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, DebugLoc dl);
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::i32 , Promote);
117 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
119 if (!UseSoftFloat && !NoImplicitFloat && 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);
130 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
132 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
133 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
135 if (!UseSoftFloat && !NoImplicitFloat) {
136 // SSE has no i16 to fp conversion, only i32
137 if (X86ScalarSSEf32) {
138 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
139 // f32 and f64 cases are Legal, f80 case is not
140 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
142 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
143 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
146 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
150 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
151 // are Legal, f80 is custom lowered.
152 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
153 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
155 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
157 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
158 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
160 if (X86ScalarSSEf32) {
161 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
162 // f32 and f64 cases are Legal, f80 case is not
163 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
165 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
166 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
169 // Handle FP_TO_UINT by promoting the destination to a larger signed
171 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
172 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
173 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
175 if (Subtarget->is64Bit()) {
176 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
177 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
179 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
180 // Expand FP_TO_UINT into a select.
181 // FIXME: We would like to use a Custom expander here eventually to do
182 // the optimal thing for SSE vs. the default expansion in the legalizer.
183 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
185 // With SSE3 we can use fisttpll to convert to a signed i64.
186 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
189 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
190 if (!X86ScalarSSEf64) {
191 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
192 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
195 // Scalar integer divide and remainder are lowered to use operations that
196 // produce two results, to match the available instructions. This exposes
197 // the two-result form to trivial CSE, which is able to combine x/y and x%y
198 // into a single instruction.
200 // Scalar integer multiply-high is also lowered to use two-result
201 // operations, to match the available instructions. However, plain multiply
202 // (low) operations are left as Legal, as there are single-result
203 // instructions for this in x86. Using the two-result multiply instructions
204 // when both high and low results are needed must be arranged by dagcombine.
205 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
206 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
207 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
208 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
209 setOperationAction(ISD::SREM , MVT::i8 , Expand);
210 setOperationAction(ISD::UREM , MVT::i8 , Expand);
211 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
212 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
213 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
214 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
215 setOperationAction(ISD::SREM , MVT::i16 , Expand);
216 setOperationAction(ISD::UREM , MVT::i16 , Expand);
217 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
218 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
219 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
220 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
221 setOperationAction(ISD::SREM , MVT::i32 , Expand);
222 setOperationAction(ISD::UREM , MVT::i32 , Expand);
223 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
224 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
225 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
226 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
227 setOperationAction(ISD::SREM , MVT::i64 , Expand);
228 setOperationAction(ISD::UREM , MVT::i64 , Expand);
230 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
231 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
232 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
233 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
234 if (Subtarget->is64Bit())
235 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
236 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
237 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
238 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
239 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
240 setOperationAction(ISD::FREM , MVT::f32 , Expand);
241 setOperationAction(ISD::FREM , MVT::f64 , Expand);
242 setOperationAction(ISD::FREM , MVT::f80 , Expand);
243 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
245 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
246 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
247 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
248 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
249 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
250 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
251 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
252 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
253 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
256 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
257 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
260 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
261 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
263 // These should be promoted to a larger select which is supported.
264 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
265 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
266 // X86 wants to expand cmov itself.
267 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
268 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
269 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
270 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
271 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
272 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
274 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
275 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
276 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
277 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
278 if (Subtarget->is64Bit()) {
279 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
280 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
282 // X86 ret instruction may pop stack.
283 setOperationAction(ISD::RET , MVT::Other, Custom);
284 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
287 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
288 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
289 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
290 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
291 if (Subtarget->is64Bit())
292 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
293 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
294 if (Subtarget->is64Bit()) {
295 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
296 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
297 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
298 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
300 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
301 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
302 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
303 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
304 if (Subtarget->is64Bit()) {
305 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
306 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
307 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
310 if (Subtarget->hasSSE1())
311 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
313 if (!Subtarget->hasSSE2())
314 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
316 // Expand certain atomics
317 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
318 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
319 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
320 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
327 if (!Subtarget->is64Bit()) {
328 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
329 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
333 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
334 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
337 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
338 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
339 // FIXME - use subtarget debug flags
340 if (!Subtarget->isTargetDarwin() &&
341 !Subtarget->isTargetELF() &&
342 !Subtarget->isTargetCygMing()) {
343 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
344 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
347 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
348 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
349 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
350 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
351 if (Subtarget->is64Bit()) {
352 setExceptionPointerRegister(X86::RAX);
353 setExceptionSelectorRegister(X86::RDX);
355 setExceptionPointerRegister(X86::EAX);
356 setExceptionSelectorRegister(X86::EDX);
358 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
359 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
361 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
363 setOperationAction(ISD::TRAP, MVT::Other, Legal);
365 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
366 setOperationAction(ISD::VASTART , MVT::Other, Custom);
367 setOperationAction(ISD::VAEND , MVT::Other, Expand);
368 if (Subtarget->is64Bit()) {
369 setOperationAction(ISD::VAARG , MVT::Other, Custom);
370 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
372 setOperationAction(ISD::VAARG , MVT::Other, Expand);
373 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
376 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
377 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
378 if (Subtarget->is64Bit())
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
380 if (Subtarget->isTargetCygMing())
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
383 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
385 if (!UseSoftFloat && X86ScalarSSEf64) {
386 // f32 and f64 use SSE.
387 // Set up the FP register classes.
388 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
389 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
391 // Use ANDPD to simulate FABS.
392 setOperationAction(ISD::FABS , MVT::f64, Custom);
393 setOperationAction(ISD::FABS , MVT::f32, Custom);
395 // Use XORP to simulate FNEG.
396 setOperationAction(ISD::FNEG , MVT::f64, Custom);
397 setOperationAction(ISD::FNEG , MVT::f32, Custom);
399 // Use ANDPD and ORPD to simulate FCOPYSIGN.
400 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
401 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
403 // We don't support sin/cos/fmod
404 setOperationAction(ISD::FSIN , MVT::f64, Expand);
405 setOperationAction(ISD::FCOS , MVT::f64, Expand);
406 setOperationAction(ISD::FSIN , MVT::f32, Expand);
407 setOperationAction(ISD::FCOS , MVT::f32, Expand);
409 // Expand FP immediates into loads from the stack, except for the special
411 addLegalFPImmediate(APFloat(+0.0)); // xorpd
412 addLegalFPImmediate(APFloat(+0.0f)); // xorps
414 // Floating truncations from f80 and extensions to f80 go through memory.
415 // If optimizing, we lie about this though and handle it in
416 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
418 setConvertAction(MVT::f32, MVT::f80, Expand);
419 setConvertAction(MVT::f64, MVT::f80, Expand);
420 setConvertAction(MVT::f80, MVT::f32, Expand);
421 setConvertAction(MVT::f80, MVT::f64, Expand);
423 } else if (!UseSoftFloat && X86ScalarSSEf32) {
424 // Use SSE for f32, x87 for f64.
425 // Set up the FP register classes.
426 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
427 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
429 // Use ANDPS to simulate FABS.
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
437 // Use ANDPS and ORPS to simulate FCOPYSIGN.
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 // We don't support sin/cos/fmod
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Special cases we handle for FP constants.
446 addLegalFPImmediate(APFloat(+0.0f)); // xorps
447 addLegalFPImmediate(APFloat(+0.0)); // FLD0
448 addLegalFPImmediate(APFloat(+1.0)); // FLD1
449 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
450 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
452 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
453 // this though and handle it in InstructionSelectPreprocess so that
454 // dagcombine2 can hack on these.
456 setConvertAction(MVT::f32, MVT::f64, Expand);
457 setConvertAction(MVT::f32, MVT::f80, Expand);
458 setConvertAction(MVT::f80, MVT::f32, Expand);
459 setConvertAction(MVT::f64, MVT::f32, Expand);
460 // And x87->x87 truncations also.
461 setConvertAction(MVT::f80, MVT::f64, Expand);
465 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
466 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
468 } else if (!UseSoftFloat) {
469 // f32 and f64 in x87.
470 // Set up the FP register classes.
471 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
472 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
474 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
475 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
476 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
477 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
479 // Floating truncations go through memory. If optimizing, we lie about
480 // this though and handle it in InstructionSelectPreprocess so that
481 // dagcombine2 can hack on these.
483 setConvertAction(MVT::f80, MVT::f32, Expand);
484 setConvertAction(MVT::f64, MVT::f32, Expand);
485 setConvertAction(MVT::f80, MVT::f64, Expand);
489 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
490 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
492 addLegalFPImmediate(APFloat(+0.0)); // FLD0
493 addLegalFPImmediate(APFloat(+1.0)); // FLD1
494 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
495 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
496 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
497 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
498 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
499 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
502 // Long double always uses X87.
504 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
505 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
506 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
509 APFloat TmpFlt(+0.0);
510 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
512 addLegalFPImmediate(TmpFlt); // FLD0
514 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
515 APFloat TmpFlt2(+1.0);
516 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt2); // FLD1
519 TmpFlt2.changeSign();
520 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
524 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
525 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
529 // Always use a library call for pow.
530 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
531 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
534 setOperationAction(ISD::FLOG, MVT::f80, Expand);
535 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
537 setOperationAction(ISD::FEXP, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
540 // First set operation action for all vector types to either promote
541 // (for widening) or expand (for scalarization). Then we will selectively
542 // turn on ones that can be effectively codegen'd.
543 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
544 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
545 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
560 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
590 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
591 // with -msoft-float, disable use of MMX as well.
592 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
593 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
594 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
595 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
596 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
597 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
599 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
600 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
601 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
602 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
604 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
605 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
606 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
607 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
609 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
610 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
612 setOperationAction(ISD::AND, MVT::v8i8, Promote);
613 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
614 setOperationAction(ISD::AND, MVT::v4i16, Promote);
615 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
616 setOperationAction(ISD::AND, MVT::v2i32, Promote);
617 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
618 setOperationAction(ISD::AND, MVT::v1i64, Legal);
620 setOperationAction(ISD::OR, MVT::v8i8, Promote);
621 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
622 setOperationAction(ISD::OR, MVT::v4i16, Promote);
623 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
624 setOperationAction(ISD::OR, MVT::v2i32, Promote);
625 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
626 setOperationAction(ISD::OR, MVT::v1i64, Legal);
628 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
629 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
630 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
631 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
632 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
633 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
634 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
636 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
637 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
638 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
639 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
640 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
641 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
642 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
643 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
644 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
646 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
648 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
657 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
658 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
659 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
660 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
662 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
664 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
665 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
666 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
667 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
668 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
669 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
672 if (!UseSoftFloat && Subtarget->hasSSE1()) {
673 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
675 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
676 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
677 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
678 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
679 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
680 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
681 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
684 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
685 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
689 if (!UseSoftFloat && Subtarget->hasSSE2()) {
690 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
692 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
693 // registers cannot be used even for integer operations.
694 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
695 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
696 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
697 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
699 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
700 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
701 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
702 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
703 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
704 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
705 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
706 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
707 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
708 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
709 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
710 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
711 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
712 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
713 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
714 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
716 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
721 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
722 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
723 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
724 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
727 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
728 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
729 MVT VT = (MVT::SimpleValueType)i;
730 // Do not attempt to custom lower non-power-of-2 vectors
731 if (!isPowerOf2_32(VT.getVectorNumElements()))
733 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
734 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
735 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
738 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
739 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
740 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
741 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
745 if (Subtarget->is64Bit()) {
746 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
747 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
750 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
751 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
752 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
753 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
754 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
755 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
756 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
757 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
758 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
759 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
760 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
761 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
764 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
766 // Custom lower v2i64 and v2f64 selects.
767 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
768 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
769 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
770 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
774 if (Subtarget->hasSSE41()) {
775 // FIXME: Do we need to handle scalar-to-vector here?
776 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
778 // i8 and i16 vectors are custom , because the source register and source
779 // source memory operand types are not the same width. f32 vectors are
780 // custom since the immediate controlling the insert encodes additional
782 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
783 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
792 if (Subtarget->is64Bit()) {
793 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
794 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
798 if (Subtarget->hasSSE42()) {
799 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
802 // We want to custom lower some of our intrinsics.
803 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
805 // Add/Sub/Mul with overflow operations are custom lowered.
806 setOperationAction(ISD::SADDO, MVT::i32, Custom);
807 setOperationAction(ISD::SADDO, MVT::i64, Custom);
808 setOperationAction(ISD::UADDO, MVT::i32, Custom);
809 setOperationAction(ISD::UADDO, MVT::i64, Custom);
810 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
811 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
812 setOperationAction(ISD::USUBO, MVT::i32, Custom);
813 setOperationAction(ISD::USUBO, MVT::i64, Custom);
814 setOperationAction(ISD::SMULO, MVT::i32, Custom);
815 setOperationAction(ISD::SMULO, MVT::i64, Custom);
816 setOperationAction(ISD::UMULO, MVT::i32, Custom);
817 setOperationAction(ISD::UMULO, MVT::i64, Custom);
819 if (!Subtarget->is64Bit()) {
820 // These libcalls are not available in 32-bit.
821 setLibcallName(RTLIB::SHL_I128, 0);
822 setLibcallName(RTLIB::SRL_I128, 0);
823 setLibcallName(RTLIB::SRA_I128, 0);
826 // We have target-specific dag combine patterns for the following nodes:
827 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
828 setTargetDAGCombine(ISD::BUILD_VECTOR);
829 setTargetDAGCombine(ISD::SELECT);
830 setTargetDAGCombine(ISD::SHL);
831 setTargetDAGCombine(ISD::SRA);
832 setTargetDAGCombine(ISD::SRL);
833 setTargetDAGCombine(ISD::STORE);
834 if (Subtarget->is64Bit())
835 setTargetDAGCombine(ISD::MUL);
837 computeRegisterProperties();
839 // FIXME: These should be based on subtarget info. Plus, the values should
840 // be smaller when we are in optimizing for size mode.
841 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
842 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
843 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
844 allowUnalignedMemoryAccesses = true; // x86 supports it!
845 setPrefLoopAlignment(16);
849 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
854 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
855 /// the desired ByVal argument alignment.
856 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
859 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
860 if (VTy->getBitWidth() == 128)
862 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
863 unsigned EltAlign = 0;
864 getMaxByValAlign(ATy->getElementType(), EltAlign);
865 if (EltAlign > MaxAlign)
867 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
868 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
869 unsigned EltAlign = 0;
870 getMaxByValAlign(STy->getElementType(i), EltAlign);
871 if (EltAlign > MaxAlign)
880 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
881 /// function arguments in the caller parameter area. For X86, aggregates
882 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
883 /// are at 4-byte boundaries.
884 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
885 if (Subtarget->is64Bit()) {
886 // Max of 8 and alignment of type.
887 unsigned TyAlign = TD->getABITypeAlignment(Ty);
894 if (Subtarget->hasSSE1())
895 getMaxByValAlign(Ty, Align);
899 /// getOptimalMemOpType - Returns the target specific optimal type for load
900 /// and store operations as a result of memset, memcpy, and memmove
901 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
904 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
905 bool isSrcConst, bool isSrcStr) const {
906 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
907 // linux. This is because the stack realignment code can't handle certain
908 // cases like PR2962. This should be removed when PR2962 is fixed.
909 if (!NoImplicitFloat && Subtarget->getStackAlignment() >= 16) {
910 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
912 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
915 if (Subtarget->is64Bit() && Size >= 8)
920 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
922 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
923 SelectionDAG &DAG) const {
924 if (usesGlobalOffsetTable())
925 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
926 if (!Subtarget->isPICStyleRIPRel())
927 // This doesn't have DebugLoc associated with it, but is not really the
928 // same as a Register.
929 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
934 //===----------------------------------------------------------------------===//
935 // Return Value Calling Convention Implementation
936 //===----------------------------------------------------------------------===//
938 #include "X86GenCallingConv.inc"
940 /// LowerRET - Lower an ISD::RET node.
941 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
942 DebugLoc dl = Op.getDebugLoc();
943 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
945 SmallVector<CCValAssign, 16> RVLocs;
946 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
947 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
948 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
949 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
951 // If this is the first return lowered for this function, add the regs to the
952 // liveout set for the function.
953 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
954 for (unsigned i = 0; i != RVLocs.size(); ++i)
955 if (RVLocs[i].isRegLoc())
956 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
958 SDValue Chain = Op.getOperand(0);
960 // Handle tail call return.
961 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
962 if (Chain.getOpcode() == X86ISD::TAILCALL) {
963 SDValue TailCall = Chain;
964 SDValue TargetAddress = TailCall.getOperand(1);
965 SDValue StackAdjustment = TailCall.getOperand(2);
966 assert(((TargetAddress.getOpcode() == ISD::Register &&
967 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
968 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
969 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
970 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
971 "Expecting an global address, external symbol, or register");
972 assert(StackAdjustment.getOpcode() == ISD::Constant &&
973 "Expecting a const value");
975 SmallVector<SDValue,8> Operands;
976 Operands.push_back(Chain.getOperand(0));
977 Operands.push_back(TargetAddress);
978 Operands.push_back(StackAdjustment);
979 // Copy registers used by the call. Last operand is a flag so it is not
981 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
982 Operands.push_back(Chain.getOperand(i));
984 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
991 SmallVector<SDValue, 6> RetOps;
992 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
993 // Operand #1 = Bytes To Pop
994 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
996 // Copy the result values into the output registers.
997 for (unsigned i = 0; i != RVLocs.size(); ++i) {
998 CCValAssign &VA = RVLocs[i];
999 assert(VA.isRegLoc() && "Can only return in registers!");
1000 SDValue ValToCopy = Op.getOperand(i*2+1);
1002 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1003 // the RET instruction and handled by the FP Stackifier.
1004 if (VA.getLocReg() == X86::ST0 ||
1005 VA.getLocReg() == X86::ST1) {
1006 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1007 // change the value to the FP stack register class.
1008 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1009 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1010 RetOps.push_back(ValToCopy);
1011 // Don't emit a copytoreg.
1015 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1016 // which is returned in RAX / RDX.
1017 if (Subtarget->is64Bit()) {
1018 MVT ValVT = ValToCopy.getValueType();
1019 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1020 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1021 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1022 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1026 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1027 Flag = Chain.getValue(1);
1030 // The x86-64 ABI for returning structs by value requires that we copy
1031 // the sret argument into %rax for the return. We saved the argument into
1032 // a virtual register in the entry block, so now we copy the value out
1034 if (Subtarget->is64Bit() &&
1035 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1036 MachineFunction &MF = DAG.getMachineFunction();
1037 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1038 unsigned Reg = FuncInfo->getSRetReturnReg();
1040 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1041 FuncInfo->setSRetReturnReg(Reg);
1043 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1045 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1046 Flag = Chain.getValue(1);
1049 RetOps[0] = Chain; // Update chain.
1051 // Add the flag if we have it.
1053 RetOps.push_back(Flag);
1055 return DAG.getNode(X86ISD::RET_FLAG, dl,
1056 MVT::Other, &RetOps[0], RetOps.size());
1060 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1061 /// appropriate copies out of appropriate physical registers. This assumes that
1062 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1063 /// being lowered. The returns a SDNode with the same number of values as the
1065 SDNode *X86TargetLowering::
1066 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1067 unsigned CallingConv, SelectionDAG &DAG) {
1069 DebugLoc dl = TheCall->getDebugLoc();
1070 // Assign locations to each value returned by this call.
1071 SmallVector<CCValAssign, 16> RVLocs;
1072 bool isVarArg = TheCall->isVarArg();
1073 bool Is64Bit = Subtarget->is64Bit();
1074 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1075 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1077 SmallVector<SDValue, 8> ResultVals;
1079 // Copy all of the result registers out of their specified physreg.
1080 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1081 CCValAssign &VA = RVLocs[i];
1082 MVT CopyVT = VA.getValVT();
1084 // If this is x86-64, and we disabled SSE, we can't return FP values
1085 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1086 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1087 cerr << "SSE register return with SSE disabled\n";
1091 // If this is a call to a function that returns an fp value on the floating
1092 // point stack, but where we prefer to use the value in xmm registers, copy
1093 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1094 if ((VA.getLocReg() == X86::ST0 ||
1095 VA.getLocReg() == X86::ST1) &&
1096 isScalarFPTypeInSSEReg(VA.getValVT())) {
1101 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1102 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1103 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1104 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1105 MVT::v2i64, InFlag).getValue(1);
1106 Val = Chain.getValue(0);
1107 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1108 Val, DAG.getConstant(0, MVT::i64));
1110 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1111 MVT::i64, InFlag).getValue(1);
1112 Val = Chain.getValue(0);
1114 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1116 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1117 CopyVT, InFlag).getValue(1);
1118 Val = Chain.getValue(0);
1120 InFlag = Chain.getValue(2);
1122 if (CopyVT != VA.getValVT()) {
1123 // Round the F80 the right size, which also moves to the appropriate xmm
1125 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1126 // This truncation won't change the value.
1127 DAG.getIntPtrConstant(1));
1130 ResultVals.push_back(Val);
1133 // Merge everything together with a MERGE_VALUES node.
1134 ResultVals.push_back(Chain);
1135 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1136 &ResultVals[0], ResultVals.size()).getNode();
1140 //===----------------------------------------------------------------------===//
1141 // C & StdCall & Fast Calling Convention implementation
1142 //===----------------------------------------------------------------------===//
1143 // StdCall calling convention seems to be standard for many Windows' API
1144 // routines and around. It differs from C calling convention just a little:
1145 // callee should clean up the stack, not caller. Symbols should be also
1146 // decorated in some fancy way :) It doesn't support any vector arguments.
1147 // For info on fast calling convention see Fast Calling Convention (tail call)
1148 // implementation LowerX86_32FastCCCallTo.
1150 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1152 static bool CallIsStructReturn(CallSDNode *TheCall) {
1153 unsigned NumOps = TheCall->getNumArgs();
1157 return TheCall->getArgFlags(0).isSRet();
1160 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1161 /// return semantics.
1162 static bool ArgsAreStructReturn(SDValue Op) {
1163 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1167 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1170 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1171 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1173 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1177 switch (CallingConv) {
1180 case CallingConv::X86_StdCall:
1181 return !Subtarget->is64Bit();
1182 case CallingConv::X86_FastCall:
1183 return !Subtarget->is64Bit();
1184 case CallingConv::Fast:
1185 return PerformTailCallOpt;
1189 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1190 /// given CallingConvention value.
1191 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1192 if (Subtarget->is64Bit()) {
1193 if (Subtarget->isTargetWin64())
1194 return CC_X86_Win64_C;
1195 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1196 return CC_X86_64_TailCall;
1201 if (CC == CallingConv::X86_FastCall)
1202 return CC_X86_32_FastCall;
1203 else if (CC == CallingConv::Fast)
1204 return CC_X86_32_FastCC;
1209 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1210 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1212 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1213 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1214 if (CC == CallingConv::X86_FastCall)
1216 else if (CC == CallingConv::X86_StdCall)
1222 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1223 /// in a register before calling.
1224 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1225 return !IsTailCall && !Is64Bit &&
1226 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1227 Subtarget->isPICStyleGOT();
1230 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1231 /// address to be loaded in a register.
1233 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1234 return !Is64Bit && IsTailCall &&
1235 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1236 Subtarget->isPICStyleGOT();
1239 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1240 /// by "Src" to address "Dst" with size and alignment information specified by
1241 /// the specific parameter attribute. The copy will be passed as a byval
1242 /// function parameter.
1244 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1245 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1247 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1248 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1249 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1252 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1253 const CCValAssign &VA,
1254 MachineFrameInfo *MFI,
1256 SDValue Root, unsigned i) {
1257 // Create the nodes corresponding to a load from this parameter slot.
1258 ISD::ArgFlagsTy Flags =
1259 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1260 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1261 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1263 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1264 // changed with more analysis.
1265 // In case of tail call optimization mark all arguments mutable. Since they
1266 // could be overwritten by lowering of arguments in case of a tail call.
1267 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1268 VA.getLocMemOffset(), isImmutable);
1269 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1270 if (Flags.isByVal())
1272 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1273 PseudoSourceValue::getFixedStack(FI), 0);
1277 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1278 MachineFunction &MF = DAG.getMachineFunction();
1279 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1280 DebugLoc dl = Op.getDebugLoc();
1282 const Function* Fn = MF.getFunction();
1283 if (Fn->hasExternalLinkage() &&
1284 Subtarget->isTargetCygMing() &&
1285 Fn->getName() == "main")
1286 FuncInfo->setForceFramePointer(true);
1288 // Decorate the function name.
1289 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1291 MachineFrameInfo *MFI = MF.getFrameInfo();
1292 SDValue Root = Op.getOperand(0);
1293 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1294 unsigned CC = MF.getFunction()->getCallingConv();
1295 bool Is64Bit = Subtarget->is64Bit();
1296 bool IsWin64 = Subtarget->isTargetWin64();
1298 assert(!(isVarArg && CC == CallingConv::Fast) &&
1299 "Var args not supported with calling convention fastcc");
1301 // Assign locations to all of the incoming arguments.
1302 SmallVector<CCValAssign, 16> ArgLocs;
1303 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1304 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1306 SmallVector<SDValue, 8> ArgValues;
1307 unsigned LastVal = ~0U;
1308 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1309 CCValAssign &VA = ArgLocs[i];
1310 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1312 assert(VA.getValNo() != LastVal &&
1313 "Don't support value assigned to multiple locs yet");
1314 LastVal = VA.getValNo();
1316 if (VA.isRegLoc()) {
1317 MVT RegVT = VA.getLocVT();
1318 TargetRegisterClass *RC = NULL;
1319 if (RegVT == MVT::i32)
1320 RC = X86::GR32RegisterClass;
1321 else if (Is64Bit && RegVT == MVT::i64)
1322 RC = X86::GR64RegisterClass;
1323 else if (RegVT == MVT::f32)
1324 RC = X86::FR32RegisterClass;
1325 else if (RegVT == MVT::f64)
1326 RC = X86::FR64RegisterClass;
1327 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1328 RC = X86::VR128RegisterClass;
1329 else if (RegVT.isVector()) {
1330 assert(RegVT.getSizeInBits() == 64);
1332 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1334 // Darwin calling convention passes MMX values in either GPRs or
1335 // XMMs in x86-64. Other targets pass them in memory.
1336 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1337 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1340 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1345 assert(0 && "Unknown argument type!");
1348 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1349 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1351 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1352 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1354 if (VA.getLocInfo() == CCValAssign::SExt)
1355 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1356 DAG.getValueType(VA.getValVT()));
1357 else if (VA.getLocInfo() == CCValAssign::ZExt)
1358 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1359 DAG.getValueType(VA.getValVT()));
1361 if (VA.getLocInfo() != CCValAssign::Full)
1362 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1364 // Handle MMX values passed in GPRs.
1365 if (Is64Bit && RegVT != VA.getLocVT()) {
1366 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1367 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1368 else if (RC == X86::VR128RegisterClass) {
1369 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1370 ArgValue, DAG.getConstant(0, MVT::i64));
1371 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1375 ArgValues.push_back(ArgValue);
1377 assert(VA.isMemLoc());
1378 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1382 // The x86-64 ABI for returning structs by value requires that we copy
1383 // the sret argument into %rax for the return. Save the argument into
1384 // a virtual register so that we can access it from the return points.
1385 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1386 MachineFunction &MF = DAG.getMachineFunction();
1387 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1388 unsigned Reg = FuncInfo->getSRetReturnReg();
1390 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1391 FuncInfo->setSRetReturnReg(Reg);
1393 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1394 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1397 unsigned StackSize = CCInfo.getNextStackOffset();
1398 // align stack specially for tail calls
1399 if (PerformTailCallOpt && CC == CallingConv::Fast)
1400 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1402 // If the function takes variable number of arguments, make a frame index for
1403 // the start of the first vararg value... for expansion of llvm.va_start.
1405 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1406 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1409 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1411 // FIXME: We should really autogenerate these arrays
1412 static const unsigned GPR64ArgRegsWin64[] = {
1413 X86::RCX, X86::RDX, X86::R8, X86::R9
1415 static const unsigned XMMArgRegsWin64[] = {
1416 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1418 static const unsigned GPR64ArgRegs64Bit[] = {
1419 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1421 static const unsigned XMMArgRegs64Bit[] = {
1422 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1423 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1425 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1428 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1429 GPR64ArgRegs = GPR64ArgRegsWin64;
1430 XMMArgRegs = XMMArgRegsWin64;
1432 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1433 GPR64ArgRegs = GPR64ArgRegs64Bit;
1434 XMMArgRegs = XMMArgRegs64Bit;
1436 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1438 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1441 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1442 "SSE register cannot be used when SSE is disabled!");
1443 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloat) &&
1444 "SSE register cannot be used when SSE is disabled!");
1445 if (UseSoftFloat || NoImplicitFloat || !Subtarget->hasSSE1())
1446 // Kernel mode asks for SSE to be disabled, so don't push them
1448 TotalNumXMMRegs = 0;
1450 // For X86-64, if there are vararg parameters that are passed via
1451 // registers, then we must store them to their spots on the stack so they
1452 // may be loaded by deferencing the result of va_next.
1453 VarArgsGPOffset = NumIntRegs * 8;
1454 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1455 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1456 TotalNumXMMRegs * 16, 16);
1458 // Store the integer parameter registers.
1459 SmallVector<SDValue, 8> MemOps;
1460 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1461 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1462 DAG.getIntPtrConstant(VarArgsGPOffset));
1463 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1464 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1465 X86::GR64RegisterClass);
1466 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1468 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1469 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1470 MemOps.push_back(Store);
1471 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1472 DAG.getIntPtrConstant(8));
1475 // Now store the XMM (fp + vector) parameter registers.
1476 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1477 DAG.getIntPtrConstant(VarArgsFPOffset));
1478 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1479 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1480 X86::VR128RegisterClass);
1481 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1483 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1484 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1485 MemOps.push_back(Store);
1486 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1487 DAG.getIntPtrConstant(16));
1489 if (!MemOps.empty())
1490 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1491 &MemOps[0], MemOps.size());
1495 ArgValues.push_back(Root);
1497 // Some CCs need callee pop.
1498 if (IsCalleePop(isVarArg, CC)) {
1499 BytesToPopOnReturn = StackSize; // Callee pops everything.
1500 BytesCallerReserves = 0;
1502 BytesToPopOnReturn = 0; // Callee pops nothing.
1503 // If this is an sret function, the return should pop the hidden pointer.
1504 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1505 BytesToPopOnReturn = 4;
1506 BytesCallerReserves = StackSize;
1510 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1511 if (CC == CallingConv::X86_FastCall)
1512 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1515 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1517 // Return the new list of results.
1518 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1519 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1523 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1524 const SDValue &StackPtr,
1525 const CCValAssign &VA,
1527 SDValue Arg, ISD::ArgFlagsTy Flags) {
1528 DebugLoc dl = TheCall->getDebugLoc();
1529 unsigned LocMemOffset = VA.getLocMemOffset();
1530 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1531 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1532 if (Flags.isByVal()) {
1533 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1535 return DAG.getStore(Chain, dl, Arg, PtrOff,
1536 PseudoSourceValue::getStack(), LocMemOffset);
1539 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1540 /// optimization is performed and it is required.
1542 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1543 SDValue &OutRetAddr,
1549 if (!IsTailCall || FPDiff==0) return Chain;
1551 // Adjust the Return address stack slot.
1552 MVT VT = getPointerTy();
1553 OutRetAddr = getReturnAddressFrameIndex(DAG);
1555 // Load the "old" Return address.
1556 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1557 return SDValue(OutRetAddr.getNode(), 1);
1560 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1561 /// optimization is performed and it is required (FPDiff!=0).
1563 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1564 SDValue Chain, SDValue RetAddrFrIdx,
1565 bool Is64Bit, int FPDiff, DebugLoc dl) {
1566 // Store the return address to the appropriate stack slot.
1567 if (!FPDiff) return Chain;
1568 // Calculate the new stack slot for the return address.
1569 int SlotSize = Is64Bit ? 8 : 4;
1570 int NewReturnAddrFI =
1571 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1572 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1573 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1574 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1575 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1579 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1580 MachineFunction &MF = DAG.getMachineFunction();
1581 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1582 SDValue Chain = TheCall->getChain();
1583 unsigned CC = TheCall->getCallingConv();
1584 bool isVarArg = TheCall->isVarArg();
1585 bool IsTailCall = TheCall->isTailCall() &&
1586 CC == CallingConv::Fast && PerformTailCallOpt;
1587 SDValue Callee = TheCall->getCallee();
1588 bool Is64Bit = Subtarget->is64Bit();
1589 bool IsStructRet = CallIsStructReturn(TheCall);
1590 DebugLoc dl = TheCall->getDebugLoc();
1592 assert(!(isVarArg && CC == CallingConv::Fast) &&
1593 "Var args not supported with calling convention fastcc");
1595 // Analyze operands of the call, assigning locations to each operand.
1596 SmallVector<CCValAssign, 16> ArgLocs;
1597 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1598 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1600 // Get a count of how many bytes are to be pushed on the stack.
1601 unsigned NumBytes = CCInfo.getNextStackOffset();
1602 if (PerformTailCallOpt && CC == CallingConv::Fast)
1603 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1607 // Lower arguments at fp - stackoffset + fpdiff.
1608 unsigned NumBytesCallerPushed =
1609 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1610 FPDiff = NumBytesCallerPushed - NumBytes;
1612 // Set the delta of movement of the returnaddr stackslot.
1613 // But only set if delta is greater than previous delta.
1614 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1615 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1618 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1620 SDValue RetAddrFrIdx;
1621 // Load return adress for tail calls.
1622 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1625 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1626 SmallVector<SDValue, 8> MemOpChains;
1629 // Walk the register/memloc assignments, inserting copies/loads. In the case
1630 // of tail call optimization arguments are handle later.
1631 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1632 CCValAssign &VA = ArgLocs[i];
1633 SDValue Arg = TheCall->getArg(i);
1634 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1635 bool isByVal = Flags.isByVal();
1637 // Promote the value if needed.
1638 switch (VA.getLocInfo()) {
1639 default: assert(0 && "Unknown loc info!");
1640 case CCValAssign::Full: break;
1641 case CCValAssign::SExt:
1642 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1644 case CCValAssign::ZExt:
1645 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1647 case CCValAssign::AExt:
1648 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1652 if (VA.isRegLoc()) {
1654 MVT RegVT = VA.getLocVT();
1655 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1656 switch (VA.getLocReg()) {
1659 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1661 // Special case: passing MMX values in GPR registers.
1662 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1665 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1666 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1667 // Special case: passing MMX values in XMM registers.
1668 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1669 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1670 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
1671 DAG.getUNDEF(MVT::v2i64), Arg,
1672 getMOVLMask(2, DAG, dl));
1677 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1679 if (!IsTailCall || (IsTailCall && isByVal)) {
1680 assert(VA.isMemLoc());
1681 if (StackPtr.getNode() == 0)
1682 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1684 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1685 Chain, Arg, Flags));
1690 if (!MemOpChains.empty())
1691 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1692 &MemOpChains[0], MemOpChains.size());
1694 // Build a sequence of copy-to-reg nodes chained together with token chain
1695 // and flag operands which copy the outgoing args into registers.
1697 // Tail call byval lowering might overwrite argument registers so in case of
1698 // tail call optimization the copies to registers are lowered later.
1700 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1701 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1702 RegsToPass[i].second, InFlag);
1703 InFlag = Chain.getValue(1);
1706 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1708 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1709 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1710 DAG.getNode(X86ISD::GlobalBaseReg,
1711 DebugLoc::getUnknownLoc(),
1714 InFlag = Chain.getValue(1);
1716 // If we are tail calling and generating PIC/GOT style code load the address
1717 // of the callee into ecx. The value in ecx is used as target of the tail
1718 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1719 // calls on PIC/GOT architectures. Normally we would just put the address of
1720 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1721 // restored (since ebx is callee saved) before jumping to the target@PLT.
1722 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1723 // Note: The actual moving to ecx is done further down.
1724 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1725 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1726 !G->getGlobal()->hasProtectedVisibility())
1727 Callee = LowerGlobalAddress(Callee, DAG);
1728 else if (isa<ExternalSymbolSDNode>(Callee))
1729 Callee = LowerExternalSymbol(Callee,DAG);
1732 if (Is64Bit && isVarArg) {
1733 // From AMD64 ABI document:
1734 // For calls that may call functions that use varargs or stdargs
1735 // (prototype-less calls or calls to functions containing ellipsis (...) in
1736 // the declaration) %al is used as hidden argument to specify the number
1737 // of SSE registers used. The contents of %al do not need to match exactly
1738 // the number of registers, but must be an ubound on the number of SSE
1739 // registers used and is in the range 0 - 8 inclusive.
1741 // FIXME: Verify this on Win64
1742 // Count the number of XMM registers allocated.
1743 static const unsigned XMMArgRegs[] = {
1744 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1745 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1747 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1748 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1749 && "SSE registers cannot be used when SSE is disabled");
1751 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1752 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1753 InFlag = Chain.getValue(1);
1757 // For tail calls lower the arguments to the 'real' stack slot.
1759 SmallVector<SDValue, 8> MemOpChains2;
1762 // Do not flag preceeding copytoreg stuff together with the following stuff.
1764 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1765 CCValAssign &VA = ArgLocs[i];
1766 if (!VA.isRegLoc()) {
1767 assert(VA.isMemLoc());
1768 SDValue Arg = TheCall->getArg(i);
1769 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1770 // Create frame index.
1771 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1772 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1773 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1774 FIN = DAG.getFrameIndex(FI, getPointerTy());
1776 if (Flags.isByVal()) {
1777 // Copy relative to framepointer.
1778 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1779 if (StackPtr.getNode() == 0)
1780 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1782 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1784 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1787 // Store relative to framepointer.
1788 MemOpChains2.push_back(
1789 DAG.getStore(Chain, dl, Arg, FIN,
1790 PseudoSourceValue::getFixedStack(FI), 0));
1795 if (!MemOpChains2.empty())
1796 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1797 &MemOpChains2[0], MemOpChains2.size());
1799 // Copy arguments to their registers.
1800 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1801 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1802 RegsToPass[i].second, InFlag);
1803 InFlag = Chain.getValue(1);
1807 // Store the return address to the appropriate stack slot.
1808 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1812 // If the callee is a GlobalAddress node (quite common, every direct call is)
1813 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1814 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1815 // We should use extra load for direct calls to dllimported functions in
1817 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1818 getTargetMachine(), true))
1819 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1821 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1822 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1823 } else if (IsTailCall) {
1824 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1826 Chain = DAG.getCopyToReg(Chain, dl,
1827 DAG.getRegister(Opc, getPointerTy()),
1829 Callee = DAG.getRegister(Opc, getPointerTy());
1830 // Add register as live out.
1831 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1834 // Returns a chain & a flag for retval copy to use.
1835 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1836 SmallVector<SDValue, 8> Ops;
1839 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1840 DAG.getIntPtrConstant(0, true), InFlag);
1841 InFlag = Chain.getValue(1);
1843 // Returns a chain & a flag for retval copy to use.
1844 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1848 Ops.push_back(Chain);
1849 Ops.push_back(Callee);
1852 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1854 // Add argument registers to the end of the list so that they are known live
1856 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1857 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1858 RegsToPass[i].second.getValueType()));
1860 // Add an implicit use GOT pointer in EBX.
1861 if (!IsTailCall && !Is64Bit &&
1862 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1863 Subtarget->isPICStyleGOT())
1864 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1866 // Add an implicit use of AL for x86 vararg functions.
1867 if (Is64Bit && isVarArg)
1868 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1870 if (InFlag.getNode())
1871 Ops.push_back(InFlag);
1874 assert(InFlag.getNode() &&
1875 "Flag must be set. Depend on flag being set in LowerRET");
1876 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1877 TheCall->getVTList(), &Ops[0], Ops.size());
1879 return SDValue(Chain.getNode(), Op.getResNo());
1882 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1883 InFlag = Chain.getValue(1);
1885 // Create the CALLSEQ_END node.
1886 unsigned NumBytesForCalleeToPush;
1887 if (IsCalleePop(isVarArg, CC))
1888 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1889 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1890 // If this is is a call to a struct-return function, the callee
1891 // pops the hidden struct pointer, so we have to push it back.
1892 // This is common for Darwin/X86, Linux & Mingw32 targets.
1893 NumBytesForCalleeToPush = 4;
1895 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1897 // Returns a flag for retval copy to use.
1898 Chain = DAG.getCALLSEQ_END(Chain,
1899 DAG.getIntPtrConstant(NumBytes, true),
1900 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1903 InFlag = Chain.getValue(1);
1905 // Handle result values, copying them out of physregs into vregs that we
1907 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1912 //===----------------------------------------------------------------------===//
1913 // Fast Calling Convention (tail call) implementation
1914 //===----------------------------------------------------------------------===//
1916 // Like std call, callee cleans arguments, convention except that ECX is
1917 // reserved for storing the tail called function address. Only 2 registers are
1918 // free for argument passing (inreg). Tail call optimization is performed
1920 // * tailcallopt is enabled
1921 // * caller/callee are fastcc
1922 // On X86_64 architecture with GOT-style position independent code only local
1923 // (within module) calls are supported at the moment.
1924 // To keep the stack aligned according to platform abi the function
1925 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1926 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1927 // If a tail called function callee has more arguments than the caller the
1928 // caller needs to make sure that there is room to move the RETADDR to. This is
1929 // achieved by reserving an area the size of the argument delta right after the
1930 // original REtADDR, but before the saved framepointer or the spilled registers
1931 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1943 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1944 /// for a 16 byte align requirement.
1945 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1946 SelectionDAG& DAG) {
1947 MachineFunction &MF = DAG.getMachineFunction();
1948 const TargetMachine &TM = MF.getTarget();
1949 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1950 unsigned StackAlignment = TFI.getStackAlignment();
1951 uint64_t AlignMask = StackAlignment - 1;
1952 int64_t Offset = StackSize;
1953 uint64_t SlotSize = TD->getPointerSize();
1954 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1955 // Number smaller than 12 so just add the difference.
1956 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1958 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1959 Offset = ((~AlignMask) & Offset) + StackAlignment +
1960 (StackAlignment-SlotSize);
1965 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1966 /// following the call is a return. A function is eligible if caller/callee
1967 /// calling conventions match, currently only fastcc supports tail calls, and
1968 /// the function CALL is immediatly followed by a RET.
1969 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1971 SelectionDAG& DAG) const {
1972 if (!PerformTailCallOpt)
1975 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1976 MachineFunction &MF = DAG.getMachineFunction();
1977 unsigned CallerCC = MF.getFunction()->getCallingConv();
1978 unsigned CalleeCC= TheCall->getCallingConv();
1979 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1980 SDValue Callee = TheCall->getCallee();
1981 // On x86/32Bit PIC/GOT tail calls are supported.
1982 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1983 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1986 // Can only do local tail calls (in same module, hidden or protected) on
1987 // x86_64 PIC/GOT at the moment.
1988 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1989 return G->getGlobal()->hasHiddenVisibility()
1990 || G->getGlobal()->hasProtectedVisibility();
1998 X86TargetLowering::createFastISel(MachineFunction &mf,
1999 MachineModuleInfo *mmo,
2001 DenseMap<const Value *, unsigned> &vm,
2002 DenseMap<const BasicBlock *,
2003 MachineBasicBlock *> &bm,
2004 DenseMap<const AllocaInst *, int> &am
2006 , SmallSet<Instruction*, 8> &cil
2009 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2017 //===----------------------------------------------------------------------===//
2018 // Other Lowering Hooks
2019 //===----------------------------------------------------------------------===//
2022 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2023 MachineFunction &MF = DAG.getMachineFunction();
2024 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2025 int ReturnAddrIndex = FuncInfo->getRAIndex();
2027 if (ReturnAddrIndex == 0) {
2028 // Set up a frame object for the return address.
2029 uint64_t SlotSize = TD->getPointerSize();
2030 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2031 FuncInfo->setRAIndex(ReturnAddrIndex);
2034 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2038 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2039 /// specific condition code, returning the condition code and the LHS/RHS of the
2040 /// comparison to make.
2041 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2042 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2044 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2045 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2046 // X > -1 -> X == 0, jump !sign.
2047 RHS = DAG.getConstant(0, RHS.getValueType());
2048 return X86::COND_NS;
2049 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2050 // X < 0 -> X == 0, jump on sign.
2052 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2054 RHS = DAG.getConstant(0, RHS.getValueType());
2055 return X86::COND_LE;
2059 switch (SetCCOpcode) {
2060 default: assert(0 && "Invalid integer condition!");
2061 case ISD::SETEQ: return X86::COND_E;
2062 case ISD::SETGT: return X86::COND_G;
2063 case ISD::SETGE: return X86::COND_GE;
2064 case ISD::SETLT: return X86::COND_L;
2065 case ISD::SETLE: return X86::COND_LE;
2066 case ISD::SETNE: return X86::COND_NE;
2067 case ISD::SETULT: return X86::COND_B;
2068 case ISD::SETUGT: return X86::COND_A;
2069 case ISD::SETULE: return X86::COND_BE;
2070 case ISD::SETUGE: return X86::COND_AE;
2074 // First determine if it is required or is profitable to flip the operands.
2076 // If LHS is a foldable load, but RHS is not, flip the condition.
2077 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2078 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2079 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2080 std::swap(LHS, RHS);
2083 switch (SetCCOpcode) {
2089 std::swap(LHS, RHS);
2093 // On a floating point condition, the flags are set as follows:
2095 // 0 | 0 | 0 | X > Y
2096 // 0 | 0 | 1 | X < Y
2097 // 1 | 0 | 0 | X == Y
2098 // 1 | 1 | 1 | unordered
2099 switch (SetCCOpcode) {
2100 default: assert(0 && "Condcode should be pre-legalized away");
2102 case ISD::SETEQ: return X86::COND_E;
2103 case ISD::SETOLT: // flipped
2105 case ISD::SETGT: return X86::COND_A;
2106 case ISD::SETOLE: // flipped
2108 case ISD::SETGE: return X86::COND_AE;
2109 case ISD::SETUGT: // flipped
2111 case ISD::SETLT: return X86::COND_B;
2112 case ISD::SETUGE: // flipped
2114 case ISD::SETLE: return X86::COND_BE;
2116 case ISD::SETNE: return X86::COND_NE;
2117 case ISD::SETUO: return X86::COND_P;
2118 case ISD::SETO: return X86::COND_NP;
2122 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2123 /// code. Current x86 isa includes the following FP cmov instructions:
2124 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2125 static bool hasFPCMov(unsigned X86CC) {
2141 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2142 /// true if Op is undef or if its value falls within the specified range (L, H].
2143 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2144 if (Op.getOpcode() == ISD::UNDEF)
2147 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2148 return (Val >= Low && Val < Hi);
2151 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2152 /// true if Op is undef or if its value equal to the specified value.
2153 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2154 if (Op.getOpcode() == ISD::UNDEF)
2156 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2159 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2160 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2161 bool X86::isPSHUFDMask(SDNode *N) {
2162 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2164 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2167 // Check if the value doesn't reference the second vector.
2168 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2169 SDValue Arg = N->getOperand(i);
2170 if (Arg.getOpcode() == ISD::UNDEF) continue;
2171 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2172 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2179 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2180 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2181 bool X86::isPSHUFHWMask(SDNode *N) {
2182 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2184 if (N->getNumOperands() != 8)
2187 // Lower quadword copied in order.
2188 for (unsigned i = 0; i != 4; ++i) {
2189 SDValue Arg = N->getOperand(i);
2190 if (Arg.getOpcode() == ISD::UNDEF) continue;
2191 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2192 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2196 // Upper quadword shuffled.
2197 for (unsigned i = 4; i != 8; ++i) {
2198 SDValue Arg = N->getOperand(i);
2199 if (Arg.getOpcode() == ISD::UNDEF) continue;
2200 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2201 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2202 if (Val < 4 || Val > 7)
2209 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2210 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2211 bool X86::isPSHUFLWMask(SDNode *N) {
2212 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2214 if (N->getNumOperands() != 8)
2217 // Upper quadword copied in order.
2218 for (unsigned i = 4; i != 8; ++i)
2219 if (!isUndefOrEqual(N->getOperand(i), i))
2222 // Lower quadword shuffled.
2223 for (unsigned i = 0; i != 4; ++i)
2224 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2230 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2231 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2232 template<class SDOperand>
2233 static bool isSHUFPMask(SDOperand *Elems, unsigned NumElems) {
2234 if (NumElems != 2 && NumElems != 4) return false;
2236 unsigned Half = NumElems / 2;
2237 for (unsigned i = 0; i < Half; ++i)
2238 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2240 for (unsigned i = Half; i < NumElems; ++i)
2241 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2247 bool X86::isSHUFPMask(SDNode *N) {
2248 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2249 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2252 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2253 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2254 /// half elements to come from vector 1 (which would equal the dest.) and
2255 /// the upper half to come from vector 2.
2256 template<class SDOperand>
2257 static bool isCommutedSHUFP(SDOperand *Ops, unsigned NumOps) {
2258 if (NumOps != 2 && NumOps != 4) return false;
2260 unsigned Half = NumOps / 2;
2261 for (unsigned i = 0; i < Half; ++i)
2262 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2264 for (unsigned i = Half; i < NumOps; ++i)
2265 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2270 static bool isCommutedSHUFP(SDNode *N) {
2271 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2272 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2275 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2276 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2277 bool X86::isMOVHLPSMask(SDNode *N) {
2278 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2280 if (N->getNumOperands() != 4)
2283 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2284 return isUndefOrEqual(N->getOperand(0), 6) &&
2285 isUndefOrEqual(N->getOperand(1), 7) &&
2286 isUndefOrEqual(N->getOperand(2), 2) &&
2287 isUndefOrEqual(N->getOperand(3), 3);
2290 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2291 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2293 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2294 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2296 if (N->getNumOperands() != 4)
2299 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2300 return isUndefOrEqual(N->getOperand(0), 2) &&
2301 isUndefOrEqual(N->getOperand(1), 3) &&
2302 isUndefOrEqual(N->getOperand(2), 2) &&
2303 isUndefOrEqual(N->getOperand(3), 3);
2306 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2307 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2308 bool X86::isMOVLPMask(SDNode *N) {
2309 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2311 unsigned NumElems = N->getNumOperands();
2312 if (NumElems != 2 && NumElems != 4)
2315 for (unsigned i = 0; i < NumElems/2; ++i)
2316 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2319 for (unsigned i = NumElems/2; i < NumElems; ++i)
2320 if (!isUndefOrEqual(N->getOperand(i), i))
2326 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2327 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2329 bool X86::isMOVHPMask(SDNode *N) {
2330 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2332 unsigned NumElems = N->getNumOperands();
2333 if (NumElems != 2 && NumElems != 4)
2336 for (unsigned i = 0; i < NumElems/2; ++i)
2337 if (!isUndefOrEqual(N->getOperand(i), i))
2340 for (unsigned i = 0; i < NumElems/2; ++i) {
2341 SDValue Arg = N->getOperand(i + NumElems/2);
2342 if (!isUndefOrEqual(Arg, i + NumElems))
2349 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2350 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2351 template<class SDOperand>
2352 bool static isUNPCKLMask(SDOperand *Elts, unsigned NumElts,
2353 bool V2IsSplat = false) {
2354 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2357 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2358 SDValue BitI = Elts[i];
2359 SDValue BitI1 = Elts[i+1];
2360 if (!isUndefOrEqual(BitI, j))
2363 if (!isUndefOrEqual(BitI1, NumElts))
2366 if (!isUndefOrEqual(BitI1, j + NumElts))
2374 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2375 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2376 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2379 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2380 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2381 template<class SDOperand>
2382 bool static isUNPCKHMask(SDOperand *Elts, unsigned NumElts,
2383 bool V2IsSplat = false) {
2384 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2387 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2388 SDValue BitI = Elts[i];
2389 SDValue BitI1 = Elts[i+1];
2390 if (!isUndefOrEqual(BitI, j + NumElts/2))
2393 if (isUndefOrEqual(BitI1, NumElts))
2396 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2404 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2405 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2406 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2409 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2410 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2412 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2413 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2415 unsigned NumElems = N->getNumOperands();
2416 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2419 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2420 SDValue BitI = N->getOperand(i);
2421 SDValue BitI1 = N->getOperand(i+1);
2423 if (!isUndefOrEqual(BitI, j))
2425 if (!isUndefOrEqual(BitI1, j))
2432 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2433 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2435 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2436 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2438 unsigned NumElems = N->getNumOperands();
2439 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2442 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2443 SDValue BitI = N->getOperand(i);
2444 SDValue BitI1 = N->getOperand(i + 1);
2446 if (!isUndefOrEqual(BitI, j))
2448 if (!isUndefOrEqual(BitI1, j))
2455 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2456 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2457 /// MOVSD, and MOVD, i.e. setting the lowest element.
2458 template<class SDOperand>
2459 static bool isMOVLMask(SDOperand *Elts, unsigned NumElts) {
2460 if (NumElts != 2 && NumElts != 4)
2463 if (!isUndefOrEqual(Elts[0], NumElts))
2466 for (unsigned i = 1; i < NumElts; ++i) {
2467 if (!isUndefOrEqual(Elts[i], i))
2474 bool X86::isMOVLMask(SDNode *N) {
2475 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2476 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2479 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2480 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2481 /// element of vector 2 and the other elements to come from vector 1 in order.
2482 template<class SDOperand>
2483 static bool isCommutedMOVL(SDOperand *Ops, unsigned NumOps,
2484 bool V2IsSplat = false,
2485 bool V2IsUndef = false) {
2486 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2489 if (!isUndefOrEqual(Ops[0], 0))
2492 for (unsigned i = 1; i < NumOps; ++i) {
2493 SDValue Arg = Ops[i];
2494 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2495 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2496 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2503 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2504 bool V2IsUndef = false) {
2505 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2506 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2507 V2IsSplat, V2IsUndef);
2510 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2511 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2512 bool X86::isMOVSHDUPMask(SDNode *N) {
2513 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2515 if (N->getNumOperands() != 4)
2518 // Expect 1, 1, 3, 3
2519 for (unsigned i = 0; i < 2; ++i) {
2520 SDValue Arg = N->getOperand(i);
2521 if (Arg.getOpcode() == ISD::UNDEF) continue;
2522 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2523 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2524 if (Val != 1) return false;
2528 for (unsigned i = 2; i < 4; ++i) {
2529 SDValue Arg = N->getOperand(i);
2530 if (Arg.getOpcode() == ISD::UNDEF) continue;
2531 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2532 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2533 if (Val != 3) return false;
2537 // Don't use movshdup if it can be done with a shufps.
2541 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2542 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2543 bool X86::isMOVSLDUPMask(SDNode *N) {
2544 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2546 if (N->getNumOperands() != 4)
2549 // Expect 0, 0, 2, 2
2550 for (unsigned i = 0; i < 2; ++i) {
2551 SDValue Arg = N->getOperand(i);
2552 if (Arg.getOpcode() == ISD::UNDEF) continue;
2553 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2554 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2555 if (Val != 0) return false;
2559 for (unsigned i = 2; i < 4; ++i) {
2560 SDValue Arg = N->getOperand(i);
2561 if (Arg.getOpcode() == ISD::UNDEF) continue;
2562 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2563 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2564 if (Val != 2) return false;
2568 // Don't use movshdup if it can be done with a shufps.
2572 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2573 /// specifies a identity operation on the LHS or RHS.
2574 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2575 unsigned NumElems = N->getNumOperands();
2576 for (unsigned i = 0; i < NumElems; ++i)
2577 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2582 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2583 /// a splat of a single element.
2584 static bool isSplatMask(SDNode *N) {
2585 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2587 // This is a splat operation if each element of the permute is the same, and
2588 // if the value doesn't reference the second vector.
2589 unsigned NumElems = N->getNumOperands();
2590 SDValue ElementBase;
2592 for (; i != NumElems; ++i) {
2593 SDValue Elt = N->getOperand(i);
2594 if (isa<ConstantSDNode>(Elt)) {
2600 if (!ElementBase.getNode())
2603 for (; i != NumElems; ++i) {
2604 SDValue Arg = N->getOperand(i);
2605 if (Arg.getOpcode() == ISD::UNDEF) continue;
2606 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2607 if (Arg != ElementBase) return false;
2610 // Make sure it is a splat of the first vector operand.
2611 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2614 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2615 /// we want to splat.
2616 static SDValue getSplatMaskEltNo(SDNode *N) {
2617 assert(isSplatMask(N) && "Not a splat mask");
2618 unsigned NumElems = N->getNumOperands();
2619 SDValue ElementBase;
2621 for (; i != NumElems; ++i) {
2622 SDValue Elt = N->getOperand(i);
2623 if (isa<ConstantSDNode>(Elt))
2626 assert(0 && " No splat value found!");
2631 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2632 /// a splat of a single element and it's a 2 or 4 element mask.
2633 bool X86::isSplatMask(SDNode *N) {
2634 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2636 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2637 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2639 return ::isSplatMask(N);
2642 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2643 /// specifies a splat of zero element.
2644 bool X86::isSplatLoMask(SDNode *N) {
2645 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2647 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2648 if (!isUndefOrEqual(N->getOperand(i), 0))
2653 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2654 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2655 bool X86::isMOVDDUPMask(SDNode *N) {
2656 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2658 unsigned e = N->getNumOperands() / 2;
2659 for (unsigned i = 0; i < e; ++i)
2660 if (!isUndefOrEqual(N->getOperand(i), i))
2662 for (unsigned i = 0; i < e; ++i)
2663 if (!isUndefOrEqual(N->getOperand(e+i), i))
2668 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2669 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2671 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2672 unsigned NumOperands = N->getNumOperands();
2673 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2675 for (unsigned i = 0; i < NumOperands; ++i) {
2677 SDValue Arg = N->getOperand(NumOperands-i-1);
2678 if (Arg.getOpcode() != ISD::UNDEF)
2679 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2680 if (Val >= NumOperands) Val -= NumOperands;
2682 if (i != NumOperands - 1)
2689 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2690 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2692 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2694 // 8 nodes, but we only care about the last 4.
2695 for (unsigned i = 7; i >= 4; --i) {
2697 SDValue Arg = N->getOperand(i);
2698 if (Arg.getOpcode() != ISD::UNDEF) {
2699 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2709 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2710 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2712 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2714 // 8 nodes, but we only care about the first 4.
2715 for (int i = 3; i >= 0; --i) {
2717 SDValue Arg = N->getOperand(i);
2718 if (Arg.getOpcode() != ISD::UNDEF)
2719 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2728 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2729 /// values in ther permute mask.
2730 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2731 SDValue &V2, SDValue &Mask,
2732 SelectionDAG &DAG) {
2733 MVT VT = Op.getValueType();
2734 MVT MaskVT = Mask.getValueType();
2735 MVT EltVT = MaskVT.getVectorElementType();
2736 unsigned NumElems = Mask.getNumOperands();
2737 SmallVector<SDValue, 8> MaskVec;
2738 DebugLoc dl = Op.getDebugLoc();
2740 for (unsigned i = 0; i != NumElems; ++i) {
2741 SDValue Arg = Mask.getOperand(i);
2742 if (Arg.getOpcode() == ISD::UNDEF) {
2743 MaskVec.push_back(DAG.getUNDEF(EltVT));
2746 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2747 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2749 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2751 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2755 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2756 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
2759 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2760 /// the two vector operands have swapped position.
2762 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG, DebugLoc dl) {
2763 MVT MaskVT = Mask.getValueType();
2764 MVT EltVT = MaskVT.getVectorElementType();
2765 unsigned NumElems = Mask.getNumOperands();
2766 SmallVector<SDValue, 8> MaskVec;
2767 for (unsigned i = 0; i != NumElems; ++i) {
2768 SDValue Arg = Mask.getOperand(i);
2769 if (Arg.getOpcode() == ISD::UNDEF) {
2770 MaskVec.push_back(DAG.getUNDEF(EltVT));
2773 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2774 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2776 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2778 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2780 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2784 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2785 /// match movhlps. The lower half elements should come from upper half of
2786 /// V1 (and in order), and the upper half elements should come from the upper
2787 /// half of V2 (and in order).
2788 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2789 unsigned NumElems = Mask->getNumOperands();
2792 for (unsigned i = 0, e = 2; i != e; ++i)
2793 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2795 for (unsigned i = 2; i != 4; ++i)
2796 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2801 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2802 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2804 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2805 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2807 N = N->getOperand(0).getNode();
2808 if (!ISD::isNON_EXTLoad(N))
2811 *LD = cast<LoadSDNode>(N);
2815 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2816 /// match movlp{s|d}. The lower half elements should come from lower half of
2817 /// V1 (and in order), and the upper half elements should come from the upper
2818 /// half of V2 (and in order). And since V1 will become the source of the
2819 /// MOVLP, it must be either a vector load or a scalar load to vector.
2820 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2821 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2823 // Is V2 is a vector load, don't do this transformation. We will try to use
2824 // load folding shufps op.
2825 if (ISD::isNON_EXTLoad(V2))
2828 unsigned NumElems = Mask->getNumOperands();
2829 if (NumElems != 2 && NumElems != 4)
2831 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2832 if (!isUndefOrEqual(Mask->getOperand(i), i))
2834 for (unsigned i = NumElems/2; i != NumElems; ++i)
2835 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2840 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2842 static bool isSplatVector(SDNode *N) {
2843 if (N->getOpcode() != ISD::BUILD_VECTOR)
2846 SDValue SplatValue = N->getOperand(0);
2847 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2848 if (N->getOperand(i) != SplatValue)
2853 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2855 static bool isUndefShuffle(SDNode *N) {
2856 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2859 SDValue V1 = N->getOperand(0);
2860 SDValue V2 = N->getOperand(1);
2861 SDValue Mask = N->getOperand(2);
2862 unsigned NumElems = Mask.getNumOperands();
2863 for (unsigned i = 0; i != NumElems; ++i) {
2864 SDValue Arg = Mask.getOperand(i);
2865 if (Arg.getOpcode() != ISD::UNDEF) {
2866 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2867 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2869 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2876 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2878 static inline bool isZeroNode(SDValue Elt) {
2879 return ((isa<ConstantSDNode>(Elt) &&
2880 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2881 (isa<ConstantFPSDNode>(Elt) &&
2882 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2885 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2886 /// to an zero vector.
2887 static bool isZeroShuffle(SDNode *N) {
2888 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2891 SDValue V1 = N->getOperand(0);
2892 SDValue V2 = N->getOperand(1);
2893 SDValue Mask = N->getOperand(2);
2894 unsigned NumElems = Mask.getNumOperands();
2895 for (unsigned i = 0; i != NumElems; ++i) {
2896 SDValue Arg = Mask.getOperand(i);
2897 if (Arg.getOpcode() == ISD::UNDEF)
2900 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2901 if (Idx < NumElems) {
2902 unsigned Opc = V1.getNode()->getOpcode();
2903 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2905 if (Opc != ISD::BUILD_VECTOR ||
2906 !isZeroNode(V1.getNode()->getOperand(Idx)))
2908 } else if (Idx >= NumElems) {
2909 unsigned Opc = V2.getNode()->getOpcode();
2910 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2912 if (Opc != ISD::BUILD_VECTOR ||
2913 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2920 /// getZeroVector - Returns a vector of specified type with all zero elements.
2922 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2924 assert(VT.isVector() && "Expected a vector type");
2926 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2927 // type. This ensures they get CSE'd.
2929 if (VT.getSizeInBits() == 64) { // MMX
2930 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2931 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2932 } else if (HasSSE2) { // SSE2
2933 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2934 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2936 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2937 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2939 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2942 /// getOnesVector - Returns a vector of specified type with all bits set.
2944 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2945 assert(VT.isVector() && "Expected a vector type");
2947 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2948 // type. This ensures they get CSE'd.
2949 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2951 if (VT.getSizeInBits() == 64) // MMX
2952 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2954 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2955 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2959 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2960 /// that point to V2 points to its first element.
2961 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2962 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2964 bool Changed = false;
2965 SmallVector<SDValue, 8> MaskVec;
2966 unsigned NumElems = Mask.getNumOperands();
2967 for (unsigned i = 0; i != NumElems; ++i) {
2968 SDValue Arg = Mask.getOperand(i);
2969 if (Arg.getOpcode() != ISD::UNDEF) {
2970 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2971 if (Val > NumElems) {
2972 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2976 MaskVec.push_back(Arg);
2980 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getDebugLoc(),
2981 Mask.getValueType(),
2982 &MaskVec[0], MaskVec.size());
2986 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2987 /// operation of specified width.
2988 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl) {
2989 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2990 MVT BaseVT = MaskVT.getVectorElementType();
2992 SmallVector<SDValue, 8> MaskVec;
2993 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2994 for (unsigned i = 1; i != NumElems; ++i)
2995 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2996 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
2997 &MaskVec[0], MaskVec.size());
3000 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
3001 /// of specified width.
3002 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG,
3004 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3005 MVT BaseVT = MaskVT.getVectorElementType();
3006 SmallVector<SDValue, 8> MaskVec;
3007 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3008 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3009 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3011 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3012 &MaskVec[0], MaskVec.size());
3015 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3016 /// of specified width.
3017 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG,
3019 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3020 MVT BaseVT = MaskVT.getVectorElementType();
3021 unsigned Half = NumElems/2;
3022 SmallVector<SDValue, 8> MaskVec;
3023 for (unsigned i = 0; i != Half; ++i) {
3024 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3025 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3027 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3028 &MaskVec[0], MaskVec.size());
3031 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
3032 /// element #0 of a vector with the specified index, leaving the rest of the
3033 /// elements in place.
3034 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
3035 SelectionDAG &DAG, DebugLoc dl) {
3036 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3037 MVT BaseVT = MaskVT.getVectorElementType();
3038 SmallVector<SDValue, 8> MaskVec;
3039 // Element #0 of the result gets the elt we are replacing.
3040 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
3041 for (unsigned i = 1; i != NumElems; ++i)
3042 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
3043 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3044 &MaskVec[0], MaskVec.size());
3047 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3048 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
3049 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
3050 MVT VT = Op.getValueType();
3053 SDValue V1 = Op.getOperand(0);
3054 SDValue Mask = Op.getOperand(2);
3055 unsigned MaskNumElems = Mask.getNumOperands();
3056 unsigned NumElems = MaskNumElems;
3057 DebugLoc dl = Op.getDebugLoc();
3058 // Special handling of v4f32 -> v4i32.
3059 if (VT != MVT::v4f32) {
3060 // Find which element we want to splat.
3061 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3062 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3063 // unpack elements to the correct location
3064 while (NumElems > 4) {
3065 if (EltNo < NumElems/2) {
3066 Mask = getUnpacklMask(MaskNumElems, DAG, dl);
3068 Mask = getUnpackhMask(MaskNumElems, DAG, dl);
3069 EltNo -= NumElems/2;
3071 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1, Mask);
3074 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3075 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3078 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3079 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3080 DAG.getUNDEF(PVT), Mask);
3081 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3084 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3085 /// load that's promoted to vector, or a load bitcasted.
3086 static bool isVectorLoad(SDValue Op) {
3087 assert(Op.getValueType().isVector() && "Expected a vector type");
3088 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3089 Op.getOpcode() == ISD::BIT_CONVERT) {
3090 return isa<LoadSDNode>(Op.getOperand(0));
3092 return isa<LoadSDNode>(Op);
3096 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3098 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3099 SelectionDAG &DAG, bool HasSSE3) {
3100 // If we have sse3 and shuffle has more than one use or input is a load, then
3101 // use movddup. Otherwise, use movlhps.
3102 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3103 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3104 MVT VT = Op.getValueType();
3107 DebugLoc dl = Op.getDebugLoc();
3108 unsigned NumElems = PVT.getVectorNumElements();
3109 if (NumElems == 2) {
3110 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3111 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3113 assert(NumElems == 4);
3114 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3115 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3116 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3117 Cst0, Cst1, Cst0, Cst1);
3120 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3121 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3122 DAG.getUNDEF(PVT), Mask);
3123 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3126 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3127 /// vector of zero or undef vector. This produces a shuffle where the low
3128 /// element of V2 is swizzled into the zero/undef vector, landing at element
3129 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3130 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3131 bool isZero, bool HasSSE2,
3132 SelectionDAG &DAG) {
3133 DebugLoc dl = V2.getDebugLoc();
3134 MVT VT = V2.getValueType();
3136 ? getZeroVector(VT, HasSSE2, DAG, dl) : DAG.getUNDEF(VT);
3137 unsigned NumElems = V2.getValueType().getVectorNumElements();
3138 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3139 MVT EVT = MaskVT.getVectorElementType();
3140 SmallVector<SDValue, 16> MaskVec;
3141 for (unsigned i = 0; i != NumElems; ++i)
3142 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3143 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3145 MaskVec.push_back(DAG.getConstant(i, EVT));
3146 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3147 &MaskVec[0], MaskVec.size());
3148 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
3151 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3152 /// a shuffle that is zero.
3154 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3155 unsigned NumElems, bool Low,
3156 SelectionDAG &DAG) {
3157 unsigned NumZeros = 0;
3158 for (unsigned i = 0; i < NumElems; ++i) {
3159 unsigned Index = Low ? i : NumElems-i-1;
3160 SDValue Idx = Mask.getOperand(Index);
3161 if (Idx.getOpcode() == ISD::UNDEF) {
3165 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3166 if (Elt.getNode() && isZeroNode(Elt))
3174 /// isVectorShift - Returns true if the shuffle can be implemented as a
3175 /// logical left or right shift of a vector.
3176 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3177 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3178 unsigned NumElems = Mask.getNumOperands();
3181 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3184 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3189 bool SeenV1 = false;
3190 bool SeenV2 = false;
3191 for (unsigned i = NumZeros; i < NumElems; ++i) {
3192 unsigned Val = isLeft ? (i - NumZeros) : i;
3193 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3194 if (Idx.getOpcode() == ISD::UNDEF)
3196 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3197 if (Index < NumElems)
3206 if (SeenV1 && SeenV2)
3209 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3215 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3217 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3218 unsigned NumNonZero, unsigned NumZero,
3219 SelectionDAG &DAG, TargetLowering &TLI) {
3223 DebugLoc dl = Op.getDebugLoc();
3226 for (unsigned i = 0; i < 16; ++i) {
3227 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3228 if (ThisIsNonZero && First) {
3230 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3232 V = DAG.getUNDEF(MVT::v8i16);
3237 SDValue ThisElt(0, 0), LastElt(0, 0);
3238 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3239 if (LastIsNonZero) {
3240 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3241 MVT::i16, Op.getOperand(i-1));
3243 if (ThisIsNonZero) {
3244 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3245 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3246 ThisElt, DAG.getConstant(8, MVT::i8));
3248 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3252 if (ThisElt.getNode())
3253 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3254 DAG.getIntPtrConstant(i/2));
3258 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3261 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3263 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3264 unsigned NumNonZero, unsigned NumZero,
3265 SelectionDAG &DAG, TargetLowering &TLI) {
3269 DebugLoc dl = Op.getDebugLoc();
3272 for (unsigned i = 0; i < 8; ++i) {
3273 bool isNonZero = (NonZeros & (1 << i)) != 0;
3277 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3279 V = DAG.getUNDEF(MVT::v8i16);
3282 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3283 MVT::v8i16, V, Op.getOperand(i),
3284 DAG.getIntPtrConstant(i));
3291 /// getVShift - Return a vector logical shift node.
3293 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3294 unsigned NumBits, SelectionDAG &DAG,
3295 const TargetLowering &TLI, DebugLoc dl) {
3296 bool isMMX = VT.getSizeInBits() == 64;
3297 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3298 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3299 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3300 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3301 DAG.getNode(Opc, dl, ShVT, SrcOp,
3302 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3306 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3307 DebugLoc dl = Op.getDebugLoc();
3308 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3309 if (ISD::isBuildVectorAllZeros(Op.getNode())
3310 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3311 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3312 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3313 // eliminated on x86-32 hosts.
3314 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3317 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3318 return getOnesVector(Op.getValueType(), DAG, dl);
3319 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3322 MVT VT = Op.getValueType();
3323 MVT EVT = VT.getVectorElementType();
3324 unsigned EVTBits = EVT.getSizeInBits();
3326 unsigned NumElems = Op.getNumOperands();
3327 unsigned NumZero = 0;
3328 unsigned NumNonZero = 0;
3329 unsigned NonZeros = 0;
3330 bool IsAllConstants = true;
3331 SmallSet<SDValue, 8> Values;
3332 for (unsigned i = 0; i < NumElems; ++i) {
3333 SDValue Elt = Op.getOperand(i);
3334 if (Elt.getOpcode() == ISD::UNDEF)
3337 if (Elt.getOpcode() != ISD::Constant &&
3338 Elt.getOpcode() != ISD::ConstantFP)
3339 IsAllConstants = false;
3340 if (isZeroNode(Elt))
3343 NonZeros |= (1 << i);
3348 if (NumNonZero == 0) {
3349 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3350 return DAG.getUNDEF(VT);
3353 // Special case for single non-zero, non-undef, element.
3354 if (NumNonZero == 1 && NumElems <= 4) {
3355 unsigned Idx = CountTrailingZeros_32(NonZeros);
3356 SDValue Item = Op.getOperand(Idx);
3358 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3359 // the value are obviously zero, truncate the value to i32 and do the
3360 // insertion that way. Only do this if the value is non-constant or if the
3361 // value is a constant being inserted into element 0. It is cheaper to do
3362 // a constant pool load than it is to do a movd + shuffle.
3363 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3364 (!IsAllConstants || Idx == 0)) {
3365 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3366 // Handle MMX and SSE both.
3367 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3368 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3370 // Truncate the value (which may itself be a constant) to i32, and
3371 // convert it to a vector with movd (S2V+shuffle to zero extend).
3372 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3373 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3374 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3375 Subtarget->hasSSE2(), DAG);
3377 // Now we have our 32-bit value zero extended in the low element of
3378 // a vector. If Idx != 0, swizzle it into place.
3381 Item, DAG.getUNDEF(Item.getValueType()),
3382 getSwapEltZeroMask(VecElts, Idx, DAG, dl)
3384 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VecVT, Ops, 3);
3386 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3390 // If we have a constant or non-constant insertion into the low element of
3391 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3392 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3393 // depending on what the source datatype is. Because we can only get here
3394 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3396 // Don't do this for i64 values on x86-32.
3397 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3398 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3399 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3400 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3401 Subtarget->hasSSE2(), DAG);
3404 // Is it a vector logical left shift?
3405 if (NumElems == 2 && Idx == 1 &&
3406 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3407 unsigned NumBits = VT.getSizeInBits();
3408 return getVShift(true, VT,
3409 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3410 VT, Op.getOperand(1)),
3411 NumBits/2, DAG, *this, dl);
3414 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3417 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3418 // is a non-constant being inserted into an element other than the low one,
3419 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3420 // movd/movss) to move this into the low element, then shuffle it into
3422 if (EVTBits == 32) {
3423 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3425 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3426 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3427 Subtarget->hasSSE2(), DAG);
3428 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3429 MVT MaskEVT = MaskVT.getVectorElementType();
3430 SmallVector<SDValue, 8> MaskVec;
3431 for (unsigned i = 0; i < NumElems; i++)
3432 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3433 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3434 &MaskVec[0], MaskVec.size());
3435 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, Item,
3436 DAG.getUNDEF(VT), Mask);
3440 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3441 if (Values.size() == 1)
3444 // A vector full of immediates; various special cases are already
3445 // handled, so this is best done with a single constant-pool load.
3449 // Let legalizer expand 2-wide build_vectors.
3450 if (EVTBits == 64) {
3451 if (NumNonZero == 1) {
3452 // One half is zero or undef.
3453 unsigned Idx = CountTrailingZeros_32(NonZeros);
3454 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3455 Op.getOperand(Idx));
3456 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3457 Subtarget->hasSSE2(), DAG);
3462 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3463 if (EVTBits == 8 && NumElems == 16) {
3464 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3466 if (V.getNode()) return V;
3469 if (EVTBits == 16 && NumElems == 8) {
3470 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3472 if (V.getNode()) return V;
3475 // If element VT is == 32 bits, turn it into a number of shuffles.
3476 SmallVector<SDValue, 8> V;
3478 if (NumElems == 4 && NumZero > 0) {
3479 for (unsigned i = 0; i < 4; ++i) {
3480 bool isZero = !(NonZeros & (1 << i));
3482 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3484 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3487 for (unsigned i = 0; i < 2; ++i) {
3488 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3491 V[i] = V[i*2]; // Must be a zero vector.
3494 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2+1], V[i*2],
3495 getMOVLMask(NumElems, DAG, dl));
3498 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3499 getMOVLMask(NumElems, DAG, dl));
3502 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3503 getUnpacklMask(NumElems, DAG, dl));
3508 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3509 MVT EVT = MaskVT.getVectorElementType();
3510 SmallVector<SDValue, 8> MaskVec;
3511 bool Reverse = (NonZeros & 0x3) == 2;
3512 for (unsigned i = 0; i < 2; ++i)
3514 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3516 MaskVec.push_back(DAG.getConstant(i, EVT));
3517 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3518 for (unsigned i = 0; i < 2; ++i)
3520 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3522 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3523 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3524 &MaskVec[0], MaskVec.size());
3525 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[0], V[1], ShufMask);
3528 if (Values.size() > 2) {
3529 // Expand into a number of unpckl*.
3531 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3532 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3533 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3534 SDValue UnpckMask = getUnpacklMask(NumElems, DAG, dl);
3535 for (unsigned i = 0; i < NumElems; ++i)
3536 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3538 while (NumElems != 0) {
3539 for (unsigned i = 0; i < NumElems; ++i)
3540 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i], V[i + NumElems],
3550 // v8i16 shuffles - Prefer shuffles in the following order:
3551 // 1. [all] pshuflw, pshufhw, optional move
3552 // 2. [ssse3] 1 x pshufb
3553 // 3. [ssse3] 2 x pshufb + 1 x por
3554 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3556 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3557 SDValue PermMask, SelectionDAG &DAG,
3558 X86TargetLowering &TLI, DebugLoc dl) {
3559 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3560 PermMask.getNode()->op_end());
3561 SmallVector<int, 8> MaskVals;
3563 // Determine if more than 1 of the words in each of the low and high quadwords
3564 // of the result come from the same quadword of one of the two inputs. Undef
3565 // mask values count as coming from any quadword, for better codegen.
3566 SmallVector<unsigned, 4> LoQuad(4);
3567 SmallVector<unsigned, 4> HiQuad(4);
3568 BitVector InputQuads(4);
3569 for (unsigned i = 0; i < 8; ++i) {
3570 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3571 SDValue Elt = MaskElts[i];
3572 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3573 cast<ConstantSDNode>(Elt)->getZExtValue();
3574 MaskVals.push_back(EltIdx);
3583 InputQuads.set(EltIdx / 4);
3586 int BestLoQuad = -1;
3587 unsigned MaxQuad = 1;
3588 for (unsigned i = 0; i < 4; ++i) {
3589 if (LoQuad[i] > MaxQuad) {
3591 MaxQuad = LoQuad[i];
3595 int BestHiQuad = -1;
3597 for (unsigned i = 0; i < 4; ++i) {
3598 if (HiQuad[i] > MaxQuad) {
3600 MaxQuad = HiQuad[i];
3604 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3605 // of the two input vectors, shuffle them into one input vector so only a
3606 // single pshufb instruction is necessary. If There are more than 2 input
3607 // quads, disable the next transformation since it does not help SSSE3.
3608 bool V1Used = InputQuads[0] || InputQuads[1];
3609 bool V2Used = InputQuads[2] || InputQuads[3];
3610 if (TLI.getSubtarget()->hasSSSE3()) {
3611 if (InputQuads.count() == 2 && V1Used && V2Used) {
3612 BestLoQuad = InputQuads.find_first();
3613 BestHiQuad = InputQuads.find_next(BestLoQuad);
3615 if (InputQuads.count() > 2) {
3621 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3622 // the shuffle mask. If a quad is scored as -1, that means that it contains
3623 // words from all 4 input quadwords.
3625 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3626 SmallVector<SDValue,8> MaskV;
3627 MaskV.push_back(DAG.getConstant(BestLoQuad < 0 ? 0 : BestLoQuad, MVT::i64));
3628 MaskV.push_back(DAG.getConstant(BestHiQuad < 0 ? 1 : BestHiQuad, MVT::i64));
3629 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, &MaskV[0], 2);
3631 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
3632 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3633 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), Mask);
3634 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3636 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3637 // source words for the shuffle, to aid later transformations.
3638 bool AllWordsInNewV = true;
3639 bool InOrder[2] = { true, true };
3640 for (unsigned i = 0; i != 8; ++i) {
3641 int idx = MaskVals[i];
3643 InOrder[i/4] = false;
3644 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3646 AllWordsInNewV = false;
3650 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3651 if (AllWordsInNewV) {
3652 for (int i = 0; i != 8; ++i) {
3653 int idx = MaskVals[i];
3656 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3657 if ((idx != i) && idx < 4)
3659 if ((idx != i) && idx > 3)
3668 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3669 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3670 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3672 for (unsigned i = 0; i != 8; ++i)
3673 MaskV.push_back((MaskVals[i] < 0) ? DAG.getUNDEF(MVT::i16)
3674 : DAG.getConstant(MaskVals[i],
3676 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3677 DAG.getUNDEF(MVT::v8i16),
3678 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i16,
3683 // If we have SSSE3, and all words of the result are from 1 input vector,
3684 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3685 // is present, fall back to case 4.
3686 if (TLI.getSubtarget()->hasSSSE3()) {
3687 SmallVector<SDValue,16> pshufbMask;
3689 // If we have elements from both input vectors, set the high bit of the
3690 // shuffle mask element to zero out elements that come from V2 in the V1
3691 // mask, and elements that come from V1 in the V2 mask, so that the two
3692 // results can be OR'd together.
3693 bool TwoInputs = V1Used && V2Used;
3694 for (unsigned i = 0; i != 8; ++i) {
3695 int EltIdx = MaskVals[i] * 2;
3696 if (TwoInputs && (EltIdx >= 16)) {
3697 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3698 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3701 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3702 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3704 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3705 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3706 DAG.getNode(ISD::BUILD_VECTOR, dl,
3707 MVT::v16i8, &pshufbMask[0], 16));
3709 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3711 // Calculate the shuffle mask for the second input, shuffle it, and
3712 // OR it with the first shuffled input.
3714 for (unsigned i = 0; i != 8; ++i) {
3715 int EltIdx = MaskVals[i] * 2;
3717 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3718 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3721 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3722 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3724 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3725 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3726 DAG.getNode(ISD::BUILD_VECTOR, dl,
3727 MVT::v16i8, &pshufbMask[0], 16));
3728 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3729 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3732 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3733 // and update MaskVals with new element order.
3734 BitVector InOrder(8);
3735 if (BestLoQuad >= 0) {
3736 SmallVector<SDValue, 8> MaskV;
3737 for (int i = 0; i != 4; ++i) {
3738 int idx = MaskVals[i];
3740 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3742 } else if ((idx / 4) == BestLoQuad) {
3743 MaskV.push_back(DAG.getConstant(idx & 3, MVT::i16));
3746 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3749 for (unsigned i = 4; i != 8; ++i)
3750 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3751 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3752 DAG.getUNDEF(MVT::v8i16),
3753 DAG.getNode(ISD::BUILD_VECTOR, dl,
3754 MVT::v8i16, &MaskV[0], 8));
3757 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3758 // and update MaskVals with the new element order.
3759 if (BestHiQuad >= 0) {
3760 SmallVector<SDValue, 8> MaskV;
3761 for (unsigned i = 0; i != 4; ++i)
3762 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3763 for (unsigned i = 4; i != 8; ++i) {
3764 int idx = MaskVals[i];
3766 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3768 } else if ((idx / 4) == BestHiQuad) {
3769 MaskV.push_back(DAG.getConstant((idx & 3) + 4, MVT::i16));
3772 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3775 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3776 DAG.getUNDEF(MVT::v8i16),
3777 DAG.getNode(ISD::BUILD_VECTOR, dl,
3778 MVT::v8i16, &MaskV[0], 8));
3781 // In case BestHi & BestLo were both -1, which means each quadword has a word
3782 // from each of the four input quadwords, calculate the InOrder bitvector now
3783 // before falling through to the insert/extract cleanup.
3784 if (BestLoQuad == -1 && BestHiQuad == -1) {
3786 for (int i = 0; i != 8; ++i)
3787 if (MaskVals[i] < 0 || MaskVals[i] == i)
3791 // The other elements are put in the right place using pextrw and pinsrw.
3792 for (unsigned i = 0; i != 8; ++i) {
3795 int EltIdx = MaskVals[i];
3798 SDValue ExtOp = (EltIdx < 8)
3799 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3800 DAG.getIntPtrConstant(EltIdx))
3801 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3802 DAG.getIntPtrConstant(EltIdx - 8));
3803 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3804 DAG.getIntPtrConstant(i));
3809 // v16i8 shuffles - Prefer shuffles in the following order:
3810 // 1. [ssse3] 1 x pshufb
3811 // 2. [ssse3] 2 x pshufb + 1 x por
3812 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3814 SDValue LowerVECTOR_SHUFFLEv16i8(SDValue V1, SDValue V2,
3815 SDValue PermMask, SelectionDAG &DAG,
3816 X86TargetLowering &TLI, DebugLoc dl) {
3817 SmallVector<SDValue, 16> MaskElts(PermMask.getNode()->op_begin(),
3818 PermMask.getNode()->op_end());
3819 SmallVector<int, 16> MaskVals;
3821 // If we have SSSE3, case 1 is generated when all result bytes come from
3822 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3823 // present, fall back to case 3.
3824 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3827 for (unsigned i = 0; i < 16; ++i) {
3828 SDValue Elt = MaskElts[i];
3829 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3830 cast<ConstantSDNode>(Elt)->getZExtValue();
3831 MaskVals.push_back(EltIdx);
3840 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3841 if (TLI.getSubtarget()->hasSSSE3()) {
3842 SmallVector<SDValue,16> pshufbMask;
3844 // If all result elements are from one input vector, then only translate
3845 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3847 // Otherwise, we have elements from both input vectors, and must zero out
3848 // elements that come from V2 in the first mask, and V1 in the second mask
3849 // so that we can OR them together.
3850 bool TwoInputs = !(V1Only || V2Only);
3851 for (unsigned i = 0; i != 16; ++i) {
3852 int EltIdx = MaskVals[i];
3853 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3854 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3857 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3859 // If all the elements are from V2, assign it to V1 and return after
3860 // building the first pshufb.
3863 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3864 DAG.getNode(ISD::BUILD_VECTOR, dl,
3865 MVT::v16i8, &pshufbMask[0], 16));
3869 // Calculate the shuffle mask for the second input, shuffle it, and
3870 // OR it with the first shuffled input.
3872 for (unsigned i = 0; i != 16; ++i) {
3873 int EltIdx = MaskVals[i];
3875 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3878 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3880 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3881 DAG.getNode(ISD::BUILD_VECTOR, dl,
3882 MVT::v16i8, &pshufbMask[0], 16));
3883 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3886 // No SSSE3 - Calculate in place words and then fix all out of place words
3887 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3888 // the 16 different words that comprise the two doublequadword input vectors.
3889 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3890 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3891 SDValue NewV = V2Only ? V2 : V1;
3892 for (int i = 0; i != 8; ++i) {
3893 int Elt0 = MaskVals[i*2];
3894 int Elt1 = MaskVals[i*2+1];
3896 // This word of the result is all undef, skip it.
3897 if (Elt0 < 0 && Elt1 < 0)
3900 // This word of the result is already in the correct place, skip it.
3901 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3903 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3906 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3907 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3910 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3911 // using a single extract together, load it and store it.
3912 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3913 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3914 DAG.getIntPtrConstant(Elt1 / 2));
3915 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3916 DAG.getIntPtrConstant(i));
3920 // If Elt1 is defined, extract it from the appropriate source. If the
3921 // source byte is not also odd, shift the extracted word left 8 bits
3922 // otherwise clear the bottom 8 bits if we need to do an or.
3924 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3925 DAG.getIntPtrConstant(Elt1 / 2));
3926 if ((Elt1 & 1) == 0)
3927 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3928 DAG.getConstant(8, TLI.getShiftAmountTy()));
3930 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3931 DAG.getConstant(0xFF00, MVT::i16));
3933 // If Elt0 is defined, extract it from the appropriate source. If the
3934 // source byte is not also even, shift the extracted word right 8 bits. If
3935 // Elt1 was also defined, OR the extracted values together before
3936 // inserting them in the result.
3938 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3939 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3940 if ((Elt0 & 1) != 0)
3941 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3942 DAG.getConstant(8, TLI.getShiftAmountTy()));
3944 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3945 DAG.getConstant(0x00FF, MVT::i16));
3946 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3949 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3950 DAG.getIntPtrConstant(i));
3952 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3955 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3956 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3957 /// done when every pair / quad of shuffle mask elements point to elements in
3958 /// the right sequence. e.g.
3959 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3961 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3963 SDValue PermMask, SelectionDAG &DAG,
3964 TargetLowering &TLI, DebugLoc dl) {
3965 unsigned NumElems = PermMask.getNumOperands();
3966 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3967 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3968 MVT MaskEltVT = MaskVT.getVectorElementType();
3970 switch (VT.getSimpleVT()) {
3971 default: assert(false && "Unexpected!");
3972 case MVT::v4f32: NewVT = MVT::v2f64; break;
3973 case MVT::v4i32: NewVT = MVT::v2i64; break;
3974 case MVT::v8i16: NewVT = MVT::v4i32; break;
3975 case MVT::v16i8: NewVT = MVT::v4i32; break;
3978 if (NewWidth == 2) {
3984 unsigned Scale = NumElems / NewWidth;
3985 SmallVector<SDValue, 8> MaskVec;
3986 for (unsigned i = 0; i < NumElems; i += Scale) {
3987 unsigned StartIdx = ~0U;
3988 for (unsigned j = 0; j < Scale; ++j) {
3989 SDValue Elt = PermMask.getOperand(i+j);
3990 if (Elt.getOpcode() == ISD::UNDEF)
3992 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3993 if (StartIdx == ~0U)
3994 StartIdx = EltIdx - (EltIdx % Scale);
3995 if (EltIdx != StartIdx + j)
3998 if (StartIdx == ~0U)
3999 MaskVec.push_back(DAG.getUNDEF(MaskEltVT));
4001 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
4004 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4005 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4006 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, NewVT, V1, V2,
4007 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4008 &MaskVec[0], MaskVec.size()));
4011 /// getVZextMovL - Return a zero-extending vector move low node.
4013 static SDValue getVZextMovL(MVT VT, MVT OpVT,
4014 SDValue SrcOp, SelectionDAG &DAG,
4015 const X86Subtarget *Subtarget, DebugLoc dl) {
4016 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4017 LoadSDNode *LD = NULL;
4018 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4019 LD = dyn_cast<LoadSDNode>(SrcOp);
4021 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4023 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4024 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
4025 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4026 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4027 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
4029 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4030 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4031 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4032 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4040 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4041 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4042 DAG.getNode(ISD::BIT_CONVERT, dl,
4046 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4049 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
4050 SDValue PermMask, MVT VT, SelectionDAG &DAG,
4052 MVT MaskVT = PermMask.getValueType();
4053 MVT MaskEVT = MaskVT.getVectorElementType();
4054 SmallVector<std::pair<int, int>, 8> Locs;
4056 SmallVector<SDValue, 8> Mask1(4, DAG.getUNDEF(MaskEVT));
4059 for (unsigned i = 0; i != 4; ++i) {
4060 SDValue Elt = PermMask.getOperand(i);
4061 if (Elt.getOpcode() == ISD::UNDEF) {
4062 Locs[i] = std::make_pair(-1, -1);
4064 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4065 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
4067 Locs[i] = std::make_pair(0, NumLo);
4071 Locs[i] = std::make_pair(1, NumHi);
4073 Mask1[2+NumHi] = Elt;
4079 if (NumLo <= 2 && NumHi <= 2) {
4080 // If no more than two elements come from either vector. This can be
4081 // implemented with two shuffles. First shuffle gather the elements.
4082 // The second shuffle, which takes the first shuffle as both of its
4083 // vector operands, put the elements into the right order.
4084 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4085 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4086 &Mask1[0], Mask1.size()));
4088 SmallVector<SDValue, 8> Mask2(4, DAG.getUNDEF(MaskEVT));
4089 for (unsigned i = 0; i != 4; ++i) {
4090 if (Locs[i].first == -1)
4093 unsigned Idx = (i < 2) ? 0 : 4;
4094 Idx += Locs[i].first * 2 + Locs[i].second;
4095 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
4099 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1,
4100 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4101 &Mask2[0], Mask2.size()));
4102 } else if (NumLo == 3 || NumHi == 3) {
4103 // Otherwise, we must have three elements from one vector, call it X, and
4104 // one element from the other, call it Y. First, use a shufps to build an
4105 // intermediate vector with the one element from Y and the element from X
4106 // that will be in the same half in the final destination (the indexes don't
4107 // matter). Then, use a shufps to build the final vector, taking the half
4108 // containing the element from Y from the intermediate, and the other half
4111 // Normalize it so the 3 elements come from V1.
4112 PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
4116 // Find the element from V2.
4118 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4119 SDValue Elt = PermMask.getOperand(HiIndex);
4120 if (Elt.getOpcode() == ISD::UNDEF)
4122 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4127 Mask1[0] = PermMask.getOperand(HiIndex);
4128 Mask1[1] = DAG.getUNDEF(MaskEVT);
4129 Mask1[2] = PermMask.getOperand(HiIndex^1);
4130 Mask1[3] = DAG.getUNDEF(MaskEVT);
4131 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4132 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &Mask1[0], 4));
4135 Mask1[0] = PermMask.getOperand(0);
4136 Mask1[1] = PermMask.getOperand(1);
4137 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
4138 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
4139 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4140 DAG.getNode(ISD::BUILD_VECTOR, dl,
4141 MaskVT, &Mask1[0], 4));
4143 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
4144 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
4145 Mask1[2] = PermMask.getOperand(2);
4146 Mask1[3] = PermMask.getOperand(3);
4147 if (Mask1[2].getOpcode() != ISD::UNDEF)
4149 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
4151 if (Mask1[3].getOpcode() != ISD::UNDEF)
4153 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
4155 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V2, V1,
4156 DAG.getNode(ISD::BUILD_VECTOR, dl,
4157 MaskVT, &Mask1[0], 4));
4161 // Break it into (shuffle shuffle_hi, shuffle_lo).
4163 SmallVector<SDValue,8> LoMask(4, DAG.getUNDEF(MaskEVT));
4164 SmallVector<SDValue,8> HiMask(4, DAG.getUNDEF(MaskEVT));
4165 SmallVector<SDValue,8> *MaskPtr = &LoMask;
4166 unsigned MaskIdx = 0;
4169 for (unsigned i = 0; i != 4; ++i) {
4176 SDValue Elt = PermMask.getOperand(i);
4177 if (Elt.getOpcode() == ISD::UNDEF) {
4178 Locs[i] = std::make_pair(-1, -1);
4179 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
4180 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4181 (*MaskPtr)[LoIdx] = Elt;
4184 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4185 (*MaskPtr)[HiIdx] = Elt;
4190 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4191 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4192 &LoMask[0], LoMask.size()));
4193 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4194 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4195 &HiMask[0], HiMask.size()));
4196 SmallVector<SDValue, 8> MaskOps;
4197 for (unsigned i = 0; i != 4; ++i) {
4198 if (Locs[i].first == -1) {
4199 MaskOps.push_back(DAG.getUNDEF(MaskEVT));
4201 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4202 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
4205 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, LoShuffle, HiShuffle,
4206 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4207 &MaskOps[0], MaskOps.size()));
4211 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4212 SDValue V1 = Op.getOperand(0);
4213 SDValue V2 = Op.getOperand(1);
4214 SDValue PermMask = Op.getOperand(2);
4215 MVT VT = Op.getValueType();
4216 DebugLoc dl = Op.getDebugLoc();
4217 unsigned NumElems = PermMask.getNumOperands();
4218 bool isMMX = VT.getSizeInBits() == 64;
4219 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4220 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4221 bool V1IsSplat = false;
4222 bool V2IsSplat = false;
4224 // FIXME: Check for legal shuffle and return?
4226 if (isUndefShuffle(Op.getNode()))
4227 return DAG.getUNDEF(VT);
4229 if (isZeroShuffle(Op.getNode()))
4230 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4232 if (isIdentityMask(PermMask.getNode()))
4234 else if (isIdentityMask(PermMask.getNode(), true))
4237 // Canonicalize movddup shuffles.
4238 if (V2IsUndef && Subtarget->hasSSE2() &&
4239 VT.getSizeInBits() == 128 &&
4240 X86::isMOVDDUPMask(PermMask.getNode()))
4241 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4243 if (isSplatMask(PermMask.getNode())) {
4244 if (isMMX || NumElems < 4) return Op;
4245 // Promote it to a v4{if}32 splat.
4246 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4249 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4251 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4252 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG,
4254 if (NewOp.getNode())
4255 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4256 LowerVECTOR_SHUFFLE(NewOp, DAG));
4257 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4258 // FIXME: Figure out a cleaner way to do this.
4259 // Try to make use of movq to zero out the top part.
4260 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4261 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4263 if (NewOp.getNode()) {
4264 SDValue NewV1 = NewOp.getOperand(0);
4265 SDValue NewV2 = NewOp.getOperand(1);
4266 SDValue NewMask = NewOp.getOperand(2);
4267 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4268 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4269 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget,
4273 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4274 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4276 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4277 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4278 DAG, Subtarget, dl);
4282 // Check if this can be converted into a logical shift.
4283 bool isLeft = false;
4286 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4287 if (isShift && ShVal.hasOneUse()) {
4288 // If the shifted value has multiple uses, it may be cheaper to use
4289 // v_set0 + movlhps or movhlps, etc.
4290 MVT EVT = VT.getVectorElementType();
4291 ShAmt *= EVT.getSizeInBits();
4292 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4295 if (X86::isMOVLMask(PermMask.getNode())) {
4298 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4299 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4304 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4305 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4306 X86::isMOVHLPSMask(PermMask.getNode()) ||
4307 X86::isMOVHPMask(PermMask.getNode()) ||
4308 X86::isMOVLPMask(PermMask.getNode())))
4311 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4312 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4313 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4316 // No better options. Use a vshl / vsrl.
4317 MVT EVT = VT.getVectorElementType();
4318 ShAmt *= EVT.getSizeInBits();
4319 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4322 bool Commuted = false;
4323 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4324 // 1,1,1,1 -> v8i16 though.
4325 V1IsSplat = isSplatVector(V1.getNode());
4326 V2IsSplat = isSplatVector(V2.getNode());
4328 // Canonicalize the splat or undef, if present, to be on the RHS.
4329 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4330 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4331 std::swap(V1IsSplat, V2IsSplat);
4332 std::swap(V1IsUndef, V2IsUndef);
4336 // FIXME: Figure out a cleaner way to do this.
4337 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4338 if (V2IsUndef) return V1;
4339 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4341 // V2 is a splat, so the mask may be malformed. That is, it may point
4342 // to any V2 element. The instruction selectior won't like this. Get
4343 // a corrected mask and commute to form a proper MOVS{S|D}.
4344 SDValue NewMask = getMOVLMask(NumElems, DAG, dl);
4345 if (NewMask.getNode() != PermMask.getNode())
4346 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4351 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4352 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4353 X86::isUNPCKLMask(PermMask.getNode()) ||
4354 X86::isUNPCKHMask(PermMask.getNode()))
4358 // Normalize mask so all entries that point to V2 points to its first
4359 // element then try to match unpck{h|l} again. If match, return a
4360 // new vector_shuffle with the corrected mask.
4361 SDValue NewMask = NormalizeMask(PermMask, DAG);
4362 if (NewMask.getNode() != PermMask.getNode()) {
4363 if (X86::isUNPCKLMask(NewMask.getNode(), true)) {
4364 SDValue NewMask = getUnpacklMask(NumElems, DAG, dl);
4365 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4366 } else if (X86::isUNPCKHMask(NewMask.getNode(), true)) {
4367 SDValue NewMask = getUnpackhMask(NumElems, DAG, dl);
4368 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4373 // Normalize the node to match x86 shuffle ops if needed
4374 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4375 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4378 // Commute is back and try unpck* again.
4379 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4380 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4381 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4382 X86::isUNPCKLMask(PermMask.getNode()) ||
4383 X86::isUNPCKHMask(PermMask.getNode()))
4387 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4388 // Try PSHUF* first, then SHUFP*.
4389 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4390 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4391 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4392 if (V2.getOpcode() != ISD::UNDEF)
4393 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1,
4394 DAG.getUNDEF(VT), PermMask);
4399 if (Subtarget->hasSSE2() &&
4400 (X86::isPSHUFDMask(PermMask.getNode()) ||
4401 X86::isPSHUFHWMask(PermMask.getNode()) ||
4402 X86::isPSHUFLWMask(PermMask.getNode()))) {
4404 if (VT == MVT::v4f32) {
4406 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT,
4407 DAG.getNode(ISD::BIT_CONVERT, dl, RVT, V1),
4408 DAG.getUNDEF(RVT), PermMask);
4409 } else if (V2.getOpcode() != ISD::UNDEF)
4410 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT, V1,
4411 DAG.getUNDEF(RVT), PermMask);
4413 Op = DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
4417 // Binary or unary shufps.
4418 if (X86::isSHUFPMask(PermMask.getNode()) ||
4419 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4423 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4424 if (VT == MVT::v8i16) {
4425 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this, dl);
4426 if (NewOp.getNode())
4430 if (VT == MVT::v16i8) {
4431 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(V1, V2, PermMask, DAG, *this, dl);
4432 if (NewOp.getNode())
4436 // Handle all 4 wide cases with a number of shuffles except for MMX.
4437 if (NumElems == 4 && !isMMX)
4438 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG, dl);
4444 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4445 SelectionDAG &DAG) {
4446 MVT VT = Op.getValueType();
4447 DebugLoc dl = Op.getDebugLoc();
4448 if (VT.getSizeInBits() == 8) {
4449 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4450 Op.getOperand(0), Op.getOperand(1));
4451 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4452 DAG.getValueType(VT));
4453 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4454 } else if (VT.getSizeInBits() == 16) {
4455 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4456 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4458 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4459 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4460 DAG.getNode(ISD::BIT_CONVERT, dl,
4464 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4465 Op.getOperand(0), Op.getOperand(1));
4466 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4467 DAG.getValueType(VT));
4468 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4469 } else if (VT == MVT::f32) {
4470 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4471 // the result back to FR32 register. It's only worth matching if the
4472 // result has a single use which is a store or a bitcast to i32. And in
4473 // the case of a store, it's not worth it if the index is a constant 0,
4474 // because a MOVSSmr can be used instead, which is smaller and faster.
4475 if (!Op.hasOneUse())
4477 SDNode *User = *Op.getNode()->use_begin();
4478 if ((User->getOpcode() != ISD::STORE ||
4479 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4480 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4481 (User->getOpcode() != ISD::BIT_CONVERT ||
4482 User->getValueType(0) != MVT::i32))
4484 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4485 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4488 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4489 } else if (VT == MVT::i32) {
4490 // ExtractPS works with constant index.
4491 if (isa<ConstantSDNode>(Op.getOperand(1)))
4499 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4500 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4503 if (Subtarget->hasSSE41()) {
4504 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4509 MVT VT = Op.getValueType();
4510 DebugLoc dl = Op.getDebugLoc();
4511 // TODO: handle v16i8.
4512 if (VT.getSizeInBits() == 16) {
4513 SDValue Vec = Op.getOperand(0);
4514 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4516 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4517 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4518 DAG.getNode(ISD::BIT_CONVERT, dl,
4521 // Transform it so it match pextrw which produces a 32-bit result.
4522 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4523 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4524 Op.getOperand(0), Op.getOperand(1));
4525 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4526 DAG.getValueType(VT));
4527 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4528 } else if (VT.getSizeInBits() == 32) {
4529 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4532 // SHUFPS the element to the lowest double word, then movss.
4533 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4534 SmallVector<SDValue, 8> IdxVec;
4536 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4538 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4540 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4542 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4543 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4544 &IdxVec[0], IdxVec.size());
4545 SDValue Vec = Op.getOperand(0);
4546 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4547 Vec, DAG.getUNDEF(Vec.getValueType()), Mask);
4548 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4549 DAG.getIntPtrConstant(0));
4550 } else if (VT.getSizeInBits() == 64) {
4551 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4552 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4553 // to match extract_elt for f64.
4554 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4558 // UNPCKHPD the element to the lowest double word, then movsd.
4559 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4560 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4561 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4562 SmallVector<SDValue, 8> IdxVec;
4563 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4565 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4566 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4567 &IdxVec[0], IdxVec.size());
4568 SDValue Vec = Op.getOperand(0);
4569 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4570 Vec, DAG.getUNDEF(Vec.getValueType()),
4572 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4573 DAG.getIntPtrConstant(0));
4580 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4581 MVT VT = Op.getValueType();
4582 MVT EVT = VT.getVectorElementType();
4583 DebugLoc dl = Op.getDebugLoc();
4585 SDValue N0 = Op.getOperand(0);
4586 SDValue N1 = Op.getOperand(1);
4587 SDValue N2 = Op.getOperand(2);
4589 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4590 isa<ConstantSDNode>(N2)) {
4591 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4593 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4595 if (N1.getValueType() != MVT::i32)
4596 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4597 if (N2.getValueType() != MVT::i32)
4598 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4599 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4600 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4601 // Bits [7:6] of the constant are the source select. This will always be
4602 // zero here. The DAG Combiner may combine an extract_elt index into these
4603 // bits. For example (insert (extract, 3), 2) could be matched by putting
4604 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4605 // Bits [5:4] of the constant are the destination select. This is the
4606 // value of the incoming immediate.
4607 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4608 // combine either bitwise AND or insert of float 0.0 to set these bits.
4609 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4610 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4611 } else if (EVT == MVT::i32) {
4612 // InsertPS works with constant index.
4613 if (isa<ConstantSDNode>(N2))
4620 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4621 MVT VT = Op.getValueType();
4622 MVT EVT = VT.getVectorElementType();
4624 if (Subtarget->hasSSE41())
4625 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4630 DebugLoc dl = Op.getDebugLoc();
4631 SDValue N0 = Op.getOperand(0);
4632 SDValue N1 = Op.getOperand(1);
4633 SDValue N2 = Op.getOperand(2);
4635 if (EVT.getSizeInBits() == 16) {
4636 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4637 // as its second argument.
4638 if (N1.getValueType() != MVT::i32)
4639 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4640 if (N2.getValueType() != MVT::i32)
4641 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4642 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4648 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4649 DebugLoc dl = Op.getDebugLoc();
4650 if (Op.getValueType() == MVT::v2f32)
4651 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4652 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4653 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4654 Op.getOperand(0))));
4656 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4657 MVT VT = MVT::v2i32;
4658 switch (Op.getValueType().getSimpleVT()) {
4665 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4666 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4669 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4670 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4671 // one of the above mentioned nodes. It has to be wrapped because otherwise
4672 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4673 // be used to form addressing mode. These wrapped nodes will be selected
4676 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4677 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4678 // FIXME there isn't really any debug info here, should come from the parent
4679 DebugLoc dl = CP->getDebugLoc();
4680 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4681 CP->getAlignment());
4682 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4683 // With PIC, the address is actually $g + Offset.
4684 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4685 !Subtarget->isPICStyleRIPRel()) {
4686 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4687 DAG.getNode(X86ISD::GlobalBaseReg,
4688 DebugLoc::getUnknownLoc(),
4697 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4699 SelectionDAG &DAG) const {
4700 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4701 bool ExtraLoadRequired =
4702 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4704 // Create the TargetGlobalAddress node, folding in the constant
4705 // offset if it is legal.
4707 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4708 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4711 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4712 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4714 // With PIC, the address is actually $g + Offset.
4715 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4716 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4717 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4721 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4722 // load the value at address GV, not the value of GV itself. This means that
4723 // the GlobalAddress must be in the base or index register of the address, not
4724 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4725 // The same applies for external symbols during PIC codegen
4726 if (ExtraLoadRequired)
4727 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4728 PseudoSourceValue::getGOT(), 0);
4730 // If there was a non-zero offset that we didn't fold, create an explicit
4733 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4734 DAG.getConstant(Offset, getPointerTy()));
4740 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4741 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4742 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4743 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4747 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4749 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4750 DebugLoc dl = GA->getDebugLoc();
4751 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4752 GA->getValueType(0),
4755 SDValue Ops[] = { Chain, TGA, *InFlag };
4756 return DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4758 SDValue Ops[] = { Chain, TGA };
4759 return DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4763 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4765 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4768 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4769 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4770 DAG.getNode(X86ISD::GlobalBaseReg,
4771 DebugLoc::getUnknownLoc(),
4773 InFlag = Chain.getValue(1);
4775 Chain = GetTLSADDR(DAG, Chain, GA, &InFlag);
4776 InFlag = Chain.getValue(1);
4778 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4779 SDValue Ops1[] = { Chain,
4780 DAG.getTargetExternalSymbol("___tls_get_addr",
4782 DAG.getRegister(X86::EAX, PtrVT),
4783 DAG.getRegister(X86::EBX, PtrVT),
4785 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
4786 InFlag = Chain.getValue(1);
4788 return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
4791 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4793 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4795 SDValue InFlag, Chain;
4796 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4798 Chain = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL);
4799 InFlag = Chain.getValue(1);
4801 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4802 SDValue Ops1[] = { Chain,
4803 DAG.getTargetExternalSymbol("__tls_get_addr",
4805 DAG.getRegister(X86::RDI, PtrVT),
4807 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
4808 InFlag = Chain.getValue(1);
4810 return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
4813 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4814 // "local exec" model.
4815 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4816 const MVT PtrVT, TLSModel::Model model,
4818 DebugLoc dl = GA->getDebugLoc();
4819 // Get the Thread Pointer
4820 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4821 DebugLoc::getUnknownLoc(), PtrVT,
4822 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4825 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4828 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4830 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4831 GA->getValueType(0),
4833 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4835 if (model == TLSModel::InitialExec)
4836 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4837 PseudoSourceValue::getGOT(), 0);
4839 // The address of the thread local variable is the add of the thread
4840 // pointer with the offset of the variable.
4841 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4845 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4846 // TODO: implement the "local dynamic" model
4847 // TODO: implement the "initial exec"model for pic executables
4848 assert(Subtarget->isTargetELF() &&
4849 "TLS not implemented for non-ELF targets");
4850 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4851 GlobalValue *GV = GA->getGlobal();
4852 TLSModel::Model model =
4853 getTLSModel (GV, getTargetMachine().getRelocationModel());
4854 if (Subtarget->is64Bit()) {
4856 case TLSModel::GeneralDynamic:
4857 case TLSModel::LocalDynamic: // not implemented
4858 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4860 case TLSModel::InitialExec:
4861 case TLSModel::LocalExec:
4862 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, true);
4866 case TLSModel::GeneralDynamic:
4867 case TLSModel::LocalDynamic: // not implemented
4868 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4870 case TLSModel::InitialExec:
4871 case TLSModel::LocalExec:
4872 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, false);
4875 assert(0 && "Unreachable");
4880 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4881 // FIXME there isn't really any debug info here
4882 DebugLoc dl = Op.getDebugLoc();
4883 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4884 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4885 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4886 // With PIC, the address is actually $g + Offset.
4887 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4888 !Subtarget->isPICStyleRIPRel()) {
4889 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4890 DAG.getNode(X86ISD::GlobalBaseReg,
4891 DebugLoc::getUnknownLoc(),
4899 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4900 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4901 // FIXME there isn't really any debug into here
4902 DebugLoc dl = JT->getDebugLoc();
4903 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4904 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4905 // With PIC, the address is actually $g + Offset.
4906 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4907 !Subtarget->isPICStyleRIPRel()) {
4908 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4909 DAG.getNode(X86ISD::GlobalBaseReg,
4910 DebugLoc::getUnknownLoc(),
4918 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4919 /// take a 2 x i32 value to shift plus a shift amount.
4920 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4921 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4922 MVT VT = Op.getValueType();
4923 unsigned VTBits = VT.getSizeInBits();
4924 DebugLoc dl = Op.getDebugLoc();
4925 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4926 SDValue ShOpLo = Op.getOperand(0);
4927 SDValue ShOpHi = Op.getOperand(1);
4928 SDValue ShAmt = Op.getOperand(2);
4929 SDValue Tmp1 = isSRA ?
4930 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4931 DAG.getConstant(VTBits - 1, MVT::i8)) :
4932 DAG.getConstant(0, VT);
4935 if (Op.getOpcode() == ISD::SHL_PARTS) {
4936 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4937 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4939 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4940 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4943 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4944 DAG.getConstant(VTBits, MVT::i8));
4945 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4946 AndNode, DAG.getConstant(0, MVT::i8));
4949 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4950 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4951 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4953 if (Op.getOpcode() == ISD::SHL_PARTS) {
4954 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4955 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4957 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4958 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4961 SDValue Ops[2] = { Lo, Hi };
4962 return DAG.getMergeValues(Ops, 2, dl);
4965 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4966 MVT SrcVT = Op.getOperand(0).getValueType();
4967 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4968 "Unknown SINT_TO_FP to lower!");
4970 // These are really Legal; caller falls through into that case.
4971 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4973 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4974 Subtarget->is64Bit())
4977 DebugLoc dl = Op.getDebugLoc();
4978 unsigned Size = SrcVT.getSizeInBits()/8;
4979 MachineFunction &MF = DAG.getMachineFunction();
4980 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4981 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4982 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4984 PseudoSourceValue::getFixedStack(SSFI), 0);
4988 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4990 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4992 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4993 SmallVector<SDValue, 8> Ops;
4994 Ops.push_back(Chain);
4995 Ops.push_back(StackSlot);
4996 Ops.push_back(DAG.getValueType(SrcVT));
4997 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4998 Tys, &Ops[0], Ops.size());
5001 Chain = Result.getValue(1);
5002 SDValue InFlag = Result.getValue(2);
5004 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5005 // shouldn't be necessary except that RFP cannot be live across
5006 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5007 MachineFunction &MF = DAG.getMachineFunction();
5008 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
5009 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5010 Tys = DAG.getVTList(MVT::Other);
5011 SmallVector<SDValue, 8> Ops;
5012 Ops.push_back(Chain);
5013 Ops.push_back(Result);
5014 Ops.push_back(StackSlot);
5015 Ops.push_back(DAG.getValueType(Op.getValueType()));
5016 Ops.push_back(InFlag);
5017 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5018 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5019 PseudoSourceValue::getFixedStack(SSFI), 0);
5025 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5026 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5027 // This algorithm is not obvious. Here it is in C code, more or less:
5029 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5030 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5031 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5033 // Copy ints to xmm registers.
5034 __m128i xh = _mm_cvtsi32_si128( hi );
5035 __m128i xl = _mm_cvtsi32_si128( lo );
5037 // Combine into low half of a single xmm register.
5038 __m128i x = _mm_unpacklo_epi32( xh, xl );
5042 // Merge in appropriate exponents to give the integer bits the right
5044 x = _mm_unpacklo_epi32( x, exp );
5046 // Subtract away the biases to deal with the IEEE-754 double precision
5048 d = _mm_sub_pd( (__m128d) x, bias );
5050 // All conversions up to here are exact. The correctly rounded result is
5051 // calculated using the current rounding mode using the following
5053 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5054 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5055 // store doesn't really need to be here (except
5056 // maybe to zero the other double)
5061 DebugLoc dl = Op.getDebugLoc();
5063 // Build some magic constants.
5064 std::vector<Constant*> CV0;
5065 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
5066 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
5067 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5068 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5069 Constant *C0 = ConstantVector::get(CV0);
5070 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5072 std::vector<Constant*> CV1;
5073 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
5074 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
5075 Constant *C1 = ConstantVector::get(CV1);
5076 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5078 SmallVector<SDValue, 4> MaskVec;
5079 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
5080 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
5081 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
5082 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
5083 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
5084 &MaskVec[0], MaskVec.size());
5085 SmallVector<SDValue, 4> MaskVec2;
5086 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
5087 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
5088 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32,
5089 &MaskVec2[0], MaskVec2.size());
5091 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5092 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5094 DAG.getIntPtrConstant(1)));
5095 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5096 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5098 DAG.getIntPtrConstant(0)));
5099 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5100 XR1, XR2, UnpcklMask);
5101 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5102 PseudoSourceValue::getConstantPool(), 0,
5104 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5105 Unpck1, CLod0, UnpcklMask);
5106 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5107 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5108 PseudoSourceValue::getConstantPool(), 0,
5110 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5112 // Add the halves; easiest way is to swap them into another reg first.
5113 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2f64,
5114 Sub, Sub, ShufMask);
5115 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5116 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5117 DAG.getIntPtrConstant(0));
5120 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5121 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5122 DebugLoc dl = Op.getDebugLoc();
5123 // FP constant to bias correct the final result.
5124 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5127 // Load the 32-bit value into an XMM register.
5128 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5129 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5131 DAG.getIntPtrConstant(0)));
5133 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5134 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5135 DAG.getIntPtrConstant(0));
5137 // Or the load with the bias.
5138 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5139 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5140 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5142 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5143 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5144 MVT::v2f64, Bias)));
5145 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5146 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5147 DAG.getIntPtrConstant(0));
5149 // Subtract the bias.
5150 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5152 // Handle final rounding.
5153 MVT DestVT = Op.getValueType();
5155 if (DestVT.bitsLT(MVT::f64)) {
5156 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5157 DAG.getIntPtrConstant(0));
5158 } else if (DestVT.bitsGT(MVT::f64)) {
5159 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5162 // Handle final rounding.
5166 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5167 SDValue N0 = Op.getOperand(0);
5168 DebugLoc dl = Op.getDebugLoc();
5170 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5171 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5172 // the optimization here.
5173 if (DAG.SignBitIsZero(N0))
5174 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5176 MVT SrcVT = N0.getValueType();
5177 if (SrcVT == MVT::i64) {
5178 // We only handle SSE2 f64 target here; caller can handle the rest.
5179 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5182 return LowerUINT_TO_FP_i64(Op, DAG);
5183 } else if (SrcVT == MVT::i32) {
5184 return LowerUINT_TO_FP_i32(Op, DAG);
5187 assert(0 && "Unknown UINT_TO_FP to lower!");
5191 std::pair<SDValue,SDValue> X86TargetLowering::
5192 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
5193 DebugLoc dl = Op.getDebugLoc();
5194 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
5195 Op.getValueType().getSimpleVT() >= MVT::i16 &&
5196 "Unknown FP_TO_SINT to lower!");
5198 // These are really Legal.
5199 if (Op.getValueType() == MVT::i32 &&
5200 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5201 return std::make_pair(SDValue(), SDValue());
5202 if (Subtarget->is64Bit() &&
5203 Op.getValueType() == MVT::i64 &&
5204 Op.getOperand(0).getValueType() != MVT::f80)
5205 return std::make_pair(SDValue(), SDValue());
5207 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5209 MachineFunction &MF = DAG.getMachineFunction();
5210 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
5211 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5212 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5214 switch (Op.getValueType().getSimpleVT()) {
5215 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5216 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5217 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5218 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5221 SDValue Chain = DAG.getEntryNode();
5222 SDValue Value = Op.getOperand(0);
5223 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5224 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5225 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5226 PseudoSourceValue::getFixedStack(SSFI), 0);
5227 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5229 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5231 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5232 Chain = Value.getValue(1);
5233 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5234 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5237 // Build the FP_TO_INT*_IN_MEM
5238 SDValue Ops[] = { Chain, Value, StackSlot };
5239 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5241 return std::make_pair(FIST, StackSlot);
5244 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5245 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
5246 SDValue FIST = Vals.first, StackSlot = Vals.second;
5247 if (FIST.getNode() == 0) return SDValue();
5250 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5251 FIST, StackSlot, NULL, 0);
5254 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5255 DebugLoc dl = Op.getDebugLoc();
5256 MVT VT = Op.getValueType();
5259 EltVT = VT.getVectorElementType();
5260 std::vector<Constant*> CV;
5261 if (EltVT == MVT::f64) {
5262 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5266 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5272 Constant *C = ConstantVector::get(CV);
5273 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5274 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5275 PseudoSourceValue::getConstantPool(), 0,
5277 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5280 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5281 DebugLoc dl = Op.getDebugLoc();
5282 MVT VT = Op.getValueType();
5284 unsigned EltNum = 1;
5285 if (VT.isVector()) {
5286 EltVT = VT.getVectorElementType();
5287 EltNum = VT.getVectorNumElements();
5289 std::vector<Constant*> CV;
5290 if (EltVT == MVT::f64) {
5291 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5295 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5301 Constant *C = ConstantVector::get(CV);
5302 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5303 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5304 PseudoSourceValue::getConstantPool(), 0,
5306 if (VT.isVector()) {
5307 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5308 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5309 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5311 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5313 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5317 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5318 SDValue Op0 = Op.getOperand(0);
5319 SDValue Op1 = Op.getOperand(1);
5320 DebugLoc dl = Op.getDebugLoc();
5321 MVT VT = Op.getValueType();
5322 MVT SrcVT = Op1.getValueType();
5324 // If second operand is smaller, extend it first.
5325 if (SrcVT.bitsLT(VT)) {
5326 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5329 // And if it is bigger, shrink it first.
5330 if (SrcVT.bitsGT(VT)) {
5331 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5335 // At this point the operands and the result should have the same
5336 // type, and that won't be f80 since that is not custom lowered.
5338 // First get the sign bit of second operand.
5339 std::vector<Constant*> CV;
5340 if (SrcVT == MVT::f64) {
5341 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5342 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5344 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5345 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5346 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5347 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5349 Constant *C = ConstantVector::get(CV);
5350 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5351 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5352 PseudoSourceValue::getConstantPool(), 0,
5354 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5356 // Shift sign bit right or left if the two operands have different types.
5357 if (SrcVT.bitsGT(VT)) {
5358 // Op0 is MVT::f32, Op1 is MVT::f64.
5359 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5360 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5361 DAG.getConstant(32, MVT::i32));
5362 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5363 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5364 DAG.getIntPtrConstant(0));
5367 // Clear first operand sign bit.
5369 if (VT == MVT::f64) {
5370 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5371 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5373 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5374 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5375 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5376 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5378 C = ConstantVector::get(CV);
5379 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5380 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5381 PseudoSourceValue::getConstantPool(), 0,
5383 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5385 // Or the value with the sign bit.
5386 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5389 /// Emit nodes that will be selected as "test Op0,Op0", or something
5391 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5392 SelectionDAG &DAG) {
5393 DebugLoc dl = Op.getDebugLoc();
5395 // CF and OF aren't always set the way we want. Determine which
5396 // of these we need.
5397 bool NeedCF = false;
5398 bool NeedOF = false;
5400 case X86::COND_A: case X86::COND_AE:
5401 case X86::COND_B: case X86::COND_BE:
5404 case X86::COND_G: case X86::COND_GE:
5405 case X86::COND_L: case X86::COND_LE:
5406 case X86::COND_O: case X86::COND_NO:
5412 // See if we can use the EFLAGS value from the operand instead of
5413 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5414 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5415 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5416 unsigned Opcode = 0;
5417 unsigned NumOperands = 0;
5418 switch (Op.getNode()->getOpcode()) {
5420 // Due to an isel shortcoming, be conservative if this add is likely to
5421 // be selected as part of a load-modify-store instruction. When the root
5422 // node in a match is a store, isel doesn't know how to remap non-chain
5423 // non-flag uses of other nodes in the match, such as the ADD in this
5424 // case. This leads to the ADD being left around and reselected, with
5425 // the result being two adds in the output.
5426 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5427 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5428 if (UI->getOpcode() == ISD::STORE)
5430 if (ConstantSDNode *C =
5431 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5432 // An add of one will be selected as an INC.
5433 if (C->getAPIntValue() == 1) {
5434 Opcode = X86ISD::INC;
5438 // An add of negative one (subtract of one) will be selected as a DEC.
5439 if (C->getAPIntValue().isAllOnesValue()) {
5440 Opcode = X86ISD::DEC;
5445 // Otherwise use a regular EFLAGS-setting add.
5446 Opcode = X86ISD::ADD;
5450 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5451 // likely to be selected as part of a load-modify-store instruction.
5452 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5453 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5454 if (UI->getOpcode() == ISD::STORE)
5456 // Otherwise use a regular EFLAGS-setting sub.
5457 Opcode = X86ISD::SUB;
5464 return SDValue(Op.getNode(), 1);
5470 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5471 SmallVector<SDValue, 4> Ops;
5472 for (unsigned i = 0; i != NumOperands; ++i)
5473 Ops.push_back(Op.getOperand(i));
5474 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5475 DAG.ReplaceAllUsesWith(Op, New);
5476 return SDValue(New.getNode(), 1);
5480 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5481 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5482 DAG.getConstant(0, Op.getValueType()));
5485 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5487 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5488 SelectionDAG &DAG) {
5489 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5490 if (C->getAPIntValue() == 0)
5491 return EmitTest(Op0, X86CC, DAG);
5493 DebugLoc dl = Op0.getDebugLoc();
5494 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5497 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5498 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5499 SDValue Op0 = Op.getOperand(0);
5500 SDValue Op1 = Op.getOperand(1);
5501 DebugLoc dl = Op.getDebugLoc();
5502 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5504 // Lower (X & (1 << N)) == 0 to BT(X, N).
5505 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5506 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5507 if (Op0.getOpcode() == ISD::AND &&
5509 Op1.getOpcode() == ISD::Constant &&
5510 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5511 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5513 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5514 if (ConstantSDNode *Op010C =
5515 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5516 if (Op010C->getZExtValue() == 1) {
5517 LHS = Op0.getOperand(0);
5518 RHS = Op0.getOperand(1).getOperand(1);
5520 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5521 if (ConstantSDNode *Op000C =
5522 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5523 if (Op000C->getZExtValue() == 1) {
5524 LHS = Op0.getOperand(1);
5525 RHS = Op0.getOperand(0).getOperand(1);
5527 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5528 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5529 SDValue AndLHS = Op0.getOperand(0);
5530 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5531 LHS = AndLHS.getOperand(0);
5532 RHS = AndLHS.getOperand(1);
5536 if (LHS.getNode()) {
5537 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5538 // instruction. Since the shift amount is in-range-or-undefined, we know
5539 // that doing a bittest on the i16 value is ok. We extend to i32 because
5540 // the encoding for the i16 version is larger than the i32 version.
5541 if (LHS.getValueType() == MVT::i8)
5542 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5544 // If the operand types disagree, extend the shift amount to match. Since
5545 // BT ignores high bits (like shifts) we can use anyextend.
5546 if (LHS.getValueType() != RHS.getValueType())
5547 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5549 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5550 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5551 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5552 DAG.getConstant(Cond, MVT::i8), BT);
5556 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5557 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5559 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5560 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5561 DAG.getConstant(X86CC, MVT::i8), Cond);
5564 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5566 SDValue Op0 = Op.getOperand(0);
5567 SDValue Op1 = Op.getOperand(1);
5568 SDValue CC = Op.getOperand(2);
5569 MVT VT = Op.getValueType();
5570 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5571 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5572 DebugLoc dl = Op.getDebugLoc();
5576 MVT VT0 = Op0.getValueType();
5577 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5578 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5581 switch (SetCCOpcode) {
5584 case ISD::SETEQ: SSECC = 0; break;
5586 case ISD::SETGT: Swap = true; // Fallthrough
5588 case ISD::SETOLT: SSECC = 1; break;
5590 case ISD::SETGE: Swap = true; // Fallthrough
5592 case ISD::SETOLE: SSECC = 2; break;
5593 case ISD::SETUO: SSECC = 3; break;
5595 case ISD::SETNE: SSECC = 4; break;
5596 case ISD::SETULE: Swap = true;
5597 case ISD::SETUGE: SSECC = 5; break;
5598 case ISD::SETULT: Swap = true;
5599 case ISD::SETUGT: SSECC = 6; break;
5600 case ISD::SETO: SSECC = 7; break;
5603 std::swap(Op0, Op1);
5605 // In the two special cases we can't handle, emit two comparisons.
5607 if (SetCCOpcode == ISD::SETUEQ) {
5609 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5610 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5611 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5613 else if (SetCCOpcode == ISD::SETONE) {
5615 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5616 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5617 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5619 assert(0 && "Illegal FP comparison");
5621 // Handle all other FP comparisons here.
5622 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5625 // We are handling one of the integer comparisons here. Since SSE only has
5626 // GT and EQ comparisons for integer, swapping operands and multiple
5627 // operations may be required for some comparisons.
5628 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5629 bool Swap = false, Invert = false, FlipSigns = false;
5631 switch (VT.getSimpleVT()) {
5633 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5634 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5635 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5636 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5639 switch (SetCCOpcode) {
5641 case ISD::SETNE: Invert = true;
5642 case ISD::SETEQ: Opc = EQOpc; break;
5643 case ISD::SETLT: Swap = true;
5644 case ISD::SETGT: Opc = GTOpc; break;
5645 case ISD::SETGE: Swap = true;
5646 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5647 case ISD::SETULT: Swap = true;
5648 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5649 case ISD::SETUGE: Swap = true;
5650 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5653 std::swap(Op0, Op1);
5655 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5656 // bits of the inputs before performing those operations.
5658 MVT EltVT = VT.getVectorElementType();
5659 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5661 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5662 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5664 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5665 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5668 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5670 // If the logical-not of the result is required, perform that now.
5672 Result = DAG.getNOT(dl, Result, VT);
5677 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5678 static bool isX86LogicalCmp(SDValue Op) {
5679 unsigned Opc = Op.getNode()->getOpcode();
5680 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5682 if (Op.getResNo() == 1 &&
5683 (Opc == X86ISD::ADD ||
5684 Opc == X86ISD::SUB ||
5685 Opc == X86ISD::SMUL ||
5686 Opc == X86ISD::UMUL ||
5687 Opc == X86ISD::INC ||
5688 Opc == X86ISD::DEC))
5694 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5695 bool addTest = true;
5696 SDValue Cond = Op.getOperand(0);
5697 DebugLoc dl = Op.getDebugLoc();
5700 if (Cond.getOpcode() == ISD::SETCC)
5701 Cond = LowerSETCC(Cond, DAG);
5703 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5704 // setting operand in place of the X86ISD::SETCC.
5705 if (Cond.getOpcode() == X86ISD::SETCC) {
5706 CC = Cond.getOperand(0);
5708 SDValue Cmp = Cond.getOperand(1);
5709 unsigned Opc = Cmp.getOpcode();
5710 MVT VT = Op.getValueType();
5712 bool IllegalFPCMov = false;
5713 if (VT.isFloatingPoint() && !VT.isVector() &&
5714 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5715 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5717 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5718 Opc == X86ISD::BT) { // FIXME
5725 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5726 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5729 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5730 SmallVector<SDValue, 4> Ops;
5731 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5732 // condition is true.
5733 Ops.push_back(Op.getOperand(2));
5734 Ops.push_back(Op.getOperand(1));
5736 Ops.push_back(Cond);
5737 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5740 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5741 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5742 // from the AND / OR.
5743 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5744 Opc = Op.getOpcode();
5745 if (Opc != ISD::OR && Opc != ISD::AND)
5747 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5748 Op.getOperand(0).hasOneUse() &&
5749 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5750 Op.getOperand(1).hasOneUse());
5753 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5754 // 1 and that the SETCC node has a single use.
5755 static bool isXor1OfSetCC(SDValue Op) {
5756 if (Op.getOpcode() != ISD::XOR)
5758 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5759 if (N1C && N1C->getAPIntValue() == 1) {
5760 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5761 Op.getOperand(0).hasOneUse();
5766 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5767 bool addTest = true;
5768 SDValue Chain = Op.getOperand(0);
5769 SDValue Cond = Op.getOperand(1);
5770 SDValue Dest = Op.getOperand(2);
5771 DebugLoc dl = Op.getDebugLoc();
5774 if (Cond.getOpcode() == ISD::SETCC)
5775 Cond = LowerSETCC(Cond, DAG);
5777 // FIXME: LowerXALUO doesn't handle these!!
5778 else if (Cond.getOpcode() == X86ISD::ADD ||
5779 Cond.getOpcode() == X86ISD::SUB ||
5780 Cond.getOpcode() == X86ISD::SMUL ||
5781 Cond.getOpcode() == X86ISD::UMUL)
5782 Cond = LowerXALUO(Cond, DAG);
5785 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5786 // setting operand in place of the X86ISD::SETCC.
5787 if (Cond.getOpcode() == X86ISD::SETCC) {
5788 CC = Cond.getOperand(0);
5790 SDValue Cmp = Cond.getOperand(1);
5791 unsigned Opc = Cmp.getOpcode();
5792 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5793 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5797 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5801 // These can only come from an arithmetic instruction with overflow,
5802 // e.g. SADDO, UADDO.
5803 Cond = Cond.getNode()->getOperand(1);
5810 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5811 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5812 if (CondOpc == ISD::OR) {
5813 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5814 // two branches instead of an explicit OR instruction with a
5816 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5817 isX86LogicalCmp(Cmp)) {
5818 CC = Cond.getOperand(0).getOperand(0);
5819 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5820 Chain, Dest, CC, Cmp);
5821 CC = Cond.getOperand(1).getOperand(0);
5825 } else { // ISD::AND
5826 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5827 // two branches instead of an explicit AND instruction with a
5828 // separate test. However, we only do this if this block doesn't
5829 // have a fall-through edge, because this requires an explicit
5830 // jmp when the condition is false.
5831 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5832 isX86LogicalCmp(Cmp) &&
5833 Op.getNode()->hasOneUse()) {
5834 X86::CondCode CCode =
5835 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5836 CCode = X86::GetOppositeBranchCondition(CCode);
5837 CC = DAG.getConstant(CCode, MVT::i8);
5838 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5839 // Look for an unconditional branch following this conditional branch.
5840 // We need this because we need to reverse the successors in order
5841 // to implement FCMP_OEQ.
5842 if (User.getOpcode() == ISD::BR) {
5843 SDValue FalseBB = User.getOperand(1);
5845 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5846 assert(NewBR == User);
5849 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5850 Chain, Dest, CC, Cmp);
5851 X86::CondCode CCode =
5852 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5853 CCode = X86::GetOppositeBranchCondition(CCode);
5854 CC = DAG.getConstant(CCode, MVT::i8);
5860 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5861 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5862 // It should be transformed during dag combiner except when the condition
5863 // is set by a arithmetics with overflow node.
5864 X86::CondCode CCode =
5865 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5866 CCode = X86::GetOppositeBranchCondition(CCode);
5867 CC = DAG.getConstant(CCode, MVT::i8);
5868 Cond = Cond.getOperand(0).getOperand(1);
5874 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5875 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5877 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5878 Chain, Dest, CC, Cond);
5882 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5883 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5884 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5885 // that the guard pages used by the OS virtual memory manager are allocated in
5886 // correct sequence.
5888 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5889 SelectionDAG &DAG) {
5890 assert(Subtarget->isTargetCygMing() &&
5891 "This should be used only on Cygwin/Mingw targets");
5892 DebugLoc dl = Op.getDebugLoc();
5895 SDValue Chain = Op.getOperand(0);
5896 SDValue Size = Op.getOperand(1);
5897 // FIXME: Ensure alignment here
5901 MVT IntPtr = getPointerTy();
5902 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5904 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5906 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5907 Flag = Chain.getValue(1);
5909 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5910 SDValue Ops[] = { Chain,
5911 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5912 DAG.getRegister(X86::EAX, IntPtr),
5913 DAG.getRegister(X86StackPtr, SPTy),
5915 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5916 Flag = Chain.getValue(1);
5918 Chain = DAG.getCALLSEQ_END(Chain,
5919 DAG.getIntPtrConstant(0, true),
5920 DAG.getIntPtrConstant(0, true),
5923 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5925 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5926 return DAG.getMergeValues(Ops1, 2, dl);
5930 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5932 SDValue Dst, SDValue Src,
5933 SDValue Size, unsigned Align,
5935 uint64_t DstSVOff) {
5936 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5938 // If not DWORD aligned or size is more than the threshold, call the library.
5939 // The libc version is likely to be faster for these cases. It can use the
5940 // address value and run time information about the CPU.
5941 if ((Align & 3) != 0 ||
5943 ConstantSize->getZExtValue() >
5944 getSubtarget()->getMaxInlineSizeThreshold()) {
5945 SDValue InFlag(0, 0);
5947 // Check to see if there is a specialized entry-point for memory zeroing.
5948 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5950 if (const char *bzeroEntry = V &&
5951 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5952 MVT IntPtr = getPointerTy();
5953 const Type *IntPtrTy = TD->getIntPtrType();
5954 TargetLowering::ArgListTy Args;
5955 TargetLowering::ArgListEntry Entry;
5957 Entry.Ty = IntPtrTy;
5958 Args.push_back(Entry);
5960 Args.push_back(Entry);
5961 std::pair<SDValue,SDValue> CallResult =
5962 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5963 CallingConv::C, false,
5964 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5965 return CallResult.second;
5968 // Otherwise have the target-independent code call memset.
5972 uint64_t SizeVal = ConstantSize->getZExtValue();
5973 SDValue InFlag(0, 0);
5976 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5977 unsigned BytesLeft = 0;
5978 bool TwoRepStos = false;
5981 uint64_t Val = ValC->getZExtValue() & 255;
5983 // If the value is a constant, then we can potentially use larger sets.
5984 switch (Align & 3) {
5985 case 2: // WORD aligned
5988 Val = (Val << 8) | Val;
5990 case 0: // DWORD aligned
5993 Val = (Val << 8) | Val;
5994 Val = (Val << 16) | Val;
5995 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5998 Val = (Val << 32) | Val;
6001 default: // Byte aligned
6004 Count = DAG.getIntPtrConstant(SizeVal);
6008 if (AVT.bitsGT(MVT::i8)) {
6009 unsigned UBytes = AVT.getSizeInBits() / 8;
6010 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6011 BytesLeft = SizeVal % UBytes;
6014 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6016 InFlag = Chain.getValue(1);
6019 Count = DAG.getIntPtrConstant(SizeVal);
6020 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6021 InFlag = Chain.getValue(1);
6024 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6027 InFlag = Chain.getValue(1);
6028 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6031 InFlag = Chain.getValue(1);
6033 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6034 SmallVector<SDValue, 8> Ops;
6035 Ops.push_back(Chain);
6036 Ops.push_back(DAG.getValueType(AVT));
6037 Ops.push_back(InFlag);
6038 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6041 InFlag = Chain.getValue(1);
6043 MVT CVT = Count.getValueType();
6044 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6045 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6046 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6049 InFlag = Chain.getValue(1);
6050 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6052 Ops.push_back(Chain);
6053 Ops.push_back(DAG.getValueType(MVT::i8));
6054 Ops.push_back(InFlag);
6055 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6056 } else if (BytesLeft) {
6057 // Handle the last 1 - 7 bytes.
6058 unsigned Offset = SizeVal - BytesLeft;
6059 MVT AddrVT = Dst.getValueType();
6060 MVT SizeVT = Size.getValueType();
6062 Chain = DAG.getMemset(Chain, dl,
6063 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6064 DAG.getConstant(Offset, AddrVT)),
6066 DAG.getConstant(BytesLeft, SizeVT),
6067 Align, DstSV, DstSVOff + Offset);
6070 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6075 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6076 SDValue Chain, SDValue Dst, SDValue Src,
6077 SDValue Size, unsigned Align,
6079 const Value *DstSV, uint64_t DstSVOff,
6080 const Value *SrcSV, uint64_t SrcSVOff) {
6081 // This requires the copy size to be a constant, preferrably
6082 // within a subtarget-specific limit.
6083 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6086 uint64_t SizeVal = ConstantSize->getZExtValue();
6087 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6090 /// If not DWORD aligned, call the library.
6091 if ((Align & 3) != 0)
6096 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6099 unsigned UBytes = AVT.getSizeInBits() / 8;
6100 unsigned CountVal = SizeVal / UBytes;
6101 SDValue Count = DAG.getIntPtrConstant(CountVal);
6102 unsigned BytesLeft = SizeVal % UBytes;
6104 SDValue InFlag(0, 0);
6105 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6108 InFlag = Chain.getValue(1);
6109 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6112 InFlag = Chain.getValue(1);
6113 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6116 InFlag = Chain.getValue(1);
6118 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6119 SmallVector<SDValue, 8> Ops;
6120 Ops.push_back(Chain);
6121 Ops.push_back(DAG.getValueType(AVT));
6122 Ops.push_back(InFlag);
6123 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6125 SmallVector<SDValue, 4> Results;
6126 Results.push_back(RepMovs);
6128 // Handle the last 1 - 7 bytes.
6129 unsigned Offset = SizeVal - BytesLeft;
6130 MVT DstVT = Dst.getValueType();
6131 MVT SrcVT = Src.getValueType();
6132 MVT SizeVT = Size.getValueType();
6133 Results.push_back(DAG.getMemcpy(Chain, dl,
6134 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6135 DAG.getConstant(Offset, DstVT)),
6136 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6137 DAG.getConstant(Offset, SrcVT)),
6138 DAG.getConstant(BytesLeft, SizeVT),
6139 Align, AlwaysInline,
6140 DstSV, DstSVOff + Offset,
6141 SrcSV, SrcSVOff + Offset));
6144 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6145 &Results[0], Results.size());
6148 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6149 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6150 DebugLoc dl = Op.getDebugLoc();
6152 if (!Subtarget->is64Bit()) {
6153 // vastart just stores the address of the VarArgsFrameIndex slot into the
6154 // memory location argument.
6155 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6156 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6160 // gp_offset (0 - 6 * 8)
6161 // fp_offset (48 - 48 + 8 * 16)
6162 // overflow_arg_area (point to parameters coming in memory).
6164 SmallVector<SDValue, 8> MemOps;
6165 SDValue FIN = Op.getOperand(1);
6167 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6168 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6170 MemOps.push_back(Store);
6173 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6174 FIN, DAG.getIntPtrConstant(4));
6175 Store = DAG.getStore(Op.getOperand(0), dl,
6176 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6178 MemOps.push_back(Store);
6180 // Store ptr to overflow_arg_area
6181 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6182 FIN, DAG.getIntPtrConstant(4));
6183 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6184 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6185 MemOps.push_back(Store);
6187 // Store ptr to reg_save_area.
6188 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6189 FIN, DAG.getIntPtrConstant(8));
6190 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6191 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6192 MemOps.push_back(Store);
6193 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6194 &MemOps[0], MemOps.size());
6197 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6198 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6199 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6200 SDValue Chain = Op.getOperand(0);
6201 SDValue SrcPtr = Op.getOperand(1);
6202 SDValue SrcSV = Op.getOperand(2);
6204 assert(0 && "VAArgInst is not yet implemented for x86-64!");
6209 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6210 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6211 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6212 SDValue Chain = Op.getOperand(0);
6213 SDValue DstPtr = Op.getOperand(1);
6214 SDValue SrcPtr = Op.getOperand(2);
6215 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6216 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6217 DebugLoc dl = Op.getDebugLoc();
6219 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6220 DAG.getIntPtrConstant(24), 8, false,
6221 DstSV, 0, SrcSV, 0);
6225 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6226 DebugLoc dl = Op.getDebugLoc();
6227 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6229 default: return SDValue(); // Don't custom lower most intrinsics.
6230 // Comparison intrinsics.
6231 case Intrinsic::x86_sse_comieq_ss:
6232 case Intrinsic::x86_sse_comilt_ss:
6233 case Intrinsic::x86_sse_comile_ss:
6234 case Intrinsic::x86_sse_comigt_ss:
6235 case Intrinsic::x86_sse_comige_ss:
6236 case Intrinsic::x86_sse_comineq_ss:
6237 case Intrinsic::x86_sse_ucomieq_ss:
6238 case Intrinsic::x86_sse_ucomilt_ss:
6239 case Intrinsic::x86_sse_ucomile_ss:
6240 case Intrinsic::x86_sse_ucomigt_ss:
6241 case Intrinsic::x86_sse_ucomige_ss:
6242 case Intrinsic::x86_sse_ucomineq_ss:
6243 case Intrinsic::x86_sse2_comieq_sd:
6244 case Intrinsic::x86_sse2_comilt_sd:
6245 case Intrinsic::x86_sse2_comile_sd:
6246 case Intrinsic::x86_sse2_comigt_sd:
6247 case Intrinsic::x86_sse2_comige_sd:
6248 case Intrinsic::x86_sse2_comineq_sd:
6249 case Intrinsic::x86_sse2_ucomieq_sd:
6250 case Intrinsic::x86_sse2_ucomilt_sd:
6251 case Intrinsic::x86_sse2_ucomile_sd:
6252 case Intrinsic::x86_sse2_ucomigt_sd:
6253 case Intrinsic::x86_sse2_ucomige_sd:
6254 case Intrinsic::x86_sse2_ucomineq_sd: {
6256 ISD::CondCode CC = ISD::SETCC_INVALID;
6259 case Intrinsic::x86_sse_comieq_ss:
6260 case Intrinsic::x86_sse2_comieq_sd:
6264 case Intrinsic::x86_sse_comilt_ss:
6265 case Intrinsic::x86_sse2_comilt_sd:
6269 case Intrinsic::x86_sse_comile_ss:
6270 case Intrinsic::x86_sse2_comile_sd:
6274 case Intrinsic::x86_sse_comigt_ss:
6275 case Intrinsic::x86_sse2_comigt_sd:
6279 case Intrinsic::x86_sse_comige_ss:
6280 case Intrinsic::x86_sse2_comige_sd:
6284 case Intrinsic::x86_sse_comineq_ss:
6285 case Intrinsic::x86_sse2_comineq_sd:
6289 case Intrinsic::x86_sse_ucomieq_ss:
6290 case Intrinsic::x86_sse2_ucomieq_sd:
6291 Opc = X86ISD::UCOMI;
6294 case Intrinsic::x86_sse_ucomilt_ss:
6295 case Intrinsic::x86_sse2_ucomilt_sd:
6296 Opc = X86ISD::UCOMI;
6299 case Intrinsic::x86_sse_ucomile_ss:
6300 case Intrinsic::x86_sse2_ucomile_sd:
6301 Opc = X86ISD::UCOMI;
6304 case Intrinsic::x86_sse_ucomigt_ss:
6305 case Intrinsic::x86_sse2_ucomigt_sd:
6306 Opc = X86ISD::UCOMI;
6309 case Intrinsic::x86_sse_ucomige_ss:
6310 case Intrinsic::x86_sse2_ucomige_sd:
6311 Opc = X86ISD::UCOMI;
6314 case Intrinsic::x86_sse_ucomineq_ss:
6315 case Intrinsic::x86_sse2_ucomineq_sd:
6316 Opc = X86ISD::UCOMI;
6321 SDValue LHS = Op.getOperand(1);
6322 SDValue RHS = Op.getOperand(2);
6323 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6324 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6325 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6326 DAG.getConstant(X86CC, MVT::i8), Cond);
6327 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6330 // Fix vector shift instructions where the last operand is a non-immediate
6332 case Intrinsic::x86_sse2_pslli_w:
6333 case Intrinsic::x86_sse2_pslli_d:
6334 case Intrinsic::x86_sse2_pslli_q:
6335 case Intrinsic::x86_sse2_psrli_w:
6336 case Intrinsic::x86_sse2_psrli_d:
6337 case Intrinsic::x86_sse2_psrli_q:
6338 case Intrinsic::x86_sse2_psrai_w:
6339 case Intrinsic::x86_sse2_psrai_d:
6340 case Intrinsic::x86_mmx_pslli_w:
6341 case Intrinsic::x86_mmx_pslli_d:
6342 case Intrinsic::x86_mmx_pslli_q:
6343 case Intrinsic::x86_mmx_psrli_w:
6344 case Intrinsic::x86_mmx_psrli_d:
6345 case Intrinsic::x86_mmx_psrli_q:
6346 case Intrinsic::x86_mmx_psrai_w:
6347 case Intrinsic::x86_mmx_psrai_d: {
6348 SDValue ShAmt = Op.getOperand(2);
6349 if (isa<ConstantSDNode>(ShAmt))
6352 unsigned NewIntNo = 0;
6353 MVT ShAmtVT = MVT::v4i32;
6355 case Intrinsic::x86_sse2_pslli_w:
6356 NewIntNo = Intrinsic::x86_sse2_psll_w;
6358 case Intrinsic::x86_sse2_pslli_d:
6359 NewIntNo = Intrinsic::x86_sse2_psll_d;
6361 case Intrinsic::x86_sse2_pslli_q:
6362 NewIntNo = Intrinsic::x86_sse2_psll_q;
6364 case Intrinsic::x86_sse2_psrli_w:
6365 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6367 case Intrinsic::x86_sse2_psrli_d:
6368 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6370 case Intrinsic::x86_sse2_psrli_q:
6371 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6373 case Intrinsic::x86_sse2_psrai_w:
6374 NewIntNo = Intrinsic::x86_sse2_psra_w;
6376 case Intrinsic::x86_sse2_psrai_d:
6377 NewIntNo = Intrinsic::x86_sse2_psra_d;
6380 ShAmtVT = MVT::v2i32;
6382 case Intrinsic::x86_mmx_pslli_w:
6383 NewIntNo = Intrinsic::x86_mmx_psll_w;
6385 case Intrinsic::x86_mmx_pslli_d:
6386 NewIntNo = Intrinsic::x86_mmx_psll_d;
6388 case Intrinsic::x86_mmx_pslli_q:
6389 NewIntNo = Intrinsic::x86_mmx_psll_q;
6391 case Intrinsic::x86_mmx_psrli_w:
6392 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6394 case Intrinsic::x86_mmx_psrli_d:
6395 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6397 case Intrinsic::x86_mmx_psrli_q:
6398 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6400 case Intrinsic::x86_mmx_psrai_w:
6401 NewIntNo = Intrinsic::x86_mmx_psra_w;
6403 case Intrinsic::x86_mmx_psrai_d:
6404 NewIntNo = Intrinsic::x86_mmx_psra_d;
6406 default: abort(); // Can't reach here.
6411 MVT VT = Op.getValueType();
6412 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6413 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6414 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6415 DAG.getConstant(NewIntNo, MVT::i32),
6416 Op.getOperand(1), ShAmt);
6421 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6422 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6423 DebugLoc dl = Op.getDebugLoc();
6426 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6428 DAG.getConstant(TD->getPointerSize(),
6429 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6430 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6431 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6436 // Just load the return address.
6437 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6438 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6439 RetAddrFI, NULL, 0);
6442 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6443 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6444 MFI->setFrameAddressIsTaken(true);
6445 MVT VT = Op.getValueType();
6446 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6447 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6448 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6449 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6451 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6455 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6456 SelectionDAG &DAG) {
6457 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6460 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6462 MachineFunction &MF = DAG.getMachineFunction();
6463 SDValue Chain = Op.getOperand(0);
6464 SDValue Offset = Op.getOperand(1);
6465 SDValue Handler = Op.getOperand(2);
6466 DebugLoc dl = Op.getDebugLoc();
6468 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6470 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6472 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6473 DAG.getIntPtrConstant(-TD->getPointerSize()));
6474 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6475 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6476 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6477 MF.getRegInfo().addLiveOut(StoreAddrReg);
6479 return DAG.getNode(X86ISD::EH_RETURN, dl,
6481 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6484 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6485 SelectionDAG &DAG) {
6486 SDValue Root = Op.getOperand(0);
6487 SDValue Trmp = Op.getOperand(1); // trampoline
6488 SDValue FPtr = Op.getOperand(2); // nested function
6489 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6490 DebugLoc dl = Op.getDebugLoc();
6492 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6494 const X86InstrInfo *TII =
6495 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6497 if (Subtarget->is64Bit()) {
6498 SDValue OutChains[6];
6500 // Large code-model.
6502 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6503 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6505 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6506 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6508 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6510 // Load the pointer to the nested function into R11.
6511 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6512 SDValue Addr = Trmp;
6513 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6516 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6517 DAG.getConstant(2, MVT::i64));
6518 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6520 // Load the 'nest' parameter value into R10.
6521 // R10 is specified in X86CallingConv.td
6522 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6523 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6524 DAG.getConstant(10, MVT::i64));
6525 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6526 Addr, TrmpAddr, 10);
6528 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6529 DAG.getConstant(12, MVT::i64));
6530 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6532 // Jump to the nested function.
6533 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6534 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6535 DAG.getConstant(20, MVT::i64));
6536 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6537 Addr, TrmpAddr, 20);
6539 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6540 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6541 DAG.getConstant(22, MVT::i64));
6542 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6546 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6547 return DAG.getMergeValues(Ops, 2, dl);
6549 const Function *Func =
6550 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6551 unsigned CC = Func->getCallingConv();
6556 assert(0 && "Unsupported calling convention");
6557 case CallingConv::C:
6558 case CallingConv::X86_StdCall: {
6559 // Pass 'nest' parameter in ECX.
6560 // Must be kept in sync with X86CallingConv.td
6563 // Check that ECX wasn't needed by an 'inreg' parameter.
6564 const FunctionType *FTy = Func->getFunctionType();
6565 const AttrListPtr &Attrs = Func->getAttributes();
6567 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6568 unsigned InRegCount = 0;
6571 for (FunctionType::param_iterator I = FTy->param_begin(),
6572 E = FTy->param_end(); I != E; ++I, ++Idx)
6573 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6574 // FIXME: should only count parameters that are lowered to integers.
6575 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6577 if (InRegCount > 2) {
6578 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6584 case CallingConv::X86_FastCall:
6585 case CallingConv::Fast:
6586 // Pass 'nest' parameter in EAX.
6587 // Must be kept in sync with X86CallingConv.td
6592 SDValue OutChains[4];
6595 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6596 DAG.getConstant(10, MVT::i32));
6597 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6599 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6600 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6601 OutChains[0] = DAG.getStore(Root, dl,
6602 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6605 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6606 DAG.getConstant(1, MVT::i32));
6607 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6609 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6610 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6611 DAG.getConstant(5, MVT::i32));
6612 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6613 TrmpAddr, 5, false, 1);
6615 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6616 DAG.getConstant(6, MVT::i32));
6617 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6620 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6621 return DAG.getMergeValues(Ops, 2, dl);
6625 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6627 The rounding mode is in bits 11:10 of FPSR, and has the following
6634 FLT_ROUNDS, on the other hand, expects the following:
6641 To perform the conversion, we do:
6642 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6645 MachineFunction &MF = DAG.getMachineFunction();
6646 const TargetMachine &TM = MF.getTarget();
6647 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6648 unsigned StackAlignment = TFI.getStackAlignment();
6649 MVT VT = Op.getValueType();
6650 DebugLoc dl = Op.getDebugLoc();
6652 // Save FP Control Word to stack slot
6653 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6654 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6656 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6657 DAG.getEntryNode(), StackSlot);
6659 // Load FP Control Word from stack slot
6660 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6662 // Transform as necessary
6664 DAG.getNode(ISD::SRL, dl, MVT::i16,
6665 DAG.getNode(ISD::AND, dl, MVT::i16,
6666 CWD, DAG.getConstant(0x800, MVT::i16)),
6667 DAG.getConstant(11, MVT::i8));
6669 DAG.getNode(ISD::SRL, dl, MVT::i16,
6670 DAG.getNode(ISD::AND, dl, MVT::i16,
6671 CWD, DAG.getConstant(0x400, MVT::i16)),
6672 DAG.getConstant(9, MVT::i8));
6675 DAG.getNode(ISD::AND, dl, MVT::i16,
6676 DAG.getNode(ISD::ADD, dl, MVT::i16,
6677 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6678 DAG.getConstant(1, MVT::i16)),
6679 DAG.getConstant(3, MVT::i16));
6682 return DAG.getNode((VT.getSizeInBits() < 16 ?
6683 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6686 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6687 MVT VT = Op.getValueType();
6689 unsigned NumBits = VT.getSizeInBits();
6690 DebugLoc dl = Op.getDebugLoc();
6692 Op = Op.getOperand(0);
6693 if (VT == MVT::i8) {
6694 // Zero extend to i32 since there is not an i8 bsr.
6696 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6699 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6700 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6701 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6703 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6704 SmallVector<SDValue, 4> Ops;
6706 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6707 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6708 Ops.push_back(Op.getValue(1));
6709 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6711 // Finally xor with NumBits-1.
6712 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6715 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6719 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6720 MVT VT = Op.getValueType();
6722 unsigned NumBits = VT.getSizeInBits();
6723 DebugLoc dl = Op.getDebugLoc();
6725 Op = Op.getOperand(0);
6726 if (VT == MVT::i8) {
6728 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6731 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6732 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6733 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6735 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6736 SmallVector<SDValue, 4> Ops;
6738 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6739 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6740 Ops.push_back(Op.getValue(1));
6741 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6744 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6748 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6749 MVT VT = Op.getValueType();
6750 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6751 DebugLoc dl = Op.getDebugLoc();
6753 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6754 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6755 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6756 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6757 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6759 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6760 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6761 // return AloBlo + AloBhi + AhiBlo;
6763 SDValue A = Op.getOperand(0);
6764 SDValue B = Op.getOperand(1);
6766 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6767 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6768 A, DAG.getConstant(32, MVT::i32));
6769 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6770 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6771 B, DAG.getConstant(32, MVT::i32));
6772 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6773 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6775 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6776 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6778 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6779 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6781 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6782 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6783 AloBhi, DAG.getConstant(32, MVT::i32));
6784 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6785 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6786 AhiBlo, DAG.getConstant(32, MVT::i32));
6787 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6788 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6793 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6794 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6795 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6796 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6797 // has only one use.
6798 SDNode *N = Op.getNode();
6799 SDValue LHS = N->getOperand(0);
6800 SDValue RHS = N->getOperand(1);
6801 unsigned BaseOp = 0;
6803 DebugLoc dl = Op.getDebugLoc();
6805 switch (Op.getOpcode()) {
6806 default: assert(0 && "Unknown ovf instruction!");
6808 // A subtract of one will be selected as a INC. Note that INC doesn't
6809 // set CF, so we can't do this for UADDO.
6810 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6811 if (C->getAPIntValue() == 1) {
6812 BaseOp = X86ISD::INC;
6816 BaseOp = X86ISD::ADD;
6820 BaseOp = X86ISD::ADD;
6824 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6825 // set CF, so we can't do this for USUBO.
6826 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6827 if (C->getAPIntValue() == 1) {
6828 BaseOp = X86ISD::DEC;
6832 BaseOp = X86ISD::SUB;
6836 BaseOp = X86ISD::SUB;
6840 BaseOp = X86ISD::SMUL;
6844 BaseOp = X86ISD::UMUL;
6849 // Also sets EFLAGS.
6850 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6851 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6854 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6855 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6857 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6861 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6862 MVT T = Op.getValueType();
6863 DebugLoc dl = Op.getDebugLoc();
6866 switch(T.getSimpleVT()) {
6868 assert(false && "Invalid value type!");
6869 case MVT::i8: Reg = X86::AL; size = 1; break;
6870 case MVT::i16: Reg = X86::AX; size = 2; break;
6871 case MVT::i32: Reg = X86::EAX; size = 4; break;
6873 assert(Subtarget->is64Bit() && "Node not type legal!");
6874 Reg = X86::RAX; size = 8;
6877 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6878 Op.getOperand(2), SDValue());
6879 SDValue Ops[] = { cpIn.getValue(0),
6882 DAG.getTargetConstant(size, MVT::i8),
6884 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6885 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6887 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6891 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6892 SelectionDAG &DAG) {
6893 assert(Subtarget->is64Bit() && "Result not type legalized?");
6894 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6895 SDValue TheChain = Op.getOperand(0);
6896 DebugLoc dl = Op.getDebugLoc();
6897 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6898 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6899 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6901 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6902 DAG.getConstant(32, MVT::i8));
6904 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6907 return DAG.getMergeValues(Ops, 2, dl);
6910 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6911 SDNode *Node = Op.getNode();
6912 DebugLoc dl = Node->getDebugLoc();
6913 MVT T = Node->getValueType(0);
6914 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6915 DAG.getConstant(0, T), Node->getOperand(2));
6916 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6917 cast<AtomicSDNode>(Node)->getMemoryVT(),
6918 Node->getOperand(0),
6919 Node->getOperand(1), negOp,
6920 cast<AtomicSDNode>(Node)->getSrcValue(),
6921 cast<AtomicSDNode>(Node)->getAlignment());
6924 /// LowerOperation - Provide custom lowering hooks for some operations.
6926 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6927 switch (Op.getOpcode()) {
6928 default: assert(0 && "Should not custom lower this!");
6929 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6930 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6931 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6932 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6933 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6934 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6935 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6936 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6937 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6938 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6939 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6940 case ISD::SHL_PARTS:
6941 case ISD::SRA_PARTS:
6942 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6943 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6944 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6945 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6946 case ISD::FABS: return LowerFABS(Op, DAG);
6947 case ISD::FNEG: return LowerFNEG(Op, DAG);
6948 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6949 case ISD::SETCC: return LowerSETCC(Op, DAG);
6950 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6951 case ISD::SELECT: return LowerSELECT(Op, DAG);
6952 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6953 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6954 case ISD::CALL: return LowerCALL(Op, DAG);
6955 case ISD::RET: return LowerRET(Op, DAG);
6956 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6957 case ISD::VASTART: return LowerVASTART(Op, DAG);
6958 case ISD::VAARG: return LowerVAARG(Op, DAG);
6959 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6960 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6961 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6962 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6963 case ISD::FRAME_TO_ARGS_OFFSET:
6964 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6965 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6966 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6967 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6968 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6969 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6970 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6971 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6977 case ISD::UMULO: return LowerXALUO(Op, DAG);
6978 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6982 void X86TargetLowering::
6983 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6984 SelectionDAG &DAG, unsigned NewOp) {
6985 MVT T = Node->getValueType(0);
6986 DebugLoc dl = Node->getDebugLoc();
6987 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6989 SDValue Chain = Node->getOperand(0);
6990 SDValue In1 = Node->getOperand(1);
6991 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6992 Node->getOperand(2), DAG.getIntPtrConstant(0));
6993 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6994 Node->getOperand(2), DAG.getIntPtrConstant(1));
6995 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6996 // have a MemOperand. Pass the info through as a normal operand.
6997 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6998 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6999 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7000 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
7001 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7002 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7003 Results.push_back(Result.getValue(2));
7006 /// ReplaceNodeResults - Replace a node with an illegal result type
7007 /// with a new node built out of custom code.
7008 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7009 SmallVectorImpl<SDValue>&Results,
7010 SelectionDAG &DAG) {
7011 DebugLoc dl = N->getDebugLoc();
7012 switch (N->getOpcode()) {
7014 assert(false && "Do not know how to custom type legalize this operation!");
7016 case ISD::FP_TO_SINT: {
7017 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
7018 SDValue FIST = Vals.first, StackSlot = Vals.second;
7019 if (FIST.getNode() != 0) {
7020 MVT VT = N->getValueType(0);
7021 // Return a load from the stack slot.
7022 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7026 case ISD::READCYCLECOUNTER: {
7027 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7028 SDValue TheChain = N->getOperand(0);
7029 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7030 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7032 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7034 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7035 SDValue Ops[] = { eax, edx };
7036 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7037 Results.push_back(edx.getValue(1));
7040 case ISD::ATOMIC_CMP_SWAP: {
7041 MVT T = N->getValueType(0);
7042 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7043 SDValue cpInL, cpInH;
7044 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7045 DAG.getConstant(0, MVT::i32));
7046 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7047 DAG.getConstant(1, MVT::i32));
7048 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7049 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7051 SDValue swapInL, swapInH;
7052 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7053 DAG.getConstant(0, MVT::i32));
7054 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7055 DAG.getConstant(1, MVT::i32));
7056 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7058 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7059 swapInL.getValue(1));
7060 SDValue Ops[] = { swapInH.getValue(0),
7062 swapInH.getValue(1) };
7063 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7064 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7065 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7066 MVT::i32, Result.getValue(1));
7067 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7068 MVT::i32, cpOutL.getValue(2));
7069 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7070 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7071 Results.push_back(cpOutH.getValue(1));
7074 case ISD::ATOMIC_LOAD_ADD:
7075 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7077 case ISD::ATOMIC_LOAD_AND:
7078 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7080 case ISD::ATOMIC_LOAD_NAND:
7081 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7083 case ISD::ATOMIC_LOAD_OR:
7084 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7086 case ISD::ATOMIC_LOAD_SUB:
7087 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7089 case ISD::ATOMIC_LOAD_XOR:
7090 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7092 case ISD::ATOMIC_SWAP:
7093 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7098 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7100 default: return NULL;
7101 case X86ISD::BSF: return "X86ISD::BSF";
7102 case X86ISD::BSR: return "X86ISD::BSR";
7103 case X86ISD::SHLD: return "X86ISD::SHLD";
7104 case X86ISD::SHRD: return "X86ISD::SHRD";
7105 case X86ISD::FAND: return "X86ISD::FAND";
7106 case X86ISD::FOR: return "X86ISD::FOR";
7107 case X86ISD::FXOR: return "X86ISD::FXOR";
7108 case X86ISD::FSRL: return "X86ISD::FSRL";
7109 case X86ISD::FILD: return "X86ISD::FILD";
7110 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7111 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7112 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7113 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7114 case X86ISD::FLD: return "X86ISD::FLD";
7115 case X86ISD::FST: return "X86ISD::FST";
7116 case X86ISD::CALL: return "X86ISD::CALL";
7117 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
7118 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7119 case X86ISD::BT: return "X86ISD::BT";
7120 case X86ISD::CMP: return "X86ISD::CMP";
7121 case X86ISD::COMI: return "X86ISD::COMI";
7122 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7123 case X86ISD::SETCC: return "X86ISD::SETCC";
7124 case X86ISD::CMOV: return "X86ISD::CMOV";
7125 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7126 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7127 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7128 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7129 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7130 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7131 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7132 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7133 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7134 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7135 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7136 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7137 case X86ISD::FMAX: return "X86ISD::FMAX";
7138 case X86ISD::FMIN: return "X86ISD::FMIN";
7139 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7140 case X86ISD::FRCP: return "X86ISD::FRCP";
7141 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7142 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7143 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7144 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7145 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7146 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7147 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7148 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7149 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7150 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7151 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7152 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7153 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7154 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7155 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7156 case X86ISD::VSHL: return "X86ISD::VSHL";
7157 case X86ISD::VSRL: return "X86ISD::VSRL";
7158 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7159 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7160 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7161 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7162 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7163 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7164 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7165 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7166 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7167 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7168 case X86ISD::ADD: return "X86ISD::ADD";
7169 case X86ISD::SUB: return "X86ISD::SUB";
7170 case X86ISD::SMUL: return "X86ISD::SMUL";
7171 case X86ISD::UMUL: return "X86ISD::UMUL";
7172 case X86ISD::INC: return "X86ISD::INC";
7173 case X86ISD::DEC: return "X86ISD::DEC";
7174 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7178 // isLegalAddressingMode - Return true if the addressing mode represented
7179 // by AM is legal for this target, for a load/store of the specified type.
7180 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7181 const Type *Ty) const {
7182 // X86 supports extremely general addressing modes.
7184 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7185 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7189 // We can only fold this if we don't need an extra load.
7190 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
7192 // If BaseGV requires a register, we cannot also have a BaseReg.
7193 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
7197 // X86-64 only supports addr of globals in small code model.
7198 if (Subtarget->is64Bit()) {
7199 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7201 // If lower 4G is not available, then we must use rip-relative addressing.
7202 if (AM.BaseOffs || AM.Scale > 1)
7213 // These scales always work.
7218 // These scales are formed with basereg+scalereg. Only accept if there is
7223 default: // Other stuff never works.
7231 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7232 if (!Ty1->isInteger() || !Ty2->isInteger())
7234 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7235 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7236 if (NumBits1 <= NumBits2)
7238 return Subtarget->is64Bit() || NumBits1 < 64;
7241 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7242 if (!VT1.isInteger() || !VT2.isInteger())
7244 unsigned NumBits1 = VT1.getSizeInBits();
7245 unsigned NumBits2 = VT2.getSizeInBits();
7246 if (NumBits1 <= NumBits2)
7248 return Subtarget->is64Bit() || NumBits1 < 64;
7251 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7252 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7253 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7256 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7257 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7258 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7261 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7262 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7263 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7264 /// are assumed to be legal.
7266 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
7267 // Only do shuffles on 128-bit vector types for now.
7268 // FIXME: pshufb, blends
7269 if (VT.getSizeInBits() == 64) return false;
7270 return (Mask.getNode()->getNumOperands() <= 4 ||
7271 isIdentityMask(Mask.getNode()) ||
7272 isIdentityMask(Mask.getNode(), true) ||
7273 isSplatMask(Mask.getNode()) ||
7274 X86::isPSHUFHWMask(Mask.getNode()) ||
7275 X86::isPSHUFLWMask(Mask.getNode()) ||
7276 X86::isUNPCKLMask(Mask.getNode()) ||
7277 X86::isUNPCKHMask(Mask.getNode()) ||
7278 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
7279 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
7283 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
7284 MVT EVT, SelectionDAG &DAG) const {
7285 unsigned NumElts = BVOps.size();
7286 // Only do shuffles on 128-bit vector types for now.
7287 if (EVT.getSizeInBits() * NumElts == 64) return false;
7288 if (NumElts == 2) return true;
7290 return (isMOVLMask(&BVOps[0], 4) ||
7291 isCommutedMOVL(&BVOps[0], 4, true) ||
7292 isSHUFPMask(&BVOps[0], 4) ||
7293 isCommutedSHUFP(&BVOps[0], 4));
7298 //===----------------------------------------------------------------------===//
7299 // X86 Scheduler Hooks
7300 //===----------------------------------------------------------------------===//
7302 // private utility function
7304 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7305 MachineBasicBlock *MBB,
7313 TargetRegisterClass *RC,
7314 bool invSrc) const {
7315 // For the atomic bitwise operator, we generate
7318 // ld t1 = [bitinstr.addr]
7319 // op t2 = t1, [bitinstr.val]
7321 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7323 // fallthrough -->nextMBB
7324 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7325 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7326 MachineFunction::iterator MBBIter = MBB;
7329 /// First build the CFG
7330 MachineFunction *F = MBB->getParent();
7331 MachineBasicBlock *thisMBB = MBB;
7332 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7333 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7334 F->insert(MBBIter, newMBB);
7335 F->insert(MBBIter, nextMBB);
7337 // Move all successors to thisMBB to nextMBB
7338 nextMBB->transferSuccessors(thisMBB);
7340 // Update thisMBB to fall through to newMBB
7341 thisMBB->addSuccessor(newMBB);
7343 // newMBB jumps to itself and fall through to nextMBB
7344 newMBB->addSuccessor(nextMBB);
7345 newMBB->addSuccessor(newMBB);
7347 // Insert instructions into newMBB based on incoming instruction
7348 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7349 "unexpected number of operands");
7350 DebugLoc dl = bInstr->getDebugLoc();
7351 MachineOperand& destOper = bInstr->getOperand(0);
7352 MachineOperand* argOpers[2 + X86AddrNumOperands];
7353 int numArgs = bInstr->getNumOperands() - 1;
7354 for (int i=0; i < numArgs; ++i)
7355 argOpers[i] = &bInstr->getOperand(i+1);
7357 // x86 address has 4 operands: base, index, scale, and displacement
7358 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7359 int valArgIndx = lastAddrIndx + 1;
7361 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7362 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7363 for (int i=0; i <= lastAddrIndx; ++i)
7364 (*MIB).addOperand(*argOpers[i]);
7366 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7368 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7373 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7374 assert((argOpers[valArgIndx]->isReg() ||
7375 argOpers[valArgIndx]->isImm()) &&
7377 if (argOpers[valArgIndx]->isReg())
7378 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7380 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7382 (*MIB).addOperand(*argOpers[valArgIndx]);
7384 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7387 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7388 for (int i=0; i <= lastAddrIndx; ++i)
7389 (*MIB).addOperand(*argOpers[i]);
7391 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7392 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7394 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7398 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7400 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7404 // private utility function: 64 bit atomics on 32 bit host.
7406 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7407 MachineBasicBlock *MBB,
7412 bool invSrc) const {
7413 // For the atomic bitwise operator, we generate
7414 // thisMBB (instructions are in pairs, except cmpxchg8b)
7415 // ld t1,t2 = [bitinstr.addr]
7417 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7418 // op t5, t6 <- out1, out2, [bitinstr.val]
7419 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7420 // mov ECX, EBX <- t5, t6
7421 // mov EAX, EDX <- t1, t2
7422 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7423 // mov t3, t4 <- EAX, EDX
7425 // result in out1, out2
7426 // fallthrough -->nextMBB
7428 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7429 const unsigned LoadOpc = X86::MOV32rm;
7430 const unsigned copyOpc = X86::MOV32rr;
7431 const unsigned NotOpc = X86::NOT32r;
7432 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7433 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7434 MachineFunction::iterator MBBIter = MBB;
7437 /// First build the CFG
7438 MachineFunction *F = MBB->getParent();
7439 MachineBasicBlock *thisMBB = MBB;
7440 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7441 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7442 F->insert(MBBIter, newMBB);
7443 F->insert(MBBIter, nextMBB);
7445 // Move all successors to thisMBB to nextMBB
7446 nextMBB->transferSuccessors(thisMBB);
7448 // Update thisMBB to fall through to newMBB
7449 thisMBB->addSuccessor(newMBB);
7451 // newMBB jumps to itself and fall through to nextMBB
7452 newMBB->addSuccessor(nextMBB);
7453 newMBB->addSuccessor(newMBB);
7455 DebugLoc dl = bInstr->getDebugLoc();
7456 // Insert instructions into newMBB based on incoming instruction
7457 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7458 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7459 "unexpected number of operands");
7460 MachineOperand& dest1Oper = bInstr->getOperand(0);
7461 MachineOperand& dest2Oper = bInstr->getOperand(1);
7462 MachineOperand* argOpers[2 + X86AddrNumOperands];
7463 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7464 argOpers[i] = &bInstr->getOperand(i+2);
7466 // x86 address has 4 operands: base, index, scale, and displacement
7467 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7469 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7470 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7471 for (int i=0; i <= lastAddrIndx; ++i)
7472 (*MIB).addOperand(*argOpers[i]);
7473 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7474 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7475 // add 4 to displacement.
7476 for (int i=0; i <= lastAddrIndx-2; ++i)
7477 (*MIB).addOperand(*argOpers[i]);
7478 MachineOperand newOp3 = *(argOpers[3]);
7480 newOp3.setImm(newOp3.getImm()+4);
7482 newOp3.setOffset(newOp3.getOffset()+4);
7483 (*MIB).addOperand(newOp3);
7484 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7486 // t3/4 are defined later, at the bottom of the loop
7487 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7488 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7489 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7490 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7491 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7492 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7494 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7495 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7497 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7498 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7504 int valArgIndx = lastAddrIndx + 1;
7505 assert((argOpers[valArgIndx]->isReg() ||
7506 argOpers[valArgIndx]->isImm()) &&
7508 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7509 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7510 if (argOpers[valArgIndx]->isReg())
7511 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7513 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7514 if (regOpcL != X86::MOV32rr)
7516 (*MIB).addOperand(*argOpers[valArgIndx]);
7517 assert(argOpers[valArgIndx + 1]->isReg() ==
7518 argOpers[valArgIndx]->isReg());
7519 assert(argOpers[valArgIndx + 1]->isImm() ==
7520 argOpers[valArgIndx]->isImm());
7521 if (argOpers[valArgIndx + 1]->isReg())
7522 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7524 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7525 if (regOpcH != X86::MOV32rr)
7527 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7529 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7531 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7534 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7536 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7539 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7540 for (int i=0; i <= lastAddrIndx; ++i)
7541 (*MIB).addOperand(*argOpers[i]);
7543 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7544 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7546 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7547 MIB.addReg(X86::EAX);
7548 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7549 MIB.addReg(X86::EDX);
7552 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7554 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7558 // private utility function
7560 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7561 MachineBasicBlock *MBB,
7562 unsigned cmovOpc) const {
7563 // For the atomic min/max operator, we generate
7566 // ld t1 = [min/max.addr]
7567 // mov t2 = [min/max.val]
7569 // cmov[cond] t2 = t1
7571 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7573 // fallthrough -->nextMBB
7575 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7576 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7577 MachineFunction::iterator MBBIter = MBB;
7580 /// First build the CFG
7581 MachineFunction *F = MBB->getParent();
7582 MachineBasicBlock *thisMBB = MBB;
7583 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7584 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7585 F->insert(MBBIter, newMBB);
7586 F->insert(MBBIter, nextMBB);
7588 // Move all successors to thisMBB to nextMBB
7589 nextMBB->transferSuccessors(thisMBB);
7591 // Update thisMBB to fall through to newMBB
7592 thisMBB->addSuccessor(newMBB);
7594 // newMBB jumps to newMBB and fall through to nextMBB
7595 newMBB->addSuccessor(nextMBB);
7596 newMBB->addSuccessor(newMBB);
7598 DebugLoc dl = mInstr->getDebugLoc();
7599 // Insert instructions into newMBB based on incoming instruction
7600 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7601 "unexpected number of operands");
7602 MachineOperand& destOper = mInstr->getOperand(0);
7603 MachineOperand* argOpers[2 + X86AddrNumOperands];
7604 int numArgs = mInstr->getNumOperands() - 1;
7605 for (int i=0; i < numArgs; ++i)
7606 argOpers[i] = &mInstr->getOperand(i+1);
7608 // x86 address has 4 operands: base, index, scale, and displacement
7609 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7610 int valArgIndx = lastAddrIndx + 1;
7612 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7613 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7614 for (int i=0; i <= lastAddrIndx; ++i)
7615 (*MIB).addOperand(*argOpers[i]);
7617 // We only support register and immediate values
7618 assert((argOpers[valArgIndx]->isReg() ||
7619 argOpers[valArgIndx]->isImm()) &&
7622 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7623 if (argOpers[valArgIndx]->isReg())
7624 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7626 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7627 (*MIB).addOperand(*argOpers[valArgIndx]);
7629 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7632 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7637 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7638 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7642 // Cmp and exchange if none has modified the memory location
7643 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7644 for (int i=0; i <= lastAddrIndx; ++i)
7645 (*MIB).addOperand(*argOpers[i]);
7647 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7648 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7650 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7651 MIB.addReg(X86::EAX);
7654 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7656 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7662 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7663 MachineBasicBlock *BB) const {
7664 DebugLoc dl = MI->getDebugLoc();
7665 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7666 switch (MI->getOpcode()) {
7667 default: assert(false && "Unexpected instr type to insert");
7668 case X86::CMOV_V1I64:
7669 case X86::CMOV_FR32:
7670 case X86::CMOV_FR64:
7671 case X86::CMOV_V4F32:
7672 case X86::CMOV_V2F64:
7673 case X86::CMOV_V2I64: {
7674 // To "insert" a SELECT_CC instruction, we actually have to insert the
7675 // diamond control-flow pattern. The incoming instruction knows the
7676 // destination vreg to set, the condition code register to branch on, the
7677 // true/false values to select between, and a branch opcode to use.
7678 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7679 MachineFunction::iterator It = BB;
7685 // cmpTY ccX, r1, r2
7687 // fallthrough --> copy0MBB
7688 MachineBasicBlock *thisMBB = BB;
7689 MachineFunction *F = BB->getParent();
7690 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7691 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7693 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7694 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7695 F->insert(It, copy0MBB);
7696 F->insert(It, sinkMBB);
7697 // Update machine-CFG edges by transferring all successors of the current
7698 // block to the new block which will contain the Phi node for the select.
7699 sinkMBB->transferSuccessors(BB);
7701 // Add the true and fallthrough blocks as its successors.
7702 BB->addSuccessor(copy0MBB);
7703 BB->addSuccessor(sinkMBB);
7706 // %FalseValue = ...
7707 // # fallthrough to sinkMBB
7710 // Update machine-CFG edges
7711 BB->addSuccessor(sinkMBB);
7714 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7717 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7718 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7719 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7721 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7725 case X86::FP32_TO_INT16_IN_MEM:
7726 case X86::FP32_TO_INT32_IN_MEM:
7727 case X86::FP32_TO_INT64_IN_MEM:
7728 case X86::FP64_TO_INT16_IN_MEM:
7729 case X86::FP64_TO_INT32_IN_MEM:
7730 case X86::FP64_TO_INT64_IN_MEM:
7731 case X86::FP80_TO_INT16_IN_MEM:
7732 case X86::FP80_TO_INT32_IN_MEM:
7733 case X86::FP80_TO_INT64_IN_MEM: {
7734 // Change the floating point control register to use "round towards zero"
7735 // mode when truncating to an integer value.
7736 MachineFunction *F = BB->getParent();
7737 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7738 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7740 // Load the old value of the high byte of the control word...
7742 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7743 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7746 // Set the high part to be round to zero...
7747 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7750 // Reload the modified control word now...
7751 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7753 // Restore the memory image of control word to original value
7754 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7757 // Get the X86 opcode to use.
7759 switch (MI->getOpcode()) {
7760 default: assert(0 && "illegal opcode!");
7761 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7762 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7763 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7764 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7765 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7766 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7767 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7768 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7769 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7773 MachineOperand &Op = MI->getOperand(0);
7775 AM.BaseType = X86AddressMode::RegBase;
7776 AM.Base.Reg = Op.getReg();
7778 AM.BaseType = X86AddressMode::FrameIndexBase;
7779 AM.Base.FrameIndex = Op.getIndex();
7781 Op = MI->getOperand(1);
7783 AM.Scale = Op.getImm();
7784 Op = MI->getOperand(2);
7786 AM.IndexReg = Op.getImm();
7787 Op = MI->getOperand(3);
7788 if (Op.isGlobal()) {
7789 AM.GV = Op.getGlobal();
7791 AM.Disp = Op.getImm();
7793 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7794 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7796 // Reload the original control word now.
7797 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7799 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7802 case X86::ATOMAND32:
7803 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7804 X86::AND32ri, X86::MOV32rm,
7805 X86::LCMPXCHG32, X86::MOV32rr,
7806 X86::NOT32r, X86::EAX,
7807 X86::GR32RegisterClass);
7809 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7810 X86::OR32ri, X86::MOV32rm,
7811 X86::LCMPXCHG32, X86::MOV32rr,
7812 X86::NOT32r, X86::EAX,
7813 X86::GR32RegisterClass);
7814 case X86::ATOMXOR32:
7815 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7816 X86::XOR32ri, X86::MOV32rm,
7817 X86::LCMPXCHG32, X86::MOV32rr,
7818 X86::NOT32r, X86::EAX,
7819 X86::GR32RegisterClass);
7820 case X86::ATOMNAND32:
7821 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7822 X86::AND32ri, X86::MOV32rm,
7823 X86::LCMPXCHG32, X86::MOV32rr,
7824 X86::NOT32r, X86::EAX,
7825 X86::GR32RegisterClass, true);
7826 case X86::ATOMMIN32:
7827 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7828 case X86::ATOMMAX32:
7829 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7830 case X86::ATOMUMIN32:
7831 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7832 case X86::ATOMUMAX32:
7833 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7835 case X86::ATOMAND16:
7836 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7837 X86::AND16ri, X86::MOV16rm,
7838 X86::LCMPXCHG16, X86::MOV16rr,
7839 X86::NOT16r, X86::AX,
7840 X86::GR16RegisterClass);
7842 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7843 X86::OR16ri, X86::MOV16rm,
7844 X86::LCMPXCHG16, X86::MOV16rr,
7845 X86::NOT16r, X86::AX,
7846 X86::GR16RegisterClass);
7847 case X86::ATOMXOR16:
7848 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7849 X86::XOR16ri, X86::MOV16rm,
7850 X86::LCMPXCHG16, X86::MOV16rr,
7851 X86::NOT16r, X86::AX,
7852 X86::GR16RegisterClass);
7853 case X86::ATOMNAND16:
7854 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7855 X86::AND16ri, X86::MOV16rm,
7856 X86::LCMPXCHG16, X86::MOV16rr,
7857 X86::NOT16r, X86::AX,
7858 X86::GR16RegisterClass, true);
7859 case X86::ATOMMIN16:
7860 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7861 case X86::ATOMMAX16:
7862 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7863 case X86::ATOMUMIN16:
7864 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7865 case X86::ATOMUMAX16:
7866 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7869 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7870 X86::AND8ri, X86::MOV8rm,
7871 X86::LCMPXCHG8, X86::MOV8rr,
7872 X86::NOT8r, X86::AL,
7873 X86::GR8RegisterClass);
7875 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7876 X86::OR8ri, X86::MOV8rm,
7877 X86::LCMPXCHG8, X86::MOV8rr,
7878 X86::NOT8r, X86::AL,
7879 X86::GR8RegisterClass);
7881 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7882 X86::XOR8ri, X86::MOV8rm,
7883 X86::LCMPXCHG8, X86::MOV8rr,
7884 X86::NOT8r, X86::AL,
7885 X86::GR8RegisterClass);
7886 case X86::ATOMNAND8:
7887 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7888 X86::AND8ri, X86::MOV8rm,
7889 X86::LCMPXCHG8, X86::MOV8rr,
7890 X86::NOT8r, X86::AL,
7891 X86::GR8RegisterClass, true);
7892 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7893 // This group is for 64-bit host.
7894 case X86::ATOMAND64:
7895 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7896 X86::AND64ri32, X86::MOV64rm,
7897 X86::LCMPXCHG64, X86::MOV64rr,
7898 X86::NOT64r, X86::RAX,
7899 X86::GR64RegisterClass);
7901 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7902 X86::OR64ri32, X86::MOV64rm,
7903 X86::LCMPXCHG64, X86::MOV64rr,
7904 X86::NOT64r, X86::RAX,
7905 X86::GR64RegisterClass);
7906 case X86::ATOMXOR64:
7907 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7908 X86::XOR64ri32, X86::MOV64rm,
7909 X86::LCMPXCHG64, X86::MOV64rr,
7910 X86::NOT64r, X86::RAX,
7911 X86::GR64RegisterClass);
7912 case X86::ATOMNAND64:
7913 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7914 X86::AND64ri32, X86::MOV64rm,
7915 X86::LCMPXCHG64, X86::MOV64rr,
7916 X86::NOT64r, X86::RAX,
7917 X86::GR64RegisterClass, true);
7918 case X86::ATOMMIN64:
7919 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7920 case X86::ATOMMAX64:
7921 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7922 case X86::ATOMUMIN64:
7923 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7924 case X86::ATOMUMAX64:
7925 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7927 // This group does 64-bit operations on a 32-bit host.
7928 case X86::ATOMAND6432:
7929 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7930 X86::AND32rr, X86::AND32rr,
7931 X86::AND32ri, X86::AND32ri,
7933 case X86::ATOMOR6432:
7934 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7935 X86::OR32rr, X86::OR32rr,
7936 X86::OR32ri, X86::OR32ri,
7938 case X86::ATOMXOR6432:
7939 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7940 X86::XOR32rr, X86::XOR32rr,
7941 X86::XOR32ri, X86::XOR32ri,
7943 case X86::ATOMNAND6432:
7944 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7945 X86::AND32rr, X86::AND32rr,
7946 X86::AND32ri, X86::AND32ri,
7948 case X86::ATOMADD6432:
7949 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7950 X86::ADD32rr, X86::ADC32rr,
7951 X86::ADD32ri, X86::ADC32ri,
7953 case X86::ATOMSUB6432:
7954 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7955 X86::SUB32rr, X86::SBB32rr,
7956 X86::SUB32ri, X86::SBB32ri,
7958 case X86::ATOMSWAP6432:
7959 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7960 X86::MOV32rr, X86::MOV32rr,
7961 X86::MOV32ri, X86::MOV32ri,
7966 //===----------------------------------------------------------------------===//
7967 // X86 Optimization Hooks
7968 //===----------------------------------------------------------------------===//
7970 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7974 const SelectionDAG &DAG,
7975 unsigned Depth) const {
7976 unsigned Opc = Op.getOpcode();
7977 assert((Opc >= ISD::BUILTIN_OP_END ||
7978 Opc == ISD::INTRINSIC_WO_CHAIN ||
7979 Opc == ISD::INTRINSIC_W_CHAIN ||
7980 Opc == ISD::INTRINSIC_VOID) &&
7981 "Should use MaskedValueIsZero if you don't know whether Op"
7982 " is a target node!");
7984 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7993 // These nodes' second result is a boolean.
7994 if (Op.getResNo() == 0)
7998 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7999 Mask.getBitWidth() - 1);
8004 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8005 /// node is a GlobalAddress + offset.
8006 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8007 GlobalValue* &GA, int64_t &Offset) const{
8008 if (N->getOpcode() == X86ISD::Wrapper) {
8009 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8010 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8011 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8015 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8018 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8019 const TargetLowering &TLI) {
8022 if (TLI.isGAPlusOffset(Base, GV, Offset))
8023 return (GV->getAlignment() >= N && (Offset % N) == 0);
8024 // DAG combine handles the stack object case.
8028 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
8029 unsigned NumElems, MVT EVT,
8031 SelectionDAG &DAG, MachineFrameInfo *MFI,
8032 const TargetLowering &TLI) {
8034 for (unsigned i = 0; i < NumElems; ++i) {
8035 SDValue Idx = PermMask.getOperand(i);
8036 if (Idx.getOpcode() == ISD::UNDEF) {
8042 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8043 if (!Elt.getNode() ||
8044 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8047 Base = Elt.getNode();
8048 if (Base->getOpcode() == ISD::UNDEF)
8052 if (Elt.getOpcode() == ISD::UNDEF)
8055 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
8056 EVT.getSizeInBits()/8, i, MFI))
8062 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8063 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8064 /// if the load addresses are consecutive, non-overlapping, and in the right
8065 /// order. In the case of v2i64, it will see if it can rewrite the
8066 /// shuffle to be an appropriate build vector so it can take advantage of
8067 // performBuildVectorCombine.
8068 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8069 const TargetLowering &TLI) {
8070 DebugLoc dl = N->getDebugLoc();
8071 MVT VT = N->getValueType(0);
8072 MVT EVT = VT.getVectorElementType();
8073 SDValue PermMask = N->getOperand(2);
8074 unsigned NumElems = PermMask.getNumOperands();
8076 // For x86-32 machines, if we see an insert and then a shuffle in a v2i64
8077 // where the upper half is 0, it is advantageous to rewrite it as a build
8078 // vector of (0, val) so it can use movq.
8079 if (VT == MVT::v2i64) {
8081 In[0] = N->getOperand(0);
8082 In[1] = N->getOperand(1);
8083 unsigned Idx0 =cast<ConstantSDNode>(PermMask.getOperand(0))->getZExtValue();
8084 unsigned Idx1 =cast<ConstantSDNode>(PermMask.getOperand(1))->getZExtValue();
8085 if (In[0].getValueType().getVectorNumElements() == NumElems &&
8086 In[Idx0/2].getOpcode() == ISD::INSERT_VECTOR_ELT &&
8087 In[Idx1/2].getOpcode() == ISD::BUILD_VECTOR) {
8088 ConstantSDNode* InsertVecIdx =
8089 dyn_cast<ConstantSDNode>(In[Idx0/2].getOperand(2));
8091 InsertVecIdx->getZExtValue() == (Idx0 % 2) &&
8092 isZeroNode(In[Idx1/2].getOperand(Idx1 % 2))) {
8093 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
8094 In[Idx0/2].getOperand(1),
8095 In[Idx1/2].getOperand(Idx1 % 2));
8100 // Try to combine a vector_shuffle into a 128-bit load.
8101 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8102 SDNode *Base = NULL;
8103 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
8107 LoadSDNode *LD = cast<LoadSDNode>(Base);
8108 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
8109 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8110 LD->getSrcValue(), LD->getSrcValueOffset(),
8112 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8113 LD->getSrcValue(), LD->getSrcValueOffset(),
8114 LD->isVolatile(), LD->getAlignment());
8117 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
8118 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
8119 TargetLowering::DAGCombinerInfo &DCI,
8120 const X86Subtarget *Subtarget,
8121 const TargetLowering &TLI) {
8122 unsigned NumOps = N->getNumOperands();
8123 DebugLoc dl = N->getDebugLoc();
8125 // Ignore single operand BUILD_VECTOR.
8129 MVT VT = N->getValueType(0);
8130 MVT EVT = VT.getVectorElementType();
8131 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
8132 // We are looking for load i64 and zero extend. We want to transform
8133 // it before legalizer has a chance to expand it. Also look for i64
8134 // BUILD_PAIR bit casted to f64.
8136 // This must be an insertion into a zero vector.
8137 SDValue HighElt = N->getOperand(1);
8138 if (!isZeroNode(HighElt))
8141 // Value must be a load.
8142 SDNode *Base = N->getOperand(0).getNode();
8143 if (!isa<LoadSDNode>(Base)) {
8144 if (Base->getOpcode() != ISD::BIT_CONVERT)
8146 Base = Base->getOperand(0).getNode();
8147 if (!isa<LoadSDNode>(Base))
8151 // Transform it into VZEXT_LOAD addr.
8152 LoadSDNode *LD = cast<LoadSDNode>(Base);
8154 // Load must not be an extload.
8155 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
8158 // Load type should legal type so we don't have to legalize it.
8159 if (!TLI.isTypeLegal(VT))
8162 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
8163 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8164 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8165 TargetLowering::TargetLoweringOpt TLO(DAG);
8166 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
8167 DCI.CommitTargetLoweringOpt(TLO);
8171 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8172 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8173 const X86Subtarget *Subtarget) {
8174 DebugLoc DL = N->getDebugLoc();
8175 SDValue Cond = N->getOperand(0);
8176 // Get the LHS/RHS of the select.
8177 SDValue LHS = N->getOperand(1);
8178 SDValue RHS = N->getOperand(2);
8180 // If we have SSE[12] support, try to form min/max nodes.
8181 if (Subtarget->hasSSE2() &&
8182 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8183 Cond.getOpcode() == ISD::SETCC) {
8184 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8186 unsigned Opcode = 0;
8187 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8190 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8193 if (!UnsafeFPMath) break;
8195 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8197 Opcode = X86ISD::FMIN;
8200 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8203 if (!UnsafeFPMath) break;
8205 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8207 Opcode = X86ISD::FMAX;
8210 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8213 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8216 if (!UnsafeFPMath) break;
8218 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8220 Opcode = X86ISD::FMIN;
8223 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8226 if (!UnsafeFPMath) break;
8228 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8230 Opcode = X86ISD::FMAX;
8236 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8239 // If this is a select between two integer constants, try to do some
8241 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8242 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8243 // Don't do this for crazy integer types.
8244 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8245 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8246 // so that TrueC (the true value) is larger than FalseC.
8247 bool NeedsCondInvert = false;
8249 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8250 // Efficiently invertible.
8251 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8252 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8253 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8254 NeedsCondInvert = true;
8255 std::swap(TrueC, FalseC);
8258 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8259 if (FalseC->getAPIntValue() == 0 &&
8260 TrueC->getAPIntValue().isPowerOf2()) {
8261 if (NeedsCondInvert) // Invert the condition if needed.
8262 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8263 DAG.getConstant(1, Cond.getValueType()));
8265 // Zero extend the condition if needed.
8266 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8268 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8269 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8270 DAG.getConstant(ShAmt, MVT::i8));
8273 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8274 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8275 if (NeedsCondInvert) // Invert the condition if needed.
8276 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8277 DAG.getConstant(1, Cond.getValueType()));
8279 // Zero extend the condition if needed.
8280 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8281 FalseC->getValueType(0), Cond);
8282 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8283 SDValue(FalseC, 0));
8286 // Optimize cases that will turn into an LEA instruction. This requires
8287 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8288 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8289 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8290 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8292 bool isFastMultiplier = false;
8294 switch ((unsigned char)Diff) {
8296 case 1: // result = add base, cond
8297 case 2: // result = lea base( , cond*2)
8298 case 3: // result = lea base(cond, cond*2)
8299 case 4: // result = lea base( , cond*4)
8300 case 5: // result = lea base(cond, cond*4)
8301 case 8: // result = lea base( , cond*8)
8302 case 9: // result = lea base(cond, cond*8)
8303 isFastMultiplier = true;
8308 if (isFastMultiplier) {
8309 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8310 if (NeedsCondInvert) // Invert the condition if needed.
8311 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8312 DAG.getConstant(1, Cond.getValueType()));
8314 // Zero extend the condition if needed.
8315 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8317 // Scale the condition by the difference.
8319 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8320 DAG.getConstant(Diff, Cond.getValueType()));
8322 // Add the base if non-zero.
8323 if (FalseC->getAPIntValue() != 0)
8324 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8325 SDValue(FalseC, 0));
8335 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8336 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8337 TargetLowering::DAGCombinerInfo &DCI) {
8338 DebugLoc DL = N->getDebugLoc();
8340 // If the flag operand isn't dead, don't touch this CMOV.
8341 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8344 // If this is a select between two integer constants, try to do some
8345 // optimizations. Note that the operands are ordered the opposite of SELECT
8347 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8348 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8349 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8350 // larger than FalseC (the false value).
8351 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8353 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8354 CC = X86::GetOppositeBranchCondition(CC);
8355 std::swap(TrueC, FalseC);
8358 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8359 // This is efficient for any integer data type (including i8/i16) and
8361 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8362 SDValue Cond = N->getOperand(3);
8363 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8364 DAG.getConstant(CC, MVT::i8), Cond);
8366 // Zero extend the condition if needed.
8367 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8369 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8370 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8371 DAG.getConstant(ShAmt, MVT::i8));
8372 if (N->getNumValues() == 2) // Dead flag value?
8373 return DCI.CombineTo(N, Cond, SDValue());
8377 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8378 // for any integer data type, including i8/i16.
8379 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8380 SDValue Cond = N->getOperand(3);
8381 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8382 DAG.getConstant(CC, MVT::i8), Cond);
8384 // Zero extend the condition if needed.
8385 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8386 FalseC->getValueType(0), Cond);
8387 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8388 SDValue(FalseC, 0));
8390 if (N->getNumValues() == 2) // Dead flag value?
8391 return DCI.CombineTo(N, Cond, SDValue());
8395 // Optimize cases that will turn into an LEA instruction. This requires
8396 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8397 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8398 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8399 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8401 bool isFastMultiplier = false;
8403 switch ((unsigned char)Diff) {
8405 case 1: // result = add base, cond
8406 case 2: // result = lea base( , cond*2)
8407 case 3: // result = lea base(cond, cond*2)
8408 case 4: // result = lea base( , cond*4)
8409 case 5: // result = lea base(cond, cond*4)
8410 case 8: // result = lea base( , cond*8)
8411 case 9: // result = lea base(cond, cond*8)
8412 isFastMultiplier = true;
8417 if (isFastMultiplier) {
8418 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8419 SDValue Cond = N->getOperand(3);
8420 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8421 DAG.getConstant(CC, MVT::i8), Cond);
8422 // Zero extend the condition if needed.
8423 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8425 // Scale the condition by the difference.
8427 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8428 DAG.getConstant(Diff, Cond.getValueType()));
8430 // Add the base if non-zero.
8431 if (FalseC->getAPIntValue() != 0)
8432 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8433 SDValue(FalseC, 0));
8434 if (N->getNumValues() == 2) // Dead flag value?
8435 return DCI.CombineTo(N, Cond, SDValue());
8445 /// PerformMulCombine - Optimize a single multiply with constant into two
8446 /// in order to implement it with two cheaper instructions, e.g.
8447 /// LEA + SHL, LEA + LEA.
8448 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8449 TargetLowering::DAGCombinerInfo &DCI) {
8450 if (DAG.getMachineFunction().
8451 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8454 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8457 MVT VT = N->getValueType(0);
8461 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8464 uint64_t MulAmt = C->getZExtValue();
8465 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8468 uint64_t MulAmt1 = 0;
8469 uint64_t MulAmt2 = 0;
8470 if ((MulAmt % 9) == 0) {
8472 MulAmt2 = MulAmt / 9;
8473 } else if ((MulAmt % 5) == 0) {
8475 MulAmt2 = MulAmt / 5;
8476 } else if ((MulAmt % 3) == 0) {
8478 MulAmt2 = MulAmt / 3;
8481 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8482 DebugLoc DL = N->getDebugLoc();
8484 if (isPowerOf2_64(MulAmt2) &&
8485 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8486 // If second multiplifer is pow2, issue it first. We want the multiply by
8487 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8489 std::swap(MulAmt1, MulAmt2);
8492 if (isPowerOf2_64(MulAmt1))
8493 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8494 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8496 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8497 DAG.getConstant(MulAmt1, VT));
8499 if (isPowerOf2_64(MulAmt2))
8500 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8501 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8503 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8504 DAG.getConstant(MulAmt2, VT));
8506 // Do not add new nodes to DAG combiner worklist.
8507 DCI.CombineTo(N, NewMul, false);
8513 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8515 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8516 const X86Subtarget *Subtarget) {
8517 // On X86 with SSE2 support, we can transform this to a vector shift if
8518 // all elements are shifted by the same amount. We can't do this in legalize
8519 // because the a constant vector is typically transformed to a constant pool
8520 // so we have no knowledge of the shift amount.
8521 if (!Subtarget->hasSSE2())
8524 MVT VT = N->getValueType(0);
8525 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8528 SDValue ShAmtOp = N->getOperand(1);
8529 MVT EltVT = VT.getVectorElementType();
8530 DebugLoc DL = N->getDebugLoc();
8532 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8533 unsigned NumElts = VT.getVectorNumElements();
8535 for (; i != NumElts; ++i) {
8536 SDValue Arg = ShAmtOp.getOperand(i);
8537 if (Arg.getOpcode() == ISD::UNDEF) continue;
8541 for (; i != NumElts; ++i) {
8542 SDValue Arg = ShAmtOp.getOperand(i);
8543 if (Arg.getOpcode() == ISD::UNDEF) continue;
8544 if (Arg != BaseShAmt) {
8548 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8549 isSplatMask(ShAmtOp.getOperand(2).getNode())) {
8550 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8551 DAG.getIntPtrConstant(0));
8555 if (EltVT.bitsGT(MVT::i32))
8556 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8557 else if (EltVT.bitsLT(MVT::i32))
8558 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8560 // The shift amount is identical so we can do a vector shift.
8561 SDValue ValOp = N->getOperand(0);
8562 switch (N->getOpcode()) {
8564 assert(0 && "Unknown shift opcode!");
8567 if (VT == MVT::v2i64)
8568 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8569 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8571 if (VT == MVT::v4i32)
8572 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8573 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8575 if (VT == MVT::v8i16)
8576 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8577 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8581 if (VT == MVT::v4i32)
8582 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8583 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8585 if (VT == MVT::v8i16)
8586 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8587 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8591 if (VT == MVT::v2i64)
8592 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8593 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8595 if (VT == MVT::v4i32)
8596 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8597 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8599 if (VT == MVT::v8i16)
8600 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8601 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8608 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8609 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8610 const X86Subtarget *Subtarget) {
8611 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8612 // the FP state in cases where an emms may be missing.
8613 // A preferable solution to the general problem is to figure out the right
8614 // places to insert EMMS. This qualifies as a quick hack.
8616 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8617 StoreSDNode *St = cast<StoreSDNode>(N);
8618 MVT VT = St->getValue().getValueType();
8619 if (VT.getSizeInBits() != 64)
8622 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2();
8623 if ((VT.isVector() ||
8624 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8625 isa<LoadSDNode>(St->getValue()) &&
8626 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8627 St->getChain().hasOneUse() && !St->isVolatile()) {
8628 SDNode* LdVal = St->getValue().getNode();
8630 int TokenFactorIndex = -1;
8631 SmallVector<SDValue, 8> Ops;
8632 SDNode* ChainVal = St->getChain().getNode();
8633 // Must be a store of a load. We currently handle two cases: the load
8634 // is a direct child, and it's under an intervening TokenFactor. It is
8635 // possible to dig deeper under nested TokenFactors.
8636 if (ChainVal == LdVal)
8637 Ld = cast<LoadSDNode>(St->getChain());
8638 else if (St->getValue().hasOneUse() &&
8639 ChainVal->getOpcode() == ISD::TokenFactor) {
8640 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8641 if (ChainVal->getOperand(i).getNode() == LdVal) {
8642 TokenFactorIndex = i;
8643 Ld = cast<LoadSDNode>(St->getValue());
8645 Ops.push_back(ChainVal->getOperand(i));
8649 if (!Ld || !ISD::isNormalLoad(Ld))
8652 // If this is not the MMX case, i.e. we are just turning i64 load/store
8653 // into f64 load/store, avoid the transformation if there are multiple
8654 // uses of the loaded value.
8655 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8658 DebugLoc LdDL = Ld->getDebugLoc();
8659 DebugLoc StDL = N->getDebugLoc();
8660 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8661 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8663 if (Subtarget->is64Bit() || F64IsLegal) {
8664 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8665 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8666 Ld->getBasePtr(), Ld->getSrcValue(),
8667 Ld->getSrcValueOffset(), Ld->isVolatile(),
8668 Ld->getAlignment());
8669 SDValue NewChain = NewLd.getValue(1);
8670 if (TokenFactorIndex != -1) {
8671 Ops.push_back(NewChain);
8672 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8675 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8676 St->getSrcValue(), St->getSrcValueOffset(),
8677 St->isVolatile(), St->getAlignment());
8680 // Otherwise, lower to two pairs of 32-bit loads / stores.
8681 SDValue LoAddr = Ld->getBasePtr();
8682 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8683 DAG.getConstant(4, MVT::i32));
8685 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8686 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8687 Ld->isVolatile(), Ld->getAlignment());
8688 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8689 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8691 MinAlign(Ld->getAlignment(), 4));
8693 SDValue NewChain = LoLd.getValue(1);
8694 if (TokenFactorIndex != -1) {
8695 Ops.push_back(LoLd);
8696 Ops.push_back(HiLd);
8697 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8701 LoAddr = St->getBasePtr();
8702 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8703 DAG.getConstant(4, MVT::i32));
8705 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8706 St->getSrcValue(), St->getSrcValueOffset(),
8707 St->isVolatile(), St->getAlignment());
8708 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8710 St->getSrcValueOffset() + 4,
8712 MinAlign(St->getAlignment(), 4));
8713 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8718 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8719 /// X86ISD::FXOR nodes.
8720 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8721 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8722 // F[X]OR(0.0, x) -> x
8723 // F[X]OR(x, 0.0) -> x
8724 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8725 if (C->getValueAPF().isPosZero())
8726 return N->getOperand(1);
8727 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8728 if (C->getValueAPF().isPosZero())
8729 return N->getOperand(0);
8733 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8734 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8735 // FAND(0.0, x) -> 0.0
8736 // FAND(x, 0.0) -> 0.0
8737 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8738 if (C->getValueAPF().isPosZero())
8739 return N->getOperand(0);
8740 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8741 if (C->getValueAPF().isPosZero())
8742 return N->getOperand(1);
8746 static SDValue PerformBTCombine(SDNode *N,
8748 TargetLowering::DAGCombinerInfo &DCI) {
8749 // BT ignores high bits in the bit index operand.
8750 SDValue Op1 = N->getOperand(1);
8751 if (Op1.hasOneUse()) {
8752 unsigned BitWidth = Op1.getValueSizeInBits();
8753 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8754 APInt KnownZero, KnownOne;
8755 TargetLowering::TargetLoweringOpt TLO(DAG);
8756 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8757 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8758 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8759 DCI.CommitTargetLoweringOpt(TLO);
8764 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8765 DAGCombinerInfo &DCI) const {
8766 SelectionDAG &DAG = DCI.DAG;
8767 switch (N->getOpcode()) {
8769 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8770 case ISD::BUILD_VECTOR:
8771 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8772 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8773 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8774 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8777 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8778 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8780 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8781 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8782 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8788 //===----------------------------------------------------------------------===//
8789 // X86 Inline Assembly Support
8790 //===----------------------------------------------------------------------===//
8792 /// getConstraintType - Given a constraint letter, return the type of
8793 /// constraint it is for this target.
8794 X86TargetLowering::ConstraintType
8795 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8796 if (Constraint.size() == 1) {
8797 switch (Constraint[0]) {
8809 return C_RegisterClass;
8817 return TargetLowering::getConstraintType(Constraint);
8820 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8821 /// with another that has more specific requirements based on the type of the
8822 /// corresponding operand.
8823 const char *X86TargetLowering::
8824 LowerXConstraint(MVT ConstraintVT) const {
8825 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8826 // 'f' like normal targets.
8827 if (ConstraintVT.isFloatingPoint()) {
8828 if (Subtarget->hasSSE2())
8830 if (Subtarget->hasSSE1())
8834 return TargetLowering::LowerXConstraint(ConstraintVT);
8837 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8838 /// vector. If it is invalid, don't add anything to Ops.
8839 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8842 std::vector<SDValue>&Ops,
8843 SelectionDAG &DAG) const {
8844 SDValue Result(0, 0);
8846 switch (Constraint) {
8849 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8850 if (C->getZExtValue() <= 31) {
8851 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8857 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8858 if (C->getZExtValue() <= 63) {
8859 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8866 if (C->getZExtValue() <= 255) {
8867 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8873 // 32-bit signed value
8874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8875 const ConstantInt *CI = C->getConstantIntValue();
8876 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8877 // Widen to 64 bits here to get it sign extended.
8878 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8881 // FIXME gcc accepts some relocatable values here too, but only in certain
8882 // memory models; it's complicated.
8887 // 32-bit unsigned value
8888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8889 const ConstantInt *CI = C->getConstantIntValue();
8890 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8891 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8895 // FIXME gcc accepts some relocatable values here too, but only in certain
8896 // memory models; it's complicated.
8900 // Literal immediates are always ok.
8901 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8902 // Widen to 64 bits here to get it sign extended.
8903 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8907 // If we are in non-pic codegen mode, we allow the address of a global (with
8908 // an optional displacement) to be used with 'i'.
8909 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8912 // Match either (GA) or (GA+C)
8914 Offset = GA->getOffset();
8915 } else if (Op.getOpcode() == ISD::ADD) {
8916 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8917 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8919 Offset = GA->getOffset()+C->getZExtValue();
8921 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8922 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8924 Offset = GA->getOffset()+C->getZExtValue();
8932 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8935 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8941 // Otherwise, not valid for this mode.
8946 if (Result.getNode()) {
8947 Ops.push_back(Result);
8950 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8954 std::vector<unsigned> X86TargetLowering::
8955 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8957 if (Constraint.size() == 1) {
8958 // FIXME: not handling fp-stack yet!
8959 switch (Constraint[0]) { // GCC X86 Constraint Letters
8960 default: break; // Unknown constraint letter
8961 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8964 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8965 else if (VT == MVT::i16)
8966 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8967 else if (VT == MVT::i8)
8968 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8969 else if (VT == MVT::i64)
8970 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8975 return std::vector<unsigned>();
8978 std::pair<unsigned, const TargetRegisterClass*>
8979 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8981 // First, see if this is a constraint that directly corresponds to an LLVM
8983 if (Constraint.size() == 1) {
8984 // GCC Constraint Letters
8985 switch (Constraint[0]) {
8987 case 'r': // GENERAL_REGS
8988 case 'R': // LEGACY_REGS
8989 case 'l': // INDEX_REGS
8991 return std::make_pair(0U, X86::GR8RegisterClass);
8993 return std::make_pair(0U, X86::GR16RegisterClass);
8994 if (VT == MVT::i32 || !Subtarget->is64Bit())
8995 return std::make_pair(0U, X86::GR32RegisterClass);
8996 return std::make_pair(0U, X86::GR64RegisterClass);
8997 case 'f': // FP Stack registers.
8998 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8999 // value to the correct fpstack register class.
9000 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9001 return std::make_pair(0U, X86::RFP32RegisterClass);
9002 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9003 return std::make_pair(0U, X86::RFP64RegisterClass);
9004 return std::make_pair(0U, X86::RFP80RegisterClass);
9005 case 'y': // MMX_REGS if MMX allowed.
9006 if (!Subtarget->hasMMX()) break;
9007 return std::make_pair(0U, X86::VR64RegisterClass);
9008 case 'Y': // SSE_REGS if SSE2 allowed
9009 if (!Subtarget->hasSSE2()) break;
9011 case 'x': // SSE_REGS if SSE1 allowed
9012 if (!Subtarget->hasSSE1()) break;
9014 switch (VT.getSimpleVT()) {
9016 // Scalar SSE types.
9019 return std::make_pair(0U, X86::FR32RegisterClass);
9022 return std::make_pair(0U, X86::FR64RegisterClass);
9030 return std::make_pair(0U, X86::VR128RegisterClass);
9036 // Use the default implementation in TargetLowering to convert the register
9037 // constraint into a member of a register class.
9038 std::pair<unsigned, const TargetRegisterClass*> Res;
9039 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9041 // Not found as a standard register?
9042 if (Res.second == 0) {
9043 // GCC calls "st(0)" just plain "st".
9044 if (StringsEqualNoCase("{st}", Constraint)) {
9045 Res.first = X86::ST0;
9046 Res.second = X86::RFP80RegisterClass;
9048 // 'A' means EAX + EDX.
9049 if (Constraint == "A") {
9050 Res.first = X86::EAX;
9051 Res.second = X86::GRADRegisterClass;
9056 // Otherwise, check to see if this is a register class of the wrong value
9057 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9058 // turn into {ax},{dx}.
9059 if (Res.second->hasType(VT))
9060 return Res; // Correct type already, nothing to do.
9062 // All of the single-register GCC register classes map their values onto
9063 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9064 // really want an 8-bit or 32-bit register, map to the appropriate register
9065 // class and return the appropriate register.
9066 if (Res.second == X86::GR16RegisterClass) {
9067 if (VT == MVT::i8) {
9068 unsigned DestReg = 0;
9069 switch (Res.first) {
9071 case X86::AX: DestReg = X86::AL; break;
9072 case X86::DX: DestReg = X86::DL; break;
9073 case X86::CX: DestReg = X86::CL; break;
9074 case X86::BX: DestReg = X86::BL; break;
9077 Res.first = DestReg;
9078 Res.second = X86::GR8RegisterClass;
9080 } else if (VT == MVT::i32) {
9081 unsigned DestReg = 0;
9082 switch (Res.first) {
9084 case X86::AX: DestReg = X86::EAX; break;
9085 case X86::DX: DestReg = X86::EDX; break;
9086 case X86::CX: DestReg = X86::ECX; break;
9087 case X86::BX: DestReg = X86::EBX; break;
9088 case X86::SI: DestReg = X86::ESI; break;
9089 case X86::DI: DestReg = X86::EDI; break;
9090 case X86::BP: DestReg = X86::EBP; break;
9091 case X86::SP: DestReg = X86::ESP; break;
9094 Res.first = DestReg;
9095 Res.second = X86::GR32RegisterClass;
9097 } else if (VT == MVT::i64) {
9098 unsigned DestReg = 0;
9099 switch (Res.first) {
9101 case X86::AX: DestReg = X86::RAX; break;
9102 case X86::DX: DestReg = X86::RDX; break;
9103 case X86::CX: DestReg = X86::RCX; break;
9104 case X86::BX: DestReg = X86::RBX; break;
9105 case X86::SI: DestReg = X86::RSI; break;
9106 case X86::DI: DestReg = X86::RDI; break;
9107 case X86::BP: DestReg = X86::RBP; break;
9108 case X86::SP: DestReg = X86::RSP; break;
9111 Res.first = DestReg;
9112 Res.second = X86::GR64RegisterClass;
9115 } else if (Res.second == X86::FR32RegisterClass ||
9116 Res.second == X86::FR64RegisterClass ||
9117 Res.second == X86::VR128RegisterClass) {
9118 // Handle references to XMM physical registers that got mapped into the
9119 // wrong class. This can happen with constraints like {xmm0} where the
9120 // target independent register mapper will just pick the first match it can
9121 // find, ignoring the required type.
9123 Res.second = X86::FR32RegisterClass;
9124 else if (VT == MVT::f64)
9125 Res.second = X86::FR64RegisterClass;
9126 else if (X86::VR128RegisterClass->hasType(VT))
9127 Res.second = X86::VR128RegisterClass;
9133 //===----------------------------------------------------------------------===//
9134 // X86 Widen vector type
9135 //===----------------------------------------------------------------------===//
9137 /// getWidenVectorType: given a vector type, returns the type to widen
9138 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9139 /// If there is no vector type that we want to widen to, returns MVT::Other
9140 /// When and where to widen is target dependent based on the cost of
9141 /// scalarizing vs using the wider vector type.
9143 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9144 assert(VT.isVector());
9145 if (isTypeLegal(VT))
9148 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9149 // type based on element type. This would speed up our search (though
9150 // it may not be worth it since the size of the list is relatively
9152 MVT EltVT = VT.getVectorElementType();
9153 unsigned NElts = VT.getVectorNumElements();
9155 // On X86, it make sense to widen any vector wider than 1
9159 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9160 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9161 MVT SVT = (MVT::SimpleValueType)nVT;
9163 if (isTypeLegal(SVT) &&
9164 SVT.getVectorElementType() == EltVT &&
9165 SVT.getVectorNumElements() > NElts)