switch (VT.getSimpleVT()) {
3756 default: assert(false && "Unexpected!");
3757 case MVT::v4f32: NewVT = MVT::v2f64; break;
3758 case MVT::v4i32: NewVT = MVT::v2i64; break;
3759 case MVT::v8i16: NewVT = MVT::v4i32; break;
3760 case MVT::v16i8: NewVT = MVT::v4i32; break;
3763 if (NewWidth == 2) {
3769 unsigned Scale = NumElems / NewWidth;
3770 SmallVector<SDValue, 8> MaskVec;
3771 for (unsigned i = 0; i < NumElems; i += Scale) {
3772 unsigned StartIdx = ~0U;
3773 for (unsigned j = 0; j < Scale; ++j) {
3774 SDValue Elt = PermMask.getOperand(i+j);
3775 if (Elt.getOpcode() == ISD::UNDEF)
3777 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3778 if (StartIdx == ~0U)
3779 StartIdx = EltIdx - (EltIdx % Scale);
3780 if (EltIdx != StartIdx + j)
3783 if (StartIdx == ~0U)
3784 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEltVT));
3786 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
3789 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3790 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3791 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3792 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3793 &MaskVec[0], MaskVec.size()));
3796 /// getVZextMovL - Return a zero-extending vector move low node.
3798 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3799 SDValue SrcOp, SelectionDAG &DAG,
3800 const X86Subtarget *Subtarget) {
3801 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3802 LoadSDNode *LD = NULL;
3803 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3804 LD = dyn_cast<LoadSDNode>(SrcOp);
3806 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3808 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3809 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3810 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3811 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3812 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3814 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3815 return DAG.getNode(ISD::BIT_CONVERT, VT,
3816 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3817 DAG.getNode(ISD::SCALAR_TO_VECTOR, OpVT,
3824 return DAG.getNode(ISD::BIT_CONVERT, VT,
3825 DAG.getNode(X86ISD::VZEXT_MOVL, OpVT,
3826 DAG.getNode(ISD::BIT_CONVERT, OpVT, SrcOp)));
3829 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3832 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
3833 SDValue PermMask, MVT VT, SelectionDAG &DAG) {
3834 MVT MaskVT = PermMask.getValueType();
3835 MVT MaskEVT = MaskVT.getVectorElementType();
3836 SmallVector<std::pair<int, int>, 8> Locs;
3838 SmallVector<SDValue, 8> Mask1(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3841 for (unsigned i = 0; i != 4; ++i) {
3842 SDValue Elt = PermMask.getOperand(i);
3843 if (Elt.getOpcode() == ISD::UNDEF) {
3844 Locs[i] = std::make_pair(-1, -1);
3846 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3847 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
3849 Locs[i] = std::make_pair(0, NumLo);
3853 Locs[i] = std::make_pair(1, NumHi);
3855 Mask1[2+NumHi] = Elt;
3861 if (NumLo <= 2 && NumHi <= 2) {
3862 // If no more than two elements come from either vector. This can be
3863 // implemented with two shuffles. First shuffle gather the elements.
3864 // The second shuffle, which takes the first shuffle as both of its
3865 // vector operands, put the elements into the right order.
3866 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3867 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3868 &Mask1[0], Mask1.size()));
3870 SmallVector<SDValue, 8> Mask2(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3871 for (unsigned i = 0; i != 4; ++i) {
3872 if (Locs[i].first == -1)
3875 unsigned Idx = (i < 2) ? 0 : 4;
3876 Idx += Locs[i].first * 2 + Locs[i].second;
3877 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3881 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3882 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3883 &Mask2[0], Mask2.size()));
3884 } else if (NumLo == 3 || NumHi == 3) {
3885 // Otherwise, we must have three elements from one vector, call it X, and
3886 // one element from the other, call it Y. First, use a shufps to build an
3887 // intermediate vector with the one element from Y and the element from X
3888 // that will be in the same half in the final destination (the indexes don't
3889 // matter). Then, use a shufps to build the final vector, taking the half
3890 // containing the element from Y from the intermediate, and the other half
3893 // Normalize it so the 3 elements come from V1.
3894 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3898 // Find the element from V2.
3900 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3901 SDValue Elt = PermMask.getOperand(HiIndex);
3902 if (Elt.getOpcode() == ISD::UNDEF)
3904 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
3909 Mask1[0] = PermMask.getOperand(HiIndex);
3910 Mask1[1] = DAG.getNode(ISD::UNDEF, MaskEVT);
3911 Mask1[2] = PermMask.getOperand(HiIndex^1);
3912 Mask1[3] = DAG.getNode(ISD::UNDEF, MaskEVT);
3913 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3914 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3917 Mask1[0] = PermMask.getOperand(0);
3918 Mask1[1] = PermMask.getOperand(1);
3919 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
3920 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
3921 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3922 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3924 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
3925 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
3926 Mask1[2] = PermMask.getOperand(2);
3927 Mask1[3] = PermMask.getOperand(3);
3928 if (Mask1[2].getOpcode() != ISD::UNDEF)
3930 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
3932 if (Mask1[3].getOpcode() != ISD::UNDEF)
3934 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
3936 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1,
3937 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &Mask1[0], 4));
3941 // Break it into (shuffle shuffle_hi, shuffle_lo).
3943 SmallVector<SDValue,8> LoMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3944 SmallVector<SDValue,8> HiMask(4, DAG.getNode(ISD::UNDEF, MaskEVT));
3945 SmallVector<SDValue,8> *MaskPtr = &LoMask;
3946 unsigned MaskIdx = 0;
3949 for (unsigned i = 0; i != 4; ++i) {
3956 SDValue Elt = PermMask.getOperand(i);
3957 if (Elt.getOpcode() == ISD::UNDEF) {
3958 Locs[i] = std::make_pair(-1, -1);
3959 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
3960 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3961 (*MaskPtr)[LoIdx] = Elt;
3964 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3965 (*MaskPtr)[HiIdx] = Elt;
3970 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3971 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3972 &LoMask[0], LoMask.size()));
3973 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3974 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3975 &HiMask[0], HiMask.size()));
3976 SmallVector<SDValue, 8> MaskOps;
3977 for (unsigned i = 0; i != 4; ++i) {
3978 if (Locs[i].first == -1) {
3979 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3981 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3982 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3985 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3986 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3987 &MaskOps[0], MaskOps.size()));
3991 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3992 SDValue V1 = Op.getOperand(0);
3993 SDValue V2 = Op.getOperand(1);
3994 SDValue PermMask = Op.getOperand(2);
3995 MVT VT = Op.getValueType();
3996 unsigned NumElems = PermMask.getNumOperands();
3997 bool isMMX = VT.getSizeInBits() == 64;
3998 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3999 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4000 bool V1IsSplat = false;
4001 bool V2IsSplat = false;
4003 if (isUndefShuffle(Op.getNode()))
4004 return DAG.getNode(ISD::UNDEF, VT);
4006 if (isZeroShuffle(Op.getNode()))
4007 return getZeroVector(VT, Subtarget->hasSSE2(), DAG);
4009 if (isIdentityMask(PermMask.getNode()))
4011 else if (isIdentityMask(PermMask.getNode(), true))
4014 // Canonicalize movddup shuffles.
4015 if (V2IsUndef && Subtarget->hasSSE2() &&
4016 VT.getSizeInBits() == 128 &&
4017 X86::isMOVDDUPMask(PermMask.getNode()))
4018 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4020 if (isSplatMask(PermMask.getNode())) {
4021 if (isMMX || NumElems < 4) return Op;
4022 // Promote it to a v4{if}32 splat.
4023 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4026 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4028 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4029 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
4030 if (NewOp.getNode())
4031 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
4032 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4033 // FIXME: Figure out a cleaner way to do this.
4034 // Try to make use of movq to zero out the top part.
4035 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4036 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4038 if (NewOp.getNode()) {
4039 SDValue NewV1 = NewOp.getOperand(0);
4040 SDValue NewV2 = NewOp.getOperand(1);
4041 SDValue NewMask = NewOp.getOperand(2);
4042 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4043 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4044 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget);
4047 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4048 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4050 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4051 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4056 // Check if this can be converted into a logical shift.
4057 bool isLeft = false;
4060 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4061 if (isShift && ShVal.hasOneUse()) {
4062 // If the shifted value has multiple uses, it may be cheaper to use
4063 // v_set0 + movlhps or movhlps, etc.
4064 MVT EVT = VT.getVectorElementType();
4065 ShAmt *= EVT.getSizeInBits();
4066 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4069 if (X86::isMOVLMask(PermMask.getNode())) {
4072 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4073 return getVZextMovL(VT, VT, V2, DAG, Subtarget);
4078 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4079 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4080 X86::isMOVHLPSMask(PermMask.getNode()) ||
4081 X86::isMOVHPMask(PermMask.getNode()) ||
4082 X86::isMOVLPMask(PermMask.getNode())))
4085 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4086 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4087 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4090 // No better options. Use a vshl / vsrl.
4091 MVT EVT = VT.getVectorElementType();
4092 ShAmt *= EVT.getSizeInBits();
4093 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this);
4096 bool Commuted = false;
4097 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4098 // 1,1,1,1 -> v8i16 though.
4099 V1IsSplat = isSplatVector(V1.getNode());
4100 V2IsSplat = isSplatVector(V2.getNode());
4102 // Canonicalize the splat or undef, if present, to be on the RHS.
4103 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4104 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4105 std::swap(V1IsSplat, V2IsSplat);
4106 std::swap(V1IsUndef, V2IsUndef);
4110 // FIXME: Figure out a cleaner way to do this.
4111 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4112 if (V2IsUndef) return V1;
4113 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4115 // V2 is a splat, so the mask may be malformed. That is, it may point
4116 // to any V2 element. The instruction selectior won't like this. Get
4117 // a corrected mask and commute to form a proper MOVS{S|D}.
4118 SDValue NewMask = getMOVLMask(NumElems, DAG);
4119 if (NewMask.getNode() != PermMask.getNode())
4120 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4125 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4126 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4127 X86::isUNPCKLMask(PermMask.getNode()) ||
4128 X86::isUNPCKHMask(PermMask.getNode()))
4132 // Normalize mask so all entries that point to V2 points to its first
4133 // element then try to match unpck{h|l} again. If match, return a
4134 // new vector_shuffle with the corrected mask.
4135 SDValue NewMask = NormalizeMask(PermMask, DAG);
4136 if (NewMask.getNode() != PermMask.getNode()) {
4137 if (X86::isUNPCKLMask(PermMask.getNode(), true)) {
4138 SDValue NewMask = getUnpacklMask(NumElems, DAG);
4139 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4140 } else if (X86::isUNPCKHMask(PermMask.getNode(), true)) {
4141 SDValue NewMask = getUnpackhMask(NumElems, DAG);
4142 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
4147 // Normalize the node to match x86 shuffle ops if needed
4148 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4149 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4152 // Commute is back and try unpck* again.
4153 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4154 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4155 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4156 X86::isUNPCKLMask(PermMask.getNode()) ||
4157 X86::isUNPCKHMask(PermMask.getNode()))
4161 // Try PSHUF* first, then SHUFP*.
4162 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4163 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4164 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4165 if (V2.getOpcode() != ISD::UNDEF)
4166 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
4167 DAG.getNode(ISD::UNDEF, VT), PermMask);
4172 if (Subtarget->hasSSE2() &&
4173 (X86::isPSHUFDMask(PermMask.getNode()) ||
4174 X86::isPSHUFHWMask(PermMask.getNode()) ||
4175 X86::isPSHUFLWMask(PermMask.getNode()))) {
4177 if (VT == MVT::v4f32) {
4179 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
4180 DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
4181 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4182 } else if (V2.getOpcode() != ISD::UNDEF)
4183 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
4184 DAG.getNode(ISD::UNDEF, RVT), PermMask);
4186 Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
4190 // Binary or unary shufps.
4191 if (X86::isSHUFPMask(PermMask.getNode()) ||
4192 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4196 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4197 if (VT == MVT::v8i16) {
4198 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
4199 if (NewOp.getNode())
4203 // Handle all 4 wide cases with a number of shuffles except for MMX.
4204 if (NumElems == 4 && !isMMX)
4205 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);
4211 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4212 SelectionDAG &DAG) {
4213 MVT VT = Op.getValueType();
4214 if (VT.getSizeInBits() == 8) {
4215 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
4216 Op.getOperand(0), Op.getOperand(1));
4217 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4218 DAG.getValueType(VT));
4219 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4220 } else if (VT.getSizeInBits() == 16) {
4221 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4222 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4224 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4225 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4226 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
4229 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
4230 Op.getOperand(0), Op.getOperand(1));
4231 SDValue Assert = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
4232 DAG.getValueType(VT));
4233 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4234 } else if (VT == MVT::f32) {
4235 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4236 // the result back to FR32 register. It's only worth matching if the
4237 // result has a single use which is a store or a bitcast to i32. And in
4238 // the case of a store, it's not worth it if the index is a constant 0,
4239 // because a MOVSSmr can be used instead, which is smaller and faster.
4240 if (!Op.hasOneUse())
4242 SDNode *User = *Op.getNode()->use_begin();
4243 if ((User->getOpcode() != ISD::STORE ||
4244 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4245 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4246 (User->getOpcode() != ISD::BIT_CONVERT ||
4247 User->getValueType(0) != MVT::i32))
4249 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4250 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
4252 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
4253 } else if (VT == MVT::i32) {
4254 // ExtractPS works with constant index.
4255 if (isa<ConstantSDNode>(Op.getOperand(1)))
4263 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4264 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4267 if (Subtarget->hasSSE41()) {
4268 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4273 MVT VT = Op.getValueType();
4274 // TODO: handle v16i8.
4275 if (VT.getSizeInBits() == 16) {
4276 SDValue Vec = Op.getOperand(0);
4277 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4279 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
4280 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
4281 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
4283 // Transform it so it match pextrw which produces a 32-bit result.
4284 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4285 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
4286 Op.getOperand(0), Op.getOperand(1));
4287 SDValue Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
4288 DAG.getValueType(VT));
4289 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
4290 } else if (VT.getSizeInBits() == 32) {
4291 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4294 // SHUFPS the element to the lowest double word, then movss.
4295 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4296 SmallVector<SDValue, 8> IdxVec;
4298 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4300 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4302 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4304 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4305 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4306 &IdxVec[0], IdxVec.size());
4307 SDValue Vec = Op.getOperand(0);
4308 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4309 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4310 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4311 DAG.getIntPtrConstant(0));
4312 } else if (VT.getSizeInBits() == 64) {
4313 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4314 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4315 // to match extract_elt for f64.
4316 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4320 // UNPCKHPD the element to the lowest double word, then movsd.
4321 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4322 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4323 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4324 SmallVector<SDValue, 8> IdxVec;
4325 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4327 push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
4328 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4329 &IdxVec[0], IdxVec.size());
4330 SDValue Vec = Op.getOperand(0);
4331 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
4332 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
4333 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
4334 DAG.getIntPtrConstant(0));
4341 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4342 MVT VT = Op.getValueType();
4343 MVT EVT = VT.getVectorElementType();
4345 SDValue N0 = Op.getOperand(0);
4346 SDValue N1 = Op.getOperand(1);
4347 SDValue N2 = Op.getOperand(2);
4349 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4350 isa<ConstantSDNode>(N2)) {
4351 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4353 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4355 if (N1.getValueType() != MVT::i32)
4356 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4357 if (N2.getValueType() != MVT::i32)
4358 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4359 return DAG.getNode(Opc, VT, N0, N1, N2);
4360 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4361 // Bits [7:6] of the constant are the source select. This will always be
4362 // zero here. The DAG Combiner may combine an extract_elt index into these
4363 // bits. For example (insert (extract, 3), 2) could be matched by putting
4364 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4365 // Bits [5:4] of the constant are the destination select. This is the
4366 // value of the incoming immediate.
4367 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4368 // combine either bitwise AND or insert of float 0.0 to set these bits.
4369 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4370 return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
4371 } else if (EVT == MVT::i32) {
4372 // InsertPS works with constant index.
4373 if (isa<ConstantSDNode>(N2))
4380 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4381 MVT VT = Op.getValueType();
4382 MVT EVT = VT.getVectorElementType();
4384 if (Subtarget->hasSSE41())
4385 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4390 SDValue N0 = Op.getOperand(0);
4391 SDValue N1 = Op.getOperand(1);
4392 SDValue N2 = Op.getOperand(2);
4394 if (EVT.getSizeInBits() == 16) {
4395 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4396 // as its second argument.
4397 if (N1.getValueType() != MVT::i32)
4398 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4399 if (N2.getValueType() != MVT::i32)
4400 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4401 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4407 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4408 if (Op.getValueType() == MVT::v2f32)
4409 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
4410 DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
4411 DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
4412 Op.getOperand(0))));
4414 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4415 MVT VT = MVT::v2i32;
4416 switch (Op.getValueType().getSimpleVT()) {
4423 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
4424 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
4427 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4428 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4429 // one of the above mentioned nodes. It has to be wrapped because otherwise
4430 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4431 // be used to form addressing mode. These wrapped nodes will be selected
4434 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4435 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4436 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
4438 CP->getAlignment());
4439 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4440 // With PIC, the address is actually $g + Offset.
4441 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4442 !Subtarget->isPICStyleRIPRel()) {
4443 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4444 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4452 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV,
4454 SelectionDAG &DAG) const {
4455 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4456 bool ExtraLoadRequired =
4457 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4459 // Create the TargetGlobalAddress node, folding in the constant
4460 // offset if it is legal.
4462 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4463 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4466 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4467 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4469 // With PIC, the address is actually $g + Offset.
4470 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4471 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4472 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4476 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4477 // load the value at address GV, not the value of GV itself. This means that
4478 // the GlobalAddress must be in the base or index register of the address, not
4479 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4480 // The same applies for external symbols during PIC codegen
4481 if (ExtraLoadRequired)
4482 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
4483 PseudoSourceValue::getGOT(), 0);
4485 // If there was a non-zero offset that we didn't fold, create an explicit
4488 Result = DAG.getNode(ISD::ADD, getPointerTy(), Result,
4489 DAG.getConstant(Offset, getPointerTy()));
4495 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4496 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4497 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4498 return LowerGlobalAddress(GV, Offset, DAG);
4501 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4503 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4506 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4507 DAG.getNode(X86ISD::GlobalBaseReg,
4509 InFlag = Chain.getValue(1);
4511 // emit leal symbol@TLSGD(,%ebx,1), %eax
4512 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4513 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4514 GA->getValueType(0),
4516 SDValue Ops[] = { Chain, TGA, InFlag };
4517 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4518 InFlag = Result.getValue(2);
4519 Chain = Result.getValue(1);
4521 // call ___tls_get_addr. This function receives its argument in
4522 // the register EAX.
4523 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4524 InFlag = Chain.getValue(1);
4526 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4527 SDValue Ops1[] = { Chain,
4528 DAG.getTargetExternalSymbol("___tls_get_addr",
4530 DAG.getRegister(X86::EAX, PtrVT),
4531 DAG.getRegister(X86::EBX, PtrVT),
4533 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4534 InFlag = Chain.getValue(1);
4536 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4539 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4541 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4543 SDValue InFlag, Chain;
4545 // emit leaq symbol@TLSGD(%rip), %rdi
4546 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4547 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4548 GA->getValueType(0),
4550 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4551 SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
4552 Chain = Result.getValue(1);
4553 InFlag = Result.getValue(2);
4555 // call __tls_get_addr. This function receives its argument in
4556 // the register RDI.
4557 Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
4558 InFlag = Chain.getValue(1);
4560 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4561 SDValue Ops1[] = { Chain,
4562 DAG.getTargetExternalSymbol("__tls_get_addr",
4564 DAG.getRegister(X86::RDI, PtrVT),
4566 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
4567 InFlag = Chain.getValue(1);
4569 return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
4572 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4573 // "local exec" model.
4574 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4576 // Get the Thread Pointer
4577 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4578 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4580 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4581 GA->getValueType(0),
4583 SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4585 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4586 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
4587 PseudoSourceValue::getGOT(), 0);
4589 // The address of the thread local variable is the add of the thread
4590 // pointer with the offset of the variable.
4591 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4595 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4596 // TODO: implement the "local dynamic" model
4597 // TODO: implement the "initial exec"model for pic executables
4598 assert(Subtarget->isTargetELF() &&
4599 "TLS not implemented for non-ELF targets");
4600 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4601 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4602 // otherwise use the "Local Exec"TLS Model
4603 if (Subtarget->is64Bit()) {
4604 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4606 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4607 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4609 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4614 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4615 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4616 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4617 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4618 // With PIC, the address is actually $g + Offset.
4619 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4620 !Subtarget->isPICStyleRIPRel()) {
4621 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4622 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4629 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4630 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4631 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4632 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4633 // With PIC, the address is actually $g + Offset.
4634 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4635 !Subtarget->isPICStyleRIPRel()) {
4636 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4637 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4644 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4645 /// take a 2 x i32 value to shift plus a shift amount.
4646 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4647 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4648 MVT VT = Op.getValueType();
4649 unsigned VTBits = VT.getSizeInBits();
4650 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4651 SDValue ShOpLo = Op.getOperand(0);
4652 SDValue ShOpHi = Op.getOperand(1);
4653 SDValue ShAmt = Op.getOperand(2);
4654 SDValue Tmp1 = isSRA ?
4655 DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
4656 DAG.getConstant(0, VT);
4659 if (Op.getOpcode() == ISD::SHL_PARTS) {
4660 Tmp2 = DAG.getNode(X86ISD::SHLD, VT, ShOpHi, ShOpLo, ShAmt);
4661 Tmp3 = DAG.getNode(ISD::SHL, VT, ShOpLo, ShAmt);
4663 Tmp2 = DAG.getNode(X86ISD::SHRD, VT, ShOpLo, ShOpHi, ShAmt);
4664 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, VT, ShOpHi, ShAmt);
4667 SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4668 DAG.getConstant(VTBits, MVT::i8));
4669 SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
4670 AndNode, DAG.getConstant(0, MVT::i8));
4673 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4674 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4675 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4677 if (Op.getOpcode() == ISD::SHL_PARTS) {
4678 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4679 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4681 Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
4682 Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
4685 SDValue Ops[2] = { Lo, Hi };
4686 return DAG.getMergeValues(Ops, 2);
4689 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4690 MVT SrcVT = Op.getOperand(0).getValueType();
4691 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4692 "Unknown SINT_TO_FP to lower!");
4694 // These are really Legal; caller falls through into that case.
4695 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4697 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4698 Subtarget->is64Bit())
4701 unsigned Size = SrcVT.getSizeInBits()/8;
4702 MachineFunction &MF = DAG.getMachineFunction();
4703 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4704 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4705 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4707 PseudoSourceValue::getFixedStack(SSFI), 0);
4711 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4713 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4715 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4716 SmallVector<SDValue, 8> Ops;
4717 Ops.push_back(Chain);
4718 Ops.push_back(StackSlot);
4719 Ops.push_back(DAG.getValueType(SrcVT));
4720 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
4721 Tys, &Ops[0], Ops.size());
4724 Chain = Result.getValue(1);
4725 SDValue InFlag = Result.getValue(2);
4727 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4728 // shouldn't be necessary except that RFP cannot be live across
4729 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4730 MachineFunction &MF = DAG.getMachineFunction();
4731 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4732 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4733 Tys = DAG.getVTList(MVT::Other);
4734 SmallVector<SDValue, 8> Ops;
4735 Ops.push_back(Chain);
4736 Ops.push_back(Result);
4737 Ops.push_back(StackSlot);
4738 Ops.push_back(DAG.getValueType(Op.getValueType()));
4739 Ops.push_back(InFlag);
4740 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4741 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
4742 PseudoSourceValue::getFixedStack(SSFI), 0);
4748 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4749 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4750 // This algorithm is not obvious. Here it is in C code, more or less:
4752 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4753 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4754 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4756 // Copy ints to xmm registers.
4757 __m128i xh = _mm_cvtsi32_si128( hi );
4758 __m128i xl = _mm_cvtsi32_si128( lo );
4760 // Combine into low half of a single xmm register.
4761 __m128i x = _mm_unpacklo_epi32( xh, xl );
4765 // Merge in appropriate exponents to give the integer bits the right
4767 x = _mm_unpacklo_epi32( x, exp );
4769 // Subtract away the biases to deal with the IEEE-754 double precision
4771 d = _mm_sub_pd( (__m128d) x, bias );
4773 // All conversions up to here are exact. The correctly rounded result is
4774 // calculated using the current rounding mode using the following
4776 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4777 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4778 // store doesn't really need to be here (except
4779 // maybe to zero the other double)
4784 // Build some magic constants.
4785 std::vector<Constant*> CV0;
4786 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4787 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4788 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4789 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4790 Constant *C0 = ConstantVector::get(CV0);
4791 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 4);
4793 std::vector<Constant*> CV1;
4794 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4795 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4796 Constant *C1 = ConstantVector::get(CV1);
4797 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 4);
4799 SmallVector<SDValue, 4> MaskVec;
4800 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
4801 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
4802 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
4803 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
4804 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, &MaskVec[0],
4806 SmallVector<SDValue, 4> MaskVec2;
4807 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
4808 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
4809 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec2[0],
4812 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4i32,
4813 DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
4815 DAG.getIntPtrConstant(1)));
4816 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4i32,
4817 DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
4819 DAG.getIntPtrConstant(0)));
4820 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32,
4821 XR1, XR2, UnpcklMask);
4822 SDValue CLod0 = DAG.getLoad(MVT::v4i32, DAG.getEntryNode(), CPIdx0,
4823 PseudoSourceValue::getConstantPool(), 0,
4825 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32,
4826 Unpck1, CLod0, UnpcklMask);
4827 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64, Unpck2);
4828 SDValue CLod1 = DAG.getLoad(MVT::v2f64, CLod0.getValue(1), CPIdx1,
4829 PseudoSourceValue::getConstantPool(), 0,
4831 SDValue Sub = DAG.getNode(ISD::FSUB, MVT::v2f64, XR2F, CLod1);
4833 // Add the halves; easiest way is to swap them into another reg first.
4834 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2f64,
4835 Sub, Sub, ShufMask);
4836 SDValue Add = DAG.getNode(ISD::FADD, MVT::v2f64, Shuf, Sub);
4837 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f64, Add,
4838 DAG.getIntPtrConstant(0));
4841 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4842 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4843 // FP constant to bias correct the final result.
4844 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4847 // Load the 32-bit value into an XMM register.
4848 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4i32,
4849 DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
4851 DAG.getIntPtrConstant(0)));
4853 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f64,
4854 DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64, Load),
4855 DAG.getIntPtrConstant(0));
4857 // Or the load with the bias.
4858 SDValue Or = DAG.getNode(ISD::OR, MVT::v2i64,
4859 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64,
4860 DAG.getNode(ISD::SCALAR_TO_VECTOR,
4862 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64,
4863 DAG.getNode(ISD::SCALAR_TO_VECTOR,
4864 MVT::v2f64, Bias)));
4865 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f64,
4866 DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64, Or),
4867 DAG.getIntPtrConstant(0));
4869 // Subtract the bias.
4870 SDValue Sub = DAG.getNode(ISD::FSUB, MVT::f64, Or, Bias);
4872 // Handle final rounding.
4873 MVT DestVT = Op.getValueType();
4875 if (DestVT.bitsLT(MVT::f64)) {
4876 return DAG.getNode(ISD::FP_ROUND, DestVT, Sub,
4877 DAG.getIntPtrConstant(0));
4878 } else if (DestVT.bitsGT(MVT::f64)) {
4879 return DAG.getNode(ISD::FP_EXTEND, DestVT, Sub);
4882 // Handle final rounding.
4886 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4887 SDValue N0 = Op.getOperand(0);
4889 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4890 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4891 // the optimization here.
4892 if (DAG.SignBitIsZero(N0))
4893 return DAG.getNode(ISD::SINT_TO_FP, Op.getValueType(), N0);
4895 MVT SrcVT = N0.getValueType();
4896 if (SrcVT == MVT::i64) {
4897 // We only handle SSE2 f64 target here; caller can handle the rest.
4898 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4901 return LowerUINT_TO_FP_i64(Op, DAG);
4902 } else if (SrcVT == MVT::i32) {
4903 return LowerUINT_TO_FP_i32(Op, DAG);
4906 assert(0 && "Unknown UINT_TO_FP to lower!");
4910 std::pair<SDValue,SDValue> X86TargetLowering::
4911 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4912 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4913 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4914 "Unknown FP_TO_SINT to lower!");
4916 // These are really Legal.
4917 if (Op.getValueType() == MVT::i32 &&
4918 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4919 return std::make_pair(SDValue(), SDValue());
4920 if (Subtarget->is64Bit() &&
4921 Op.getValueType() == MVT::i64 &&
4922 Op.getOperand(0).getValueType() != MVT::f80)
4923 return std::make_pair(SDValue(), SDValue());
4925 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4927 MachineFunction &MF = DAG.getMachineFunction();
4928 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4929 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4930 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4932 switch (Op.getValueType().getSimpleVT()) {
4933 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4934 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4935 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4936 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4939 SDValue Chain = DAG.getEntryNode();
4940 SDValue Value = Op.getOperand(0);
4941 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4942 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4943 Chain = DAG.getStore(Chain, Value, StackSlot,
4944 PseudoSourceValue::getFixedStack(SSFI), 0);
4945 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4947 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4949 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4950 Chain = Value.getValue(1);
4951 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4952 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4955 // Build the FP_TO_INT*_IN_MEM
4956 SDValue Ops[] = { Chain, Value, StackSlot };
4957 SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4959 return std::make_pair(FIST, StackSlot);
4962 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4963 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4964 SDValue FIST = Vals.first, StackSlot = Vals.second;
4965 if (FIST.getNode() == 0) return SDValue();
4968 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4971 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4972 MVT VT = Op.getValueType();
4975 EltVT = VT.getVectorElementType();
4976 std::vector<Constant*> CV;
4977 if (EltVT == MVT::f64) {
4978 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4982 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4988 Constant *C = ConstantVector::get(CV);
4989 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4990 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
4991 PseudoSourceValue::getConstantPool(), 0,
4993 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4996 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4997 MVT VT = Op.getValueType();
4999 unsigned EltNum = 1;
5000 if (VT.isVector()) {
5001 EltVT = VT.getVectorElementType();
5002 EltNum = VT.getVectorNumElements();
5004 std::vector<Constant*> CV;
5005 if (EltVT == MVT::f64) {
5006 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5010 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5016 Constant *C = ConstantVector::get(CV);
5017 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5018 SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
5019 PseudoSourceValue::getConstantPool(), 0,
5021 if (VT.isVector()) {
5022 return DAG.getNode(ISD::BIT_CONVERT, VT,
5023 DAG.getNode(ISD::XOR, MVT::v2i64,
5024 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
5025 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
5027 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
5031 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5032 SDValue Op0 = Op.getOperand(0);
5033 SDValue Op1 = Op.getOperand(1);
5034 MVT VT = Op.getValueType();
5035 MVT SrcVT = Op1.getValueType();
5037 // If second operand is smaller, extend it first.
5038 if (SrcVT.bitsLT(VT)) {
5039 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
5042 // And if it is bigger, shrink it first.
5043 if (SrcVT.bitsGT(VT)) {
5044 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
5048 // At this point the operands and the result should have the same
5049 // type, and that won't be f80 since that is not custom lowered.
5051 // First get the sign bit of second operand.
5052 std::vector<Constant*> CV;
5053 if (SrcVT == MVT::f64) {
5054 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5055 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5057 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5058 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5059 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5060 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5062 Constant *C = ConstantVector::get(CV);
5063 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5064 SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
5065 PseudoSourceValue::getConstantPool(), 0,
5067 SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
5069 // Shift sign bit right or left if the two operands have different types.
5070 if (SrcVT.bitsGT(VT)) {
5071 // Op0 is MVT::f32, Op1 is MVT::f64.
5072 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
5073 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
5074 DAG.getConstant(32, MVT::i32));
5075 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
5076 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
5077 DAG.getIntPtrConstant(0));
5080 // Clear first operand sign bit.
5082 if (VT == MVT::f64) {
5083 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5084 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5086 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5087 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5088 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5089 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5091 C = ConstantVector::get(CV);
5092 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
5093 SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
5094 PseudoSourceValue::getConstantPool(), 0,
5096 SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
5098 // Or the value with the sign bit.
5099 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
5102 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5103 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5104 SDValue Op0 = Op.getOperand(0);
5105 SDValue Op1 = Op.getOperand(1);
5106 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5108 // Lower (X & (1 << N)) == 0 to BT.
5109 // Lower ((X >>u N) & 1) != 0 to BT.
5110 // Lower ((X >>s N) & 1) != 0 to BT.
5111 if (Op0.getOpcode() == ISD::AND &&
5113 Op1.getOpcode() == ISD::Constant &&
5114 Op0.getOperand(1).getOpcode() == ISD::Constant &&
5115 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5116 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5117 ConstantSDNode *CmpRHS = cast<ConstantSDNode>(Op1);
5118 SDValue AndLHS = Op0.getOperand(0);
5119 if (CmpRHS->getZExtValue() == 0 && AndRHS->getZExtValue() == 1 &&
5120 AndLHS.getOpcode() == ISD::SRL) {
5121 SDValue LHS = AndLHS.getOperand(0);
5122 SDValue RHS = AndLHS.getOperand(1);
5124 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5125 // instruction. Since the shift amount is in-range-or-undefined, we know
5126 // that doing a bittest on the i16 value is ok. We extend to i32 because
5127 // the encoding for the i16 version is larger than the i32 version.
5128 if (LHS.getValueType() == MVT::i8)
5129 LHS = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, LHS);
5131 // If the operand types disagree, extend the shift amount to match. Since
5132 // BT ignores high bits (like shifts) we can use anyextend.
5133 if (LHS.getValueType() != RHS.getValueType())
5134 RHS = DAG.getNode(ISD::ANY_EXTEND, LHS.getValueType(), RHS);
5136 SDValue BT = DAG.getNode(X86ISD::BT, MVT::i32, LHS, RHS);
5137 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5138 return DAG.getNode(X86ISD::SETCC, MVT::i8,
5139 DAG.getConstant(Cond, MVT::i8), BT);
5143 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5144 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5146 SDValue Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
5147 return DAG.getNode(X86ISD::SETCC, MVT::i8,
5148 DAG.getConstant(X86CC, MVT::i8), Cond);
5151 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5153 SDValue Op0 = Op.getOperand(0);
5154 SDValue Op1 = Op.getOperand(1);
5155 SDValue CC = Op.getOperand(2);
5156 MVT VT = Op.getValueType();
5157 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5158 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5162 MVT VT0 = Op0.getValueType();
5163 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5164 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5167 switch (SetCCOpcode) {
5170 case ISD::SETEQ: SSECC = 0; break;
5172 case ISD::SETGT: Swap = true; // Fallthrough
5174 case ISD::SETOLT: SSECC = 1; break;
5176 case ISD::SETGE: Swap = true; // Fallthrough
5178 case ISD::SETOLE: SSECC = 2; break;
5179 case ISD::SETUO: SSECC = 3; break;
5181 case ISD::SETNE: SSECC = 4; break;
5182 case ISD::SETULE: Swap = true;
5183 case ISD::SETUGE: SSECC = 5; break;
5184 case ISD::SETULT: Swap = true;
5185 case ISD::SETUGT: SSECC = 6; break;
5186 case ISD::SETO: SSECC = 7; break;
5189 std::swap(Op0, Op1);
5191 // In the two special cases we can't handle, emit two comparisons.
5193 if (SetCCOpcode == ISD::SETUEQ) {
5195 UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5196 EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5197 return DAG.getNode(ISD::OR, VT, UNORD, EQ);
5199 else if (SetCCOpcode == ISD::SETONE) {
5201 ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5202 NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5203 return DAG.getNode(ISD::AND, VT, ORD, NEQ);
5205 assert(0 && "Illegal FP comparison");
5207 // Handle all other FP comparisons here.
5208 return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5211 // We are handling one of the integer comparisons here. Since SSE only has
5212 // GT and EQ comparisons for integer, swapping operands and multiple
5213 // operations may be required for some comparisons.
5214 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5215 bool Swap = false, Invert = false, FlipSigns = false;
5217 switch (VT.getSimpleVT()) {
5219 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5220 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5221 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5222 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5225 switch (SetCCOpcode) {
5227 case ISD::SETNE: Invert = true;
5228 case ISD::SETEQ: Opc = EQOpc; break;
5229 case ISD::SETLT: Swap = true;
5230 case ISD::SETGT: Opc = GTOpc; break;
5231 case ISD::SETGE: Swap = true;
5232 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5233 case ISD::SETULT: Swap = true;
5234 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5235 case ISD::SETUGE: Swap = true;
5236 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5239 std::swap(Op0, Op1);
5241 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5242 // bits of the inputs before performing those operations.
5244 MVT EltVT = VT.getVectorElementType();
5245 SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
5246 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5247 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
5249 Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
5250 Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
5253 SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);
5255 // If the logical-not of the result is required, perform that now.
5257 MVT EltVT = VT.getVectorElementType();
5258 SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
5259 std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
5260 SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
5262 Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
5267 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5268 static bool isX86LogicalCmp(unsigned Opc) {
5269 return Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI;
5272 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5273 bool addTest = true;
5274 SDValue Cond = Op.getOperand(0);
5277 if (Cond.getOpcode() == ISD::SETCC)
5278 Cond = LowerSETCC(Cond, DAG);
5280 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5281 // setting operand in place of the X86ISD::SETCC.
5282 if (Cond.getOpcode() == X86ISD::SETCC) {
5283 CC = Cond.getOperand(0);
5285 SDValue Cmp = Cond.getOperand(1);
5286 unsigned Opc = Cmp.getOpcode();
5287 MVT VT = Op.getValueType();
5289 bool IllegalFPCMov = false;
5290 if (VT.isFloatingPoint() && !VT.isVector() &&
5291 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5292 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5294 if (isX86LogicalCmp(Opc) && !IllegalFPCMov) {
5301 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5302 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5305 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5307 SmallVector<SDValue, 4> Ops;
5308 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5309 // condition is true.
5310 Ops.push_back(Op.getOperand(2));
5311 Ops.push_back(Op.getOperand(1));
5313 Ops.push_back(Cond);
5314 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
5317 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5318 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5319 // from the AND / OR.
5320 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5321 Opc = Op.getOpcode();
5322 if (Opc != ISD::OR && Opc != ISD::AND)
5324 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5325 Op.getOperand(0).hasOneUse() &&
5326 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5327 Op.getOperand(1).hasOneUse());
5330 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5331 bool addTest = true;
5332 SDValue Chain = Op.getOperand(0);
5333 SDValue Cond = Op.getOperand(1);
5334 SDValue Dest = Op.getOperand(2);
5337 if (Cond.getOpcode() == ISD::SETCC)
5338 Cond = LowerSETCC(Cond, DAG);
5340 // FIXME: LowerXALUO doesn't handle these!!
5341 else if (Cond.getOpcode() == X86ISD::ADD ||
5342 Cond.getOpcode() == X86ISD::SUB ||
5343 Cond.getOpcode() == X86ISD::SMUL ||
5344 Cond.getOpcode() == X86ISD::UMUL)
5345 Cond = LowerXALUO(Cond, DAG);
5348 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5349 // setting operand in place of the X86ISD::SETCC.
5350 if (Cond.getOpcode() == X86ISD::SETCC) {
5351 CC = Cond.getOperand(0);
5353 SDValue Cmp = Cond.getOperand(1);
5354 unsigned Opc = Cmp.getOpcode();
5355 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5356 if (isX86LogicalCmp(Opc) || Opc == X86ISD::BT) {
5360 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5364 // These can only come from an arithmetic instruction with overflow,
5365 // e.g. SADDO, UADDO.
5366 Cond = Cond.getNode()->getOperand(1);
5373 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5374 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5375 unsigned Opc = Cmp.getOpcode();
5376 if (CondOpc == ISD::OR) {
5377 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5378 // two branches instead of an explicit OR instruction with a
5380 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5381 isX86LogicalCmp(Opc)) {
5382 CC = Cond.getOperand(0).getOperand(0);
5383 Chain = DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5384 Chain, Dest, CC, Cmp);
5385 CC = Cond.getOperand(1).getOperand(0);
5389 } else { // ISD::AND
5390 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5391 // two branches instead of an explicit AND instruction with a
5392 // separate test. However, we only do this if this block doesn't
5393 // have a fall-through edge, because this requires an explicit
5394 // jmp when the condition is false.
5395 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5396 isX86LogicalCmp(Opc) &&
5397 Op.getNode()->hasOneUse()) {
5398 X86::CondCode CCode =
5399 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5400 CCode = X86::GetOppositeBranchCondition(CCode);
5401 CC = DAG.getConstant(CCode, MVT::i8);
5402 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5403 // Look for an unconditional branch following this conditional branch.
5404 // We need this because we need to reverse the successors in order
5405 // to implement FCMP_OEQ.
5406 if (User.getOpcode() == ISD::BR) {
5407 SDValue FalseBB = User.getOperand(1);
5409 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5410 assert(NewBR == User);
5413 Chain = DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5414 Chain, Dest, CC, Cmp);
5415 X86::CondCode CCode =
5416 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5417 CCode = X86::GetOppositeBranchCondition(CCode);
5418 CC = DAG.getConstant(CCode, MVT::i8);
5428 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5429 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
5431 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
5432 Chain, Dest, CC, Cond);
5436 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5437 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5438 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5439 // that the guard pages used by the OS virtual memory manager are allocated in
5440 // correct sequence.
5442 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5443 SelectionDAG &DAG) {
5444 assert(Subtarget->isTargetCygMing() &&
5445 "This should be used only on Cygwin/Mingw targets");
5448 SDValue Chain = Op.getOperand(0);
5449 SDValue Size = Op.getOperand(1);
5450 // FIXME: Ensure alignment here
5454 MVT IntPtr = getPointerTy();
5455 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5457 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5459 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
5460 Flag = Chain.getValue(1);
5462 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5463 SDValue Ops[] = { Chain,
5464 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5465 DAG.getRegister(X86::EAX, IntPtr),
5466 DAG.getRegister(X86StackPtr, SPTy),
5468 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 5);
5469 Flag = Chain.getValue(1);
5471 Chain = DAG.getCALLSEQ_END(Chain,
5472 DAG.getIntPtrConstant(0, true),
5473 DAG.getIntPtrConstant(0, true),
5476 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
5478 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5479 return DAG.getMergeValues(Ops1, 2);
5483 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
5485 SDValue Dst, SDValue Src,
5486 SDValue Size, unsigned Align,
5488 uint64_t DstSVOff) {
5489 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5491 // If not DWORD aligned or size is more than the threshold, call the library.
5492 // The libc version is likely to be faster for these cases. It can use the
5493 // address value and run time information about the CPU.
5494 if ((Align & 3) != 0 ||
5496 ConstantSize->getZExtValue() >
5497 getSubtarget()->getMaxInlineSizeThreshold()) {
5498 SDValue InFlag(0, 0);
5500 // Check to see if there is a specialized entry-point for memory zeroing.
5501 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5503 if (const char *bzeroEntry = V &&
5504 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5505 MVT IntPtr = getPointerTy();
5506 const Type *IntPtrTy = TD->getIntPtrType();
5507 TargetLowering::ArgListTy Args;
5508 TargetLowering::ArgListEntry Entry;
5510 Entry.Ty = IntPtrTy;
5511 Args.push_back(Entry);
5513 Args.push_back(Entry);
5514 std::pair<SDValue,SDValue> CallResult =
5515 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5516 CallingConv::C, false,
5517 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG);
5518 return CallResult.second;
5521 // Otherwise have the target-independent code call memset.
5525 uint64_t SizeVal = ConstantSize->getZExtValue();
5526 SDValue InFlag(0, 0);
5529 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5530 unsigned BytesLeft = 0;
5531 bool TwoRepStos = false;
5534 uint64_t Val = ValC->getZExtValue() & 255;
5536 // If the value is a constant, then we can potentially use larger sets.
5537 switch (Align & 3) {
5538 case 2: // WORD aligned
5541 Val = (Val << 8) | Val;
5543 case 0: // DWORD aligned
5546 Val = (Val << 8) | Val;
5547 Val = (Val << 16) | Val;
5548 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5551 Val = (Val << 32) | Val;
5554 default: // Byte aligned
5557 Count = DAG.getIntPtrConstant(SizeVal);
5561 if (AVT.bitsGT(MVT::i8)) {
5562 unsigned UBytes = AVT.getSizeInBits() / 8;
5563 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5564 BytesLeft = SizeVal % UBytes;
5567 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
5569 InFlag = Chain.getValue(1);
5572 Count = DAG.getIntPtrConstant(SizeVal);
5573 Chain = DAG.getCopyToReg(Chain, X86::AL, Src, InFlag);
5574 InFlag = Chain.getValue(1);
5577 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5579 InFlag = Chain.getValue(1);
5580 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5582 InFlag = Chain.getValue(1);
5584 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5585 SmallVector<SDValue, 8> Ops;
5586 Ops.push_back(Chain);
5587 Ops.push_back(DAG.getValueType(AVT));
5588 Ops.push_back(InFlag);
5589 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5592 InFlag = Chain.getValue(1);
5594 MVT CVT = Count.getValueType();
5595 SDValue Left = DAG.getNode(ISD::AND, CVT, Count,
5596 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5597 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
5599 InFlag = Chain.getValue(1);
5600 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5602 Ops.push_back(Chain);
5603 Ops.push_back(DAG.getValueType(MVT::i8));
5604 Ops.push_back(InFlag);
5605 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
5606 } else if (BytesLeft) {
5607 // Handle the last 1 - 7 bytes.
5608 unsigned Offset = SizeVal - BytesLeft;
5609 MVT AddrVT = Dst.getValueType();
5610 MVT SizeVT = Size.getValueType();
5612 Chain = DAG.getMemset(Chain,
5613 DAG.getNode(ISD::ADD, AddrVT, Dst,
5614 DAG.getConstant(Offset, AddrVT)),
5616 DAG.getConstant(BytesLeft, SizeVT),
5617 Align, DstSV, DstSVOff + Offset);
5620 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5625 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
5626 SDValue Chain, SDValue Dst, SDValue Src,
5627 SDValue Size, unsigned Align,
5629 const Value *DstSV, uint64_t DstSVOff,
5630 const Value *SrcSV, uint64_t SrcSVOff) {
5631 // This requires the copy size to be a constant, preferrably
5632 // within a subtarget-specific limit.
5633 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5636 uint64_t SizeVal = ConstantSize->getZExtValue();
5637 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5640 /// If not DWORD aligned, call the library.
5641 if ((Align & 3) != 0)
5646 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5649 unsigned UBytes = AVT.getSizeInBits() / 8;
5650 unsigned CountVal = SizeVal / UBytes;
5651 SDValue Count = DAG.getIntPtrConstant(CountVal);
5652 unsigned BytesLeft = SizeVal % UBytes;
5654 SDValue InFlag(0, 0);
5655 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
5657 InFlag = Chain.getValue(1);
5658 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
5660 InFlag = Chain.getValue(1);
5661 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
5663 InFlag = Chain.getValue(1);
5665 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5666 SmallVector<SDValue, 8> Ops;
5667 Ops.push_back(Chain);
5668 Ops.push_back(DAG.getValueType(AVT));
5669 Ops.push_back(InFlag);
5670 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
5672 SmallVector<SDValue, 4> Results;
5673 Results.push_back(RepMovs);
5675 // Handle the last 1 - 7 bytes.
5676 unsigned Offset = SizeVal - BytesLeft;
5677 MVT DstVT = Dst.getValueType();
5678 MVT SrcVT = Src.getValueType();
5679 MVT SizeVT = Size.getValueType();
5680 Results.push_back(DAG.getMemcpy(Chain,
5681 DAG.getNode(ISD::ADD, DstVT, Dst,
5682 DAG.getConstant(Offset, DstVT)),
5683 DAG.getNode(ISD::ADD, SrcVT, Src,
5684 DAG.getConstant(Offset, SrcVT)),
5685 DAG.getConstant(BytesLeft, SizeVT),
5686 Align, AlwaysInline,
5687 DstSV, DstSVOff + Offset,
5688 SrcSV, SrcSVOff + Offset));
5691 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());
5694 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5695 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5697 if (!Subtarget->is64Bit()) {
5698 // vastart just stores the address of the VarArgsFrameIndex slot into the
5699 // memory location argument.
5700 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5701 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV, 0);
5705 // gp_offset (0 - 6 * 8)
5706 // fp_offset (48 - 48 + 8 * 16)
5707 // overflow_arg_area (point to parameters coming in memory).
5709 SmallVector<SDValue, 8> MemOps;
5710 SDValue FIN = Op.getOperand(1);
5712 SDValue Store = DAG.getStore(Op.getOperand(0),
5713 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5715 MemOps.push_back(Store);
5718 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5719 Store = DAG.getStore(Op.getOperand(0),
5720 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5722 MemOps.push_back(Store);
5724 // Store ptr to overflow_arg_area
5725 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
5726 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5727 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV, 0);
5728 MemOps.push_back(Store);
5730 // Store ptr to reg_save_area.
5731 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
5732 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5733 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV, 0);
5734 MemOps.push_back(Store);
5735 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5738 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5739 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5740 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5741 SDValue Chain = Op.getOperand(0);
5742 SDValue SrcPtr = Op.getOperand(1);
5743 SDValue SrcSV = Op.getOperand(2);
5745 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5750 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5751 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5752 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5753 SDValue Chain = Op.getOperand(0);
5754 SDValue DstPtr = Op.getOperand(1);
5755 SDValue SrcPtr = Op.getOperand(2);
5756 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5757 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5759 return DAG.getMemcpy(Chain, DstPtr, SrcPtr,
5760 DAG.getIntPtrConstant(24), 8, false,
5761 DstSV, 0, SrcSV, 0);
5765 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5766 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5768 default: return SDValue(); // Don't custom lower most intrinsics.
5769 // Comparison intrinsics.
5770 case Intrinsic::x86_sse_comieq_ss:
5771 case Intrinsic::x86_sse_comilt_ss:
5772 case Intrinsic::x86_sse_comile_ss:
5773 case Intrinsic::x86_sse_comigt_ss:
5774 case Intrinsic::x86_sse_comige_ss:
5775 case Intrinsic::x86_sse_comineq_ss:
5776 case Intrinsic::x86_sse_ucomieq_ss:
5777 case Intrinsic::x86_sse_ucomilt_ss:
5778 case Intrinsic::x86_sse_ucomile_ss:
5779 case Intrinsic::x86_sse_ucomigt_ss:
5780 case Intrinsic::x86_sse_ucomige_ss:
5781 case Intrinsic::x86_sse_ucomineq_ss:
5782 case Intrinsic::x86_sse2_comieq_sd:
5783 case Intrinsic::x86_sse2_comilt_sd:
5784 case Intrinsic::x86_sse2_comile_sd:
5785 case Intrinsic::x86_sse2_comigt_sd:
5786 case Intrinsic::x86_sse2_comige_sd:
5787 case Intrinsic::x86_sse2_comineq_sd:
5788 case Intrinsic::x86_sse2_ucomieq_sd:
5789 case Intrinsic::x86_sse2_ucomilt_sd:
5790 case Intrinsic::x86_sse2_ucomile_sd:
5791 case Intrinsic::x86_sse2_ucomigt_sd:
5792 case Intrinsic::x86_sse2_ucomige_sd:
5793 case Intrinsic::x86_sse2_ucomineq_sd: {
5795 ISD::CondCode CC = ISD::SETCC_INVALID;
5798 case Intrinsic::x86_sse_comieq_ss:
5799 case Intrinsic::x86_sse2_comieq_sd:
5803 case Intrinsic::x86_sse_comilt_ss:
5804 case Intrinsic::x86_sse2_comilt_sd:
5808 case Intrinsic::x86_sse_comile_ss:
5809 case Intrinsic::x86_sse2_comile_sd:
5813 case Intrinsic::x86_sse_comigt_ss:
5814 case Intrinsic::x86_sse2_comigt_sd:
5818 case Intrinsic::x86_sse_comige_ss:
5819 case Intrinsic::x86_sse2_comige_sd:
5823 case Intrinsic::x86_sse_comineq_ss:
5824 case Intrinsic::x86_sse2_comineq_sd:
5828 case Intrinsic::x86_sse_ucomieq_ss:
5829 case Intrinsic::x86_sse2_ucomieq_sd:
5830 Opc = X86ISD::UCOMI;
5833 case Intrinsic::x86_sse_ucomilt_ss:
5834 case Intrinsic::x86_sse2_ucomilt_sd:
5835 Opc = X86ISD::UCOMI;
5838 case Intrinsic::x86_sse_ucomile_ss:
5839 case Intrinsic::x86_sse2_ucomile_sd:
5840 Opc = X86ISD::UCOMI;
5843 case Intrinsic::x86_sse_ucomigt_ss:
5844 case Intrinsic::x86_sse2_ucomigt_sd:
5845 Opc = X86ISD::UCOMI;
5848 case Intrinsic::x86_sse_ucomige_ss:
5849 case Intrinsic::x86_sse2_ucomige_sd:
5850 Opc = X86ISD::UCOMI;
5853 case Intrinsic::x86_sse_ucomineq_ss:
5854 case Intrinsic::x86_sse2_ucomineq_sd:
5855 Opc = X86ISD::UCOMI;
5860 SDValue LHS = Op.getOperand(1);
5861 SDValue RHS = Op.getOperand(2);
5862 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
5863 SDValue Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5864 SDValue SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5865 DAG.getConstant(X86CC, MVT::i8), Cond);
5866 return DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, SetCC);
5869 // Fix vector shift instructions where the last operand is a non-immediate
5871 case Intrinsic::x86_sse2_pslli_w:
5872 case Intrinsic::x86_sse2_pslli_d:
5873 case Intrinsic::x86_sse2_pslli_q:
5874 case Intrinsic::x86_sse2_psrli_w:
5875 case Intrinsic::x86_sse2_psrli_d:
5876 case Intrinsic::x86_sse2_psrli_q:
5877 case Intrinsic::x86_sse2_psrai_w:
5878 case Intrinsic::x86_sse2_psrai_d:
5879 case Intrinsic::x86_mmx_pslli_w:
5880 case Intrinsic::x86_mmx_pslli_d:
5881 case Intrinsic::x86_mmx_pslli_q:
5882 case Intrinsic::x86_mmx_psrli_w:
5883 case Intrinsic::x86_mmx_psrli_d:
5884 case Intrinsic::x86_mmx_psrli_q:
5885 case Intrinsic::x86_mmx_psrai_w:
5886 case Intrinsic::x86_mmx_psrai_d: {
5887 SDValue ShAmt = Op.getOperand(2);
5888 if (isa<ConstantSDNode>(ShAmt))
5891 unsigned NewIntNo = 0;
5892 MVT ShAmtVT = MVT::v4i32;
5894 case Intrinsic::x86_sse2_pslli_w:
5895 NewIntNo = Intrinsic::x86_sse2_psll_w;
5897 case Intrinsic::x86_sse2_pslli_d:
5898 NewIntNo = Intrinsic::x86_sse2_psll_d;
5900 case Intrinsic::x86_sse2_pslli_q:
5901 NewIntNo = Intrinsic::x86_sse2_psll_q;
5903 case Intrinsic::x86_sse2_psrli_w:
5904 NewIntNo = Intrinsic::x86_sse2_psrl_w;
5906 case Intrinsic::x86_sse2_psrli_d:
5907 NewIntNo = Intrinsic::x86_sse2_psrl_d;
5909 case Intrinsic::x86_sse2_psrli_q:
5910 NewIntNo = Intrinsic::x86_sse2_psrl_q;
5912 case Intrinsic::x86_sse2_psrai_w:
5913 NewIntNo = Intrinsic::x86_sse2_psra_w;
5915 case Intrinsic::x86_sse2_psrai_d:
5916 NewIntNo = Intrinsic::x86_sse2_psra_d;
5919 ShAmtVT = MVT::v2i32;
5921 case Intrinsic::x86_mmx_pslli_w:
5922 NewIntNo = Intrinsic::x86_mmx_psll_w;
5924 case Intrinsic::x86_mmx_pslli_d:
5925 NewIntNo = Intrinsic::x86_mmx_psll_d;
5927 case Intrinsic::x86_mmx_pslli_q:
5928 NewIntNo = Intrinsic::x86_mmx_psll_q;
5930 case Intrinsic::x86_mmx_psrli_w:
5931 NewIntNo = Intrinsic::x86_mmx_psrl_w;
5933 case Intrinsic::x86_mmx_psrli_d:
5934 NewIntNo = Intrinsic::x86_mmx_psrl_d;
5936 case Intrinsic::x86_mmx_psrli_q:
5937 NewIntNo = Intrinsic::x86_mmx_psrl_q;
5939 case Intrinsic::x86_mmx_psrai_w:
5940 NewIntNo = Intrinsic::x86_mmx_psra_w;
5942 case Intrinsic::x86_mmx_psrai_d:
5943 NewIntNo = Intrinsic::x86_mmx_psra_d;
5945 default: abort(); // Can't reach here.
5950 MVT VT = Op.getValueType();
5951 ShAmt = DAG.getNode(ISD::BIT_CONVERT, VT,
5952 DAG.getNode(ISD::SCALAR_TO_VECTOR, ShAmtVT, ShAmt));
5953 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
5954 DAG.getConstant(NewIntNo, MVT::i32),
5955 Op.getOperand(1), ShAmt);
5960 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5961 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5964 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5966 DAG.getConstant(TD->getPointerSize(),
5967 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
5968 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(),
5969 DAG.getNode(ISD::ADD, getPointerTy(), FrameAddr, Offset),
5973 // Just load the return address.
5974 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
5975 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5978 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5979 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5980 MFI->setFrameAddressIsTaken(true);
5981 MVT VT = Op.getValueType();
5982 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5983 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
5984 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), FrameReg, VT);
5986 FrameAddr = DAG.getLoad(VT, DAG.getEntryNode(), FrameAddr, NULL, 0);
5990 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
5991 SelectionDAG &DAG) {
5992 return DAG.getIntPtrConstant(2*TD->getPointerSize());
5995 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
5997 MachineFunction &MF = DAG.getMachineFunction();
5998 SDValue Chain = Op.getOperand(0);
5999 SDValue Offset = Op.getOperand(1);
6000 SDValue Handler = Op.getOperand(2);
6002 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6004 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6006 SDValue StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
6007 DAG.getIntPtrConstant(-TD->getPointerSize()));
6008 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
6009 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
6010 Chain = DAG.getCopyToReg(Chain, StoreAddrReg, StoreAddr);
6011 MF.getRegInfo().addLiveOut(StoreAddrReg);
6013 return DAG.getNode(X86ISD::EH_RETURN,
6015 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6018 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6019 SelectionDAG &DAG) {
6020 SDValue Root = Op.getOperand(0);
6021 SDValue Trmp = Op.getOperand(1); // trampoline
6022 SDValue FPtr = Op.getOperand(2); // nested function
6023 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6025 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6027 const X86InstrInfo *TII =
6028 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6030 if (Subtarget->is64Bit()) {
6031 SDValue OutChains[6];
6033 // Large code-model.
6035 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6036 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6038 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6039 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6041 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6043 // Load the pointer to the nested function into R11.
6044 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6045 SDValue Addr = Trmp;
6046 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
6049 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
6050 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpAddr, 2, false, 2);
6052 // Load the 'nest' parameter value into R10.
6053 // R10 is specified in X86CallingConv.td
6054 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6055 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
6056 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
6059 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
6060 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 12, false, 2);
6062 // Jump to the nested function.
6063 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6064 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
6065 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
6068 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6069 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
6070 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
6074 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
6075 return DAG.getMergeValues(Ops, 2);
6077 const Function *Func =
6078 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6079 unsigned CC = Func->getCallingConv();
6084 assert(0 && "Unsupported calling convention");
6085 case CallingConv::C:
6086 case CallingConv::X86_StdCall: {
6087 // Pass 'nest' parameter in ECX.
6088 // Must be kept in sync with X86CallingConv.td
6091 // Check that ECX wasn't needed by an 'inreg' parameter.
6092 const FunctionType *FTy = Func->getFunctionType();
6093 const AttrListPtr &Attrs = Func->getAttributes();
6095 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6096 unsigned InRegCount = 0;
6099 for (FunctionType::param_iterator I = FTy->param_begin(),
6100 E = FTy->param_end(); I != E; ++I, ++Idx)
6101 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6102 // FIXME: should only count parameters that are lowered to integers.
6103 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6105 if (InRegCount > 2) {
6106 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6112 case CallingConv::X86_FastCall:
6113 case CallingConv::Fast:
6114 // Pass 'nest' parameter in EAX.
6115 // Must be kept in sync with X86CallingConv.td
6120 SDValue OutChains[4];
6123 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
6124 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
6126 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6127 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6128 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6131 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
6132 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpAddr, 1, false, 1);
6134 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6135 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
6136 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
6137 TrmpAddr, 5, false, 1);
6139 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
6140 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpAddr, 6, false, 1);
6143 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
6144 return DAG.getMergeValues(Ops, 2);
6148 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6150 The rounding mode is in bits 11:10 of FPSR, and has the following
6157 FLT_ROUNDS, on the other hand, expects the following:
6164 To perform the conversion, we do:
6165 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6168 MachineFunction &MF = DAG.getMachineFunction();
6169 const TargetMachine &TM = MF.getTarget();
6170 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6171 unsigned StackAlignment = TFI.getStackAlignment();
6172 MVT VT = Op.getValueType();
6174 // Save FP Control Word to stack slot
6175 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6176 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6178 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
6179 DAG.getEntryNode(), StackSlot);
6181 // Load FP Control Word from stack slot
6182 SDValue CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
6184 // Transform as necessary
6186 DAG.getNode(ISD::SRL, MVT::i16,
6187 DAG.getNode(ISD::AND, MVT::i16,
6188 CWD, DAG.getConstant(0x800, MVT::i16)),
6189 DAG.getConstant(11, MVT::i8));
6191 DAG.getNode(ISD::SRL, MVT::i16,
6192 DAG.getNode(ISD::AND, MVT::i16,
6193 CWD, DAG.getConstant(0x400, MVT::i16)),
6194 DAG.getConstant(9, MVT::i8));
6197 DAG.getNode(ISD::AND, MVT::i16,
6198 DAG.getNode(ISD::ADD, MVT::i16,
6199 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
6200 DAG.getConstant(1, MVT::i16)),
6201 DAG.getConstant(3, MVT::i16));
6204 return DAG.getNode((VT.getSizeInBits() < 16 ?
6205 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
6208 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6209 MVT VT = Op.getValueType();
6211 unsigned NumBits = VT.getSizeInBits();
6213 Op = Op.getOperand(0);
6214 if (VT == MVT::i8) {
6215 // Zero extend to i32 since there is not an i8 bsr.
6217 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
6220 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6221 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6222 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
6224 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6225 SmallVector<SDValue, 4> Ops;
6227 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6228 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6229 Ops.push_back(Op.getValue(1));
6230 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
6232 // Finally xor with NumBits-1.
6233 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6236 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
6240 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6241 MVT VT = Op.getValueType();
6243 unsigned NumBits = VT.getSizeInBits();
6245 Op = Op.getOperand(0);
6246 if (VT == MVT::i8) {
6248 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
6251 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6252 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6253 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
6255 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6256 SmallVector<SDValue, 4> Ops;
6258 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6259 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6260 Ops.push_back(Op.getValue(1));
6261 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
6264 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
6268 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6269 MVT VT = Op.getValueType();
6270 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6272 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6273 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6274 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6275 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6276 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6278 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6279 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6280 // return AloBlo + AloBhi + AhiBlo;
6282 SDValue A = Op.getOperand(0);
6283 SDValue B = Op.getOperand(1);
6285 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6286 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6287 A, DAG.getConstant(32, MVT::i32));
6288 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6289 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6290 B, DAG.getConstant(32, MVT::i32));
6291 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6292 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6294 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6295 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6297 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6298 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6300 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6301 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6302 AloBhi, DAG.getConstant(32, MVT::i32));
6303 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VT,
6304 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6305 AhiBlo, DAG.getConstant(32, MVT::i32));
6306 SDValue Res = DAG.getNode(ISD::ADD, VT, AloBlo, AloBhi);
6307 Res = DAG.getNode(ISD::ADD, VT, Res, AhiBlo);
6312 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6313 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6314 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6315 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6316 // has only one use.
6317 SDNode *N = Op.getNode();
6318 SDValue LHS = N->getOperand(0);
6319 SDValue RHS = N->getOperand(1);
6320 unsigned BaseOp = 0;
6323 switch (Op.getOpcode()) {
6324 default: assert(0 && "Unknown ovf instruction!");
6326 BaseOp = X86ISD::ADD;
6330 BaseOp = X86ISD::ADD;
6334 BaseOp = X86ISD::SUB;
6338 BaseOp = X86ISD::SUB;
6342 BaseOp = X86ISD::SMUL;
6346 BaseOp = X86ISD::UMUL;
6351 // Also sets EFLAGS.
6352 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6353 SDValue Sum = DAG.getNode(BaseOp, VTs, LHS, RHS);
6356 DAG.getNode(X86ISD::SETCC, N->getValueType(1),
6357 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6359 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6363 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6364 MVT T = Op.getValueType();
6367 switch(T.getSimpleVT()) {
6369 assert(false && "Invalid value type!");
6370 case MVT::i8: Reg = X86::AL; size = 1; break;
6371 case MVT::i16: Reg = X86::AX; size = 2; break;
6372 case MVT::i32: Reg = X86::EAX; size = 4; break;
6374 assert(Subtarget->is64Bit() && "Node not type legal!");
6375 Reg = X86::RAX; size = 8;
6378 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), Reg,
6379 Op.getOperand(2), SDValue());
6380 SDValue Ops[] = { cpIn.getValue(0),
6383 DAG.getTargetConstant(size, MVT::i8),
6385 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6386 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, Tys, Ops, 5);
6388 DAG.getCopyFromReg(Result.getValue(0), Reg, T, Result.getValue(1));
6392 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6393 SelectionDAG &DAG) {
6394 assert(Subtarget->is64Bit() && "Result not type legalized?");
6395 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6396 SDValue TheChain = Op.getOperand(0);
6397 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
6398 SDValue rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
6399 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX, MVT::i64,
6401 SDValue Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
6402 DAG.getConstant(32, MVT::i8));
6404 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp),
6407 return DAG.getMergeValues(Ops, 2);
6410 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6411 SDNode *Node = Op.getNode();
6412 MVT T = Node->getValueType(0);
6413 SDValue negOp = DAG.getNode(ISD::SUB, T,
6414 DAG.getConstant(0, T), Node->getOperand(2));
6415 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD,
6416 cast<AtomicSDNode>(Node)->getMemoryVT(),
6417 Node->getOperand(0),
6418 Node->getOperand(1), negOp,
6419 cast<AtomicSDNode>(Node)->getSrcValue(),
6420 cast<AtomicSDNode>(Node)->getAlignment());
6423 /// LowerOperation - Provide custom lowering hooks for some operations.
6425 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6426 switch (Op.getOpcode()) {
6427 default: assert(0 && "Should not custom lower this!");
6428 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6429 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6430 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6431 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6432 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6433 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6434 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6435 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6436 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6437 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6438 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6439 case ISD::SHL_PARTS:
6440 case ISD::SRA_PARTS:
6441 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6442 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6443 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6444 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6445 case ISD::FABS: return LowerFABS(Op, DAG);
6446 case ISD::FNEG: return LowerFNEG(Op, DAG);
6447 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6448 case ISD::SETCC: return LowerSETCC(Op, DAG);
6449 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6450 case ISD::SELECT: return LowerSELECT(Op, DAG);
6451 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6452 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6453 case ISD::CALL: return LowerCALL(Op, DAG);
6454 case ISD::RET: return LowerRET(Op, DAG);
6455 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6456 case ISD::VASTART: return LowerVASTART(Op, DAG);
6457 case ISD::VAARG: return LowerVAARG(Op, DAG);
6458 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6459 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6460 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6461 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6462 case ISD::FRAME_TO_ARGS_OFFSET:
6463 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6464 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6465 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6466 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6467 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6468 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6469 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6470 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6476 case ISD::UMULO: return LowerXALUO(Op, DAG);
6477 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6481 void X86TargetLowering::
6482 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6483 SelectionDAG &DAG, unsigned NewOp) {
6484 MVT T = Node->getValueType(0);
6485 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6487 SDValue Chain = Node->getOperand(0);
6488 SDValue In1 = Node->getOperand(1);
6489 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
6490 Node->getOperand(2), DAG.getIntPtrConstant(0));
6491 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
6492 Node->getOperand(2), DAG.getIntPtrConstant(1));
6493 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6494 // have a MemOperand. Pass the info through as a normal operand.
6495 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6496 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6497 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6498 SDValue Result = DAG.getNode(NewOp, Tys, Ops, 5);
6499 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6500 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2));
6501 Results.push_back(Result.getValue(2));
6504 /// ReplaceNodeResults - Replace a node with an illegal result type
6505 /// with a new node built out of custom code.
6506 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6507 SmallVectorImpl<SDValue>&Results,
6508 SelectionDAG &DAG) {
6509 switch (N->getOpcode()) {
6511 assert(false && "Do not know how to custom type legalize this operation!");
6513 case ISD::FP_TO_SINT: {
6514 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
6515 SDValue FIST = Vals.first, StackSlot = Vals.second;
6516 if (FIST.getNode() != 0) {
6517 MVT VT = N->getValueType(0);
6518 // Return a load from the stack slot.
6519 Results.push_back(DAG.getLoad(VT, FIST, StackSlot, NULL, 0));
6523 case ISD::READCYCLECOUNTER: {
6524 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6525 SDValue TheChain = N->getOperand(0);
6526 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
6527 SDValue eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
6528 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX, MVT::i32,
6530 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6531 SDValue Ops[] = { eax, edx };
6532 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2));
6533 Results.push_back(edx.getValue(1));
6536 case ISD::ATOMIC_CMP_SWAP: {
6537 MVT T = N->getValueType(0);
6538 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6539 SDValue cpInL, cpInH;
6540 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(2),
6541 DAG.getConstant(0, MVT::i32));
6542 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(2),
6543 DAG.getConstant(1, MVT::i32));
6544 cpInL = DAG.getCopyToReg(N->getOperand(0), X86::EAX, cpInL, SDValue());
6545 cpInH = DAG.getCopyToReg(cpInL.getValue(0), X86::EDX, cpInH,
6547 SDValue swapInL, swapInH;
6548 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(3),
6549 DAG.getConstant(0, MVT::i32));
6550 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(3),
6551 DAG.getConstant(1, MVT::i32));
6552 swapInL = DAG.getCopyToReg(cpInH.getValue(0), X86::EBX, swapInL,
6554 swapInH = DAG.getCopyToReg(swapInL.getValue(0), X86::ECX, swapInH,
6555 swapInL.getValue(1));
6556 SDValue Ops[] = { swapInH.getValue(0),
6558 swapInH.getValue(1) };
6559 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6560 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, Tys, Ops, 3);
6561 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), X86::EAX, MVT::i32,
6562 Result.getValue(1));
6563 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), X86::EDX, MVT::i32,
6564 cpOutL.getValue(2));
6565 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6566 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OpsF, 2));
6567 Results.push_back(cpOutH.getValue(1));
6570 case ISD::ATOMIC_LOAD_ADD:
6571 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6573 case ISD::ATOMIC_LOAD_AND:
6574 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6576 case ISD::ATOMIC_LOAD_NAND:
6577 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6579 case ISD::ATOMIC_LOAD_OR:
6580 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6582 case ISD::ATOMIC_LOAD_SUB:
6583 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6585 case ISD::ATOMIC_LOAD_XOR:
6586 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6588 case ISD::ATOMIC_SWAP:
6589 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6594 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6596 default: return NULL;
6597 case X86ISD::BSF: return "X86ISD::BSF";
6598 case X86ISD::BSR: return "X86ISD::BSR";
6599 case X86ISD::SHLD: return "X86ISD::SHLD";
6600 case X86ISD::SHRD: return "X86ISD::SHRD";
6601 case X86ISD::FAND: return "X86ISD::FAND";
6602 case X86ISD::FOR: return "X86ISD::FOR";
6603 case X86ISD::FXOR: return "X86ISD::FXOR";
6604 case X86ISD::FSRL: return "X86ISD::FSRL";
6605 case X86ISD::FILD: return "X86ISD::FILD";
6606 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6607 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6608 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6609 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6610 case X86ISD::FLD: return "X86ISD::FLD";
6611 case X86ISD::FST: return "X86ISD::FST";
6612 case X86ISD::CALL: return "X86ISD::CALL";
6613 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6614 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6615 case X86ISD::BT: return "X86ISD::BT";
6616 case X86ISD::CMP: return "X86ISD::CMP";
6617 case X86ISD::COMI: return "X86ISD::COMI";
6618 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6619 case X86ISD::SETCC: return "X86ISD::SETCC";
6620 case X86ISD::CMOV: return "X86ISD::CMOV";
6621 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6622 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6623 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6624 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6625 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6626 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6627 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6628 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6629 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6630 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6631 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6632 case X86ISD::FMAX: return "X86ISD::FMAX";
6633 case X86ISD::FMIN: return "X86ISD::FMIN";
6634 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6635 case X86ISD::FRCP: return "X86ISD::FRCP";
6636 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6637 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
6638 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6639 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6640 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6641 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6642 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6643 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6644 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6645 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6646 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6647 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6648 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6649 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6650 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6651 case X86ISD::VSHL: return "X86ISD::VSHL";
6652 case X86ISD::VSRL: return "X86ISD::VSRL";
6653 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6654 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6655 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6656 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6657 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6658 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6659 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6660 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6661 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6662 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6663 case X86ISD::ADD: return "X86ISD::ADD";
6664 case X86ISD::SUB: return "X86ISD::SUB";
6665 case X86ISD::SMUL: return "X86ISD::SMUL";
6666 case X86ISD::UMUL: return "X86ISD::UMUL";
6670 // isLegalAddressingMode - Return true if the addressing mode represented
6671 // by AM is legal for this target, for a load/store of the specified type.
6672 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6673 const Type *Ty) const {
6674 // X86 supports extremely general addressing modes.
6676 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6677 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6681 // We can only fold this if we don't need an extra load.
6682 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6684 // If BaseGV requires a register, we cannot also have a BaseReg.
6685 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6689 // X86-64 only supports addr of globals in small code model.
6690 if (Subtarget->is64Bit()) {
6691 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6693 // If lower 4G is not available, then we must use rip-relative addressing.
6694 if (AM.BaseOffs || AM.Scale > 1)
6705 // These scales always work.
6710 // These scales are formed with basereg+scalereg. Only accept if there is
6715 default: // Other stuff never works.
6723 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6724 if (!Ty1->isInteger() || !Ty2->isInteger())
6726 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6727 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6728 if (NumBits1 <= NumBits2)
6730 return Subtarget->is64Bit() || NumBits1 < 64;
6733 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6734 if (!VT1.isInteger() || !VT2.isInteger())
6736 unsigned NumBits1 = VT1.getSizeInBits();
6737 unsigned NumBits2 = VT2.getSizeInBits();
6738 if (NumBits1 <= NumBits2)
6740 return Subtarget->is64Bit() || NumBits1 < 64;
6743 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6744 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6745 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6746 /// are assumed to be legal.
6748 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
6749 // Only do shuffles on 128-bit vector types for now.
6750 if (VT.getSizeInBits() == 64) return false;
6751 return (Mask.getNode()->getNumOperands() <= 4 ||
6752 isIdentityMask(Mask.getNode()) ||
6753 isIdentityMask(Mask.getNode(), true) ||
6754 isSplatMask(Mask.getNode()) ||
6755 isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
6756 X86::isUNPCKLMask(Mask.getNode()) ||
6757 X86::isUNPCKHMask(Mask.getNode()) ||
6758 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
6759 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
6763 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
6764 MVT EVT, SelectionDAG &DAG) const {
6765 unsigned NumElts = BVOps.size();
6766 // Only do shuffles on 128-bit vector types for now.
6767 if (EVT.getSizeInBits() * NumElts == 64) return false;
6768 if (NumElts == 2) return true;
6770 return (isMOVLMask(&BVOps[0], 4) ||
6771 isCommutedMOVL(&BVOps[0], 4, true) ||
6772 isSHUFPMask(&BVOps[0], 4) ||
6773 isCommutedSHUFP(&BVOps[0], 4));
6778 //===----------------------------------------------------------------------===//
6779 // X86 Scheduler Hooks
6780 //===----------------------------------------------------------------------===//
6782 // private utility function
6784 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6785 MachineBasicBlock *MBB,
6793 TargetRegisterClass *RC,
6795 // For the atomic bitwise operator, we generate
6798 // ld t1 = [bitinstr.addr]
6799 // op t2 = t1, [bitinstr.val]
6801 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6803 // fallthrough -->nextMBB
6804 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6805 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6806 MachineFunction::iterator MBBIter = MBB;
6809 /// First build the CFG
6810 MachineFunction *F = MBB->getParent();
6811 MachineBasicBlock *thisMBB = MBB;
6812 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6813 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6814 F->insert(MBBIter, newMBB);
6815 F->insert(MBBIter, nextMBB);
6817 // Move all successors to thisMBB to nextMBB
6818 nextMBB->transferSuccessors(thisMBB);
6820 // Update thisMBB to fall through to newMBB
6821 thisMBB->addSuccessor(newMBB);
6823 // newMBB jumps to itself and fall through to nextMBB
6824 newMBB->addSuccessor(nextMBB);
6825 newMBB->addSuccessor(newMBB);
6827 // Insert instructions into newMBB based on incoming instruction
6828 assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
6829 MachineOperand& destOper = bInstr->getOperand(0);
6830 MachineOperand* argOpers[6];
6831 int numArgs = bInstr->getNumOperands() - 1;
6832 for (int i=0; i < numArgs; ++i)
6833 argOpers[i] = &bInstr->getOperand(i+1);
6835 // x86 address has 4 operands: base, index, scale, and displacement
6836 int lastAddrIndx = 3; // [0,3]
6839 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6840 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(LoadOpc), t1);
6841 for (int i=0; i <= lastAddrIndx; ++i)
6842 (*MIB).addOperand(*argOpers[i]);
6844 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
6846 MIB = BuildMI(newMBB, TII->get(notOpc), tt).addReg(t1);
6851 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6852 assert((argOpers[valArgIndx]->isReg() ||
6853 argOpers[valArgIndx]->isImm()) &&
6855 if (argOpers[valArgIndx]->isReg())
6856 MIB = BuildMI(newMBB, TII->get(regOpc), t2);
6858 MIB = BuildMI(newMBB, TII->get(immOpc), t2);
6860 (*MIB).addOperand(*argOpers[valArgIndx]);
6862 MIB = BuildMI(newMBB, TII->get(copyOpc), EAXreg);
6865 MIB = BuildMI(newMBB, TII->get(CXchgOpc));
6866 for (int i=0; i <= lastAddrIndx; ++i)
6867 (*MIB).addOperand(*argOpers[i]);
6869 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
6870 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
6872 MIB = BuildMI(newMBB, TII->get(copyOpc), destOper.getReg());
6876 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
6878 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
6882 // private utility function: 64 bit atomics on 32 bit host.
6884 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
6885 MachineBasicBlock *MBB,
6891 // For the atomic bitwise operator, we generate
6892 // thisMBB (instructions are in pairs, except cmpxchg8b)
6893 // ld t1,t2 = [bitinstr.addr]
6895 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
6896 // op t5, t6 <- out1, out2, [bitinstr.val]
6897 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
6898 // mov ECX, EBX <- t5, t6
6899 // mov EAX, EDX <- t1, t2
6900 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
6901 // mov t3, t4 <- EAX, EDX
6903 // result in out1, out2
6904 // fallthrough -->nextMBB
6906 const TargetRegisterClass *RC = X86::GR32RegisterClass;
6907 const unsigned LoadOpc = X86::MOV32rm;
6908 const unsigned copyOpc = X86::MOV32rr;
6909 const unsigned NotOpc = X86::NOT32r;
6910 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6911 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6912 MachineFunction::iterator MBBIter = MBB;
6915 /// First build the CFG
6916 MachineFunction *F = MBB->getParent();
6917 MachineBasicBlock *thisMBB = MBB;
6918 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6919 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6920 F->insert(MBBIter, newMBB);
6921 F->insert(MBBIter, nextMBB);
6923 // Move all successors to thisMBB to nextMBB
6924 nextMBB->transferSuccessors(thisMBB);
6926 // Update thisMBB to fall through to newMBB
6927 thisMBB->addSuccessor(newMBB);
6929 // newMBB jumps to itself and fall through to nextMBB
6930 newMBB->addSuccessor(nextMBB);
6931 newMBB->addSuccessor(newMBB);
6933 // Insert instructions into newMBB based on incoming instruction
6934 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
6935 assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
6936 MachineOperand& dest1Oper = bInstr->getOperand(0);
6937 MachineOperand& dest2Oper = bInstr->getOperand(1);
6938 MachineOperand* argOpers[6];
6939 for (int i=0; i < 6; ++i)
6940 argOpers[i] = &bInstr->getOperand(i+2);
6942 // x86 address has 4 operands: base, index, scale, and displacement
6943 int lastAddrIndx = 3; // [0,3]
6945 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
6946 MachineInstrBuilder MIB = BuildMI(thisMBB, TII->get(LoadOpc), t1);
6947 for (int i=0; i <= lastAddrIndx; ++i)
6948 (*MIB).addOperand(*argOpers[i]);
6949 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
6950 MIB = BuildMI(thisMBB, TII->get(LoadOpc), t2);
6951 // add 4 to displacement.
6952 for (int i=0; i <= lastAddrIndx-1; ++i)
6953 (*MIB).addOperand(*argOpers[i]);
6954 MachineOperand newOp3 = *(argOpers[3]);
6956 newOp3.setImm(newOp3.getImm()+4);
6958 newOp3.setOffset(newOp3.getOffset()+4);
6959 (*MIB).addOperand(newOp3);
6961 // t3/4 are defined later, at the bottom of the loop
6962 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
6963 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
6964 BuildMI(newMBB, TII->get(X86::PHI), dest1Oper.getReg())
6965 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
6966 BuildMI(newMBB, TII->get(X86::PHI), dest2Oper.getReg())
6967 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
6969 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
6970 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
6972 MIB = BuildMI(newMBB, TII->get(NotOpc), tt1).addReg(t1);
6973 MIB = BuildMI(newMBB, TII->get(NotOpc), tt2).addReg(t2);
6979 assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
6981 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
6982 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
6983 if (argOpers[4]->isReg())
6984 MIB = BuildMI(newMBB, TII->get(regOpcL), t5);
6986 MIB = BuildMI(newMBB, TII->get(immOpcL), t5);
6987 if (regOpcL != X86::MOV32rr)
6989 (*MIB).addOperand(*argOpers[4]);
6990 assert(argOpers[5]->isReg() == argOpers[4]->isReg());
6991 assert(argOpers[5]->isImm() == argOpers[4]->isImm());
6992 if (argOpers[5]->isReg())
6993 MIB = BuildMI(newMBB, TII->get(regOpcH), t6);
6995 MIB = BuildMI(newMBB, TII->get(immOpcH), t6);
6996 if (regOpcH != X86::MOV32rr)
6998 (*MIB).addOperand(*argOpers[5]);
7000 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EAX);
7002 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EDX);
7005 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::EBX);
7007 MIB = BuildMI(newMBB, TII->get(copyOpc), X86::ECX);
7010 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG8B));
7011 for (int i=0; i <= lastAddrIndx; ++i)
7012 (*MIB).addOperand(*argOpers[i]);
7014 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7015 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7017 MIB = BuildMI(newMBB, TII->get(copyOpc), t3);
7018 MIB.addReg(X86::EAX);
7019 MIB = BuildMI(newMBB, TII->get(copyOpc), t4);
7020 MIB.addReg(X86::EDX);
7023 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
7025 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7029 // private utility function
7031 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7032 MachineBasicBlock *MBB,
7034 // For the atomic min/max operator, we generate
7037 // ld t1 = [min/max.addr]
7038 // mov t2 = [min/max.val]
7040 // cmov[cond] t2 = t1
7042 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7044 // fallthrough -->nextMBB
7046 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7047 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7048 MachineFunction::iterator MBBIter = MBB;
7051 /// First build the CFG
7052 MachineFunction *F = MBB->getParent();
7053 MachineBasicBlock *thisMBB = MBB;
7054 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7055 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7056 F->insert(MBBIter, newMBB);
7057 F->insert(MBBIter, nextMBB);
7059 // Move all successors to thisMBB to nextMBB
7060 nextMBB->transferSuccessors(thisMBB);
7062 // Update thisMBB to fall through to newMBB
7063 thisMBB->addSuccessor(newMBB);
7065 // newMBB jumps to newMBB and fall through to nextMBB
7066 newMBB->addSuccessor(nextMBB);
7067 newMBB->addSuccessor(newMBB);
7069 // Insert instructions into newMBB based on incoming instruction
7070 assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
7071 MachineOperand& destOper = mInstr->getOperand(0);
7072 MachineOperand* argOpers[6];
7073 int numArgs = mInstr->getNumOperands() - 1;
7074 for (int i=0; i < numArgs; ++i)
7075 argOpers[i] = &mInstr->getOperand(i+1);
7077 // x86 address has 4 operands: base, index, scale, and displacement
7078 int lastAddrIndx = 3; // [0,3]
7081 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7082 MachineInstrBuilder MIB = BuildMI(newMBB, TII->get(X86::MOV32rm), t1);
7083 for (int i=0; i <= lastAddrIndx; ++i)
7084 (*MIB).addOperand(*argOpers[i]);
7086 // We only support register and immediate values
7087 assert((argOpers[valArgIndx]->isReg() ||
7088 argOpers[valArgIndx]->isImm()) &&
7091 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7092 if (argOpers[valArgIndx]->isReg())
7093 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
7095 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), t2);
7096 (*MIB).addOperand(*argOpers[valArgIndx]);
7098 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), X86::EAX);
7101 MIB = BuildMI(newMBB, TII->get(X86::CMP32rr));
7106 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7107 MIB = BuildMI(newMBB, TII->get(cmovOpc),t3);
7111 // Cmp and exchange if none has modified the memory location
7112 MIB = BuildMI(newMBB, TII->get(X86::LCMPXCHG32));
7113 for (int i=0; i <= lastAddrIndx; ++i)
7114 (*MIB).addOperand(*argOpers[i]);
7116 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7117 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7119 MIB = BuildMI(newMBB, TII->get(X86::MOV32rr), destOper.getReg());
7120 MIB.addReg(X86::EAX);
7123 BuildMI(newMBB, TII->get(X86::JNE)).addMBB(newMBB);
7125 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7131 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7132 MachineBasicBlock *BB) {
7133 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7134 switch (MI->getOpcode()) {
7135 default: assert(false && "Unexpected instr type to insert");
7136 case X86::CMOV_V1I64:
7137 case X86::CMOV_FR32:
7138 case X86::CMOV_FR64:
7139 case X86::CMOV_V4F32:
7140 case X86::CMOV_V2F64:
7141 case X86::CMOV_V2I64: {
7142 // To "insert" a SELECT_CC instruction, we actually have to insert the
7143 // diamond control-flow pattern. The incoming instruction knows the
7144 // destination vreg to set, the condition code register to branch on, the
7145 // true/false values to select between, and a branch opcode to use.
7146 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7147 MachineFunction::iterator It = BB;
7153 // cmpTY ccX, r1, r2
7155 // fallthrough --> copy0MBB
7156 MachineBasicBlock *thisMBB = BB;
7157 MachineFunction *F = BB->getParent();
7158 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7159 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7161 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7162 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
7163 F->insert(It, copy0MBB);
7164 F->insert(It, sinkMBB);
7165 // Update machine-CFG edges by transferring all successors of the current
7166 // block to the new block which will contain the Phi node for the select.
7167 sinkMBB->transferSuccessors(BB);
7169 // Add the true and fallthrough blocks as its successors.
7170 BB->addSuccessor(copy0MBB);
7171 BB->addSuccessor(sinkMBB);
7174 // %FalseValue = ...
7175 // # fallthrough to sinkMBB
7178 // Update machine-CFG edges
7179 BB->addSuccessor(sinkMBB);
7182 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7185 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
7186 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7187 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7189 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7193 case X86::FP32_TO_INT16_IN_MEM:
7194 case X86::FP32_TO_INT32_IN_MEM:
7195 case X86::FP32_TO_INT64_IN_MEM:
7196 case X86::FP64_TO_INT16_IN_MEM:
7197 case X86::FP64_TO_INT32_IN_MEM:
7198 case X86::FP64_TO_INT64_IN_MEM:
7199 case X86::FP80_TO_INT16_IN_MEM:
7200 case X86::FP80_TO_INT32_IN_MEM:
7201 case X86::FP80_TO_INT64_IN_MEM: {
7202 // Change the floating point control register to use "round towards zero"
7203 // mode when truncating to an integer value.
7204 MachineFunction *F = BB->getParent();
7205 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7206 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7208 // Load the old value of the high byte of the control word...
7210 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7211 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
7213 // Set the high part to be round to zero...
7214 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
7217 // Reload the modified control word now...
7218 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
7220 // Restore the memory image of control word to original value
7221 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
7224 // Get the X86 opcode to use.
7226 switch (MI->getOpcode()) {
7227 default: assert(0 && "illegal opcode!");
7228 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7229 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7230 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7231 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7232 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7233 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7234 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7235 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7236 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7240 MachineOperand &Op = MI->getOperand(0);
7242 AM.BaseType = X86AddressMode::RegBase;
7243 AM.Base.Reg = Op.getReg();
7245 AM.BaseType = X86AddressMode::FrameIndexBase;
7246 AM.Base.FrameIndex = Op.getIndex();
7248 Op = MI->getOperand(1);
7250 AM.Scale = Op.getImm();
7251 Op = MI->getOperand(2);
7253 AM.IndexReg = Op.getImm();
7254 Op = MI->getOperand(3);
7255 if (Op.isGlobal()) {
7256 AM.GV = Op.getGlobal();
7258 AM.Disp = Op.getImm();
7260 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
7261 .addReg(MI->getOperand(4).getReg());
7263 // Reload the original control word now.
7264 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
7266 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7269 case X86::ATOMAND32:
7270 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7271 X86::AND32ri, X86::MOV32rm,
7272 X86::LCMPXCHG32, X86::MOV32rr,
7273 X86::NOT32r, X86::EAX,
7274 X86::GR32RegisterClass);
7276 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7277 X86::OR32ri, X86::MOV32rm,
7278 X86::LCMPXCHG32, X86::MOV32rr,
7279 X86::NOT32r, X86::EAX,
7280 X86::GR32RegisterClass);
7281 case X86::ATOMXOR32:
7282 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7283 X86::XOR32ri, X86::MOV32rm,
7284 X86::LCMPXCHG32, X86::MOV32rr,
7285 X86::NOT32r, X86::EAX,
7286 X86::GR32RegisterClass);
7287 case X86::ATOMNAND32:
7288 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7289 X86::AND32ri, X86::MOV32rm,
7290 X86::LCMPXCHG32, X86::MOV32rr,
7291 X86::NOT32r, X86::EAX,
7292 X86::GR32RegisterClass, true);
7293 case X86::ATOMMIN32:
7294 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7295 case X86::ATOMMAX32:
7296 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7297 case X86::ATOMUMIN32:
7298 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7299 case X86::ATOMUMAX32:
7300 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7302 case X86::ATOMAND16:
7303 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7304 X86::AND16ri, X86::MOV16rm,
7305 X86::LCMPXCHG16, X86::MOV16rr,
7306 X86::NOT16r, X86::AX,
7307 X86::GR16RegisterClass);
7309 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7310 X86::OR16ri, X86::MOV16rm,
7311 X86::LCMPXCHG16, X86::MOV16rr,
7312 X86::NOT16r, X86::AX,
7313 X86::GR16RegisterClass);
7314 case X86::ATOMXOR16:
7315 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7316 X86::XOR16ri, X86::MOV16rm,
7317 X86::LCMPXCHG16, X86::MOV16rr,
7318 X86::NOT16r, X86::AX,
7319 X86::GR16RegisterClass);
7320 case X86::ATOMNAND16:
7321 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7322 X86::AND16ri, X86::MOV16rm,
7323 X86::LCMPXCHG16, X86::MOV16rr,
7324 X86::NOT16r, X86::AX,
7325 X86::GR16RegisterClass, true);
7326 case X86::ATOMMIN16:
7327 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7328 case X86::ATOMMAX16:
7329 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7330 case X86::ATOMUMIN16:
7331 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7332 case X86::ATOMUMAX16:
7333 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7336 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7337 X86::AND8ri, X86::MOV8rm,
7338 X86::LCMPXCHG8, X86::MOV8rr,
7339 X86::NOT8r, X86::AL,
7340 X86::GR8RegisterClass);
7342 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7343 X86::OR8ri, X86::MOV8rm,
7344 X86::LCMPXCHG8, X86::MOV8rr,
7345 X86::NOT8r, X86::AL,
7346 X86::GR8RegisterClass);
7348 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7349 X86::XOR8ri, X86::MOV8rm,
7350 X86::LCMPXCHG8, X86::MOV8rr,
7351 X86::NOT8r, X86::AL,
7352 X86::GR8RegisterClass);
7353 case X86::ATOMNAND8:
7354 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7355 X86::AND8ri, X86::MOV8rm,
7356 X86::LCMPXCHG8, X86::MOV8rr,
7357 X86::NOT8r, X86::AL,
7358 X86::GR8RegisterClass, true);
7359 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7360 // This group is for 64-bit host.
7361 case X86::ATOMAND64:
7362 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7363 X86::AND64ri32, X86::MOV64rm,
7364 X86::LCMPXCHG64, X86::MOV64rr,
7365 X86::NOT64r, X86::RAX,
7366 X86::GR64RegisterClass);
7368 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7369 X86::OR64ri32, X86::MOV64rm,
7370 X86::LCMPXCHG64, X86::MOV64rr,
7371 X86::NOT64r, X86::RAX,
7372 X86::GR64RegisterClass);
7373 case X86::ATOMXOR64:
7374 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7375 X86::XOR64ri32, X86::MOV64rm,
7376 X86::LCMPXCHG64, X86::MOV64rr,
7377 X86::NOT64r, X86::RAX,
7378 X86::GR64RegisterClass);
7379 case X86::ATOMNAND64:
7380 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7381 X86::AND64ri32, X86::MOV64rm,
7382 X86::LCMPXCHG64, X86::MOV64rr,
7383 X86::NOT64r, X86::RAX,
7384 X86::GR64RegisterClass, true);
7385 case X86::ATOMMIN64:
7386 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7387 case X86::ATOMMAX64:
7388 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7389 case X86::ATOMUMIN64:
7390 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7391 case X86::ATOMUMAX64:
7392 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7394 // This group does 64-bit operations on a 32-bit host.
7395 case X86::ATOMAND6432:
7396 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7397 X86::AND32rr, X86::AND32rr,
7398 X86::AND32ri, X86::AND32ri,
7400 case X86::ATOMOR6432:
7401 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7402 X86::OR32rr, X86::OR32rr,
7403 X86::OR32ri, X86::OR32ri,
7405 case X86::ATOMXOR6432:
7406 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7407 X86::XOR32rr, X86::XOR32rr,
7408 X86::XOR32ri, X86::XOR32ri,
7410 case X86::ATOMNAND6432:
7411 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7412 X86::AND32rr, X86::AND32rr,
7413 X86::AND32ri, X86::AND32ri,
7415 case X86::ATOMADD6432:
7416 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7417 X86::ADD32rr, X86::ADC32rr,
7418 X86::ADD32ri, X86::ADC32ri,
7420 case X86::ATOMSUB6432:
7421 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7422 X86::SUB32rr, X86::SBB32rr,
7423 X86::SUB32ri, X86::SBB32ri,
7425 case X86::ATOMSWAP6432:
7426 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7427 X86::MOV32rr, X86::MOV32rr,
7428 X86::MOV32ri, X86::MOV32ri,
7433 //===----------------------------------------------------------------------===//
7434 // X86 Optimization Hooks
7435 //===----------------------------------------------------------------------===//
7437 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7441 const SelectionDAG &DAG,
7442 unsigned Depth) const {
7443 unsigned Opc = Op.getOpcode();
7444 assert((Opc >= ISD::BUILTIN_OP_END ||
7445 Opc == ISD::INTRINSIC_WO_CHAIN ||
7446 Opc == ISD::INTRINSIC_W_CHAIN ||
7447 Opc == ISD::INTRINSIC_VOID) &&
7448 "Should use MaskedValueIsZero if you don't know whether Op"
7449 " is a target node!");
7451 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7455 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7456 Mask.getBitWidth() - 1);
7461 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7462 /// node is a GlobalAddress + offset.
7463 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7464 GlobalValue* &GA, int64_t &Offset) const{
7465 if (N->getOpcode() == X86ISD::Wrapper) {
7466 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7467 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7468 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7472 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7475 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7476 const TargetLowering &TLI) {
7479 if (TLI.isGAPlusOffset(Base, GV, Offset))
7480 return (GV->getAlignment() >= N && (Offset % N) == 0);
7481 // DAG combine handles the stack object case.
7485 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
7486 unsigned NumElems, MVT EVT,
7488 SelectionDAG &DAG, MachineFrameInfo *MFI,
7489 const TargetLowering &TLI) {
7491 for (unsigned i = 0; i < NumElems; ++i) {
7492 SDValue Idx = PermMask.getOperand(i);
7493 if (Idx.getOpcode() == ISD::UNDEF) {
7499 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7500 if (!Elt.getNode() ||
7501 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7504 Base = Elt.getNode();
7505 if (Base->getOpcode() == ISD::UNDEF)
7509 if (Elt.getOpcode() == ISD::UNDEF)
7512 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7513 EVT.getSizeInBits()/8, i, MFI))
7519 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7520 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7521 /// if the load addresses are consecutive, non-overlapping, and in the right
7523 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7524 const TargetLowering &TLI) {
7525 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7526 MVT VT = N->getValueType(0);
7527 MVT EVT = VT.getVectorElementType();
7528 SDValue PermMask = N->getOperand(2);
7529 unsigned NumElems = PermMask.getNumOperands();
7530 SDNode *Base = NULL;
7531 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
7535 LoadSDNode *LD = cast<LoadSDNode>(Base);
7536 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7537 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7538 LD->getSrcValueOffset(), LD->isVolatile());
7539 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
7540 LD->getSrcValueOffset(), LD->isVolatile(),
7541 LD->getAlignment());
7544 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7545 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7546 const X86Subtarget *Subtarget,
7547 const TargetLowering &TLI) {
7548 unsigned NumOps = N->getNumOperands();
7550 // Ignore single operand BUILD_VECTOR.
7554 MVT VT = N->getValueType(0);
7555 MVT EVT = VT.getVectorElementType();
7556 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7557 // We are looking for load i64 and zero extend. We want to transform
7558 // it before legalizer has a chance to expand it. Also look for i64
7559 // BUILD_PAIR bit casted to f64.
7561 // This must be an insertion into a zero vector.
7562 SDValue HighElt = N->getOperand(1);
7563 if (!isZeroNode(HighElt))
7566 // Value must be a load.
7567 SDNode *Base = N->getOperand(0).getNode();
7568 if (!isa<LoadSDNode>(Base)) {
7569 if (Base->getOpcode() != ISD::BIT_CONVERT)
7571 Base = Base->getOperand(0).getNode();
7572 if (!isa<LoadSDNode>(Base))
7576 // Transform it into VZEXT_LOAD addr.
7577 LoadSDNode *LD = cast<LoadSDNode>(Base);
7579 // Load must not be an extload.
7580 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7583 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7584 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7585 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, Tys, Ops, 2);
7586 DAG.ReplaceAllUsesOfValueWith(SDValue(Base, 1), ResNode.getValue(1));
7590 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7591 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7592 const X86Subtarget *Subtarget) {
7593 SDValue Cond = N->getOperand(0);
7595 // If we have SSE[12] support, try to form min/max nodes.
7596 if (Subtarget->hasSSE2() &&
7597 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
7598 if (Cond.getOpcode() == ISD::SETCC) {
7599 // Get the LHS/RHS of the select.
7600 SDValue LHS = N->getOperand(1);
7601 SDValue RHS = N->getOperand(2);
7602 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7604 unsigned Opcode = 0;
7605 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7608 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7611 if (!UnsafeFPMath) break;
7613 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7615 Opcode = X86ISD::FMIN;
7618 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7621 if (!UnsafeFPMath) break;
7623 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7625 Opcode = X86ISD::FMAX;
7628 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7631 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7634 if (!UnsafeFPMath) break;
7636 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7638 Opcode = X86ISD::FMIN;
7641 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7644 if (!UnsafeFPMath) break;
7646 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7648 Opcode = X86ISD::FMAX;
7654 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
7662 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
7663 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
7664 const X86Subtarget *Subtarget) {
7665 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
7666 // the FP state in cases where an emms may be missing.
7667 // A preferable solution to the general problem is to figure out the right
7668 // places to insert EMMS. This qualifies as a quick hack.
7669 StoreSDNode *St = cast<StoreSDNode>(N);
7670 if (St->getValue().getValueType().isVector() &&
7671 St->getValue().getValueType().getSizeInBits() == 64 &&
7672 isa<LoadSDNode>(St->getValue()) &&
7673 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
7674 St->getChain().hasOneUse() && !St->isVolatile()) {
7675 SDNode* LdVal = St->getValue().getNode();
7677 int TokenFactorIndex = -1;
7678 SmallVector<SDValue, 8> Ops;
7679 SDNode* ChainVal = St->getChain().getNode();
7680 // Must be a store of a load. We currently handle two cases: the load
7681 // is a direct child, and it's under an intervening TokenFactor. It is
7682 // possible to dig deeper under nested TokenFactors.
7683 if (ChainVal == LdVal)
7684 Ld = cast<LoadSDNode>(St->getChain());
7685 else if (St->getValue().hasOneUse() &&
7686 ChainVal->getOpcode() == ISD::TokenFactor) {
7687 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
7688 if (ChainVal->getOperand(i).getNode() == LdVal) {
7689 TokenFactorIndex = i;
7690 Ld = cast<LoadSDNode>(St->getValue());
7692 Ops.push_back(ChainVal->getOperand(i));
7696 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
7697 if (Subtarget->is64Bit()) {
7698 SDValue NewLd = DAG.getLoad(MVT::i64, Ld->getChain(),
7699 Ld->getBasePtr(), Ld->getSrcValue(),
7700 Ld->getSrcValueOffset(), Ld->isVolatile(),
7701 Ld->getAlignment());
7702 SDValue NewChain = NewLd.getValue(1);
7703 if (TokenFactorIndex != -1) {
7704 Ops.push_back(NewChain);
7705 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7708 return DAG.getStore(NewChain, NewLd, St->getBasePtr(),
7709 St->getSrcValue(), St->getSrcValueOffset(),
7710 St->isVolatile(), St->getAlignment());
7713 // Otherwise, lower to two 32-bit copies.
7714 SDValue LoAddr = Ld->getBasePtr();
7715 SDValue HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7716 DAG.getConstant(4, MVT::i32));
7718 SDValue LoLd = DAG.getLoad(MVT::i32, Ld->getChain(), LoAddr,
7719 Ld->getSrcValue(), Ld->getSrcValueOffset(),
7720 Ld->isVolatile(), Ld->getAlignment());
7721 SDValue HiLd = DAG.getLoad(MVT::i32, Ld->getChain(), HiAddr,
7722 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
7724 MinAlign(Ld->getAlignment(), 4));
7726 SDValue NewChain = LoLd.getValue(1);
7727 if (TokenFactorIndex != -1) {
7728 Ops.push_back(LoLd);
7729 Ops.push_back(HiLd);
7730 NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0],
7734 LoAddr = St->getBasePtr();
7735 HiAddr = DAG.getNode(ISD::ADD, MVT::i32, LoAddr,
7736 DAG.getConstant(4, MVT::i32));
7738 SDValue LoSt = DAG.getStore(NewChain, LoLd, LoAddr,
7739 St->getSrcValue(), St->getSrcValueOffset(),
7740 St->isVolatile(), St->getAlignment());
7741 SDValue HiSt = DAG.getStore(NewChain, HiLd, HiAddr,
7743 St->getSrcValueOffset() + 4,
7745 MinAlign(St->getAlignment(), 4));
7746 return DAG.getNode(ISD::TokenFactor, MVT::Other, LoSt, HiSt);
7752 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
7753 /// X86ISD::FXOR nodes.
7754 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
7755 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
7756 // F[X]OR(0.0, x) -> x
7757 // F[X]OR(x, 0.0) -> x
7758 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7759 if (C->getValueAPF().isPosZero())
7760 return N->getOperand(1);
7761 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7762 if (C->getValueAPF().isPosZero())
7763 return N->getOperand(0);
7767 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
7768 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
7769 // FAND(0.0, x) -> 0.0
7770 // FAND(x, 0.0) -> 0.0
7771 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
7772 if (C->getValueAPF().isPosZero())
7773 return N->getOperand(0);
7774 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
7775 if (C->getValueAPF().isPosZero())
7776 return N->getOperand(1);
7781 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
7782 DAGCombinerInfo &DCI) const {
7783 SelectionDAG &DAG = DCI.DAG;
7784 switch (N->getOpcode()) {
7786 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
7787 case ISD::BUILD_VECTOR:
7788 return PerformBuildVectorCombine(N, DAG, Subtarget, *this);
7789 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
7790 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
7792 case X86ISD::FOR: return PerformFORCombine(N, DAG);
7793 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
7799 //===----------------------------------------------------------------------===//
7800 // X86 Inline Assembly Support
7801 //===----------------------------------------------------------------------===//
7803 /// getConstraintType - Given a constraint letter, return the type of
7804 /// constraint it is for this target.
7805 X86TargetLowering::ConstraintType
7806 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
7807 if (Constraint.size() == 1) {
7808 switch (Constraint[0]) {
7820 return C_RegisterClass;
7825 return TargetLowering::getConstraintType(Constraint);
7828 /// LowerXConstraint - try to replace an X constraint, which matches anything,
7829 /// with another that has more specific requirements based on the type of the
7830 /// corresponding operand.
7831 const char *X86TargetLowering::
7832 LowerXConstraint(MVT ConstraintVT) const {
7833 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
7834 // 'f' like normal targets.
7835 if (ConstraintVT.isFloatingPoint()) {
7836 if (Subtarget->hasSSE2())
7838 if (Subtarget->hasSSE1())
7842 return TargetLowering::LowerXConstraint(ConstraintVT);
7845 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7846 /// vector. If it is invalid, don't add anything to Ops.
7847 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7850 std::vector<SDValue>&Ops,
7851 SelectionDAG &DAG) const {
7852 SDValue Result(0, 0);
7854 switch (Constraint) {
7857 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7858 if (C->getZExtValue() <= 31) {
7859 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7866 if (C->getZExtValue() <= 63) {
7867 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7873 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
7874 if (C->getZExtValue() <= 255) {
7875 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
7881 // Literal immediates are always ok.
7882 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
7883 Result = DAG.getTargetConstant(CST->getZExtValue(), Op.getValueType());
7887 // If we are in non-pic codegen mode, we allow the address of a global (with
7888 // an optional displacement) to be used with 'i'.
7889 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
7892 // Match either (GA) or (GA+C)
7894 Offset = GA->getOffset();
7895 } else if (Op.getOpcode() == ISD::ADD) {
7896 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7897 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7899 Offset = GA->getOffset()+C->getZExtValue();
7901 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7902 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
7904 Offset = GA->getOffset()+C->getZExtValue();
7912 Op = LowerGlobalAddress(GA->getGlobal(), Offset, DAG);
7914 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
7920 // Otherwise, not valid for this mode.
7925 if (Result.getNode()) {
7926 Ops.push_back(Result);
7929 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
7933 std::vector<unsigned> X86TargetLowering::
7934 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7936 if (Constraint.size() == 1) {
7937 // FIXME: not handling fp-stack yet!
7938 switch (Constraint[0]) { // GCC X86 Constraint Letters
7939 default: break; // Unknown constraint letter
7940 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
7943 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
7944 else if (VT == MVT::i16)
7945 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
7946 else if (VT == MVT::i8)
7947 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
7948 else if (VT == MVT::i64)
7949 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
7954 return std::vector<unsigned>();
7957 std::pair<unsigned, const TargetRegisterClass*>
7958 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7960 // First, see if this is a constraint that directly corresponds to an LLVM
7962 if (Constraint.size() == 1) {
7963 // GCC Constraint Letters
7964 switch (Constraint[0]) {
7966 case 'r': // GENERAL_REGS
7967 case 'R': // LEGACY_REGS
7968 case 'l': // INDEX_REGS
7970 return std::make_pair(0U, X86::GR8RegisterClass);
7972 return std::make_pair(0U, X86::GR16RegisterClass);
7973 if (VT == MVT::i32 || !Subtarget->is64Bit())
7974 return std::make_pair(0U, X86::GR32RegisterClass);
7975 return std::make_pair(0U, X86::GR64RegisterClass);
7976 case 'f': // FP Stack registers.
7977 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
7978 // value to the correct fpstack register class.
7979 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
7980 return std::make_pair(0U, X86::RFP32RegisterClass);
7981 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
7982 return std::make_pair(0U, X86::RFP64RegisterClass);
7983 return std::make_pair(0U, X86::RFP80RegisterClass);
7984 case 'y': // MMX_REGS if MMX allowed.
7985 if (!Subtarget->hasMMX()) break;
7986 return std::make_pair(0U, X86::VR64RegisterClass);
7987 case 'Y': // SSE_REGS if SSE2 allowed
7988 if (!Subtarget->hasSSE2()) break;
7990 case 'x': // SSE_REGS if SSE1 allowed
7991 if (!Subtarget->hasSSE1()) break;
7993 switch (VT.getSimpleVT()) {
7995 // Scalar SSE types.
7998 return std::make_pair(0U, X86::FR32RegisterClass);
8001 return std::make_pair(0U, X86::FR64RegisterClass);
8009 return std::make_pair(0U, X86::VR128RegisterClass);
8015 // Use the default implementation in TargetLowering to convert the register
8016 // constraint into a member of a register class.
8017 std::pair<unsigned, const TargetRegisterClass*> Res;
8018 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8020 // Not found as a standard register?
8021 if (Res.second == 0) {
8022 // GCC calls "st(0)" just plain "st".
8023 if (StringsEqualNoCase("{st}", Constraint)) {
8024 Res.first = X86::ST0;
8025 Res.second = X86::RFP80RegisterClass;
8027 // 'A' means EAX + EDX.
8028 if (Constraint == "A") {
8029 Res.first = X86::EAX;
8030 Res.second = X86::GRADRegisterClass;
8035 // Otherwise, check to see if this is a register class of the wrong value
8036 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8037 // turn into {ax},{dx}.
8038 if (Res.second->hasType(VT))
8039 return Res; // Correct type already, nothing to do.
8041 // All of the single-register GCC register classes map their values onto
8042 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8043 // really want an 8-bit or 32-bit register, map to the appropriate register
8044 // class and return the appropriate register.
8045 if (Res.second == X86::GR16RegisterClass) {
8046 if (VT == MVT::i8) {
8047 unsigned DestReg = 0;
8048 switch (Res.first) {
8050 case X86::AX: DestReg = X86::AL; break;
8051 case X86::DX: DestReg = X86::DL; break;
8052 case X86::CX: DestReg = X86::CL; break;
8053 case X86::BX: DestReg = X86::BL; break;
8056 Res.first = DestReg;
8057 Res.second = Res.second = X86::GR8RegisterClass;
8059 } else if (VT == MVT::i32) {
8060 unsigned DestReg = 0;
8061 switch (Res.first) {
8063 case X86::AX: DestReg = X86::EAX; break;
8064 case X86::DX: DestReg = X86::EDX; break;
8065 case X86::CX: DestReg = X86::ECX; break;
8066 case X86::BX: DestReg = X86::EBX; break;
8067 case X86::SI: DestReg = X86::ESI; break;
8068 case X86::DI: DestReg = X86::EDI; break;
8069 case X86::BP: DestReg = X86::EBP; break;
8070 case X86::SP: DestReg = X86::ESP; break;
8073 Res.first = DestReg;
8074 Res.second = Res.second = X86::GR32RegisterClass;
8076 } else if (VT == MVT::i64) {
8077 unsigned DestReg = 0;
8078 switch (Res.first) {
8080 case X86::AX: DestReg = X86::RAX; break;
8081 case X86::DX: DestReg = X86::RDX; break;
8082 case X86::CX: DestReg = X86::RCX; break;
8083 case X86::BX: DestReg = X86::RBX; break;
8084 case X86::SI: DestReg = X86::RSI; break;
8085 case X86::DI: DestReg = X86::RDI; break;
8086 case X86::BP: DestReg = X86::RBP; break;
8087 case X86::SP: DestReg = X86::RSP; break;
8090 Res.first = DestReg;
8091 Res.second = Res.second = X86::GR64RegisterClass;
8094 } else if (Res.second == X86::FR32RegisterClass ||
8095 Res.second == X86::FR64RegisterClass ||
8096 Res.second == X86::VR128RegisterClass) {
8097 // Handle references to XMM physical registers that got mapped into the
8098 // wrong class. This can happen with constraints like {xmm0} where the
8099 // target independent register mapper will just pick the first match it can
8100 // find, ignoring the required type.
8102 Res.second = X86::FR32RegisterClass;
8103 else if (VT == MVT::f64)
8104 Res.second = X86::FR64RegisterClass;
8105 else if (X86::VR128RegisterClass->hasType(VT))
8106 Res.second = X86::VR128RegisterClass;
8112 //===----------------------------------------------------------------------===//
8113 // X86 Widen vector type
8114 //===----------------------------------------------------------------------===//
8116 /// getWidenVectorType: given a vector type, returns the type to widen
8117 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8118 /// If there is no vector type that we want to widen to, returns MVT::Other
8119 /// When and where to widen is target dependent based on the cost of
8120 /// scalarizing vs using the wider vector type.
8122 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8123 assert(VT.isVector());
8124 if (isTypeLegal(VT))
8127 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8128 // type based on element type. This would speed up our search (though
8129 // it may not be worth it since the size of the list is relatively
8131 MVT EltVT = VT.getVectorElementType();
8132 unsigned NElts = VT.getVectorNumElements();
8134 // On X86, it make sense to widen any vector wider than 1
8138 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8139 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8140 MVT SVT = (MVT::SimpleValueType)nVT;
8142 if (isTypeLegal(SVT) &&
8143 SVT.getVectorElementType() == EltVT &&
8144 SVT.getVectorNumElements() > NElts)