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 "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalAlias.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/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/PseudoSourceValue.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/StringExtras.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/raw_ostream.h"
45 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
47 // Forward declarations.
48 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
51 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
52 : TargetLowering(TM) {
53 Subtarget = &TM.getSubtarget<X86Subtarget>();
54 X86ScalarSSEf64 = Subtarget->hasSSE2();
55 X86ScalarSSEf32 = Subtarget->hasSSE1();
56 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
58 RegInfo = TM.getRegisterInfo();
61 // Set up the TargetLowering object.
63 // X86 is weird, it always uses i8 for shift amounts and setcc results.
64 setShiftAmountType(MVT::i8);
65 setBooleanContents(ZeroOrOneBooleanContent);
66 setSchedulingPreference(SchedulingForRegPressure);
67 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
68 setStackPointerRegisterToSaveRestore(X86StackPtr);
70 if (Subtarget->isTargetDarwin()) {
71 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
72 setUseUnderscoreSetJmp(false);
73 setUseUnderscoreLongJmp(false);
74 } else if (Subtarget->isTargetMingw()) {
75 // MS runtime is weird: it exports _setjmp, but longjmp!
76 setUseUnderscoreSetJmp(true);
77 setUseUnderscoreLongJmp(false);
79 setUseUnderscoreSetJmp(true);
80 setUseUnderscoreLongJmp(true);
83 // Set up the register classes.
84 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
85 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
86 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
87 if (Subtarget->is64Bit())
88 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
90 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
92 // We don't accept any truncstore of integer registers.
93 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
94 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
95 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
96 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
97 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
98 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
100 // SETOEQ and SETUNE require checking two conditions.
101 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
102 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
103 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
104 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
105 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
106 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
108 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
110 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
111 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
112 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
114 if (Subtarget->is64Bit()) {
115 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
116 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
117 } else if (!UseSoftFloat) {
118 if (X86ScalarSSEf64) {
119 // We have an impenetrably clever algorithm for ui64->double only.
120 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
122 // We have an algorithm for SSE2, and we turn this into a 64-bit
123 // FILD for other targets.
124 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
127 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
129 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
130 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
133 // SSE has no i16 to fp conversion, only i32
134 if (X86ScalarSSEf32) {
135 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
136 // f32 and f64 cases are Legal, f80 case is not
137 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
139 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
140 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
143 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
144 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
147 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
148 // are Legal, f80 is custom lowered.
149 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
150 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
152 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
154 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
155 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
157 if (X86ScalarSSEf32) {
158 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
159 // f32 and f64 cases are Legal, f80 case is not
160 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
162 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
163 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
166 // Handle FP_TO_UINT by promoting the destination to a larger signed
168 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
169 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
170 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
172 if (Subtarget->is64Bit()) {
173 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
174 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
175 } else if (!UseSoftFloat) {
176 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
177 // Expand FP_TO_UINT into a select.
178 // FIXME: We would like to use a Custom expander here eventually to do
179 // the optimal thing for SSE vs. the default expansion in the legalizer.
180 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
182 // With SSE3 we can use fisttpll to convert to a signed i64; without
183 // SSE, we're stuck with a fistpll.
184 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
187 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
188 if (!X86ScalarSSEf64) {
189 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
190 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
193 // Scalar integer divide and remainder are lowered to use operations that
194 // produce two results, to match the available instructions. This exposes
195 // the two-result form to trivial CSE, which is able to combine x/y and x%y
196 // into a single instruction.
198 // Scalar integer multiply-high is also lowered to use two-result
199 // operations, to match the available instructions. However, plain multiply
200 // (low) operations are left as Legal, as there are single-result
201 // instructions for this in x86. Using the two-result multiply instructions
202 // when both high and low results are needed must be arranged by dagcombine.
203 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
204 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
205 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
206 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
207 setOperationAction(ISD::SREM , MVT::i8 , Expand);
208 setOperationAction(ISD::UREM , MVT::i8 , Expand);
209 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
210 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
211 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
212 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
213 setOperationAction(ISD::SREM , MVT::i16 , Expand);
214 setOperationAction(ISD::UREM , MVT::i16 , Expand);
215 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
216 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
217 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
218 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
219 setOperationAction(ISD::SREM , MVT::i32 , Expand);
220 setOperationAction(ISD::UREM , MVT::i32 , Expand);
221 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
222 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
223 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
224 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
225 setOperationAction(ISD::SREM , MVT::i64 , Expand);
226 setOperationAction(ISD::UREM , MVT::i64 , Expand);
228 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
229 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
230 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
231 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
232 if (Subtarget->is64Bit())
233 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
234 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
235 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
236 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
237 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
238 setOperationAction(ISD::FREM , MVT::f32 , Expand);
239 setOperationAction(ISD::FREM , MVT::f64 , Expand);
240 setOperationAction(ISD::FREM , MVT::f80 , Expand);
241 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
243 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
244 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
245 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
246 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
247 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
248 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
249 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
250 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
251 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
252 if (Subtarget->is64Bit()) {
253 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
254 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
255 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
258 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
259 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
261 // These should be promoted to a larger select which is supported.
262 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
263 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
264 // X86 wants to expand cmov itself.
265 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
266 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
267 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
268 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
269 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
270 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
271 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
272 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
273 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
274 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
275 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
276 if (Subtarget->is64Bit()) {
277 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
278 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
280 // X86 ret instruction may pop stack.
281 setOperationAction(ISD::RET , MVT::Other, Custom);
282 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
285 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
286 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
287 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
288 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
289 if (Subtarget->is64Bit())
290 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
291 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
292 if (Subtarget->is64Bit()) {
293 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
294 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
295 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
296 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
298 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
299 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
300 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
301 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
302 if (Subtarget->is64Bit()) {
303 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
304 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
305 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
308 if (Subtarget->hasSSE1())
309 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
311 if (!Subtarget->hasSSE2())
312 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
314 // Expand certain atomics
315 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
316 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
317 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
318 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
320 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
321 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
322 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
325 if (!Subtarget->is64Bit()) {
326 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
327 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
328 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
329 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
335 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
336 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
337 // FIXME - use subtarget debug flags
338 if (!Subtarget->isTargetDarwin() &&
339 !Subtarget->isTargetELF() &&
340 !Subtarget->isTargetCygMing()) {
341 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
342 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
345 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
346 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
347 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
348 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
349 if (Subtarget->is64Bit()) {
350 setExceptionPointerRegister(X86::RAX);
351 setExceptionSelectorRegister(X86::RDX);
353 setExceptionPointerRegister(X86::EAX);
354 setExceptionSelectorRegister(X86::EDX);
356 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
357 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
359 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
361 setOperationAction(ISD::TRAP, MVT::Other, Legal);
363 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
364 setOperationAction(ISD::VASTART , MVT::Other, Custom);
365 setOperationAction(ISD::VAEND , MVT::Other, Expand);
366 if (Subtarget->is64Bit()) {
367 setOperationAction(ISD::VAARG , MVT::Other, Custom);
368 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
370 setOperationAction(ISD::VAARG , MVT::Other, Expand);
371 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
374 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
375 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
376 if (Subtarget->is64Bit())
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
378 if (Subtarget->isTargetCygMing())
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
383 if (!UseSoftFloat && X86ScalarSSEf64) {
384 // f32 and f64 use SSE.
385 // Set up the FP register classes.
386 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
387 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
389 // Use ANDPD to simulate FABS.
390 setOperationAction(ISD::FABS , MVT::f64, Custom);
391 setOperationAction(ISD::FABS , MVT::f32, Custom);
393 // Use XORP to simulate FNEG.
394 setOperationAction(ISD::FNEG , MVT::f64, Custom);
395 setOperationAction(ISD::FNEG , MVT::f32, Custom);
397 // Use ANDPD and ORPD to simulate FCOPYSIGN.
398 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
399 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
401 // We don't support sin/cos/fmod
402 setOperationAction(ISD::FSIN , MVT::f64, Expand);
403 setOperationAction(ISD::FCOS , MVT::f64, Expand);
404 setOperationAction(ISD::FSIN , MVT::f32, Expand);
405 setOperationAction(ISD::FCOS , MVT::f32, Expand);
407 // Expand FP immediates into loads from the stack, except for the special
409 addLegalFPImmediate(APFloat(+0.0)); // xorpd
410 addLegalFPImmediate(APFloat(+0.0f)); // xorps
411 } else if (!UseSoftFloat && X86ScalarSSEf32) {
412 // Use SSE for f32, x87 for f64.
413 // Set up the FP register classes.
414 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
415 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
417 // Use ANDPS to simulate FABS.
418 setOperationAction(ISD::FABS , MVT::f32, Custom);
420 // Use XORP to simulate FNEG.
421 setOperationAction(ISD::FNEG , MVT::f32, Custom);
423 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
425 // Use ANDPS and ORPS to simulate FCOPYSIGN.
426 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
427 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
429 // We don't support sin/cos/fmod
430 setOperationAction(ISD::FSIN , MVT::f32, Expand);
431 setOperationAction(ISD::FCOS , MVT::f32, Expand);
433 // Special cases we handle for FP constants.
434 addLegalFPImmediate(APFloat(+0.0f)); // xorps
435 addLegalFPImmediate(APFloat(+0.0)); // FLD0
436 addLegalFPImmediate(APFloat(+1.0)); // FLD1
437 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
438 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
441 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
442 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
444 } else if (!UseSoftFloat) {
445 // f32 and f64 in x87.
446 // Set up the FP register classes.
447 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
448 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
450 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
451 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
452 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
453 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
456 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
457 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
459 addLegalFPImmediate(APFloat(+0.0)); // FLD0
460 addLegalFPImmediate(APFloat(+1.0)); // FLD1
461 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
462 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
463 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
464 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
465 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
466 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
469 // Long double always uses X87.
471 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
472 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
476 APFloat TmpFlt(+0.0);
477 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
479 addLegalFPImmediate(TmpFlt); // FLD0
481 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
482 APFloat TmpFlt2(+1.0);
483 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
485 addLegalFPImmediate(TmpFlt2); // FLD1
486 TmpFlt2.changeSign();
487 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
491 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
492 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
496 // Always use a library call for pow.
497 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
498 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
499 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
501 setOperationAction(ISD::FLOG, MVT::f80, Expand);
502 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
503 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
504 setOperationAction(ISD::FEXP, MVT::f80, Expand);
505 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
507 // First set operation action for all vector types to either promote
508 // (for widening) or expand (for scalarization). Then we will selectively
509 // turn on ones that can be effectively codegen'd.
510 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
511 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
512 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
513 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
514 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
515 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
516 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
517 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
524 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
525 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
527 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
528 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
562 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
563 // with -msoft-float, disable use of MMX as well.
564 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
565 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
566 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
567 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
568 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
569 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
571 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
572 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
573 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
574 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
576 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
577 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
578 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
579 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
581 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
582 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
584 setOperationAction(ISD::AND, MVT::v8i8, Promote);
585 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
586 setOperationAction(ISD::AND, MVT::v4i16, Promote);
587 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
588 setOperationAction(ISD::AND, MVT::v2i32, Promote);
589 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
590 setOperationAction(ISD::AND, MVT::v1i64, Legal);
592 setOperationAction(ISD::OR, MVT::v8i8, Promote);
593 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
594 setOperationAction(ISD::OR, MVT::v4i16, Promote);
595 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
596 setOperationAction(ISD::OR, MVT::v2i32, Promote);
597 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
598 setOperationAction(ISD::OR, MVT::v1i64, Legal);
600 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
601 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
602 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
603 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
604 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
605 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
606 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
608 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
609 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
610 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
611 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
612 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
613 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
614 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
615 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
616 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
619 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
620 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
621 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
622 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
624 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
625 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
626 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
629 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
630 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
631 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
632 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
634 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
636 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
637 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
638 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
639 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
640 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
641 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
644 if (!UseSoftFloat && Subtarget->hasSSE1()) {
645 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
647 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
648 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
649 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
650 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
651 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
652 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
653 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
654 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
656 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
657 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
658 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
661 if (!UseSoftFloat && Subtarget->hasSSE2()) {
662 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
664 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
665 // registers cannot be used even for integer operations.
666 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
667 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
668 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
669 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
671 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
672 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
673 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
674 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
675 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
676 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
677 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
678 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
679 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
680 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
681 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
682 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
683 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
684 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
685 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
686 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
690 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
691 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
694 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
696 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
697 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
699 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
700 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
701 MVT VT = (MVT::SimpleValueType)i;
702 // Do not attempt to custom lower non-power-of-2 vectors
703 if (!isPowerOf2_32(VT.getVectorNumElements()))
705 // Do not attempt to custom lower non-128-bit vectors
706 if (!VT.is128BitVector())
708 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
709 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
710 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
713 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
714 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
715 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
716 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
717 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
718 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
720 if (Subtarget->is64Bit()) {
721 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
722 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
725 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
726 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
727 MVT VT = (MVT::SimpleValueType)i;
729 // Do not attempt to promote non-128-bit vectors
730 if (!VT.is128BitVector()) {
733 setOperationAction(ISD::AND, VT, Promote);
734 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
735 setOperationAction(ISD::OR, VT, Promote);
736 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
737 setOperationAction(ISD::XOR, VT, Promote);
738 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
739 setOperationAction(ISD::LOAD, VT, Promote);
740 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
741 setOperationAction(ISD::SELECT, VT, Promote);
742 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
745 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
747 // Custom lower v2i64 and v2f64 selects.
748 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
749 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
750 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
751 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
753 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
754 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
755 if (!DisableMMX && Subtarget->hasMMX()) {
756 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
757 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
761 if (Subtarget->hasSSE41()) {
762 // FIXME: Do we need to handle scalar-to-vector here?
763 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
765 // i8 and i16 vectors are custom , because the source register and source
766 // source memory operand types are not the same width. f32 vectors are
767 // custom since the immediate controlling the insert encodes additional
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
770 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
771 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
772 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
774 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
775 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
776 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
777 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
779 if (Subtarget->is64Bit()) {
780 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
781 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
785 if (Subtarget->hasSSE42()) {
786 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
789 if (!UseSoftFloat && Subtarget->hasAVX()) {
790 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
791 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
792 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
793 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
795 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
796 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
797 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
798 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
799 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
800 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
801 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
802 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
803 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
804 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
805 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
806 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
807 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
808 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
809 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
811 // Operations to consider commented out -v16i16 v32i8
812 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
813 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
814 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
815 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
816 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
817 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
818 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
819 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
820 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
821 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
822 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
823 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
824 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
825 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
827 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
828 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
829 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
830 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
832 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
833 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
834 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
835 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
838 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
839 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
840 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
841 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
843 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
846 // Not sure we want to do this since there are no 256-bit integer
849 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
850 // This includes 256-bit vectors
851 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
852 MVT VT = (MVT::SimpleValueType)i;
854 // Do not attempt to custom lower non-power-of-2 vectors
855 if (!isPowerOf2_32(VT.getVectorNumElements()))
858 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
859 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
860 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
863 if (Subtarget->is64Bit()) {
864 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
870 // Not sure we want to do this since there are no 256-bit integer
873 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
874 // Including 256-bit vectors
875 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
876 MVT VT = (MVT::SimpleValueType)i;
878 if (!VT.is256BitVector()) {
881 setOperationAction(ISD::AND, VT, Promote);
882 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
883 setOperationAction(ISD::OR, VT, Promote);
884 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
885 setOperationAction(ISD::XOR, VT, Promote);
886 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
887 setOperationAction(ISD::LOAD, VT, Promote);
888 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
889 setOperationAction(ISD::SELECT, VT, Promote);
890 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
893 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
897 // We want to custom lower some of our intrinsics.
898 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
900 // Add/Sub/Mul with overflow operations are custom lowered.
901 setOperationAction(ISD::SADDO, MVT::i32, Custom);
902 setOperationAction(ISD::SADDO, MVT::i64, Custom);
903 setOperationAction(ISD::UADDO, MVT::i32, Custom);
904 setOperationAction(ISD::UADDO, MVT::i64, Custom);
905 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
906 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
907 setOperationAction(ISD::USUBO, MVT::i32, Custom);
908 setOperationAction(ISD::USUBO, MVT::i64, Custom);
909 setOperationAction(ISD::SMULO, MVT::i32, Custom);
910 setOperationAction(ISD::SMULO, MVT::i64, Custom);
912 if (!Subtarget->is64Bit()) {
913 // These libcalls are not available in 32-bit.
914 setLibcallName(RTLIB::SHL_I128, 0);
915 setLibcallName(RTLIB::SRL_I128, 0);
916 setLibcallName(RTLIB::SRA_I128, 0);
919 // We have target-specific dag combine patterns for the following nodes:
920 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
921 setTargetDAGCombine(ISD::BUILD_VECTOR);
922 setTargetDAGCombine(ISD::SELECT);
923 setTargetDAGCombine(ISD::SHL);
924 setTargetDAGCombine(ISD::SRA);
925 setTargetDAGCombine(ISD::SRL);
926 setTargetDAGCombine(ISD::STORE);
927 setTargetDAGCombine(ISD::MEMBARRIER);
928 if (Subtarget->is64Bit())
929 setTargetDAGCombine(ISD::MUL);
931 computeRegisterProperties();
933 // FIXME: These should be based on subtarget info. Plus, the values should
934 // be smaller when we are in optimizing for size mode.
935 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
936 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
937 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
938 allowUnalignedMemoryAccesses = true; // x86 supports it!
939 setPrefLoopAlignment(16);
940 benefitFromCodePlacementOpt = true;
944 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
949 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
950 /// the desired ByVal argument alignment.
951 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
954 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
955 if (VTy->getBitWidth() == 128)
957 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
958 unsigned EltAlign = 0;
959 getMaxByValAlign(ATy->getElementType(), EltAlign);
960 if (EltAlign > MaxAlign)
962 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
963 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
964 unsigned EltAlign = 0;
965 getMaxByValAlign(STy->getElementType(i), EltAlign);
966 if (EltAlign > MaxAlign)
975 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
976 /// function arguments in the caller parameter area. For X86, aggregates
977 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
978 /// are at 4-byte boundaries.
979 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
980 if (Subtarget->is64Bit()) {
981 // Max of 8 and alignment of type.
982 unsigned TyAlign = TD->getABITypeAlignment(Ty);
989 if (Subtarget->hasSSE1())
990 getMaxByValAlign(Ty, Align);
994 /// getOptimalMemOpType - Returns the target specific optimal type for load
995 /// and store operations as a result of memset, memcpy, and memmove
996 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
999 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1000 bool isSrcConst, bool isSrcStr,
1001 SelectionDAG &DAG) const {
1002 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1003 // linux. This is because the stack realignment code can't handle certain
1004 // cases like PR2962. This should be removed when PR2962 is fixed.
1005 const Function *F = DAG.getMachineFunction().getFunction();
1006 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1007 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1008 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1010 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1013 if (Subtarget->is64Bit() && Size >= 8)
1018 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1020 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1021 SelectionDAG &DAG) const {
1022 if (usesGlobalOffsetTable())
1023 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1024 if (!Subtarget->is64Bit())
1025 // This doesn't have DebugLoc associated with it, but is not really the
1026 // same as a Register.
1027 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1032 /// getFunctionAlignment - Return the Log2 alignment of this function.
1033 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1034 return F->hasFnAttr(Attribute::OptimizeForSize) ? 1 : 4;
1037 //===----------------------------------------------------------------------===//
1038 // Return Value Calling Convention Implementation
1039 //===----------------------------------------------------------------------===//
1041 #include "X86GenCallingConv.inc"
1043 /// LowerRET - Lower an ISD::RET node.
1044 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1045 DebugLoc dl = Op.getDebugLoc();
1046 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
1048 SmallVector<CCValAssign, 16> RVLocs;
1049 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1050 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1051 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
1052 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
1054 // If this is the first return lowered for this function, add the regs to the
1055 // liveout set for the function.
1056 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1057 for (unsigned i = 0; i != RVLocs.size(); ++i)
1058 if (RVLocs[i].isRegLoc())
1059 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1061 SDValue Chain = Op.getOperand(0);
1063 // Handle tail call return.
1064 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
1065 if (Chain.getOpcode() == X86ISD::TAILCALL) {
1066 SDValue TailCall = Chain;
1067 SDValue TargetAddress = TailCall.getOperand(1);
1068 SDValue StackAdjustment = TailCall.getOperand(2);
1069 assert(((TargetAddress.getOpcode() == ISD::Register &&
1070 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
1071 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
1072 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
1073 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
1074 "Expecting an global address, external symbol, or register");
1075 assert(StackAdjustment.getOpcode() == ISD::Constant &&
1076 "Expecting a const value");
1078 SmallVector<SDValue,8> Operands;
1079 Operands.push_back(Chain.getOperand(0));
1080 Operands.push_back(TargetAddress);
1081 Operands.push_back(StackAdjustment);
1082 // Copy registers used by the call. Last operand is a flag so it is not
1084 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
1085 Operands.push_back(Chain.getOperand(i));
1087 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
1094 SmallVector<SDValue, 6> RetOps;
1095 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1096 // Operand #1 = Bytes To Pop
1097 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1099 // Copy the result values into the output registers.
1100 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1101 CCValAssign &VA = RVLocs[i];
1102 assert(VA.isRegLoc() && "Can only return in registers!");
1103 SDValue ValToCopy = Op.getOperand(i*2+1);
1105 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1106 // the RET instruction and handled by the FP Stackifier.
1107 if (VA.getLocReg() == X86::ST0 ||
1108 VA.getLocReg() == X86::ST1) {
1109 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1110 // change the value to the FP stack register class.
1111 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1112 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1113 RetOps.push_back(ValToCopy);
1114 // Don't emit a copytoreg.
1118 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1119 // which is returned in RAX / RDX.
1120 if (Subtarget->is64Bit()) {
1121 MVT ValVT = ValToCopy.getValueType();
1122 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1123 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1124 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1125 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1129 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1130 Flag = Chain.getValue(1);
1133 // The x86-64 ABI for returning structs by value requires that we copy
1134 // the sret argument into %rax for the return. We saved the argument into
1135 // a virtual register in the entry block, so now we copy the value out
1137 if (Subtarget->is64Bit() &&
1138 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1139 MachineFunction &MF = DAG.getMachineFunction();
1140 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1141 unsigned Reg = FuncInfo->getSRetReturnReg();
1143 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1144 FuncInfo->setSRetReturnReg(Reg);
1146 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1148 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1149 Flag = Chain.getValue(1);
1152 RetOps[0] = Chain; // Update chain.
1154 // Add the flag if we have it.
1156 RetOps.push_back(Flag);
1158 return DAG.getNode(X86ISD::RET_FLAG, dl,
1159 MVT::Other, &RetOps[0], RetOps.size());
1163 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1164 /// appropriate copies out of appropriate physical registers. This assumes that
1165 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1166 /// being lowered. The returns a SDNode with the same number of values as the
1168 SDNode *X86TargetLowering::
1169 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1170 unsigned CallingConv, SelectionDAG &DAG) {
1172 DebugLoc dl = TheCall->getDebugLoc();
1173 // Assign locations to each value returned by this call.
1174 SmallVector<CCValAssign, 16> RVLocs;
1175 bool isVarArg = TheCall->isVarArg();
1176 bool Is64Bit = Subtarget->is64Bit();
1177 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1178 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1180 SmallVector<SDValue, 8> ResultVals;
1182 // Copy all of the result registers out of their specified physreg.
1183 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1184 CCValAssign &VA = RVLocs[i];
1185 MVT CopyVT = VA.getValVT();
1187 // If this is x86-64, and we disabled SSE, we can't return FP values
1188 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1189 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1190 llvm_report_error("SSE register return with SSE disabled");
1193 // If this is a call to a function that returns an fp value on the floating
1194 // point stack, but where we prefer to use the value in xmm registers, copy
1195 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1196 if ((VA.getLocReg() == X86::ST0 ||
1197 VA.getLocReg() == X86::ST1) &&
1198 isScalarFPTypeInSSEReg(VA.getValVT())) {
1203 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1204 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1205 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1206 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1207 MVT::v2i64, InFlag).getValue(1);
1208 Val = Chain.getValue(0);
1209 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1210 Val, DAG.getConstant(0, MVT::i64));
1212 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1213 MVT::i64, InFlag).getValue(1);
1214 Val = Chain.getValue(0);
1216 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1218 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1219 CopyVT, InFlag).getValue(1);
1220 Val = Chain.getValue(0);
1222 InFlag = Chain.getValue(2);
1224 if (CopyVT != VA.getValVT()) {
1225 // Round the F80 the right size, which also moves to the appropriate xmm
1227 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1228 // This truncation won't change the value.
1229 DAG.getIntPtrConstant(1));
1232 ResultVals.push_back(Val);
1235 // Merge everything together with a MERGE_VALUES node.
1236 ResultVals.push_back(Chain);
1237 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1238 &ResultVals[0], ResultVals.size()).getNode();
1242 //===----------------------------------------------------------------------===//
1243 // C & StdCall & Fast Calling Convention implementation
1244 //===----------------------------------------------------------------------===//
1245 // StdCall calling convention seems to be standard for many Windows' API
1246 // routines and around. It differs from C calling convention just a little:
1247 // callee should clean up the stack, not caller. Symbols should be also
1248 // decorated in some fancy way :) It doesn't support any vector arguments.
1249 // For info on fast calling convention see Fast Calling Convention (tail call)
1250 // implementation LowerX86_32FastCCCallTo.
1252 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1254 static bool CallIsStructReturn(CallSDNode *TheCall) {
1255 unsigned NumOps = TheCall->getNumArgs();
1259 return TheCall->getArgFlags(0).isSRet();
1262 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1263 /// return semantics.
1264 static bool ArgsAreStructReturn(SDValue Op) {
1265 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1269 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1272 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1273 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1275 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1279 switch (CallingConv) {
1282 case CallingConv::X86_StdCall:
1283 return !Subtarget->is64Bit();
1284 case CallingConv::X86_FastCall:
1285 return !Subtarget->is64Bit();
1286 case CallingConv::Fast:
1287 return PerformTailCallOpt;
1291 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1292 /// given CallingConvention value.
1293 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1294 if (Subtarget->is64Bit()) {
1295 if (Subtarget->isTargetWin64())
1296 return CC_X86_Win64_C;
1301 if (CC == CallingConv::X86_FastCall)
1302 return CC_X86_32_FastCall;
1303 else if (CC == CallingConv::Fast)
1304 return CC_X86_32_FastCC;
1309 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1310 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1312 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1313 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1314 if (CC == CallingConv::X86_FastCall)
1316 else if (CC == CallingConv::X86_StdCall)
1322 /// isUsingGOT - Return true if the target uses a GOT for PIC, and if we're in
1324 static bool isUsingGOT(const TargetMachine &TM) {
1325 const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
1326 return TM.getRelocationModel() == Reloc::PIC_ &&
1327 Subtarget.isPICStyleGOT();
1330 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1331 /// by "Src" to address "Dst" with size and alignment information specified by
1332 /// the specific parameter attribute. The copy will be passed as a byval
1333 /// function parameter.
1335 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1336 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1338 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1339 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1340 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1343 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1344 const CCValAssign &VA,
1345 MachineFrameInfo *MFI,
1347 SDValue Root, unsigned i) {
1348 // Create the nodes corresponding to a load from this parameter slot.
1349 ISD::ArgFlagsTy Flags =
1350 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1351 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1352 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1354 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1355 // changed with more analysis.
1356 // In case of tail call optimization mark all arguments mutable. Since they
1357 // could be overwritten by lowering of arguments in case of a tail call.
1358 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1359 VA.getLocMemOffset(), isImmutable);
1360 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1361 if (Flags.isByVal())
1363 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1364 PseudoSourceValue::getFixedStack(FI), 0);
1368 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1369 MachineFunction &MF = DAG.getMachineFunction();
1370 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1371 DebugLoc dl = Op.getDebugLoc();
1373 const Function* Fn = MF.getFunction();
1374 if (Fn->hasExternalLinkage() &&
1375 Subtarget->isTargetCygMing() &&
1376 Fn->getName() == "main")
1377 FuncInfo->setForceFramePointer(true);
1379 // Decorate the function name.
1380 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1382 MachineFrameInfo *MFI = MF.getFrameInfo();
1383 SDValue Root = Op.getOperand(0);
1384 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1385 unsigned CC = MF.getFunction()->getCallingConv();
1386 bool Is64Bit = Subtarget->is64Bit();
1387 bool IsWin64 = Subtarget->isTargetWin64();
1389 assert(!(isVarArg && CC == CallingConv::Fast) &&
1390 "Var args not supported with calling convention fastcc");
1392 // Assign locations to all of the incoming arguments.
1393 SmallVector<CCValAssign, 16> ArgLocs;
1394 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1395 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1397 SmallVector<SDValue, 8> ArgValues;
1398 unsigned LastVal = ~0U;
1399 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1400 CCValAssign &VA = ArgLocs[i];
1401 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1403 assert(VA.getValNo() != LastVal &&
1404 "Don't support value assigned to multiple locs yet");
1405 LastVal = VA.getValNo();
1407 if (VA.isRegLoc()) {
1408 MVT RegVT = VA.getLocVT();
1409 TargetRegisterClass *RC = NULL;
1410 if (RegVT == MVT::i32)
1411 RC = X86::GR32RegisterClass;
1412 else if (Is64Bit && RegVT == MVT::i64)
1413 RC = X86::GR64RegisterClass;
1414 else if (RegVT == MVT::f32)
1415 RC = X86::FR32RegisterClass;
1416 else if (RegVT == MVT::f64)
1417 RC = X86::FR64RegisterClass;
1418 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1419 RC = X86::VR128RegisterClass;
1420 else if (RegVT.isVector()) {
1421 assert(RegVT.getSizeInBits() == 64);
1423 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1425 // Darwin calling convention passes MMX values in either GPRs or
1426 // XMMs in x86-64. Other targets pass them in memory.
1427 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1428 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1431 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1436 assert(0 && "Unknown argument type!");
1439 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1440 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1442 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1443 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1445 if (VA.getLocInfo() == CCValAssign::SExt)
1446 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1447 DAG.getValueType(VA.getValVT()));
1448 else if (VA.getLocInfo() == CCValAssign::ZExt)
1449 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1450 DAG.getValueType(VA.getValVT()));
1452 if (VA.getLocInfo() != CCValAssign::Full)
1453 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1455 // Handle MMX values passed in GPRs.
1456 if (Is64Bit && RegVT != VA.getLocVT()) {
1457 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1458 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1459 else if (RC == X86::VR128RegisterClass) {
1460 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1461 ArgValue, DAG.getConstant(0, MVT::i64));
1462 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1466 ArgValues.push_back(ArgValue);
1468 assert(VA.isMemLoc());
1469 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1473 // The x86-64 ABI for returning structs by value requires that we copy
1474 // the sret argument into %rax for the return. Save the argument into
1475 // a virtual register so that we can access it from the return points.
1476 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1477 MachineFunction &MF = DAG.getMachineFunction();
1478 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1479 unsigned Reg = FuncInfo->getSRetReturnReg();
1481 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1482 FuncInfo->setSRetReturnReg(Reg);
1484 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1485 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1488 unsigned StackSize = CCInfo.getNextStackOffset();
1489 // align stack specially for tail calls
1490 if (PerformTailCallOpt && CC == CallingConv::Fast)
1491 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1493 // If the function takes variable number of arguments, make a frame index for
1494 // the start of the first vararg value... for expansion of llvm.va_start.
1496 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1497 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1500 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1502 // FIXME: We should really autogenerate these arrays
1503 static const unsigned GPR64ArgRegsWin64[] = {
1504 X86::RCX, X86::RDX, X86::R8, X86::R9
1506 static const unsigned XMMArgRegsWin64[] = {
1507 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1509 static const unsigned GPR64ArgRegs64Bit[] = {
1510 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1512 static const unsigned XMMArgRegs64Bit[] = {
1513 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1514 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1516 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1519 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1520 GPR64ArgRegs = GPR64ArgRegsWin64;
1521 XMMArgRegs = XMMArgRegsWin64;
1523 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1524 GPR64ArgRegs = GPR64ArgRegs64Bit;
1525 XMMArgRegs = XMMArgRegs64Bit;
1527 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1529 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1532 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1533 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1534 "SSE register cannot be used when SSE is disabled!");
1535 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1536 "SSE register cannot be used when SSE is disabled!");
1537 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1538 // Kernel mode asks for SSE to be disabled, so don't push them
1540 TotalNumXMMRegs = 0;
1542 // For X86-64, if there are vararg parameters that are passed via
1543 // registers, then we must store them to their spots on the stack so they
1544 // may be loaded by deferencing the result of va_next.
1545 VarArgsGPOffset = NumIntRegs * 8;
1546 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1547 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1548 TotalNumXMMRegs * 16, 16);
1550 // Store the integer parameter registers.
1551 SmallVector<SDValue, 8> MemOps;
1552 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1553 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1554 DAG.getIntPtrConstant(VarArgsGPOffset));
1555 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1556 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1557 X86::GR64RegisterClass);
1558 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1560 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1561 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1562 MemOps.push_back(Store);
1563 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1564 DAG.getIntPtrConstant(8));
1567 // Now store the XMM (fp + vector) parameter registers.
1568 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1569 DAG.getIntPtrConstant(VarArgsFPOffset));
1570 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1571 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1572 X86::VR128RegisterClass);
1573 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1575 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1576 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1577 MemOps.push_back(Store);
1578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1579 DAG.getIntPtrConstant(16));
1581 if (!MemOps.empty())
1582 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1583 &MemOps[0], MemOps.size());
1587 ArgValues.push_back(Root);
1589 // Some CCs need callee pop.
1590 if (IsCalleePop(isVarArg, CC)) {
1591 BytesToPopOnReturn = StackSize; // Callee pops everything.
1592 BytesCallerReserves = 0;
1594 BytesToPopOnReturn = 0; // Callee pops nothing.
1595 // If this is an sret function, the return should pop the hidden pointer.
1596 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1597 BytesToPopOnReturn = 4;
1598 BytesCallerReserves = StackSize;
1602 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1603 if (CC == CallingConv::X86_FastCall)
1604 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1607 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1609 // Return the new list of results.
1610 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1611 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1615 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1616 const SDValue &StackPtr,
1617 const CCValAssign &VA,
1619 SDValue Arg, ISD::ArgFlagsTy Flags) {
1620 DebugLoc dl = TheCall->getDebugLoc();
1621 unsigned LocMemOffset = VA.getLocMemOffset();
1622 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1623 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1624 if (Flags.isByVal()) {
1625 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1627 return DAG.getStore(Chain, dl, Arg, PtrOff,
1628 PseudoSourceValue::getStack(), LocMemOffset);
1631 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1632 /// optimization is performed and it is required.
1634 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1635 SDValue &OutRetAddr,
1641 if (!IsTailCall || FPDiff==0) return Chain;
1643 // Adjust the Return address stack slot.
1644 MVT VT = getPointerTy();
1645 OutRetAddr = getReturnAddressFrameIndex(DAG);
1647 // Load the "old" Return address.
1648 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1649 return SDValue(OutRetAddr.getNode(), 1);
1652 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1653 /// optimization is performed and it is required (FPDiff!=0).
1655 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1656 SDValue Chain, SDValue RetAddrFrIdx,
1657 bool Is64Bit, int FPDiff, DebugLoc dl) {
1658 // Store the return address to the appropriate stack slot.
1659 if (!FPDiff) return Chain;
1660 // Calculate the new stack slot for the return address.
1661 int SlotSize = Is64Bit ? 8 : 4;
1662 int NewReturnAddrFI =
1663 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1664 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1665 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1666 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1667 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1671 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1672 MachineFunction &MF = DAG.getMachineFunction();
1673 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1674 SDValue Chain = TheCall->getChain();
1675 unsigned CC = TheCall->getCallingConv();
1676 bool isVarArg = TheCall->isVarArg();
1677 bool IsTailCall = TheCall->isTailCall() &&
1678 CC == CallingConv::Fast && PerformTailCallOpt;
1679 SDValue Callee = TheCall->getCallee();
1680 bool Is64Bit = Subtarget->is64Bit();
1681 bool IsStructRet = CallIsStructReturn(TheCall);
1682 DebugLoc dl = TheCall->getDebugLoc();
1684 assert(!(isVarArg && CC == CallingConv::Fast) &&
1685 "Var args not supported with calling convention fastcc");
1687 // Analyze operands of the call, assigning locations to each operand.
1688 SmallVector<CCValAssign, 16> ArgLocs;
1689 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1690 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1692 // Get a count of how many bytes are to be pushed on the stack.
1693 unsigned NumBytes = CCInfo.getNextStackOffset();
1694 if (PerformTailCallOpt && CC == CallingConv::Fast)
1695 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1699 // Lower arguments at fp - stackoffset + fpdiff.
1700 unsigned NumBytesCallerPushed =
1701 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1702 FPDiff = NumBytesCallerPushed - NumBytes;
1704 // Set the delta of movement of the returnaddr stackslot.
1705 // But only set if delta is greater than previous delta.
1706 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1707 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1710 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1712 SDValue RetAddrFrIdx;
1713 // Load return adress for tail calls.
1714 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1717 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1718 SmallVector<SDValue, 8> MemOpChains;
1721 // Walk the register/memloc assignments, inserting copies/loads. In the case
1722 // of tail call optimization arguments are handle later.
1723 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1724 CCValAssign &VA = ArgLocs[i];
1725 SDValue Arg = TheCall->getArg(i);
1726 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1727 bool isByVal = Flags.isByVal();
1729 // Promote the value if needed.
1730 switch (VA.getLocInfo()) {
1731 default: assert(0 && "Unknown loc info!");
1732 case CCValAssign::Full: break;
1733 case CCValAssign::SExt:
1734 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1736 case CCValAssign::ZExt:
1737 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1739 case CCValAssign::AExt:
1740 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1744 if (VA.isRegLoc()) {
1746 MVT RegVT = VA.getLocVT();
1747 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1748 switch (VA.getLocReg()) {
1751 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1753 // Special case: passing MMX values in GPR registers.
1754 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1757 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1758 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1759 // Special case: passing MMX values in XMM registers.
1760 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1761 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1762 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1767 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1769 if (!IsTailCall || (IsTailCall && isByVal)) {
1770 assert(VA.isMemLoc());
1771 if (StackPtr.getNode() == 0)
1772 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1774 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1775 Chain, Arg, Flags));
1780 if (!MemOpChains.empty())
1781 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1782 &MemOpChains[0], MemOpChains.size());
1784 // Build a sequence of copy-to-reg nodes chained together with token chain
1785 // and flag operands which copy the outgoing args into registers.
1787 // Tail call byval lowering might overwrite argument registers so in case of
1788 // tail call optimization the copies to registers are lowered later.
1790 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1791 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1792 RegsToPass[i].second, InFlag);
1793 InFlag = Chain.getValue(1);
1797 if (isUsingGOT(getTargetMachine())) {
1798 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1801 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1802 DAG.getNode(X86ISD::GlobalBaseReg,
1803 DebugLoc::getUnknownLoc(),
1806 InFlag = Chain.getValue(1);
1808 // If we are tail calling and generating PIC/GOT style code load the
1809 // address of the callee into ECX. The value in ecx is used as target of
1810 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1811 // for tail calls on PIC/GOT architectures. Normally we would just put the
1812 // address of GOT into ebx and then call target@PLT. But for tail calls
1813 // ebx would be restored (since ebx is callee saved) before jumping to the
1816 // Note: The actual moving to ECX is done further down.
1817 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1818 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1819 !G->getGlobal()->hasProtectedVisibility())
1820 Callee = LowerGlobalAddress(Callee, DAG);
1821 else if (isa<ExternalSymbolSDNode>(Callee))
1822 Callee = LowerExternalSymbol(Callee,DAG);
1826 if (Is64Bit && isVarArg) {
1827 // From AMD64 ABI document:
1828 // For calls that may call functions that use varargs or stdargs
1829 // (prototype-less calls or calls to functions containing ellipsis (...) in
1830 // the declaration) %al is used as hidden argument to specify the number
1831 // of SSE registers used. The contents of %al do not need to match exactly
1832 // the number of registers, but must be an ubound on the number of SSE
1833 // registers used and is in the range 0 - 8 inclusive.
1835 // FIXME: Verify this on Win64
1836 // Count the number of XMM registers allocated.
1837 static const unsigned XMMArgRegs[] = {
1838 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1839 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1841 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1842 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1843 && "SSE registers cannot be used when SSE is disabled");
1845 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1846 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1847 InFlag = Chain.getValue(1);
1851 // For tail calls lower the arguments to the 'real' stack slot.
1853 SmallVector<SDValue, 8> MemOpChains2;
1856 // Do not flag preceeding copytoreg stuff together with the following stuff.
1858 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1859 CCValAssign &VA = ArgLocs[i];
1860 if (!VA.isRegLoc()) {
1861 assert(VA.isMemLoc());
1862 SDValue Arg = TheCall->getArg(i);
1863 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1864 // Create frame index.
1865 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1866 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1867 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1868 FIN = DAG.getFrameIndex(FI, getPointerTy());
1870 if (Flags.isByVal()) {
1871 // Copy relative to framepointer.
1872 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1873 if (StackPtr.getNode() == 0)
1874 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1876 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1878 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1881 // Store relative to framepointer.
1882 MemOpChains2.push_back(
1883 DAG.getStore(Chain, dl, Arg, FIN,
1884 PseudoSourceValue::getFixedStack(FI), 0));
1889 if (!MemOpChains2.empty())
1890 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1891 &MemOpChains2[0], MemOpChains2.size());
1893 // Copy arguments to their registers.
1894 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1895 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1896 RegsToPass[i].second, InFlag);
1897 InFlag = Chain.getValue(1);
1901 // Store the return address to the appropriate stack slot.
1902 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1906 // If the callee is a GlobalAddress node (quite common, every direct call is)
1907 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1908 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1909 // We should use extra load for direct calls to dllimported functions in
1911 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1912 getTargetMachine(), true))
1913 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1915 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1916 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1917 } else if (IsTailCall) {
1918 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1920 Chain = DAG.getCopyToReg(Chain, dl,
1921 DAG.getRegister(Opc, getPointerTy()),
1923 Callee = DAG.getRegister(Opc, getPointerTy());
1924 // Add register as live out.
1925 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1928 // Returns a chain & a flag for retval copy to use.
1929 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1930 SmallVector<SDValue, 8> Ops;
1933 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1934 DAG.getIntPtrConstant(0, true), InFlag);
1935 InFlag = Chain.getValue(1);
1937 // Returns a chain & a flag for retval copy to use.
1938 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1942 Ops.push_back(Chain);
1943 Ops.push_back(Callee);
1946 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1948 // Add argument registers to the end of the list so that they are known live
1950 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1951 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1952 RegsToPass[i].second.getValueType()));
1954 // Add an implicit use GOT pointer in EBX.
1955 if (!IsTailCall && isUsingGOT(getTargetMachine()))
1956 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1958 // Add an implicit use of AL for x86 vararg functions.
1959 if (Is64Bit && isVarArg)
1960 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1962 if (InFlag.getNode())
1963 Ops.push_back(InFlag);
1966 assert(InFlag.getNode() &&
1967 "Flag must be set. Depend on flag being set in LowerRET");
1968 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1969 TheCall->getVTList(), &Ops[0], Ops.size());
1971 return SDValue(Chain.getNode(), Op.getResNo());
1974 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1975 InFlag = Chain.getValue(1);
1977 // Create the CALLSEQ_END node.
1978 unsigned NumBytesForCalleeToPush;
1979 if (IsCalleePop(isVarArg, CC))
1980 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1981 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1982 // If this is is a call to a struct-return function, the callee
1983 // pops the hidden struct pointer, so we have to push it back.
1984 // This is common for Darwin/X86, Linux & Mingw32 targets.
1985 NumBytesForCalleeToPush = 4;
1987 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1989 // Returns a flag for retval copy to use.
1990 Chain = DAG.getCALLSEQ_END(Chain,
1991 DAG.getIntPtrConstant(NumBytes, true),
1992 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1995 InFlag = Chain.getValue(1);
1997 // Handle result values, copying them out of physregs into vregs that we
1999 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
2004 //===----------------------------------------------------------------------===//
2005 // Fast Calling Convention (tail call) implementation
2006 //===----------------------------------------------------------------------===//
2008 // Like std call, callee cleans arguments, convention except that ECX is
2009 // reserved for storing the tail called function address. Only 2 registers are
2010 // free for argument passing (inreg). Tail call optimization is performed
2012 // * tailcallopt is enabled
2013 // * caller/callee are fastcc
2014 // On X86_64 architecture with GOT-style position independent code only local
2015 // (within module) calls are supported at the moment.
2016 // To keep the stack aligned according to platform abi the function
2017 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2018 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2019 // If a tail called function callee has more arguments than the caller the
2020 // caller needs to make sure that there is room to move the RETADDR to. This is
2021 // achieved by reserving an area the size of the argument delta right after the
2022 // original REtADDR, but before the saved framepointer or the spilled registers
2023 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2035 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2036 /// for a 16 byte align requirement.
2037 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2038 SelectionDAG& DAG) {
2039 MachineFunction &MF = DAG.getMachineFunction();
2040 const TargetMachine &TM = MF.getTarget();
2041 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2042 unsigned StackAlignment = TFI.getStackAlignment();
2043 uint64_t AlignMask = StackAlignment - 1;
2044 int64_t Offset = StackSize;
2045 uint64_t SlotSize = TD->getPointerSize();
2046 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2047 // Number smaller than 12 so just add the difference.
2048 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2050 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2051 Offset = ((~AlignMask) & Offset) + StackAlignment +
2052 (StackAlignment-SlotSize);
2057 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2058 /// following the call is a return. A function is eligible if caller/callee
2059 /// calling conventions match, currently only fastcc supports tail calls, and
2060 /// the function CALL is immediatly followed by a RET.
2061 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2063 SelectionDAG& DAG) const {
2064 if (!PerformTailCallOpt)
2067 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2068 MachineFunction &MF = DAG.getMachineFunction();
2069 unsigned CallerCC = MF.getFunction()->getCallingConv();
2070 unsigned CalleeCC= TheCall->getCallingConv();
2071 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2072 SDValue Callee = TheCall->getCallee();
2073 // On x86/32Bit PIC/GOT tail calls are supported.
2074 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
2075 !Subtarget->isPICStyleGOT() || !Subtarget->is64Bit())
2078 // Can only do local tail calls (in same module, hidden or protected) on
2079 // x86_64 PIC/GOT at the moment.
2080 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2081 return G->getGlobal()->hasHiddenVisibility()
2082 || G->getGlobal()->hasProtectedVisibility();
2090 X86TargetLowering::createFastISel(MachineFunction &mf,
2091 MachineModuleInfo *mmo,
2093 DenseMap<const Value *, unsigned> &vm,
2094 DenseMap<const BasicBlock *,
2095 MachineBasicBlock *> &bm,
2096 DenseMap<const AllocaInst *, int> &am
2098 , SmallSet<Instruction*, 8> &cil
2101 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2109 //===----------------------------------------------------------------------===//
2110 // Other Lowering Hooks
2111 //===----------------------------------------------------------------------===//
2114 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2115 MachineFunction &MF = DAG.getMachineFunction();
2116 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2117 int ReturnAddrIndex = FuncInfo->getRAIndex();
2119 if (ReturnAddrIndex == 0) {
2120 // Set up a frame object for the return address.
2121 uint64_t SlotSize = TD->getPointerSize();
2122 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2123 FuncInfo->setRAIndex(ReturnAddrIndex);
2126 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2130 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2131 /// specific condition code, returning the condition code and the LHS/RHS of the
2132 /// comparison to make.
2133 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2134 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2136 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2137 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2138 // X > -1 -> X == 0, jump !sign.
2139 RHS = DAG.getConstant(0, RHS.getValueType());
2140 return X86::COND_NS;
2141 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2142 // X < 0 -> X == 0, jump on sign.
2144 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2146 RHS = DAG.getConstant(0, RHS.getValueType());
2147 return X86::COND_LE;
2151 switch (SetCCOpcode) {
2152 default: assert(0 && "Invalid integer condition!");
2153 case ISD::SETEQ: return X86::COND_E;
2154 case ISD::SETGT: return X86::COND_G;
2155 case ISD::SETGE: return X86::COND_GE;
2156 case ISD::SETLT: return X86::COND_L;
2157 case ISD::SETLE: return X86::COND_LE;
2158 case ISD::SETNE: return X86::COND_NE;
2159 case ISD::SETULT: return X86::COND_B;
2160 case ISD::SETUGT: return X86::COND_A;
2161 case ISD::SETULE: return X86::COND_BE;
2162 case ISD::SETUGE: return X86::COND_AE;
2166 // First determine if it is required or is profitable to flip the operands.
2168 // If LHS is a foldable load, but RHS is not, flip the condition.
2169 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2170 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2171 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2172 std::swap(LHS, RHS);
2175 switch (SetCCOpcode) {
2181 std::swap(LHS, RHS);
2185 // On a floating point condition, the flags are set as follows:
2187 // 0 | 0 | 0 | X > Y
2188 // 0 | 0 | 1 | X < Y
2189 // 1 | 0 | 0 | X == Y
2190 // 1 | 1 | 1 | unordered
2191 switch (SetCCOpcode) {
2192 default: assert(0 && "Condcode should be pre-legalized away");
2194 case ISD::SETEQ: return X86::COND_E;
2195 case ISD::SETOLT: // flipped
2197 case ISD::SETGT: return X86::COND_A;
2198 case ISD::SETOLE: // flipped
2200 case ISD::SETGE: return X86::COND_AE;
2201 case ISD::SETUGT: // flipped
2203 case ISD::SETLT: return X86::COND_B;
2204 case ISD::SETUGE: // flipped
2206 case ISD::SETLE: return X86::COND_BE;
2208 case ISD::SETNE: return X86::COND_NE;
2209 case ISD::SETUO: return X86::COND_P;
2210 case ISD::SETO: return X86::COND_NP;
2214 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2215 /// code. Current x86 isa includes the following FP cmov instructions:
2216 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2217 static bool hasFPCMov(unsigned X86CC) {
2233 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2234 /// the specified range (L, H].
2235 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2236 return (Val < 0) || (Val >= Low && Val < Hi);
2239 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2240 /// specified value.
2241 static bool isUndefOrEqual(int Val, int CmpVal) {
2242 if (Val < 0 || Val == CmpVal)
2247 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2248 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2249 /// the second operand.
2250 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2251 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2252 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2253 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2254 return (Mask[0] < 2 && Mask[1] < 2);
2258 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2259 SmallVector<int, 8> M;
2261 return ::isPSHUFDMask(M, N->getValueType(0));
2264 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2265 /// is suitable for input to PSHUFHW.
2266 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2267 if (VT != MVT::v8i16)
2270 // Lower quadword copied in order or undef.
2271 for (int i = 0; i != 4; ++i)
2272 if (Mask[i] >= 0 && Mask[i] != i)
2275 // Upper quadword shuffled.
2276 for (int i = 4; i != 8; ++i)
2277 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2283 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2284 SmallVector<int, 8> M;
2286 return ::isPSHUFHWMask(M, N->getValueType(0));
2289 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2290 /// is suitable for input to PSHUFLW.
2291 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2292 if (VT != MVT::v8i16)
2295 // Upper quadword copied in order.
2296 for (int i = 4; i != 8; ++i)
2297 if (Mask[i] >= 0 && Mask[i] != i)
2300 // Lower quadword shuffled.
2301 for (int i = 0; i != 4; ++i)
2308 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2309 SmallVector<int, 8> M;
2311 return ::isPSHUFLWMask(M, N->getValueType(0));
2314 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2315 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2316 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2317 int NumElems = VT.getVectorNumElements();
2318 if (NumElems != 2 && NumElems != 4)
2321 int Half = NumElems / 2;
2322 for (int i = 0; i < Half; ++i)
2323 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2325 for (int i = Half; i < NumElems; ++i)
2326 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2332 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2333 SmallVector<int, 8> M;
2335 return ::isSHUFPMask(M, N->getValueType(0));
2338 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2339 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2340 /// half elements to come from vector 1 (which would equal the dest.) and
2341 /// the upper half to come from vector 2.
2342 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2343 int NumElems = VT.getVectorNumElements();
2345 if (NumElems != 2 && NumElems != 4)
2348 int Half = NumElems / 2;
2349 for (int i = 0; i < Half; ++i)
2350 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2352 for (int i = Half; i < NumElems; ++i)
2353 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2358 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2359 SmallVector<int, 8> M;
2361 return isCommutedSHUFPMask(M, N->getValueType(0));
2364 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2365 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2366 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2367 if (N->getValueType(0).getVectorNumElements() != 4)
2370 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2371 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2372 isUndefOrEqual(N->getMaskElt(1), 7) &&
2373 isUndefOrEqual(N->getMaskElt(2), 2) &&
2374 isUndefOrEqual(N->getMaskElt(3), 3);
2377 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2378 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2379 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2380 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2382 if (NumElems != 2 && NumElems != 4)
2385 for (unsigned i = 0; i < NumElems/2; ++i)
2386 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2389 for (unsigned i = NumElems/2; i < NumElems; ++i)
2390 if (!isUndefOrEqual(N->getMaskElt(i), i))
2396 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2397 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2399 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2400 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2402 if (NumElems != 2 && NumElems != 4)
2405 for (unsigned i = 0; i < NumElems/2; ++i)
2406 if (!isUndefOrEqual(N->getMaskElt(i), i))
2409 for (unsigned i = 0; i < NumElems/2; ++i)
2410 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2416 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2417 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2419 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2420 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2425 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2426 isUndefOrEqual(N->getMaskElt(1), 3) &&
2427 isUndefOrEqual(N->getMaskElt(2), 2) &&
2428 isUndefOrEqual(N->getMaskElt(3), 3);
2431 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2432 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2433 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2434 bool V2IsSplat = false) {
2435 int NumElts = VT.getVectorNumElements();
2436 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2439 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2441 int BitI1 = Mask[i+1];
2442 if (!isUndefOrEqual(BitI, j))
2445 if (!isUndefOrEqual(BitI1, NumElts))
2448 if (!isUndefOrEqual(BitI1, j + NumElts))
2455 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2456 SmallVector<int, 8> M;
2458 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2461 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2462 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2463 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2464 bool V2IsSplat = false) {
2465 int NumElts = VT.getVectorNumElements();
2466 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2469 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2471 int BitI1 = Mask[i+1];
2472 if (!isUndefOrEqual(BitI, j + NumElts/2))
2475 if (isUndefOrEqual(BitI1, NumElts))
2478 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2485 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2486 SmallVector<int, 8> M;
2488 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2491 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2492 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2494 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2495 int NumElems = VT.getVectorNumElements();
2496 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2499 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2501 int BitI1 = Mask[i+1];
2502 if (!isUndefOrEqual(BitI, j))
2504 if (!isUndefOrEqual(BitI1, j))
2510 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2511 SmallVector<int, 8> M;
2513 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2516 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2517 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2519 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2520 int NumElems = VT.getVectorNumElements();
2521 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2524 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2526 int BitI1 = Mask[i+1];
2527 if (!isUndefOrEqual(BitI, j))
2529 if (!isUndefOrEqual(BitI1, j))
2535 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2536 SmallVector<int, 8> M;
2538 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2541 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2542 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2543 /// MOVSD, and MOVD, i.e. setting the lowest element.
2544 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2545 if (VT.getVectorElementType().getSizeInBits() < 32)
2548 int NumElts = VT.getVectorNumElements();
2550 if (!isUndefOrEqual(Mask[0], NumElts))
2553 for (int i = 1; i < NumElts; ++i)
2554 if (!isUndefOrEqual(Mask[i], i))
2560 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2561 SmallVector<int, 8> M;
2563 return ::isMOVLMask(M, N->getValueType(0));
2566 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2567 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2568 /// element of vector 2 and the other elements to come from vector 1 in order.
2569 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2570 bool V2IsSplat = false, bool V2IsUndef = false) {
2571 int NumOps = VT.getVectorNumElements();
2572 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2575 if (!isUndefOrEqual(Mask[0], 0))
2578 for (int i = 1; i < NumOps; ++i)
2579 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2580 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2581 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2587 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2588 bool V2IsUndef = false) {
2589 SmallVector<int, 8> M;
2591 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2594 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2595 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2596 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2597 if (N->getValueType(0).getVectorNumElements() != 4)
2600 // Expect 1, 1, 3, 3
2601 for (unsigned i = 0; i < 2; ++i) {
2602 int Elt = N->getMaskElt(i);
2603 if (Elt >= 0 && Elt != 1)
2608 for (unsigned i = 2; i < 4; ++i) {
2609 int Elt = N->getMaskElt(i);
2610 if (Elt >= 0 && Elt != 3)
2615 // Don't use movshdup if it can be done with a shufps.
2616 // FIXME: verify that matching u, u, 3, 3 is what we want.
2620 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2621 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2622 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2623 if (N->getValueType(0).getVectorNumElements() != 4)
2626 // Expect 0, 0, 2, 2
2627 for (unsigned i = 0; i < 2; ++i)
2628 if (N->getMaskElt(i) > 0)
2632 for (unsigned i = 2; i < 4; ++i) {
2633 int Elt = N->getMaskElt(i);
2634 if (Elt >= 0 && Elt != 2)
2639 // Don't use movsldup if it can be done with a shufps.
2643 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2644 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2645 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2646 int e = N->getValueType(0).getVectorNumElements() / 2;
2648 for (int i = 0; i < e; ++i)
2649 if (!isUndefOrEqual(N->getMaskElt(i), i))
2651 for (int i = 0; i < e; ++i)
2652 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2657 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2658 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2660 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2661 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2662 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2664 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2666 for (int i = 0; i < NumOperands; ++i) {
2667 int Val = SVOp->getMaskElt(NumOperands-i-1);
2668 if (Val < 0) Val = 0;
2669 if (Val >= NumOperands) Val -= NumOperands;
2671 if (i != NumOperands - 1)
2677 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2678 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2680 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2681 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2683 // 8 nodes, but we only care about the last 4.
2684 for (unsigned i = 7; i >= 4; --i) {
2685 int Val = SVOp->getMaskElt(i);
2694 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2695 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2697 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2698 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2700 // 8 nodes, but we only care about the first 4.
2701 for (int i = 3; i >= 0; --i) {
2702 int Val = SVOp->getMaskElt(i);
2711 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2712 /// their permute mask.
2713 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2714 SelectionDAG &DAG) {
2715 MVT VT = SVOp->getValueType(0);
2716 unsigned NumElems = VT.getVectorNumElements();
2717 SmallVector<int, 8> MaskVec;
2719 for (unsigned i = 0; i != NumElems; ++i) {
2720 int idx = SVOp->getMaskElt(i);
2722 MaskVec.push_back(idx);
2723 else if (idx < (int)NumElems)
2724 MaskVec.push_back(idx + NumElems);
2726 MaskVec.push_back(idx - NumElems);
2728 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2729 SVOp->getOperand(0), &MaskVec[0]);
2732 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2733 /// the two vector operands have swapped position.
2734 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2735 unsigned NumElems = VT.getVectorNumElements();
2736 for (unsigned i = 0; i != NumElems; ++i) {
2740 else if (idx < (int)NumElems)
2741 Mask[i] = idx + NumElems;
2743 Mask[i] = idx - NumElems;
2747 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2748 /// match movhlps. The lower half elements should come from upper half of
2749 /// V1 (and in order), and the upper half elements should come from the upper
2750 /// half of V2 (and in order).
2751 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2752 if (Op->getValueType(0).getVectorNumElements() != 4)
2754 for (unsigned i = 0, e = 2; i != e; ++i)
2755 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2757 for (unsigned i = 2; i != 4; ++i)
2758 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2763 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2764 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2766 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2767 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2769 N = N->getOperand(0).getNode();
2770 if (!ISD::isNON_EXTLoad(N))
2773 *LD = cast<LoadSDNode>(N);
2777 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2778 /// match movlp{s|d}. The lower half elements should come from lower half of
2779 /// V1 (and in order), and the upper half elements should come from the upper
2780 /// half of V2 (and in order). And since V1 will become the source of the
2781 /// MOVLP, it must be either a vector load or a scalar load to vector.
2782 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2783 ShuffleVectorSDNode *Op) {
2784 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2786 // Is V2 is a vector load, don't do this transformation. We will try to use
2787 // load folding shufps op.
2788 if (ISD::isNON_EXTLoad(V2))
2791 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2793 if (NumElems != 2 && NumElems != 4)
2795 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2796 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2798 for (unsigned i = NumElems/2; i != NumElems; ++i)
2799 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2804 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2806 static bool isSplatVector(SDNode *N) {
2807 if (N->getOpcode() != ISD::BUILD_VECTOR)
2810 SDValue SplatValue = N->getOperand(0);
2811 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2812 if (N->getOperand(i) != SplatValue)
2817 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2819 static inline bool isZeroNode(SDValue Elt) {
2820 return ((isa<ConstantSDNode>(Elt) &&
2821 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2822 (isa<ConstantFPSDNode>(Elt) &&
2823 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2826 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2827 /// to an zero vector.
2828 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2829 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2830 SDValue V1 = N->getOperand(0);
2831 SDValue V2 = N->getOperand(1);
2832 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2833 for (unsigned i = 0; i != NumElems; ++i) {
2834 int Idx = N->getMaskElt(i);
2835 if (Idx >= (int)NumElems) {
2836 unsigned Opc = V2.getOpcode();
2837 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2839 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2841 } else if (Idx >= 0) {
2842 unsigned Opc = V1.getOpcode();
2843 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2845 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2852 /// getZeroVector - Returns a vector of specified type with all zero elements.
2854 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2856 assert(VT.isVector() && "Expected a vector type");
2858 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2859 // type. This ensures they get CSE'd.
2861 if (VT.getSizeInBits() == 64) { // MMX
2862 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2863 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2864 } else if (HasSSE2) { // SSE2
2865 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2866 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2868 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2869 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2871 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2874 /// getOnesVector - Returns a vector of specified type with all bits set.
2876 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2877 assert(VT.isVector() && "Expected a vector type");
2879 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2880 // type. This ensures they get CSE'd.
2881 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2883 if (VT.getSizeInBits() == 64) // MMX
2884 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2886 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2887 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2891 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2892 /// that point to V2 points to its first element.
2893 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2894 MVT VT = SVOp->getValueType(0);
2895 unsigned NumElems = VT.getVectorNumElements();
2897 bool Changed = false;
2898 SmallVector<int, 8> MaskVec;
2899 SVOp->getMask(MaskVec);
2901 for (unsigned i = 0; i != NumElems; ++i) {
2902 if (MaskVec[i] > (int)NumElems) {
2903 MaskVec[i] = NumElems;
2908 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2909 SVOp->getOperand(1), &MaskVec[0]);
2910 return SDValue(SVOp, 0);
2913 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2914 /// operation of specified width.
2915 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2917 unsigned NumElems = VT.getVectorNumElements();
2918 SmallVector<int, 8> Mask;
2919 Mask.push_back(NumElems);
2920 for (unsigned i = 1; i != NumElems; ++i)
2922 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2925 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2926 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2928 unsigned NumElems = VT.getVectorNumElements();
2929 SmallVector<int, 8> Mask;
2930 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2932 Mask.push_back(i + NumElems);
2934 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2937 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2938 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2940 unsigned NumElems = VT.getVectorNumElements();
2941 unsigned Half = NumElems/2;
2942 SmallVector<int, 8> Mask;
2943 for (unsigned i = 0; i != Half; ++i) {
2944 Mask.push_back(i + Half);
2945 Mask.push_back(i + NumElems + Half);
2947 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2950 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2951 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2953 if (SV->getValueType(0).getVectorNumElements() <= 4)
2954 return SDValue(SV, 0);
2956 MVT PVT = MVT::v4f32;
2957 MVT VT = SV->getValueType(0);
2958 DebugLoc dl = SV->getDebugLoc();
2959 SDValue V1 = SV->getOperand(0);
2960 int NumElems = VT.getVectorNumElements();
2961 int EltNo = SV->getSplatIndex();
2963 // unpack elements to the correct location
2964 while (NumElems > 4) {
2965 if (EltNo < NumElems/2) {
2966 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2968 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2969 EltNo -= NumElems/2;
2974 // Perform the splat.
2975 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2976 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2977 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2978 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2981 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2982 /// vector of zero or undef vector. This produces a shuffle where the low
2983 /// element of V2 is swizzled into the zero/undef vector, landing at element
2984 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2985 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2986 bool isZero, bool HasSSE2,
2987 SelectionDAG &DAG) {
2988 MVT VT = V2.getValueType();
2990 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
2991 unsigned NumElems = VT.getVectorNumElements();
2992 SmallVector<int, 16> MaskVec;
2993 for (unsigned i = 0; i != NumElems; ++i)
2994 // If this is the insertion idx, put the low elt of V2 here.
2995 MaskVec.push_back(i == Idx ? NumElems : i);
2996 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
2999 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3000 /// a shuffle that is zero.
3002 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3003 bool Low, SelectionDAG &DAG) {
3004 unsigned NumZeros = 0;
3005 for (int i = 0; i < NumElems; ++i) {
3006 unsigned Index = Low ? i : NumElems-i-1;
3007 int Idx = SVOp->getMaskElt(Index);
3012 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3013 if (Elt.getNode() && isZeroNode(Elt))
3021 /// isVectorShift - Returns true if the shuffle can be implemented as a
3022 /// logical left or right shift of a vector.
3023 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3024 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3025 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3026 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3029 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3032 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3036 bool SeenV1 = false;
3037 bool SeenV2 = false;
3038 for (int i = NumZeros; i < NumElems; ++i) {
3039 int Val = isLeft ? (i - NumZeros) : i;
3040 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3052 if (SeenV1 && SeenV2)
3055 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3061 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3063 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3064 unsigned NumNonZero, unsigned NumZero,
3065 SelectionDAG &DAG, TargetLowering &TLI) {
3069 DebugLoc dl = Op.getDebugLoc();
3072 for (unsigned i = 0; i < 16; ++i) {
3073 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3074 if (ThisIsNonZero && First) {
3076 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3078 V = DAG.getUNDEF(MVT::v8i16);
3083 SDValue ThisElt(0, 0), LastElt(0, 0);
3084 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3085 if (LastIsNonZero) {
3086 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3087 MVT::i16, Op.getOperand(i-1));
3089 if (ThisIsNonZero) {
3090 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3091 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3092 ThisElt, DAG.getConstant(8, MVT::i8));
3094 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3098 if (ThisElt.getNode())
3099 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3100 DAG.getIntPtrConstant(i/2));
3104 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3107 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3109 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3110 unsigned NumNonZero, unsigned NumZero,
3111 SelectionDAG &DAG, TargetLowering &TLI) {
3115 DebugLoc dl = Op.getDebugLoc();
3118 for (unsigned i = 0; i < 8; ++i) {
3119 bool isNonZero = (NonZeros & (1 << i)) != 0;
3123 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3125 V = DAG.getUNDEF(MVT::v8i16);
3128 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3129 MVT::v8i16, V, Op.getOperand(i),
3130 DAG.getIntPtrConstant(i));
3137 /// getVShift - Return a vector logical shift node.
3139 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3140 unsigned NumBits, SelectionDAG &DAG,
3141 const TargetLowering &TLI, DebugLoc dl) {
3142 bool isMMX = VT.getSizeInBits() == 64;
3143 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3144 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3145 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3146 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3147 DAG.getNode(Opc, dl, ShVT, SrcOp,
3148 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3152 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3153 DebugLoc dl = Op.getDebugLoc();
3154 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3155 if (ISD::isBuildVectorAllZeros(Op.getNode())
3156 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3157 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3158 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3159 // eliminated on x86-32 hosts.
3160 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3163 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3164 return getOnesVector(Op.getValueType(), DAG, dl);
3165 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3168 MVT VT = Op.getValueType();
3169 MVT EVT = VT.getVectorElementType();
3170 unsigned EVTBits = EVT.getSizeInBits();
3172 unsigned NumElems = Op.getNumOperands();
3173 unsigned NumZero = 0;
3174 unsigned NumNonZero = 0;
3175 unsigned NonZeros = 0;
3176 bool IsAllConstants = true;
3177 SmallSet<SDValue, 8> Values;
3178 for (unsigned i = 0; i < NumElems; ++i) {
3179 SDValue Elt = Op.getOperand(i);
3180 if (Elt.getOpcode() == ISD::UNDEF)
3183 if (Elt.getOpcode() != ISD::Constant &&
3184 Elt.getOpcode() != ISD::ConstantFP)
3185 IsAllConstants = false;
3186 if (isZeroNode(Elt))
3189 NonZeros |= (1 << i);
3194 if (NumNonZero == 0) {
3195 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3196 return DAG.getUNDEF(VT);
3199 // Special case for single non-zero, non-undef, element.
3200 if (NumNonZero == 1) {
3201 unsigned Idx = CountTrailingZeros_32(NonZeros);
3202 SDValue Item = Op.getOperand(Idx);
3204 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3205 // the value are obviously zero, truncate the value to i32 and do the
3206 // insertion that way. Only do this if the value is non-constant or if the
3207 // value is a constant being inserted into element 0. It is cheaper to do
3208 // a constant pool load than it is to do a movd + shuffle.
3209 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3210 (!IsAllConstants || Idx == 0)) {
3211 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3212 // Handle MMX and SSE both.
3213 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3214 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3216 // Truncate the value (which may itself be a constant) to i32, and
3217 // convert it to a vector with movd (S2V+shuffle to zero extend).
3218 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3219 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3220 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3221 Subtarget->hasSSE2(), DAG);
3223 // Now we have our 32-bit value zero extended in the low element of
3224 // a vector. If Idx != 0, swizzle it into place.
3226 SmallVector<int, 4> Mask;
3227 Mask.push_back(Idx);
3228 for (unsigned i = 1; i != VecElts; ++i)
3230 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3231 DAG.getUNDEF(Item.getValueType()),
3234 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3238 // If we have a constant or non-constant insertion into the low element of
3239 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3240 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3241 // depending on what the source datatype is.
3244 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3245 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3246 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3247 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3248 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3249 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3251 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3252 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3253 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3254 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3255 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3256 Subtarget->hasSSE2(), DAG);
3257 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3261 // Is it a vector logical left shift?
3262 if (NumElems == 2 && Idx == 1 &&
3263 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3264 unsigned NumBits = VT.getSizeInBits();
3265 return getVShift(true, VT,
3266 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3267 VT, Op.getOperand(1)),
3268 NumBits/2, DAG, *this, dl);
3271 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3274 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3275 // is a non-constant being inserted into an element other than the low one,
3276 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3277 // movd/movss) to move this into the low element, then shuffle it into
3279 if (EVTBits == 32) {
3280 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3282 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3283 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3284 Subtarget->hasSSE2(), DAG);
3285 SmallVector<int, 8> MaskVec;
3286 for (unsigned i = 0; i < NumElems; i++)
3287 MaskVec.push_back(i == Idx ? 0 : 1);
3288 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3292 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3293 if (Values.size() == 1)
3296 // A vector full of immediates; various special cases are already
3297 // handled, so this is best done with a single constant-pool load.
3301 // Let legalizer expand 2-wide build_vectors.
3302 if (EVTBits == 64) {
3303 if (NumNonZero == 1) {
3304 // One half is zero or undef.
3305 unsigned Idx = CountTrailingZeros_32(NonZeros);
3306 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3307 Op.getOperand(Idx));
3308 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3309 Subtarget->hasSSE2(), DAG);
3314 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3315 if (EVTBits == 8 && NumElems == 16) {
3316 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3318 if (V.getNode()) return V;
3321 if (EVTBits == 16 && NumElems == 8) {
3322 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3324 if (V.getNode()) return V;
3327 // If element VT is == 32 bits, turn it into a number of shuffles.
3328 SmallVector<SDValue, 8> V;
3330 if (NumElems == 4 && NumZero > 0) {
3331 for (unsigned i = 0; i < 4; ++i) {
3332 bool isZero = !(NonZeros & (1 << i));
3334 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3336 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3339 for (unsigned i = 0; i < 2; ++i) {
3340 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3343 V[i] = V[i*2]; // Must be a zero vector.
3346 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3349 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3352 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3357 SmallVector<int, 8> MaskVec;
3358 bool Reverse = (NonZeros & 0x3) == 2;
3359 for (unsigned i = 0; i < 2; ++i)
3360 MaskVec.push_back(Reverse ? 1-i : i);
3361 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3362 for (unsigned i = 0; i < 2; ++i)
3363 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3364 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3367 if (Values.size() > 2) {
3368 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3369 // values to be inserted is equal to the number of elements, in which case
3370 // use the unpack code below in the hopes of matching the consecutive elts
3371 // load merge pattern for shuffles.
3372 // FIXME: We could probably just check that here directly.
3373 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3374 getSubtarget()->hasSSE41()) {
3375 V[0] = DAG.getUNDEF(VT);
3376 for (unsigned i = 0; i < NumElems; ++i)
3377 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3378 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3379 Op.getOperand(i), DAG.getIntPtrConstant(i));
3382 // Expand into a number of unpckl*.
3384 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3385 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3386 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3387 for (unsigned i = 0; i < NumElems; ++i)
3388 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3390 while (NumElems != 0) {
3391 for (unsigned i = 0; i < NumElems; ++i)
3392 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3401 // v8i16 shuffles - Prefer shuffles in the following order:
3402 // 1. [all] pshuflw, pshufhw, optional move
3403 // 2. [ssse3] 1 x pshufb
3404 // 3. [ssse3] 2 x pshufb + 1 x por
3405 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3407 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3408 SelectionDAG &DAG, X86TargetLowering &TLI) {
3409 SDValue V1 = SVOp->getOperand(0);
3410 SDValue V2 = SVOp->getOperand(1);
3411 DebugLoc dl = SVOp->getDebugLoc();
3412 SmallVector<int, 8> MaskVals;
3414 // Determine if more than 1 of the words in each of the low and high quadwords
3415 // of the result come from the same quadword of one of the two inputs. Undef
3416 // mask values count as coming from any quadword, for better codegen.
3417 SmallVector<unsigned, 4> LoQuad(4);
3418 SmallVector<unsigned, 4> HiQuad(4);
3419 BitVector InputQuads(4);
3420 for (unsigned i = 0; i < 8; ++i) {
3421 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3422 int EltIdx = SVOp->getMaskElt(i);
3423 MaskVals.push_back(EltIdx);
3432 InputQuads.set(EltIdx / 4);
3435 int BestLoQuad = -1;
3436 unsigned MaxQuad = 1;
3437 for (unsigned i = 0; i < 4; ++i) {
3438 if (LoQuad[i] > MaxQuad) {
3440 MaxQuad = LoQuad[i];
3444 int BestHiQuad = -1;
3446 for (unsigned i = 0; i < 4; ++i) {
3447 if (HiQuad[i] > MaxQuad) {
3449 MaxQuad = HiQuad[i];
3453 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3454 // of the two input vectors, shuffle them into one input vector so only a
3455 // single pshufb instruction is necessary. If There are more than 2 input
3456 // quads, disable the next transformation since it does not help SSSE3.
3457 bool V1Used = InputQuads[0] || InputQuads[1];
3458 bool V2Used = InputQuads[2] || InputQuads[3];
3459 if (TLI.getSubtarget()->hasSSSE3()) {
3460 if (InputQuads.count() == 2 && V1Used && V2Used) {
3461 BestLoQuad = InputQuads.find_first();
3462 BestHiQuad = InputQuads.find_next(BestLoQuad);
3464 if (InputQuads.count() > 2) {
3470 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3471 // the shuffle mask. If a quad is scored as -1, that means that it contains
3472 // words from all 4 input quadwords.
3474 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3475 SmallVector<int, 8> MaskV;
3476 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3477 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3478 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3479 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3480 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3481 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3483 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3484 // source words for the shuffle, to aid later transformations.
3485 bool AllWordsInNewV = true;
3486 bool InOrder[2] = { true, true };
3487 for (unsigned i = 0; i != 8; ++i) {
3488 int idx = MaskVals[i];
3490 InOrder[i/4] = false;
3491 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3493 AllWordsInNewV = false;
3497 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3498 if (AllWordsInNewV) {
3499 for (int i = 0; i != 8; ++i) {
3500 int idx = MaskVals[i];
3503 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3504 if ((idx != i) && idx < 4)
3506 if ((idx != i) && idx > 3)
3515 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3516 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3517 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3518 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3519 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3523 // If we have SSSE3, and all words of the result are from 1 input vector,
3524 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3525 // is present, fall back to case 4.
3526 if (TLI.getSubtarget()->hasSSSE3()) {
3527 SmallVector<SDValue,16> pshufbMask;
3529 // If we have elements from both input vectors, set the high bit of the
3530 // shuffle mask element to zero out elements that come from V2 in the V1
3531 // mask, and elements that come from V1 in the V2 mask, so that the two
3532 // results can be OR'd together.
3533 bool TwoInputs = V1Used && V2Used;
3534 for (unsigned i = 0; i != 8; ++i) {
3535 int EltIdx = MaskVals[i] * 2;
3536 if (TwoInputs && (EltIdx >= 16)) {
3537 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3538 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3541 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3542 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3544 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3545 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3546 DAG.getNode(ISD::BUILD_VECTOR, dl,
3547 MVT::v16i8, &pshufbMask[0], 16));
3549 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3551 // Calculate the shuffle mask for the second input, shuffle it, and
3552 // OR it with the first shuffled input.
3554 for (unsigned i = 0; i != 8; ++i) {
3555 int EltIdx = MaskVals[i] * 2;
3557 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3558 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3561 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3562 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3564 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3565 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3566 DAG.getNode(ISD::BUILD_VECTOR, dl,
3567 MVT::v16i8, &pshufbMask[0], 16));
3568 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3569 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3572 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3573 // and update MaskVals with new element order.
3574 BitVector InOrder(8);
3575 if (BestLoQuad >= 0) {
3576 SmallVector<int, 8> MaskV;
3577 for (int i = 0; i != 4; ++i) {
3578 int idx = MaskVals[i];
3580 MaskV.push_back(-1);
3582 } else if ((idx / 4) == BestLoQuad) {
3583 MaskV.push_back(idx & 3);
3586 MaskV.push_back(-1);
3589 for (unsigned i = 4; i != 8; ++i)
3591 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3595 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3596 // and update MaskVals with the new element order.
3597 if (BestHiQuad >= 0) {
3598 SmallVector<int, 8> MaskV;
3599 for (unsigned i = 0; i != 4; ++i)
3601 for (unsigned i = 4; i != 8; ++i) {
3602 int idx = MaskVals[i];
3604 MaskV.push_back(-1);
3606 } else if ((idx / 4) == BestHiQuad) {
3607 MaskV.push_back((idx & 3) + 4);
3610 MaskV.push_back(-1);
3613 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3617 // In case BestHi & BestLo were both -1, which means each quadword has a word
3618 // from each of the four input quadwords, calculate the InOrder bitvector now
3619 // before falling through to the insert/extract cleanup.
3620 if (BestLoQuad == -1 && BestHiQuad == -1) {
3622 for (int i = 0; i != 8; ++i)
3623 if (MaskVals[i] < 0 || MaskVals[i] == i)
3627 // The other elements are put in the right place using pextrw and pinsrw.
3628 for (unsigned i = 0; i != 8; ++i) {
3631 int EltIdx = MaskVals[i];
3634 SDValue ExtOp = (EltIdx < 8)
3635 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3636 DAG.getIntPtrConstant(EltIdx))
3637 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3638 DAG.getIntPtrConstant(EltIdx - 8));
3639 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3640 DAG.getIntPtrConstant(i));
3645 // v16i8 shuffles - Prefer shuffles in the following order:
3646 // 1. [ssse3] 1 x pshufb
3647 // 2. [ssse3] 2 x pshufb + 1 x por
3648 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3650 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3651 SelectionDAG &DAG, X86TargetLowering &TLI) {
3652 SDValue V1 = SVOp->getOperand(0);
3653 SDValue V2 = SVOp->getOperand(1);
3654 DebugLoc dl = SVOp->getDebugLoc();
3655 SmallVector<int, 16> MaskVals;
3656 SVOp->getMask(MaskVals);
3658 // If we have SSSE3, case 1 is generated when all result bytes come from
3659 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3660 // present, fall back to case 3.
3661 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3664 for (unsigned i = 0; i < 16; ++i) {
3665 int EltIdx = MaskVals[i];
3674 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3675 if (TLI.getSubtarget()->hasSSSE3()) {
3676 SmallVector<SDValue,16> pshufbMask;
3678 // If all result elements are from one input vector, then only translate
3679 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3681 // Otherwise, we have elements from both input vectors, and must zero out
3682 // elements that come from V2 in the first mask, and V1 in the second mask
3683 // so that we can OR them together.
3684 bool TwoInputs = !(V1Only || V2Only);
3685 for (unsigned i = 0; i != 16; ++i) {
3686 int EltIdx = MaskVals[i];
3687 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3688 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3691 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3693 // If all the elements are from V2, assign it to V1 and return after
3694 // building the first pshufb.
3697 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3698 DAG.getNode(ISD::BUILD_VECTOR, dl,
3699 MVT::v16i8, &pshufbMask[0], 16));
3703 // Calculate the shuffle mask for the second input, shuffle it, and
3704 // OR it with the first shuffled input.
3706 for (unsigned i = 0; i != 16; ++i) {
3707 int EltIdx = MaskVals[i];
3709 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3712 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3714 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3715 DAG.getNode(ISD::BUILD_VECTOR, dl,
3716 MVT::v16i8, &pshufbMask[0], 16));
3717 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3720 // No SSSE3 - Calculate in place words and then fix all out of place words
3721 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3722 // the 16 different words that comprise the two doublequadword input vectors.
3723 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3724 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3725 SDValue NewV = V2Only ? V2 : V1;
3726 for (int i = 0; i != 8; ++i) {
3727 int Elt0 = MaskVals[i*2];
3728 int Elt1 = MaskVals[i*2+1];
3730 // This word of the result is all undef, skip it.
3731 if (Elt0 < 0 && Elt1 < 0)
3734 // This word of the result is already in the correct place, skip it.
3735 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3737 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3740 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3741 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3744 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3745 // using a single extract together, load it and store it.
3746 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3747 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3748 DAG.getIntPtrConstant(Elt1 / 2));
3749 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3750 DAG.getIntPtrConstant(i));
3754 // If Elt1 is defined, extract it from the appropriate source. If the
3755 // source byte is not also odd, shift the extracted word left 8 bits
3756 // otherwise clear the bottom 8 bits if we need to do an or.
3758 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3759 DAG.getIntPtrConstant(Elt1 / 2));
3760 if ((Elt1 & 1) == 0)
3761 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3762 DAG.getConstant(8, TLI.getShiftAmountTy()));
3764 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3765 DAG.getConstant(0xFF00, MVT::i16));
3767 // If Elt0 is defined, extract it from the appropriate source. If the
3768 // source byte is not also even, shift the extracted word right 8 bits. If
3769 // Elt1 was also defined, OR the extracted values together before
3770 // inserting them in the result.
3772 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3773 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3774 if ((Elt0 & 1) != 0)
3775 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3776 DAG.getConstant(8, TLI.getShiftAmountTy()));
3778 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3779 DAG.getConstant(0x00FF, MVT::i16));
3780 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3783 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3784 DAG.getIntPtrConstant(i));
3786 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3789 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3790 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3791 /// done when every pair / quad of shuffle mask elements point to elements in
3792 /// the right sequence. e.g.
3793 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3795 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3797 TargetLowering &TLI, DebugLoc dl) {
3798 MVT VT = SVOp->getValueType(0);
3799 SDValue V1 = SVOp->getOperand(0);
3800 SDValue V2 = SVOp->getOperand(1);
3801 unsigned NumElems = VT.getVectorNumElements();
3802 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3803 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3804 MVT MaskEltVT = MaskVT.getVectorElementType();
3806 switch (VT.getSimpleVT()) {
3807 default: assert(false && "Unexpected!");
3808 case MVT::v4f32: NewVT = MVT::v2f64; break;
3809 case MVT::v4i32: NewVT = MVT::v2i64; break;
3810 case MVT::v8i16: NewVT = MVT::v4i32; break;
3811 case MVT::v16i8: NewVT = MVT::v4i32; break;
3814 if (NewWidth == 2) {
3820 int Scale = NumElems / NewWidth;
3821 SmallVector<int, 8> MaskVec;
3822 for (unsigned i = 0; i < NumElems; i += Scale) {
3824 for (int j = 0; j < Scale; ++j) {
3825 int EltIdx = SVOp->getMaskElt(i+j);
3829 StartIdx = EltIdx - (EltIdx % Scale);
3830 if (EltIdx != StartIdx + j)
3834 MaskVec.push_back(-1);
3836 MaskVec.push_back(StartIdx / Scale);
3839 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3840 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3841 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3844 /// getVZextMovL - Return a zero-extending vector move low node.
3846 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3847 SDValue SrcOp, SelectionDAG &DAG,
3848 const X86Subtarget *Subtarget, DebugLoc dl) {
3849 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3850 LoadSDNode *LD = NULL;
3851 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3852 LD = dyn_cast<LoadSDNode>(SrcOp);
3854 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3856 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3857 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3858 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3859 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3860 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3862 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3863 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3864 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3865 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3873 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3874 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3875 DAG.getNode(ISD::BIT_CONVERT, dl,
3879 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3882 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3883 SDValue V1 = SVOp->getOperand(0);
3884 SDValue V2 = SVOp->getOperand(1);
3885 DebugLoc dl = SVOp->getDebugLoc();
3886 MVT VT = SVOp->getValueType(0);
3888 SmallVector<std::pair<int, int>, 8> Locs;
3890 SmallVector<int, 8> Mask1(4U, -1);
3891 SmallVector<int, 8> PermMask;
3892 SVOp->getMask(PermMask);
3896 for (unsigned i = 0; i != 4; ++i) {
3897 int Idx = PermMask[i];
3899 Locs[i] = std::make_pair(-1, -1);
3901 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3903 Locs[i] = std::make_pair(0, NumLo);
3907 Locs[i] = std::make_pair(1, NumHi);
3909 Mask1[2+NumHi] = Idx;
3915 if (NumLo <= 2 && NumHi <= 2) {
3916 // If no more than two elements come from either vector. This can be
3917 // implemented with two shuffles. First shuffle gather the elements.
3918 // The second shuffle, which takes the first shuffle as both of its
3919 // vector operands, put the elements into the right order.
3920 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3922 SmallVector<int, 8> Mask2(4U, -1);
3924 for (unsigned i = 0; i != 4; ++i) {
3925 if (Locs[i].first == -1)
3928 unsigned Idx = (i < 2) ? 0 : 4;
3929 Idx += Locs[i].first * 2 + Locs[i].second;
3934 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3935 } else if (NumLo == 3 || NumHi == 3) {
3936 // Otherwise, we must have three elements from one vector, call it X, and
3937 // one element from the other, call it Y. First, use a shufps to build an
3938 // intermediate vector with the one element from Y and the element from X
3939 // that will be in the same half in the final destination (the indexes don't
3940 // matter). Then, use a shufps to build the final vector, taking the half
3941 // containing the element from Y from the intermediate, and the other half
3944 // Normalize it so the 3 elements come from V1.
3945 CommuteVectorShuffleMask(PermMask, VT);
3949 // Find the element from V2.
3951 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3952 int Val = PermMask[HiIndex];
3959 Mask1[0] = PermMask[HiIndex];
3961 Mask1[2] = PermMask[HiIndex^1];
3963 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3966 Mask1[0] = PermMask[0];
3967 Mask1[1] = PermMask[1];
3968 Mask1[2] = HiIndex & 1 ? 6 : 4;
3969 Mask1[3] = HiIndex & 1 ? 4 : 6;
3970 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3972 Mask1[0] = HiIndex & 1 ? 2 : 0;
3973 Mask1[1] = HiIndex & 1 ? 0 : 2;
3974 Mask1[2] = PermMask[2];
3975 Mask1[3] = PermMask[3];
3980 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3984 // Break it into (shuffle shuffle_hi, shuffle_lo).
3986 SmallVector<int,8> LoMask(4U, -1);
3987 SmallVector<int,8> HiMask(4U, -1);
3989 SmallVector<int,8> *MaskPtr = &LoMask;
3990 unsigned MaskIdx = 0;
3993 for (unsigned i = 0; i != 4; ++i) {
4000 int Idx = PermMask[i];
4002 Locs[i] = std::make_pair(-1, -1);
4003 } else if (Idx < 4) {
4004 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4005 (*MaskPtr)[LoIdx] = Idx;
4008 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4009 (*MaskPtr)[HiIdx] = Idx;
4014 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4015 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4016 SmallVector<int, 8> MaskOps;
4017 for (unsigned i = 0; i != 4; ++i) {
4018 if (Locs[i].first == -1) {
4019 MaskOps.push_back(-1);
4021 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4022 MaskOps.push_back(Idx);
4025 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4029 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4030 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4031 SDValue V1 = Op.getOperand(0);
4032 SDValue V2 = Op.getOperand(1);
4033 MVT VT = Op.getValueType();
4034 DebugLoc dl = Op.getDebugLoc();
4035 unsigned NumElems = VT.getVectorNumElements();
4036 bool isMMX = VT.getSizeInBits() == 64;
4037 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4038 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4039 bool V1IsSplat = false;
4040 bool V2IsSplat = false;
4042 if (isZeroShuffle(SVOp))
4043 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4045 // Promote splats to v4f32.
4046 if (SVOp->isSplat()) {
4047 if (isMMX || NumElems < 4)
4049 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4052 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4054 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4055 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4056 if (NewOp.getNode())
4057 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4058 LowerVECTOR_SHUFFLE(NewOp, DAG));
4059 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4060 // FIXME: Figure out a cleaner way to do this.
4061 // Try to make use of movq to zero out the top part.
4062 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4063 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4064 if (NewOp.getNode()) {
4065 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4066 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4067 DAG, Subtarget, dl);
4069 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4070 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4071 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4072 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4073 DAG, Subtarget, dl);
4077 if (X86::isPSHUFDMask(SVOp))
4080 // Check if this can be converted into a logical shift.
4081 bool isLeft = false;
4084 bool isShift = getSubtarget()->hasSSE2() &&
4085 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4086 if (isShift && ShVal.hasOneUse()) {
4087 // If the shifted value has multiple uses, it may be cheaper to use
4088 // v_set0 + movlhps or movhlps, etc.
4089 MVT EVT = VT.getVectorElementType();
4090 ShAmt *= EVT.getSizeInBits();
4091 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4094 if (X86::isMOVLMask(SVOp)) {
4097 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4098 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4103 // FIXME: fold these into legal mask.
4104 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4105 X86::isMOVSLDUPMask(SVOp) ||
4106 X86::isMOVHLPSMask(SVOp) ||
4107 X86::isMOVHPMask(SVOp) ||
4108 X86::isMOVLPMask(SVOp)))
4111 if (ShouldXformToMOVHLPS(SVOp) ||
4112 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4113 return CommuteVectorShuffle(SVOp, DAG);
4116 // No better options. Use a vshl / vsrl.
4117 MVT EVT = VT.getVectorElementType();
4118 ShAmt *= EVT.getSizeInBits();
4119 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4122 bool Commuted = false;
4123 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4124 // 1,1,1,1 -> v8i16 though.
4125 V1IsSplat = isSplatVector(V1.getNode());
4126 V2IsSplat = isSplatVector(V2.getNode());
4128 // Canonicalize the splat or undef, if present, to be on the RHS.
4129 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4130 Op = CommuteVectorShuffle(SVOp, DAG);
4131 SVOp = cast<ShuffleVectorSDNode>(Op);
4132 V1 = SVOp->getOperand(0);
4133 V2 = SVOp->getOperand(1);
4134 std::swap(V1IsSplat, V2IsSplat);
4135 std::swap(V1IsUndef, V2IsUndef);
4139 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4140 // Shuffling low element of v1 into undef, just return v1.
4143 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4144 // the instruction selector will not match, so get a canonical MOVL with
4145 // swapped operands to undo the commute.
4146 return getMOVL(DAG, dl, VT, V2, V1);
4149 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4150 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4151 X86::isUNPCKLMask(SVOp) ||
4152 X86::isUNPCKHMask(SVOp))
4156 // Normalize mask so all entries that point to V2 points to its first
4157 // element then try to match unpck{h|l} again. If match, return a
4158 // new vector_shuffle with the corrected mask.
4159 SDValue NewMask = NormalizeMask(SVOp, DAG);
4160 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4161 if (NSVOp != SVOp) {
4162 if (X86::isUNPCKLMask(NSVOp, true)) {
4164 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4171 // Commute is back and try unpck* again.
4172 // FIXME: this seems wrong.
4173 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4174 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4175 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4176 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4177 X86::isUNPCKLMask(NewSVOp) ||
4178 X86::isUNPCKHMask(NewSVOp))
4182 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4184 // Normalize the node to match x86 shuffle ops if needed
4185 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4186 return CommuteVectorShuffle(SVOp, DAG);
4188 // Check for legal shuffle and return?
4189 SmallVector<int, 16> PermMask;
4190 SVOp->getMask(PermMask);
4191 if (isShuffleMaskLegal(PermMask, VT))
4194 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4195 if (VT == MVT::v8i16) {
4196 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4197 if (NewOp.getNode())
4201 if (VT == MVT::v16i8) {
4202 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4203 if (NewOp.getNode())
4207 // Handle all 4 wide cases with a number of shuffles except for MMX.
4208 if (NumElems == 4 && !isMMX)
4209 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4215 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4216 SelectionDAG &DAG) {
4217 MVT VT = Op.getValueType();
4218 DebugLoc dl = Op.getDebugLoc();
4219 if (VT.getSizeInBits() == 8) {
4220 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4221 Op.getOperand(0), Op.getOperand(1));
4222 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4223 DAG.getValueType(VT));
4224 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4225 } else if (VT.getSizeInBits() == 16) {
4226 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4227 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4229 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4230 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4231 DAG.getNode(ISD::BIT_CONVERT, dl,
4235 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4236 Op.getOperand(0), Op.getOperand(1));
4237 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4238 DAG.getValueType(VT));
4239 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4240 } else if (VT == MVT::f32) {
4241 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4242 // the result back to FR32 register. It's only worth matching if the
4243 // result has a single use which is a store or a bitcast to i32. And in
4244 // the case of a store, it's not worth it if the index is a constant 0,
4245 // because a MOVSSmr can be used instead, which is smaller and faster.
4246 if (!Op.hasOneUse())
4248 SDNode *User = *Op.getNode()->use_begin();
4249 if ((User->getOpcode() != ISD::STORE ||
4250 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4251 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4252 (User->getOpcode() != ISD::BIT_CONVERT ||
4253 User->getValueType(0) != MVT::i32))
4255 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4256 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4259 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4260 } else if (VT == MVT::i32) {
4261 // ExtractPS works with constant index.
4262 if (isa<ConstantSDNode>(Op.getOperand(1)))
4270 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4271 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4274 if (Subtarget->hasSSE41()) {
4275 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4280 MVT VT = Op.getValueType();
4281 DebugLoc dl = Op.getDebugLoc();
4282 // TODO: handle v16i8.
4283 if (VT.getSizeInBits() == 16) {
4284 SDValue Vec = Op.getOperand(0);
4285 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4287 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4288 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4289 DAG.getNode(ISD::BIT_CONVERT, dl,
4292 // Transform it so it match pextrw which produces a 32-bit result.
4293 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4294 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4295 Op.getOperand(0), Op.getOperand(1));
4296 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4297 DAG.getValueType(VT));
4298 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4299 } else if (VT.getSizeInBits() == 32) {
4300 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4304 // SHUFPS the element to the lowest double word, then movss.
4305 int Mask[4] = { Idx, -1, -1, -1 };
4306 MVT VVT = Op.getOperand(0).getValueType();
4307 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4308 DAG.getUNDEF(VVT), Mask);
4309 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4310 DAG.getIntPtrConstant(0));
4311 } else if (VT.getSizeInBits() == 64) {
4312 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4313 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4314 // to match extract_elt for f64.
4315 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4319 // UNPCKHPD the element to the lowest double word, then movsd.
4320 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4321 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4322 int Mask[2] = { 1, -1 };
4323 MVT VVT = Op.getOperand(0).getValueType();
4324 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4325 DAG.getUNDEF(VVT), Mask);
4326 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4327 DAG.getIntPtrConstant(0));
4334 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4335 MVT VT = Op.getValueType();
4336 MVT EVT = VT.getVectorElementType();
4337 DebugLoc dl = Op.getDebugLoc();
4339 SDValue N0 = Op.getOperand(0);
4340 SDValue N1 = Op.getOperand(1);
4341 SDValue N2 = Op.getOperand(2);
4343 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4344 isa<ConstantSDNode>(N2)) {
4345 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4347 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4349 if (N1.getValueType() != MVT::i32)
4350 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4351 if (N2.getValueType() != MVT::i32)
4352 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4353 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4354 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4355 // Bits [7:6] of the constant are the source select. This will always be
4356 // zero here. The DAG Combiner may combine an extract_elt index into these
4357 // bits. For example (insert (extract, 3), 2) could be matched by putting
4358 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4359 // Bits [5:4] of the constant are the destination select. This is the
4360 // value of the incoming immediate.
4361 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4362 // combine either bitwise AND or insert of float 0.0 to set these bits.
4363 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4364 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4365 } else if (EVT == MVT::i32) {
4366 // InsertPS works with constant index.
4367 if (isa<ConstantSDNode>(N2))
4374 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4375 MVT VT = Op.getValueType();
4376 MVT EVT = VT.getVectorElementType();
4378 if (Subtarget->hasSSE41())
4379 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4384 DebugLoc dl = Op.getDebugLoc();
4385 SDValue N0 = Op.getOperand(0);
4386 SDValue N1 = Op.getOperand(1);
4387 SDValue N2 = Op.getOperand(2);
4389 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4390 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4391 // as its second argument.
4392 if (N1.getValueType() != MVT::i32)
4393 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4394 if (N2.getValueType() != MVT::i32)
4395 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4396 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4402 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4403 DebugLoc dl = Op.getDebugLoc();
4404 if (Op.getValueType() == MVT::v2f32)
4405 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4406 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4407 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4408 Op.getOperand(0))));
4410 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4411 MVT VT = MVT::v2i32;
4412 switch (Op.getValueType().getSimpleVT()) {
4419 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4420 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4423 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4424 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4425 // one of the above mentioned nodes. It has to be wrapped because otherwise
4426 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4427 // be used to form addressing mode. These wrapped nodes will be selected
4430 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4431 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4433 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4435 unsigned char OpFlag = 0;
4436 unsigned WrapperKind = X86ISD::Wrapper;
4438 if (Subtarget->is64Bit() &&
4439 getTargetMachine().getCodeModel() == CodeModel::Small) {
4440 WrapperKind = X86ISD::WrapperRIP;
4441 } else if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4442 if (Subtarget->isPICStyleStub())
4443 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4444 else if (Subtarget->isPICStyleGOT())
4445 OpFlag = X86II::MO_GOTOFF;
4448 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4450 CP->getOffset(), OpFlag);
4451 DebugLoc DL = CP->getDebugLoc();
4452 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4453 // With PIC, the address is actually $g + Offset.
4455 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4456 DAG.getNode(X86ISD::GlobalBaseReg,
4457 DebugLoc::getUnknownLoc(), getPointerTy()),
4464 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4465 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4467 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4469 unsigned char OpFlag = 0;
4470 unsigned WrapperKind = X86ISD::Wrapper;
4472 if (Subtarget->is64Bit()) {
4473 WrapperKind = X86ISD::WrapperRIP;
4474 } else if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4475 if (Subtarget->isPICStyleStub())
4476 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4477 else if (Subtarget->isPICStyleGOT())
4478 OpFlag = X86II::MO_GOTOFF;
4481 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4483 DebugLoc DL = JT->getDebugLoc();
4484 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4486 // With PIC, the address is actually $g + Offset.
4488 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4489 DAG.getNode(X86ISD::GlobalBaseReg,
4490 DebugLoc::getUnknownLoc(), getPointerTy()),
4498 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4499 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4501 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4503 unsigned char OpFlag = 0;
4504 unsigned WrapperKind = X86ISD::Wrapper;
4505 if (Subtarget->is64Bit()) {
4506 WrapperKind = X86ISD::WrapperRIP;
4507 } else if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4508 if (Subtarget->isPICStyleStub())
4509 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4510 else if (Subtarget->isPICStyleGOT())
4511 OpFlag = X86II::MO_GOTOFF;
4514 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4516 DebugLoc DL = Op.getDebugLoc();
4517 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4520 // With PIC, the address is actually $g + Offset.
4521 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4522 !Subtarget->is64Bit()) {
4523 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4524 DAG.getNode(X86ISD::GlobalBaseReg,
4525 DebugLoc::getUnknownLoc(),
4534 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4536 SelectionDAG &DAG) const {
4537 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4538 bool ExtraLoadRequired =
4539 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4541 // Create the TargetGlobalAddress node, folding in the constant
4542 // offset if it is legal.
4544 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4545 // A direct static reference to a global.
4546 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4549 unsigned char OpFlags = 0;
4551 if (GV->hasDLLImportLinkage())
4552 OpFlags = X86II::MO_DLLIMPORT;
4553 else if (Subtarget->isPICStyleRIPRel() &&
4554 getTargetMachine().getRelocationModel() != Reloc::Static) {
4555 if (ExtraLoadRequired)
4556 OpFlags = X86II::MO_GOTPCREL;
4557 } else if (Subtarget->isPICStyleGOT() &&
4558 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4559 if (ExtraLoadRequired)
4560 OpFlags = X86II::MO_GOT;
4562 OpFlags = X86II::MO_GOTOFF;
4565 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4568 if (Subtarget->is64Bit() &&
4569 getTargetMachine().getCodeModel() == CodeModel::Small)
4570 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4572 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4574 // With PIC, the address is actually $g + Offset.
4575 if (IsPic && !Subtarget->is64Bit()) {
4576 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4577 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4581 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4582 // load the value at address GV, not the value of GV itself. This means that
4583 // the GlobalAddress must be in the base or index register of the address, not
4584 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4585 // The same applies for external symbols during PIC codegen
4586 if (ExtraLoadRequired)
4587 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4588 PseudoSourceValue::getGOT(), 0);
4590 // If there was a non-zero offset that we didn't fold, create an explicit
4593 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4594 DAG.getConstant(Offset, getPointerTy()));
4600 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4601 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4602 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4603 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4607 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4608 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4609 unsigned char OperandFlags) {
4610 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4611 DebugLoc dl = GA->getDebugLoc();
4612 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4613 GA->getValueType(0),
4617 SDValue Ops[] = { Chain, TGA, *InFlag };
4618 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4620 SDValue Ops[] = { Chain, TGA };
4621 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4623 SDValue Flag = Chain.getValue(1);
4624 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4627 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4629 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4632 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4633 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4634 DAG.getNode(X86ISD::GlobalBaseReg,
4635 DebugLoc::getUnknownLoc(),
4637 InFlag = Chain.getValue(1);
4639 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4642 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4644 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4646 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4647 X86::RAX, X86II::MO_TLSGD);
4650 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4651 // "local exec" model.
4652 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4653 const MVT PtrVT, TLSModel::Model model,
4655 DebugLoc dl = GA->getDebugLoc();
4656 // Get the Thread Pointer
4657 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4658 DebugLoc::getUnknownLoc(), PtrVT,
4659 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4662 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4665 unsigned char OperandFlags = 0;
4666 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4668 unsigned WrapperKind = X86ISD::Wrapper;
4669 if (model == TLSModel::LocalExec) {
4670 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4671 } else if (is64Bit) {
4672 assert(model == TLSModel::InitialExec);
4673 OperandFlags = X86II::MO_GOTTPOFF;
4674 WrapperKind = X86ISD::WrapperRIP;
4676 assert(model == TLSModel::InitialExec);
4677 OperandFlags = X86II::MO_INDNTPOFF;
4680 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4682 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4683 GA->getOffset(), OperandFlags);
4684 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4686 if (model == TLSModel::InitialExec)
4687 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4688 PseudoSourceValue::getGOT(), 0);
4690 // The address of the thread local variable is the add of the thread
4691 // pointer with the offset of the variable.
4692 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4696 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4697 // TODO: implement the "local dynamic" model
4698 // TODO: implement the "initial exec"model for pic executables
4699 assert(Subtarget->isTargetELF() &&
4700 "TLS not implemented for non-ELF targets");
4701 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4702 const GlobalValue *GV = GA->getGlobal();
4704 // If GV is an alias then use the aliasee for determining
4705 // thread-localness.
4706 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4707 GV = GA->resolveAliasedGlobal(false);
4709 TLSModel::Model model = getTLSModel(GV,
4710 getTargetMachine().getRelocationModel());
4713 case TLSModel::GeneralDynamic:
4714 case TLSModel::LocalDynamic: // not implemented
4715 if (Subtarget->is64Bit())
4716 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4717 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4719 case TLSModel::InitialExec:
4720 case TLSModel::LocalExec:
4721 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4722 Subtarget->is64Bit());
4725 assert(0 && "Unreachable");
4730 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4731 /// take a 2 x i32 value to shift plus a shift amount.
4732 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4733 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4734 MVT VT = Op.getValueType();
4735 unsigned VTBits = VT.getSizeInBits();
4736 DebugLoc dl = Op.getDebugLoc();
4737 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4738 SDValue ShOpLo = Op.getOperand(0);
4739 SDValue ShOpHi = Op.getOperand(1);
4740 SDValue ShAmt = Op.getOperand(2);
4741 SDValue Tmp1 = isSRA ?
4742 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4743 DAG.getConstant(VTBits - 1, MVT::i8)) :
4744 DAG.getConstant(0, VT);
4747 if (Op.getOpcode() == ISD::SHL_PARTS) {
4748 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4749 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4751 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4752 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4755 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4756 DAG.getConstant(VTBits, MVT::i8));
4757 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4758 AndNode, DAG.getConstant(0, MVT::i8));
4761 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4762 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4763 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4765 if (Op.getOpcode() == ISD::SHL_PARTS) {
4766 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4767 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4769 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4770 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4773 SDValue Ops[2] = { Lo, Hi };
4774 return DAG.getMergeValues(Ops, 2, dl);
4777 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4778 MVT SrcVT = Op.getOperand(0).getValueType();
4780 if (SrcVT.isVector()) {
4781 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4787 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4788 "Unknown SINT_TO_FP to lower!");
4790 // These are really Legal; return the operand so the caller accepts it as
4792 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4794 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4795 Subtarget->is64Bit()) {
4799 DebugLoc dl = Op.getDebugLoc();
4800 unsigned Size = SrcVT.getSizeInBits()/8;
4801 MachineFunction &MF = DAG.getMachineFunction();
4802 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4803 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4804 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4806 PseudoSourceValue::getFixedStack(SSFI), 0);
4807 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4810 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4812 SelectionDAG &DAG) {
4814 DebugLoc dl = Op.getDebugLoc();
4816 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4818 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4820 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4821 SmallVector<SDValue, 8> Ops;
4822 Ops.push_back(Chain);
4823 Ops.push_back(StackSlot);
4824 Ops.push_back(DAG.getValueType(SrcVT));
4825 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4826 Tys, &Ops[0], Ops.size());
4829 Chain = Result.getValue(1);
4830 SDValue InFlag = Result.getValue(2);
4832 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4833 // shouldn't be necessary except that RFP cannot be live across
4834 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4835 MachineFunction &MF = DAG.getMachineFunction();
4836 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4837 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4838 Tys = DAG.getVTList(MVT::Other);
4839 SmallVector<SDValue, 8> Ops;
4840 Ops.push_back(Chain);
4841 Ops.push_back(Result);
4842 Ops.push_back(StackSlot);
4843 Ops.push_back(DAG.getValueType(Op.getValueType()));
4844 Ops.push_back(InFlag);
4845 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4846 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4847 PseudoSourceValue::getFixedStack(SSFI), 0);
4853 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4854 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4855 // This algorithm is not obvious. Here it is in C code, more or less:
4857 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4858 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4859 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4861 // Copy ints to xmm registers.
4862 __m128i xh = _mm_cvtsi32_si128( hi );
4863 __m128i xl = _mm_cvtsi32_si128( lo );
4865 // Combine into low half of a single xmm register.
4866 __m128i x = _mm_unpacklo_epi32( xh, xl );
4870 // Merge in appropriate exponents to give the integer bits the right
4872 x = _mm_unpacklo_epi32( x, exp );
4874 // Subtract away the biases to deal with the IEEE-754 double precision
4876 d = _mm_sub_pd( (__m128d) x, bias );
4878 // All conversions up to here are exact. The correctly rounded result is
4879 // calculated using the current rounding mode using the following
4881 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4882 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4883 // store doesn't really need to be here (except
4884 // maybe to zero the other double)
4889 DebugLoc dl = Op.getDebugLoc();
4891 // Build some magic constants.
4892 std::vector<Constant*> CV0;
4893 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4894 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4895 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4896 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4897 Constant *C0 = ConstantVector::get(CV0);
4898 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4900 std::vector<Constant*> CV1;
4901 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4902 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4903 Constant *C1 = ConstantVector::get(CV1);
4904 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4906 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4907 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4909 DAG.getIntPtrConstant(1)));
4910 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4911 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4913 DAG.getIntPtrConstant(0)));
4914 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4915 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4916 PseudoSourceValue::getConstantPool(), 0,
4918 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4919 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4920 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4921 PseudoSourceValue::getConstantPool(), 0,
4923 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4925 // Add the halves; easiest way is to swap them into another reg first.
4926 int ShufMask[2] = { 1, -1 };
4927 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4928 DAG.getUNDEF(MVT::v2f64), ShufMask);
4929 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4930 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4931 DAG.getIntPtrConstant(0));
4934 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4935 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4936 DebugLoc dl = Op.getDebugLoc();
4937 // FP constant to bias correct the final result.
4938 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4941 // Load the 32-bit value into an XMM register.
4942 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4943 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4945 DAG.getIntPtrConstant(0)));
4947 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4948 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4949 DAG.getIntPtrConstant(0));
4951 // Or the load with the bias.
4952 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4953 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4954 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4956 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4957 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4958 MVT::v2f64, Bias)));
4959 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4960 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4961 DAG.getIntPtrConstant(0));
4963 // Subtract the bias.
4964 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4966 // Handle final rounding.
4967 MVT DestVT = Op.getValueType();
4969 if (DestVT.bitsLT(MVT::f64)) {
4970 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4971 DAG.getIntPtrConstant(0));
4972 } else if (DestVT.bitsGT(MVT::f64)) {
4973 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4976 // Handle final rounding.
4980 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4981 SDValue N0 = Op.getOperand(0);
4982 DebugLoc dl = Op.getDebugLoc();
4984 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4985 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4986 // the optimization here.
4987 if (DAG.SignBitIsZero(N0))
4988 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4990 MVT SrcVT = N0.getValueType();
4991 if (SrcVT == MVT::i64) {
4992 // We only handle SSE2 f64 target here; caller can expand the rest.
4993 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4996 return LowerUINT_TO_FP_i64(Op, DAG);
4997 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
4998 return LowerUINT_TO_FP_i32(Op, DAG);
5001 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5003 // Make a 64-bit buffer, and use it to build an FILD.
5004 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5005 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5006 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5007 getPointerTy(), StackSlot, WordOff);
5008 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5009 StackSlot, NULL, 0);
5010 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5011 OffsetSlot, NULL, 0);
5012 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5015 std::pair<SDValue,SDValue> X86TargetLowering::
5016 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5017 DebugLoc dl = Op.getDebugLoc();
5019 MVT DstTy = Op.getValueType();
5022 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5026 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5027 DstTy.getSimpleVT() >= MVT::i16 &&
5028 "Unknown FP_TO_SINT to lower!");
5030 // These are really Legal.
5031 if (DstTy == MVT::i32 &&
5032 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5033 return std::make_pair(SDValue(), SDValue());
5034 if (Subtarget->is64Bit() &&
5035 DstTy == MVT::i64 &&
5036 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5037 return std::make_pair(SDValue(), SDValue());
5039 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5041 MachineFunction &MF = DAG.getMachineFunction();
5042 unsigned MemSize = DstTy.getSizeInBits()/8;
5043 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5044 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5047 switch (DstTy.getSimpleVT()) {
5048 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5049 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5050 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5051 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5054 SDValue Chain = DAG.getEntryNode();
5055 SDValue Value = Op.getOperand(0);
5056 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5057 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5058 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5059 PseudoSourceValue::getFixedStack(SSFI), 0);
5060 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5062 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5064 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5065 Chain = Value.getValue(1);
5066 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5067 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5070 // Build the FP_TO_INT*_IN_MEM
5071 SDValue Ops[] = { Chain, Value, StackSlot };
5072 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5074 return std::make_pair(FIST, StackSlot);
5077 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5078 if (Op.getValueType().isVector()) {
5079 if (Op.getValueType() == MVT::v2i32 &&
5080 Op.getOperand(0).getValueType() == MVT::v2f64) {
5086 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5087 SDValue FIST = Vals.first, StackSlot = Vals.second;
5088 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5089 if (FIST.getNode() == 0) return Op;
5092 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5093 FIST, StackSlot, NULL, 0);
5096 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5097 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5098 SDValue FIST = Vals.first, StackSlot = Vals.second;
5099 assert(FIST.getNode() && "Unexpected failure");
5102 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5103 FIST, StackSlot, NULL, 0);
5106 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5107 DebugLoc dl = Op.getDebugLoc();
5108 MVT VT = Op.getValueType();
5111 EltVT = VT.getVectorElementType();
5112 std::vector<Constant*> CV;
5113 if (EltVT == MVT::f64) {
5114 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5118 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5124 Constant *C = ConstantVector::get(CV);
5125 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5126 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5127 PseudoSourceValue::getConstantPool(), 0,
5129 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5132 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5133 DebugLoc dl = Op.getDebugLoc();
5134 MVT VT = Op.getValueType();
5136 unsigned EltNum = 1;
5137 if (VT.isVector()) {
5138 EltVT = VT.getVectorElementType();
5139 EltNum = VT.getVectorNumElements();
5141 std::vector<Constant*> CV;
5142 if (EltVT == MVT::f64) {
5143 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5147 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5153 Constant *C = ConstantVector::get(CV);
5154 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5155 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5156 PseudoSourceValue::getConstantPool(), 0,
5158 if (VT.isVector()) {
5159 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5160 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5161 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5163 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5165 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5169 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5170 SDValue Op0 = Op.getOperand(0);
5171 SDValue Op1 = Op.getOperand(1);
5172 DebugLoc dl = Op.getDebugLoc();
5173 MVT VT = Op.getValueType();
5174 MVT SrcVT = Op1.getValueType();
5176 // If second operand is smaller, extend it first.
5177 if (SrcVT.bitsLT(VT)) {
5178 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5181 // And if it is bigger, shrink it first.
5182 if (SrcVT.bitsGT(VT)) {
5183 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5187 // At this point the operands and the result should have the same
5188 // type, and that won't be f80 since that is not custom lowered.
5190 // First get the sign bit of second operand.
5191 std::vector<Constant*> CV;
5192 if (SrcVT == MVT::f64) {
5193 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5194 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5196 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5197 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5198 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5199 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5201 Constant *C = ConstantVector::get(CV);
5202 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5203 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5204 PseudoSourceValue::getConstantPool(), 0,
5206 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5208 // Shift sign bit right or left if the two operands have different types.
5209 if (SrcVT.bitsGT(VT)) {
5210 // Op0 is MVT::f32, Op1 is MVT::f64.
5211 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5212 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5213 DAG.getConstant(32, MVT::i32));
5214 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5215 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5216 DAG.getIntPtrConstant(0));
5219 // Clear first operand sign bit.
5221 if (VT == MVT::f64) {
5222 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5223 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5225 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5226 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5227 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5228 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5230 C = ConstantVector::get(CV);
5231 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5232 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5233 PseudoSourceValue::getConstantPool(), 0,
5235 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5237 // Or the value with the sign bit.
5238 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5241 /// Emit nodes that will be selected as "test Op0,Op0", or something
5243 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5244 SelectionDAG &DAG) {
5245 DebugLoc dl = Op.getDebugLoc();
5247 // CF and OF aren't always set the way we want. Determine which
5248 // of these we need.
5249 bool NeedCF = false;
5250 bool NeedOF = false;
5252 case X86::COND_A: case X86::COND_AE:
5253 case X86::COND_B: case X86::COND_BE:
5256 case X86::COND_G: case X86::COND_GE:
5257 case X86::COND_L: case X86::COND_LE:
5258 case X86::COND_O: case X86::COND_NO:
5264 // See if we can use the EFLAGS value from the operand instead of
5265 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5266 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5267 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5268 unsigned Opcode = 0;
5269 unsigned NumOperands = 0;
5270 switch (Op.getNode()->getOpcode()) {
5272 // Due to an isel shortcoming, be conservative if this add is likely to
5273 // be selected as part of a load-modify-store instruction. When the root
5274 // node in a match is a store, isel doesn't know how to remap non-chain
5275 // non-flag uses of other nodes in the match, such as the ADD in this
5276 // case. This leads to the ADD being left around and reselected, with
5277 // the result being two adds in the output.
5278 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5279 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5280 if (UI->getOpcode() == ISD::STORE)
5282 if (ConstantSDNode *C =
5283 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5284 // An add of one will be selected as an INC.
5285 if (C->getAPIntValue() == 1) {
5286 Opcode = X86ISD::INC;
5290 // An add of negative one (subtract of one) will be selected as a DEC.
5291 if (C->getAPIntValue().isAllOnesValue()) {
5292 Opcode = X86ISD::DEC;
5297 // Otherwise use a regular EFLAGS-setting add.
5298 Opcode = X86ISD::ADD;
5302 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5303 // likely to be selected as part of a load-modify-store instruction.
5304 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5305 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5306 if (UI->getOpcode() == ISD::STORE)
5308 // Otherwise use a regular EFLAGS-setting sub.
5309 Opcode = X86ISD::SUB;
5316 return SDValue(Op.getNode(), 1);
5322 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5323 SmallVector<SDValue, 4> Ops;
5324 for (unsigned i = 0; i != NumOperands; ++i)
5325 Ops.push_back(Op.getOperand(i));
5326 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5327 DAG.ReplaceAllUsesWith(Op, New);
5328 return SDValue(New.getNode(), 1);
5332 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5333 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5334 DAG.getConstant(0, Op.getValueType()));
5337 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5339 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5340 SelectionDAG &DAG) {
5341 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5342 if (C->getAPIntValue() == 0)
5343 return EmitTest(Op0, X86CC, DAG);
5345 DebugLoc dl = Op0.getDebugLoc();
5346 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5349 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5350 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5351 SDValue Op0 = Op.getOperand(0);
5352 SDValue Op1 = Op.getOperand(1);
5353 DebugLoc dl = Op.getDebugLoc();
5354 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5356 // Lower (X & (1 << N)) == 0 to BT(X, N).
5357 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5358 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5359 if (Op0.getOpcode() == ISD::AND &&
5361 Op1.getOpcode() == ISD::Constant &&
5362 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5363 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5365 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5366 if (ConstantSDNode *Op010C =
5367 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5368 if (Op010C->getZExtValue() == 1) {
5369 LHS = Op0.getOperand(0);
5370 RHS = Op0.getOperand(1).getOperand(1);
5372 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5373 if (ConstantSDNode *Op000C =
5374 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5375 if (Op000C->getZExtValue() == 1) {
5376 LHS = Op0.getOperand(1);
5377 RHS = Op0.getOperand(0).getOperand(1);
5379 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5380 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5381 SDValue AndLHS = Op0.getOperand(0);
5382 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5383 LHS = AndLHS.getOperand(0);
5384 RHS = AndLHS.getOperand(1);
5388 if (LHS.getNode()) {
5389 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5390 // instruction. Since the shift amount is in-range-or-undefined, we know
5391 // that doing a bittest on the i16 value is ok. We extend to i32 because
5392 // the encoding for the i16 version is larger than the i32 version.
5393 if (LHS.getValueType() == MVT::i8)
5394 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5396 // If the operand types disagree, extend the shift amount to match. Since
5397 // BT ignores high bits (like shifts) we can use anyextend.
5398 if (LHS.getValueType() != RHS.getValueType())
5399 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5401 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5402 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5403 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5404 DAG.getConstant(Cond, MVT::i8), BT);
5408 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5409 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5411 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5412 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5413 DAG.getConstant(X86CC, MVT::i8), Cond);
5416 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5418 SDValue Op0 = Op.getOperand(0);
5419 SDValue Op1 = Op.getOperand(1);
5420 SDValue CC = Op.getOperand(2);
5421 MVT VT = Op.getValueType();
5422 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5423 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5424 DebugLoc dl = Op.getDebugLoc();
5428 MVT VT0 = Op0.getValueType();
5429 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5430 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5433 switch (SetCCOpcode) {
5436 case ISD::SETEQ: SSECC = 0; break;
5438 case ISD::SETGT: Swap = true; // Fallthrough
5440 case ISD::SETOLT: SSECC = 1; break;
5442 case ISD::SETGE: Swap = true; // Fallthrough
5444 case ISD::SETOLE: SSECC = 2; break;
5445 case ISD::SETUO: SSECC = 3; break;
5447 case ISD::SETNE: SSECC = 4; break;
5448 case ISD::SETULE: Swap = true;
5449 case ISD::SETUGE: SSECC = 5; break;
5450 case ISD::SETULT: Swap = true;
5451 case ISD::SETUGT: SSECC = 6; break;
5452 case ISD::SETO: SSECC = 7; break;
5455 std::swap(Op0, Op1);
5457 // In the two special cases we can't handle, emit two comparisons.
5459 if (SetCCOpcode == ISD::SETUEQ) {
5461 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5462 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5463 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5465 else if (SetCCOpcode == ISD::SETONE) {
5467 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5468 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5469 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5471 assert(0 && "Illegal FP comparison");
5473 // Handle all other FP comparisons here.
5474 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5477 // We are handling one of the integer comparisons here. Since SSE only has
5478 // GT and EQ comparisons for integer, swapping operands and multiple
5479 // operations may be required for some comparisons.
5480 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5481 bool Swap = false, Invert = false, FlipSigns = false;
5483 switch (VT.getSimpleVT()) {
5485 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5486 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5487 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5488 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5491 switch (SetCCOpcode) {
5493 case ISD::SETNE: Invert = true;
5494 case ISD::SETEQ: Opc = EQOpc; break;
5495 case ISD::SETLT: Swap = true;
5496 case ISD::SETGT: Opc = GTOpc; break;
5497 case ISD::SETGE: Swap = true;
5498 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5499 case ISD::SETULT: Swap = true;
5500 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5501 case ISD::SETUGE: Swap = true;
5502 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5505 std::swap(Op0, Op1);
5507 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5508 // bits of the inputs before performing those operations.
5510 MVT EltVT = VT.getVectorElementType();
5511 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5513 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5514 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5516 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5517 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5520 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5522 // If the logical-not of the result is required, perform that now.
5524 Result = DAG.getNOT(dl, Result, VT);
5529 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5530 static bool isX86LogicalCmp(SDValue Op) {
5531 unsigned Opc = Op.getNode()->getOpcode();
5532 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5534 if (Op.getResNo() == 1 &&
5535 (Opc == X86ISD::ADD ||
5536 Opc == X86ISD::SUB ||
5537 Opc == X86ISD::SMUL ||
5538 Opc == X86ISD::UMUL ||
5539 Opc == X86ISD::INC ||
5540 Opc == X86ISD::DEC))
5546 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5547 bool addTest = true;
5548 SDValue Cond = Op.getOperand(0);
5549 DebugLoc dl = Op.getDebugLoc();
5552 if (Cond.getOpcode() == ISD::SETCC)
5553 Cond = LowerSETCC(Cond, DAG);
5555 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5556 // setting operand in place of the X86ISD::SETCC.
5557 if (Cond.getOpcode() == X86ISD::SETCC) {
5558 CC = Cond.getOperand(0);
5560 SDValue Cmp = Cond.getOperand(1);
5561 unsigned Opc = Cmp.getOpcode();
5562 MVT VT = Op.getValueType();
5564 bool IllegalFPCMov = false;
5565 if (VT.isFloatingPoint() && !VT.isVector() &&
5566 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5567 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5569 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5570 Opc == X86ISD::BT) { // FIXME
5577 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5578 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5581 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5582 SmallVector<SDValue, 4> Ops;
5583 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5584 // condition is true.
5585 Ops.push_back(Op.getOperand(2));
5586 Ops.push_back(Op.getOperand(1));
5588 Ops.push_back(Cond);
5589 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5592 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5593 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5594 // from the AND / OR.
5595 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5596 Opc = Op.getOpcode();
5597 if (Opc != ISD::OR && Opc != ISD::AND)
5599 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5600 Op.getOperand(0).hasOneUse() &&
5601 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5602 Op.getOperand(1).hasOneUse());
5605 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5606 // 1 and that the SETCC node has a single use.
5607 static bool isXor1OfSetCC(SDValue Op) {
5608 if (Op.getOpcode() != ISD::XOR)
5610 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5611 if (N1C && N1C->getAPIntValue() == 1) {
5612 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5613 Op.getOperand(0).hasOneUse();
5618 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5619 bool addTest = true;
5620 SDValue Chain = Op.getOperand(0);
5621 SDValue Cond = Op.getOperand(1);
5622 SDValue Dest = Op.getOperand(2);
5623 DebugLoc dl = Op.getDebugLoc();
5626 if (Cond.getOpcode() == ISD::SETCC)
5627 Cond = LowerSETCC(Cond, DAG);
5629 // FIXME: LowerXALUO doesn't handle these!!
5630 else if (Cond.getOpcode() == X86ISD::ADD ||
5631 Cond.getOpcode() == X86ISD::SUB ||
5632 Cond.getOpcode() == X86ISD::SMUL ||
5633 Cond.getOpcode() == X86ISD::UMUL)
5634 Cond = LowerXALUO(Cond, DAG);
5637 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5638 // setting operand in place of the X86ISD::SETCC.
5639 if (Cond.getOpcode() == X86ISD::SETCC) {
5640 CC = Cond.getOperand(0);
5642 SDValue Cmp = Cond.getOperand(1);
5643 unsigned Opc = Cmp.getOpcode();
5644 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5645 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5649 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5653 // These can only come from an arithmetic instruction with overflow,
5654 // e.g. SADDO, UADDO.
5655 Cond = Cond.getNode()->getOperand(1);
5662 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5663 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5664 if (CondOpc == ISD::OR) {
5665 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5666 // two branches instead of an explicit OR instruction with a
5668 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5669 isX86LogicalCmp(Cmp)) {
5670 CC = Cond.getOperand(0).getOperand(0);
5671 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5672 Chain, Dest, CC, Cmp);
5673 CC = Cond.getOperand(1).getOperand(0);
5677 } else { // ISD::AND
5678 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5679 // two branches instead of an explicit AND instruction with a
5680 // separate test. However, we only do this if this block doesn't
5681 // have a fall-through edge, because this requires an explicit
5682 // jmp when the condition is false.
5683 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5684 isX86LogicalCmp(Cmp) &&
5685 Op.getNode()->hasOneUse()) {
5686 X86::CondCode CCode =
5687 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5688 CCode = X86::GetOppositeBranchCondition(CCode);
5689 CC = DAG.getConstant(CCode, MVT::i8);
5690 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5691 // Look for an unconditional branch following this conditional branch.
5692 // We need this because we need to reverse the successors in order
5693 // to implement FCMP_OEQ.
5694 if (User.getOpcode() == ISD::BR) {
5695 SDValue FalseBB = User.getOperand(1);
5697 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5698 assert(NewBR == User);
5701 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5702 Chain, Dest, CC, Cmp);
5703 X86::CondCode CCode =
5704 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5705 CCode = X86::GetOppositeBranchCondition(CCode);
5706 CC = DAG.getConstant(CCode, MVT::i8);
5712 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5713 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5714 // It should be transformed during dag combiner except when the condition
5715 // is set by a arithmetics with overflow node.
5716 X86::CondCode CCode =
5717 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5718 CCode = X86::GetOppositeBranchCondition(CCode);
5719 CC = DAG.getConstant(CCode, MVT::i8);
5720 Cond = Cond.getOperand(0).getOperand(1);
5726 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5727 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5729 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5730 Chain, Dest, CC, Cond);
5734 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5735 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5736 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5737 // that the guard pages used by the OS virtual memory manager are allocated in
5738 // correct sequence.
5740 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5741 SelectionDAG &DAG) {
5742 assert(Subtarget->isTargetCygMing() &&
5743 "This should be used only on Cygwin/Mingw targets");
5744 DebugLoc dl = Op.getDebugLoc();
5747 SDValue Chain = Op.getOperand(0);
5748 SDValue Size = Op.getOperand(1);
5749 // FIXME: Ensure alignment here
5753 MVT IntPtr = getPointerTy();
5754 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5756 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5758 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5759 Flag = Chain.getValue(1);
5761 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5762 SDValue Ops[] = { Chain,
5763 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5764 DAG.getRegister(X86::EAX, IntPtr),
5765 DAG.getRegister(X86StackPtr, SPTy),
5767 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5768 Flag = Chain.getValue(1);
5770 Chain = DAG.getCALLSEQ_END(Chain,
5771 DAG.getIntPtrConstant(0, true),
5772 DAG.getIntPtrConstant(0, true),
5775 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5777 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5778 return DAG.getMergeValues(Ops1, 2, dl);
5782 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5784 SDValue Dst, SDValue Src,
5785 SDValue Size, unsigned Align,
5787 uint64_t DstSVOff) {
5788 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5790 // If not DWORD aligned or size is more than the threshold, call the library.
5791 // The libc version is likely to be faster for these cases. It can use the
5792 // address value and run time information about the CPU.
5793 if ((Align & 3) != 0 ||
5795 ConstantSize->getZExtValue() >
5796 getSubtarget()->getMaxInlineSizeThreshold()) {
5797 SDValue InFlag(0, 0);
5799 // Check to see if there is a specialized entry-point for memory zeroing.
5800 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5802 if (const char *bzeroEntry = V &&
5803 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5804 MVT IntPtr = getPointerTy();
5805 const Type *IntPtrTy = TD->getIntPtrType();
5806 TargetLowering::ArgListTy Args;
5807 TargetLowering::ArgListEntry Entry;
5809 Entry.Ty = IntPtrTy;
5810 Args.push_back(Entry);
5812 Args.push_back(Entry);
5813 std::pair<SDValue,SDValue> CallResult =
5814 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5815 0, CallingConv::C, false,
5816 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5817 return CallResult.second;
5820 // Otherwise have the target-independent code call memset.
5824 uint64_t SizeVal = ConstantSize->getZExtValue();
5825 SDValue InFlag(0, 0);
5828 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5829 unsigned BytesLeft = 0;
5830 bool TwoRepStos = false;
5833 uint64_t Val = ValC->getZExtValue() & 255;
5835 // If the value is a constant, then we can potentially use larger sets.
5836 switch (Align & 3) {
5837 case 2: // WORD aligned
5840 Val = (Val << 8) | Val;
5842 case 0: // DWORD aligned
5845 Val = (Val << 8) | Val;
5846 Val = (Val << 16) | Val;
5847 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5850 Val = (Val << 32) | Val;
5853 default: // Byte aligned
5856 Count = DAG.getIntPtrConstant(SizeVal);
5860 if (AVT.bitsGT(MVT::i8)) {
5861 unsigned UBytes = AVT.getSizeInBits() / 8;
5862 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5863 BytesLeft = SizeVal % UBytes;
5866 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5868 InFlag = Chain.getValue(1);
5871 Count = DAG.getIntPtrConstant(SizeVal);
5872 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5873 InFlag = Chain.getValue(1);
5876 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5879 InFlag = Chain.getValue(1);
5880 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5883 InFlag = Chain.getValue(1);
5885 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5886 SmallVector<SDValue, 8> Ops;
5887 Ops.push_back(Chain);
5888 Ops.push_back(DAG.getValueType(AVT));
5889 Ops.push_back(InFlag);
5890 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5893 InFlag = Chain.getValue(1);
5895 MVT CVT = Count.getValueType();
5896 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5897 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5898 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5901 InFlag = Chain.getValue(1);
5902 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5904 Ops.push_back(Chain);
5905 Ops.push_back(DAG.getValueType(MVT::i8));
5906 Ops.push_back(InFlag);
5907 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5908 } else if (BytesLeft) {
5909 // Handle the last 1 - 7 bytes.
5910 unsigned Offset = SizeVal - BytesLeft;
5911 MVT AddrVT = Dst.getValueType();
5912 MVT SizeVT = Size.getValueType();
5914 Chain = DAG.getMemset(Chain, dl,
5915 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5916 DAG.getConstant(Offset, AddrVT)),
5918 DAG.getConstant(BytesLeft, SizeVT),
5919 Align, DstSV, DstSVOff + Offset);
5922 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5927 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5928 SDValue Chain, SDValue Dst, SDValue Src,
5929 SDValue Size, unsigned Align,
5931 const Value *DstSV, uint64_t DstSVOff,
5932 const Value *SrcSV, uint64_t SrcSVOff) {
5933 // This requires the copy size to be a constant, preferrably
5934 // within a subtarget-specific limit.
5935 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5938 uint64_t SizeVal = ConstantSize->getZExtValue();
5939 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5942 /// If not DWORD aligned, call the library.
5943 if ((Align & 3) != 0)
5948 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5951 unsigned UBytes = AVT.getSizeInBits() / 8;
5952 unsigned CountVal = SizeVal / UBytes;
5953 SDValue Count = DAG.getIntPtrConstant(CountVal);
5954 unsigned BytesLeft = SizeVal % UBytes;
5956 SDValue InFlag(0, 0);
5957 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5960 InFlag = Chain.getValue(1);
5961 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5964 InFlag = Chain.getValue(1);
5965 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5968 InFlag = Chain.getValue(1);
5970 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5971 SmallVector<SDValue, 8> Ops;
5972 Ops.push_back(Chain);
5973 Ops.push_back(DAG.getValueType(AVT));
5974 Ops.push_back(InFlag);
5975 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5977 SmallVector<SDValue, 4> Results;
5978 Results.push_back(RepMovs);
5980 // Handle the last 1 - 7 bytes.
5981 unsigned Offset = SizeVal - BytesLeft;
5982 MVT DstVT = Dst.getValueType();
5983 MVT SrcVT = Src.getValueType();
5984 MVT SizeVT = Size.getValueType();
5985 Results.push_back(DAG.getMemcpy(Chain, dl,
5986 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5987 DAG.getConstant(Offset, DstVT)),
5988 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5989 DAG.getConstant(Offset, SrcVT)),
5990 DAG.getConstant(BytesLeft, SizeVT),
5991 Align, AlwaysInline,
5992 DstSV, DstSVOff + Offset,
5993 SrcSV, SrcSVOff + Offset));
5996 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5997 &Results[0], Results.size());
6000 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6001 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6002 DebugLoc dl = Op.getDebugLoc();
6004 if (!Subtarget->is64Bit()) {
6005 // vastart just stores the address of the VarArgsFrameIndex slot into the
6006 // memory location argument.
6007 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6008 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6012 // gp_offset (0 - 6 * 8)
6013 // fp_offset (48 - 48 + 8 * 16)
6014 // overflow_arg_area (point to parameters coming in memory).
6016 SmallVector<SDValue, 8> MemOps;
6017 SDValue FIN = Op.getOperand(1);
6019 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6020 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6022 MemOps.push_back(Store);
6025 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6026 FIN, DAG.getIntPtrConstant(4));
6027 Store = DAG.getStore(Op.getOperand(0), dl,
6028 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6030 MemOps.push_back(Store);
6032 // Store ptr to overflow_arg_area
6033 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6034 FIN, DAG.getIntPtrConstant(4));
6035 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6036 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6037 MemOps.push_back(Store);
6039 // Store ptr to reg_save_area.
6040 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6041 FIN, DAG.getIntPtrConstant(8));
6042 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6043 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6044 MemOps.push_back(Store);
6045 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6046 &MemOps[0], MemOps.size());
6049 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6050 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6051 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6052 SDValue Chain = Op.getOperand(0);
6053 SDValue SrcPtr = Op.getOperand(1);
6054 SDValue SrcSV = Op.getOperand(2);
6056 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6060 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6061 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6062 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6063 SDValue Chain = Op.getOperand(0);
6064 SDValue DstPtr = Op.getOperand(1);
6065 SDValue SrcPtr = Op.getOperand(2);
6066 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6067 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6068 DebugLoc dl = Op.getDebugLoc();
6070 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6071 DAG.getIntPtrConstant(24), 8, false,
6072 DstSV, 0, SrcSV, 0);
6076 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6077 DebugLoc dl = Op.getDebugLoc();
6078 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6080 default: return SDValue(); // Don't custom lower most intrinsics.
6081 // Comparison intrinsics.
6082 case Intrinsic::x86_sse_comieq_ss:
6083 case Intrinsic::x86_sse_comilt_ss:
6084 case Intrinsic::x86_sse_comile_ss:
6085 case Intrinsic::x86_sse_comigt_ss:
6086 case Intrinsic::x86_sse_comige_ss:
6087 case Intrinsic::x86_sse_comineq_ss:
6088 case Intrinsic::x86_sse_ucomieq_ss:
6089 case Intrinsic::x86_sse_ucomilt_ss:
6090 case Intrinsic::x86_sse_ucomile_ss:
6091 case Intrinsic::x86_sse_ucomigt_ss:
6092 case Intrinsic::x86_sse_ucomige_ss:
6093 case Intrinsic::x86_sse_ucomineq_ss:
6094 case Intrinsic::x86_sse2_comieq_sd:
6095 case Intrinsic::x86_sse2_comilt_sd:
6096 case Intrinsic::x86_sse2_comile_sd:
6097 case Intrinsic::x86_sse2_comigt_sd:
6098 case Intrinsic::x86_sse2_comige_sd:
6099 case Intrinsic::x86_sse2_comineq_sd:
6100 case Intrinsic::x86_sse2_ucomieq_sd:
6101 case Intrinsic::x86_sse2_ucomilt_sd:
6102 case Intrinsic::x86_sse2_ucomile_sd:
6103 case Intrinsic::x86_sse2_ucomigt_sd:
6104 case Intrinsic::x86_sse2_ucomige_sd:
6105 case Intrinsic::x86_sse2_ucomineq_sd: {
6107 ISD::CondCode CC = ISD::SETCC_INVALID;
6110 case Intrinsic::x86_sse_comieq_ss:
6111 case Intrinsic::x86_sse2_comieq_sd:
6115 case Intrinsic::x86_sse_comilt_ss:
6116 case Intrinsic::x86_sse2_comilt_sd:
6120 case Intrinsic::x86_sse_comile_ss:
6121 case Intrinsic::x86_sse2_comile_sd:
6125 case Intrinsic::x86_sse_comigt_ss:
6126 case Intrinsic::x86_sse2_comigt_sd:
6130 case Intrinsic::x86_sse_comige_ss:
6131 case Intrinsic::x86_sse2_comige_sd:
6135 case Intrinsic::x86_sse_comineq_ss:
6136 case Intrinsic::x86_sse2_comineq_sd:
6140 case Intrinsic::x86_sse_ucomieq_ss:
6141 case Intrinsic::x86_sse2_ucomieq_sd:
6142 Opc = X86ISD::UCOMI;
6145 case Intrinsic::x86_sse_ucomilt_ss:
6146 case Intrinsic::x86_sse2_ucomilt_sd:
6147 Opc = X86ISD::UCOMI;
6150 case Intrinsic::x86_sse_ucomile_ss:
6151 case Intrinsic::x86_sse2_ucomile_sd:
6152 Opc = X86ISD::UCOMI;
6155 case Intrinsic::x86_sse_ucomigt_ss:
6156 case Intrinsic::x86_sse2_ucomigt_sd:
6157 Opc = X86ISD::UCOMI;
6160 case Intrinsic::x86_sse_ucomige_ss:
6161 case Intrinsic::x86_sse2_ucomige_sd:
6162 Opc = X86ISD::UCOMI;
6165 case Intrinsic::x86_sse_ucomineq_ss:
6166 case Intrinsic::x86_sse2_ucomineq_sd:
6167 Opc = X86ISD::UCOMI;
6172 SDValue LHS = Op.getOperand(1);
6173 SDValue RHS = Op.getOperand(2);
6174 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6175 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6176 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6177 DAG.getConstant(X86CC, MVT::i8), Cond);
6178 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6181 // Fix vector shift instructions where the last operand is a non-immediate
6183 case Intrinsic::x86_sse2_pslli_w:
6184 case Intrinsic::x86_sse2_pslli_d:
6185 case Intrinsic::x86_sse2_pslli_q:
6186 case Intrinsic::x86_sse2_psrli_w:
6187 case Intrinsic::x86_sse2_psrli_d:
6188 case Intrinsic::x86_sse2_psrli_q:
6189 case Intrinsic::x86_sse2_psrai_w:
6190 case Intrinsic::x86_sse2_psrai_d:
6191 case Intrinsic::x86_mmx_pslli_w:
6192 case Intrinsic::x86_mmx_pslli_d:
6193 case Intrinsic::x86_mmx_pslli_q:
6194 case Intrinsic::x86_mmx_psrli_w:
6195 case Intrinsic::x86_mmx_psrli_d:
6196 case Intrinsic::x86_mmx_psrli_q:
6197 case Intrinsic::x86_mmx_psrai_w:
6198 case Intrinsic::x86_mmx_psrai_d: {
6199 SDValue ShAmt = Op.getOperand(2);
6200 if (isa<ConstantSDNode>(ShAmt))
6203 unsigned NewIntNo = 0;
6204 MVT ShAmtVT = MVT::v4i32;
6206 case Intrinsic::x86_sse2_pslli_w:
6207 NewIntNo = Intrinsic::x86_sse2_psll_w;
6209 case Intrinsic::x86_sse2_pslli_d:
6210 NewIntNo = Intrinsic::x86_sse2_psll_d;
6212 case Intrinsic::x86_sse2_pslli_q:
6213 NewIntNo = Intrinsic::x86_sse2_psll_q;
6215 case Intrinsic::x86_sse2_psrli_w:
6216 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6218 case Intrinsic::x86_sse2_psrli_d:
6219 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6221 case Intrinsic::x86_sse2_psrli_q:
6222 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6224 case Intrinsic::x86_sse2_psrai_w:
6225 NewIntNo = Intrinsic::x86_sse2_psra_w;
6227 case Intrinsic::x86_sse2_psrai_d:
6228 NewIntNo = Intrinsic::x86_sse2_psra_d;
6231 ShAmtVT = MVT::v2i32;
6233 case Intrinsic::x86_mmx_pslli_w:
6234 NewIntNo = Intrinsic::x86_mmx_psll_w;
6236 case Intrinsic::x86_mmx_pslli_d:
6237 NewIntNo = Intrinsic::x86_mmx_psll_d;
6239 case Intrinsic::x86_mmx_pslli_q:
6240 NewIntNo = Intrinsic::x86_mmx_psll_q;
6242 case Intrinsic::x86_mmx_psrli_w:
6243 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6245 case Intrinsic::x86_mmx_psrli_d:
6246 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6248 case Intrinsic::x86_mmx_psrli_q:
6249 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6251 case Intrinsic::x86_mmx_psrai_w:
6252 NewIntNo = Intrinsic::x86_mmx_psra_w;
6254 case Intrinsic::x86_mmx_psrai_d:
6255 NewIntNo = Intrinsic::x86_mmx_psra_d;
6257 default: LLVM_UNREACHABLE("Impossible intrinsic"); // Can't reach here.
6262 MVT VT = Op.getValueType();
6263 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6264 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6265 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6266 DAG.getConstant(NewIntNo, MVT::i32),
6267 Op.getOperand(1), ShAmt);
6272 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6273 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6274 DebugLoc dl = Op.getDebugLoc();
6277 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6279 DAG.getConstant(TD->getPointerSize(),
6280 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6281 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6282 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6287 // Just load the return address.
6288 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6289 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6290 RetAddrFI, NULL, 0);
6293 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6294 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6295 MFI->setFrameAddressIsTaken(true);
6296 MVT VT = Op.getValueType();
6297 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6298 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6299 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6300 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6302 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6306 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6307 SelectionDAG &DAG) {
6308 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6311 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6313 MachineFunction &MF = DAG.getMachineFunction();
6314 SDValue Chain = Op.getOperand(0);
6315 SDValue Offset = Op.getOperand(1);
6316 SDValue Handler = Op.getOperand(2);
6317 DebugLoc dl = Op.getDebugLoc();
6319 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6321 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6323 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6324 DAG.getIntPtrConstant(-TD->getPointerSize()));
6325 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6326 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6327 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6328 MF.getRegInfo().addLiveOut(StoreAddrReg);
6330 return DAG.getNode(X86ISD::EH_RETURN, dl,
6332 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6335 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6336 SelectionDAG &DAG) {
6337 SDValue Root = Op.getOperand(0);
6338 SDValue Trmp = Op.getOperand(1); // trampoline
6339 SDValue FPtr = Op.getOperand(2); // nested function
6340 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6341 DebugLoc dl = Op.getDebugLoc();
6343 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6345 const X86InstrInfo *TII =
6346 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6348 if (Subtarget->is64Bit()) {
6349 SDValue OutChains[6];
6351 // Large code-model.
6353 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6354 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6356 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6357 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6359 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6361 // Load the pointer to the nested function into R11.
6362 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6363 SDValue Addr = Trmp;
6364 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6367 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6368 DAG.getConstant(2, MVT::i64));
6369 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6371 // Load the 'nest' parameter value into R10.
6372 // R10 is specified in X86CallingConv.td
6373 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6374 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6375 DAG.getConstant(10, MVT::i64));
6376 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6377 Addr, TrmpAddr, 10);
6379 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6380 DAG.getConstant(12, MVT::i64));
6381 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6383 // Jump to the nested function.
6384 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6385 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6386 DAG.getConstant(20, MVT::i64));
6387 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6388 Addr, TrmpAddr, 20);
6390 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6391 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6392 DAG.getConstant(22, MVT::i64));
6393 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6397 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6398 return DAG.getMergeValues(Ops, 2, dl);
6400 const Function *Func =
6401 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6402 unsigned CC = Func->getCallingConv();
6407 assert(0 && "Unsupported calling convention");
6408 case CallingConv::C:
6409 case CallingConv::X86_StdCall: {
6410 // Pass 'nest' parameter in ECX.
6411 // Must be kept in sync with X86CallingConv.td
6414 // Check that ECX wasn't needed by an 'inreg' parameter.
6415 const FunctionType *FTy = Func->getFunctionType();
6416 const AttrListPtr &Attrs = Func->getAttributes();
6418 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6419 unsigned InRegCount = 0;
6422 for (FunctionType::param_iterator I = FTy->param_begin(),
6423 E = FTy->param_end(); I != E; ++I, ++Idx)
6424 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6425 // FIXME: should only count parameters that are lowered to integers.
6426 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6428 if (InRegCount > 2) {
6429 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6434 case CallingConv::X86_FastCall:
6435 case CallingConv::Fast:
6436 // Pass 'nest' parameter in EAX.
6437 // Must be kept in sync with X86CallingConv.td
6442 SDValue OutChains[4];
6445 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6446 DAG.getConstant(10, MVT::i32));
6447 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6449 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6450 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6451 OutChains[0] = DAG.getStore(Root, dl,
6452 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6455 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6456 DAG.getConstant(1, MVT::i32));
6457 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6459 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6460 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6461 DAG.getConstant(5, MVT::i32));
6462 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6463 TrmpAddr, 5, false, 1);
6465 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6466 DAG.getConstant(6, MVT::i32));
6467 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6470 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6471 return DAG.getMergeValues(Ops, 2, dl);
6475 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6477 The rounding mode is in bits 11:10 of FPSR, and has the following
6484 FLT_ROUNDS, on the other hand, expects the following:
6491 To perform the conversion, we do:
6492 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6495 MachineFunction &MF = DAG.getMachineFunction();
6496 const TargetMachine &TM = MF.getTarget();
6497 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6498 unsigned StackAlignment = TFI.getStackAlignment();
6499 MVT VT = Op.getValueType();
6500 DebugLoc dl = Op.getDebugLoc();
6502 // Save FP Control Word to stack slot
6503 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6504 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6506 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6507 DAG.getEntryNode(), StackSlot);
6509 // Load FP Control Word from stack slot
6510 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6512 // Transform as necessary
6514 DAG.getNode(ISD::SRL, dl, MVT::i16,
6515 DAG.getNode(ISD::AND, dl, MVT::i16,
6516 CWD, DAG.getConstant(0x800, MVT::i16)),
6517 DAG.getConstant(11, MVT::i8));
6519 DAG.getNode(ISD::SRL, dl, MVT::i16,
6520 DAG.getNode(ISD::AND, dl, MVT::i16,
6521 CWD, DAG.getConstant(0x400, MVT::i16)),
6522 DAG.getConstant(9, MVT::i8));
6525 DAG.getNode(ISD::AND, dl, MVT::i16,
6526 DAG.getNode(ISD::ADD, dl, MVT::i16,
6527 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6528 DAG.getConstant(1, MVT::i16)),
6529 DAG.getConstant(3, MVT::i16));
6532 return DAG.getNode((VT.getSizeInBits() < 16 ?
6533 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6536 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6537 MVT VT = Op.getValueType();
6539 unsigned NumBits = VT.getSizeInBits();
6540 DebugLoc dl = Op.getDebugLoc();
6542 Op = Op.getOperand(0);
6543 if (VT == MVT::i8) {
6544 // Zero extend to i32 since there is not an i8 bsr.
6546 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6549 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6550 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6551 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6553 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6554 SmallVector<SDValue, 4> Ops;
6556 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6557 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6558 Ops.push_back(Op.getValue(1));
6559 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6561 // Finally xor with NumBits-1.
6562 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6565 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6569 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6570 MVT VT = Op.getValueType();
6572 unsigned NumBits = VT.getSizeInBits();
6573 DebugLoc dl = Op.getDebugLoc();
6575 Op = Op.getOperand(0);
6576 if (VT == MVT::i8) {
6578 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6581 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6582 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6583 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6585 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6586 SmallVector<SDValue, 4> Ops;
6588 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6589 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6590 Ops.push_back(Op.getValue(1));
6591 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6594 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6598 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6599 MVT VT = Op.getValueType();
6600 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6601 DebugLoc dl = Op.getDebugLoc();
6603 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6604 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6605 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6606 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6607 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6609 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6610 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6611 // return AloBlo + AloBhi + AhiBlo;
6613 SDValue A = Op.getOperand(0);
6614 SDValue B = Op.getOperand(1);
6616 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6617 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6618 A, DAG.getConstant(32, MVT::i32));
6619 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6620 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6621 B, DAG.getConstant(32, MVT::i32));
6622 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6623 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6625 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6626 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6628 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6629 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6631 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6632 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6633 AloBhi, DAG.getConstant(32, MVT::i32));
6634 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6635 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6636 AhiBlo, DAG.getConstant(32, MVT::i32));
6637 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6638 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6643 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6644 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6645 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6646 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6647 // has only one use.
6648 SDNode *N = Op.getNode();
6649 SDValue LHS = N->getOperand(0);
6650 SDValue RHS = N->getOperand(1);
6651 unsigned BaseOp = 0;
6653 DebugLoc dl = Op.getDebugLoc();
6655 switch (Op.getOpcode()) {
6656 default: assert(0 && "Unknown ovf instruction!");
6658 // A subtract of one will be selected as a INC. Note that INC doesn't
6659 // set CF, so we can't do this for UADDO.
6660 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6661 if (C->getAPIntValue() == 1) {
6662 BaseOp = X86ISD::INC;
6666 BaseOp = X86ISD::ADD;
6670 BaseOp = X86ISD::ADD;
6674 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6675 // set CF, so we can't do this for USUBO.
6676 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6677 if (C->getAPIntValue() == 1) {
6678 BaseOp = X86ISD::DEC;
6682 BaseOp = X86ISD::SUB;
6686 BaseOp = X86ISD::SUB;
6690 BaseOp = X86ISD::SMUL;
6694 BaseOp = X86ISD::UMUL;
6699 // Also sets EFLAGS.
6700 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6701 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6704 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6705 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6707 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6711 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6712 MVT T = Op.getValueType();
6713 DebugLoc dl = Op.getDebugLoc();
6716 switch(T.getSimpleVT()) {
6718 assert(false && "Invalid value type!");
6719 case MVT::i8: Reg = X86::AL; size = 1; break;
6720 case MVT::i16: Reg = X86::AX; size = 2; break;
6721 case MVT::i32: Reg = X86::EAX; size = 4; break;
6723 assert(Subtarget->is64Bit() && "Node not type legal!");
6724 Reg = X86::RAX; size = 8;
6727 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6728 Op.getOperand(2), SDValue());
6729 SDValue Ops[] = { cpIn.getValue(0),
6732 DAG.getTargetConstant(size, MVT::i8),
6734 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6735 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6737 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6741 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6742 SelectionDAG &DAG) {
6743 assert(Subtarget->is64Bit() && "Result not type legalized?");
6744 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6745 SDValue TheChain = Op.getOperand(0);
6746 DebugLoc dl = Op.getDebugLoc();
6747 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6748 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6749 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6751 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6752 DAG.getConstant(32, MVT::i8));
6754 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6757 return DAG.getMergeValues(Ops, 2, dl);
6760 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6761 SDNode *Node = Op.getNode();
6762 DebugLoc dl = Node->getDebugLoc();
6763 MVT T = Node->getValueType(0);
6764 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6765 DAG.getConstant(0, T), Node->getOperand(2));
6766 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6767 cast<AtomicSDNode>(Node)->getMemoryVT(),
6768 Node->getOperand(0),
6769 Node->getOperand(1), negOp,
6770 cast<AtomicSDNode>(Node)->getSrcValue(),
6771 cast<AtomicSDNode>(Node)->getAlignment());
6774 /// LowerOperation - Provide custom lowering hooks for some operations.
6776 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6777 switch (Op.getOpcode()) {
6778 default: assert(0 && "Should not custom lower this!");
6779 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6780 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6781 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6782 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6783 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6784 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6785 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6786 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6787 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6788 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6789 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6790 case ISD::SHL_PARTS:
6791 case ISD::SRA_PARTS:
6792 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6793 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6794 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6795 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6796 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6797 case ISD::FABS: return LowerFABS(Op, DAG);
6798 case ISD::FNEG: return LowerFNEG(Op, DAG);
6799 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6800 case ISD::SETCC: return LowerSETCC(Op, DAG);
6801 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6802 case ISD::SELECT: return LowerSELECT(Op, DAG);
6803 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6804 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6805 case ISD::CALL: return LowerCALL(Op, DAG);
6806 case ISD::RET: return LowerRET(Op, DAG);
6807 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6808 case ISD::VASTART: return LowerVASTART(Op, DAG);
6809 case ISD::VAARG: return LowerVAARG(Op, DAG);
6810 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6811 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6812 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6813 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6814 case ISD::FRAME_TO_ARGS_OFFSET:
6815 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6816 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6817 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6818 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6819 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6820 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6821 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6822 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6828 case ISD::UMULO: return LowerXALUO(Op, DAG);
6829 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6833 void X86TargetLowering::
6834 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6835 SelectionDAG &DAG, unsigned NewOp) {
6836 MVT T = Node->getValueType(0);
6837 DebugLoc dl = Node->getDebugLoc();
6838 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6840 SDValue Chain = Node->getOperand(0);
6841 SDValue In1 = Node->getOperand(1);
6842 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6843 Node->getOperand(2), DAG.getIntPtrConstant(0));
6844 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6845 Node->getOperand(2), DAG.getIntPtrConstant(1));
6846 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6847 // have a MemOperand. Pass the info through as a normal operand.
6848 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6849 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6850 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6851 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6852 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6853 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6854 Results.push_back(Result.getValue(2));
6857 /// ReplaceNodeResults - Replace a node with an illegal result type
6858 /// with a new node built out of custom code.
6859 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6860 SmallVectorImpl<SDValue>&Results,
6861 SelectionDAG &DAG) {
6862 DebugLoc dl = N->getDebugLoc();
6863 switch (N->getOpcode()) {
6865 assert(false && "Do not know how to custom type legalize this operation!");
6867 case ISD::FP_TO_SINT: {
6868 std::pair<SDValue,SDValue> Vals =
6869 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6870 SDValue FIST = Vals.first, StackSlot = Vals.second;
6871 if (FIST.getNode() != 0) {
6872 MVT VT = N->getValueType(0);
6873 // Return a load from the stack slot.
6874 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6878 case ISD::READCYCLECOUNTER: {
6879 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6880 SDValue TheChain = N->getOperand(0);
6881 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6882 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6884 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6886 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6887 SDValue Ops[] = { eax, edx };
6888 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6889 Results.push_back(edx.getValue(1));
6892 case ISD::ATOMIC_CMP_SWAP: {
6893 MVT T = N->getValueType(0);
6894 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6895 SDValue cpInL, cpInH;
6896 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6897 DAG.getConstant(0, MVT::i32));
6898 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6899 DAG.getConstant(1, MVT::i32));
6900 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6901 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6903 SDValue swapInL, swapInH;
6904 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6905 DAG.getConstant(0, MVT::i32));
6906 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6907 DAG.getConstant(1, MVT::i32));
6908 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6910 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6911 swapInL.getValue(1));
6912 SDValue Ops[] = { swapInH.getValue(0),
6914 swapInH.getValue(1) };
6915 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6916 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6917 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6918 MVT::i32, Result.getValue(1));
6919 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6920 MVT::i32, cpOutL.getValue(2));
6921 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6922 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6923 Results.push_back(cpOutH.getValue(1));
6926 case ISD::ATOMIC_LOAD_ADD:
6927 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6929 case ISD::ATOMIC_LOAD_AND:
6930 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6932 case ISD::ATOMIC_LOAD_NAND:
6933 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6935 case ISD::ATOMIC_LOAD_OR:
6936 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6938 case ISD::ATOMIC_LOAD_SUB:
6939 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6941 case ISD::ATOMIC_LOAD_XOR:
6942 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6944 case ISD::ATOMIC_SWAP:
6945 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6950 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6952 default: return NULL;
6953 case X86ISD::BSF: return "X86ISD::BSF";
6954 case X86ISD::BSR: return "X86ISD::BSR";
6955 case X86ISD::SHLD: return "X86ISD::SHLD";
6956 case X86ISD::SHRD: return "X86ISD::SHRD";
6957 case X86ISD::FAND: return "X86ISD::FAND";
6958 case X86ISD::FOR: return "X86ISD::FOR";
6959 case X86ISD::FXOR: return "X86ISD::FXOR";
6960 case X86ISD::FSRL: return "X86ISD::FSRL";
6961 case X86ISD::FILD: return "X86ISD::FILD";
6962 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6963 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6964 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6965 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6966 case X86ISD::FLD: return "X86ISD::FLD";
6967 case X86ISD::FST: return "X86ISD::FST";
6968 case X86ISD::CALL: return "X86ISD::CALL";
6969 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6970 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6971 case X86ISD::BT: return "X86ISD::BT";
6972 case X86ISD::CMP: return "X86ISD::CMP";
6973 case X86ISD::COMI: return "X86ISD::COMI";
6974 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6975 case X86ISD::SETCC: return "X86ISD::SETCC";
6976 case X86ISD::CMOV: return "X86ISD::CMOV";
6977 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6978 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6979 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6980 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6981 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6982 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6983 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
6984 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6985 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6986 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6987 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6988 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6989 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6990 case X86ISD::FMAX: return "X86ISD::FMAX";
6991 case X86ISD::FMIN: return "X86ISD::FMIN";
6992 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6993 case X86ISD::FRCP: return "X86ISD::FRCP";
6994 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6995 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
6996 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6997 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6998 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6999 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7000 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7001 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7002 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7003 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7004 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7005 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7006 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7007 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7008 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7009 case X86ISD::VSHL: return "X86ISD::VSHL";
7010 case X86ISD::VSRL: return "X86ISD::VSRL";
7011 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7012 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7013 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7014 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7015 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7016 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7017 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7018 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7019 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7020 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7021 case X86ISD::ADD: return "X86ISD::ADD";
7022 case X86ISD::SUB: return "X86ISD::SUB";
7023 case X86ISD::SMUL: return "X86ISD::SMUL";
7024 case X86ISD::UMUL: return "X86ISD::UMUL";
7025 case X86ISD::INC: return "X86ISD::INC";
7026 case X86ISD::DEC: return "X86ISD::DEC";
7027 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7031 // isLegalAddressingMode - Return true if the addressing mode represented
7032 // by AM is legal for this target, for a load/store of the specified type.
7033 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7034 const Type *Ty) const {
7035 // X86 supports extremely general addressing modes.
7037 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7038 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7042 // We can only fold this if we don't need an extra load.
7043 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
7045 // If BaseGV requires a register, we cannot also have a BaseReg.
7046 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
7050 // X86-64 only supports addr of globals in small code model.
7051 if (Subtarget->is64Bit()) {
7052 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7054 // If lower 4G is not available, then we must use rip-relative addressing.
7055 if (AM.BaseOffs || AM.Scale > 1)
7066 // These scales always work.
7071 // These scales are formed with basereg+scalereg. Only accept if there is
7076 default: // Other stuff never works.
7084 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7085 if (!Ty1->isInteger() || !Ty2->isInteger())
7087 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7088 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7089 if (NumBits1 <= NumBits2)
7091 return Subtarget->is64Bit() || NumBits1 < 64;
7094 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7095 if (!VT1.isInteger() || !VT2.isInteger())
7097 unsigned NumBits1 = VT1.getSizeInBits();
7098 unsigned NumBits2 = VT2.getSizeInBits();
7099 if (NumBits1 <= NumBits2)
7101 return Subtarget->is64Bit() || NumBits1 < 64;
7104 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7105 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7106 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7109 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7110 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7111 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7114 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
7115 // i16 instructions are longer (0x66 prefix) and potentially slower.
7116 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7119 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7120 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7121 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7122 /// are assumed to be legal.
7124 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7126 // Only do shuffles on 128-bit vector types for now.
7127 if (VT.getSizeInBits() == 64)
7130 // FIXME: pshufb, blends, palignr, shifts.
7131 return (VT.getVectorNumElements() == 2 ||
7132 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7133 isMOVLMask(M, VT) ||
7134 isSHUFPMask(M, VT) ||
7135 isPSHUFDMask(M, VT) ||
7136 isPSHUFHWMask(M, VT) ||
7137 isPSHUFLWMask(M, VT) ||
7138 isUNPCKLMask(M, VT) ||
7139 isUNPCKHMask(M, VT) ||
7140 isUNPCKL_v_undef_Mask(M, VT) ||
7141 isUNPCKH_v_undef_Mask(M, VT));
7145 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7147 unsigned NumElts = VT.getVectorNumElements();
7148 // FIXME: This collection of masks seems suspect.
7151 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7152 return (isMOVLMask(Mask, VT) ||
7153 isCommutedMOVLMask(Mask, VT, true) ||
7154 isSHUFPMask(Mask, VT) ||
7155 isCommutedSHUFPMask(Mask, VT));
7160 //===----------------------------------------------------------------------===//
7161 // X86 Scheduler Hooks
7162 //===----------------------------------------------------------------------===//
7164 // private utility function
7166 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7167 MachineBasicBlock *MBB,
7175 TargetRegisterClass *RC,
7176 bool invSrc) const {
7177 // For the atomic bitwise operator, we generate
7180 // ld t1 = [bitinstr.addr]
7181 // op t2 = t1, [bitinstr.val]
7183 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7185 // fallthrough -->nextMBB
7186 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7187 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7188 MachineFunction::iterator MBBIter = MBB;
7191 /// First build the CFG
7192 MachineFunction *F = MBB->getParent();
7193 MachineBasicBlock *thisMBB = MBB;
7194 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7195 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7196 F->insert(MBBIter, newMBB);
7197 F->insert(MBBIter, nextMBB);
7199 // Move all successors to thisMBB to nextMBB
7200 nextMBB->transferSuccessors(thisMBB);
7202 // Update thisMBB to fall through to newMBB
7203 thisMBB->addSuccessor(newMBB);
7205 // newMBB jumps to itself and fall through to nextMBB
7206 newMBB->addSuccessor(nextMBB);
7207 newMBB->addSuccessor(newMBB);
7209 // Insert instructions into newMBB based on incoming instruction
7210 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7211 "unexpected number of operands");
7212 DebugLoc dl = bInstr->getDebugLoc();
7213 MachineOperand& destOper = bInstr->getOperand(0);
7214 MachineOperand* argOpers[2 + X86AddrNumOperands];
7215 int numArgs = bInstr->getNumOperands() - 1;
7216 for (int i=0; i < numArgs; ++i)
7217 argOpers[i] = &bInstr->getOperand(i+1);
7219 // x86 address has 4 operands: base, index, scale, and displacement
7220 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7221 int valArgIndx = lastAddrIndx + 1;
7223 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7224 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7225 for (int i=0; i <= lastAddrIndx; ++i)
7226 (*MIB).addOperand(*argOpers[i]);
7228 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7230 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7235 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7236 assert((argOpers[valArgIndx]->isReg() ||
7237 argOpers[valArgIndx]->isImm()) &&
7239 if (argOpers[valArgIndx]->isReg())
7240 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7242 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7244 (*MIB).addOperand(*argOpers[valArgIndx]);
7246 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7249 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7250 for (int i=0; i <= lastAddrIndx; ++i)
7251 (*MIB).addOperand(*argOpers[i]);
7253 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7254 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7256 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7260 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7262 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7266 // private utility function: 64 bit atomics on 32 bit host.
7268 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7269 MachineBasicBlock *MBB,
7274 bool invSrc) const {
7275 // For the atomic bitwise operator, we generate
7276 // thisMBB (instructions are in pairs, except cmpxchg8b)
7277 // ld t1,t2 = [bitinstr.addr]
7279 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7280 // op t5, t6 <- out1, out2, [bitinstr.val]
7281 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7282 // mov ECX, EBX <- t5, t6
7283 // mov EAX, EDX <- t1, t2
7284 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7285 // mov t3, t4 <- EAX, EDX
7287 // result in out1, out2
7288 // fallthrough -->nextMBB
7290 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7291 const unsigned LoadOpc = X86::MOV32rm;
7292 const unsigned copyOpc = X86::MOV32rr;
7293 const unsigned NotOpc = X86::NOT32r;
7294 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7295 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7296 MachineFunction::iterator MBBIter = MBB;
7299 /// First build the CFG
7300 MachineFunction *F = MBB->getParent();
7301 MachineBasicBlock *thisMBB = MBB;
7302 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7303 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7304 F->insert(MBBIter, newMBB);
7305 F->insert(MBBIter, nextMBB);
7307 // Move all successors to thisMBB to nextMBB
7308 nextMBB->transferSuccessors(thisMBB);
7310 // Update thisMBB to fall through to newMBB
7311 thisMBB->addSuccessor(newMBB);
7313 // newMBB jumps to itself and fall through to nextMBB
7314 newMBB->addSuccessor(nextMBB);
7315 newMBB->addSuccessor(newMBB);
7317 DebugLoc dl = bInstr->getDebugLoc();
7318 // Insert instructions into newMBB based on incoming instruction
7319 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7320 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7321 "unexpected number of operands");
7322 MachineOperand& dest1Oper = bInstr->getOperand(0);
7323 MachineOperand& dest2Oper = bInstr->getOperand(1);
7324 MachineOperand* argOpers[2 + X86AddrNumOperands];
7325 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7326 argOpers[i] = &bInstr->getOperand(i+2);
7328 // x86 address has 4 operands: base, index, scale, and displacement
7329 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7331 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7332 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7333 for (int i=0; i <= lastAddrIndx; ++i)
7334 (*MIB).addOperand(*argOpers[i]);
7335 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7336 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7337 // add 4 to displacement.
7338 for (int i=0; i <= lastAddrIndx-2; ++i)
7339 (*MIB).addOperand(*argOpers[i]);
7340 MachineOperand newOp3 = *(argOpers[3]);
7342 newOp3.setImm(newOp3.getImm()+4);
7344 newOp3.setOffset(newOp3.getOffset()+4);
7345 (*MIB).addOperand(newOp3);
7346 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7348 // t3/4 are defined later, at the bottom of the loop
7349 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7350 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7351 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7352 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7353 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7354 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7356 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7357 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7359 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7360 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7366 int valArgIndx = lastAddrIndx + 1;
7367 assert((argOpers[valArgIndx]->isReg() ||
7368 argOpers[valArgIndx]->isImm()) &&
7370 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7371 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7372 if (argOpers[valArgIndx]->isReg())
7373 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7375 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7376 if (regOpcL != X86::MOV32rr)
7378 (*MIB).addOperand(*argOpers[valArgIndx]);
7379 assert(argOpers[valArgIndx + 1]->isReg() ==
7380 argOpers[valArgIndx]->isReg());
7381 assert(argOpers[valArgIndx + 1]->isImm() ==
7382 argOpers[valArgIndx]->isImm());
7383 if (argOpers[valArgIndx + 1]->isReg())
7384 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7386 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7387 if (regOpcH != X86::MOV32rr)
7389 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7391 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7393 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7396 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7398 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7401 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7402 for (int i=0; i <= lastAddrIndx; ++i)
7403 (*MIB).addOperand(*argOpers[i]);
7405 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7406 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7408 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7409 MIB.addReg(X86::EAX);
7410 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7411 MIB.addReg(X86::EDX);
7414 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7416 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7420 // private utility function
7422 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7423 MachineBasicBlock *MBB,
7424 unsigned cmovOpc) const {
7425 // For the atomic min/max operator, we generate
7428 // ld t1 = [min/max.addr]
7429 // mov t2 = [min/max.val]
7431 // cmov[cond] t2 = t1
7433 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7435 // fallthrough -->nextMBB
7437 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7438 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7439 MachineFunction::iterator MBBIter = MBB;
7442 /// First build the CFG
7443 MachineFunction *F = MBB->getParent();
7444 MachineBasicBlock *thisMBB = MBB;
7445 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7446 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7447 F->insert(MBBIter, newMBB);
7448 F->insert(MBBIter, nextMBB);
7450 // Move all successors to thisMBB to nextMBB
7451 nextMBB->transferSuccessors(thisMBB);
7453 // Update thisMBB to fall through to newMBB
7454 thisMBB->addSuccessor(newMBB);
7456 // newMBB jumps to newMBB and fall through to nextMBB
7457 newMBB->addSuccessor(nextMBB);
7458 newMBB->addSuccessor(newMBB);
7460 DebugLoc dl = mInstr->getDebugLoc();
7461 // Insert instructions into newMBB based on incoming instruction
7462 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7463 "unexpected number of operands");
7464 MachineOperand& destOper = mInstr->getOperand(0);
7465 MachineOperand* argOpers[2 + X86AddrNumOperands];
7466 int numArgs = mInstr->getNumOperands() - 1;
7467 for (int i=0; i < numArgs; ++i)
7468 argOpers[i] = &mInstr->getOperand(i+1);
7470 // x86 address has 4 operands: base, index, scale, and displacement
7471 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7472 int valArgIndx = lastAddrIndx + 1;
7474 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7475 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7476 for (int i=0; i <= lastAddrIndx; ++i)
7477 (*MIB).addOperand(*argOpers[i]);
7479 // We only support register and immediate values
7480 assert((argOpers[valArgIndx]->isReg() ||
7481 argOpers[valArgIndx]->isImm()) &&
7484 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7485 if (argOpers[valArgIndx]->isReg())
7486 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7488 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7489 (*MIB).addOperand(*argOpers[valArgIndx]);
7491 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7494 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7499 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7500 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7504 // Cmp and exchange if none has modified the memory location
7505 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7506 for (int i=0; i <= lastAddrIndx; ++i)
7507 (*MIB).addOperand(*argOpers[i]);
7509 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7510 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7512 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7513 MIB.addReg(X86::EAX);
7516 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7518 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7524 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7525 MachineBasicBlock *BB) const {
7526 DebugLoc dl = MI->getDebugLoc();
7527 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7528 switch (MI->getOpcode()) {
7529 default: assert(false && "Unexpected instr type to insert");
7530 case X86::CMOV_V1I64:
7531 case X86::CMOV_FR32:
7532 case X86::CMOV_FR64:
7533 case X86::CMOV_V4F32:
7534 case X86::CMOV_V2F64:
7535 case X86::CMOV_V2I64: {
7536 // To "insert" a SELECT_CC instruction, we actually have to insert the
7537 // diamond control-flow pattern. The incoming instruction knows the
7538 // destination vreg to set, the condition code register to branch on, the
7539 // true/false values to select between, and a branch opcode to use.
7540 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7541 MachineFunction::iterator It = BB;
7547 // cmpTY ccX, r1, r2
7549 // fallthrough --> copy0MBB
7550 MachineBasicBlock *thisMBB = BB;
7551 MachineFunction *F = BB->getParent();
7552 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7553 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7555 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7556 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7557 F->insert(It, copy0MBB);
7558 F->insert(It, sinkMBB);
7559 // Update machine-CFG edges by transferring all successors of the current
7560 // block to the new block which will contain the Phi node for the select.
7561 sinkMBB->transferSuccessors(BB);
7563 // Add the true and fallthrough blocks as its successors.
7564 BB->addSuccessor(copy0MBB);
7565 BB->addSuccessor(sinkMBB);
7568 // %FalseValue = ...
7569 // # fallthrough to sinkMBB
7572 // Update machine-CFG edges
7573 BB->addSuccessor(sinkMBB);
7576 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7579 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7580 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7581 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7583 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7587 case X86::FP32_TO_INT16_IN_MEM:
7588 case X86::FP32_TO_INT32_IN_MEM:
7589 case X86::FP32_TO_INT64_IN_MEM:
7590 case X86::FP64_TO_INT16_IN_MEM:
7591 case X86::FP64_TO_INT32_IN_MEM:
7592 case X86::FP64_TO_INT64_IN_MEM:
7593 case X86::FP80_TO_INT16_IN_MEM:
7594 case X86::FP80_TO_INT32_IN_MEM:
7595 case X86::FP80_TO_INT64_IN_MEM: {
7596 // Change the floating point control register to use "round towards zero"
7597 // mode when truncating to an integer value.
7598 MachineFunction *F = BB->getParent();
7599 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7600 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7602 // Load the old value of the high byte of the control word...
7604 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7605 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7608 // Set the high part to be round to zero...
7609 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7612 // Reload the modified control word now...
7613 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7615 // Restore the memory image of control word to original value
7616 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7619 // Get the X86 opcode to use.
7621 switch (MI->getOpcode()) {
7622 default: assert(0 && "illegal opcode!");
7623 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7624 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7625 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7626 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7627 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7628 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7629 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7630 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7631 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7635 MachineOperand &Op = MI->getOperand(0);
7637 AM.BaseType = X86AddressMode::RegBase;
7638 AM.Base.Reg = Op.getReg();
7640 AM.BaseType = X86AddressMode::FrameIndexBase;
7641 AM.Base.FrameIndex = Op.getIndex();
7643 Op = MI->getOperand(1);
7645 AM.Scale = Op.getImm();
7646 Op = MI->getOperand(2);
7648 AM.IndexReg = Op.getImm();
7649 Op = MI->getOperand(3);
7650 if (Op.isGlobal()) {
7651 AM.GV = Op.getGlobal();
7653 AM.Disp = Op.getImm();
7655 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7656 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7658 // Reload the original control word now.
7659 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7661 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7664 case X86::ATOMAND32:
7665 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7666 X86::AND32ri, X86::MOV32rm,
7667 X86::LCMPXCHG32, X86::MOV32rr,
7668 X86::NOT32r, X86::EAX,
7669 X86::GR32RegisterClass);
7671 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7672 X86::OR32ri, X86::MOV32rm,
7673 X86::LCMPXCHG32, X86::MOV32rr,
7674 X86::NOT32r, X86::EAX,
7675 X86::GR32RegisterClass);
7676 case X86::ATOMXOR32:
7677 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7678 X86::XOR32ri, X86::MOV32rm,
7679 X86::LCMPXCHG32, X86::MOV32rr,
7680 X86::NOT32r, X86::EAX,
7681 X86::GR32RegisterClass);
7682 case X86::ATOMNAND32:
7683 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7684 X86::AND32ri, X86::MOV32rm,
7685 X86::LCMPXCHG32, X86::MOV32rr,
7686 X86::NOT32r, X86::EAX,
7687 X86::GR32RegisterClass, true);
7688 case X86::ATOMMIN32:
7689 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7690 case X86::ATOMMAX32:
7691 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7692 case X86::ATOMUMIN32:
7693 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7694 case X86::ATOMUMAX32:
7695 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7697 case X86::ATOMAND16:
7698 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7699 X86::AND16ri, X86::MOV16rm,
7700 X86::LCMPXCHG16, X86::MOV16rr,
7701 X86::NOT16r, X86::AX,
7702 X86::GR16RegisterClass);
7704 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7705 X86::OR16ri, X86::MOV16rm,
7706 X86::LCMPXCHG16, X86::MOV16rr,
7707 X86::NOT16r, X86::AX,
7708 X86::GR16RegisterClass);
7709 case X86::ATOMXOR16:
7710 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7711 X86::XOR16ri, X86::MOV16rm,
7712 X86::LCMPXCHG16, X86::MOV16rr,
7713 X86::NOT16r, X86::AX,
7714 X86::GR16RegisterClass);
7715 case X86::ATOMNAND16:
7716 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7717 X86::AND16ri, X86::MOV16rm,
7718 X86::LCMPXCHG16, X86::MOV16rr,
7719 X86::NOT16r, X86::AX,
7720 X86::GR16RegisterClass, true);
7721 case X86::ATOMMIN16:
7722 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7723 case X86::ATOMMAX16:
7724 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7725 case X86::ATOMUMIN16:
7726 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7727 case X86::ATOMUMAX16:
7728 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7731 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7732 X86::AND8ri, X86::MOV8rm,
7733 X86::LCMPXCHG8, X86::MOV8rr,
7734 X86::NOT8r, X86::AL,
7735 X86::GR8RegisterClass);
7737 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7738 X86::OR8ri, X86::MOV8rm,
7739 X86::LCMPXCHG8, X86::MOV8rr,
7740 X86::NOT8r, X86::AL,
7741 X86::GR8RegisterClass);
7743 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7744 X86::XOR8ri, X86::MOV8rm,
7745 X86::LCMPXCHG8, X86::MOV8rr,
7746 X86::NOT8r, X86::AL,
7747 X86::GR8RegisterClass);
7748 case X86::ATOMNAND8:
7749 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7750 X86::AND8ri, X86::MOV8rm,
7751 X86::LCMPXCHG8, X86::MOV8rr,
7752 X86::NOT8r, X86::AL,
7753 X86::GR8RegisterClass, true);
7754 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7755 // This group is for 64-bit host.
7756 case X86::ATOMAND64:
7757 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7758 X86::AND64ri32, X86::MOV64rm,
7759 X86::LCMPXCHG64, X86::MOV64rr,
7760 X86::NOT64r, X86::RAX,
7761 X86::GR64RegisterClass);
7763 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7764 X86::OR64ri32, X86::MOV64rm,
7765 X86::LCMPXCHG64, X86::MOV64rr,
7766 X86::NOT64r, X86::RAX,
7767 X86::GR64RegisterClass);
7768 case X86::ATOMXOR64:
7769 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7770 X86::XOR64ri32, X86::MOV64rm,
7771 X86::LCMPXCHG64, X86::MOV64rr,
7772 X86::NOT64r, X86::RAX,
7773 X86::GR64RegisterClass);
7774 case X86::ATOMNAND64:
7775 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7776 X86::AND64ri32, X86::MOV64rm,
7777 X86::LCMPXCHG64, X86::MOV64rr,
7778 X86::NOT64r, X86::RAX,
7779 X86::GR64RegisterClass, true);
7780 case X86::ATOMMIN64:
7781 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7782 case X86::ATOMMAX64:
7783 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7784 case X86::ATOMUMIN64:
7785 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7786 case X86::ATOMUMAX64:
7787 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7789 // This group does 64-bit operations on a 32-bit host.
7790 case X86::ATOMAND6432:
7791 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7792 X86::AND32rr, X86::AND32rr,
7793 X86::AND32ri, X86::AND32ri,
7795 case X86::ATOMOR6432:
7796 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7797 X86::OR32rr, X86::OR32rr,
7798 X86::OR32ri, X86::OR32ri,
7800 case X86::ATOMXOR6432:
7801 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7802 X86::XOR32rr, X86::XOR32rr,
7803 X86::XOR32ri, X86::XOR32ri,
7805 case X86::ATOMNAND6432:
7806 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7807 X86::AND32rr, X86::AND32rr,
7808 X86::AND32ri, X86::AND32ri,
7810 case X86::ATOMADD6432:
7811 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7812 X86::ADD32rr, X86::ADC32rr,
7813 X86::ADD32ri, X86::ADC32ri,
7815 case X86::ATOMSUB6432:
7816 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7817 X86::SUB32rr, X86::SBB32rr,
7818 X86::SUB32ri, X86::SBB32ri,
7820 case X86::ATOMSWAP6432:
7821 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7822 X86::MOV32rr, X86::MOV32rr,
7823 X86::MOV32ri, X86::MOV32ri,
7828 //===----------------------------------------------------------------------===//
7829 // X86 Optimization Hooks
7830 //===----------------------------------------------------------------------===//
7832 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7836 const SelectionDAG &DAG,
7837 unsigned Depth) const {
7838 unsigned Opc = Op.getOpcode();
7839 assert((Opc >= ISD::BUILTIN_OP_END ||
7840 Opc == ISD::INTRINSIC_WO_CHAIN ||
7841 Opc == ISD::INTRINSIC_W_CHAIN ||
7842 Opc == ISD::INTRINSIC_VOID) &&
7843 "Should use MaskedValueIsZero if you don't know whether Op"
7844 " is a target node!");
7846 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7855 // These nodes' second result is a boolean.
7856 if (Op.getResNo() == 0)
7860 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7861 Mask.getBitWidth() - 1);
7866 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7867 /// node is a GlobalAddress + offset.
7868 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7869 GlobalValue* &GA, int64_t &Offset) const{
7870 if (N->getOpcode() == X86ISD::Wrapper) {
7871 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7872 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7873 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7877 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7880 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7881 const TargetLowering &TLI) {
7884 if (TLI.isGAPlusOffset(Base, GV, Offset))
7885 return (GV->getAlignment() >= N && (Offset % N) == 0);
7886 // DAG combine handles the stack object case.
7890 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7891 MVT EVT, LoadSDNode *&LDBase,
7892 unsigned &LastLoadedElt,
7893 SelectionDAG &DAG, MachineFrameInfo *MFI,
7894 const TargetLowering &TLI) {
7896 LastLoadedElt = -1U;
7897 for (unsigned i = 0; i < NumElems; ++i) {
7898 if (N->getMaskElt(i) < 0) {
7904 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7905 if (!Elt.getNode() ||
7906 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7909 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7911 LDBase = cast<LoadSDNode>(Elt.getNode());
7915 if (Elt.getOpcode() == ISD::UNDEF)
7918 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7919 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7926 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7927 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7928 /// if the load addresses are consecutive, non-overlapping, and in the right
7929 /// order. In the case of v2i64, it will see if it can rewrite the
7930 /// shuffle to be an appropriate build vector so it can take advantage of
7931 // performBuildVectorCombine.
7932 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7933 const TargetLowering &TLI) {
7934 DebugLoc dl = N->getDebugLoc();
7935 MVT VT = N->getValueType(0);
7936 MVT EVT = VT.getVectorElementType();
7937 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7938 unsigned NumElems = VT.getVectorNumElements();
7940 if (VT.getSizeInBits() != 128)
7943 // Try to combine a vector_shuffle into a 128-bit load.
7944 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7945 LoadSDNode *LD = NULL;
7946 unsigned LastLoadedElt;
7947 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7951 if (LastLoadedElt == NumElems - 1) {
7952 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7953 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7954 LD->getSrcValue(), LD->getSrcValueOffset(),
7956 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7957 LD->getSrcValue(), LD->getSrcValueOffset(),
7958 LD->isVolatile(), LD->getAlignment());
7959 } else if (NumElems == 4 && LastLoadedElt == 1) {
7960 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7961 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7962 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7963 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
7968 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7969 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7970 const X86Subtarget *Subtarget) {
7971 DebugLoc DL = N->getDebugLoc();
7972 SDValue Cond = N->getOperand(0);
7973 // Get the LHS/RHS of the select.
7974 SDValue LHS = N->getOperand(1);
7975 SDValue RHS = N->getOperand(2);
7977 // If we have SSE[12] support, try to form min/max nodes.
7978 if (Subtarget->hasSSE2() &&
7979 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7980 Cond.getOpcode() == ISD::SETCC) {
7981 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7983 unsigned Opcode = 0;
7984 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7987 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7990 if (!UnsafeFPMath) break;
7992 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7994 Opcode = X86ISD::FMIN;
7997 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8000 if (!UnsafeFPMath) break;
8002 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8004 Opcode = X86ISD::FMAX;
8007 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8010 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8013 if (!UnsafeFPMath) break;
8015 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8017 Opcode = X86ISD::FMIN;
8020 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8023 if (!UnsafeFPMath) break;
8025 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8027 Opcode = X86ISD::FMAX;
8033 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8036 // If this is a select between two integer constants, try to do some
8038 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8039 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8040 // Don't do this for crazy integer types.
8041 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8042 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8043 // so that TrueC (the true value) is larger than FalseC.
8044 bool NeedsCondInvert = false;
8046 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8047 // Efficiently invertible.
8048 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8049 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8050 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8051 NeedsCondInvert = true;
8052 std::swap(TrueC, FalseC);
8055 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8056 if (FalseC->getAPIntValue() == 0 &&
8057 TrueC->getAPIntValue().isPowerOf2()) {
8058 if (NeedsCondInvert) // Invert the condition if needed.
8059 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8060 DAG.getConstant(1, Cond.getValueType()));
8062 // Zero extend the condition if needed.
8063 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8065 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8066 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8067 DAG.getConstant(ShAmt, MVT::i8));
8070 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8071 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8072 if (NeedsCondInvert) // Invert the condition if needed.
8073 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8074 DAG.getConstant(1, Cond.getValueType()));
8076 // Zero extend the condition if needed.
8077 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8078 FalseC->getValueType(0), Cond);
8079 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8080 SDValue(FalseC, 0));
8083 // Optimize cases that will turn into an LEA instruction. This requires
8084 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8085 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8086 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8087 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8089 bool isFastMultiplier = false;
8091 switch ((unsigned char)Diff) {
8093 case 1: // result = add base, cond
8094 case 2: // result = lea base( , cond*2)
8095 case 3: // result = lea base(cond, cond*2)
8096 case 4: // result = lea base( , cond*4)
8097 case 5: // result = lea base(cond, cond*4)
8098 case 8: // result = lea base( , cond*8)
8099 case 9: // result = lea base(cond, cond*8)
8100 isFastMultiplier = true;
8105 if (isFastMultiplier) {
8106 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8107 if (NeedsCondInvert) // Invert the condition if needed.
8108 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8109 DAG.getConstant(1, Cond.getValueType()));
8111 // Zero extend the condition if needed.
8112 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8114 // Scale the condition by the difference.
8116 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8117 DAG.getConstant(Diff, Cond.getValueType()));
8119 // Add the base if non-zero.
8120 if (FalseC->getAPIntValue() != 0)
8121 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8122 SDValue(FalseC, 0));
8132 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8133 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8134 TargetLowering::DAGCombinerInfo &DCI) {
8135 DebugLoc DL = N->getDebugLoc();
8137 // If the flag operand isn't dead, don't touch this CMOV.
8138 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8141 // If this is a select between two integer constants, try to do some
8142 // optimizations. Note that the operands are ordered the opposite of SELECT
8144 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8145 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8146 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8147 // larger than FalseC (the false value).
8148 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8150 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8151 CC = X86::GetOppositeBranchCondition(CC);
8152 std::swap(TrueC, FalseC);
8155 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8156 // This is efficient for any integer data type (including i8/i16) and
8158 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8159 SDValue Cond = N->getOperand(3);
8160 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8161 DAG.getConstant(CC, MVT::i8), Cond);
8163 // Zero extend the condition if needed.
8164 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8166 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8167 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8168 DAG.getConstant(ShAmt, MVT::i8));
8169 if (N->getNumValues() == 2) // Dead flag value?
8170 return DCI.CombineTo(N, Cond, SDValue());
8174 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8175 // for any integer data type, including i8/i16.
8176 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8177 SDValue Cond = N->getOperand(3);
8178 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8179 DAG.getConstant(CC, MVT::i8), Cond);
8181 // Zero extend the condition if needed.
8182 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8183 FalseC->getValueType(0), Cond);
8184 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8185 SDValue(FalseC, 0));
8187 if (N->getNumValues() == 2) // Dead flag value?
8188 return DCI.CombineTo(N, Cond, SDValue());
8192 // Optimize cases that will turn into an LEA instruction. This requires
8193 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8194 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8195 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8196 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8198 bool isFastMultiplier = false;
8200 switch ((unsigned char)Diff) {
8202 case 1: // result = add base, cond
8203 case 2: // result = lea base( , cond*2)
8204 case 3: // result = lea base(cond, cond*2)
8205 case 4: // result = lea base( , cond*4)
8206 case 5: // result = lea base(cond, cond*4)
8207 case 8: // result = lea base( , cond*8)
8208 case 9: // result = lea base(cond, cond*8)
8209 isFastMultiplier = true;
8214 if (isFastMultiplier) {
8215 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8216 SDValue Cond = N->getOperand(3);
8217 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8218 DAG.getConstant(CC, MVT::i8), Cond);
8219 // Zero extend the condition if needed.
8220 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8222 // Scale the condition by the difference.
8224 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8225 DAG.getConstant(Diff, Cond.getValueType()));
8227 // Add the base if non-zero.
8228 if (FalseC->getAPIntValue() != 0)
8229 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8230 SDValue(FalseC, 0));
8231 if (N->getNumValues() == 2) // Dead flag value?
8232 return DCI.CombineTo(N, Cond, SDValue());
8242 /// PerformMulCombine - Optimize a single multiply with constant into two
8243 /// in order to implement it with two cheaper instructions, e.g.
8244 /// LEA + SHL, LEA + LEA.
8245 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8246 TargetLowering::DAGCombinerInfo &DCI) {
8247 if (DAG.getMachineFunction().
8248 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8251 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8254 MVT VT = N->getValueType(0);
8258 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8261 uint64_t MulAmt = C->getZExtValue();
8262 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8265 uint64_t MulAmt1 = 0;
8266 uint64_t MulAmt2 = 0;
8267 if ((MulAmt % 9) == 0) {
8269 MulAmt2 = MulAmt / 9;
8270 } else if ((MulAmt % 5) == 0) {
8272 MulAmt2 = MulAmt / 5;
8273 } else if ((MulAmt % 3) == 0) {
8275 MulAmt2 = MulAmt / 3;
8278 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8279 DebugLoc DL = N->getDebugLoc();
8281 if (isPowerOf2_64(MulAmt2) &&
8282 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8283 // If second multiplifer is pow2, issue it first. We want the multiply by
8284 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8286 std::swap(MulAmt1, MulAmt2);
8289 if (isPowerOf2_64(MulAmt1))
8290 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8291 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8293 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8294 DAG.getConstant(MulAmt1, VT));
8296 if (isPowerOf2_64(MulAmt2))
8297 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8298 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8300 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8301 DAG.getConstant(MulAmt2, VT));
8303 // Do not add new nodes to DAG combiner worklist.
8304 DCI.CombineTo(N, NewMul, false);
8310 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8312 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8313 const X86Subtarget *Subtarget) {
8314 // On X86 with SSE2 support, we can transform this to a vector shift if
8315 // all elements are shifted by the same amount. We can't do this in legalize
8316 // because the a constant vector is typically transformed to a constant pool
8317 // so we have no knowledge of the shift amount.
8318 if (!Subtarget->hasSSE2())
8321 MVT VT = N->getValueType(0);
8322 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8325 SDValue ShAmtOp = N->getOperand(1);
8326 MVT EltVT = VT.getVectorElementType();
8327 DebugLoc DL = N->getDebugLoc();
8329 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8330 unsigned NumElts = VT.getVectorNumElements();
8332 for (; i != NumElts; ++i) {
8333 SDValue Arg = ShAmtOp.getOperand(i);
8334 if (Arg.getOpcode() == ISD::UNDEF) continue;
8338 for (; i != NumElts; ++i) {
8339 SDValue Arg = ShAmtOp.getOperand(i);
8340 if (Arg.getOpcode() == ISD::UNDEF) continue;
8341 if (Arg != BaseShAmt) {
8345 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8346 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8347 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8348 DAG.getIntPtrConstant(0));
8352 if (EltVT.bitsGT(MVT::i32))
8353 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8354 else if (EltVT.bitsLT(MVT::i32))
8355 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8357 // The shift amount is identical so we can do a vector shift.
8358 SDValue ValOp = N->getOperand(0);
8359 switch (N->getOpcode()) {
8361 assert(0 && "Unknown shift opcode!");
8364 if (VT == MVT::v2i64)
8365 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8366 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8368 if (VT == MVT::v4i32)
8369 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8370 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8372 if (VT == MVT::v8i16)
8373 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8374 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8378 if (VT == MVT::v4i32)
8379 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8380 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8382 if (VT == MVT::v8i16)
8383 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8384 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8388 if (VT == MVT::v2i64)
8389 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8390 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8392 if (VT == MVT::v4i32)
8393 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8394 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8396 if (VT == MVT::v8i16)
8397 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8398 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8405 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8406 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8407 const X86Subtarget *Subtarget) {
8408 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8409 // the FP state in cases where an emms may be missing.
8410 // A preferable solution to the general problem is to figure out the right
8411 // places to insert EMMS. This qualifies as a quick hack.
8413 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8414 StoreSDNode *St = cast<StoreSDNode>(N);
8415 MVT VT = St->getValue().getValueType();
8416 if (VT.getSizeInBits() != 64)
8419 const Function *F = DAG.getMachineFunction().getFunction();
8420 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8421 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8422 && Subtarget->hasSSE2();
8423 if ((VT.isVector() ||
8424 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8425 isa<LoadSDNode>(St->getValue()) &&
8426 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8427 St->getChain().hasOneUse() && !St->isVolatile()) {
8428 SDNode* LdVal = St->getValue().getNode();
8430 int TokenFactorIndex = -1;
8431 SmallVector<SDValue, 8> Ops;
8432 SDNode* ChainVal = St->getChain().getNode();
8433 // Must be a store of a load. We currently handle two cases: the load
8434 // is a direct child, and it's under an intervening TokenFactor. It is
8435 // possible to dig deeper under nested TokenFactors.
8436 if (ChainVal == LdVal)
8437 Ld = cast<LoadSDNode>(St->getChain());
8438 else if (St->getValue().hasOneUse() &&
8439 ChainVal->getOpcode() == ISD::TokenFactor) {
8440 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8441 if (ChainVal->getOperand(i).getNode() == LdVal) {
8442 TokenFactorIndex = i;
8443 Ld = cast<LoadSDNode>(St->getValue());
8445 Ops.push_back(ChainVal->getOperand(i));
8449 if (!Ld || !ISD::isNormalLoad(Ld))
8452 // If this is not the MMX case, i.e. we are just turning i64 load/store
8453 // into f64 load/store, avoid the transformation if there are multiple
8454 // uses of the loaded value.
8455 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8458 DebugLoc LdDL = Ld->getDebugLoc();
8459 DebugLoc StDL = N->getDebugLoc();
8460 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8461 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8463 if (Subtarget->is64Bit() || F64IsLegal) {
8464 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8465 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8466 Ld->getBasePtr(), Ld->getSrcValue(),
8467 Ld->getSrcValueOffset(), Ld->isVolatile(),
8468 Ld->getAlignment());
8469 SDValue NewChain = NewLd.getValue(1);
8470 if (TokenFactorIndex != -1) {
8471 Ops.push_back(NewChain);
8472 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8475 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8476 St->getSrcValue(), St->getSrcValueOffset(),
8477 St->isVolatile(), St->getAlignment());
8480 // Otherwise, lower to two pairs of 32-bit loads / stores.
8481 SDValue LoAddr = Ld->getBasePtr();
8482 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8483 DAG.getConstant(4, MVT::i32));
8485 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8486 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8487 Ld->isVolatile(), Ld->getAlignment());
8488 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8489 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8491 MinAlign(Ld->getAlignment(), 4));
8493 SDValue NewChain = LoLd.getValue(1);
8494 if (TokenFactorIndex != -1) {
8495 Ops.push_back(LoLd);
8496 Ops.push_back(HiLd);
8497 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8501 LoAddr = St->getBasePtr();
8502 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8503 DAG.getConstant(4, MVT::i32));
8505 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8506 St->getSrcValue(), St->getSrcValueOffset(),
8507 St->isVolatile(), St->getAlignment());
8508 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8510 St->getSrcValueOffset() + 4,
8512 MinAlign(St->getAlignment(), 4));
8513 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8518 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8519 /// X86ISD::FXOR nodes.
8520 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8521 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8522 // F[X]OR(0.0, x) -> x
8523 // F[X]OR(x, 0.0) -> x
8524 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8525 if (C->getValueAPF().isPosZero())
8526 return N->getOperand(1);
8527 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8528 if (C->getValueAPF().isPosZero())
8529 return N->getOperand(0);
8533 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8534 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8535 // FAND(0.0, x) -> 0.0
8536 // FAND(x, 0.0) -> 0.0
8537 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8538 if (C->getValueAPF().isPosZero())
8539 return N->getOperand(0);
8540 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8541 if (C->getValueAPF().isPosZero())
8542 return N->getOperand(1);
8546 static SDValue PerformBTCombine(SDNode *N,
8548 TargetLowering::DAGCombinerInfo &DCI) {
8549 // BT ignores high bits in the bit index operand.
8550 SDValue Op1 = N->getOperand(1);
8551 if (Op1.hasOneUse()) {
8552 unsigned BitWidth = Op1.getValueSizeInBits();
8553 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8554 APInt KnownZero, KnownOne;
8555 TargetLowering::TargetLoweringOpt TLO(DAG);
8556 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8557 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8558 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8559 DCI.CommitTargetLoweringOpt(TLO);
8564 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8565 SDValue Op = N->getOperand(0);
8566 if (Op.getOpcode() == ISD::BIT_CONVERT)
8567 Op = Op.getOperand(0);
8568 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8569 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8570 VT.getVectorElementType().getSizeInBits() ==
8571 OpVT.getVectorElementType().getSizeInBits()) {
8572 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8577 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8578 // Locked instructions, in turn, have implicit fence semantics (all memory
8579 // operations are flushed before issuing the locked instruction, and the
8580 // are not buffered), so we can fold away the common pattern of
8581 // fence-atomic-fence.
8582 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8583 SDValue atomic = N->getOperand(0);
8584 switch (atomic.getOpcode()) {
8585 case ISD::ATOMIC_CMP_SWAP:
8586 case ISD::ATOMIC_SWAP:
8587 case ISD::ATOMIC_LOAD_ADD:
8588 case ISD::ATOMIC_LOAD_SUB:
8589 case ISD::ATOMIC_LOAD_AND:
8590 case ISD::ATOMIC_LOAD_OR:
8591 case ISD::ATOMIC_LOAD_XOR:
8592 case ISD::ATOMIC_LOAD_NAND:
8593 case ISD::ATOMIC_LOAD_MIN:
8594 case ISD::ATOMIC_LOAD_MAX:
8595 case ISD::ATOMIC_LOAD_UMIN:
8596 case ISD::ATOMIC_LOAD_UMAX:
8602 SDValue fence = atomic.getOperand(0);
8603 if (fence.getOpcode() != ISD::MEMBARRIER)
8606 switch (atomic.getOpcode()) {
8607 case ISD::ATOMIC_CMP_SWAP:
8608 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8609 atomic.getOperand(1), atomic.getOperand(2),
8610 atomic.getOperand(3));
8611 case ISD::ATOMIC_SWAP:
8612 case ISD::ATOMIC_LOAD_ADD:
8613 case ISD::ATOMIC_LOAD_SUB:
8614 case ISD::ATOMIC_LOAD_AND:
8615 case ISD::ATOMIC_LOAD_OR:
8616 case ISD::ATOMIC_LOAD_XOR:
8617 case ISD::ATOMIC_LOAD_NAND:
8618 case ISD::ATOMIC_LOAD_MIN:
8619 case ISD::ATOMIC_LOAD_MAX:
8620 case ISD::ATOMIC_LOAD_UMIN:
8621 case ISD::ATOMIC_LOAD_UMAX:
8622 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8623 atomic.getOperand(1), atomic.getOperand(2));
8629 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8630 DAGCombinerInfo &DCI) const {
8631 SelectionDAG &DAG = DCI.DAG;
8632 switch (N->getOpcode()) {
8634 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8635 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8636 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8637 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8640 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8641 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8643 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8644 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8645 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8646 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8647 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8653 //===----------------------------------------------------------------------===//
8654 // X86 Inline Assembly Support
8655 //===----------------------------------------------------------------------===//
8657 /// getConstraintType - Given a constraint letter, return the type of
8658 /// constraint it is for this target.
8659 X86TargetLowering::ConstraintType
8660 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8661 if (Constraint.size() == 1) {
8662 switch (Constraint[0]) {
8674 return C_RegisterClass;
8682 return TargetLowering::getConstraintType(Constraint);
8685 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8686 /// with another that has more specific requirements based on the type of the
8687 /// corresponding operand.
8688 const char *X86TargetLowering::
8689 LowerXConstraint(MVT ConstraintVT) const {
8690 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8691 // 'f' like normal targets.
8692 if (ConstraintVT.isFloatingPoint()) {
8693 if (Subtarget->hasSSE2())
8695 if (Subtarget->hasSSE1())
8699 return TargetLowering::LowerXConstraint(ConstraintVT);
8702 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8703 /// vector. If it is invalid, don't add anything to Ops.
8704 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8707 std::vector<SDValue>&Ops,
8708 SelectionDAG &DAG) const {
8709 SDValue Result(0, 0);
8711 switch (Constraint) {
8714 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8715 if (C->getZExtValue() <= 31) {
8716 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8722 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8723 if (C->getZExtValue() <= 63) {
8724 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8730 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8731 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8732 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8738 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8739 if (C->getZExtValue() <= 255) {
8740 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8746 // 32-bit signed value
8747 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8748 const ConstantInt *CI = C->getConstantIntValue();
8749 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8750 // Widen to 64 bits here to get it sign extended.
8751 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8754 // FIXME gcc accepts some relocatable values here too, but only in certain
8755 // memory models; it's complicated.
8760 // 32-bit unsigned value
8761 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8762 const ConstantInt *CI = C->getConstantIntValue();
8763 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8764 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8768 // FIXME gcc accepts some relocatable values here too, but only in certain
8769 // memory models; it's complicated.
8773 // Literal immediates are always ok.
8774 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8775 // Widen to 64 bits here to get it sign extended.
8776 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8780 // If we are in non-pic codegen mode, we allow the address of a global (with
8781 // an optional displacement) to be used with 'i'.
8782 GlobalAddressSDNode *GA = 0;
8785 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8787 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8788 Offset += GA->getOffset();
8790 } else if (Op.getOpcode() == ISD::ADD) {
8791 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8792 Offset += C->getZExtValue();
8793 Op = Op.getOperand(0);
8796 } else if (Op.getOpcode() == ISD::SUB) {
8797 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8798 Offset += -C->getZExtValue();
8799 Op = Op.getOperand(0);
8804 // Otherwise, this isn't something we can handle, reject it.
8807 // If we require an extra load to get this address, as in PIC mode, we
8809 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(),
8810 getTargetMachine(), false))
8814 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
8816 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8823 if (Result.getNode()) {
8824 Ops.push_back(Result);
8827 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8831 std::vector<unsigned> X86TargetLowering::
8832 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8834 if (Constraint.size() == 1) {
8835 // FIXME: not handling fp-stack yet!
8836 switch (Constraint[0]) { // GCC X86 Constraint Letters
8837 default: break; // Unknown constraint letter
8838 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8841 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8842 else if (VT == MVT::i16)
8843 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8844 else if (VT == MVT::i8)
8845 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8846 else if (VT == MVT::i64)
8847 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8852 return std::vector<unsigned>();
8855 std::pair<unsigned, const TargetRegisterClass*>
8856 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8858 // First, see if this is a constraint that directly corresponds to an LLVM
8860 if (Constraint.size() == 1) {
8861 // GCC Constraint Letters
8862 switch (Constraint[0]) {
8864 case 'r': // GENERAL_REGS
8865 case 'R': // LEGACY_REGS
8866 case 'l': // INDEX_REGS
8868 return std::make_pair(0U, X86::GR8RegisterClass);
8870 return std::make_pair(0U, X86::GR16RegisterClass);
8871 if (VT == MVT::i32 || !Subtarget->is64Bit())
8872 return std::make_pair(0U, X86::GR32RegisterClass);
8873 return std::make_pair(0U, X86::GR64RegisterClass);
8874 case 'f': // FP Stack registers.
8875 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8876 // value to the correct fpstack register class.
8877 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8878 return std::make_pair(0U, X86::RFP32RegisterClass);
8879 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8880 return std::make_pair(0U, X86::RFP64RegisterClass);
8881 return std::make_pair(0U, X86::RFP80RegisterClass);
8882 case 'y': // MMX_REGS if MMX allowed.
8883 if (!Subtarget->hasMMX()) break;
8884 return std::make_pair(0U, X86::VR64RegisterClass);
8885 case 'Y': // SSE_REGS if SSE2 allowed
8886 if (!Subtarget->hasSSE2()) break;
8888 case 'x': // SSE_REGS if SSE1 allowed
8889 if (!Subtarget->hasSSE1()) break;
8891 switch (VT.getSimpleVT()) {
8893 // Scalar SSE types.
8896 return std::make_pair(0U, X86::FR32RegisterClass);
8899 return std::make_pair(0U, X86::FR64RegisterClass);
8907 return std::make_pair(0U, X86::VR128RegisterClass);
8913 // Use the default implementation in TargetLowering to convert the register
8914 // constraint into a member of a register class.
8915 std::pair<unsigned, const TargetRegisterClass*> Res;
8916 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8918 // Not found as a standard register?
8919 if (Res.second == 0) {
8920 // GCC calls "st(0)" just plain "st".
8921 if (StringsEqualNoCase("{st}", Constraint)) {
8922 Res.first = X86::ST0;
8923 Res.second = X86::RFP80RegisterClass;
8925 // 'A' means EAX + EDX.
8926 if (Constraint == "A") {
8927 Res.first = X86::EAX;
8928 Res.second = X86::GRADRegisterClass;
8933 // Otherwise, check to see if this is a register class of the wrong value
8934 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8935 // turn into {ax},{dx}.
8936 if (Res.second->hasType(VT))
8937 return Res; // Correct type already, nothing to do.
8939 // All of the single-register GCC register classes map their values onto
8940 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8941 // really want an 8-bit or 32-bit register, map to the appropriate register
8942 // class and return the appropriate register.
8943 if (Res.second == X86::GR16RegisterClass) {
8944 if (VT == MVT::i8) {
8945 unsigned DestReg = 0;
8946 switch (Res.first) {
8948 case X86::AX: DestReg = X86::AL; break;
8949 case X86::DX: DestReg = X86::DL; break;
8950 case X86::CX: DestReg = X86::CL; break;
8951 case X86::BX: DestReg = X86::BL; break;
8954 Res.first = DestReg;
8955 Res.second = X86::GR8RegisterClass;
8957 } else if (VT == MVT::i32) {
8958 unsigned DestReg = 0;
8959 switch (Res.first) {
8961 case X86::AX: DestReg = X86::EAX; break;
8962 case X86::DX: DestReg = X86::EDX; break;
8963 case X86::CX: DestReg = X86::ECX; break;
8964 case X86::BX: DestReg = X86::EBX; break;
8965 case X86::SI: DestReg = X86::ESI; break;
8966 case X86::DI: DestReg = X86::EDI; break;
8967 case X86::BP: DestReg = X86::EBP; break;
8968 case X86::SP: DestReg = X86::ESP; break;
8971 Res.first = DestReg;
8972 Res.second = X86::GR32RegisterClass;
8974 } else if (VT == MVT::i64) {
8975 unsigned DestReg = 0;
8976 switch (Res.first) {
8978 case X86::AX: DestReg = X86::RAX; break;
8979 case X86::DX: DestReg = X86::RDX; break;
8980 case X86::CX: DestReg = X86::RCX; break;
8981 case X86::BX: DestReg = X86::RBX; break;
8982 case X86::SI: DestReg = X86::RSI; break;
8983 case X86::DI: DestReg = X86::RDI; break;
8984 case X86::BP: DestReg = X86::RBP; break;
8985 case X86::SP: DestReg = X86::RSP; break;
8988 Res.first = DestReg;
8989 Res.second = X86::GR64RegisterClass;
8992 } else if (Res.second == X86::FR32RegisterClass ||
8993 Res.second == X86::FR64RegisterClass ||
8994 Res.second == X86::VR128RegisterClass) {
8995 // Handle references to XMM physical registers that got mapped into the
8996 // wrong class. This can happen with constraints like {xmm0} where the
8997 // target independent register mapper will just pick the first match it can
8998 // find, ignoring the required type.
9000 Res.second = X86::FR32RegisterClass;
9001 else if (VT == MVT::f64)
9002 Res.second = X86::FR64RegisterClass;
9003 else if (X86::VR128RegisterClass->hasType(VT))
9004 Res.second = X86::VR128RegisterClass;
9010 //===----------------------------------------------------------------------===//
9011 // X86 Widen vector type
9012 //===----------------------------------------------------------------------===//
9014 /// getWidenVectorType: given a vector type, returns the type to widen
9015 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9016 /// If there is no vector type that we want to widen to, returns MVT::Other
9017 /// When and where to widen is target dependent based on the cost of
9018 /// scalarizing vs using the wider vector type.
9020 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9021 assert(VT.isVector());
9022 if (isTypeLegal(VT))
9025 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9026 // type based on element type. This would speed up our search (though
9027 // it may not be worth it since the size of the list is relatively
9029 MVT EltVT = VT.getVectorElementType();
9030 unsigned NElts = VT.getVectorNumElements();
9032 // On X86, it make sense to widen any vector wider than 1
9036 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9037 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9038 MVT SVT = (MVT::SimpleValueType)nVT;
9040 if (isTypeLegal(SVT) &&
9041 SVT.getVectorElementType() == EltVT &&
9042 SVT.getVectorNumElements() > NElts)