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"
44 const int X86AddrNumOperands = 4;
47 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
49 // Forward declarations.
50 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl);
52 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
53 : TargetLowering(TM) {
54 Subtarget = &TM.getSubtarget<X86Subtarget>();
55 X86ScalarSSEf64 = Subtarget->hasSSE2();
56 X86ScalarSSEf32 = Subtarget->hasSSE1();
57 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
61 RegInfo = TM.getRegisterInfo();
64 // Set up the TargetLowering object.
66 // X86 is weird, it always uses i8 for shift amounts and setcc results.
67 setShiftAmountType(MVT::i8);
68 setBooleanContents(ZeroOrOneBooleanContent);
69 setSchedulingPreference(SchedulingForRegPressure);
70 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
71 setStackPointerRegisterToSaveRestore(X86StackPtr);
73 if (Subtarget->isTargetDarwin()) {
74 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
75 setUseUnderscoreSetJmp(false);
76 setUseUnderscoreLongJmp(false);
77 } else if (Subtarget->isTargetMingw()) {
78 // MS runtime is weird: it exports _setjmp, but longjmp!
79 setUseUnderscoreSetJmp(true);
80 setUseUnderscoreLongJmp(false);
82 setUseUnderscoreSetJmp(true);
83 setUseUnderscoreLongJmp(true);
86 // Set up the register classes.
87 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
88 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
89 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
90 if (Subtarget->is64Bit())
91 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
93 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
95 // We don't accept any truncstore of integer registers.
96 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
97 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
98 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
99 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
100 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
101 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
103 // SETOEQ and SETUNE require checking two conditions.
104 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
105 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
106 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
107 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
108 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
109 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
111 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
113 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
114 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
115 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
117 if (Subtarget->is64Bit()) {
118 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
119 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
121 if (!UseSoftFloat && !NoImplicitFloat && X86ScalarSSEf64) {
122 // We have an impenetrably clever algorithm for ui64->double only.
123 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
125 // We have faster algorithm for ui32->single only.
126 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
128 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
132 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
134 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
135 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
137 if (!UseSoftFloat && !NoImplicitFloat) {
138 // SSE has no i16 to fp conversion, only i32
139 if (X86ScalarSSEf32) {
140 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
141 // f32 and f64 cases are Legal, f80 case is not
142 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
144 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
145 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
148 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
149 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
152 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
153 // are Legal, f80 is custom lowered.
154 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
155 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
157 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
159 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
160 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
162 if (X86ScalarSSEf32) {
163 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
164 // f32 and f64 cases are Legal, f80 case is not
165 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
167 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
168 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
171 // Handle FP_TO_UINT by promoting the destination to a larger signed
173 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
174 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
175 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
177 if (Subtarget->is64Bit()) {
178 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
179 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
181 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
182 // Expand FP_TO_UINT into a select.
183 // FIXME: We would like to use a Custom expander here eventually to do
184 // the optimal thing for SSE vs. the default expansion in the legalizer.
185 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
187 // With SSE3 we can use fisttpll to convert to a signed i64.
188 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
191 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
192 if (!X86ScalarSSEf64) {
193 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
194 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
197 // Scalar integer divide and remainder are lowered to use operations that
198 // produce two results, to match the available instructions. This exposes
199 // the two-result form to trivial CSE, which is able to combine x/y and x%y
200 // into a single instruction.
202 // Scalar integer multiply-high is also lowered to use two-result
203 // operations, to match the available instructions. However, plain multiply
204 // (low) operations are left as Legal, as there are single-result
205 // instructions for this in x86. Using the two-result multiply instructions
206 // when both high and low results are needed must be arranged by dagcombine.
207 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
208 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
209 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
210 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
211 setOperationAction(ISD::SREM , MVT::i8 , Expand);
212 setOperationAction(ISD::UREM , MVT::i8 , Expand);
213 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
214 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
215 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
216 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
217 setOperationAction(ISD::SREM , MVT::i16 , Expand);
218 setOperationAction(ISD::UREM , MVT::i16 , Expand);
219 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
220 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
221 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
222 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
223 setOperationAction(ISD::SREM , MVT::i32 , Expand);
224 setOperationAction(ISD::UREM , MVT::i32 , Expand);
225 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
226 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
227 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
228 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
229 setOperationAction(ISD::SREM , MVT::i64 , Expand);
230 setOperationAction(ISD::UREM , MVT::i64 , Expand);
232 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
233 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
234 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
235 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
236 if (Subtarget->is64Bit())
237 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
238 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
239 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
240 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
241 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
242 setOperationAction(ISD::FREM , MVT::f32 , Expand);
243 setOperationAction(ISD::FREM , MVT::f64 , Expand);
244 setOperationAction(ISD::FREM , MVT::f80 , Expand);
245 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
247 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
248 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
249 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
250 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
251 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
252 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
253 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
254 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
255 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
258 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
259 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
262 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
263 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
265 // These should be promoted to a larger select which is supported.
266 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
267 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
268 // X86 wants to expand cmov itself.
269 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
270 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
271 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
272 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
273 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
274 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
275 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
276 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
277 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
278 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
279 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
280 if (Subtarget->is64Bit()) {
281 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
282 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
284 // X86 ret instruction may pop stack.
285 setOperationAction(ISD::RET , MVT::Other, Custom);
286 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
289 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
290 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
291 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
292 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
293 if (Subtarget->is64Bit())
294 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
295 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
298 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
299 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
300 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
302 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
303 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
304 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
305 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
306 if (Subtarget->is64Bit()) {
307 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
308 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
309 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
312 if (Subtarget->hasSSE1())
313 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
315 if (!Subtarget->hasSSE2())
316 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
318 // Expand certain atomics
319 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
320 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
321 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
322 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
327 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
329 if (!Subtarget->is64Bit()) {
330 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
333 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
334 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
335 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
336 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
339 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
340 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
341 // FIXME - use subtarget debug flags
342 if (!Subtarget->isTargetDarwin() &&
343 !Subtarget->isTargetELF() &&
344 !Subtarget->isTargetCygMing()) {
345 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
346 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
349 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
350 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
351 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
352 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
353 if (Subtarget->is64Bit()) {
354 setExceptionPointerRegister(X86::RAX);
355 setExceptionSelectorRegister(X86::RDX);
357 setExceptionPointerRegister(X86::EAX);
358 setExceptionSelectorRegister(X86::EDX);
360 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
361 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
363 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
365 setOperationAction(ISD::TRAP, MVT::Other, Legal);
367 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
368 setOperationAction(ISD::VASTART , MVT::Other, Custom);
369 setOperationAction(ISD::VAEND , MVT::Other, Expand);
370 if (Subtarget->is64Bit()) {
371 setOperationAction(ISD::VAARG , MVT::Other, Custom);
372 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
374 setOperationAction(ISD::VAARG , MVT::Other, Expand);
375 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
378 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
379 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
380 if (Subtarget->is64Bit())
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
382 if (Subtarget->isTargetCygMing())
383 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
385 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
387 if (!UseSoftFloat && X86ScalarSSEf64) {
388 // f32 and f64 use SSE.
389 // Set up the FP register classes.
390 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
391 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
393 // Use ANDPD to simulate FABS.
394 setOperationAction(ISD::FABS , MVT::f64, Custom);
395 setOperationAction(ISD::FABS , MVT::f32, Custom);
397 // Use XORP to simulate FNEG.
398 setOperationAction(ISD::FNEG , MVT::f64, Custom);
399 setOperationAction(ISD::FNEG , MVT::f32, Custom);
401 // Use ANDPD and ORPD to simulate FCOPYSIGN.
402 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
403 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
405 // We don't support sin/cos/fmod
406 setOperationAction(ISD::FSIN , MVT::f64, Expand);
407 setOperationAction(ISD::FCOS , MVT::f64, Expand);
408 setOperationAction(ISD::FSIN , MVT::f32, Expand);
409 setOperationAction(ISD::FCOS , MVT::f32, Expand);
411 // Expand FP immediates into loads from the stack, except for the special
413 addLegalFPImmediate(APFloat(+0.0)); // xorpd
414 addLegalFPImmediate(APFloat(+0.0f)); // xorps
416 // Floating truncations from f80 and extensions to f80 go through memory.
417 // If optimizing, we lie about this though and handle it in
418 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
420 setConvertAction(MVT::f32, MVT::f80, Expand);
421 setConvertAction(MVT::f64, MVT::f80, Expand);
422 setConvertAction(MVT::f80, MVT::f32, Expand);
423 setConvertAction(MVT::f80, MVT::f64, Expand);
425 } else if (!UseSoftFloat && X86ScalarSSEf32) {
426 // Use SSE for f32, x87 for f64.
427 // Set up the FP register classes.
428 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
429 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
431 // Use ANDPS to simulate FABS.
432 setOperationAction(ISD::FABS , MVT::f32, Custom);
434 // Use XORP to simulate FNEG.
435 setOperationAction(ISD::FNEG , MVT::f32, Custom);
437 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
439 // Use ANDPS and ORPS to simulate FCOPYSIGN.
440 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
441 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
443 // We don't support sin/cos/fmod
444 setOperationAction(ISD::FSIN , MVT::f32, Expand);
445 setOperationAction(ISD::FCOS , MVT::f32, Expand);
447 // Special cases we handle for FP constants.
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 addLegalFPImmediate(APFloat(+0.0)); // FLD0
450 addLegalFPImmediate(APFloat(+1.0)); // FLD1
451 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
452 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
454 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
455 // this though and handle it in InstructionSelectPreprocess so that
456 // dagcombine2 can hack on these.
458 setConvertAction(MVT::f32, MVT::f64, Expand);
459 setConvertAction(MVT::f32, MVT::f80, Expand);
460 setConvertAction(MVT::f80, MVT::f32, Expand);
461 setConvertAction(MVT::f64, MVT::f32, Expand);
462 // And x87->x87 truncations also.
463 setConvertAction(MVT::f80, MVT::f64, Expand);
467 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
468 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
470 } else if (!UseSoftFloat) {
471 // f32 and f64 in x87.
472 // Set up the FP register classes.
473 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
474 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
476 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
477 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
478 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
479 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
481 // Floating truncations go through memory. If optimizing, we lie about
482 // this though and handle it in InstructionSelectPreprocess so that
483 // dagcombine2 can hack on these.
485 setConvertAction(MVT::f80, MVT::f32, Expand);
486 setConvertAction(MVT::f64, MVT::f32, Expand);
487 setConvertAction(MVT::f80, MVT::f64, Expand);
491 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
492 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
494 addLegalFPImmediate(APFloat(+0.0)); // FLD0
495 addLegalFPImmediate(APFloat(+1.0)); // FLD1
496 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
497 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
498 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
504 // Long double always uses X87.
506 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
507 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
508 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
511 APFloat TmpFlt(+0.0);
512 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
514 addLegalFPImmediate(TmpFlt); // FLD0
516 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
517 APFloat TmpFlt2(+1.0);
518 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
520 addLegalFPImmediate(TmpFlt2); // FLD1
521 TmpFlt2.changeSign();
522 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
526 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
527 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
531 // Always use a library call for pow.
532 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
533 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
534 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
536 setOperationAction(ISD::FLOG, MVT::f80, Expand);
537 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
538 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
539 setOperationAction(ISD::FEXP, MVT::f80, Expand);
540 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
542 // First set operation action for all vector types to either promote
543 // (for widening) or expand (for scalarization). Then we will selectively
544 // turn on ones that can be effectively codegen'd.
545 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
546 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
547 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
562 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
592 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
593 // with -msoft-float, disable use of MMX as well.
594 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
595 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
596 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
597 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
598 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
599 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
601 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
602 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
603 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
604 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
606 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
607 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
608 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
609 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
611 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
612 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
614 setOperationAction(ISD::AND, MVT::v8i8, Promote);
615 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
616 setOperationAction(ISD::AND, MVT::v4i16, Promote);
617 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
618 setOperationAction(ISD::AND, MVT::v2i32, Promote);
619 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
620 setOperationAction(ISD::AND, MVT::v1i64, Legal);
622 setOperationAction(ISD::OR, MVT::v8i8, Promote);
623 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
624 setOperationAction(ISD::OR, MVT::v4i16, Promote);
625 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
626 setOperationAction(ISD::OR, MVT::v2i32, Promote);
627 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
628 setOperationAction(ISD::OR, MVT::v1i64, Legal);
630 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
631 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
632 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
633 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
634 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
635 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
636 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
638 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
639 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
640 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
641 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
642 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
643 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
644 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
645 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
646 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
648 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
652 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
656 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
657 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
659 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
660 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
661 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
662 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
664 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
666 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
667 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
668 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
669 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
670 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
671 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
674 if (!UseSoftFloat && Subtarget->hasSSE1()) {
675 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
677 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
678 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
679 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
680 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
681 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
682 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
683 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
684 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
686 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
687 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
691 if (!UseSoftFloat && Subtarget->hasSSE2()) {
692 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
694 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
695 // registers cannot be used even for integer operations.
696 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
697 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
698 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
699 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
701 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
702 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
703 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
704 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
705 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
706 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
707 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
708 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
709 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
710 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
711 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
712 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
713 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
714 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
715 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
716 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
720 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
721 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
723 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
724 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
726 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
727 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
729 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
730 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
731 MVT VT = (MVT::SimpleValueType)i;
732 // Do not attempt to custom lower non-power-of-2 vectors
733 if (!isPowerOf2_32(VT.getVectorNumElements()))
735 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
736 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
737 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
740 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
741 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
742 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
743 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
744 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
745 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
747 if (Subtarget->is64Bit()) {
748 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
749 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
752 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
753 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
754 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
755 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
756 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
757 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
758 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
759 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
760 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
761 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
762 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
763 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
766 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
768 // Custom lower v2i64 and v2f64 selects.
769 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
770 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
771 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
772 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
776 if (Subtarget->hasSSE41()) {
777 // FIXME: Do we need to handle scalar-to-vector here?
778 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
780 // i8 and i16 vectors are custom , because the source register and source
781 // source memory operand types are not the same width. f32 vectors are
782 // custom since the immediate controlling the insert encodes additional
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
787 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
794 if (Subtarget->is64Bit()) {
795 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
800 if (Subtarget->hasSSE42()) {
801 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
804 // We want to custom lower some of our intrinsics.
805 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
807 // Add/Sub/Mul with overflow operations are custom lowered.
808 setOperationAction(ISD::SADDO, MVT::i32, Custom);
809 setOperationAction(ISD::SADDO, MVT::i64, Custom);
810 setOperationAction(ISD::UADDO, MVT::i32, Custom);
811 setOperationAction(ISD::UADDO, MVT::i64, Custom);
812 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
813 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
814 setOperationAction(ISD::USUBO, MVT::i32, Custom);
815 setOperationAction(ISD::USUBO, MVT::i64, Custom);
816 setOperationAction(ISD::SMULO, MVT::i32, Custom);
817 setOperationAction(ISD::SMULO, MVT::i64, Custom);
818 setOperationAction(ISD::UMULO, MVT::i32, Custom);
819 setOperationAction(ISD::UMULO, MVT::i64, Custom);
821 // We have target-specific dag combine patterns for the following nodes:
822 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
823 setTargetDAGCombine(ISD::BUILD_VECTOR);
824 setTargetDAGCombine(ISD::SELECT);
825 setTargetDAGCombine(ISD::SHL);
826 setTargetDAGCombine(ISD::SRA);
827 setTargetDAGCombine(ISD::SRL);
828 setTargetDAGCombine(ISD::STORE);
829 if (Subtarget->is64Bit())
830 setTargetDAGCombine(ISD::MUL);
832 computeRegisterProperties();
834 // FIXME: These should be based on subtarget info. Plus, the values should
835 // be smaller when we are in optimizing for size mode.
836 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
837 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
838 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
839 allowUnalignedMemoryAccesses = true; // x86 supports it!
840 setPrefLoopAlignment(16);
844 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
849 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
850 /// the desired ByVal argument alignment.
851 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
854 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
855 if (VTy->getBitWidth() == 128)
857 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
858 unsigned EltAlign = 0;
859 getMaxByValAlign(ATy->getElementType(), EltAlign);
860 if (EltAlign > MaxAlign)
862 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
863 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
864 unsigned EltAlign = 0;
865 getMaxByValAlign(STy->getElementType(i), EltAlign);
866 if (EltAlign > MaxAlign)
875 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
876 /// function arguments in the caller parameter area. For X86, aggregates
877 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
878 /// are at 4-byte boundaries.
879 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
880 if (Subtarget->is64Bit()) {
881 // Max of 8 and alignment of type.
882 unsigned TyAlign = TD->getABITypeAlignment(Ty);
889 if (Subtarget->hasSSE1())
890 getMaxByValAlign(Ty, Align);
894 /// getOptimalMemOpType - Returns the target specific optimal type for load
895 /// and store operations as a result of memset, memcpy, and memmove
896 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
899 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
900 bool isSrcConst, bool isSrcStr) const {
901 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
902 // linux. This is because the stack realignment code can't handle certain
903 // cases like PR2962. This should be removed when PR2962 is fixed.
904 if (!NoImplicitFloat && Subtarget->getStackAlignment() >= 16) {
905 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
907 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
910 if (Subtarget->is64Bit() && Size >= 8)
915 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
917 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
918 SelectionDAG &DAG) const {
919 if (usesGlobalOffsetTable())
920 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
921 if (!Subtarget->isPICStyleRIPRel())
922 // This doesn't have DebugLoc associated with it, but is not really the
923 // same as a Register.
924 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
929 //===----------------------------------------------------------------------===//
930 // Return Value Calling Convention Implementation
931 //===----------------------------------------------------------------------===//
933 #include "X86GenCallingConv.inc"
935 /// LowerRET - Lower an ISD::RET node.
936 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
937 DebugLoc dl = Op.getDebugLoc();
938 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
940 SmallVector<CCValAssign, 16> RVLocs;
941 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
942 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
943 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
944 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
946 // If this is the first return lowered for this function, add the regs to the
947 // liveout set for the function.
948 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
949 for (unsigned i = 0; i != RVLocs.size(); ++i)
950 if (RVLocs[i].isRegLoc())
951 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
953 SDValue Chain = Op.getOperand(0);
955 // Handle tail call return.
956 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
957 if (Chain.getOpcode() == X86ISD::TAILCALL) {
958 SDValue TailCall = Chain;
959 SDValue TargetAddress = TailCall.getOperand(1);
960 SDValue StackAdjustment = TailCall.getOperand(2);
961 assert(((TargetAddress.getOpcode() == ISD::Register &&
962 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
963 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
964 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
965 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
966 "Expecting an global address, external symbol, or register");
967 assert(StackAdjustment.getOpcode() == ISD::Constant &&
968 "Expecting a const value");
970 SmallVector<SDValue,8> Operands;
971 Operands.push_back(Chain.getOperand(0));
972 Operands.push_back(TargetAddress);
973 Operands.push_back(StackAdjustment);
974 // Copy registers used by the call. Last operand is a flag so it is not
976 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
977 Operands.push_back(Chain.getOperand(i));
979 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
986 SmallVector<SDValue, 6> RetOps;
987 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
988 // Operand #1 = Bytes To Pop
989 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
991 // Copy the result values into the output registers.
992 for (unsigned i = 0; i != RVLocs.size(); ++i) {
993 CCValAssign &VA = RVLocs[i];
994 assert(VA.isRegLoc() && "Can only return in registers!");
995 SDValue ValToCopy = Op.getOperand(i*2+1);
997 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
998 // the RET instruction and handled by the FP Stackifier.
999 if (VA.getLocReg() == X86::ST0 ||
1000 VA.getLocReg() == X86::ST1) {
1001 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1002 // change the value to the FP stack register class.
1003 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1004 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1005 RetOps.push_back(ValToCopy);
1006 // Don't emit a copytoreg.
1010 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1011 // which is returned in RAX / RDX.
1012 if (Subtarget->is64Bit()) {
1013 MVT ValVT = ValToCopy.getValueType();
1014 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1015 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1016 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1017 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1021 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1022 Flag = Chain.getValue(1);
1025 // The x86-64 ABI for returning structs by value requires that we copy
1026 // the sret argument into %rax for the return. We saved the argument into
1027 // a virtual register in the entry block, so now we copy the value out
1029 if (Subtarget->is64Bit() &&
1030 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1031 MachineFunction &MF = DAG.getMachineFunction();
1032 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1033 unsigned Reg = FuncInfo->getSRetReturnReg();
1035 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1036 FuncInfo->setSRetReturnReg(Reg);
1038 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1040 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1041 Flag = Chain.getValue(1);
1044 RetOps[0] = Chain; // Update chain.
1046 // Add the flag if we have it.
1048 RetOps.push_back(Flag);
1050 return DAG.getNode(X86ISD::RET_FLAG, dl,
1051 MVT::Other, &RetOps[0], RetOps.size());
1055 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1056 /// appropriate copies out of appropriate physical registers. This assumes that
1057 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1058 /// being lowered. The returns a SDNode with the same number of values as the
1060 SDNode *X86TargetLowering::
1061 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1062 unsigned CallingConv, SelectionDAG &DAG) {
1064 DebugLoc dl = TheCall->getDebugLoc();
1065 // Assign locations to each value returned by this call.
1066 SmallVector<CCValAssign, 16> RVLocs;
1067 bool isVarArg = TheCall->isVarArg();
1068 bool Is64Bit = Subtarget->is64Bit();
1069 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1070 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1072 SmallVector<SDValue, 8> ResultVals;
1074 // Copy all of the result registers out of their specified physreg.
1075 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1076 CCValAssign &VA = RVLocs[i];
1077 MVT CopyVT = VA.getValVT();
1079 // If this is x86-64, and we disabled SSE, we can't return FP values
1080 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1081 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1082 cerr << "SSE register return with SSE disabled\n";
1086 // If this is a call to a function that returns an fp value on the floating
1087 // point stack, but where we prefer to use the value in xmm registers, copy
1088 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1089 if ((VA.getLocReg() == X86::ST0 ||
1090 VA.getLocReg() == X86::ST1) &&
1091 isScalarFPTypeInSSEReg(VA.getValVT())) {
1096 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1097 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1098 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1099 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1100 MVT::v2i64, InFlag).getValue(1);
1101 Val = Chain.getValue(0);
1102 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1103 Val, DAG.getConstant(0, MVT::i64));
1105 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1106 MVT::i64, InFlag).getValue(1);
1107 Val = Chain.getValue(0);
1109 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1111 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1112 CopyVT, InFlag).getValue(1);
1113 Val = Chain.getValue(0);
1115 InFlag = Chain.getValue(2);
1117 if (CopyVT != VA.getValVT()) {
1118 // Round the F80 the right size, which also moves to the appropriate xmm
1120 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1121 // This truncation won't change the value.
1122 DAG.getIntPtrConstant(1));
1125 ResultVals.push_back(Val);
1128 // Merge everything together with a MERGE_VALUES node.
1129 ResultVals.push_back(Chain);
1130 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1131 &ResultVals[0], ResultVals.size()).getNode();
1135 //===----------------------------------------------------------------------===//
1136 // C & StdCall & Fast Calling Convention implementation
1137 //===----------------------------------------------------------------------===//
1138 // StdCall calling convention seems to be standard for many Windows' API
1139 // routines and around. It differs from C calling convention just a little:
1140 // callee should clean up the stack, not caller. Symbols should be also
1141 // decorated in some fancy way :) It doesn't support any vector arguments.
1142 // For info on fast calling convention see Fast Calling Convention (tail call)
1143 // implementation LowerX86_32FastCCCallTo.
1145 /// AddLiveIn - This helper function adds the specified physical register to the
1146 /// MachineFunction as a live in value. It also creates a corresponding virtual
1147 /// register for it.
1148 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
1149 const TargetRegisterClass *RC) {
1150 assert(RC->contains(PReg) && "Not the correct regclass!");
1151 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1152 MF.getRegInfo().addLiveIn(PReg, VReg);
1156 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1158 static bool CallIsStructReturn(CallSDNode *TheCall) {
1159 unsigned NumOps = TheCall->getNumArgs();
1163 return TheCall->getArgFlags(0).isSRet();
1166 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1167 /// return semantics.
1168 static bool ArgsAreStructReturn(SDValue Op) {
1169 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1173 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1176 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1177 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1179 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1183 switch (CallingConv) {
1186 case CallingConv::X86_StdCall:
1187 return !Subtarget->is64Bit();
1188 case CallingConv::X86_FastCall:
1189 return !Subtarget->is64Bit();
1190 case CallingConv::Fast:
1191 return PerformTailCallOpt;
1195 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1196 /// given CallingConvention value.
1197 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1198 if (Subtarget->is64Bit()) {
1199 if (Subtarget->isTargetWin64())
1200 return CC_X86_Win64_C;
1201 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1202 return CC_X86_64_TailCall;
1207 if (CC == CallingConv::X86_FastCall)
1208 return CC_X86_32_FastCall;
1209 else if (CC == CallingConv::Fast)
1210 return CC_X86_32_FastCC;
1215 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1216 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1218 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1219 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1220 if (CC == CallingConv::X86_FastCall)
1222 else if (CC == CallingConv::X86_StdCall)
1228 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1229 /// in a register before calling.
1230 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1231 return !IsTailCall && !Is64Bit &&
1232 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1233 Subtarget->isPICStyleGOT();
1236 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1237 /// address to be loaded in a register.
1239 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1240 return !Is64Bit && IsTailCall &&
1241 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1242 Subtarget->isPICStyleGOT();
1245 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1246 /// by "Src" to address "Dst" with size and alignment information specified by
1247 /// the specific parameter attribute. The copy will be passed as a byval
1248 /// function parameter.
1250 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1251 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1253 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1254 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1255 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1258 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1259 const CCValAssign &VA,
1260 MachineFrameInfo *MFI,
1262 SDValue Root, unsigned i) {
1263 // Create the nodes corresponding to a load from this parameter slot.
1264 ISD::ArgFlagsTy Flags =
1265 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1266 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1267 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1269 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1270 // changed with more analysis.
1271 // In case of tail call optimization mark all arguments mutable. Since they
1272 // could be overwritten by lowering of arguments in case of a tail call.
1273 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1274 VA.getLocMemOffset(), isImmutable);
1275 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1276 if (Flags.isByVal())
1278 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1279 PseudoSourceValue::getFixedStack(FI), 0);
1283 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1284 MachineFunction &MF = DAG.getMachineFunction();
1285 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1286 DebugLoc dl = Op.getDebugLoc();
1288 const Function* Fn = MF.getFunction();
1289 if (Fn->hasExternalLinkage() &&
1290 Subtarget->isTargetCygMing() &&
1291 Fn->getName() == "main")
1292 FuncInfo->setForceFramePointer(true);
1294 // Decorate the function name.
1295 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1297 MachineFrameInfo *MFI = MF.getFrameInfo();
1298 SDValue Root = Op.getOperand(0);
1299 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1300 unsigned CC = MF.getFunction()->getCallingConv();
1301 bool Is64Bit = Subtarget->is64Bit();
1302 bool IsWin64 = Subtarget->isTargetWin64();
1304 assert(!(isVarArg && CC == CallingConv::Fast) &&
1305 "Var args not supported with calling convention fastcc");
1307 // Assign locations to all of the incoming arguments.
1308 SmallVector<CCValAssign, 16> ArgLocs;
1309 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1310 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1312 SmallVector<SDValue, 8> ArgValues;
1313 unsigned LastVal = ~0U;
1314 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1315 CCValAssign &VA = ArgLocs[i];
1316 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1318 assert(VA.getValNo() != LastVal &&
1319 "Don't support value assigned to multiple locs yet");
1320 LastVal = VA.getValNo();
1322 if (VA.isRegLoc()) {
1323 MVT RegVT = VA.getLocVT();
1324 TargetRegisterClass *RC = NULL;
1325 if (RegVT == MVT::i32)
1326 RC = X86::GR32RegisterClass;
1327 else if (Is64Bit && RegVT == MVT::i64)
1328 RC = X86::GR64RegisterClass;
1329 else if (RegVT == MVT::f32)
1330 RC = X86::FR32RegisterClass;
1331 else if (RegVT == MVT::f64)
1332 RC = X86::FR64RegisterClass;
1333 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1334 RC = X86::VR128RegisterClass;
1335 else if (RegVT.isVector()) {
1336 assert(RegVT.getSizeInBits() == 64);
1338 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1340 // Darwin calling convention passes MMX values in either GPRs or
1341 // XMMs in x86-64. Other targets pass them in memory.
1342 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1343 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1346 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1351 assert(0 && "Unknown argument type!");
1354 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1355 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1357 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1358 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1360 if (VA.getLocInfo() == CCValAssign::SExt)
1361 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1362 DAG.getValueType(VA.getValVT()));
1363 else if (VA.getLocInfo() == CCValAssign::ZExt)
1364 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1365 DAG.getValueType(VA.getValVT()));
1367 if (VA.getLocInfo() != CCValAssign::Full)
1368 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1370 // Handle MMX values passed in GPRs.
1371 if (Is64Bit && RegVT != VA.getLocVT()) {
1372 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1373 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1374 else if (RC == X86::VR128RegisterClass) {
1375 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1376 ArgValue, DAG.getConstant(0, MVT::i64));
1377 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1381 ArgValues.push_back(ArgValue);
1383 assert(VA.isMemLoc());
1384 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1388 // The x86-64 ABI for returning structs by value requires that we copy
1389 // the sret argument into %rax for the return. Save the argument into
1390 // a virtual register so that we can access it from the return points.
1391 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1392 MachineFunction &MF = DAG.getMachineFunction();
1393 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1394 unsigned Reg = FuncInfo->getSRetReturnReg();
1396 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1397 FuncInfo->setSRetReturnReg(Reg);
1399 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1400 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1403 unsigned StackSize = CCInfo.getNextStackOffset();
1404 // align stack specially for tail calls
1405 if (PerformTailCallOpt && CC == CallingConv::Fast)
1406 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1408 // If the function takes variable number of arguments, make a frame index for
1409 // the start of the first vararg value... for expansion of llvm.va_start.
1411 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1412 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1415 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1417 // FIXME: We should really autogenerate these arrays
1418 static const unsigned GPR64ArgRegsWin64[] = {
1419 X86::RCX, X86::RDX, X86::R8, X86::R9
1421 static const unsigned XMMArgRegsWin64[] = {
1422 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1424 static const unsigned GPR64ArgRegs64Bit[] = {
1425 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1427 static const unsigned XMMArgRegs64Bit[] = {
1428 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1429 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1431 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1434 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1435 GPR64ArgRegs = GPR64ArgRegsWin64;
1436 XMMArgRegs = XMMArgRegsWin64;
1438 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1439 GPR64ArgRegs = GPR64ArgRegs64Bit;
1440 XMMArgRegs = XMMArgRegs64Bit;
1442 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1444 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1447 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1448 "SSE register cannot be used when SSE is disabled!");
1449 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloat) &&
1450 "SSE register cannot be used when SSE is disabled!");
1451 if (UseSoftFloat || NoImplicitFloat || !Subtarget->hasSSE1())
1452 // Kernel mode asks for SSE to be disabled, so don't push them
1454 TotalNumXMMRegs = 0;
1456 // For X86-64, if there are vararg parameters that are passed via
1457 // registers, then we must store them to their spots on the stack so they
1458 // may be loaded by deferencing the result of va_next.
1459 VarArgsGPOffset = NumIntRegs * 8;
1460 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1461 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1462 TotalNumXMMRegs * 16, 16);
1464 // Store the integer parameter registers.
1465 SmallVector<SDValue, 8> MemOps;
1466 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1467 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1468 DAG.getIntPtrConstant(VarArgsGPOffset));
1469 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1470 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1471 X86::GR64RegisterClass);
1472 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1474 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1475 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1476 MemOps.push_back(Store);
1477 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1478 DAG.getIntPtrConstant(8));
1481 // Now store the XMM (fp + vector) parameter registers.
1482 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1483 DAG.getIntPtrConstant(VarArgsFPOffset));
1484 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1485 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1486 X86::VR128RegisterClass);
1487 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1489 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1490 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1491 MemOps.push_back(Store);
1492 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1493 DAG.getIntPtrConstant(16));
1495 if (!MemOps.empty())
1496 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1497 &MemOps[0], MemOps.size());
1501 ArgValues.push_back(Root);
1503 // Some CCs need callee pop.
1504 if (IsCalleePop(isVarArg, CC)) {
1505 BytesToPopOnReturn = StackSize; // Callee pops everything.
1506 BytesCallerReserves = 0;
1508 BytesToPopOnReturn = 0; // Callee pops nothing.
1509 // If this is an sret function, the return should pop the hidden pointer.
1510 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1511 BytesToPopOnReturn = 4;
1512 BytesCallerReserves = StackSize;
1516 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1517 if (CC == CallingConv::X86_FastCall)
1518 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1521 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1523 // Return the new list of results.
1524 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1525 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1529 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1530 const SDValue &StackPtr,
1531 const CCValAssign &VA,
1533 SDValue Arg, ISD::ArgFlagsTy Flags) {
1534 DebugLoc dl = TheCall->getDebugLoc();
1535 unsigned LocMemOffset = VA.getLocMemOffset();
1536 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1537 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1538 if (Flags.isByVal()) {
1539 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1541 return DAG.getStore(Chain, dl, Arg, PtrOff,
1542 PseudoSourceValue::getStack(), LocMemOffset);
1545 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1546 /// optimization is performed and it is required.
1548 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1549 SDValue &OutRetAddr,
1555 if (!IsTailCall || FPDiff==0) return Chain;
1557 // Adjust the Return address stack slot.
1558 MVT VT = getPointerTy();
1559 OutRetAddr = getReturnAddressFrameIndex(DAG);
1561 // Load the "old" Return address.
1562 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1563 return SDValue(OutRetAddr.getNode(), 1);
1566 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1567 /// optimization is performed and it is required (FPDiff!=0).
1569 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1570 SDValue Chain, SDValue RetAddrFrIdx,
1571 bool Is64Bit, int FPDiff, DebugLoc dl) {
1572 // Store the return address to the appropriate stack slot.
1573 if (!FPDiff) return Chain;
1574 // Calculate the new stack slot for the return address.
1575 int SlotSize = Is64Bit ? 8 : 4;
1576 int NewReturnAddrFI =
1577 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1578 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1579 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1580 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1581 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1585 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1586 MachineFunction &MF = DAG.getMachineFunction();
1587 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1588 SDValue Chain = TheCall->getChain();
1589 unsigned CC = TheCall->getCallingConv();
1590 bool isVarArg = TheCall->isVarArg();
1591 bool IsTailCall = TheCall->isTailCall() &&
1592 CC == CallingConv::Fast && PerformTailCallOpt;
1593 SDValue Callee = TheCall->getCallee();
1594 bool Is64Bit = Subtarget->is64Bit();
1595 bool IsStructRet = CallIsStructReturn(TheCall);
1596 DebugLoc dl = TheCall->getDebugLoc();
1598 assert(!(isVarArg && CC == CallingConv::Fast) &&
1599 "Var args not supported with calling convention fastcc");
1601 // Analyze operands of the call, assigning locations to each operand.
1602 SmallVector<CCValAssign, 16> ArgLocs;
1603 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1604 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1606 // Get a count of how many bytes are to be pushed on the stack.
1607 unsigned NumBytes = CCInfo.getNextStackOffset();
1608 if (PerformTailCallOpt && CC == CallingConv::Fast)
1609 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1613 // Lower arguments at fp - stackoffset + fpdiff.
1614 unsigned NumBytesCallerPushed =
1615 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1616 FPDiff = NumBytesCallerPushed - NumBytes;
1618 // Set the delta of movement of the returnaddr stackslot.
1619 // But only set if delta is greater than previous delta.
1620 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1621 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1624 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1626 SDValue RetAddrFrIdx;
1627 // Load return adress for tail calls.
1628 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1631 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1632 SmallVector<SDValue, 8> MemOpChains;
1635 // Walk the register/memloc assignments, inserting copies/loads. In the case
1636 // of tail call optimization arguments are handle later.
1637 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1638 CCValAssign &VA = ArgLocs[i];
1639 SDValue Arg = TheCall->getArg(i);
1640 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1641 bool isByVal = Flags.isByVal();
1643 // Promote the value if needed.
1644 switch (VA.getLocInfo()) {
1645 default: assert(0 && "Unknown loc info!");
1646 case CCValAssign::Full: break;
1647 case CCValAssign::SExt:
1648 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1650 case CCValAssign::ZExt:
1651 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1653 case CCValAssign::AExt:
1654 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1658 if (VA.isRegLoc()) {
1660 MVT RegVT = VA.getLocVT();
1661 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1662 switch (VA.getLocReg()) {
1665 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1667 // Special case: passing MMX values in GPR registers.
1668 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1671 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1672 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1673 // Special case: passing MMX values in XMM registers.
1674 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1675 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1676 Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
1677 DAG.getUNDEF(MVT::v2i64), Arg,
1678 getMOVLMask(2, DAG, dl));
1683 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1685 if (!IsTailCall || (IsTailCall && isByVal)) {
1686 assert(VA.isMemLoc());
1687 if (StackPtr.getNode() == 0)
1688 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1690 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1691 Chain, Arg, Flags));
1696 if (!MemOpChains.empty())
1697 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1698 &MemOpChains[0], MemOpChains.size());
1700 // Build a sequence of copy-to-reg nodes chained together with token chain
1701 // and flag operands which copy the outgoing args into registers.
1703 // Tail call byval lowering might overwrite argument registers so in case of
1704 // tail call optimization the copies to registers are lowered later.
1706 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1707 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1708 RegsToPass[i].second, InFlag);
1709 InFlag = Chain.getValue(1);
1712 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1714 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1715 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1716 DAG.getNode(X86ISD::GlobalBaseReg,
1717 DebugLoc::getUnknownLoc(),
1720 InFlag = Chain.getValue(1);
1722 // If we are tail calling and generating PIC/GOT style code load the address
1723 // of the callee into ecx. The value in ecx is used as target of the tail
1724 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1725 // calls on PIC/GOT architectures. Normally we would just put the address of
1726 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1727 // restored (since ebx is callee saved) before jumping to the target@PLT.
1728 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1729 // Note: The actual moving to ecx is done further down.
1730 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1731 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1732 !G->getGlobal()->hasProtectedVisibility())
1733 Callee = LowerGlobalAddress(Callee, DAG);
1734 else if (isa<ExternalSymbolSDNode>(Callee))
1735 Callee = LowerExternalSymbol(Callee,DAG);
1738 if (Is64Bit && isVarArg) {
1739 // From AMD64 ABI document:
1740 // For calls that may call functions that use varargs or stdargs
1741 // (prototype-less calls or calls to functions containing ellipsis (...) in
1742 // the declaration) %al is used as hidden argument to specify the number
1743 // of SSE registers used. The contents of %al do not need to match exactly
1744 // the number of registers, but must be an ubound on the number of SSE
1745 // registers used and is in the range 0 - 8 inclusive.
1747 // FIXME: Verify this on Win64
1748 // Count the number of XMM registers allocated.
1749 static const unsigned XMMArgRegs[] = {
1750 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1751 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1753 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1754 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1755 && "SSE registers cannot be used when SSE is disabled");
1757 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1758 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1759 InFlag = Chain.getValue(1);
1763 // For tail calls lower the arguments to the 'real' stack slot.
1765 SmallVector<SDValue, 8> MemOpChains2;
1768 // Do not flag preceeding copytoreg stuff together with the following stuff.
1770 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1771 CCValAssign &VA = ArgLocs[i];
1772 if (!VA.isRegLoc()) {
1773 assert(VA.isMemLoc());
1774 SDValue Arg = TheCall->getArg(i);
1775 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1776 // Create frame index.
1777 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1778 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1779 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1780 FIN = DAG.getFrameIndex(FI, getPointerTy());
1782 if (Flags.isByVal()) {
1783 // Copy relative to framepointer.
1784 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1785 if (StackPtr.getNode() == 0)
1786 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1788 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1790 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1793 // Store relative to framepointer.
1794 MemOpChains2.push_back(
1795 DAG.getStore(Chain, dl, Arg, FIN,
1796 PseudoSourceValue::getFixedStack(FI), 0));
1801 if (!MemOpChains2.empty())
1802 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1803 &MemOpChains2[0], MemOpChains2.size());
1805 // Copy arguments to their registers.
1806 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1807 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1808 RegsToPass[i].second, InFlag);
1809 InFlag = Chain.getValue(1);
1813 // Store the return address to the appropriate stack slot.
1814 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1818 // If the callee is a GlobalAddress node (quite common, every direct call is)
1819 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1820 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1821 // We should use extra load for direct calls to dllimported functions in
1823 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1824 getTargetMachine(), true))
1825 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1827 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1828 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1829 } else if (IsTailCall) {
1830 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1832 Chain = DAG.getCopyToReg(Chain, dl,
1833 DAG.getRegister(Opc, getPointerTy()),
1835 Callee = DAG.getRegister(Opc, getPointerTy());
1836 // Add register as live out.
1837 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1840 // Returns a chain & a flag for retval copy to use.
1841 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1842 SmallVector<SDValue, 8> Ops;
1845 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1846 DAG.getIntPtrConstant(0, true), InFlag);
1847 InFlag = Chain.getValue(1);
1849 // Returns a chain & a flag for retval copy to use.
1850 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1854 Ops.push_back(Chain);
1855 Ops.push_back(Callee);
1858 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1860 // Add argument registers to the end of the list so that they are known live
1862 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1863 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1864 RegsToPass[i].second.getValueType()));
1866 // Add an implicit use GOT pointer in EBX.
1867 if (!IsTailCall && !Is64Bit &&
1868 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1869 Subtarget->isPICStyleGOT())
1870 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1872 // Add an implicit use of AL for x86 vararg functions.
1873 if (Is64Bit && isVarArg)
1874 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1876 if (InFlag.getNode())
1877 Ops.push_back(InFlag);
1880 assert(InFlag.getNode() &&
1881 "Flag must be set. Depend on flag being set in LowerRET");
1882 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1883 TheCall->getVTList(), &Ops[0], Ops.size());
1885 return SDValue(Chain.getNode(), Op.getResNo());
1888 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1889 InFlag = Chain.getValue(1);
1891 // Create the CALLSEQ_END node.
1892 unsigned NumBytesForCalleeToPush;
1893 if (IsCalleePop(isVarArg, CC))
1894 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1895 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1896 // If this is is a call to a struct-return function, the callee
1897 // pops the hidden struct pointer, so we have to push it back.
1898 // This is common for Darwin/X86, Linux & Mingw32 targets.
1899 NumBytesForCalleeToPush = 4;
1901 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1903 // Returns a flag for retval copy to use.
1904 Chain = DAG.getCALLSEQ_END(Chain,
1905 DAG.getIntPtrConstant(NumBytes, true),
1906 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1909 InFlag = Chain.getValue(1);
1911 // Handle result values, copying them out of physregs into vregs that we
1913 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1918 //===----------------------------------------------------------------------===//
1919 // Fast Calling Convention (tail call) implementation
1920 //===----------------------------------------------------------------------===//
1922 // Like std call, callee cleans arguments, convention except that ECX is
1923 // reserved for storing the tail called function address. Only 2 registers are
1924 // free for argument passing (inreg). Tail call optimization is performed
1926 // * tailcallopt is enabled
1927 // * caller/callee are fastcc
1928 // On X86_64 architecture with GOT-style position independent code only local
1929 // (within module) calls are supported at the moment.
1930 // To keep the stack aligned according to platform abi the function
1931 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1932 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1933 // If a tail called function callee has more arguments than the caller the
1934 // caller needs to make sure that there is room to move the RETADDR to. This is
1935 // achieved by reserving an area the size of the argument delta right after the
1936 // original REtADDR, but before the saved framepointer or the spilled registers
1937 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1949 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1950 /// for a 16 byte align requirement.
1951 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1952 SelectionDAG& DAG) {
1953 MachineFunction &MF = DAG.getMachineFunction();
1954 const TargetMachine &TM = MF.getTarget();
1955 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1956 unsigned StackAlignment = TFI.getStackAlignment();
1957 uint64_t AlignMask = StackAlignment - 1;
1958 int64_t Offset = StackSize;
1959 uint64_t SlotSize = TD->getPointerSize();
1960 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1961 // Number smaller than 12 so just add the difference.
1962 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1964 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1965 Offset = ((~AlignMask) & Offset) + StackAlignment +
1966 (StackAlignment-SlotSize);
1971 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1972 /// following the call is a return. A function is eligible if caller/callee
1973 /// calling conventions match, currently only fastcc supports tail calls, and
1974 /// the function CALL is immediatly followed by a RET.
1975 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1977 SelectionDAG& DAG) const {
1978 if (!PerformTailCallOpt)
1981 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1982 MachineFunction &MF = DAG.getMachineFunction();
1983 unsigned CallerCC = MF.getFunction()->getCallingConv();
1984 unsigned CalleeCC= TheCall->getCallingConv();
1985 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1986 SDValue Callee = TheCall->getCallee();
1987 // On x86/32Bit PIC/GOT tail calls are supported.
1988 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1989 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1992 // Can only do local tail calls (in same module, hidden or protected) on
1993 // x86_64 PIC/GOT at the moment.
1994 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1995 return G->getGlobal()->hasHiddenVisibility()
1996 || G->getGlobal()->hasProtectedVisibility();
2004 X86TargetLowering::createFastISel(MachineFunction &mf,
2005 MachineModuleInfo *mmo,
2007 DenseMap<const Value *, unsigned> &vm,
2008 DenseMap<const BasicBlock *,
2009 MachineBasicBlock *> &bm,
2010 DenseMap<const AllocaInst *, int> &am
2012 , SmallSet<Instruction*, 8> &cil
2015 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2023 //===----------------------------------------------------------------------===//
2024 // Other Lowering Hooks
2025 //===----------------------------------------------------------------------===//
2028 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2029 MachineFunction &MF = DAG.getMachineFunction();
2030 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2031 int ReturnAddrIndex = FuncInfo->getRAIndex();
2033 if (ReturnAddrIndex == 0) {
2034 // Set up a frame object for the return address.
2035 uint64_t SlotSize = TD->getPointerSize();
2036 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2037 FuncInfo->setRAIndex(ReturnAddrIndex);
2040 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2044 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2045 /// specific condition code, returning the condition code and the LHS/RHS of the
2046 /// comparison to make.
2047 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2048 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2050 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2051 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2052 // X > -1 -> X == 0, jump !sign.
2053 RHS = DAG.getConstant(0, RHS.getValueType());
2054 return X86::COND_NS;
2055 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2056 // X < 0 -> X == 0, jump on sign.
2058 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2060 RHS = DAG.getConstant(0, RHS.getValueType());
2061 return X86::COND_LE;
2065 switch (SetCCOpcode) {
2066 default: assert(0 && "Invalid integer condition!");
2067 case ISD::SETEQ: return X86::COND_E;
2068 case ISD::SETGT: return X86::COND_G;
2069 case ISD::SETGE: return X86::COND_GE;
2070 case ISD::SETLT: return X86::COND_L;
2071 case ISD::SETLE: return X86::COND_LE;
2072 case ISD::SETNE: return X86::COND_NE;
2073 case ISD::SETULT: return X86::COND_B;
2074 case ISD::SETUGT: return X86::COND_A;
2075 case ISD::SETULE: return X86::COND_BE;
2076 case ISD::SETUGE: return X86::COND_AE;
2080 // First determine if it is required or is profitable to flip the operands.
2082 // If LHS is a foldable load, but RHS is not, flip the condition.
2083 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2084 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2085 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2086 std::swap(LHS, RHS);
2089 switch (SetCCOpcode) {
2095 std::swap(LHS, RHS);
2099 // On a floating point condition, the flags are set as follows:
2101 // 0 | 0 | 0 | X > Y
2102 // 0 | 0 | 1 | X < Y
2103 // 1 | 0 | 0 | X == Y
2104 // 1 | 1 | 1 | unordered
2105 switch (SetCCOpcode) {
2106 default: assert(0 && "Condcode should be pre-legalized away");
2108 case ISD::SETEQ: return X86::COND_E;
2109 case ISD::SETOLT: // flipped
2111 case ISD::SETGT: return X86::COND_A;
2112 case ISD::SETOLE: // flipped
2114 case ISD::SETGE: return X86::COND_AE;
2115 case ISD::SETUGT: // flipped
2117 case ISD::SETLT: return X86::COND_B;
2118 case ISD::SETUGE: // flipped
2120 case ISD::SETLE: return X86::COND_BE;
2122 case ISD::SETNE: return X86::COND_NE;
2123 case ISD::SETUO: return X86::COND_P;
2124 case ISD::SETO: return X86::COND_NP;
2128 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2129 /// code. Current x86 isa includes the following FP cmov instructions:
2130 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2131 static bool hasFPCMov(unsigned X86CC) {
2147 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2148 /// true if Op is undef or if its value falls within the specified range (L, H].
2149 static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
2150 if (Op.getOpcode() == ISD::UNDEF)
2153 unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
2154 return (Val >= Low && Val < Hi);
2157 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2158 /// true if Op is undef or if its value equal to the specified value.
2159 static bool isUndefOrEqual(SDValue Op, unsigned Val) {
2160 if (Op.getOpcode() == ISD::UNDEF)
2162 return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
2165 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2166 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2167 bool X86::isPSHUFDMask(SDNode *N) {
2168 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2170 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2173 // Check if the value doesn't reference the second vector.
2174 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2175 SDValue Arg = N->getOperand(i);
2176 if (Arg.getOpcode() == ISD::UNDEF) continue;
2177 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2178 if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
2185 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2186 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2187 bool X86::isPSHUFHWMask(SDNode *N) {
2188 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2190 if (N->getNumOperands() != 8)
2193 // Lower quadword copied in order.
2194 for (unsigned i = 0; i != 4; ++i) {
2195 SDValue Arg = N->getOperand(i);
2196 if (Arg.getOpcode() == ISD::UNDEF) continue;
2197 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2198 if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
2202 // Upper quadword shuffled.
2203 for (unsigned i = 4; i != 8; ++i) {
2204 SDValue Arg = N->getOperand(i);
2205 if (Arg.getOpcode() == ISD::UNDEF) continue;
2206 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2207 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2208 if (Val < 4 || Val > 7)
2215 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2216 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2217 bool X86::isPSHUFLWMask(SDNode *N) {
2218 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2220 if (N->getNumOperands() != 8)
2223 // Upper quadword copied in order.
2224 for (unsigned i = 4; i != 8; ++i)
2225 if (!isUndefOrEqual(N->getOperand(i), i))
2228 // Lower quadword shuffled.
2229 for (unsigned i = 0; i != 4; ++i)
2230 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2236 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2237 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2238 template<class SDOperand>
2239 static bool isSHUFPMask(SDOperand *Elems, unsigned NumElems) {
2240 if (NumElems != 2 && NumElems != 4) return false;
2242 unsigned Half = NumElems / 2;
2243 for (unsigned i = 0; i < Half; ++i)
2244 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2246 for (unsigned i = Half; i < NumElems; ++i)
2247 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2253 bool X86::isSHUFPMask(SDNode *N) {
2254 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2255 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2258 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2259 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2260 /// half elements to come from vector 1 (which would equal the dest.) and
2261 /// the upper half to come from vector 2.
2262 template<class SDOperand>
2263 static bool isCommutedSHUFP(SDOperand *Ops, unsigned NumOps) {
2264 if (NumOps != 2 && NumOps != 4) return false;
2266 unsigned Half = NumOps / 2;
2267 for (unsigned i = 0; i < Half; ++i)
2268 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2270 for (unsigned i = Half; i < NumOps; ++i)
2271 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2276 static bool isCommutedSHUFP(SDNode *N) {
2277 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2278 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2281 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2282 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2283 bool X86::isMOVHLPSMask(SDNode *N) {
2284 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2286 if (N->getNumOperands() != 4)
2289 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2290 return isUndefOrEqual(N->getOperand(0), 6) &&
2291 isUndefOrEqual(N->getOperand(1), 7) &&
2292 isUndefOrEqual(N->getOperand(2), 2) &&
2293 isUndefOrEqual(N->getOperand(3), 3);
2296 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2297 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2299 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2300 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2302 if (N->getNumOperands() != 4)
2305 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2306 return isUndefOrEqual(N->getOperand(0), 2) &&
2307 isUndefOrEqual(N->getOperand(1), 3) &&
2308 isUndefOrEqual(N->getOperand(2), 2) &&
2309 isUndefOrEqual(N->getOperand(3), 3);
2312 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2313 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2314 bool X86::isMOVLPMask(SDNode *N) {
2315 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2317 unsigned NumElems = N->getNumOperands();
2318 if (NumElems != 2 && NumElems != 4)
2321 for (unsigned i = 0; i < NumElems/2; ++i)
2322 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2325 for (unsigned i = NumElems/2; i < NumElems; ++i)
2326 if (!isUndefOrEqual(N->getOperand(i), i))
2332 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2333 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2335 bool X86::isMOVHPMask(SDNode *N) {
2336 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2338 unsigned NumElems = N->getNumOperands();
2339 if (NumElems != 2 && NumElems != 4)
2342 for (unsigned i = 0; i < NumElems/2; ++i)
2343 if (!isUndefOrEqual(N->getOperand(i), i))
2346 for (unsigned i = 0; i < NumElems/2; ++i) {
2347 SDValue Arg = N->getOperand(i + NumElems/2);
2348 if (!isUndefOrEqual(Arg, i + NumElems))
2355 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2356 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2357 template<class SDOperand>
2358 bool static isUNPCKLMask(SDOperand *Elts, unsigned NumElts,
2359 bool V2IsSplat = false) {
2360 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2363 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2364 SDValue BitI = Elts[i];
2365 SDValue BitI1 = Elts[i+1];
2366 if (!isUndefOrEqual(BitI, j))
2369 if (!isUndefOrEqual(BitI1, NumElts))
2372 if (!isUndefOrEqual(BitI1, j + NumElts))
2380 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2381 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2382 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2385 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2386 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2387 template<class SDOperand>
2388 bool static isUNPCKHMask(SDOperand *Elts, unsigned NumElts,
2389 bool V2IsSplat = false) {
2390 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2393 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2394 SDValue BitI = Elts[i];
2395 SDValue BitI1 = Elts[i+1];
2396 if (!isUndefOrEqual(BitI, j + NumElts/2))
2399 if (isUndefOrEqual(BitI1, NumElts))
2402 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2410 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2411 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2412 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2415 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2416 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2418 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2419 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2421 unsigned NumElems = N->getNumOperands();
2422 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2425 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2426 SDValue BitI = N->getOperand(i);
2427 SDValue BitI1 = N->getOperand(i+1);
2429 if (!isUndefOrEqual(BitI, j))
2431 if (!isUndefOrEqual(BitI1, j))
2438 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2439 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2441 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2442 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2444 unsigned NumElems = N->getNumOperands();
2445 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2448 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2449 SDValue BitI = N->getOperand(i);
2450 SDValue BitI1 = N->getOperand(i + 1);
2452 if (!isUndefOrEqual(BitI, j))
2454 if (!isUndefOrEqual(BitI1, j))
2461 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2462 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2463 /// MOVSD, and MOVD, i.e. setting the lowest element.
2464 template<class SDOperand>
2465 static bool isMOVLMask(SDOperand *Elts, unsigned NumElts) {
2466 if (NumElts != 2 && NumElts != 4)
2469 if (!isUndefOrEqual(Elts[0], NumElts))
2472 for (unsigned i = 1; i < NumElts; ++i) {
2473 if (!isUndefOrEqual(Elts[i], i))
2480 bool X86::isMOVLMask(SDNode *N) {
2481 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2482 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2485 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2486 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2487 /// element of vector 2 and the other elements to come from vector 1 in order.
2488 template<class SDOperand>
2489 static bool isCommutedMOVL(SDOperand *Ops, unsigned NumOps,
2490 bool V2IsSplat = false,
2491 bool V2IsUndef = false) {
2492 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2495 if (!isUndefOrEqual(Ops[0], 0))
2498 for (unsigned i = 1; i < NumOps; ++i) {
2499 SDValue Arg = Ops[i];
2500 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2501 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2502 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2509 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2510 bool V2IsUndef = false) {
2511 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2512 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2513 V2IsSplat, V2IsUndef);
2516 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2517 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2518 bool X86::isMOVSHDUPMask(SDNode *N) {
2519 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2521 if (N->getNumOperands() != 4)
2524 // Expect 1, 1, 3, 3
2525 for (unsigned i = 0; i < 2; ++i) {
2526 SDValue Arg = N->getOperand(i);
2527 if (Arg.getOpcode() == ISD::UNDEF) continue;
2528 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2529 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2530 if (Val != 1) return false;
2534 for (unsigned i = 2; i < 4; ++i) {
2535 SDValue Arg = N->getOperand(i);
2536 if (Arg.getOpcode() == ISD::UNDEF) continue;
2537 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2538 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2539 if (Val != 3) return false;
2543 // Don't use movshdup if it can be done with a shufps.
2547 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2548 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2549 bool X86::isMOVSLDUPMask(SDNode *N) {
2550 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2552 if (N->getNumOperands() != 4)
2555 // Expect 0, 0, 2, 2
2556 for (unsigned i = 0; i < 2; ++i) {
2557 SDValue Arg = N->getOperand(i);
2558 if (Arg.getOpcode() == ISD::UNDEF) continue;
2559 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2560 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2561 if (Val != 0) return false;
2565 for (unsigned i = 2; i < 4; ++i) {
2566 SDValue Arg = N->getOperand(i);
2567 if (Arg.getOpcode() == ISD::UNDEF) continue;
2568 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2569 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2570 if (Val != 2) return false;
2574 // Don't use movshdup if it can be done with a shufps.
2578 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2579 /// specifies a identity operation on the LHS or RHS.
2580 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2581 unsigned NumElems = N->getNumOperands();
2582 for (unsigned i = 0; i < NumElems; ++i)
2583 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2588 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2589 /// a splat of a single element.
2590 static bool isSplatMask(SDNode *N) {
2591 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2593 // This is a splat operation if each element of the permute is the same, and
2594 // if the value doesn't reference the second vector.
2595 unsigned NumElems = N->getNumOperands();
2596 SDValue ElementBase;
2598 for (; i != NumElems; ++i) {
2599 SDValue Elt = N->getOperand(i);
2600 if (isa<ConstantSDNode>(Elt)) {
2606 if (!ElementBase.getNode())
2609 for (; i != NumElems; ++i) {
2610 SDValue Arg = N->getOperand(i);
2611 if (Arg.getOpcode() == ISD::UNDEF) continue;
2612 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2613 if (Arg != ElementBase) return false;
2616 // Make sure it is a splat of the first vector operand.
2617 return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
2620 /// getSplatMaskEltNo - Given a splat mask, return the index to the element
2621 /// we want to splat.
2622 static SDValue getSplatMaskEltNo(SDNode *N) {
2623 assert(isSplatMask(N) && "Not a splat mask");
2624 unsigned NumElems = N->getNumOperands();
2625 SDValue ElementBase;
2627 for (; i != NumElems; ++i) {
2628 SDValue Elt = N->getOperand(i);
2629 if (isa<ConstantSDNode>(Elt))
2632 assert(0 && " No splat value found!");
2637 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2638 /// a splat of a single element and it's a 2 or 4 element mask.
2639 bool X86::isSplatMask(SDNode *N) {
2640 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2642 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2643 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2645 return ::isSplatMask(N);
2648 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2649 /// specifies a splat of zero element.
2650 bool X86::isSplatLoMask(SDNode *N) {
2651 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2653 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2654 if (!isUndefOrEqual(N->getOperand(i), 0))
2659 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2660 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2661 bool X86::isMOVDDUPMask(SDNode *N) {
2662 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2664 unsigned e = N->getNumOperands() / 2;
2665 for (unsigned i = 0; i < e; ++i)
2666 if (!isUndefOrEqual(N->getOperand(i), i))
2668 for (unsigned i = 0; i < e; ++i)
2669 if (!isUndefOrEqual(N->getOperand(e+i), i))
2674 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2675 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2677 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2678 unsigned NumOperands = N->getNumOperands();
2679 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2681 for (unsigned i = 0; i < NumOperands; ++i) {
2683 SDValue Arg = N->getOperand(NumOperands-i-1);
2684 if (Arg.getOpcode() != ISD::UNDEF)
2685 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2686 if (Val >= NumOperands) Val -= NumOperands;
2688 if (i != NumOperands - 1)
2695 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2696 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2698 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2700 // 8 nodes, but we only care about the last 4.
2701 for (unsigned i = 7; i >= 4; --i) {
2703 SDValue Arg = N->getOperand(i);
2704 if (Arg.getOpcode() != ISD::UNDEF) {
2705 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2715 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2716 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2718 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2720 // 8 nodes, but we only care about the first 4.
2721 for (int i = 3; i >= 0; --i) {
2723 SDValue Arg = N->getOperand(i);
2724 if (Arg.getOpcode() != ISD::UNDEF)
2725 Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2734 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2735 /// values in ther permute mask.
2736 static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
2737 SDValue &V2, SDValue &Mask,
2738 SelectionDAG &DAG) {
2739 MVT VT = Op.getValueType();
2740 MVT MaskVT = Mask.getValueType();
2741 MVT EltVT = MaskVT.getVectorElementType();
2742 unsigned NumElems = Mask.getNumOperands();
2743 SmallVector<SDValue, 8> MaskVec;
2744 DebugLoc dl = Op.getDebugLoc();
2746 for (unsigned i = 0; i != NumElems; ++i) {
2747 SDValue Arg = Mask.getOperand(i);
2748 if (Arg.getOpcode() == ISD::UNDEF) {
2749 MaskVec.push_back(DAG.getUNDEF(EltVT));
2752 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2753 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2755 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2757 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2761 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2762 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
2765 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2766 /// the two vector operands have swapped position.
2768 SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG, DebugLoc dl) {
2769 MVT MaskVT = Mask.getValueType();
2770 MVT EltVT = MaskVT.getVectorElementType();
2771 unsigned NumElems = Mask.getNumOperands();
2772 SmallVector<SDValue, 8> MaskVec;
2773 for (unsigned i = 0; i != NumElems; ++i) {
2774 SDValue Arg = Mask.getOperand(i);
2775 if (Arg.getOpcode() == ISD::UNDEF) {
2776 MaskVec.push_back(DAG.getUNDEF(EltVT));
2779 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2780 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2782 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2784 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2786 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
2790 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2791 /// match movhlps. The lower half elements should come from upper half of
2792 /// V1 (and in order), and the upper half elements should come from the upper
2793 /// half of V2 (and in order).
2794 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2795 unsigned NumElems = Mask->getNumOperands();
2798 for (unsigned i = 0, e = 2; i != e; ++i)
2799 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2801 for (unsigned i = 2; i != 4; ++i)
2802 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2807 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2808 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2810 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2811 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2813 N = N->getOperand(0).getNode();
2814 if (!ISD::isNON_EXTLoad(N))
2817 *LD = cast<LoadSDNode>(N);
2821 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2822 /// match movlp{s|d}. The lower half elements should come from lower half of
2823 /// V1 (and in order), and the upper half elements should come from the upper
2824 /// half of V2 (and in order). And since V1 will become the source of the
2825 /// MOVLP, it must be either a vector load or a scalar load to vector.
2826 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2827 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2829 // Is V2 is a vector load, don't do this transformation. We will try to use
2830 // load folding shufps op.
2831 if (ISD::isNON_EXTLoad(V2))
2834 unsigned NumElems = Mask->getNumOperands();
2835 if (NumElems != 2 && NumElems != 4)
2837 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2838 if (!isUndefOrEqual(Mask->getOperand(i), i))
2840 for (unsigned i = NumElems/2; i != NumElems; ++i)
2841 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2846 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2848 static bool isSplatVector(SDNode *N) {
2849 if (N->getOpcode() != ISD::BUILD_VECTOR)
2852 SDValue SplatValue = N->getOperand(0);
2853 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2854 if (N->getOperand(i) != SplatValue)
2859 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2861 static bool isUndefShuffle(SDNode *N) {
2862 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2865 SDValue V1 = N->getOperand(0);
2866 SDValue V2 = N->getOperand(1);
2867 SDValue Mask = N->getOperand(2);
2868 unsigned NumElems = Mask.getNumOperands();
2869 for (unsigned i = 0; i != NumElems; ++i) {
2870 SDValue Arg = Mask.getOperand(i);
2871 if (Arg.getOpcode() != ISD::UNDEF) {
2872 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2873 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2875 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2882 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2884 static inline bool isZeroNode(SDValue Elt) {
2885 return ((isa<ConstantSDNode>(Elt) &&
2886 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2887 (isa<ConstantFPSDNode>(Elt) &&
2888 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2891 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2892 /// to an zero vector.
2893 static bool isZeroShuffle(SDNode *N) {
2894 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2897 SDValue V1 = N->getOperand(0);
2898 SDValue V2 = N->getOperand(1);
2899 SDValue Mask = N->getOperand(2);
2900 unsigned NumElems = Mask.getNumOperands();
2901 for (unsigned i = 0; i != NumElems; ++i) {
2902 SDValue Arg = Mask.getOperand(i);
2903 if (Arg.getOpcode() == ISD::UNDEF)
2906 unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2907 if (Idx < NumElems) {
2908 unsigned Opc = V1.getNode()->getOpcode();
2909 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2911 if (Opc != ISD::BUILD_VECTOR ||
2912 !isZeroNode(V1.getNode()->getOperand(Idx)))
2914 } else if (Idx >= NumElems) {
2915 unsigned Opc = V2.getNode()->getOpcode();
2916 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2918 if (Opc != ISD::BUILD_VECTOR ||
2919 !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
2926 /// getZeroVector - Returns a vector of specified type with all zero elements.
2928 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2930 assert(VT.isVector() && "Expected a vector type");
2932 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2933 // type. This ensures they get CSE'd.
2935 if (VT.getSizeInBits() == 64) { // MMX
2936 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2937 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2938 } else if (HasSSE2) { // SSE2
2939 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2940 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2942 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2943 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2945 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2948 /// getOnesVector - Returns a vector of specified type with all bits set.
2950 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2951 assert(VT.isVector() && "Expected a vector type");
2953 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2954 // type. This ensures they get CSE'd.
2955 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2957 if (VT.getSizeInBits() == 64) // MMX
2958 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2960 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2961 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2965 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2966 /// that point to V2 points to its first element.
2967 static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
2968 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2970 bool Changed = false;
2971 SmallVector<SDValue, 8> MaskVec;
2972 unsigned NumElems = Mask.getNumOperands();
2973 for (unsigned i = 0; i != NumElems; ++i) {
2974 SDValue Arg = Mask.getOperand(i);
2975 if (Arg.getOpcode() != ISD::UNDEF) {
2976 unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
2977 if (Val > NumElems) {
2978 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2982 MaskVec.push_back(Arg);
2986 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getDebugLoc(),
2987 Mask.getValueType(),
2988 &MaskVec[0], MaskVec.size());
2992 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2993 /// operation of specified width.
2994 static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl) {
2995 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2996 MVT BaseVT = MaskVT.getVectorElementType();
2998 SmallVector<SDValue, 8> MaskVec;
2999 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
3000 for (unsigned i = 1; i != NumElems; ++i)
3001 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3002 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3003 &MaskVec[0], MaskVec.size());
3006 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
3007 /// of specified width.
3008 static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG,
3010 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3011 MVT BaseVT = MaskVT.getVectorElementType();
3012 SmallVector<SDValue, 8> MaskVec;
3013 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3014 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3015 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3017 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3018 &MaskVec[0], MaskVec.size());
3021 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3022 /// of specified width.
3023 static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG,
3025 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3026 MVT BaseVT = MaskVT.getVectorElementType();
3027 unsigned Half = NumElems/2;
3028 SmallVector<SDValue, 8> MaskVec;
3029 for (unsigned i = 0; i != Half; ++i) {
3030 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3031 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3033 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3034 &MaskVec[0], MaskVec.size());
3037 /// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
3038 /// element #0 of a vector with the specified index, leaving the rest of the
3039 /// elements in place.
3040 static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
3041 SelectionDAG &DAG, DebugLoc dl) {
3042 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3043 MVT BaseVT = MaskVT.getVectorElementType();
3044 SmallVector<SDValue, 8> MaskVec;
3045 // Element #0 of the result gets the elt we are replacing.
3046 MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
3047 for (unsigned i = 1; i != NumElems; ++i)
3048 MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
3049 return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3050 &MaskVec[0], MaskVec.size());
3053 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3054 static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
3055 MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
3056 MVT VT = Op.getValueType();
3059 SDValue V1 = Op.getOperand(0);
3060 SDValue Mask = Op.getOperand(2);
3061 unsigned MaskNumElems = Mask.getNumOperands();
3062 unsigned NumElems = MaskNumElems;
3063 DebugLoc dl = Op.getDebugLoc();
3064 // Special handling of v4f32 -> v4i32.
3065 if (VT != MVT::v4f32) {
3066 // Find which element we want to splat.
3067 SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
3068 unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
3069 // unpack elements to the correct location
3070 while (NumElems > 4) {
3071 if (EltNo < NumElems/2) {
3072 Mask = getUnpacklMask(MaskNumElems, DAG, dl);
3074 Mask = getUnpackhMask(MaskNumElems, DAG, dl);
3075 EltNo -= NumElems/2;
3077 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1, Mask);
3080 SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
3081 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3084 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3085 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3086 DAG.getUNDEF(PVT), Mask);
3087 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3090 /// isVectorLoad - Returns true if the node is a vector load, a scalar
3091 /// load that's promoted to vector, or a load bitcasted.
3092 static bool isVectorLoad(SDValue Op) {
3093 assert(Op.getValueType().isVector() && "Expected a vector type");
3094 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
3095 Op.getOpcode() == ISD::BIT_CONVERT) {
3096 return isa<LoadSDNode>(Op.getOperand(0));
3098 return isa<LoadSDNode>(Op);
3102 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
3104 static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
3105 SelectionDAG &DAG, bool HasSSE3) {
3106 // If we have sse3 and shuffle has more than one use or input is a load, then
3107 // use movddup. Otherwise, use movlhps.
3108 bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
3109 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
3110 MVT VT = Op.getValueType();
3113 DebugLoc dl = Op.getDebugLoc();
3114 unsigned NumElems = PVT.getVectorNumElements();
3115 if (NumElems == 2) {
3116 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3117 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3119 assert(NumElems == 4);
3120 SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
3121 SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
3122 Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3123 Cst0, Cst1, Cst0, Cst1);
3126 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3127 SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
3128 DAG.getUNDEF(PVT), Mask);
3129 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
3132 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3133 /// vector of zero or undef vector. This produces a shuffle where the low
3134 /// element of V2 is swizzled into the zero/undef vector, landing at element
3135 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3136 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3137 bool isZero, bool HasSSE2,
3138 SelectionDAG &DAG) {
3139 DebugLoc dl = V2.getDebugLoc();
3140 MVT VT = V2.getValueType();
3142 ? getZeroVector(VT, HasSSE2, DAG, dl) : DAG.getUNDEF(VT);
3143 unsigned NumElems = V2.getValueType().getVectorNumElements();
3144 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3145 MVT EVT = MaskVT.getVectorElementType();
3146 SmallVector<SDValue, 16> MaskVec;
3147 for (unsigned i = 0; i != NumElems; ++i)
3148 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3149 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3151 MaskVec.push_back(DAG.getConstant(i, EVT));
3152 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3153 &MaskVec[0], MaskVec.size());
3154 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
3157 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3158 /// a shuffle that is zero.
3160 unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
3161 unsigned NumElems, bool Low,
3162 SelectionDAG &DAG) {
3163 unsigned NumZeros = 0;
3164 for (unsigned i = 0; i < NumElems; ++i) {
3165 unsigned Index = Low ? i : NumElems-i-1;
3166 SDValue Idx = Mask.getOperand(Index);
3167 if (Idx.getOpcode() == ISD::UNDEF) {
3171 SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
3172 if (Elt.getNode() && isZeroNode(Elt))
3180 /// isVectorShift - Returns true if the shuffle can be implemented as a
3181 /// logical left or right shift of a vector.
3182 static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
3183 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3184 unsigned NumElems = Mask.getNumOperands();
3187 unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
3190 NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
3195 bool SeenV1 = false;
3196 bool SeenV2 = false;
3197 for (unsigned i = NumZeros; i < NumElems; ++i) {
3198 unsigned Val = isLeft ? (i - NumZeros) : i;
3199 SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
3200 if (Idx.getOpcode() == ISD::UNDEF)
3202 unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
3203 if (Index < NumElems)
3212 if (SeenV1 && SeenV2)
3215 ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
3221 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3223 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3224 unsigned NumNonZero, unsigned NumZero,
3225 SelectionDAG &DAG, TargetLowering &TLI) {
3229 DebugLoc dl = Op.getDebugLoc();
3232 for (unsigned i = 0; i < 16; ++i) {
3233 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3234 if (ThisIsNonZero && First) {
3236 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3238 V = DAG.getUNDEF(MVT::v8i16);
3243 SDValue ThisElt(0, 0), LastElt(0, 0);
3244 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3245 if (LastIsNonZero) {
3246 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3247 MVT::i16, Op.getOperand(i-1));
3249 if (ThisIsNonZero) {
3250 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3251 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3252 ThisElt, DAG.getConstant(8, MVT::i8));
3254 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3258 if (ThisElt.getNode())
3259 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3260 DAG.getIntPtrConstant(i/2));
3264 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3267 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3269 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3270 unsigned NumNonZero, unsigned NumZero,
3271 SelectionDAG &DAG, TargetLowering &TLI) {
3275 DebugLoc dl = Op.getDebugLoc();
3278 for (unsigned i = 0; i < 8; ++i) {
3279 bool isNonZero = (NonZeros & (1 << i)) != 0;
3283 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3285 V = DAG.getUNDEF(MVT::v8i16);
3288 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3289 MVT::v8i16, V, Op.getOperand(i),
3290 DAG.getIntPtrConstant(i));
3297 /// getVShift - Return a vector logical shift node.
3299 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3300 unsigned NumBits, SelectionDAG &DAG,
3301 const TargetLowering &TLI, DebugLoc dl) {
3302 bool isMMX = VT.getSizeInBits() == 64;
3303 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3304 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3305 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3306 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3307 DAG.getNode(Opc, dl, ShVT, SrcOp,
3308 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3312 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3313 DebugLoc dl = Op.getDebugLoc();
3314 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3315 if (ISD::isBuildVectorAllZeros(Op.getNode())
3316 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3317 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3318 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3319 // eliminated on x86-32 hosts.
3320 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3323 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3324 return getOnesVector(Op.getValueType(), DAG, dl);
3325 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3328 MVT VT = Op.getValueType();
3329 MVT EVT = VT.getVectorElementType();
3330 unsigned EVTBits = EVT.getSizeInBits();
3332 unsigned NumElems = Op.getNumOperands();
3333 unsigned NumZero = 0;
3334 unsigned NumNonZero = 0;
3335 unsigned NonZeros = 0;
3336 bool IsAllConstants = true;
3337 SmallSet<SDValue, 8> Values;
3338 for (unsigned i = 0; i < NumElems; ++i) {
3339 SDValue Elt = Op.getOperand(i);
3340 if (Elt.getOpcode() == ISD::UNDEF)
3343 if (Elt.getOpcode() != ISD::Constant &&
3344 Elt.getOpcode() != ISD::ConstantFP)
3345 IsAllConstants = false;
3346 if (isZeroNode(Elt))
3349 NonZeros |= (1 << i);
3354 if (NumNonZero == 0) {
3355 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3356 return DAG.getUNDEF(VT);
3359 // Special case for single non-zero, non-undef, element.
3360 if (NumNonZero == 1 && NumElems <= 4) {
3361 unsigned Idx = CountTrailingZeros_32(NonZeros);
3362 SDValue Item = Op.getOperand(Idx);
3364 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3365 // the value are obviously zero, truncate the value to i32 and do the
3366 // insertion that way. Only do this if the value is non-constant or if the
3367 // value is a constant being inserted into element 0. It is cheaper to do
3368 // a constant pool load than it is to do a movd + shuffle.
3369 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3370 (!IsAllConstants || Idx == 0)) {
3371 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3372 // Handle MMX and SSE both.
3373 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3374 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3376 // Truncate the value (which may itself be a constant) to i32, and
3377 // convert it to a vector with movd (S2V+shuffle to zero extend).
3378 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3379 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3380 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3381 Subtarget->hasSSE2(), DAG);
3383 // Now we have our 32-bit value zero extended in the low element of
3384 // a vector. If Idx != 0, swizzle it into place.
3387 Item, DAG.getUNDEF(Item.getValueType()),
3388 getSwapEltZeroMask(VecElts, Idx, DAG, dl)
3390 Item = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VecVT, Ops, 3);
3392 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3396 // If we have a constant or non-constant insertion into the low element of
3397 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3398 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3399 // depending on what the source datatype is. Because we can only get here
3400 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3402 // Don't do this for i64 values on x86-32.
3403 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3404 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3405 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3406 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3407 Subtarget->hasSSE2(), DAG);
3410 // Is it a vector logical left shift?
3411 if (NumElems == 2 && Idx == 1 &&
3412 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3413 unsigned NumBits = VT.getSizeInBits();
3414 return getVShift(true, VT,
3415 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3416 VT, Op.getOperand(1)),
3417 NumBits/2, DAG, *this, dl);
3420 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3423 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3424 // is a non-constant being inserted into an element other than the low one,
3425 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3426 // movd/movss) to move this into the low element, then shuffle it into
3428 if (EVTBits == 32) {
3429 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3431 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3432 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3433 Subtarget->hasSSE2(), DAG);
3434 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3435 MVT MaskEVT = MaskVT.getVectorElementType();
3436 SmallVector<SDValue, 8> MaskVec;
3437 for (unsigned i = 0; i < NumElems; i++)
3438 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3439 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3440 &MaskVec[0], MaskVec.size());
3441 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, Item,
3442 DAG.getUNDEF(VT), Mask);
3446 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3447 if (Values.size() == 1)
3450 // A vector full of immediates; various special cases are already
3451 // handled, so this is best done with a single constant-pool load.
3455 // Let legalizer expand 2-wide build_vectors.
3456 if (EVTBits == 64) {
3457 if (NumNonZero == 1) {
3458 // One half is zero or undef.
3459 unsigned Idx = CountTrailingZeros_32(NonZeros);
3460 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3461 Op.getOperand(Idx));
3462 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3463 Subtarget->hasSSE2(), DAG);
3468 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3469 if (EVTBits == 8 && NumElems == 16) {
3470 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3472 if (V.getNode()) return V;
3475 if (EVTBits == 16 && NumElems == 8) {
3476 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3478 if (V.getNode()) return V;
3481 // If element VT is == 32 bits, turn it into a number of shuffles.
3482 SmallVector<SDValue, 8> V;
3484 if (NumElems == 4 && NumZero > 0) {
3485 for (unsigned i = 0; i < 4; ++i) {
3486 bool isZero = !(NonZeros & (1 << i));
3488 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3490 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3493 for (unsigned i = 0; i < 2; ++i) {
3494 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3497 V[i] = V[i*2]; // Must be a zero vector.
3500 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2+1], V[i*2],
3501 getMOVLMask(NumElems, DAG, dl));
3504 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3505 getMOVLMask(NumElems, DAG, dl));
3508 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
3509 getUnpacklMask(NumElems, DAG, dl));
3514 MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3515 MVT EVT = MaskVT.getVectorElementType();
3516 SmallVector<SDValue, 8> MaskVec;
3517 bool Reverse = (NonZeros & 0x3) == 2;
3518 for (unsigned i = 0; i < 2; ++i)
3520 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3522 MaskVec.push_back(DAG.getConstant(i, EVT));
3523 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3524 for (unsigned i = 0; i < 2; ++i)
3526 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3528 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3529 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
3530 &MaskVec[0], MaskVec.size());
3531 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[0], V[1], ShufMask);
3534 if (Values.size() > 2) {
3535 // Expand into a number of unpckl*.
3537 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3538 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3539 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3540 SDValue UnpckMask = getUnpacklMask(NumElems, DAG, dl);
3541 for (unsigned i = 0; i < NumElems; ++i)
3542 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3544 while (NumElems != 0) {
3545 for (unsigned i = 0; i < NumElems; ++i)
3546 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i], V[i + NumElems],
3556 // v8i16 shuffles - Prefer shuffles in the following order:
3557 // 1. [all] pshuflw, pshufhw, optional move
3558 // 2. [ssse3] 1 x pshufb
3559 // 3. [ssse3] 2 x pshufb + 1 x por
3560 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3562 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
3563 SDValue PermMask, SelectionDAG &DAG,
3564 X86TargetLowering &TLI, DebugLoc dl) {
3565 SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
3566 PermMask.getNode()->op_end());
3567 SmallVector<int, 8> MaskVals;
3569 // Determine if more than 1 of the words in each of the low and high quadwords
3570 // of the result come from the same quadword of one of the two inputs. Undef
3571 // mask values count as coming from any quadword, for better codegen.
3572 SmallVector<unsigned, 4> LoQuad(4);
3573 SmallVector<unsigned, 4> HiQuad(4);
3574 BitVector InputQuads(4);
3575 for (unsigned i = 0; i < 8; ++i) {
3576 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3577 SDValue Elt = MaskElts[i];
3578 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3579 cast<ConstantSDNode>(Elt)->getZExtValue();
3580 MaskVals.push_back(EltIdx);
3589 InputQuads.set(EltIdx / 4);
3592 int BestLoQuad = -1;
3593 unsigned MaxQuad = 1;
3594 for (unsigned i = 0; i < 4; ++i) {
3595 if (LoQuad[i] > MaxQuad) {
3597 MaxQuad = LoQuad[i];
3601 int BestHiQuad = -1;
3603 for (unsigned i = 0; i < 4; ++i) {
3604 if (HiQuad[i] > MaxQuad) {
3606 MaxQuad = HiQuad[i];
3610 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3611 // of the two input vectors, shuffle them into one input vector so only a
3612 // single pshufb instruction is necessary. If There are more than 2 input
3613 // quads, disable the next transformation since it does not help SSSE3.
3614 bool V1Used = InputQuads[0] || InputQuads[1];
3615 bool V2Used = InputQuads[2] || InputQuads[3];
3616 if (TLI.getSubtarget()->hasSSSE3()) {
3617 if (InputQuads.count() == 2 && V1Used && V2Used) {
3618 BestLoQuad = InputQuads.find_first();
3619 BestHiQuad = InputQuads.find_next(BestLoQuad);
3621 if (InputQuads.count() > 2) {
3627 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3628 // the shuffle mask. If a quad is scored as -1, that means that it contains
3629 // words from all 4 input quadwords.
3631 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3632 SmallVector<SDValue,8> MaskV;
3633 MaskV.push_back(DAG.getConstant(BestLoQuad < 0 ? 0 : BestLoQuad, MVT::i64));
3634 MaskV.push_back(DAG.getConstant(BestHiQuad < 0 ? 1 : BestHiQuad, MVT::i64));
3635 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, &MaskV[0], 2);
3637 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
3638 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3639 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), Mask);
3640 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3642 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3643 // source words for the shuffle, to aid later transformations.
3644 bool AllWordsInNewV = true;
3645 bool InOrder[2] = { true, true };
3646 for (unsigned i = 0; i != 8; ++i) {
3647 int idx = MaskVals[i];
3649 InOrder[i/4] = false;
3650 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3652 AllWordsInNewV = false;
3656 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3657 if (AllWordsInNewV) {
3658 for (int i = 0; i != 8; ++i) {
3659 int idx = MaskVals[i];
3662 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3663 if ((idx != i) && idx < 4)
3665 if ((idx != i) && idx > 3)
3674 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3675 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3676 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3678 for (unsigned i = 0; i != 8; ++i)
3679 MaskV.push_back((MaskVals[i] < 0) ? DAG.getUNDEF(MVT::i16)
3680 : DAG.getConstant(MaskVals[i],
3682 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3683 DAG.getUNDEF(MVT::v8i16),
3684 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i16,
3689 // If we have SSSE3, and all words of the result are from 1 input vector,
3690 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3691 // is present, fall back to case 4.
3692 if (TLI.getSubtarget()->hasSSSE3()) {
3693 SmallVector<SDValue,16> pshufbMask;
3695 // If we have elements from both input vectors, set the high bit of the
3696 // shuffle mask element to zero out elements that come from V2 in the V1
3697 // mask, and elements that come from V1 in the V2 mask, so that the two
3698 // results can be OR'd together.
3699 bool TwoInputs = V1Used && V2Used;
3700 for (unsigned i = 0; i != 8; ++i) {
3701 int EltIdx = MaskVals[i] * 2;
3702 if (TwoInputs && (EltIdx >= 16)) {
3703 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3704 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3707 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3708 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3710 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3711 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3712 DAG.getNode(ISD::BUILD_VECTOR, dl,
3713 MVT::v16i8, &pshufbMask[0], 16));
3715 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3717 // Calculate the shuffle mask for the second input, shuffle it, and
3718 // OR it with the first shuffled input.
3720 for (unsigned i = 0; i != 8; ++i) {
3721 int EltIdx = MaskVals[i] * 2;
3723 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3724 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3727 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3728 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3730 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3731 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3732 DAG.getNode(ISD::BUILD_VECTOR, dl,
3733 MVT::v16i8, &pshufbMask[0], 16));
3734 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3735 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3738 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3739 // and update MaskVals with new element order.
3740 BitVector InOrder(8);
3741 if (BestLoQuad >= 0) {
3742 SmallVector<SDValue, 8> MaskV;
3743 for (int i = 0; i != 4; ++i) {
3744 int idx = MaskVals[i];
3746 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3748 } else if ((idx / 4) == BestLoQuad) {
3749 MaskV.push_back(DAG.getConstant(idx & 3, MVT::i16));
3752 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3755 for (unsigned i = 4; i != 8; ++i)
3756 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3757 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3758 DAG.getUNDEF(MVT::v8i16),
3759 DAG.getNode(ISD::BUILD_VECTOR, dl,
3760 MVT::v8i16, &MaskV[0], 8));
3763 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3764 // and update MaskVals with the new element order.
3765 if (BestHiQuad >= 0) {
3766 SmallVector<SDValue, 8> MaskV;
3767 for (unsigned i = 0; i != 4; ++i)
3768 MaskV.push_back(DAG.getConstant(i, MVT::i16));
3769 for (unsigned i = 4; i != 8; ++i) {
3770 int idx = MaskVals[i];
3772 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3774 } else if ((idx / 4) == BestHiQuad) {
3775 MaskV.push_back(DAG.getConstant((idx & 3) + 4, MVT::i16));
3778 MaskV.push_back(DAG.getUNDEF(MVT::i16));
3781 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, NewV,
3782 DAG.getUNDEF(MVT::v8i16),
3783 DAG.getNode(ISD::BUILD_VECTOR, dl,
3784 MVT::v8i16, &MaskV[0], 8));
3787 // In case BestHi & BestLo were both -1, which means each quadword has a word
3788 // from each of the four input quadwords, calculate the InOrder bitvector now
3789 // before falling through to the insert/extract cleanup.
3790 if (BestLoQuad == -1 && BestHiQuad == -1) {
3792 for (int i = 0; i != 8; ++i)
3793 if (MaskVals[i] < 0 || MaskVals[i] == i)
3797 // The other elements are put in the right place using pextrw and pinsrw.
3798 for (unsigned i = 0; i != 8; ++i) {
3801 int EltIdx = MaskVals[i];
3804 SDValue ExtOp = (EltIdx < 8)
3805 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3806 DAG.getIntPtrConstant(EltIdx))
3807 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3808 DAG.getIntPtrConstant(EltIdx - 8));
3809 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3810 DAG.getIntPtrConstant(i));
3815 // v16i8 shuffles - Prefer shuffles in the following order:
3816 // 1. [ssse3] 1 x pshufb
3817 // 2. [ssse3] 2 x pshufb + 1 x por
3818 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3820 SDValue LowerVECTOR_SHUFFLEv16i8(SDValue V1, SDValue V2,
3821 SDValue PermMask, SelectionDAG &DAG,
3822 X86TargetLowering &TLI, DebugLoc dl) {
3823 SmallVector<SDValue, 16> MaskElts(PermMask.getNode()->op_begin(),
3824 PermMask.getNode()->op_end());
3825 SmallVector<int, 16> MaskVals;
3827 // If we have SSSE3, case 1 is generated when all result bytes come from
3828 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3829 // present, fall back to case 3.
3830 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3833 for (unsigned i = 0; i < 16; ++i) {
3834 SDValue Elt = MaskElts[i];
3835 int EltIdx = Elt.getOpcode() == ISD::UNDEF ? -1 :
3836 cast<ConstantSDNode>(Elt)->getZExtValue();
3837 MaskVals.push_back(EltIdx);
3846 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3847 if (TLI.getSubtarget()->hasSSSE3()) {
3848 SmallVector<SDValue,16> pshufbMask;
3850 // If all result elements are from one input vector, then only translate
3851 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3853 // Otherwise, we have elements from both input vectors, and must zero out
3854 // elements that come from V2 in the first mask, and V1 in the second mask
3855 // so that we can OR them together.
3856 bool TwoInputs = !(V1Only || V2Only);
3857 for (unsigned i = 0; i != 16; ++i) {
3858 int EltIdx = MaskVals[i];
3859 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3860 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3863 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3865 // If all the elements are from V2, assign it to V1 and return after
3866 // building the first pshufb.
3869 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3870 DAG.getNode(ISD::BUILD_VECTOR, dl,
3871 MVT::v16i8, &pshufbMask[0], 16));
3875 // Calculate the shuffle mask for the second input, shuffle it, and
3876 // OR it with the first shuffled input.
3878 for (unsigned i = 0; i != 16; ++i) {
3879 int EltIdx = MaskVals[i];
3881 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3884 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3886 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3887 DAG.getNode(ISD::BUILD_VECTOR, dl,
3888 MVT::v16i8, &pshufbMask[0], 16));
3889 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3892 // No SSSE3 - Calculate in place words and then fix all out of place words
3893 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3894 // the 16 different words that comprise the two doublequadword input vectors.
3895 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3896 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3897 SDValue NewV = V2Only ? V2 : V1;
3898 for (int i = 0; i != 8; ++i) {
3899 int Elt0 = MaskVals[i*2];
3900 int Elt1 = MaskVals[i*2+1];
3902 // This word of the result is all undef, skip it.
3903 if (Elt0 < 0 && Elt1 < 0)
3906 // This word of the result is already in the correct place, skip it.
3907 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3909 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3912 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3913 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3916 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3917 // using a single extract together, load it and store it.
3918 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3919 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3920 DAG.getIntPtrConstant(Elt1 / 2));
3921 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3922 DAG.getIntPtrConstant(i));
3926 // If Elt1 is defined, extract it from the appropriate source. If the
3927 // source byte is not also odd, shift the extracted word left 8 bits
3928 // otherwise clear the bottom 8 bits if we need to do an or.
3930 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3931 DAG.getIntPtrConstant(Elt1 / 2));
3932 if ((Elt1 & 1) == 0)
3933 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3934 DAG.getConstant(8, TLI.getShiftAmountTy()));
3936 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3937 DAG.getConstant(0xFF00, MVT::i16));
3939 // If Elt0 is defined, extract it from the appropriate source. If the
3940 // source byte is not also even, shift the extracted word right 8 bits. If
3941 // Elt1 was also defined, OR the extracted values together before
3942 // inserting them in the result.
3944 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3945 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3946 if ((Elt0 & 1) != 0)
3947 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3948 DAG.getConstant(8, TLI.getShiftAmountTy()));
3950 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3951 DAG.getConstant(0x00FF, MVT::i16));
3952 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3955 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3956 DAG.getIntPtrConstant(i));
3958 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3961 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3962 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3963 /// done when every pair / quad of shuffle mask elements point to elements in
3964 /// the right sequence. e.g.
3965 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3967 SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
3969 SDValue PermMask, SelectionDAG &DAG,
3970 TargetLowering &TLI, DebugLoc dl) {
3971 unsigned NumElems = PermMask.getNumOperands();
3972 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3973 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3974 MVT MaskEltVT = MaskVT.getVectorElementType();
3976 switch (VT.getSimpleVT()) {
3977 default: assert(false && "Unexpected!");
3978 case MVT::v4f32: NewVT = MVT::v2f64; break;
3979 case MVT::v4i32: NewVT = MVT::v2i64; break;
3980 case MVT::v8i16: NewVT = MVT::v4i32; break;
3981 case MVT::v16i8: NewVT = MVT::v4i32; break;
3984 if (NewWidth == 2) {
3990 unsigned Scale = NumElems / NewWidth;
3991 SmallVector<SDValue, 8> MaskVec;
3992 for (unsigned i = 0; i < NumElems; i += Scale) {
3993 unsigned StartIdx = ~0U;
3994 for (unsigned j = 0; j < Scale; ++j) {
3995 SDValue Elt = PermMask.getOperand(i+j);
3996 if (Elt.getOpcode() == ISD::UNDEF)
3998 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
3999 if (StartIdx == ~0U)
4000 StartIdx = EltIdx - (EltIdx % Scale);
4001 if (EltIdx != StartIdx + j)
4004 if (StartIdx == ~0U)
4005 MaskVec.push_back(DAG.getUNDEF(MaskEltVT));
4007 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
4010 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4011 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4012 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, NewVT, V1, V2,
4013 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4014 &MaskVec[0], MaskVec.size()));
4017 /// getVZextMovL - Return a zero-extending vector move low node.
4019 static SDValue getVZextMovL(MVT VT, MVT OpVT,
4020 SDValue SrcOp, SelectionDAG &DAG,
4021 const X86Subtarget *Subtarget, DebugLoc dl) {
4022 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4023 LoadSDNode *LD = NULL;
4024 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4025 LD = dyn_cast<LoadSDNode>(SrcOp);
4027 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4029 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4030 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
4031 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4032 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4033 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
4035 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4036 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4037 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4038 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4046 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4047 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4048 DAG.getNode(ISD::BIT_CONVERT, dl,
4052 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4055 LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
4056 SDValue PermMask, MVT VT, SelectionDAG &DAG,
4058 MVT MaskVT = PermMask.getValueType();
4059 MVT MaskEVT = MaskVT.getVectorElementType();
4060 SmallVector<std::pair<int, int>, 8> Locs;
4062 SmallVector<SDValue, 8> Mask1(4, DAG.getUNDEF(MaskEVT));
4065 for (unsigned i = 0; i != 4; ++i) {
4066 SDValue Elt = PermMask.getOperand(i);
4067 if (Elt.getOpcode() == ISD::UNDEF) {
4068 Locs[i] = std::make_pair(-1, -1);
4070 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4071 assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
4073 Locs[i] = std::make_pair(0, NumLo);
4077 Locs[i] = std::make_pair(1, NumHi);
4079 Mask1[2+NumHi] = Elt;
4085 if (NumLo <= 2 && NumHi <= 2) {
4086 // If no more than two elements come from either vector. This can be
4087 // implemented with two shuffles. First shuffle gather the elements.
4088 // The second shuffle, which takes the first shuffle as both of its
4089 // vector operands, put the elements into the right order.
4090 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4091 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4092 &Mask1[0], Mask1.size()));
4094 SmallVector<SDValue, 8> Mask2(4, DAG.getUNDEF(MaskEVT));
4095 for (unsigned i = 0; i != 4; ++i) {
4096 if (Locs[i].first == -1)
4099 unsigned Idx = (i < 2) ? 0 : 4;
4100 Idx += Locs[i].first * 2 + Locs[i].second;
4101 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
4105 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1,
4106 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4107 &Mask2[0], Mask2.size()));
4108 } else if (NumLo == 3 || NumHi == 3) {
4109 // Otherwise, we must have three elements from one vector, call it X, and
4110 // one element from the other, call it Y. First, use a shufps to build an
4111 // intermediate vector with the one element from Y and the element from X
4112 // that will be in the same half in the final destination (the indexes don't
4113 // matter). Then, use a shufps to build the final vector, taking the half
4114 // containing the element from Y from the intermediate, and the other half
4117 // Normalize it so the 3 elements come from V1.
4118 PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
4122 // Find the element from V2.
4124 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4125 SDValue Elt = PermMask.getOperand(HiIndex);
4126 if (Elt.getOpcode() == ISD::UNDEF)
4128 unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
4133 Mask1[0] = PermMask.getOperand(HiIndex);
4134 Mask1[1] = DAG.getUNDEF(MaskEVT);
4135 Mask1[2] = PermMask.getOperand(HiIndex^1);
4136 Mask1[3] = DAG.getUNDEF(MaskEVT);
4137 V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4138 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &Mask1[0], 4));
4141 Mask1[0] = PermMask.getOperand(0);
4142 Mask1[1] = PermMask.getOperand(1);
4143 Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
4144 Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
4145 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4146 DAG.getNode(ISD::BUILD_VECTOR, dl,
4147 MaskVT, &Mask1[0], 4));
4149 Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
4150 Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
4151 Mask1[2] = PermMask.getOperand(2);
4152 Mask1[3] = PermMask.getOperand(3);
4153 if (Mask1[2].getOpcode() != ISD::UNDEF)
4155 DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
4157 if (Mask1[3].getOpcode() != ISD::UNDEF)
4159 DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
4161 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V2, V1,
4162 DAG.getNode(ISD::BUILD_VECTOR, dl,
4163 MaskVT, &Mask1[0], 4));
4167 // Break it into (shuffle shuffle_hi, shuffle_lo).
4169 SmallVector<SDValue,8> LoMask(4, DAG.getUNDEF(MaskEVT));
4170 SmallVector<SDValue,8> HiMask(4, DAG.getUNDEF(MaskEVT));
4171 SmallVector<SDValue,8> *MaskPtr = &LoMask;
4172 unsigned MaskIdx = 0;
4175 for (unsigned i = 0; i != 4; ++i) {
4182 SDValue Elt = PermMask.getOperand(i);
4183 if (Elt.getOpcode() == ISD::UNDEF) {
4184 Locs[i] = std::make_pair(-1, -1);
4185 } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
4186 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4187 (*MaskPtr)[LoIdx] = Elt;
4190 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4191 (*MaskPtr)[HiIdx] = Elt;
4196 SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4197 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4198 &LoMask[0], LoMask.size()));
4199 SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
4200 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4201 &HiMask[0], HiMask.size()));
4202 SmallVector<SDValue, 8> MaskOps;
4203 for (unsigned i = 0; i != 4; ++i) {
4204 if (Locs[i].first == -1) {
4205 MaskOps.push_back(DAG.getUNDEF(MaskEVT));
4207 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4208 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
4211 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, LoShuffle, HiShuffle,
4212 DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4213 &MaskOps[0], MaskOps.size()));
4217 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4218 SDValue V1 = Op.getOperand(0);
4219 SDValue V2 = Op.getOperand(1);
4220 SDValue PermMask = Op.getOperand(2);
4221 MVT VT = Op.getValueType();
4222 DebugLoc dl = Op.getDebugLoc();
4223 unsigned NumElems = PermMask.getNumOperands();
4224 bool isMMX = VT.getSizeInBits() == 64;
4225 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4226 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4227 bool V1IsSplat = false;
4228 bool V2IsSplat = false;
4230 // FIXME: Check for legal shuffle and return?
4232 if (isUndefShuffle(Op.getNode()))
4233 return DAG.getUNDEF(VT);
4235 if (isZeroShuffle(Op.getNode()))
4236 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4238 if (isIdentityMask(PermMask.getNode()))
4240 else if (isIdentityMask(PermMask.getNode(), true))
4243 // Canonicalize movddup shuffles.
4244 if (V2IsUndef && Subtarget->hasSSE2() &&
4245 VT.getSizeInBits() == 128 &&
4246 X86::isMOVDDUPMask(PermMask.getNode()))
4247 return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
4249 if (isSplatMask(PermMask.getNode())) {
4250 if (isMMX || NumElems < 4) return Op;
4251 // Promote it to a v4{if}32 splat.
4252 return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
4255 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4257 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4258 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG,
4260 if (NewOp.getNode())
4261 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4262 LowerVECTOR_SHUFFLE(NewOp, DAG));
4263 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4264 // FIXME: Figure out a cleaner way to do this.
4265 // Try to make use of movq to zero out the top part.
4266 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4267 SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4269 if (NewOp.getNode()) {
4270 SDValue NewV1 = NewOp.getOperand(0);
4271 SDValue NewV2 = NewOp.getOperand(1);
4272 SDValue NewMask = NewOp.getOperand(2);
4273 if (isCommutedMOVL(NewMask.getNode(), true, false)) {
4274 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
4275 return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget,
4279 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4280 SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
4282 if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
4283 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4284 DAG, Subtarget, dl);
4288 // Check if this can be converted into a logical shift.
4289 bool isLeft = false;
4292 bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
4293 if (isShift && ShVal.hasOneUse()) {
4294 // If the shifted value has multiple uses, it may be cheaper to use
4295 // v_set0 + movlhps or movhlps, etc.
4296 MVT EVT = VT.getVectorElementType();
4297 ShAmt *= EVT.getSizeInBits();
4298 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4301 if (X86::isMOVLMask(PermMask.getNode())) {
4304 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4305 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4310 if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
4311 X86::isMOVSLDUPMask(PermMask.getNode()) ||
4312 X86::isMOVHLPSMask(PermMask.getNode()) ||
4313 X86::isMOVHPMask(PermMask.getNode()) ||
4314 X86::isMOVLPMask(PermMask.getNode())))
4317 if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
4318 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
4319 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4322 // No better options. Use a vshl / vsrl.
4323 MVT EVT = VT.getVectorElementType();
4324 ShAmt *= EVT.getSizeInBits();
4325 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4328 bool Commuted = false;
4329 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4330 // 1,1,1,1 -> v8i16 though.
4331 V1IsSplat = isSplatVector(V1.getNode());
4332 V2IsSplat = isSplatVector(V2.getNode());
4334 // Canonicalize the splat or undef, if present, to be on the RHS.
4335 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4336 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4337 std::swap(V1IsSplat, V2IsSplat);
4338 std::swap(V1IsUndef, V2IsUndef);
4342 // FIXME: Figure out a cleaner way to do this.
4343 if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
4344 if (V2IsUndef) return V1;
4345 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4347 // V2 is a splat, so the mask may be malformed. That is, it may point
4348 // to any V2 element. The instruction selectior won't like this. Get
4349 // a corrected mask and commute to form a proper MOVS{S|D}.
4350 SDValue NewMask = getMOVLMask(NumElems, DAG, dl);
4351 if (NewMask.getNode() != PermMask.getNode())
4352 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4357 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4358 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4359 X86::isUNPCKLMask(PermMask.getNode()) ||
4360 X86::isUNPCKHMask(PermMask.getNode()))
4364 // Normalize mask so all entries that point to V2 points to its first
4365 // element then try to match unpck{h|l} again. If match, return a
4366 // new vector_shuffle with the corrected mask.
4367 SDValue NewMask = NormalizeMask(PermMask, DAG);
4368 if (NewMask.getNode() != PermMask.getNode()) {
4369 if (X86::isUNPCKLMask(NewMask.getNode(), true)) {
4370 SDValue NewMask = getUnpacklMask(NumElems, DAG, dl);
4371 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4372 } else if (X86::isUNPCKHMask(NewMask.getNode(), true)) {
4373 SDValue NewMask = getUnpackhMask(NumElems, DAG, dl);
4374 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
4379 // Normalize the node to match x86 shuffle ops if needed
4380 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
4381 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4384 // Commute is back and try unpck* again.
4385 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
4386 if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
4387 X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
4388 X86::isUNPCKLMask(PermMask.getNode()) ||
4389 X86::isUNPCKHMask(PermMask.getNode()))
4393 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4394 // Try PSHUF* first, then SHUFP*.
4395 // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
4396 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
4397 if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
4398 if (V2.getOpcode() != ISD::UNDEF)
4399 return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1,
4400 DAG.getUNDEF(VT), PermMask);
4405 if (Subtarget->hasSSE2() &&
4406 (X86::isPSHUFDMask(PermMask.getNode()) ||
4407 X86::isPSHUFHWMask(PermMask.getNode()) ||
4408 X86::isPSHUFLWMask(PermMask.getNode()))) {
4410 if (VT == MVT::v4f32) {
4412 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT,
4413 DAG.getNode(ISD::BIT_CONVERT, dl, RVT, V1),
4414 DAG.getUNDEF(RVT), PermMask);
4415 } else if (V2.getOpcode() != ISD::UNDEF)
4416 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT, V1,
4417 DAG.getUNDEF(RVT), PermMask);
4419 Op = DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
4423 // Binary or unary shufps.
4424 if (X86::isSHUFPMask(PermMask.getNode()) ||
4425 (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
4429 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4430 if (VT == MVT::v8i16) {
4431 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this, dl);
4432 if (NewOp.getNode())
4436 if (VT == MVT::v16i8) {
4437 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(V1, V2, PermMask, DAG, *this, dl);
4438 if (NewOp.getNode())
4442 // Handle all 4 wide cases with a number of shuffles except for MMX.
4443 if (NumElems == 4 && !isMMX)
4444 return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG, dl);
4450 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4451 SelectionDAG &DAG) {
4452 MVT VT = Op.getValueType();
4453 DebugLoc dl = Op.getDebugLoc();
4454 if (VT.getSizeInBits() == 8) {
4455 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4456 Op.getOperand(0), Op.getOperand(1));
4457 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4458 DAG.getValueType(VT));
4459 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4460 } else if (VT.getSizeInBits() == 16) {
4461 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4462 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4464 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4465 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4466 DAG.getNode(ISD::BIT_CONVERT, dl,
4470 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4471 Op.getOperand(0), Op.getOperand(1));
4472 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4473 DAG.getValueType(VT));
4474 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4475 } else if (VT == MVT::f32) {
4476 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4477 // the result back to FR32 register. It's only worth matching if the
4478 // result has a single use which is a store or a bitcast to i32. And in
4479 // the case of a store, it's not worth it if the index is a constant 0,
4480 // because a MOVSSmr can be used instead, which is smaller and faster.
4481 if (!Op.hasOneUse())
4483 SDNode *User = *Op.getNode()->use_begin();
4484 if ((User->getOpcode() != ISD::STORE ||
4485 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4486 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4487 (User->getOpcode() != ISD::BIT_CONVERT ||
4488 User->getValueType(0) != MVT::i32))
4490 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4491 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4494 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4495 } else if (VT == MVT::i32) {
4496 // ExtractPS works with constant index.
4497 if (isa<ConstantSDNode>(Op.getOperand(1)))
4505 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4506 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4509 if (Subtarget->hasSSE41()) {
4510 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4515 MVT VT = Op.getValueType();
4516 DebugLoc dl = Op.getDebugLoc();
4517 // TODO: handle v16i8.
4518 if (VT.getSizeInBits() == 16) {
4519 SDValue Vec = Op.getOperand(0);
4520 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4522 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4523 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4524 DAG.getNode(ISD::BIT_CONVERT, dl,
4527 // Transform it so it match pextrw which produces a 32-bit result.
4528 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4529 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4530 Op.getOperand(0), Op.getOperand(1));
4531 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4532 DAG.getValueType(VT));
4533 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4534 } else if (VT.getSizeInBits() == 32) {
4535 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4538 // SHUFPS the element to the lowest double word, then movss.
4539 MVT MaskVT = MVT::getIntVectorWithNumElements(4);
4540 SmallVector<SDValue, 8> IdxVec;
4542 push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
4544 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4546 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4548 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4549 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4550 &IdxVec[0], IdxVec.size());
4551 SDValue Vec = Op.getOperand(0);
4552 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4553 Vec, DAG.getUNDEF(Vec.getValueType()), Mask);
4554 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4555 DAG.getIntPtrConstant(0));
4556 } else if (VT.getSizeInBits() == 64) {
4557 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4558 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4559 // to match extract_elt for f64.
4560 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4564 // UNPCKHPD the element to the lowest double word, then movsd.
4565 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4566 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4567 MVT MaskVT = MVT::getIntVectorWithNumElements(2);
4568 SmallVector<SDValue, 8> IdxVec;
4569 IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
4571 push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
4572 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
4573 &IdxVec[0], IdxVec.size());
4574 SDValue Vec = Op.getOperand(0);
4575 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
4576 Vec, DAG.getUNDEF(Vec.getValueType()),
4578 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4579 DAG.getIntPtrConstant(0));
4586 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4587 MVT VT = Op.getValueType();
4588 MVT EVT = VT.getVectorElementType();
4589 DebugLoc dl = Op.getDebugLoc();
4591 SDValue N0 = Op.getOperand(0);
4592 SDValue N1 = Op.getOperand(1);
4593 SDValue N2 = Op.getOperand(2);
4595 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4596 isa<ConstantSDNode>(N2)) {
4597 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4599 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4601 if (N1.getValueType() != MVT::i32)
4602 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4603 if (N2.getValueType() != MVT::i32)
4604 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4605 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4606 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4607 // Bits [7:6] of the constant are the source select. This will always be
4608 // zero here. The DAG Combiner may combine an extract_elt index into these
4609 // bits. For example (insert (extract, 3), 2) could be matched by putting
4610 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4611 // Bits [5:4] of the constant are the destination select. This is the
4612 // value of the incoming immediate.
4613 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4614 // combine either bitwise AND or insert of float 0.0 to set these bits.
4615 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4616 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4617 } else if (EVT == MVT::i32) {
4618 // InsertPS works with constant index.
4619 if (isa<ConstantSDNode>(N2))
4626 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4627 MVT VT = Op.getValueType();
4628 MVT EVT = VT.getVectorElementType();
4630 if (Subtarget->hasSSE41())
4631 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4636 DebugLoc dl = Op.getDebugLoc();
4637 SDValue N0 = Op.getOperand(0);
4638 SDValue N1 = Op.getOperand(1);
4639 SDValue N2 = Op.getOperand(2);
4641 if (EVT.getSizeInBits() == 16) {
4642 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4643 // as its second argument.
4644 if (N1.getValueType() != MVT::i32)
4645 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4646 if (N2.getValueType() != MVT::i32)
4647 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4648 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4654 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4655 DebugLoc dl = Op.getDebugLoc();
4656 if (Op.getValueType() == MVT::v2f32)
4657 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4658 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4659 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4660 Op.getOperand(0))));
4662 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4663 MVT VT = MVT::v2i32;
4664 switch (Op.getValueType().getSimpleVT()) {
4671 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4672 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4675 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4676 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4677 // one of the above mentioned nodes. It has to be wrapped because otherwise
4678 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4679 // be used to form addressing mode. These wrapped nodes will be selected
4682 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4683 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4684 // FIXME there isn't really any debug info here, should come from the parent
4685 DebugLoc dl = CP->getDebugLoc();
4686 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4687 CP->getAlignment());
4688 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4689 // With PIC, the address is actually $g + Offset.
4690 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4691 !Subtarget->isPICStyleRIPRel()) {
4692 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4693 DAG.getNode(X86ISD::GlobalBaseReg,
4694 DebugLoc::getUnknownLoc(),
4703 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4705 SelectionDAG &DAG) const {
4706 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4707 bool ExtraLoadRequired =
4708 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4710 // Create the TargetGlobalAddress node, folding in the constant
4711 // offset if it is legal.
4713 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4714 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4717 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4718 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4720 // With PIC, the address is actually $g + Offset.
4721 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4722 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4723 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4727 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4728 // load the value at address GV, not the value of GV itself. This means that
4729 // the GlobalAddress must be in the base or index register of the address, not
4730 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4731 // The same applies for external symbols during PIC codegen
4732 if (ExtraLoadRequired)
4733 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4734 PseudoSourceValue::getGOT(), 0);
4736 // If there was a non-zero offset that we didn't fold, create an explicit
4739 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4740 DAG.getConstant(Offset, getPointerTy()));
4746 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4747 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4748 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4749 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4752 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4754 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4757 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4758 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4759 DAG.getNode(X86ISD::GlobalBaseReg,
4760 DebugLoc::getUnknownLoc(),
4762 InFlag = Chain.getValue(1);
4764 // emit leal symbol@TLSGD(,%ebx,1), %eax
4765 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4766 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4767 GA->getValueType(0),
4769 SDValue Ops[] = { Chain, TGA, InFlag };
4770 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4771 InFlag = Result.getValue(2);
4772 Chain = Result.getValue(1);
4774 // call ___tls_get_addr. This function receives its argument in
4775 // the register EAX.
4776 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Result, InFlag);
4777 InFlag = Chain.getValue(1);
4779 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4780 SDValue Ops1[] = { Chain,
4781 DAG.getTargetExternalSymbol("___tls_get_addr",
4783 DAG.getRegister(X86::EAX, PtrVT),
4784 DAG.getRegister(X86::EBX, PtrVT),
4786 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
4787 InFlag = Chain.getValue(1);
4789 return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
4792 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4794 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4796 SDValue InFlag, Chain;
4797 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4799 // emit leaq symbol@TLSGD(%rip), %rdi
4800 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4801 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4802 GA->getValueType(0),
4804 SDValue Ops[] = { DAG.getEntryNode(), TGA};
4805 SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4806 Chain = Result.getValue(1);
4807 InFlag = Result.getValue(2);
4809 // call __tls_get_addr. This function receives its argument in
4810 // the register RDI.
4811 Chain = DAG.getCopyToReg(Chain, dl, X86::RDI, Result, InFlag);
4812 InFlag = Chain.getValue(1);
4814 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4815 SDValue Ops1[] = { Chain,
4816 DAG.getTargetExternalSymbol("__tls_get_addr",
4818 DAG.getRegister(X86::RDI, PtrVT),
4820 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
4821 InFlag = Chain.getValue(1);
4823 return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
4826 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4827 // "local exec" model.
4828 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4829 const MVT PtrVT, TLSModel::Model model) {
4830 DebugLoc dl = GA->getDebugLoc();
4831 // Get the Thread Pointer
4832 SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER,
4833 DebugLoc::getUnknownLoc(), PtrVT);
4834 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4836 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4837 GA->getValueType(0),
4839 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4841 if (model == TLSModel::InitialExec)
4842 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4843 PseudoSourceValue::getGOT(), 0);
4845 // The address of the thread local variable is the add of the thread
4846 // pointer with the offset of the variable.
4847 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4851 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4852 // TODO: implement the "local dynamic" model
4853 // TODO: implement the "initial exec"model for pic executables
4854 assert(Subtarget->isTargetELF() &&
4855 "TLS not implemented for non-ELF targets");
4856 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4857 GlobalValue *GV = GA->getGlobal();
4858 TLSModel::Model model =
4859 getTLSModel (GV, getTargetMachine().getRelocationModel());
4860 if (Subtarget->is64Bit()) {
4862 case TLSModel::GeneralDynamic:
4863 case TLSModel::LocalDynamic: // not implemented
4864 case TLSModel::InitialExec: // not implemented
4865 case TLSModel::LocalExec: // not implemented
4866 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4868 assert (0 && "Unknown TLS model");
4872 case TLSModel::GeneralDynamic:
4873 case TLSModel::LocalDynamic: // not implemented
4874 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4876 case TLSModel::InitialExec:
4877 case TLSModel::LocalExec:
4878 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model);
4880 assert (0 && "Unknown TLS model");
4886 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4887 // FIXME there isn't really any debug info here
4888 DebugLoc dl = Op.getDebugLoc();
4889 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4890 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4891 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4892 // With PIC, the address is actually $g + Offset.
4893 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4894 !Subtarget->isPICStyleRIPRel()) {
4895 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4896 DAG.getNode(X86ISD::GlobalBaseReg,
4897 DebugLoc::getUnknownLoc(),
4905 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4906 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4907 // FIXME there isn't really any debug into here
4908 DebugLoc dl = JT->getDebugLoc();
4909 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4910 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4911 // With PIC, the address is actually $g + Offset.
4912 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4913 !Subtarget->isPICStyleRIPRel()) {
4914 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4915 DAG.getNode(X86ISD::GlobalBaseReg,
4916 DebugLoc::getUnknownLoc(),
4924 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4925 /// take a 2 x i32 value to shift plus a shift amount.
4926 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4927 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4928 MVT VT = Op.getValueType();
4929 unsigned VTBits = VT.getSizeInBits();
4930 DebugLoc dl = Op.getDebugLoc();
4931 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4932 SDValue ShOpLo = Op.getOperand(0);
4933 SDValue ShOpHi = Op.getOperand(1);
4934 SDValue ShAmt = Op.getOperand(2);
4935 SDValue Tmp1 = isSRA ?
4936 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4937 DAG.getConstant(VTBits - 1, MVT::i8)) :
4938 DAG.getConstant(0, VT);
4941 if (Op.getOpcode() == ISD::SHL_PARTS) {
4942 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4943 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4945 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4946 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4949 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4950 DAG.getConstant(VTBits, MVT::i8));
4951 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4952 AndNode, DAG.getConstant(0, MVT::i8));
4955 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4956 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4957 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4959 if (Op.getOpcode() == ISD::SHL_PARTS) {
4960 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4961 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4963 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4964 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4967 SDValue Ops[2] = { Lo, Hi };
4968 return DAG.getMergeValues(Ops, 2, dl);
4971 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4972 MVT SrcVT = Op.getOperand(0).getValueType();
4973 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4974 "Unknown SINT_TO_FP to lower!");
4976 // These are really Legal; caller falls through into that case.
4977 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4979 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4980 Subtarget->is64Bit())
4983 DebugLoc dl = Op.getDebugLoc();
4984 unsigned Size = SrcVT.getSizeInBits()/8;
4985 MachineFunction &MF = DAG.getMachineFunction();
4986 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4987 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4988 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4990 PseudoSourceValue::getFixedStack(SSFI), 0);
4994 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4996 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4998 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4999 SmallVector<SDValue, 8> Ops;
5000 Ops.push_back(Chain);
5001 Ops.push_back(StackSlot);
5002 Ops.push_back(DAG.getValueType(SrcVT));
5003 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5004 Tys, &Ops[0], Ops.size());
5007 Chain = Result.getValue(1);
5008 SDValue InFlag = Result.getValue(2);
5010 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5011 // shouldn't be necessary except that RFP cannot be live across
5012 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5013 MachineFunction &MF = DAG.getMachineFunction();
5014 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
5015 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5016 Tys = DAG.getVTList(MVT::Other);
5017 SmallVector<SDValue, 8> Ops;
5018 Ops.push_back(Chain);
5019 Ops.push_back(Result);
5020 Ops.push_back(StackSlot);
5021 Ops.push_back(DAG.getValueType(Op.getValueType()));
5022 Ops.push_back(InFlag);
5023 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5024 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5025 PseudoSourceValue::getFixedStack(SSFI), 0);
5031 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5032 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5033 // This algorithm is not obvious. Here it is in C code, more or less:
5035 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5036 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5037 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5039 // Copy ints to xmm registers.
5040 __m128i xh = _mm_cvtsi32_si128( hi );
5041 __m128i xl = _mm_cvtsi32_si128( lo );
5043 // Combine into low half of a single xmm register.
5044 __m128i x = _mm_unpacklo_epi32( xh, xl );
5048 // Merge in appropriate exponents to give the integer bits the right
5050 x = _mm_unpacklo_epi32( x, exp );
5052 // Subtract away the biases to deal with the IEEE-754 double precision
5054 d = _mm_sub_pd( (__m128d) x, bias );
5056 // All conversions up to here are exact. The correctly rounded result is
5057 // calculated using the current rounding mode using the following
5059 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5060 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5061 // store doesn't really need to be here (except
5062 // maybe to zero the other double)
5067 DebugLoc dl = Op.getDebugLoc();
5069 // Build some magic constants.
5070 std::vector<Constant*> CV0;
5071 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
5072 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
5073 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5074 CV0.push_back(ConstantInt::get(APInt(32, 0)));
5075 Constant *C0 = ConstantVector::get(CV0);
5076 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5078 std::vector<Constant*> CV1;
5079 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
5080 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
5081 Constant *C1 = ConstantVector::get(CV1);
5082 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5084 SmallVector<SDValue, 4> MaskVec;
5085 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
5086 MaskVec.push_back(DAG.getConstant(4, MVT::i32));
5087 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
5088 MaskVec.push_back(DAG.getConstant(5, MVT::i32));
5089 SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
5090 &MaskVec[0], MaskVec.size());
5091 SmallVector<SDValue, 4> MaskVec2;
5092 MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
5093 MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
5094 SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32,
5095 &MaskVec2[0], MaskVec2.size());
5097 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5098 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5100 DAG.getIntPtrConstant(1)));
5101 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5102 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5104 DAG.getIntPtrConstant(0)));
5105 SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5106 XR1, XR2, UnpcklMask);
5107 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5108 PseudoSourceValue::getConstantPool(), 0,
5110 SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
5111 Unpck1, CLod0, UnpcklMask);
5112 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5113 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5114 PseudoSourceValue::getConstantPool(), 0,
5116 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5118 // Add the halves; easiest way is to swap them into another reg first.
5119 SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2f64,
5120 Sub, Sub, ShufMask);
5121 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5122 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5123 DAG.getIntPtrConstant(0));
5126 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5127 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5128 DebugLoc dl = Op.getDebugLoc();
5129 // FP constant to bias correct the final result.
5130 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5133 // Load the 32-bit value into an XMM register.
5134 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5135 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5137 DAG.getIntPtrConstant(0)));
5139 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5140 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5141 DAG.getIntPtrConstant(0));
5143 // Or the load with the bias.
5144 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5145 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5146 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5148 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5149 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5150 MVT::v2f64, Bias)));
5151 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5152 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5153 DAG.getIntPtrConstant(0));
5155 // Subtract the bias.
5156 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5158 // Handle final rounding.
5159 MVT DestVT = Op.getValueType();
5161 if (DestVT.bitsLT(MVT::f64)) {
5162 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5163 DAG.getIntPtrConstant(0));
5164 } else if (DestVT.bitsGT(MVT::f64)) {
5165 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5168 // Handle final rounding.
5172 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5173 SDValue N0 = Op.getOperand(0);
5174 DebugLoc dl = Op.getDebugLoc();
5176 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5177 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5178 // the optimization here.
5179 if (DAG.SignBitIsZero(N0))
5180 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5182 MVT SrcVT = N0.getValueType();
5183 if (SrcVT == MVT::i64) {
5184 // We only handle SSE2 f64 target here; caller can handle the rest.
5185 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5188 return LowerUINT_TO_FP_i64(Op, DAG);
5189 } else if (SrcVT == MVT::i32) {
5190 return LowerUINT_TO_FP_i32(Op, DAG);
5193 assert(0 && "Unknown UINT_TO_FP to lower!");
5197 std::pair<SDValue,SDValue> X86TargetLowering::
5198 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
5199 DebugLoc dl = Op.getDebugLoc();
5200 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
5201 Op.getValueType().getSimpleVT() >= MVT::i16 &&
5202 "Unknown FP_TO_SINT to lower!");
5204 // These are really Legal.
5205 if (Op.getValueType() == MVT::i32 &&
5206 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5207 return std::make_pair(SDValue(), SDValue());
5208 if (Subtarget->is64Bit() &&
5209 Op.getValueType() == MVT::i64 &&
5210 Op.getOperand(0).getValueType() != MVT::f80)
5211 return std::make_pair(SDValue(), SDValue());
5213 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5215 MachineFunction &MF = DAG.getMachineFunction();
5216 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
5217 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5218 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5220 switch (Op.getValueType().getSimpleVT()) {
5221 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5222 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5223 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5224 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5227 SDValue Chain = DAG.getEntryNode();
5228 SDValue Value = Op.getOperand(0);
5229 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5230 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5231 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5232 PseudoSourceValue::getFixedStack(SSFI), 0);
5233 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5235 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5237 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5238 Chain = Value.getValue(1);
5239 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5240 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5243 // Build the FP_TO_INT*_IN_MEM
5244 SDValue Ops[] = { Chain, Value, StackSlot };
5245 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5247 return std::make_pair(FIST, StackSlot);
5250 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5251 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
5252 SDValue FIST = Vals.first, StackSlot = Vals.second;
5253 if (FIST.getNode() == 0) return SDValue();
5256 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5257 FIST, StackSlot, NULL, 0);
5260 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5261 DebugLoc dl = Op.getDebugLoc();
5262 MVT VT = Op.getValueType();
5265 EltVT = VT.getVectorElementType();
5266 std::vector<Constant*> CV;
5267 if (EltVT == MVT::f64) {
5268 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5272 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5278 Constant *C = ConstantVector::get(CV);
5279 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5280 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5281 PseudoSourceValue::getConstantPool(), 0,
5283 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5286 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5287 DebugLoc dl = Op.getDebugLoc();
5288 MVT VT = Op.getValueType();
5290 unsigned EltNum = 1;
5291 if (VT.isVector()) {
5292 EltVT = VT.getVectorElementType();
5293 EltNum = VT.getVectorNumElements();
5295 std::vector<Constant*> CV;
5296 if (EltVT == MVT::f64) {
5297 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5301 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5307 Constant *C = ConstantVector::get(CV);
5308 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5309 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5310 PseudoSourceValue::getConstantPool(), 0,
5312 if (VT.isVector()) {
5313 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5314 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5315 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5317 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5319 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5323 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5324 SDValue Op0 = Op.getOperand(0);
5325 SDValue Op1 = Op.getOperand(1);
5326 DebugLoc dl = Op.getDebugLoc();
5327 MVT VT = Op.getValueType();
5328 MVT SrcVT = Op1.getValueType();
5330 // If second operand is smaller, extend it first.
5331 if (SrcVT.bitsLT(VT)) {
5332 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5335 // And if it is bigger, shrink it first.
5336 if (SrcVT.bitsGT(VT)) {
5337 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5341 // At this point the operands and the result should have the same
5342 // type, and that won't be f80 since that is not custom lowered.
5344 // First get the sign bit of second operand.
5345 std::vector<Constant*> CV;
5346 if (SrcVT == MVT::f64) {
5347 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5348 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5350 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5351 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5352 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5353 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5355 Constant *C = ConstantVector::get(CV);
5356 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5357 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5358 PseudoSourceValue::getConstantPool(), 0,
5360 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5362 // Shift sign bit right or left if the two operands have different types.
5363 if (SrcVT.bitsGT(VT)) {
5364 // Op0 is MVT::f32, Op1 is MVT::f64.
5365 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5366 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5367 DAG.getConstant(32, MVT::i32));
5368 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5369 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5370 DAG.getIntPtrConstant(0));
5373 // Clear first operand sign bit.
5375 if (VT == MVT::f64) {
5376 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5377 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5379 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5380 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5381 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5382 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5384 C = ConstantVector::get(CV);
5385 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5386 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5387 PseudoSourceValue::getConstantPool(), 0,
5389 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5391 // Or the value with the sign bit.
5392 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5395 /// Emit nodes that will be selected as "test Op0,Op0", or something
5397 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5398 SelectionDAG &DAG) {
5399 DebugLoc dl = Op.getDebugLoc();
5401 // CF and OF aren't always set the way we want. Determine which
5402 // of these we need.
5403 bool NeedCF = false;
5404 bool NeedOF = false;
5406 case X86::COND_A: case X86::COND_AE:
5407 case X86::COND_B: case X86::COND_BE:
5410 case X86::COND_G: case X86::COND_GE:
5411 case X86::COND_L: case X86::COND_LE:
5412 case X86::COND_O: case X86::COND_NO:
5418 // See if we can use the EFLAGS value from the operand instead of
5419 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5420 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5421 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5422 unsigned Opcode = 0;
5423 unsigned NumOperands = 0;
5424 switch (Op.getNode()->getOpcode()) {
5426 // Due to an isel shortcoming, be conservative if this add is likely to
5427 // be selected as part of a load-modify-store instruction. When the root
5428 // node in a match is a store, isel doesn't know how to remap non-chain
5429 // non-flag uses of other nodes in the match, such as the ADD in this
5430 // case. This leads to the ADD being left around and reselected, with
5431 // the result being two adds in the output.
5432 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5433 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5434 if (UI->getOpcode() == ISD::STORE)
5436 if (ConstantSDNode *C =
5437 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5438 // An add of one will be selected as an INC.
5439 if (C->getAPIntValue() == 1) {
5440 Opcode = X86ISD::INC;
5444 // An add of negative one (subtract of one) will be selected as a DEC.
5445 if (C->getAPIntValue().isAllOnesValue()) {
5446 Opcode = X86ISD::DEC;
5451 // Otherwise use a regular EFLAGS-setting add.
5452 Opcode = X86ISD::ADD;
5456 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5457 // likely to be selected as part of a load-modify-store instruction.
5458 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5459 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5460 if (UI->getOpcode() == ISD::STORE)
5462 // Otherwise use a regular EFLAGS-setting sub.
5463 Opcode = X86ISD::SUB;
5470 return SDValue(Op.getNode(), 1);
5476 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::i32);
5477 SmallVector<SDValue, 4> Ops;
5478 for (unsigned i = 0; i != NumOperands; ++i)
5479 Ops.push_back(Op.getOperand(i));
5480 SDValue New = DAG.getNode(Opcode, dl, VTs, 2, &Ops[0], NumOperands);
5481 DAG.ReplaceAllUsesWith(Op, New);
5482 return SDValue(New.getNode(), 1);
5486 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5487 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5488 DAG.getConstant(0, Op.getValueType()));
5491 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5493 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5494 SelectionDAG &DAG) {
5495 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5496 if (C->getAPIntValue() == 0)
5497 return EmitTest(Op0, X86CC, DAG);
5499 DebugLoc dl = Op0.getDebugLoc();
5500 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5503 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5504 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5505 SDValue Op0 = Op.getOperand(0);
5506 SDValue Op1 = Op.getOperand(1);
5507 DebugLoc dl = Op.getDebugLoc();
5508 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5510 // Lower (X & (1 << N)) == 0 to BT(X, N).
5511 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5512 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5513 if (Op0.getOpcode() == ISD::AND &&
5515 Op1.getOpcode() == ISD::Constant &&
5516 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5517 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5519 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5520 if (ConstantSDNode *Op010C =
5521 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5522 if (Op010C->getZExtValue() == 1) {
5523 LHS = Op0.getOperand(0);
5524 RHS = Op0.getOperand(1).getOperand(1);
5526 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5527 if (ConstantSDNode *Op000C =
5528 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5529 if (Op000C->getZExtValue() == 1) {
5530 LHS = Op0.getOperand(1);
5531 RHS = Op0.getOperand(0).getOperand(1);
5533 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5534 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5535 SDValue AndLHS = Op0.getOperand(0);
5536 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5537 LHS = AndLHS.getOperand(0);
5538 RHS = AndLHS.getOperand(1);
5542 if (LHS.getNode()) {
5543 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5544 // instruction. Since the shift amount is in-range-or-undefined, we know
5545 // that doing a bittest on the i16 value is ok. We extend to i32 because
5546 // the encoding for the i16 version is larger than the i32 version.
5547 if (LHS.getValueType() == MVT::i8)
5548 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5550 // If the operand types disagree, extend the shift amount to match. Since
5551 // BT ignores high bits (like shifts) we can use anyextend.
5552 if (LHS.getValueType() != RHS.getValueType())
5553 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5555 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5556 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5557 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5558 DAG.getConstant(Cond, MVT::i8), BT);
5562 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5563 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5565 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5566 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5567 DAG.getConstant(X86CC, MVT::i8), Cond);
5570 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5572 SDValue Op0 = Op.getOperand(0);
5573 SDValue Op1 = Op.getOperand(1);
5574 SDValue CC = Op.getOperand(2);
5575 MVT VT = Op.getValueType();
5576 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5577 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5578 DebugLoc dl = Op.getDebugLoc();
5582 MVT VT0 = Op0.getValueType();
5583 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5584 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5587 switch (SetCCOpcode) {
5590 case ISD::SETEQ: SSECC = 0; break;
5592 case ISD::SETGT: Swap = true; // Fallthrough
5594 case ISD::SETOLT: SSECC = 1; break;
5596 case ISD::SETGE: Swap = true; // Fallthrough
5598 case ISD::SETOLE: SSECC = 2; break;
5599 case ISD::SETUO: SSECC = 3; break;
5601 case ISD::SETNE: SSECC = 4; break;
5602 case ISD::SETULE: Swap = true;
5603 case ISD::SETUGE: SSECC = 5; break;
5604 case ISD::SETULT: Swap = true;
5605 case ISD::SETUGT: SSECC = 6; break;
5606 case ISD::SETO: SSECC = 7; break;
5609 std::swap(Op0, Op1);
5611 // In the two special cases we can't handle, emit two comparisons.
5613 if (SetCCOpcode == ISD::SETUEQ) {
5615 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5616 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5617 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5619 else if (SetCCOpcode == ISD::SETONE) {
5621 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5622 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5623 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5625 assert(0 && "Illegal FP comparison");
5627 // Handle all other FP comparisons here.
5628 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5631 // We are handling one of the integer comparisons here. Since SSE only has
5632 // GT and EQ comparisons for integer, swapping operands and multiple
5633 // operations may be required for some comparisons.
5634 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5635 bool Swap = false, Invert = false, FlipSigns = false;
5637 switch (VT.getSimpleVT()) {
5639 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5640 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5641 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5642 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5645 switch (SetCCOpcode) {
5647 case ISD::SETNE: Invert = true;
5648 case ISD::SETEQ: Opc = EQOpc; break;
5649 case ISD::SETLT: Swap = true;
5650 case ISD::SETGT: Opc = GTOpc; break;
5651 case ISD::SETGE: Swap = true;
5652 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5653 case ISD::SETULT: Swap = true;
5654 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5655 case ISD::SETUGE: Swap = true;
5656 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5659 std::swap(Op0, Op1);
5661 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5662 // bits of the inputs before performing those operations.
5664 MVT EltVT = VT.getVectorElementType();
5665 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5667 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5668 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5670 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5671 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5674 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5676 // If the logical-not of the result is required, perform that now.
5678 Result = DAG.getNOT(dl, Result, VT);
5683 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5684 static bool isX86LogicalCmp(SDValue Op) {
5685 unsigned Opc = Op.getNode()->getOpcode();
5686 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5688 if (Op.getResNo() == 1 &&
5689 (Opc == X86ISD::ADD ||
5690 Opc == X86ISD::SUB ||
5691 Opc == X86ISD::SMUL ||
5692 Opc == X86ISD::UMUL ||
5693 Opc == X86ISD::INC ||
5694 Opc == X86ISD::DEC))
5700 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5701 bool addTest = true;
5702 SDValue Cond = Op.getOperand(0);
5703 DebugLoc dl = Op.getDebugLoc();
5706 if (Cond.getOpcode() == ISD::SETCC)
5707 Cond = LowerSETCC(Cond, DAG);
5709 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5710 // setting operand in place of the X86ISD::SETCC.
5711 if (Cond.getOpcode() == X86ISD::SETCC) {
5712 CC = Cond.getOperand(0);
5714 SDValue Cmp = Cond.getOperand(1);
5715 unsigned Opc = Cmp.getOpcode();
5716 MVT VT = Op.getValueType();
5718 bool IllegalFPCMov = false;
5719 if (VT.isFloatingPoint() && !VT.isVector() &&
5720 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5721 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5723 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5724 Opc == X86ISD::BT) { // FIXME
5731 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5732 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5735 const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
5737 SmallVector<SDValue, 4> Ops;
5738 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5739 // condition is true.
5740 Ops.push_back(Op.getOperand(2));
5741 Ops.push_back(Op.getOperand(1));
5743 Ops.push_back(Cond);
5744 return DAG.getNode(X86ISD::CMOV, dl, VTs, 2, &Ops[0], Ops.size());
5747 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5748 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5749 // from the AND / OR.
5750 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5751 Opc = Op.getOpcode();
5752 if (Opc != ISD::OR && Opc != ISD::AND)
5754 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5755 Op.getOperand(0).hasOneUse() &&
5756 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5757 Op.getOperand(1).hasOneUse());
5760 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5761 // 1 and that the SETCC node has a single use.
5762 static bool isXor1OfSetCC(SDValue Op) {
5763 if (Op.getOpcode() != ISD::XOR)
5765 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5766 if (N1C && N1C->getAPIntValue() == 1) {
5767 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5768 Op.getOperand(0).hasOneUse();
5773 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5774 bool addTest = true;
5775 SDValue Chain = Op.getOperand(0);
5776 SDValue Cond = Op.getOperand(1);
5777 SDValue Dest = Op.getOperand(2);
5778 DebugLoc dl = Op.getDebugLoc();
5781 if (Cond.getOpcode() == ISD::SETCC)
5782 Cond = LowerSETCC(Cond, DAG);
5784 // FIXME: LowerXALUO doesn't handle these!!
5785 else if (Cond.getOpcode() == X86ISD::ADD ||
5786 Cond.getOpcode() == X86ISD::SUB ||
5787 Cond.getOpcode() == X86ISD::SMUL ||
5788 Cond.getOpcode() == X86ISD::UMUL)
5789 Cond = LowerXALUO(Cond, DAG);
5792 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5793 // setting operand in place of the X86ISD::SETCC.
5794 if (Cond.getOpcode() == X86ISD::SETCC) {
5795 CC = Cond.getOperand(0);
5797 SDValue Cmp = Cond.getOperand(1);
5798 unsigned Opc = Cmp.getOpcode();
5799 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5800 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5804 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5808 // These can only come from an arithmetic instruction with overflow,
5809 // e.g. SADDO, UADDO.
5810 Cond = Cond.getNode()->getOperand(1);
5817 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5818 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5819 if (CondOpc == ISD::OR) {
5820 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5821 // two branches instead of an explicit OR instruction with a
5823 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5824 isX86LogicalCmp(Cmp)) {
5825 CC = Cond.getOperand(0).getOperand(0);
5826 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5827 Chain, Dest, CC, Cmp);
5828 CC = Cond.getOperand(1).getOperand(0);
5832 } else { // ISD::AND
5833 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5834 // two branches instead of an explicit AND instruction with a
5835 // separate test. However, we only do this if this block doesn't
5836 // have a fall-through edge, because this requires an explicit
5837 // jmp when the condition is false.
5838 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5839 isX86LogicalCmp(Cmp) &&
5840 Op.getNode()->hasOneUse()) {
5841 X86::CondCode CCode =
5842 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5843 CCode = X86::GetOppositeBranchCondition(CCode);
5844 CC = DAG.getConstant(CCode, MVT::i8);
5845 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5846 // Look for an unconditional branch following this conditional branch.
5847 // We need this because we need to reverse the successors in order
5848 // to implement FCMP_OEQ.
5849 if (User.getOpcode() == ISD::BR) {
5850 SDValue FalseBB = User.getOperand(1);
5852 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5853 assert(NewBR == User);
5856 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5857 Chain, Dest, CC, Cmp);
5858 X86::CondCode CCode =
5859 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5860 CCode = X86::GetOppositeBranchCondition(CCode);
5861 CC = DAG.getConstant(CCode, MVT::i8);
5867 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5868 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5869 // It should be transformed during dag combiner except when the condition
5870 // is set by a arithmetics with overflow node.
5871 X86::CondCode CCode =
5872 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5873 CCode = X86::GetOppositeBranchCondition(CCode);
5874 CC = DAG.getConstant(CCode, MVT::i8);
5875 Cond = Cond.getOperand(0).getOperand(1);
5881 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5882 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5884 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5885 Chain, Dest, CC, Cond);
5889 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5890 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5891 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5892 // that the guard pages used by the OS virtual memory manager are allocated in
5893 // correct sequence.
5895 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5896 SelectionDAG &DAG) {
5897 assert(Subtarget->isTargetCygMing() &&
5898 "This should be used only on Cygwin/Mingw targets");
5899 DebugLoc dl = Op.getDebugLoc();
5902 SDValue Chain = Op.getOperand(0);
5903 SDValue Size = Op.getOperand(1);
5904 // FIXME: Ensure alignment here
5908 MVT IntPtr = getPointerTy();
5909 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5911 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5913 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5914 Flag = Chain.getValue(1);
5916 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5917 SDValue Ops[] = { Chain,
5918 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5919 DAG.getRegister(X86::EAX, IntPtr),
5920 DAG.getRegister(X86StackPtr, SPTy),
5922 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5923 Flag = Chain.getValue(1);
5925 Chain = DAG.getCALLSEQ_END(Chain,
5926 DAG.getIntPtrConstant(0, true),
5927 DAG.getIntPtrConstant(0, true),
5930 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5932 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5933 return DAG.getMergeValues(Ops1, 2, dl);
5937 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5939 SDValue Dst, SDValue Src,
5940 SDValue Size, unsigned Align,
5942 uint64_t DstSVOff) {
5943 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5945 // If not DWORD aligned or size is more than the threshold, call the library.
5946 // The libc version is likely to be faster for these cases. It can use the
5947 // address value and run time information about the CPU.
5948 if ((Align & 3) != 0 ||
5950 ConstantSize->getZExtValue() >
5951 getSubtarget()->getMaxInlineSizeThreshold()) {
5952 SDValue InFlag(0, 0);
5954 // Check to see if there is a specialized entry-point for memory zeroing.
5955 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5957 if (const char *bzeroEntry = V &&
5958 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5959 MVT IntPtr = getPointerTy();
5960 const Type *IntPtrTy = TD->getIntPtrType();
5961 TargetLowering::ArgListTy Args;
5962 TargetLowering::ArgListEntry Entry;
5964 Entry.Ty = IntPtrTy;
5965 Args.push_back(Entry);
5967 Args.push_back(Entry);
5968 std::pair<SDValue,SDValue> CallResult =
5969 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5970 CallingConv::C, false,
5971 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5972 return CallResult.second;
5975 // Otherwise have the target-independent code call memset.
5979 uint64_t SizeVal = ConstantSize->getZExtValue();
5980 SDValue InFlag(0, 0);
5983 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5984 unsigned BytesLeft = 0;
5985 bool TwoRepStos = false;
5988 uint64_t Val = ValC->getZExtValue() & 255;
5990 // If the value is a constant, then we can potentially use larger sets.
5991 switch (Align & 3) {
5992 case 2: // WORD aligned
5995 Val = (Val << 8) | Val;
5997 case 0: // DWORD aligned
6000 Val = (Val << 8) | Val;
6001 Val = (Val << 16) | Val;
6002 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6005 Val = (Val << 32) | Val;
6008 default: // Byte aligned
6011 Count = DAG.getIntPtrConstant(SizeVal);
6015 if (AVT.bitsGT(MVT::i8)) {
6016 unsigned UBytes = AVT.getSizeInBits() / 8;
6017 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6018 BytesLeft = SizeVal % UBytes;
6021 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6023 InFlag = Chain.getValue(1);
6026 Count = DAG.getIntPtrConstant(SizeVal);
6027 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6028 InFlag = Chain.getValue(1);
6031 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6034 InFlag = Chain.getValue(1);
6035 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6038 InFlag = Chain.getValue(1);
6040 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6041 SmallVector<SDValue, 8> Ops;
6042 Ops.push_back(Chain);
6043 Ops.push_back(DAG.getValueType(AVT));
6044 Ops.push_back(InFlag);
6045 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6048 InFlag = Chain.getValue(1);
6050 MVT CVT = Count.getValueType();
6051 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6052 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6053 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6056 InFlag = Chain.getValue(1);
6057 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6059 Ops.push_back(Chain);
6060 Ops.push_back(DAG.getValueType(MVT::i8));
6061 Ops.push_back(InFlag);
6062 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6063 } else if (BytesLeft) {
6064 // Handle the last 1 - 7 bytes.
6065 unsigned Offset = SizeVal - BytesLeft;
6066 MVT AddrVT = Dst.getValueType();
6067 MVT SizeVT = Size.getValueType();
6069 Chain = DAG.getMemset(Chain, dl,
6070 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6071 DAG.getConstant(Offset, AddrVT)),
6073 DAG.getConstant(BytesLeft, SizeVT),
6074 Align, DstSV, DstSVOff + Offset);
6077 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6082 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6083 SDValue Chain, SDValue Dst, SDValue Src,
6084 SDValue Size, unsigned Align,
6086 const Value *DstSV, uint64_t DstSVOff,
6087 const Value *SrcSV, uint64_t SrcSVOff) {
6088 // This requires the copy size to be a constant, preferrably
6089 // within a subtarget-specific limit.
6090 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6093 uint64_t SizeVal = ConstantSize->getZExtValue();
6094 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6097 /// If not DWORD aligned, call the library.
6098 if ((Align & 3) != 0)
6103 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6106 unsigned UBytes = AVT.getSizeInBits() / 8;
6107 unsigned CountVal = SizeVal / UBytes;
6108 SDValue Count = DAG.getIntPtrConstant(CountVal);
6109 unsigned BytesLeft = SizeVal % UBytes;
6111 SDValue InFlag(0, 0);
6112 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6115 InFlag = Chain.getValue(1);
6116 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6119 InFlag = Chain.getValue(1);
6120 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6123 InFlag = Chain.getValue(1);
6125 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6126 SmallVector<SDValue, 8> Ops;
6127 Ops.push_back(Chain);
6128 Ops.push_back(DAG.getValueType(AVT));
6129 Ops.push_back(InFlag);
6130 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6132 SmallVector<SDValue, 4> Results;
6133 Results.push_back(RepMovs);
6135 // Handle the last 1 - 7 bytes.
6136 unsigned Offset = SizeVal - BytesLeft;
6137 MVT DstVT = Dst.getValueType();
6138 MVT SrcVT = Src.getValueType();
6139 MVT SizeVT = Size.getValueType();
6140 Results.push_back(DAG.getMemcpy(Chain, dl,
6141 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6142 DAG.getConstant(Offset, DstVT)),
6143 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6144 DAG.getConstant(Offset, SrcVT)),
6145 DAG.getConstant(BytesLeft, SizeVT),
6146 Align, AlwaysInline,
6147 DstSV, DstSVOff + Offset,
6148 SrcSV, SrcSVOff + Offset));
6151 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6152 &Results[0], Results.size());
6155 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6156 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6157 DebugLoc dl = Op.getDebugLoc();
6159 if (!Subtarget->is64Bit()) {
6160 // vastart just stores the address of the VarArgsFrameIndex slot into the
6161 // memory location argument.
6162 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6163 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6167 // gp_offset (0 - 6 * 8)
6168 // fp_offset (48 - 48 + 8 * 16)
6169 // overflow_arg_area (point to parameters coming in memory).
6171 SmallVector<SDValue, 8> MemOps;
6172 SDValue FIN = Op.getOperand(1);
6174 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6175 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6177 MemOps.push_back(Store);
6180 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6181 FIN, DAG.getIntPtrConstant(4));
6182 Store = DAG.getStore(Op.getOperand(0), dl,
6183 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6185 MemOps.push_back(Store);
6187 // Store ptr to overflow_arg_area
6188 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6189 FIN, DAG.getIntPtrConstant(4));
6190 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6191 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6192 MemOps.push_back(Store);
6194 // Store ptr to reg_save_area.
6195 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6196 FIN, DAG.getIntPtrConstant(8));
6197 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6198 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6199 MemOps.push_back(Store);
6200 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6201 &MemOps[0], MemOps.size());
6204 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6205 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6206 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6207 SDValue Chain = Op.getOperand(0);
6208 SDValue SrcPtr = Op.getOperand(1);
6209 SDValue SrcSV = Op.getOperand(2);
6211 assert(0 && "VAArgInst is not yet implemented for x86-64!");
6216 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6217 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6218 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6219 SDValue Chain = Op.getOperand(0);
6220 SDValue DstPtr = Op.getOperand(1);
6221 SDValue SrcPtr = Op.getOperand(2);
6222 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6223 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6224 DebugLoc dl = Op.getDebugLoc();
6226 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6227 DAG.getIntPtrConstant(24), 8, false,
6228 DstSV, 0, SrcSV, 0);
6232 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6233 DebugLoc dl = Op.getDebugLoc();
6234 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6236 default: return SDValue(); // Don't custom lower most intrinsics.
6237 // Comparison intrinsics.
6238 case Intrinsic::x86_sse_comieq_ss:
6239 case Intrinsic::x86_sse_comilt_ss:
6240 case Intrinsic::x86_sse_comile_ss:
6241 case Intrinsic::x86_sse_comigt_ss:
6242 case Intrinsic::x86_sse_comige_ss:
6243 case Intrinsic::x86_sse_comineq_ss:
6244 case Intrinsic::x86_sse_ucomieq_ss:
6245 case Intrinsic::x86_sse_ucomilt_ss:
6246 case Intrinsic::x86_sse_ucomile_ss:
6247 case Intrinsic::x86_sse_ucomigt_ss:
6248 case Intrinsic::x86_sse_ucomige_ss:
6249 case Intrinsic::x86_sse_ucomineq_ss:
6250 case Intrinsic::x86_sse2_comieq_sd:
6251 case Intrinsic::x86_sse2_comilt_sd:
6252 case Intrinsic::x86_sse2_comile_sd:
6253 case Intrinsic::x86_sse2_comigt_sd:
6254 case Intrinsic::x86_sse2_comige_sd:
6255 case Intrinsic::x86_sse2_comineq_sd:
6256 case Intrinsic::x86_sse2_ucomieq_sd:
6257 case Intrinsic::x86_sse2_ucomilt_sd:
6258 case Intrinsic::x86_sse2_ucomile_sd:
6259 case Intrinsic::x86_sse2_ucomigt_sd:
6260 case Intrinsic::x86_sse2_ucomige_sd:
6261 case Intrinsic::x86_sse2_ucomineq_sd: {
6263 ISD::CondCode CC = ISD::SETCC_INVALID;
6266 case Intrinsic::x86_sse_comieq_ss:
6267 case Intrinsic::x86_sse2_comieq_sd:
6271 case Intrinsic::x86_sse_comilt_ss:
6272 case Intrinsic::x86_sse2_comilt_sd:
6276 case Intrinsic::x86_sse_comile_ss:
6277 case Intrinsic::x86_sse2_comile_sd:
6281 case Intrinsic::x86_sse_comigt_ss:
6282 case Intrinsic::x86_sse2_comigt_sd:
6286 case Intrinsic::x86_sse_comige_ss:
6287 case Intrinsic::x86_sse2_comige_sd:
6291 case Intrinsic::x86_sse_comineq_ss:
6292 case Intrinsic::x86_sse2_comineq_sd:
6296 case Intrinsic::x86_sse_ucomieq_ss:
6297 case Intrinsic::x86_sse2_ucomieq_sd:
6298 Opc = X86ISD::UCOMI;
6301 case Intrinsic::x86_sse_ucomilt_ss:
6302 case Intrinsic::x86_sse2_ucomilt_sd:
6303 Opc = X86ISD::UCOMI;
6306 case Intrinsic::x86_sse_ucomile_ss:
6307 case Intrinsic::x86_sse2_ucomile_sd:
6308 Opc = X86ISD::UCOMI;
6311 case Intrinsic::x86_sse_ucomigt_ss:
6312 case Intrinsic::x86_sse2_ucomigt_sd:
6313 Opc = X86ISD::UCOMI;
6316 case Intrinsic::x86_sse_ucomige_ss:
6317 case Intrinsic::x86_sse2_ucomige_sd:
6318 Opc = X86ISD::UCOMI;
6321 case Intrinsic::x86_sse_ucomineq_ss:
6322 case Intrinsic::x86_sse2_ucomineq_sd:
6323 Opc = X86ISD::UCOMI;
6328 SDValue LHS = Op.getOperand(1);
6329 SDValue RHS = Op.getOperand(2);
6330 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6331 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6332 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6333 DAG.getConstant(X86CC, MVT::i8), Cond);
6334 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6337 // Fix vector shift instructions where the last operand is a non-immediate
6339 case Intrinsic::x86_sse2_pslli_w:
6340 case Intrinsic::x86_sse2_pslli_d:
6341 case Intrinsic::x86_sse2_pslli_q:
6342 case Intrinsic::x86_sse2_psrli_w:
6343 case Intrinsic::x86_sse2_psrli_d:
6344 case Intrinsic::x86_sse2_psrli_q:
6345 case Intrinsic::x86_sse2_psrai_w:
6346 case Intrinsic::x86_sse2_psrai_d:
6347 case Intrinsic::x86_mmx_pslli_w:
6348 case Intrinsic::x86_mmx_pslli_d:
6349 case Intrinsic::x86_mmx_pslli_q:
6350 case Intrinsic::x86_mmx_psrli_w:
6351 case Intrinsic::x86_mmx_psrli_d:
6352 case Intrinsic::x86_mmx_psrli_q:
6353 case Intrinsic::x86_mmx_psrai_w:
6354 case Intrinsic::x86_mmx_psrai_d: {
6355 SDValue ShAmt = Op.getOperand(2);
6356 if (isa<ConstantSDNode>(ShAmt))
6359 unsigned NewIntNo = 0;
6360 MVT ShAmtVT = MVT::v4i32;
6362 case Intrinsic::x86_sse2_pslli_w:
6363 NewIntNo = Intrinsic::x86_sse2_psll_w;
6365 case Intrinsic::x86_sse2_pslli_d:
6366 NewIntNo = Intrinsic::x86_sse2_psll_d;
6368 case Intrinsic::x86_sse2_pslli_q:
6369 NewIntNo = Intrinsic::x86_sse2_psll_q;
6371 case Intrinsic::x86_sse2_psrli_w:
6372 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6374 case Intrinsic::x86_sse2_psrli_d:
6375 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6377 case Intrinsic::x86_sse2_psrli_q:
6378 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6380 case Intrinsic::x86_sse2_psrai_w:
6381 NewIntNo = Intrinsic::x86_sse2_psra_w;
6383 case Intrinsic::x86_sse2_psrai_d:
6384 NewIntNo = Intrinsic::x86_sse2_psra_d;
6387 ShAmtVT = MVT::v2i32;
6389 case Intrinsic::x86_mmx_pslli_w:
6390 NewIntNo = Intrinsic::x86_mmx_psll_w;
6392 case Intrinsic::x86_mmx_pslli_d:
6393 NewIntNo = Intrinsic::x86_mmx_psll_d;
6395 case Intrinsic::x86_mmx_pslli_q:
6396 NewIntNo = Intrinsic::x86_mmx_psll_q;
6398 case Intrinsic::x86_mmx_psrli_w:
6399 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6401 case Intrinsic::x86_mmx_psrli_d:
6402 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6404 case Intrinsic::x86_mmx_psrli_q:
6405 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6407 case Intrinsic::x86_mmx_psrai_w:
6408 NewIntNo = Intrinsic::x86_mmx_psra_w;
6410 case Intrinsic::x86_mmx_psrai_d:
6411 NewIntNo = Intrinsic::x86_mmx_psra_d;
6413 default: abort(); // Can't reach here.
6418 MVT VT = Op.getValueType();
6419 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6420 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6421 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6422 DAG.getConstant(NewIntNo, MVT::i32),
6423 Op.getOperand(1), ShAmt);
6428 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6429 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6430 DebugLoc dl = Op.getDebugLoc();
6433 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6435 DAG.getConstant(TD->getPointerSize(),
6436 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6437 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6438 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6443 // Just load the return address.
6444 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6445 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6446 RetAddrFI, NULL, 0);
6449 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6450 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6451 MFI->setFrameAddressIsTaken(true);
6452 MVT VT = Op.getValueType();
6453 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6454 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6455 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6456 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6458 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6462 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6463 SelectionDAG &DAG) {
6464 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6467 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6469 MachineFunction &MF = DAG.getMachineFunction();
6470 SDValue Chain = Op.getOperand(0);
6471 SDValue Offset = Op.getOperand(1);
6472 SDValue Handler = Op.getOperand(2);
6473 DebugLoc dl = Op.getDebugLoc();
6475 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6477 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6479 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6480 DAG.getIntPtrConstant(-TD->getPointerSize()));
6481 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6482 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6483 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6484 MF.getRegInfo().addLiveOut(StoreAddrReg);
6486 return DAG.getNode(X86ISD::EH_RETURN, dl,
6488 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6491 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6492 SelectionDAG &DAG) {
6493 SDValue Root = Op.getOperand(0);
6494 SDValue Trmp = Op.getOperand(1); // trampoline
6495 SDValue FPtr = Op.getOperand(2); // nested function
6496 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6497 DebugLoc dl = Op.getDebugLoc();
6499 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6501 const X86InstrInfo *TII =
6502 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6504 if (Subtarget->is64Bit()) {
6505 SDValue OutChains[6];
6507 // Large code-model.
6509 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6510 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6512 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6513 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6515 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6517 // Load the pointer to the nested function into R11.
6518 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6519 SDValue Addr = Trmp;
6520 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6523 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6524 DAG.getConstant(2, MVT::i64));
6525 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6527 // Load the 'nest' parameter value into R10.
6528 // R10 is specified in X86CallingConv.td
6529 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6530 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6531 DAG.getConstant(10, MVT::i64));
6532 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6533 Addr, TrmpAddr, 10);
6535 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6536 DAG.getConstant(12, MVT::i64));
6537 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6539 // Jump to the nested function.
6540 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6541 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6542 DAG.getConstant(20, MVT::i64));
6543 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6544 Addr, TrmpAddr, 20);
6546 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6547 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6548 DAG.getConstant(22, MVT::i64));
6549 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6553 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6554 return DAG.getMergeValues(Ops, 2, dl);
6556 const Function *Func =
6557 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6558 unsigned CC = Func->getCallingConv();
6563 assert(0 && "Unsupported calling convention");
6564 case CallingConv::C:
6565 case CallingConv::X86_StdCall: {
6566 // Pass 'nest' parameter in ECX.
6567 // Must be kept in sync with X86CallingConv.td
6570 // Check that ECX wasn't needed by an 'inreg' parameter.
6571 const FunctionType *FTy = Func->getFunctionType();
6572 const AttrListPtr &Attrs = Func->getAttributes();
6574 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6575 unsigned InRegCount = 0;
6578 for (FunctionType::param_iterator I = FTy->param_begin(),
6579 E = FTy->param_end(); I != E; ++I, ++Idx)
6580 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6581 // FIXME: should only count parameters that are lowered to integers.
6582 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6584 if (InRegCount > 2) {
6585 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6591 case CallingConv::X86_FastCall:
6592 case CallingConv::Fast:
6593 // Pass 'nest' parameter in EAX.
6594 // Must be kept in sync with X86CallingConv.td
6599 SDValue OutChains[4];
6602 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6603 DAG.getConstant(10, MVT::i32));
6604 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6606 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6607 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6608 OutChains[0] = DAG.getStore(Root, dl,
6609 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6612 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6613 DAG.getConstant(1, MVT::i32));
6614 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6616 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6617 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6618 DAG.getConstant(5, MVT::i32));
6619 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6620 TrmpAddr, 5, false, 1);
6622 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6623 DAG.getConstant(6, MVT::i32));
6624 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6627 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6628 return DAG.getMergeValues(Ops, 2, dl);
6632 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6634 The rounding mode is in bits 11:10 of FPSR, and has the following
6641 FLT_ROUNDS, on the other hand, expects the following:
6648 To perform the conversion, we do:
6649 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6652 MachineFunction &MF = DAG.getMachineFunction();
6653 const TargetMachine &TM = MF.getTarget();
6654 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6655 unsigned StackAlignment = TFI.getStackAlignment();
6656 MVT VT = Op.getValueType();
6657 DebugLoc dl = Op.getDebugLoc();
6659 // Save FP Control Word to stack slot
6660 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6661 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6663 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6664 DAG.getEntryNode(), StackSlot);
6666 // Load FP Control Word from stack slot
6667 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6669 // Transform as necessary
6671 DAG.getNode(ISD::SRL, dl, MVT::i16,
6672 DAG.getNode(ISD::AND, dl, MVT::i16,
6673 CWD, DAG.getConstant(0x800, MVT::i16)),
6674 DAG.getConstant(11, MVT::i8));
6676 DAG.getNode(ISD::SRL, dl, MVT::i16,
6677 DAG.getNode(ISD::AND, dl, MVT::i16,
6678 CWD, DAG.getConstant(0x400, MVT::i16)),
6679 DAG.getConstant(9, MVT::i8));
6682 DAG.getNode(ISD::AND, dl, MVT::i16,
6683 DAG.getNode(ISD::ADD, dl, MVT::i16,
6684 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6685 DAG.getConstant(1, MVT::i16)),
6686 DAG.getConstant(3, MVT::i16));
6689 return DAG.getNode((VT.getSizeInBits() < 16 ?
6690 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6693 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6694 MVT VT = Op.getValueType();
6696 unsigned NumBits = VT.getSizeInBits();
6697 DebugLoc dl = Op.getDebugLoc();
6699 Op = Op.getOperand(0);
6700 if (VT == MVT::i8) {
6701 // Zero extend to i32 since there is not an i8 bsr.
6703 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6706 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6707 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6708 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6710 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6711 SmallVector<SDValue, 4> Ops;
6713 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6714 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6715 Ops.push_back(Op.getValue(1));
6716 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6718 // Finally xor with NumBits-1.
6719 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6722 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6726 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6727 MVT VT = Op.getValueType();
6729 unsigned NumBits = VT.getSizeInBits();
6730 DebugLoc dl = Op.getDebugLoc();
6732 Op = Op.getOperand(0);
6733 if (VT == MVT::i8) {
6735 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6738 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6739 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6740 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6742 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6743 SmallVector<SDValue, 4> Ops;
6745 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6746 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6747 Ops.push_back(Op.getValue(1));
6748 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6751 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6755 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6756 MVT VT = Op.getValueType();
6757 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6758 DebugLoc dl = Op.getDebugLoc();
6760 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6761 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6762 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6763 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6764 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6766 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6767 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6768 // return AloBlo + AloBhi + AhiBlo;
6770 SDValue A = Op.getOperand(0);
6771 SDValue B = Op.getOperand(1);
6773 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6774 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6775 A, DAG.getConstant(32, MVT::i32));
6776 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6777 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6778 B, DAG.getConstant(32, MVT::i32));
6779 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6780 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6782 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6783 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6785 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6786 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6788 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6789 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6790 AloBhi, DAG.getConstant(32, MVT::i32));
6791 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6792 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6793 AhiBlo, DAG.getConstant(32, MVT::i32));
6794 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6795 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6800 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6801 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6802 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6803 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6804 // has only one use.
6805 SDNode *N = Op.getNode();
6806 SDValue LHS = N->getOperand(0);
6807 SDValue RHS = N->getOperand(1);
6808 unsigned BaseOp = 0;
6810 DebugLoc dl = Op.getDebugLoc();
6812 switch (Op.getOpcode()) {
6813 default: assert(0 && "Unknown ovf instruction!");
6815 // A subtract of one will be selected as a INC. Note that INC doesn't
6816 // set CF, so we can't do this for UADDO.
6817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6818 if (C->getAPIntValue() == 1) {
6819 BaseOp = X86ISD::INC;
6823 BaseOp = X86ISD::ADD;
6827 BaseOp = X86ISD::ADD;
6831 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6832 // set CF, so we can't do this for USUBO.
6833 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6834 if (C->getAPIntValue() == 1) {
6835 BaseOp = X86ISD::DEC;
6839 BaseOp = X86ISD::SUB;
6843 BaseOp = X86ISD::SUB;
6847 BaseOp = X86ISD::SMUL;
6851 BaseOp = X86ISD::UMUL;
6856 // Also sets EFLAGS.
6857 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6858 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6861 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6862 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6864 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6868 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6869 MVT T = Op.getValueType();
6870 DebugLoc dl = Op.getDebugLoc();
6873 switch(T.getSimpleVT()) {
6875 assert(false && "Invalid value type!");
6876 case MVT::i8: Reg = X86::AL; size = 1; break;
6877 case MVT::i16: Reg = X86::AX; size = 2; break;
6878 case MVT::i32: Reg = X86::EAX; size = 4; break;
6880 assert(Subtarget->is64Bit() && "Node not type legal!");
6881 Reg = X86::RAX; size = 8;
6884 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6885 Op.getOperand(2), SDValue());
6886 SDValue Ops[] = { cpIn.getValue(0),
6889 DAG.getTargetConstant(size, MVT::i8),
6891 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6892 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6894 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6898 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6899 SelectionDAG &DAG) {
6900 assert(Subtarget->is64Bit() && "Result not type legalized?");
6901 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6902 SDValue TheChain = Op.getOperand(0);
6903 DebugLoc dl = Op.getDebugLoc();
6904 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6905 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6906 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6908 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6909 DAG.getConstant(32, MVT::i8));
6911 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6914 return DAG.getMergeValues(Ops, 2, dl);
6917 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6918 SDNode *Node = Op.getNode();
6919 DebugLoc dl = Node->getDebugLoc();
6920 MVT T = Node->getValueType(0);
6921 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6922 DAG.getConstant(0, T), Node->getOperand(2));
6923 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6924 cast<AtomicSDNode>(Node)->getMemoryVT(),
6925 Node->getOperand(0),
6926 Node->getOperand(1), negOp,
6927 cast<AtomicSDNode>(Node)->getSrcValue(),
6928 cast<AtomicSDNode>(Node)->getAlignment());
6931 /// LowerOperation - Provide custom lowering hooks for some operations.
6933 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6934 switch (Op.getOpcode()) {
6935 default: assert(0 && "Should not custom lower this!");
6936 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6937 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6938 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6939 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6940 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6941 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6942 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6943 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6944 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6945 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6946 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6947 case ISD::SHL_PARTS:
6948 case ISD::SRA_PARTS:
6949 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6950 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6951 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6952 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6953 case ISD::FABS: return LowerFABS(Op, DAG);
6954 case ISD::FNEG: return LowerFNEG(Op, DAG);
6955 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6956 case ISD::SETCC: return LowerSETCC(Op, DAG);
6957 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6958 case ISD::SELECT: return LowerSELECT(Op, DAG);
6959 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6960 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6961 case ISD::CALL: return LowerCALL(Op, DAG);
6962 case ISD::RET: return LowerRET(Op, DAG);
6963 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6964 case ISD::VASTART: return LowerVASTART(Op, DAG);
6965 case ISD::VAARG: return LowerVAARG(Op, DAG);
6966 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6967 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6968 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6969 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6970 case ISD::FRAME_TO_ARGS_OFFSET:
6971 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6972 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6973 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6974 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6975 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6976 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6977 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6978 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6984 case ISD::UMULO: return LowerXALUO(Op, DAG);
6985 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6989 void X86TargetLowering::
6990 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6991 SelectionDAG &DAG, unsigned NewOp) {
6992 MVT T = Node->getValueType(0);
6993 DebugLoc dl = Node->getDebugLoc();
6994 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6996 SDValue Chain = Node->getOperand(0);
6997 SDValue In1 = Node->getOperand(1);
6998 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6999 Node->getOperand(2), DAG.getIntPtrConstant(0));
7000 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7001 Node->getOperand(2), DAG.getIntPtrConstant(1));
7002 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
7003 // have a MemOperand. Pass the info through as a normal operand.
7004 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
7005 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
7006 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7007 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
7008 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7009 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7010 Results.push_back(Result.getValue(2));
7013 /// ReplaceNodeResults - Replace a node with an illegal result type
7014 /// with a new node built out of custom code.
7015 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7016 SmallVectorImpl<SDValue>&Results,
7017 SelectionDAG &DAG) {
7018 DebugLoc dl = N->getDebugLoc();
7019 switch (N->getOpcode()) {
7021 assert(false && "Do not know how to custom type legalize this operation!");
7023 case ISD::FP_TO_SINT: {
7024 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
7025 SDValue FIST = Vals.first, StackSlot = Vals.second;
7026 if (FIST.getNode() != 0) {
7027 MVT VT = N->getValueType(0);
7028 // Return a load from the stack slot.
7029 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7033 case ISD::READCYCLECOUNTER: {
7034 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7035 SDValue TheChain = N->getOperand(0);
7036 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7037 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7039 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7041 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7042 SDValue Ops[] = { eax, edx };
7043 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7044 Results.push_back(edx.getValue(1));
7047 case ISD::ATOMIC_CMP_SWAP: {
7048 MVT T = N->getValueType(0);
7049 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7050 SDValue cpInL, cpInH;
7051 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7052 DAG.getConstant(0, MVT::i32));
7053 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7054 DAG.getConstant(1, MVT::i32));
7055 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7056 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7058 SDValue swapInL, swapInH;
7059 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7060 DAG.getConstant(0, MVT::i32));
7061 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7062 DAG.getConstant(1, MVT::i32));
7063 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7065 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7066 swapInL.getValue(1));
7067 SDValue Ops[] = { swapInH.getValue(0),
7069 swapInH.getValue(1) };
7070 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7071 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7072 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7073 MVT::i32, Result.getValue(1));
7074 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7075 MVT::i32, cpOutL.getValue(2));
7076 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7077 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7078 Results.push_back(cpOutH.getValue(1));
7081 case ISD::ATOMIC_LOAD_ADD:
7082 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7084 case ISD::ATOMIC_LOAD_AND:
7085 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7087 case ISD::ATOMIC_LOAD_NAND:
7088 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7090 case ISD::ATOMIC_LOAD_OR:
7091 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7093 case ISD::ATOMIC_LOAD_SUB:
7094 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7096 case ISD::ATOMIC_LOAD_XOR:
7097 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7099 case ISD::ATOMIC_SWAP:
7100 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7105 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7107 default: return NULL;
7108 case X86ISD::BSF: return "X86ISD::BSF";
7109 case X86ISD::BSR: return "X86ISD::BSR";
7110 case X86ISD::SHLD: return "X86ISD::SHLD";
7111 case X86ISD::SHRD: return "X86ISD::SHRD";
7112 case X86ISD::FAND: return "X86ISD::FAND";
7113 case X86ISD::FOR: return "X86ISD::FOR";
7114 case X86ISD::FXOR: return "X86ISD::FXOR";
7115 case X86ISD::FSRL: return "X86ISD::FSRL";
7116 case X86ISD::FILD: return "X86ISD::FILD";
7117 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7118 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7119 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7120 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7121 case X86ISD::FLD: return "X86ISD::FLD";
7122 case X86ISD::FST: return "X86ISD::FST";
7123 case X86ISD::CALL: return "X86ISD::CALL";
7124 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
7125 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7126 case X86ISD::BT: return "X86ISD::BT";
7127 case X86ISD::CMP: return "X86ISD::CMP";
7128 case X86ISD::COMI: return "X86ISD::COMI";
7129 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7130 case X86ISD::SETCC: return "X86ISD::SETCC";
7131 case X86ISD::CMOV: return "X86ISD::CMOV";
7132 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7133 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7134 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7135 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7136 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7137 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7138 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7139 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7140 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7141 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7142 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7143 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7144 case X86ISD::FMAX: return "X86ISD::FMAX";
7145 case X86ISD::FMIN: return "X86ISD::FMIN";
7146 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7147 case X86ISD::FRCP: return "X86ISD::FRCP";
7148 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7149 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
7150 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7151 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7152 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7153 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7154 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7155 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7156 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7157 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7158 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7159 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7160 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7161 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7162 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7163 case X86ISD::VSHL: return "X86ISD::VSHL";
7164 case X86ISD::VSRL: return "X86ISD::VSRL";
7165 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7166 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7167 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7168 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7169 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7170 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7171 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7172 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7173 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7174 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7175 case X86ISD::ADD: return "X86ISD::ADD";
7176 case X86ISD::SUB: return "X86ISD::SUB";
7177 case X86ISD::SMUL: return "X86ISD::SMUL";
7178 case X86ISD::UMUL: return "X86ISD::UMUL";
7179 case X86ISD::INC: return "X86ISD::INC";
7180 case X86ISD::DEC: return "X86ISD::DEC";
7184 // isLegalAddressingMode - Return true if the addressing mode represented
7185 // by AM is legal for this target, for a load/store of the specified type.
7186 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7187 const Type *Ty) const {
7188 // X86 supports extremely general addressing modes.
7190 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7191 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7195 // We can only fold this if we don't need an extra load.
7196 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
7198 // If BaseGV requires a register, we cannot also have a BaseReg.
7199 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
7203 // X86-64 only supports addr of globals in small code model.
7204 if (Subtarget->is64Bit()) {
7205 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7207 // If lower 4G is not available, then we must use rip-relative addressing.
7208 if (AM.BaseOffs || AM.Scale > 1)
7219 // These scales always work.
7224 // These scales are formed with basereg+scalereg. Only accept if there is
7229 default: // Other stuff never works.
7237 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7238 if (!Ty1->isInteger() || !Ty2->isInteger())
7240 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7241 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7242 if (NumBits1 <= NumBits2)
7244 return Subtarget->is64Bit() || NumBits1 < 64;
7247 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7248 if (!VT1.isInteger() || !VT2.isInteger())
7250 unsigned NumBits1 = VT1.getSizeInBits();
7251 unsigned NumBits2 = VT2.getSizeInBits();
7252 if (NumBits1 <= NumBits2)
7254 return Subtarget->is64Bit() || NumBits1 < 64;
7257 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7258 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7259 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7260 /// are assumed to be legal.
7262 X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
7263 // Only do shuffles on 128-bit vector types for now.
7264 // FIXME: pshufb, blends
7265 if (VT.getSizeInBits() == 64) return false;
7266 return (Mask.getNode()->getNumOperands() <= 4 ||
7267 isIdentityMask(Mask.getNode()) ||
7268 isIdentityMask(Mask.getNode(), true) ||
7269 isSplatMask(Mask.getNode()) ||
7270 X86::isPSHUFHWMask(Mask.getNode()) ||
7271 X86::isPSHUFLWMask(Mask.getNode()) ||
7272 X86::isUNPCKLMask(Mask.getNode()) ||
7273 X86::isUNPCKHMask(Mask.getNode()) ||
7274 X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
7275 X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
7279 X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
7280 MVT EVT, SelectionDAG &DAG) const {
7281 unsigned NumElts = BVOps.size();
7282 // Only do shuffles on 128-bit vector types for now.
7283 if (EVT.getSizeInBits() * NumElts == 64) return false;
7284 if (NumElts == 2) return true;
7286 return (isMOVLMask(&BVOps[0], 4) ||
7287 isCommutedMOVL(&BVOps[0], 4, true) ||
7288 isSHUFPMask(&BVOps[0], 4) ||
7289 isCommutedSHUFP(&BVOps[0], 4));
7294 //===----------------------------------------------------------------------===//
7295 // X86 Scheduler Hooks
7296 //===----------------------------------------------------------------------===//
7298 // private utility function
7300 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7301 MachineBasicBlock *MBB,
7309 TargetRegisterClass *RC,
7310 bool invSrc) const {
7311 // For the atomic bitwise operator, we generate
7314 // ld t1 = [bitinstr.addr]
7315 // op t2 = t1, [bitinstr.val]
7317 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7319 // fallthrough -->nextMBB
7320 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7321 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7322 MachineFunction::iterator MBBIter = MBB;
7325 /// First build the CFG
7326 MachineFunction *F = MBB->getParent();
7327 MachineBasicBlock *thisMBB = MBB;
7328 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7329 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7330 F->insert(MBBIter, newMBB);
7331 F->insert(MBBIter, nextMBB);
7333 // Move all successors to thisMBB to nextMBB
7334 nextMBB->transferSuccessors(thisMBB);
7336 // Update thisMBB to fall through to newMBB
7337 thisMBB->addSuccessor(newMBB);
7339 // newMBB jumps to itself and fall through to nextMBB
7340 newMBB->addSuccessor(nextMBB);
7341 newMBB->addSuccessor(newMBB);
7343 // Insert instructions into newMBB based on incoming instruction
7344 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7345 "unexpected number of operands");
7346 DebugLoc dl = bInstr->getDebugLoc();
7347 MachineOperand& destOper = bInstr->getOperand(0);
7348 MachineOperand* argOpers[2 + X86AddrNumOperands];
7349 int numArgs = bInstr->getNumOperands() - 1;
7350 for (int i=0; i < numArgs; ++i)
7351 argOpers[i] = &bInstr->getOperand(i+1);
7353 // x86 address has 4 operands: base, index, scale, and displacement
7354 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7355 int valArgIndx = lastAddrIndx + 1;
7357 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7358 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7359 for (int i=0; i <= lastAddrIndx; ++i)
7360 (*MIB).addOperand(*argOpers[i]);
7362 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7364 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7369 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7370 assert((argOpers[valArgIndx]->isReg() ||
7371 argOpers[valArgIndx]->isImm()) &&
7373 if (argOpers[valArgIndx]->isReg())
7374 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7376 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7378 (*MIB).addOperand(*argOpers[valArgIndx]);
7380 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7383 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7384 for (int i=0; i <= lastAddrIndx; ++i)
7385 (*MIB).addOperand(*argOpers[i]);
7387 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7388 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7390 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7394 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7396 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7400 // private utility function: 64 bit atomics on 32 bit host.
7402 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7403 MachineBasicBlock *MBB,
7408 bool invSrc) const {
7409 // For the atomic bitwise operator, we generate
7410 // thisMBB (instructions are in pairs, except cmpxchg8b)
7411 // ld t1,t2 = [bitinstr.addr]
7413 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7414 // op t5, t6 <- out1, out2, [bitinstr.val]
7415 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7416 // mov ECX, EBX <- t5, t6
7417 // mov EAX, EDX <- t1, t2
7418 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7419 // mov t3, t4 <- EAX, EDX
7421 // result in out1, out2
7422 // fallthrough -->nextMBB
7424 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7425 const unsigned LoadOpc = X86::MOV32rm;
7426 const unsigned copyOpc = X86::MOV32rr;
7427 const unsigned NotOpc = X86::NOT32r;
7428 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7429 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7430 MachineFunction::iterator MBBIter = MBB;
7433 /// First build the CFG
7434 MachineFunction *F = MBB->getParent();
7435 MachineBasicBlock *thisMBB = MBB;
7436 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7437 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7438 F->insert(MBBIter, newMBB);
7439 F->insert(MBBIter, nextMBB);
7441 // Move all successors to thisMBB to nextMBB
7442 nextMBB->transferSuccessors(thisMBB);
7444 // Update thisMBB to fall through to newMBB
7445 thisMBB->addSuccessor(newMBB);
7447 // newMBB jumps to itself and fall through to nextMBB
7448 newMBB->addSuccessor(nextMBB);
7449 newMBB->addSuccessor(newMBB);
7451 DebugLoc dl = bInstr->getDebugLoc();
7452 // Insert instructions into newMBB based on incoming instruction
7453 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7454 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7455 "unexpected number of operands");
7456 MachineOperand& dest1Oper = bInstr->getOperand(0);
7457 MachineOperand& dest2Oper = bInstr->getOperand(1);
7458 MachineOperand* argOpers[2 + X86AddrNumOperands];
7459 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7460 argOpers[i] = &bInstr->getOperand(i+2);
7462 // x86 address has 4 operands: base, index, scale, and displacement
7463 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7465 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7466 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7467 for (int i=0; i <= lastAddrIndx; ++i)
7468 (*MIB).addOperand(*argOpers[i]);
7469 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7470 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7471 // add 4 to displacement.
7472 for (int i=0; i <= lastAddrIndx-1; ++i)
7473 (*MIB).addOperand(*argOpers[i]);
7474 MachineOperand newOp3 = *(argOpers[3]);
7476 newOp3.setImm(newOp3.getImm()+4);
7478 newOp3.setOffset(newOp3.getOffset()+4);
7479 (*MIB).addOperand(newOp3);
7481 // t3/4 are defined later, at the bottom of the loop
7482 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7483 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7484 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7485 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7486 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7487 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7489 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7490 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7492 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7493 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7499 int valArgIndx = lastAddrIndx + 1;
7500 assert((argOpers[valArgIndx]->isReg() ||
7501 argOpers[valArgIndx]->isImm()) &&
7503 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7504 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7505 if (argOpers[valArgIndx]->isReg())
7506 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7508 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7509 if (regOpcL != X86::MOV32rr)
7511 (*MIB).addOperand(*argOpers[valArgIndx]);
7512 assert(argOpers[valArgIndx + 1]->isReg() ==
7513 argOpers[valArgIndx]->isReg());
7514 assert(argOpers[valArgIndx + 1]->isImm() ==
7515 argOpers[valArgIndx]->isImm());
7516 if (argOpers[valArgIndx + 1]->isReg())
7517 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7519 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7520 if (regOpcH != X86::MOV32rr)
7522 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7524 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7526 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7529 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7531 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7534 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7535 for (int i=0; i <= lastAddrIndx; ++i)
7536 (*MIB).addOperand(*argOpers[i]);
7538 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7539 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7541 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7542 MIB.addReg(X86::EAX);
7543 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7544 MIB.addReg(X86::EDX);
7547 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7549 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7553 // private utility function
7555 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7556 MachineBasicBlock *MBB,
7557 unsigned cmovOpc) const {
7558 // For the atomic min/max operator, we generate
7561 // ld t1 = [min/max.addr]
7562 // mov t2 = [min/max.val]
7564 // cmov[cond] t2 = t1
7566 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7568 // fallthrough -->nextMBB
7570 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7571 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7572 MachineFunction::iterator MBBIter = MBB;
7575 /// First build the CFG
7576 MachineFunction *F = MBB->getParent();
7577 MachineBasicBlock *thisMBB = MBB;
7578 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7579 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7580 F->insert(MBBIter, newMBB);
7581 F->insert(MBBIter, nextMBB);
7583 // Move all successors to thisMBB to nextMBB
7584 nextMBB->transferSuccessors(thisMBB);
7586 // Update thisMBB to fall through to newMBB
7587 thisMBB->addSuccessor(newMBB);
7589 // newMBB jumps to newMBB and fall through to nextMBB
7590 newMBB->addSuccessor(nextMBB);
7591 newMBB->addSuccessor(newMBB);
7593 DebugLoc dl = mInstr->getDebugLoc();
7594 // Insert instructions into newMBB based on incoming instruction
7595 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7596 "unexpected number of operands");
7597 MachineOperand& destOper = mInstr->getOperand(0);
7598 MachineOperand* argOpers[2 + X86AddrNumOperands];
7599 int numArgs = mInstr->getNumOperands() - 1;
7600 for (int i=0; i < numArgs; ++i)
7601 argOpers[i] = &mInstr->getOperand(i+1);
7603 // x86 address has 4 operands: base, index, scale, and displacement
7604 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7605 int valArgIndx = lastAddrIndx + 1;
7607 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7608 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7609 for (int i=0; i <= lastAddrIndx; ++i)
7610 (*MIB).addOperand(*argOpers[i]);
7612 // We only support register and immediate values
7613 assert((argOpers[valArgIndx]->isReg() ||
7614 argOpers[valArgIndx]->isImm()) &&
7617 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7618 if (argOpers[valArgIndx]->isReg())
7619 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7621 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7622 (*MIB).addOperand(*argOpers[valArgIndx]);
7624 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7627 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7632 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7633 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7637 // Cmp and exchange if none has modified the memory location
7638 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7639 for (int i=0; i <= lastAddrIndx; ++i)
7640 (*MIB).addOperand(*argOpers[i]);
7642 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7643 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7645 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7646 MIB.addReg(X86::EAX);
7649 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7651 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7657 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7658 MachineBasicBlock *BB) const {
7659 DebugLoc dl = MI->getDebugLoc();
7660 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7661 switch (MI->getOpcode()) {
7662 default: assert(false && "Unexpected instr type to insert");
7663 case X86::CMOV_V1I64:
7664 case X86::CMOV_FR32:
7665 case X86::CMOV_FR64:
7666 case X86::CMOV_V4F32:
7667 case X86::CMOV_V2F64:
7668 case X86::CMOV_V2I64: {
7669 // To "insert" a SELECT_CC instruction, we actually have to insert the
7670 // diamond control-flow pattern. The incoming instruction knows the
7671 // destination vreg to set, the condition code register to branch on, the
7672 // true/false values to select between, and a branch opcode to use.
7673 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7674 MachineFunction::iterator It = BB;
7680 // cmpTY ccX, r1, r2
7682 // fallthrough --> copy0MBB
7683 MachineBasicBlock *thisMBB = BB;
7684 MachineFunction *F = BB->getParent();
7685 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7686 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7688 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7689 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7690 F->insert(It, copy0MBB);
7691 F->insert(It, sinkMBB);
7692 // Update machine-CFG edges by transferring all successors of the current
7693 // block to the new block which will contain the Phi node for the select.
7694 sinkMBB->transferSuccessors(BB);
7696 // Add the true and fallthrough blocks as its successors.
7697 BB->addSuccessor(copy0MBB);
7698 BB->addSuccessor(sinkMBB);
7701 // %FalseValue = ...
7702 // # fallthrough to sinkMBB
7705 // Update machine-CFG edges
7706 BB->addSuccessor(sinkMBB);
7709 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7712 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7713 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7714 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7716 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7720 case X86::FP32_TO_INT16_IN_MEM:
7721 case X86::FP32_TO_INT32_IN_MEM:
7722 case X86::FP32_TO_INT64_IN_MEM:
7723 case X86::FP64_TO_INT16_IN_MEM:
7724 case X86::FP64_TO_INT32_IN_MEM:
7725 case X86::FP64_TO_INT64_IN_MEM:
7726 case X86::FP80_TO_INT16_IN_MEM:
7727 case X86::FP80_TO_INT32_IN_MEM:
7728 case X86::FP80_TO_INT64_IN_MEM: {
7729 // Change the floating point control register to use "round towards zero"
7730 // mode when truncating to an integer value.
7731 MachineFunction *F = BB->getParent();
7732 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7733 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7735 // Load the old value of the high byte of the control word...
7737 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7738 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7741 // Set the high part to be round to zero...
7742 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7745 // Reload the modified control word now...
7746 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7748 // Restore the memory image of control word to original value
7749 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7752 // Get the X86 opcode to use.
7754 switch (MI->getOpcode()) {
7755 default: assert(0 && "illegal opcode!");
7756 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7757 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7758 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7759 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7760 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7761 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7762 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7763 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7764 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7768 MachineOperand &Op = MI->getOperand(0);
7770 AM.BaseType = X86AddressMode::RegBase;
7771 AM.Base.Reg = Op.getReg();
7773 AM.BaseType = X86AddressMode::FrameIndexBase;
7774 AM.Base.FrameIndex = Op.getIndex();
7776 Op = MI->getOperand(1);
7778 AM.Scale = Op.getImm();
7779 Op = MI->getOperand(2);
7781 AM.IndexReg = Op.getImm();
7782 Op = MI->getOperand(3);
7783 if (Op.isGlobal()) {
7784 AM.GV = Op.getGlobal();
7786 AM.Disp = Op.getImm();
7788 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7789 .addReg(MI->getOperand(4).getReg());
7791 // Reload the original control word now.
7792 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7794 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7797 case X86::ATOMAND32:
7798 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7799 X86::AND32ri, X86::MOV32rm,
7800 X86::LCMPXCHG32, X86::MOV32rr,
7801 X86::NOT32r, X86::EAX,
7802 X86::GR32RegisterClass);
7804 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7805 X86::OR32ri, X86::MOV32rm,
7806 X86::LCMPXCHG32, X86::MOV32rr,
7807 X86::NOT32r, X86::EAX,
7808 X86::GR32RegisterClass);
7809 case X86::ATOMXOR32:
7810 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7811 X86::XOR32ri, X86::MOV32rm,
7812 X86::LCMPXCHG32, X86::MOV32rr,
7813 X86::NOT32r, X86::EAX,
7814 X86::GR32RegisterClass);
7815 case X86::ATOMNAND32:
7816 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7817 X86::AND32ri, X86::MOV32rm,
7818 X86::LCMPXCHG32, X86::MOV32rr,
7819 X86::NOT32r, X86::EAX,
7820 X86::GR32RegisterClass, true);
7821 case X86::ATOMMIN32:
7822 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7823 case X86::ATOMMAX32:
7824 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7825 case X86::ATOMUMIN32:
7826 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7827 case X86::ATOMUMAX32:
7828 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7830 case X86::ATOMAND16:
7831 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7832 X86::AND16ri, X86::MOV16rm,
7833 X86::LCMPXCHG16, X86::MOV16rr,
7834 X86::NOT16r, X86::AX,
7835 X86::GR16RegisterClass);
7837 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7838 X86::OR16ri, X86::MOV16rm,
7839 X86::LCMPXCHG16, X86::MOV16rr,
7840 X86::NOT16r, X86::AX,
7841 X86::GR16RegisterClass);
7842 case X86::ATOMXOR16:
7843 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7844 X86::XOR16ri, X86::MOV16rm,
7845 X86::LCMPXCHG16, X86::MOV16rr,
7846 X86::NOT16r, X86::AX,
7847 X86::GR16RegisterClass);
7848 case X86::ATOMNAND16:
7849 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7850 X86::AND16ri, X86::MOV16rm,
7851 X86::LCMPXCHG16, X86::MOV16rr,
7852 X86::NOT16r, X86::AX,
7853 X86::GR16RegisterClass, true);
7854 case X86::ATOMMIN16:
7855 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7856 case X86::ATOMMAX16:
7857 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7858 case X86::ATOMUMIN16:
7859 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7860 case X86::ATOMUMAX16:
7861 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7864 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7865 X86::AND8ri, X86::MOV8rm,
7866 X86::LCMPXCHG8, X86::MOV8rr,
7867 X86::NOT8r, X86::AL,
7868 X86::GR8RegisterClass);
7870 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7871 X86::OR8ri, X86::MOV8rm,
7872 X86::LCMPXCHG8, X86::MOV8rr,
7873 X86::NOT8r, X86::AL,
7874 X86::GR8RegisterClass);
7876 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7877 X86::XOR8ri, X86::MOV8rm,
7878 X86::LCMPXCHG8, X86::MOV8rr,
7879 X86::NOT8r, X86::AL,
7880 X86::GR8RegisterClass);
7881 case X86::ATOMNAND8:
7882 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7883 X86::AND8ri, X86::MOV8rm,
7884 X86::LCMPXCHG8, X86::MOV8rr,
7885 X86::NOT8r, X86::AL,
7886 X86::GR8RegisterClass, true);
7887 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7888 // This group is for 64-bit host.
7889 case X86::ATOMAND64:
7890 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7891 X86::AND64ri32, X86::MOV64rm,
7892 X86::LCMPXCHG64, X86::MOV64rr,
7893 X86::NOT64r, X86::RAX,
7894 X86::GR64RegisterClass);
7896 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7897 X86::OR64ri32, X86::MOV64rm,
7898 X86::LCMPXCHG64, X86::MOV64rr,
7899 X86::NOT64r, X86::RAX,
7900 X86::GR64RegisterClass);
7901 case X86::ATOMXOR64:
7902 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7903 X86::XOR64ri32, X86::MOV64rm,
7904 X86::LCMPXCHG64, X86::MOV64rr,
7905 X86::NOT64r, X86::RAX,
7906 X86::GR64RegisterClass);
7907 case X86::ATOMNAND64:
7908 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7909 X86::AND64ri32, X86::MOV64rm,
7910 X86::LCMPXCHG64, X86::MOV64rr,
7911 X86::NOT64r, X86::RAX,
7912 X86::GR64RegisterClass, true);
7913 case X86::ATOMMIN64:
7914 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7915 case X86::ATOMMAX64:
7916 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7917 case X86::ATOMUMIN64:
7918 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7919 case X86::ATOMUMAX64:
7920 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7922 // This group does 64-bit operations on a 32-bit host.
7923 case X86::ATOMAND6432:
7924 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7925 X86::AND32rr, X86::AND32rr,
7926 X86::AND32ri, X86::AND32ri,
7928 case X86::ATOMOR6432:
7929 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7930 X86::OR32rr, X86::OR32rr,
7931 X86::OR32ri, X86::OR32ri,
7933 case X86::ATOMXOR6432:
7934 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7935 X86::XOR32rr, X86::XOR32rr,
7936 X86::XOR32ri, X86::XOR32ri,
7938 case X86::ATOMNAND6432:
7939 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7940 X86::AND32rr, X86::AND32rr,
7941 X86::AND32ri, X86::AND32ri,
7943 case X86::ATOMADD6432:
7944 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7945 X86::ADD32rr, X86::ADC32rr,
7946 X86::ADD32ri, X86::ADC32ri,
7948 case X86::ATOMSUB6432:
7949 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7950 X86::SUB32rr, X86::SBB32rr,
7951 X86::SUB32ri, X86::SBB32ri,
7953 case X86::ATOMSWAP6432:
7954 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7955 X86::MOV32rr, X86::MOV32rr,
7956 X86::MOV32ri, X86::MOV32ri,
7961 //===----------------------------------------------------------------------===//
7962 // X86 Optimization Hooks
7963 //===----------------------------------------------------------------------===//
7965 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7969 const SelectionDAG &DAG,
7970 unsigned Depth) const {
7971 unsigned Opc = Op.getOpcode();
7972 assert((Opc >= ISD::BUILTIN_OP_END ||
7973 Opc == ISD::INTRINSIC_WO_CHAIN ||
7974 Opc == ISD::INTRINSIC_W_CHAIN ||
7975 Opc == ISD::INTRINSIC_VOID) &&
7976 "Should use MaskedValueIsZero if you don't know whether Op"
7977 " is a target node!");
7979 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7988 // These nodes' second result is a boolean.
7989 if (Op.getResNo() == 0)
7993 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7994 Mask.getBitWidth() - 1);
7999 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8000 /// node is a GlobalAddress + offset.
8001 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8002 GlobalValue* &GA, int64_t &Offset) const{
8003 if (N->getOpcode() == X86ISD::Wrapper) {
8004 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8005 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8006 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8010 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8013 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8014 const TargetLowering &TLI) {
8017 if (TLI.isGAPlusOffset(Base, GV, Offset))
8018 return (GV->getAlignment() >= N && (Offset % N) == 0);
8019 // DAG combine handles the stack object case.
8023 static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
8024 unsigned NumElems, MVT EVT,
8026 SelectionDAG &DAG, MachineFrameInfo *MFI,
8027 const TargetLowering &TLI) {
8029 for (unsigned i = 0; i < NumElems; ++i) {
8030 SDValue Idx = PermMask.getOperand(i);
8031 if (Idx.getOpcode() == ISD::UNDEF) {
8037 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8038 if (!Elt.getNode() ||
8039 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8042 Base = Elt.getNode();
8043 if (Base->getOpcode() == ISD::UNDEF)
8047 if (Elt.getOpcode() == ISD::UNDEF)
8050 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
8051 EVT.getSizeInBits()/8, i, MFI))
8057 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8058 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8059 /// if the load addresses are consecutive, non-overlapping, and in the right
8061 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8062 const TargetLowering &TLI) {
8063 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8064 DebugLoc dl = N->getDebugLoc();
8065 MVT VT = N->getValueType(0);
8066 MVT EVT = VT.getVectorElementType();
8067 SDValue PermMask = N->getOperand(2);
8068 unsigned NumElems = PermMask.getNumOperands();
8069 SDNode *Base = NULL;
8070 if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
8074 LoadSDNode *LD = cast<LoadSDNode>(Base);
8075 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
8076 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8077 LD->getSrcValue(), LD->getSrcValueOffset(),
8079 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8080 LD->getSrcValue(), LD->getSrcValueOffset(),
8081 LD->isVolatile(), LD->getAlignment());
8084 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
8085 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
8086 TargetLowering::DAGCombinerInfo &DCI,
8087 const X86Subtarget *Subtarget,
8088 const TargetLowering &TLI) {
8089 unsigned NumOps = N->getNumOperands();
8090 DebugLoc dl = N->getDebugLoc();
8092 // Ignore single operand BUILD_VECTOR.
8096 MVT VT = N->getValueType(0);
8097 MVT EVT = VT.getVectorElementType();
8098 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
8099 // We are looking for load i64 and zero extend. We want to transform
8100 // it before legalizer has a chance to expand it. Also look for i64
8101 // BUILD_PAIR bit casted to f64.
8103 // This must be an insertion into a zero vector.
8104 SDValue HighElt = N->getOperand(1);
8105 if (!isZeroNode(HighElt))
8108 // Value must be a load.
8109 SDNode *Base = N->getOperand(0).getNode();
8110 if (!isa<LoadSDNode>(Base)) {
8111 if (Base->getOpcode() != ISD::BIT_CONVERT)
8113 Base = Base->getOperand(0).getNode();
8114 if (!isa<LoadSDNode>(Base))
8118 // Transform it into VZEXT_LOAD addr.
8119 LoadSDNode *LD = cast<LoadSDNode>(Base);
8121 // Load must not be an extload.
8122 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
8125 // Load type should legal type so we don't have to legalize it.
8126 if (!TLI.isTypeLegal(VT))
8129 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
8130 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8131 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8132 TargetLowering::TargetLoweringOpt TLO(DAG);
8133 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
8134 DCI.CommitTargetLoweringOpt(TLO);
8138 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8139 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8140 const X86Subtarget *Subtarget) {
8141 DebugLoc DL = N->getDebugLoc();
8142 SDValue Cond = N->getOperand(0);
8143 // Get the LHS/RHS of the select.
8144 SDValue LHS = N->getOperand(1);
8145 SDValue RHS = N->getOperand(2);
8147 // If we have SSE[12] support, try to form min/max nodes.
8148 if (Subtarget->hasSSE2() &&
8149 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8150 Cond.getOpcode() == ISD::SETCC) {
8151 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8153 unsigned Opcode = 0;
8154 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8157 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8160 if (!UnsafeFPMath) break;
8162 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8164 Opcode = X86ISD::FMIN;
8167 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8170 if (!UnsafeFPMath) break;
8172 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8174 Opcode = X86ISD::FMAX;
8177 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8180 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8183 if (!UnsafeFPMath) break;
8185 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8187 Opcode = X86ISD::FMIN;
8190 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8193 if (!UnsafeFPMath) break;
8195 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8197 Opcode = X86ISD::FMAX;
8203 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8206 // If this is a select between two integer constants, try to do some
8208 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8209 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8210 // Don't do this for crazy integer types.
8211 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8212 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8213 // so that TrueC (the true value) is larger than FalseC.
8214 bool NeedsCondInvert = false;
8216 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8217 // Efficiently invertible.
8218 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8219 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8220 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8221 NeedsCondInvert = true;
8222 std::swap(TrueC, FalseC);
8225 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8226 if (FalseC->getAPIntValue() == 0 &&
8227 TrueC->getAPIntValue().isPowerOf2()) {
8228 if (NeedsCondInvert) // Invert the condition if needed.
8229 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8230 DAG.getConstant(1, Cond.getValueType()));
8232 // Zero extend the condition if needed.
8233 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8235 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8236 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8237 DAG.getConstant(ShAmt, MVT::i8));
8240 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8241 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8242 if (NeedsCondInvert) // Invert the condition if needed.
8243 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8244 DAG.getConstant(1, Cond.getValueType()));
8246 // Zero extend the condition if needed.
8247 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8248 FalseC->getValueType(0), Cond);
8249 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8250 SDValue(FalseC, 0));
8253 // Optimize cases that will turn into an LEA instruction. This requires
8254 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8255 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8256 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8257 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8259 bool isFastMultiplier = false;
8261 switch ((unsigned char)Diff) {
8263 case 1: // result = add base, cond
8264 case 2: // result = lea base( , cond*2)
8265 case 3: // result = lea base(cond, cond*2)
8266 case 4: // result = lea base( , cond*4)
8267 case 5: // result = lea base(cond, cond*4)
8268 case 8: // result = lea base( , cond*8)
8269 case 9: // result = lea base(cond, cond*8)
8270 isFastMultiplier = true;
8275 if (isFastMultiplier) {
8276 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8277 if (NeedsCondInvert) // Invert the condition if needed.
8278 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8279 DAG.getConstant(1, Cond.getValueType()));
8281 // Zero extend the condition if needed.
8282 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8284 // Scale the condition by the difference.
8286 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8287 DAG.getConstant(Diff, Cond.getValueType()));
8289 // Add the base if non-zero.
8290 if (FalseC->getAPIntValue() != 0)
8291 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8292 SDValue(FalseC, 0));
8302 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8303 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8304 TargetLowering::DAGCombinerInfo &DCI) {
8305 DebugLoc DL = N->getDebugLoc();
8307 // If the flag operand isn't dead, don't touch this CMOV.
8308 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8311 // If this is a select between two integer constants, try to do some
8312 // optimizations. Note that the operands are ordered the opposite of SELECT
8314 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8315 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8316 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8317 // larger than FalseC (the false value).
8318 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8320 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8321 CC = X86::GetOppositeBranchCondition(CC);
8322 std::swap(TrueC, FalseC);
8325 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8326 // This is efficient for any integer data type (including i8/i16) and
8328 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8329 SDValue Cond = N->getOperand(3);
8330 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8331 DAG.getConstant(CC, MVT::i8), Cond);
8333 // Zero extend the condition if needed.
8334 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8336 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8337 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8338 DAG.getConstant(ShAmt, MVT::i8));
8339 if (N->getNumValues() == 2) // Dead flag value?
8340 return DCI.CombineTo(N, Cond, SDValue());
8344 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8345 // for any integer data type, including i8/i16.
8346 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8347 SDValue Cond = N->getOperand(3);
8348 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8349 DAG.getConstant(CC, MVT::i8), Cond);
8351 // Zero extend the condition if needed.
8352 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8353 FalseC->getValueType(0), Cond);
8354 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8355 SDValue(FalseC, 0));
8357 if (N->getNumValues() == 2) // Dead flag value?
8358 return DCI.CombineTo(N, Cond, SDValue());
8362 // Optimize cases that will turn into an LEA instruction. This requires
8363 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8364 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8365 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8366 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8368 bool isFastMultiplier = false;
8370 switch ((unsigned char)Diff) {
8372 case 1: // result = add base, cond
8373 case 2: // result = lea base( , cond*2)
8374 case 3: // result = lea base(cond, cond*2)
8375 case 4: // result = lea base( , cond*4)
8376 case 5: // result = lea base(cond, cond*4)
8377 case 8: // result = lea base( , cond*8)
8378 case 9: // result = lea base(cond, cond*8)
8379 isFastMultiplier = true;
8384 if (isFastMultiplier) {
8385 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8386 SDValue Cond = N->getOperand(3);
8387 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8388 DAG.getConstant(CC, MVT::i8), Cond);
8389 // Zero extend the condition if needed.
8390 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8392 // Scale the condition by the difference.
8394 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8395 DAG.getConstant(Diff, Cond.getValueType()));
8397 // Add the base if non-zero.
8398 if (FalseC->getAPIntValue() != 0)
8399 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8400 SDValue(FalseC, 0));
8401 if (N->getNumValues() == 2) // Dead flag value?
8402 return DCI.CombineTo(N, Cond, SDValue());
8412 /// PerformMulCombine - Optimize a single multiply with constant into two
8413 /// in order to implement it with two cheaper instructions, e.g.
8414 /// LEA + SHL, LEA + LEA.
8415 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8416 TargetLowering::DAGCombinerInfo &DCI) {
8417 if (DAG.getMachineFunction().
8418 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8421 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8424 MVT VT = N->getValueType(0);
8428 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8431 uint64_t MulAmt = C->getZExtValue();
8432 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8435 uint64_t MulAmt1 = 0;
8436 uint64_t MulAmt2 = 0;
8437 if ((MulAmt % 9) == 0) {
8439 MulAmt2 = MulAmt / 9;
8440 } else if ((MulAmt % 5) == 0) {
8442 MulAmt2 = MulAmt / 5;
8443 } else if ((MulAmt % 3) == 0) {
8445 MulAmt2 = MulAmt / 3;
8448 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8449 DebugLoc DL = N->getDebugLoc();
8451 if (isPowerOf2_64(MulAmt2) &&
8452 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8453 // If second multiplifer is pow2, issue it first. We want the multiply by
8454 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8456 std::swap(MulAmt1, MulAmt2);
8459 if (isPowerOf2_64(MulAmt1))
8460 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8461 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8463 NewMul = DAG.getNode(ISD::MUL, DL, VT, N->getOperand(0),
8464 DAG.getConstant(MulAmt1, VT));
8466 if (isPowerOf2_64(MulAmt2))
8467 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8468 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8470 NewMul = DAG.getNode(ISD::MUL, DL, VT, NewMul,
8471 DAG.getConstant(MulAmt2, VT));
8473 // Do not add new nodes to DAG combiner worklist.
8474 DCI.CombineTo(N, NewMul, false);
8480 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8482 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8483 const X86Subtarget *Subtarget) {
8484 // On X86 with SSE2 support, we can transform this to a vector shift if
8485 // all elements are shifted by the same amount. We can't do this in legalize
8486 // because the a constant vector is typically transformed to a constant pool
8487 // so we have no knowledge of the shift amount.
8488 if (!Subtarget->hasSSE2())
8491 MVT VT = N->getValueType(0);
8492 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8495 SDValue ShAmtOp = N->getOperand(1);
8496 MVT EltVT = VT.getVectorElementType();
8497 DebugLoc DL = N->getDebugLoc();
8499 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8500 unsigned NumElts = VT.getVectorNumElements();
8502 for (; i != NumElts; ++i) {
8503 SDValue Arg = ShAmtOp.getOperand(i);
8504 if (Arg.getOpcode() == ISD::UNDEF) continue;
8508 for (; i != NumElts; ++i) {
8509 SDValue Arg = ShAmtOp.getOperand(i);
8510 if (Arg.getOpcode() == ISD::UNDEF) continue;
8511 if (Arg != BaseShAmt) {
8515 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8516 isSplatMask(ShAmtOp.getOperand(2).getNode())) {
8517 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8518 DAG.getIntPtrConstant(0));
8522 if (EltVT.bitsGT(MVT::i32))
8523 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8524 else if (EltVT.bitsLT(MVT::i32))
8525 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8527 // The shift amount is identical so we can do a vector shift.
8528 SDValue ValOp = N->getOperand(0);
8529 switch (N->getOpcode()) {
8531 assert(0 && "Unknown shift opcode!");
8534 if (VT == MVT::v2i64)
8535 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8536 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8538 if (VT == MVT::v4i32)
8539 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8540 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8542 if (VT == MVT::v8i16)
8543 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8544 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8548 if (VT == MVT::v4i32)
8549 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8550 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8552 if (VT == MVT::v8i16)
8553 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8554 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8558 if (VT == MVT::v2i64)
8559 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8560 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8562 if (VT == MVT::v4i32)
8563 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8564 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8566 if (VT == MVT::v8i16)
8567 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8568 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8575 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8576 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8577 const X86Subtarget *Subtarget) {
8578 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8579 // the FP state in cases where an emms may be missing.
8580 // A preferable solution to the general problem is to figure out the right
8581 // places to insert EMMS. This qualifies as a quick hack.
8583 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8584 StoreSDNode *St = cast<StoreSDNode>(N);
8585 MVT VT = St->getValue().getValueType();
8586 if (VT.getSizeInBits() != 64)
8589 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2();
8590 if ((VT.isVector() ||
8591 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8592 isa<LoadSDNode>(St->getValue()) &&
8593 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8594 St->getChain().hasOneUse() && !St->isVolatile()) {
8595 SDNode* LdVal = St->getValue().getNode();
8597 int TokenFactorIndex = -1;
8598 SmallVector<SDValue, 8> Ops;
8599 SDNode* ChainVal = St->getChain().getNode();
8600 // Must be a store of a load. We currently handle two cases: the load
8601 // is a direct child, and it's under an intervening TokenFactor. It is
8602 // possible to dig deeper under nested TokenFactors.
8603 if (ChainVal == LdVal)
8604 Ld = cast<LoadSDNode>(St->getChain());
8605 else if (St->getValue().hasOneUse() &&
8606 ChainVal->getOpcode() == ISD::TokenFactor) {
8607 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8608 if (ChainVal->getOperand(i).getNode() == LdVal) {
8609 TokenFactorIndex = i;
8610 Ld = cast<LoadSDNode>(St->getValue());
8612 Ops.push_back(ChainVal->getOperand(i));
8616 if (!Ld || !ISD::isNormalLoad(Ld))
8619 // If this is not the MMX case, i.e. we are just turning i64 load/store
8620 // into f64 load/store, avoid the transformation if there are multiple
8621 // uses of the loaded value.
8622 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8625 DebugLoc LdDL = Ld->getDebugLoc();
8626 DebugLoc StDL = N->getDebugLoc();
8627 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8628 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8630 if (Subtarget->is64Bit() || F64IsLegal) {
8631 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8632 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8633 Ld->getBasePtr(), Ld->getSrcValue(),
8634 Ld->getSrcValueOffset(), Ld->isVolatile(),
8635 Ld->getAlignment());
8636 SDValue NewChain = NewLd.getValue(1);
8637 if (TokenFactorIndex != -1) {
8638 Ops.push_back(NewChain);
8639 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8642 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8643 St->getSrcValue(), St->getSrcValueOffset(),
8644 St->isVolatile(), St->getAlignment());
8647 // Otherwise, lower to two pairs of 32-bit loads / stores.
8648 SDValue LoAddr = Ld->getBasePtr();
8649 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8650 DAG.getConstant(4, MVT::i32));
8652 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8653 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8654 Ld->isVolatile(), Ld->getAlignment());
8655 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8656 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8658 MinAlign(Ld->getAlignment(), 4));
8660 SDValue NewChain = LoLd.getValue(1);
8661 if (TokenFactorIndex != -1) {
8662 Ops.push_back(LoLd);
8663 Ops.push_back(HiLd);
8664 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8668 LoAddr = St->getBasePtr();
8669 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8670 DAG.getConstant(4, MVT::i32));
8672 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8673 St->getSrcValue(), St->getSrcValueOffset(),
8674 St->isVolatile(), St->getAlignment());
8675 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8677 St->getSrcValueOffset() + 4,
8679 MinAlign(St->getAlignment(), 4));
8680 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8685 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8686 /// X86ISD::FXOR nodes.
8687 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8688 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8689 // F[X]OR(0.0, x) -> x
8690 // F[X]OR(x, 0.0) -> x
8691 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8692 if (C->getValueAPF().isPosZero())
8693 return N->getOperand(1);
8694 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8695 if (C->getValueAPF().isPosZero())
8696 return N->getOperand(0);
8700 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8701 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8702 // FAND(0.0, x) -> 0.0
8703 // FAND(x, 0.0) -> 0.0
8704 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8705 if (C->getValueAPF().isPosZero())
8706 return N->getOperand(0);
8707 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8708 if (C->getValueAPF().isPosZero())
8709 return N->getOperand(1);
8713 static SDValue PerformBTCombine(SDNode *N,
8715 TargetLowering::DAGCombinerInfo &DCI) {
8716 // BT ignores high bits in the bit index operand.
8717 SDValue Op1 = N->getOperand(1);
8718 if (Op1.hasOneUse()) {
8719 unsigned BitWidth = Op1.getValueSizeInBits();
8720 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8721 APInt KnownZero, KnownOne;
8722 TargetLowering::TargetLoweringOpt TLO(DAG);
8723 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8724 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8725 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8726 DCI.CommitTargetLoweringOpt(TLO);
8731 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8732 DAGCombinerInfo &DCI) const {
8733 SelectionDAG &DAG = DCI.DAG;
8734 switch (N->getOpcode()) {
8736 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8737 case ISD::BUILD_VECTOR:
8738 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8739 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8740 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8741 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8744 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8745 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8747 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8748 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8749 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8755 //===----------------------------------------------------------------------===//
8756 // X86 Inline Assembly Support
8757 //===----------------------------------------------------------------------===//
8759 /// getConstraintType - Given a constraint letter, return the type of
8760 /// constraint it is for this target.
8761 X86TargetLowering::ConstraintType
8762 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8763 if (Constraint.size() == 1) {
8764 switch (Constraint[0]) {
8776 return C_RegisterClass;
8784 return TargetLowering::getConstraintType(Constraint);
8787 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8788 /// with another that has more specific requirements based on the type of the
8789 /// corresponding operand.
8790 const char *X86TargetLowering::
8791 LowerXConstraint(MVT ConstraintVT) const {
8792 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8793 // 'f' like normal targets.
8794 if (ConstraintVT.isFloatingPoint()) {
8795 if (Subtarget->hasSSE2())
8797 if (Subtarget->hasSSE1())
8801 return TargetLowering::LowerXConstraint(ConstraintVT);
8804 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8805 /// vector. If it is invalid, don't add anything to Ops.
8806 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8809 std::vector<SDValue>&Ops,
8810 SelectionDAG &DAG) const {
8811 SDValue Result(0, 0);
8813 switch (Constraint) {
8816 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8817 if (C->getZExtValue() <= 31) {
8818 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8825 if (C->getZExtValue() <= 63) {
8826 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8832 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8833 if (C->getZExtValue() <= 255) {
8834 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8840 // 32-bit signed value
8841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8842 const ConstantInt *CI = C->getConstantIntValue();
8843 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8844 // Widen to 64 bits here to get it sign extended.
8845 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8848 // FIXME gcc accepts some relocatable values here too, but only in certain
8849 // memory models; it's complicated.
8854 // 32-bit unsigned value
8855 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8856 const ConstantInt *CI = C->getConstantIntValue();
8857 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8858 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8862 // FIXME gcc accepts some relocatable values here too, but only in certain
8863 // memory models; it's complicated.
8867 // Literal immediates are always ok.
8868 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8869 // Widen to 64 bits here to get it sign extended.
8870 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8874 // If we are in non-pic codegen mode, we allow the address of a global (with
8875 // an optional displacement) to be used with 'i'.
8876 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8879 // Match either (GA) or (GA+C)
8881 Offset = GA->getOffset();
8882 } else if (Op.getOpcode() == ISD::ADD) {
8883 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8884 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8886 Offset = GA->getOffset()+C->getZExtValue();
8888 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8889 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8891 Offset = GA->getOffset()+C->getZExtValue();
8899 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8902 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8908 // Otherwise, not valid for this mode.
8913 if (Result.getNode()) {
8914 Ops.push_back(Result);
8917 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8921 std::vector<unsigned> X86TargetLowering::
8922 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8924 if (Constraint.size() == 1) {
8925 // FIXME: not handling fp-stack yet!
8926 switch (Constraint[0]) { // GCC X86 Constraint Letters
8927 default: break; // Unknown constraint letter
8928 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8931 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8932 else if (VT == MVT::i16)
8933 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8934 else if (VT == MVT::i8)
8935 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8936 else if (VT == MVT::i64)
8937 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8942 return std::vector<unsigned>();
8945 std::pair<unsigned, const TargetRegisterClass*>
8946 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8948 // First, see if this is a constraint that directly corresponds to an LLVM
8950 if (Constraint.size() == 1) {
8951 // GCC Constraint Letters
8952 switch (Constraint[0]) {
8954 case 'r': // GENERAL_REGS
8955 case 'R': // LEGACY_REGS
8956 case 'l': // INDEX_REGS
8958 return std::make_pair(0U, X86::GR8RegisterClass);
8960 return std::make_pair(0U, X86::GR16RegisterClass);
8961 if (VT == MVT::i32 || !Subtarget->is64Bit())
8962 return std::make_pair(0U, X86::GR32RegisterClass);
8963 return std::make_pair(0U, X86::GR64RegisterClass);
8964 case 'f': // FP Stack registers.
8965 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8966 // value to the correct fpstack register class.
8967 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8968 return std::make_pair(0U, X86::RFP32RegisterClass);
8969 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8970 return std::make_pair(0U, X86::RFP64RegisterClass);
8971 return std::make_pair(0U, X86::RFP80RegisterClass);
8972 case 'y': // MMX_REGS if MMX allowed.
8973 if (!Subtarget->hasMMX()) break;
8974 return std::make_pair(0U, X86::VR64RegisterClass);
8975 case 'Y': // SSE_REGS if SSE2 allowed
8976 if (!Subtarget->hasSSE2()) break;
8978 case 'x': // SSE_REGS if SSE1 allowed
8979 if (!Subtarget->hasSSE1()) break;
8981 switch (VT.getSimpleVT()) {
8983 // Scalar SSE types.
8986 return std::make_pair(0U, X86::FR32RegisterClass);
8989 return std::make_pair(0U, X86::FR64RegisterClass);
8997 return std::make_pair(0U, X86::VR128RegisterClass);
9003 // Use the default implementation in TargetLowering to convert the register
9004 // constraint into a member of a register class.
9005 std::pair<unsigned, const TargetRegisterClass*> Res;
9006 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9008 // Not found as a standard register?
9009 if (Res.second == 0) {
9010 // GCC calls "st(0)" just plain "st".
9011 if (StringsEqualNoCase("{st}", Constraint)) {
9012 Res.first = X86::ST0;
9013 Res.second = X86::RFP80RegisterClass;
9015 // 'A' means EAX + EDX.
9016 if (Constraint == "A") {
9017 Res.first = X86::EAX;
9018 Res.second = X86::GRADRegisterClass;
9023 // Otherwise, check to see if this is a register class of the wrong value
9024 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9025 // turn into {ax},{dx}.
9026 if (Res.second->hasType(VT))
9027 return Res; // Correct type already, nothing to do.
9029 // All of the single-register GCC register classes map their values onto
9030 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9031 // really want an 8-bit or 32-bit register, map to the appropriate register
9032 // class and return the appropriate register.
9033 if (Res.second == X86::GR16RegisterClass) {
9034 if (VT == MVT::i8) {
9035 unsigned DestReg = 0;
9036 switch (Res.first) {
9038 case X86::AX: DestReg = X86::AL; break;
9039 case X86::DX: DestReg = X86::DL; break;
9040 case X86::CX: DestReg = X86::CL; break;
9041 case X86::BX: DestReg = X86::BL; break;
9044 Res.first = DestReg;
9045 Res.second = Res.second = X86::GR8RegisterClass;
9047 } else if (VT == MVT::i32) {
9048 unsigned DestReg = 0;
9049 switch (Res.first) {
9051 case X86::AX: DestReg = X86::EAX; break;
9052 case X86::DX: DestReg = X86::EDX; break;
9053 case X86::CX: DestReg = X86::ECX; break;
9054 case X86::BX: DestReg = X86::EBX; break;
9055 case X86::SI: DestReg = X86::ESI; break;
9056 case X86::DI: DestReg = X86::EDI; break;
9057 case X86::BP: DestReg = X86::EBP; break;
9058 case X86::SP: DestReg = X86::ESP; break;
9061 Res.first = DestReg;
9062 Res.second = Res.second = X86::GR32RegisterClass;
9064 } else if (VT == MVT::i64) {
9065 unsigned DestReg = 0;
9066 switch (Res.first) {
9068 case X86::AX: DestReg = X86::RAX; break;
9069 case X86::DX: DestReg = X86::RDX; break;
9070 case X86::CX: DestReg = X86::RCX; break;
9071 case X86::BX: DestReg = X86::RBX; break;
9072 case X86::SI: DestReg = X86::RSI; break;
9073 case X86::DI: DestReg = X86::RDI; break;
9074 case X86::BP: DestReg = X86::RBP; break;
9075 case X86::SP: DestReg = X86::RSP; break;
9078 Res.first = DestReg;
9079 Res.second = Res.second = X86::GR64RegisterClass;
9082 } else if (Res.second == X86::FR32RegisterClass ||
9083 Res.second == X86::FR64RegisterClass ||
9084 Res.second == X86::VR128RegisterClass) {
9085 // Handle references to XMM physical registers that got mapped into the
9086 // wrong class. This can happen with constraints like {xmm0} where the
9087 // target independent register mapper will just pick the first match it can
9088 // find, ignoring the required type.
9090 Res.second = X86::FR32RegisterClass;
9091 else if (VT == MVT::f64)
9092 Res.second = X86::FR64RegisterClass;
9093 else if (X86::VR128RegisterClass->hasType(VT))
9094 Res.second = X86::VR128RegisterClass;
9100 //===----------------------------------------------------------------------===//
9101 // X86 Widen vector type
9102 //===----------------------------------------------------------------------===//
9104 /// getWidenVectorType: given a vector type, returns the type to widen
9105 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9106 /// If there is no vector type that we want to widen to, returns MVT::Other
9107 /// When and where to widen is target dependent based on the cost of
9108 /// scalarizing vs using the wider vector type.
9110 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9111 assert(VT.isVector());
9112 if (isTypeLegal(VT))
9115 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9116 // type based on element type. This would speed up our search (though
9117 // it may not be worth it since the size of the list is relatively
9119 MVT EltVT = VT.getVectorElementType();
9120 unsigned NElts = VT.getVectorNumElements();
9122 // On X86, it make sense to widen any vector wider than 1
9126 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9127 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9128 MVT SVT = (MVT::SimpleValueType)nVT;
9130 if (isTypeLegal(SVT) &&
9131 SVT.getVectorElementType() == EltVT &&
9132 SVT.getVectorNumElements() > NElts)