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, DAG.getContext());
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(),
1178 RVLocs, DAG.getContext());
1179 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1181 SmallVector<SDValue, 8> ResultVals;
1183 // Copy all of the result registers out of their specified physreg.
1184 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1185 CCValAssign &VA = RVLocs[i];
1186 MVT CopyVT = VA.getValVT();
1188 // If this is x86-64, and we disabled SSE, we can't return FP values
1189 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1190 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1191 llvm_report_error("SSE register return with SSE disabled");
1194 // If this is a call to a function that returns an fp value on the floating
1195 // point stack, but where we prefer to use the value in xmm registers, copy
1196 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1197 if ((VA.getLocReg() == X86::ST0 ||
1198 VA.getLocReg() == X86::ST1) &&
1199 isScalarFPTypeInSSEReg(VA.getValVT())) {
1204 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1205 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1206 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1207 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1208 MVT::v2i64, InFlag).getValue(1);
1209 Val = Chain.getValue(0);
1210 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1211 Val, DAG.getConstant(0, MVT::i64));
1213 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1214 MVT::i64, InFlag).getValue(1);
1215 Val = Chain.getValue(0);
1217 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1219 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1220 CopyVT, InFlag).getValue(1);
1221 Val = Chain.getValue(0);
1223 InFlag = Chain.getValue(2);
1225 if (CopyVT != VA.getValVT()) {
1226 // Round the F80 the right size, which also moves to the appropriate xmm
1228 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1229 // This truncation won't change the value.
1230 DAG.getIntPtrConstant(1));
1233 ResultVals.push_back(Val);
1236 // Merge everything together with a MERGE_VALUES node.
1237 ResultVals.push_back(Chain);
1238 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1239 &ResultVals[0], ResultVals.size()).getNode();
1243 //===----------------------------------------------------------------------===//
1244 // C & StdCall & Fast Calling Convention implementation
1245 //===----------------------------------------------------------------------===//
1246 // StdCall calling convention seems to be standard for many Windows' API
1247 // routines and around. It differs from C calling convention just a little:
1248 // callee should clean up the stack, not caller. Symbols should be also
1249 // decorated in some fancy way :) It doesn't support any vector arguments.
1250 // For info on fast calling convention see Fast Calling Convention (tail call)
1251 // implementation LowerX86_32FastCCCallTo.
1253 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1255 static bool CallIsStructReturn(CallSDNode *TheCall) {
1256 unsigned NumOps = TheCall->getNumArgs();
1260 return TheCall->getArgFlags(0).isSRet();
1263 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1264 /// return semantics.
1265 static bool ArgsAreStructReturn(SDValue Op) {
1266 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1270 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1273 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1274 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1276 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1280 switch (CallingConv) {
1283 case CallingConv::X86_StdCall:
1284 return !Subtarget->is64Bit();
1285 case CallingConv::X86_FastCall:
1286 return !Subtarget->is64Bit();
1287 case CallingConv::Fast:
1288 return PerformTailCallOpt;
1292 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1293 /// given CallingConvention value.
1294 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1295 if (Subtarget->is64Bit()) {
1296 if (Subtarget->isTargetWin64())
1297 return CC_X86_Win64_C;
1302 if (CC == CallingConv::X86_FastCall)
1303 return CC_X86_32_FastCall;
1304 else if (CC == CallingConv::Fast)
1305 return CC_X86_32_FastCC;
1310 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1311 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1313 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1314 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1315 if (CC == CallingConv::X86_FastCall)
1317 else if (CC == CallingConv::X86_StdCall)
1323 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1324 /// by "Src" to address "Dst" with size and alignment information specified by
1325 /// the specific parameter attribute. The copy will be passed as a byval
1326 /// function parameter.
1328 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1329 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1331 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1332 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1333 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1336 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1337 const CCValAssign &VA,
1338 MachineFrameInfo *MFI,
1340 SDValue Root, unsigned i) {
1341 // Create the nodes corresponding to a load from this parameter slot.
1342 ISD::ArgFlagsTy Flags =
1343 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1344 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1345 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1347 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1348 // changed with more analysis.
1349 // In case of tail call optimization mark all arguments mutable. Since they
1350 // could be overwritten by lowering of arguments in case of a tail call.
1351 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1352 VA.getLocMemOffset(), isImmutable);
1353 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1354 if (Flags.isByVal())
1356 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1357 PseudoSourceValue::getFixedStack(FI), 0);
1361 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1362 MachineFunction &MF = DAG.getMachineFunction();
1363 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1364 DebugLoc dl = Op.getDebugLoc();
1366 const Function* Fn = MF.getFunction();
1367 if (Fn->hasExternalLinkage() &&
1368 Subtarget->isTargetCygMing() &&
1369 Fn->getName() == "main")
1370 FuncInfo->setForceFramePointer(true);
1372 // Decorate the function name.
1373 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1375 MachineFrameInfo *MFI = MF.getFrameInfo();
1376 SDValue Root = Op.getOperand(0);
1377 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1378 unsigned CC = MF.getFunction()->getCallingConv();
1379 bool Is64Bit = Subtarget->is64Bit();
1380 bool IsWin64 = Subtarget->isTargetWin64();
1382 assert(!(isVarArg && CC == CallingConv::Fast) &&
1383 "Var args not supported with calling convention fastcc");
1385 // Assign locations to all of the incoming arguments.
1386 SmallVector<CCValAssign, 16> ArgLocs;
1387 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, DAG.getContext());
1388 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1390 SmallVector<SDValue, 8> ArgValues;
1391 unsigned LastVal = ~0U;
1392 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1393 CCValAssign &VA = ArgLocs[i];
1394 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1396 assert(VA.getValNo() != LastVal &&
1397 "Don't support value assigned to multiple locs yet");
1398 LastVal = VA.getValNo();
1400 if (VA.isRegLoc()) {
1401 MVT RegVT = VA.getLocVT();
1402 TargetRegisterClass *RC = NULL;
1403 if (RegVT == MVT::i32)
1404 RC = X86::GR32RegisterClass;
1405 else if (Is64Bit && RegVT == MVT::i64)
1406 RC = X86::GR64RegisterClass;
1407 else if (RegVT == MVT::f32)
1408 RC = X86::FR32RegisterClass;
1409 else if (RegVT == MVT::f64)
1410 RC = X86::FR64RegisterClass;
1411 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1412 RC = X86::VR128RegisterClass;
1413 else if (RegVT.isVector()) {
1414 assert(RegVT.getSizeInBits() == 64);
1416 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1418 // Darwin calling convention passes MMX values in either GPRs or
1419 // XMMs in x86-64. Other targets pass them in memory.
1420 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1421 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1424 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1429 assert(0 && "Unknown argument type!");
1432 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1433 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1435 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1436 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1438 if (VA.getLocInfo() == CCValAssign::SExt)
1439 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1440 DAG.getValueType(VA.getValVT()));
1441 else if (VA.getLocInfo() == CCValAssign::ZExt)
1442 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1443 DAG.getValueType(VA.getValVT()));
1445 if (VA.getLocInfo() != CCValAssign::Full)
1446 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1448 // Handle MMX values passed in GPRs.
1449 if (Is64Bit && RegVT != VA.getLocVT()) {
1450 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1451 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1452 else if (RC == X86::VR128RegisterClass) {
1453 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1454 ArgValue, DAG.getConstant(0, MVT::i64));
1455 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1459 ArgValues.push_back(ArgValue);
1461 assert(VA.isMemLoc());
1462 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1466 // The x86-64 ABI for returning structs by value requires that we copy
1467 // the sret argument into %rax for the return. Save the argument into
1468 // a virtual register so that we can access it from the return points.
1469 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1470 MachineFunction &MF = DAG.getMachineFunction();
1471 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1472 unsigned Reg = FuncInfo->getSRetReturnReg();
1474 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1475 FuncInfo->setSRetReturnReg(Reg);
1477 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1478 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1481 unsigned StackSize = CCInfo.getNextStackOffset();
1482 // align stack specially for tail calls
1483 if (PerformTailCallOpt && CC == CallingConv::Fast)
1484 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1486 // If the function takes variable number of arguments, make a frame index for
1487 // the start of the first vararg value... for expansion of llvm.va_start.
1489 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1490 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1493 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1495 // FIXME: We should really autogenerate these arrays
1496 static const unsigned GPR64ArgRegsWin64[] = {
1497 X86::RCX, X86::RDX, X86::R8, X86::R9
1499 static const unsigned XMMArgRegsWin64[] = {
1500 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1502 static const unsigned GPR64ArgRegs64Bit[] = {
1503 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1505 static const unsigned XMMArgRegs64Bit[] = {
1506 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1507 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1509 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1512 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1513 GPR64ArgRegs = GPR64ArgRegsWin64;
1514 XMMArgRegs = XMMArgRegsWin64;
1516 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1517 GPR64ArgRegs = GPR64ArgRegs64Bit;
1518 XMMArgRegs = XMMArgRegs64Bit;
1520 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1522 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1525 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1526 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1527 "SSE register cannot be used when SSE is disabled!");
1528 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1529 "SSE register cannot be used when SSE is disabled!");
1530 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1531 // Kernel mode asks for SSE to be disabled, so don't push them
1533 TotalNumXMMRegs = 0;
1535 // For X86-64, if there are vararg parameters that are passed via
1536 // registers, then we must store them to their spots on the stack so they
1537 // may be loaded by deferencing the result of va_next.
1538 VarArgsGPOffset = NumIntRegs * 8;
1539 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1540 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1541 TotalNumXMMRegs * 16, 16);
1543 // Store the integer parameter registers.
1544 SmallVector<SDValue, 8> MemOps;
1545 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1546 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1547 DAG.getIntPtrConstant(VarArgsGPOffset));
1548 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1549 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1550 X86::GR64RegisterClass);
1551 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1553 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1554 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1555 MemOps.push_back(Store);
1556 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1557 DAG.getIntPtrConstant(8));
1560 // Now store the XMM (fp + vector) parameter registers.
1561 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1562 DAG.getIntPtrConstant(VarArgsFPOffset));
1563 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1564 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1565 X86::VR128RegisterClass);
1566 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1568 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1569 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1570 MemOps.push_back(Store);
1571 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1572 DAG.getIntPtrConstant(16));
1574 if (!MemOps.empty())
1575 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1576 &MemOps[0], MemOps.size());
1580 ArgValues.push_back(Root);
1582 // Some CCs need callee pop.
1583 if (IsCalleePop(isVarArg, CC)) {
1584 BytesToPopOnReturn = StackSize; // Callee pops everything.
1585 BytesCallerReserves = 0;
1587 BytesToPopOnReturn = 0; // Callee pops nothing.
1588 // If this is an sret function, the return should pop the hidden pointer.
1589 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1590 BytesToPopOnReturn = 4;
1591 BytesCallerReserves = StackSize;
1595 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1596 if (CC == CallingConv::X86_FastCall)
1597 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1600 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1602 // Return the new list of results.
1603 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1604 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1608 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1609 const SDValue &StackPtr,
1610 const CCValAssign &VA,
1612 SDValue Arg, ISD::ArgFlagsTy Flags) {
1613 DebugLoc dl = TheCall->getDebugLoc();
1614 unsigned LocMemOffset = VA.getLocMemOffset();
1615 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1616 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1617 if (Flags.isByVal()) {
1618 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1620 return DAG.getStore(Chain, dl, Arg, PtrOff,
1621 PseudoSourceValue::getStack(), LocMemOffset);
1624 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1625 /// optimization is performed and it is required.
1627 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1628 SDValue &OutRetAddr,
1634 if (!IsTailCall || FPDiff==0) return Chain;
1636 // Adjust the Return address stack slot.
1637 MVT VT = getPointerTy();
1638 OutRetAddr = getReturnAddressFrameIndex(DAG);
1640 // Load the "old" Return address.
1641 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1642 return SDValue(OutRetAddr.getNode(), 1);
1645 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1646 /// optimization is performed and it is required (FPDiff!=0).
1648 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1649 SDValue Chain, SDValue RetAddrFrIdx,
1650 bool Is64Bit, int FPDiff, DebugLoc dl) {
1651 // Store the return address to the appropriate stack slot.
1652 if (!FPDiff) return Chain;
1653 // Calculate the new stack slot for the return address.
1654 int SlotSize = Is64Bit ? 8 : 4;
1655 int NewReturnAddrFI =
1656 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1657 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1658 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1659 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1660 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1664 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1665 MachineFunction &MF = DAG.getMachineFunction();
1666 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1667 SDValue Chain = TheCall->getChain();
1668 unsigned CC = TheCall->getCallingConv();
1669 bool isVarArg = TheCall->isVarArg();
1670 bool IsTailCall = TheCall->isTailCall() &&
1671 CC == CallingConv::Fast && PerformTailCallOpt;
1672 SDValue Callee = TheCall->getCallee();
1673 bool Is64Bit = Subtarget->is64Bit();
1674 bool IsStructRet = CallIsStructReturn(TheCall);
1675 DebugLoc dl = TheCall->getDebugLoc();
1677 assert(!(isVarArg && CC == CallingConv::Fast) &&
1678 "Var args not supported with calling convention fastcc");
1680 // Analyze operands of the call, assigning locations to each operand.
1681 SmallVector<CCValAssign, 16> ArgLocs;
1682 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, DAG.getContext());
1683 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1685 // Get a count of how many bytes are to be pushed on the stack.
1686 unsigned NumBytes = CCInfo.getNextStackOffset();
1687 if (PerformTailCallOpt && CC == CallingConv::Fast)
1688 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1692 // Lower arguments at fp - stackoffset + fpdiff.
1693 unsigned NumBytesCallerPushed =
1694 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1695 FPDiff = NumBytesCallerPushed - NumBytes;
1697 // Set the delta of movement of the returnaddr stackslot.
1698 // But only set if delta is greater than previous delta.
1699 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1700 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1703 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1705 SDValue RetAddrFrIdx;
1706 // Load return adress for tail calls.
1707 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1710 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1711 SmallVector<SDValue, 8> MemOpChains;
1714 // Walk the register/memloc assignments, inserting copies/loads. In the case
1715 // of tail call optimization arguments are handle later.
1716 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1717 CCValAssign &VA = ArgLocs[i];
1718 SDValue Arg = TheCall->getArg(i);
1719 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1720 bool isByVal = Flags.isByVal();
1722 // Promote the value if needed.
1723 switch (VA.getLocInfo()) {
1724 default: assert(0 && "Unknown loc info!");
1725 case CCValAssign::Full: break;
1726 case CCValAssign::SExt:
1727 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1729 case CCValAssign::ZExt:
1730 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1732 case CCValAssign::AExt:
1733 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1737 if (VA.isRegLoc()) {
1739 MVT RegVT = VA.getLocVT();
1740 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1741 switch (VA.getLocReg()) {
1744 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1746 // Special case: passing MMX values in GPR registers.
1747 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1750 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1751 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1752 // Special case: passing MMX values in XMM registers.
1753 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1754 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1755 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1760 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1762 if (!IsTailCall || (IsTailCall && isByVal)) {
1763 assert(VA.isMemLoc());
1764 if (StackPtr.getNode() == 0)
1765 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1767 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1768 Chain, Arg, Flags));
1773 if (!MemOpChains.empty())
1774 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1775 &MemOpChains[0], MemOpChains.size());
1777 // Build a sequence of copy-to-reg nodes chained together with token chain
1778 // and flag operands which copy the outgoing args into registers.
1780 // Tail call byval lowering might overwrite argument registers so in case of
1781 // tail call optimization the copies to registers are lowered later.
1783 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1784 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1785 RegsToPass[i].second, InFlag);
1786 InFlag = Chain.getValue(1);
1790 if (Subtarget->isPICStyleGOT()) {
1791 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1794 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1795 DAG.getNode(X86ISD::GlobalBaseReg,
1796 DebugLoc::getUnknownLoc(),
1799 InFlag = Chain.getValue(1);
1801 // If we are tail calling and generating PIC/GOT style code load the
1802 // address of the callee into ECX. The value in ecx is used as target of
1803 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1804 // for tail calls on PIC/GOT architectures. Normally we would just put the
1805 // address of GOT into ebx and then call target@PLT. But for tail calls
1806 // ebx would be restored (since ebx is callee saved) before jumping to the
1809 // Note: The actual moving to ECX is done further down.
1810 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1811 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1812 !G->getGlobal()->hasProtectedVisibility())
1813 Callee = LowerGlobalAddress(Callee, DAG);
1814 else if (isa<ExternalSymbolSDNode>(Callee))
1815 Callee = LowerExternalSymbol(Callee, DAG);
1819 if (Is64Bit && isVarArg) {
1820 // From AMD64 ABI document:
1821 // For calls that may call functions that use varargs or stdargs
1822 // (prototype-less calls or calls to functions containing ellipsis (...) in
1823 // the declaration) %al is used as hidden argument to specify the number
1824 // of SSE registers used. The contents of %al do not need to match exactly
1825 // the number of registers, but must be an ubound on the number of SSE
1826 // registers used and is in the range 0 - 8 inclusive.
1828 // FIXME: Verify this on Win64
1829 // Count the number of XMM registers allocated.
1830 static const unsigned XMMArgRegs[] = {
1831 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1832 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1834 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1835 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1836 && "SSE registers cannot be used when SSE is disabled");
1838 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1839 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1840 InFlag = Chain.getValue(1);
1844 // For tail calls lower the arguments to the 'real' stack slot.
1846 SmallVector<SDValue, 8> MemOpChains2;
1849 // Do not flag preceeding copytoreg stuff together with the following stuff.
1851 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1852 CCValAssign &VA = ArgLocs[i];
1853 if (!VA.isRegLoc()) {
1854 assert(VA.isMemLoc());
1855 SDValue Arg = TheCall->getArg(i);
1856 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1857 // Create frame index.
1858 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1859 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1860 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1861 FIN = DAG.getFrameIndex(FI, getPointerTy());
1863 if (Flags.isByVal()) {
1864 // Copy relative to framepointer.
1865 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1866 if (StackPtr.getNode() == 0)
1867 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1869 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1871 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1874 // Store relative to framepointer.
1875 MemOpChains2.push_back(
1876 DAG.getStore(Chain, dl, Arg, FIN,
1877 PseudoSourceValue::getFixedStack(FI), 0));
1882 if (!MemOpChains2.empty())
1883 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1884 &MemOpChains2[0], MemOpChains2.size());
1886 // Copy arguments to their registers.
1887 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1888 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1889 RegsToPass[i].second, InFlag);
1890 InFlag = Chain.getValue(1);
1894 // Store the return address to the appropriate stack slot.
1895 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1899 // If the callee is a GlobalAddress node (quite common, every direct call is)
1900 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1901 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1902 // We should use extra load for direct calls to dllimported functions in
1904 GlobalValue *GV = G->getGlobal();
1905 if (!GV->hasDLLImportLinkage()) {
1906 unsigned char OpFlags = 0;
1908 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1909 // external symbols most go through the PLT in PIC mode. If the symbol
1910 // has hidden or protected visibility, or if it is static or local, then
1911 // we don't need to use the PLT - we can directly call it.
1912 if (Subtarget->isTargetELF() &&
1913 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1914 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1915 OpFlags = X86II::MO_PLT;
1916 } else if (Subtarget->isPICStyleStub() &&
1917 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1918 Subtarget->getDarwinVers() < 9) {
1919 // PC-relative references to external symbols should go through $stub,
1920 // unless we're building with the leopard linker or later, which
1921 // automatically synthesizes these stubs.
1922 OpFlags = X86II::MO_DARWIN_STUB;
1925 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1926 G->getOffset(), OpFlags);
1928 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1929 unsigned char OpFlags = 0;
1931 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1932 // symbols should go through the PLT.
1933 if (Subtarget->isTargetELF() &&
1934 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1935 OpFlags = X86II::MO_PLT;
1936 } else if (Subtarget->isPICStyleStub() &&
1937 Subtarget->getDarwinVers() < 9) {
1938 // PC-relative references to external symbols should go through $stub,
1939 // unless we're building with the leopard linker or later, which
1940 // automatically synthesizes these stubs.
1941 OpFlags = X86II::MO_DARWIN_STUB;
1944 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1946 } else if (IsTailCall) {
1947 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1949 Chain = DAG.getCopyToReg(Chain, dl,
1950 DAG.getRegister(Opc, getPointerTy()),
1952 Callee = DAG.getRegister(Opc, getPointerTy());
1953 // Add register as live out.
1954 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1957 // Returns a chain & a flag for retval copy to use.
1958 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1959 SmallVector<SDValue, 8> Ops;
1962 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1963 DAG.getIntPtrConstant(0, true), InFlag);
1964 InFlag = Chain.getValue(1);
1966 // Returns a chain & a flag for retval copy to use.
1967 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1971 Ops.push_back(Chain);
1972 Ops.push_back(Callee);
1975 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1977 // Add argument registers to the end of the list so that they are known live
1979 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1980 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1981 RegsToPass[i].second.getValueType()));
1983 // Add an implicit use GOT pointer in EBX.
1984 if (!IsTailCall && Subtarget->isPICStyleGOT())
1985 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1987 // Add an implicit use of AL for x86 vararg functions.
1988 if (Is64Bit && isVarArg)
1989 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1991 if (InFlag.getNode())
1992 Ops.push_back(InFlag);
1995 assert(InFlag.getNode() &&
1996 "Flag must be set. Depend on flag being set in LowerRET");
1997 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1998 TheCall->getVTList(), &Ops[0], Ops.size());
2000 return SDValue(Chain.getNode(), Op.getResNo());
2003 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2004 InFlag = Chain.getValue(1);
2006 // Create the CALLSEQ_END node.
2007 unsigned NumBytesForCalleeToPush;
2008 if (IsCalleePop(isVarArg, CC))
2009 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2010 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
2011 // If this is is a call to a struct-return function, the callee
2012 // pops the hidden struct pointer, so we have to push it back.
2013 // This is common for Darwin/X86, Linux & Mingw32 targets.
2014 NumBytesForCalleeToPush = 4;
2016 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2018 // Returns a flag for retval copy to use.
2019 Chain = DAG.getCALLSEQ_END(Chain,
2020 DAG.getIntPtrConstant(NumBytes, true),
2021 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2024 InFlag = Chain.getValue(1);
2026 // Handle result values, copying them out of physregs into vregs that we
2028 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
2033 //===----------------------------------------------------------------------===//
2034 // Fast Calling Convention (tail call) implementation
2035 //===----------------------------------------------------------------------===//
2037 // Like std call, callee cleans arguments, convention except that ECX is
2038 // reserved for storing the tail called function address. Only 2 registers are
2039 // free for argument passing (inreg). Tail call optimization is performed
2041 // * tailcallopt is enabled
2042 // * caller/callee are fastcc
2043 // On X86_64 architecture with GOT-style position independent code only local
2044 // (within module) calls are supported at the moment.
2045 // To keep the stack aligned according to platform abi the function
2046 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2047 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2048 // If a tail called function callee has more arguments than the caller the
2049 // caller needs to make sure that there is room to move the RETADDR to. This is
2050 // achieved by reserving an area the size of the argument delta right after the
2051 // original REtADDR, but before the saved framepointer or the spilled registers
2052 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2064 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2065 /// for a 16 byte align requirement.
2066 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2067 SelectionDAG& DAG) {
2068 MachineFunction &MF = DAG.getMachineFunction();
2069 const TargetMachine &TM = MF.getTarget();
2070 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2071 unsigned StackAlignment = TFI.getStackAlignment();
2072 uint64_t AlignMask = StackAlignment - 1;
2073 int64_t Offset = StackSize;
2074 uint64_t SlotSize = TD->getPointerSize();
2075 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2076 // Number smaller than 12 so just add the difference.
2077 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2079 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2080 Offset = ((~AlignMask) & Offset) + StackAlignment +
2081 (StackAlignment-SlotSize);
2086 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2087 /// following the call is a return. A function is eligible if caller/callee
2088 /// calling conventions match, currently only fastcc supports tail calls, and
2089 /// the function CALL is immediatly followed by a RET.
2090 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2092 SelectionDAG& DAG) const {
2093 if (!PerformTailCallOpt)
2096 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2098 DAG.getMachineFunction().getFunction()->getCallingConv();
2099 unsigned CalleeCC = TheCall->getCallingConv();
2100 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC)
2108 X86TargetLowering::createFastISel(MachineFunction &mf,
2109 MachineModuleInfo *mmo,
2111 DenseMap<const Value *, unsigned> &vm,
2112 DenseMap<const BasicBlock *,
2113 MachineBasicBlock *> &bm,
2114 DenseMap<const AllocaInst *, int> &am
2116 , SmallSet<Instruction*, 8> &cil
2119 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2127 //===----------------------------------------------------------------------===//
2128 // Other Lowering Hooks
2129 //===----------------------------------------------------------------------===//
2132 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2133 MachineFunction &MF = DAG.getMachineFunction();
2134 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2135 int ReturnAddrIndex = FuncInfo->getRAIndex();
2137 if (ReturnAddrIndex == 0) {
2138 // Set up a frame object for the return address.
2139 uint64_t SlotSize = TD->getPointerSize();
2140 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2141 FuncInfo->setRAIndex(ReturnAddrIndex);
2144 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2148 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2149 /// specific condition code, returning the condition code and the LHS/RHS of the
2150 /// comparison to make.
2151 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2152 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2154 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2155 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2156 // X > -1 -> X == 0, jump !sign.
2157 RHS = DAG.getConstant(0, RHS.getValueType());
2158 return X86::COND_NS;
2159 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2160 // X < 0 -> X == 0, jump on sign.
2162 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2164 RHS = DAG.getConstant(0, RHS.getValueType());
2165 return X86::COND_LE;
2169 switch (SetCCOpcode) {
2170 default: assert(0 && "Invalid integer condition!");
2171 case ISD::SETEQ: return X86::COND_E;
2172 case ISD::SETGT: return X86::COND_G;
2173 case ISD::SETGE: return X86::COND_GE;
2174 case ISD::SETLT: return X86::COND_L;
2175 case ISD::SETLE: return X86::COND_LE;
2176 case ISD::SETNE: return X86::COND_NE;
2177 case ISD::SETULT: return X86::COND_B;
2178 case ISD::SETUGT: return X86::COND_A;
2179 case ISD::SETULE: return X86::COND_BE;
2180 case ISD::SETUGE: return X86::COND_AE;
2184 // First determine if it is required or is profitable to flip the operands.
2186 // If LHS is a foldable load, but RHS is not, flip the condition.
2187 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2188 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2189 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2190 std::swap(LHS, RHS);
2193 switch (SetCCOpcode) {
2199 std::swap(LHS, RHS);
2203 // On a floating point condition, the flags are set as follows:
2205 // 0 | 0 | 0 | X > Y
2206 // 0 | 0 | 1 | X < Y
2207 // 1 | 0 | 0 | X == Y
2208 // 1 | 1 | 1 | unordered
2209 switch (SetCCOpcode) {
2210 default: assert(0 && "Condcode should be pre-legalized away");
2212 case ISD::SETEQ: return X86::COND_E;
2213 case ISD::SETOLT: // flipped
2215 case ISD::SETGT: return X86::COND_A;
2216 case ISD::SETOLE: // flipped
2218 case ISD::SETGE: return X86::COND_AE;
2219 case ISD::SETUGT: // flipped
2221 case ISD::SETLT: return X86::COND_B;
2222 case ISD::SETUGE: // flipped
2224 case ISD::SETLE: return X86::COND_BE;
2226 case ISD::SETNE: return X86::COND_NE;
2227 case ISD::SETUO: return X86::COND_P;
2228 case ISD::SETO: return X86::COND_NP;
2232 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2233 /// code. Current x86 isa includes the following FP cmov instructions:
2234 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2235 static bool hasFPCMov(unsigned X86CC) {
2251 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2252 /// the specified range (L, H].
2253 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2254 return (Val < 0) || (Val >= Low && Val < Hi);
2257 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2258 /// specified value.
2259 static bool isUndefOrEqual(int Val, int CmpVal) {
2260 if (Val < 0 || Val == CmpVal)
2265 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2266 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2267 /// the second operand.
2268 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2269 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2270 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2271 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2272 return (Mask[0] < 2 && Mask[1] < 2);
2276 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2277 SmallVector<int, 8> M;
2279 return ::isPSHUFDMask(M, N->getValueType(0));
2282 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2283 /// is suitable for input to PSHUFHW.
2284 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2285 if (VT != MVT::v8i16)
2288 // Lower quadword copied in order or undef.
2289 for (int i = 0; i != 4; ++i)
2290 if (Mask[i] >= 0 && Mask[i] != i)
2293 // Upper quadword shuffled.
2294 for (int i = 4; i != 8; ++i)
2295 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2301 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2302 SmallVector<int, 8> M;
2304 return ::isPSHUFHWMask(M, N->getValueType(0));
2307 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2308 /// is suitable for input to PSHUFLW.
2309 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2310 if (VT != MVT::v8i16)
2313 // Upper quadword copied in order.
2314 for (int i = 4; i != 8; ++i)
2315 if (Mask[i] >= 0 && Mask[i] != i)
2318 // Lower quadword shuffled.
2319 for (int i = 0; i != 4; ++i)
2326 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2327 SmallVector<int, 8> M;
2329 return ::isPSHUFLWMask(M, N->getValueType(0));
2332 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2333 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2334 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2335 int NumElems = VT.getVectorNumElements();
2336 if (NumElems != 2 && NumElems != 4)
2339 int Half = NumElems / 2;
2340 for (int i = 0; i < Half; ++i)
2341 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2343 for (int i = Half; i < NumElems; ++i)
2344 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2350 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2351 SmallVector<int, 8> M;
2353 return ::isSHUFPMask(M, N->getValueType(0));
2356 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2357 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2358 /// half elements to come from vector 1 (which would equal the dest.) and
2359 /// the upper half to come from vector 2.
2360 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2361 int NumElems = VT.getVectorNumElements();
2363 if (NumElems != 2 && NumElems != 4)
2366 int Half = NumElems / 2;
2367 for (int i = 0; i < Half; ++i)
2368 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2370 for (int i = Half; i < NumElems; ++i)
2371 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2376 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2377 SmallVector<int, 8> M;
2379 return isCommutedSHUFPMask(M, N->getValueType(0));
2382 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2383 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2384 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2385 if (N->getValueType(0).getVectorNumElements() != 4)
2388 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2389 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2390 isUndefOrEqual(N->getMaskElt(1), 7) &&
2391 isUndefOrEqual(N->getMaskElt(2), 2) &&
2392 isUndefOrEqual(N->getMaskElt(3), 3);
2395 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2396 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2397 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2398 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2400 if (NumElems != 2 && NumElems != 4)
2403 for (unsigned i = 0; i < NumElems/2; ++i)
2404 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2407 for (unsigned i = NumElems/2; i < NumElems; ++i)
2408 if (!isUndefOrEqual(N->getMaskElt(i), i))
2414 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2415 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2417 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2418 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2420 if (NumElems != 2 && NumElems != 4)
2423 for (unsigned i = 0; i < NumElems/2; ++i)
2424 if (!isUndefOrEqual(N->getMaskElt(i), i))
2427 for (unsigned i = 0; i < NumElems/2; ++i)
2428 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2434 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2435 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2437 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2438 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2443 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2444 isUndefOrEqual(N->getMaskElt(1), 3) &&
2445 isUndefOrEqual(N->getMaskElt(2), 2) &&
2446 isUndefOrEqual(N->getMaskElt(3), 3);
2449 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2450 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2451 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2452 bool V2IsSplat = false) {
2453 int NumElts = VT.getVectorNumElements();
2454 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2457 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2459 int BitI1 = Mask[i+1];
2460 if (!isUndefOrEqual(BitI, j))
2463 if (!isUndefOrEqual(BitI1, NumElts))
2466 if (!isUndefOrEqual(BitI1, j + NumElts))
2473 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2474 SmallVector<int, 8> M;
2476 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2479 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2480 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2481 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2482 bool V2IsSplat = false) {
2483 int NumElts = VT.getVectorNumElements();
2484 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2487 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2489 int BitI1 = Mask[i+1];
2490 if (!isUndefOrEqual(BitI, j + NumElts/2))
2493 if (isUndefOrEqual(BitI1, NumElts))
2496 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2503 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2504 SmallVector<int, 8> M;
2506 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2509 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2510 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2512 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2513 int NumElems = VT.getVectorNumElements();
2514 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2517 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2519 int BitI1 = Mask[i+1];
2520 if (!isUndefOrEqual(BitI, j))
2522 if (!isUndefOrEqual(BitI1, j))
2528 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2529 SmallVector<int, 8> M;
2531 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2534 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2535 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2537 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2538 int NumElems = VT.getVectorNumElements();
2539 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2542 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2544 int BitI1 = Mask[i+1];
2545 if (!isUndefOrEqual(BitI, j))
2547 if (!isUndefOrEqual(BitI1, j))
2553 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2554 SmallVector<int, 8> M;
2556 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2559 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2560 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2561 /// MOVSD, and MOVD, i.e. setting the lowest element.
2562 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2563 if (VT.getVectorElementType().getSizeInBits() < 32)
2566 int NumElts = VT.getVectorNumElements();
2568 if (!isUndefOrEqual(Mask[0], NumElts))
2571 for (int i = 1; i < NumElts; ++i)
2572 if (!isUndefOrEqual(Mask[i], i))
2578 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2579 SmallVector<int, 8> M;
2581 return ::isMOVLMask(M, N->getValueType(0));
2584 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2585 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2586 /// element of vector 2 and the other elements to come from vector 1 in order.
2587 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2588 bool V2IsSplat = false, bool V2IsUndef = false) {
2589 int NumOps = VT.getVectorNumElements();
2590 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2593 if (!isUndefOrEqual(Mask[0], 0))
2596 for (int i = 1; i < NumOps; ++i)
2597 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2598 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2599 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2605 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2606 bool V2IsUndef = false) {
2607 SmallVector<int, 8> M;
2609 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2612 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2613 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2614 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2615 if (N->getValueType(0).getVectorNumElements() != 4)
2618 // Expect 1, 1, 3, 3
2619 for (unsigned i = 0; i < 2; ++i) {
2620 int Elt = N->getMaskElt(i);
2621 if (Elt >= 0 && Elt != 1)
2626 for (unsigned i = 2; i < 4; ++i) {
2627 int Elt = N->getMaskElt(i);
2628 if (Elt >= 0 && Elt != 3)
2633 // Don't use movshdup if it can be done with a shufps.
2634 // FIXME: verify that matching u, u, 3, 3 is what we want.
2638 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2639 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2640 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2641 if (N->getValueType(0).getVectorNumElements() != 4)
2644 // Expect 0, 0, 2, 2
2645 for (unsigned i = 0; i < 2; ++i)
2646 if (N->getMaskElt(i) > 0)
2650 for (unsigned i = 2; i < 4; ++i) {
2651 int Elt = N->getMaskElt(i);
2652 if (Elt >= 0 && Elt != 2)
2657 // Don't use movsldup if it can be done with a shufps.
2661 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2662 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2663 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2664 int e = N->getValueType(0).getVectorNumElements() / 2;
2666 for (int i = 0; i < e; ++i)
2667 if (!isUndefOrEqual(N->getMaskElt(i), i))
2669 for (int i = 0; i < e; ++i)
2670 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2675 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2676 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2678 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2679 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2680 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2682 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2684 for (int i = 0; i < NumOperands; ++i) {
2685 int Val = SVOp->getMaskElt(NumOperands-i-1);
2686 if (Val < 0) Val = 0;
2687 if (Val >= NumOperands) Val -= NumOperands;
2689 if (i != NumOperands - 1)
2695 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2696 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2698 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2699 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2701 // 8 nodes, but we only care about the last 4.
2702 for (unsigned i = 7; i >= 4; --i) {
2703 int Val = SVOp->getMaskElt(i);
2712 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2713 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2715 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2716 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2718 // 8 nodes, but we only care about the first 4.
2719 for (int i = 3; i >= 0; --i) {
2720 int Val = SVOp->getMaskElt(i);
2729 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2730 /// their permute mask.
2731 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2732 SelectionDAG &DAG) {
2733 MVT VT = SVOp->getValueType(0);
2734 unsigned NumElems = VT.getVectorNumElements();
2735 SmallVector<int, 8> MaskVec;
2737 for (unsigned i = 0; i != NumElems; ++i) {
2738 int idx = SVOp->getMaskElt(i);
2740 MaskVec.push_back(idx);
2741 else if (idx < (int)NumElems)
2742 MaskVec.push_back(idx + NumElems);
2744 MaskVec.push_back(idx - NumElems);
2746 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2747 SVOp->getOperand(0), &MaskVec[0]);
2750 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2751 /// the two vector operands have swapped position.
2752 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2753 unsigned NumElems = VT.getVectorNumElements();
2754 for (unsigned i = 0; i != NumElems; ++i) {
2758 else if (idx < (int)NumElems)
2759 Mask[i] = idx + NumElems;
2761 Mask[i] = idx - NumElems;
2765 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2766 /// match movhlps. The lower half elements should come from upper half of
2767 /// V1 (and in order), and the upper half elements should come from the upper
2768 /// half of V2 (and in order).
2769 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2770 if (Op->getValueType(0).getVectorNumElements() != 4)
2772 for (unsigned i = 0, e = 2; i != e; ++i)
2773 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2775 for (unsigned i = 2; i != 4; ++i)
2776 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2781 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2782 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2784 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2785 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2787 N = N->getOperand(0).getNode();
2788 if (!ISD::isNON_EXTLoad(N))
2791 *LD = cast<LoadSDNode>(N);
2795 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2796 /// match movlp{s|d}. The lower half elements should come from lower half of
2797 /// V1 (and in order), and the upper half elements should come from the upper
2798 /// half of V2 (and in order). And since V1 will become the source of the
2799 /// MOVLP, it must be either a vector load or a scalar load to vector.
2800 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2801 ShuffleVectorSDNode *Op) {
2802 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2804 // Is V2 is a vector load, don't do this transformation. We will try to use
2805 // load folding shufps op.
2806 if (ISD::isNON_EXTLoad(V2))
2809 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2811 if (NumElems != 2 && NumElems != 4)
2813 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2814 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2816 for (unsigned i = NumElems/2; i != NumElems; ++i)
2817 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2822 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2824 static bool isSplatVector(SDNode *N) {
2825 if (N->getOpcode() != ISD::BUILD_VECTOR)
2828 SDValue SplatValue = N->getOperand(0);
2829 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2830 if (N->getOperand(i) != SplatValue)
2835 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2837 static inline bool isZeroNode(SDValue Elt) {
2838 return ((isa<ConstantSDNode>(Elt) &&
2839 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2840 (isa<ConstantFPSDNode>(Elt) &&
2841 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2844 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2845 /// to an zero vector.
2846 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2847 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2848 SDValue V1 = N->getOperand(0);
2849 SDValue V2 = N->getOperand(1);
2850 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2851 for (unsigned i = 0; i != NumElems; ++i) {
2852 int Idx = N->getMaskElt(i);
2853 if (Idx >= (int)NumElems) {
2854 unsigned Opc = V2.getOpcode();
2855 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2857 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2859 } else if (Idx >= 0) {
2860 unsigned Opc = V1.getOpcode();
2861 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2863 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2870 /// getZeroVector - Returns a vector of specified type with all zero elements.
2872 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2874 assert(VT.isVector() && "Expected a vector type");
2876 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2877 // type. This ensures they get CSE'd.
2879 if (VT.getSizeInBits() == 64) { // MMX
2880 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2881 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2882 } else if (HasSSE2) { // SSE2
2883 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2884 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2886 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2887 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2889 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2892 /// getOnesVector - Returns a vector of specified type with all bits set.
2894 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2895 assert(VT.isVector() && "Expected a vector type");
2897 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2898 // type. This ensures they get CSE'd.
2899 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2901 if (VT.getSizeInBits() == 64) // MMX
2902 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2904 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2905 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2909 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2910 /// that point to V2 points to its first element.
2911 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2912 MVT VT = SVOp->getValueType(0);
2913 unsigned NumElems = VT.getVectorNumElements();
2915 bool Changed = false;
2916 SmallVector<int, 8> MaskVec;
2917 SVOp->getMask(MaskVec);
2919 for (unsigned i = 0; i != NumElems; ++i) {
2920 if (MaskVec[i] > (int)NumElems) {
2921 MaskVec[i] = NumElems;
2926 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2927 SVOp->getOperand(1), &MaskVec[0]);
2928 return SDValue(SVOp, 0);
2931 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2932 /// operation of specified width.
2933 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2935 unsigned NumElems = VT.getVectorNumElements();
2936 SmallVector<int, 8> Mask;
2937 Mask.push_back(NumElems);
2938 for (unsigned i = 1; i != NumElems; ++i)
2940 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2943 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2944 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2946 unsigned NumElems = VT.getVectorNumElements();
2947 SmallVector<int, 8> Mask;
2948 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2950 Mask.push_back(i + NumElems);
2952 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2955 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2956 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2958 unsigned NumElems = VT.getVectorNumElements();
2959 unsigned Half = NumElems/2;
2960 SmallVector<int, 8> Mask;
2961 for (unsigned i = 0; i != Half; ++i) {
2962 Mask.push_back(i + Half);
2963 Mask.push_back(i + NumElems + Half);
2965 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2968 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2969 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2971 if (SV->getValueType(0).getVectorNumElements() <= 4)
2972 return SDValue(SV, 0);
2974 MVT PVT = MVT::v4f32;
2975 MVT VT = SV->getValueType(0);
2976 DebugLoc dl = SV->getDebugLoc();
2977 SDValue V1 = SV->getOperand(0);
2978 int NumElems = VT.getVectorNumElements();
2979 int EltNo = SV->getSplatIndex();
2981 // unpack elements to the correct location
2982 while (NumElems > 4) {
2983 if (EltNo < NumElems/2) {
2984 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2986 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2987 EltNo -= NumElems/2;
2992 // Perform the splat.
2993 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2994 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2995 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2996 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2999 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3000 /// vector of zero or undef vector. This produces a shuffle where the low
3001 /// element of V2 is swizzled into the zero/undef vector, landing at element
3002 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3003 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3004 bool isZero, bool HasSSE2,
3005 SelectionDAG &DAG) {
3006 MVT VT = V2.getValueType();
3008 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3009 unsigned NumElems = VT.getVectorNumElements();
3010 SmallVector<int, 16> MaskVec;
3011 for (unsigned i = 0; i != NumElems; ++i)
3012 // If this is the insertion idx, put the low elt of V2 here.
3013 MaskVec.push_back(i == Idx ? NumElems : i);
3014 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3017 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3018 /// a shuffle that is zero.
3020 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3021 bool Low, SelectionDAG &DAG) {
3022 unsigned NumZeros = 0;
3023 for (int i = 0; i < NumElems; ++i) {
3024 unsigned Index = Low ? i : NumElems-i-1;
3025 int Idx = SVOp->getMaskElt(Index);
3030 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3031 if (Elt.getNode() && isZeroNode(Elt))
3039 /// isVectorShift - Returns true if the shuffle can be implemented as a
3040 /// logical left or right shift of a vector.
3041 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3042 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3043 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3044 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3047 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3050 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3054 bool SeenV1 = false;
3055 bool SeenV2 = false;
3056 for (int i = NumZeros; i < NumElems; ++i) {
3057 int Val = isLeft ? (i - NumZeros) : i;
3058 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3070 if (SeenV1 && SeenV2)
3073 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3079 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3081 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3082 unsigned NumNonZero, unsigned NumZero,
3083 SelectionDAG &DAG, TargetLowering &TLI) {
3087 DebugLoc dl = Op.getDebugLoc();
3090 for (unsigned i = 0; i < 16; ++i) {
3091 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3092 if (ThisIsNonZero && First) {
3094 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3096 V = DAG.getUNDEF(MVT::v8i16);
3101 SDValue ThisElt(0, 0), LastElt(0, 0);
3102 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3103 if (LastIsNonZero) {
3104 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3105 MVT::i16, Op.getOperand(i-1));
3107 if (ThisIsNonZero) {
3108 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3109 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3110 ThisElt, DAG.getConstant(8, MVT::i8));
3112 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3116 if (ThisElt.getNode())
3117 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3118 DAG.getIntPtrConstant(i/2));
3122 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3125 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3127 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3128 unsigned NumNonZero, unsigned NumZero,
3129 SelectionDAG &DAG, TargetLowering &TLI) {
3133 DebugLoc dl = Op.getDebugLoc();
3136 for (unsigned i = 0; i < 8; ++i) {
3137 bool isNonZero = (NonZeros & (1 << i)) != 0;
3141 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3143 V = DAG.getUNDEF(MVT::v8i16);
3146 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3147 MVT::v8i16, V, Op.getOperand(i),
3148 DAG.getIntPtrConstant(i));
3155 /// getVShift - Return a vector logical shift node.
3157 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3158 unsigned NumBits, SelectionDAG &DAG,
3159 const TargetLowering &TLI, DebugLoc dl) {
3160 bool isMMX = VT.getSizeInBits() == 64;
3161 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3162 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3163 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3164 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3165 DAG.getNode(Opc, dl, ShVT, SrcOp,
3166 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3170 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3171 DebugLoc dl = Op.getDebugLoc();
3172 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3173 if (ISD::isBuildVectorAllZeros(Op.getNode())
3174 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3175 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3176 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3177 // eliminated on x86-32 hosts.
3178 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3181 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3182 return getOnesVector(Op.getValueType(), DAG, dl);
3183 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3186 MVT VT = Op.getValueType();
3187 MVT EVT = VT.getVectorElementType();
3188 unsigned EVTBits = EVT.getSizeInBits();
3190 unsigned NumElems = Op.getNumOperands();
3191 unsigned NumZero = 0;
3192 unsigned NumNonZero = 0;
3193 unsigned NonZeros = 0;
3194 bool IsAllConstants = true;
3195 SmallSet<SDValue, 8> Values;
3196 for (unsigned i = 0; i < NumElems; ++i) {
3197 SDValue Elt = Op.getOperand(i);
3198 if (Elt.getOpcode() == ISD::UNDEF)
3201 if (Elt.getOpcode() != ISD::Constant &&
3202 Elt.getOpcode() != ISD::ConstantFP)
3203 IsAllConstants = false;
3204 if (isZeroNode(Elt))
3207 NonZeros |= (1 << i);
3212 if (NumNonZero == 0) {
3213 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3214 return DAG.getUNDEF(VT);
3217 // Special case for single non-zero, non-undef, element.
3218 if (NumNonZero == 1) {
3219 unsigned Idx = CountTrailingZeros_32(NonZeros);
3220 SDValue Item = Op.getOperand(Idx);
3222 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3223 // the value are obviously zero, truncate the value to i32 and do the
3224 // insertion that way. Only do this if the value is non-constant or if the
3225 // value is a constant being inserted into element 0. It is cheaper to do
3226 // a constant pool load than it is to do a movd + shuffle.
3227 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3228 (!IsAllConstants || Idx == 0)) {
3229 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3230 // Handle MMX and SSE both.
3231 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3232 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3234 // Truncate the value (which may itself be a constant) to i32, and
3235 // convert it to a vector with movd (S2V+shuffle to zero extend).
3236 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3237 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3238 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3239 Subtarget->hasSSE2(), DAG);
3241 // Now we have our 32-bit value zero extended in the low element of
3242 // a vector. If Idx != 0, swizzle it into place.
3244 SmallVector<int, 4> Mask;
3245 Mask.push_back(Idx);
3246 for (unsigned i = 1; i != VecElts; ++i)
3248 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3249 DAG.getUNDEF(Item.getValueType()),
3252 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3256 // If we have a constant or non-constant insertion into the low element of
3257 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3258 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3259 // depending on what the source datatype is.
3262 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3263 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3264 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3265 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3266 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3267 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3269 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3270 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3271 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3272 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3273 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3274 Subtarget->hasSSE2(), DAG);
3275 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3279 // Is it a vector logical left shift?
3280 if (NumElems == 2 && Idx == 1 &&
3281 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3282 unsigned NumBits = VT.getSizeInBits();
3283 return getVShift(true, VT,
3284 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3285 VT, Op.getOperand(1)),
3286 NumBits/2, DAG, *this, dl);
3289 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3292 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3293 // is a non-constant being inserted into an element other than the low one,
3294 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3295 // movd/movss) to move this into the low element, then shuffle it into
3297 if (EVTBits == 32) {
3298 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3300 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3301 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3302 Subtarget->hasSSE2(), DAG);
3303 SmallVector<int, 8> MaskVec;
3304 for (unsigned i = 0; i < NumElems; i++)
3305 MaskVec.push_back(i == Idx ? 0 : 1);
3306 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3310 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3311 if (Values.size() == 1)
3314 // A vector full of immediates; various special cases are already
3315 // handled, so this is best done with a single constant-pool load.
3319 // Let legalizer expand 2-wide build_vectors.
3320 if (EVTBits == 64) {
3321 if (NumNonZero == 1) {
3322 // One half is zero or undef.
3323 unsigned Idx = CountTrailingZeros_32(NonZeros);
3324 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3325 Op.getOperand(Idx));
3326 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3327 Subtarget->hasSSE2(), DAG);
3332 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3333 if (EVTBits == 8 && NumElems == 16) {
3334 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3336 if (V.getNode()) return V;
3339 if (EVTBits == 16 && NumElems == 8) {
3340 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3342 if (V.getNode()) return V;
3345 // If element VT is == 32 bits, turn it into a number of shuffles.
3346 SmallVector<SDValue, 8> V;
3348 if (NumElems == 4 && NumZero > 0) {
3349 for (unsigned i = 0; i < 4; ++i) {
3350 bool isZero = !(NonZeros & (1 << i));
3352 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3354 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3357 for (unsigned i = 0; i < 2; ++i) {
3358 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3361 V[i] = V[i*2]; // Must be a zero vector.
3364 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3367 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3370 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3375 SmallVector<int, 8> MaskVec;
3376 bool Reverse = (NonZeros & 0x3) == 2;
3377 for (unsigned i = 0; i < 2; ++i)
3378 MaskVec.push_back(Reverse ? 1-i : i);
3379 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3380 for (unsigned i = 0; i < 2; ++i)
3381 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3382 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3385 if (Values.size() > 2) {
3386 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3387 // values to be inserted is equal to the number of elements, in which case
3388 // use the unpack code below in the hopes of matching the consecutive elts
3389 // load merge pattern for shuffles.
3390 // FIXME: We could probably just check that here directly.
3391 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3392 getSubtarget()->hasSSE41()) {
3393 V[0] = DAG.getUNDEF(VT);
3394 for (unsigned i = 0; i < NumElems; ++i)
3395 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3396 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3397 Op.getOperand(i), DAG.getIntPtrConstant(i));
3400 // Expand into a number of unpckl*.
3402 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3403 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3404 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3405 for (unsigned i = 0; i < NumElems; ++i)
3406 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3408 while (NumElems != 0) {
3409 for (unsigned i = 0; i < NumElems; ++i)
3410 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3419 // v8i16 shuffles - Prefer shuffles in the following order:
3420 // 1. [all] pshuflw, pshufhw, optional move
3421 // 2. [ssse3] 1 x pshufb
3422 // 3. [ssse3] 2 x pshufb + 1 x por
3423 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3425 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3426 SelectionDAG &DAG, X86TargetLowering &TLI) {
3427 SDValue V1 = SVOp->getOperand(0);
3428 SDValue V2 = SVOp->getOperand(1);
3429 DebugLoc dl = SVOp->getDebugLoc();
3430 SmallVector<int, 8> MaskVals;
3432 // Determine if more than 1 of the words in each of the low and high quadwords
3433 // of the result come from the same quadword of one of the two inputs. Undef
3434 // mask values count as coming from any quadword, for better codegen.
3435 SmallVector<unsigned, 4> LoQuad(4);
3436 SmallVector<unsigned, 4> HiQuad(4);
3437 BitVector InputQuads(4);
3438 for (unsigned i = 0; i < 8; ++i) {
3439 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3440 int EltIdx = SVOp->getMaskElt(i);
3441 MaskVals.push_back(EltIdx);
3450 InputQuads.set(EltIdx / 4);
3453 int BestLoQuad = -1;
3454 unsigned MaxQuad = 1;
3455 for (unsigned i = 0; i < 4; ++i) {
3456 if (LoQuad[i] > MaxQuad) {
3458 MaxQuad = LoQuad[i];
3462 int BestHiQuad = -1;
3464 for (unsigned i = 0; i < 4; ++i) {
3465 if (HiQuad[i] > MaxQuad) {
3467 MaxQuad = HiQuad[i];
3471 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3472 // of the two input vectors, shuffle them into one input vector so only a
3473 // single pshufb instruction is necessary. If There are more than 2 input
3474 // quads, disable the next transformation since it does not help SSSE3.
3475 bool V1Used = InputQuads[0] || InputQuads[1];
3476 bool V2Used = InputQuads[2] || InputQuads[3];
3477 if (TLI.getSubtarget()->hasSSSE3()) {
3478 if (InputQuads.count() == 2 && V1Used && V2Used) {
3479 BestLoQuad = InputQuads.find_first();
3480 BestHiQuad = InputQuads.find_next(BestLoQuad);
3482 if (InputQuads.count() > 2) {
3488 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3489 // the shuffle mask. If a quad is scored as -1, that means that it contains
3490 // words from all 4 input quadwords.
3492 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3493 SmallVector<int, 8> MaskV;
3494 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3495 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3496 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3497 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3498 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3499 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3501 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3502 // source words for the shuffle, to aid later transformations.
3503 bool AllWordsInNewV = true;
3504 bool InOrder[2] = { true, true };
3505 for (unsigned i = 0; i != 8; ++i) {
3506 int idx = MaskVals[i];
3508 InOrder[i/4] = false;
3509 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3511 AllWordsInNewV = false;
3515 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3516 if (AllWordsInNewV) {
3517 for (int i = 0; i != 8; ++i) {
3518 int idx = MaskVals[i];
3521 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3522 if ((idx != i) && idx < 4)
3524 if ((idx != i) && idx > 3)
3533 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3534 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3535 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3536 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3537 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3541 // If we have SSSE3, and all words of the result are from 1 input vector,
3542 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3543 // is present, fall back to case 4.
3544 if (TLI.getSubtarget()->hasSSSE3()) {
3545 SmallVector<SDValue,16> pshufbMask;
3547 // If we have elements from both input vectors, set the high bit of the
3548 // shuffle mask element to zero out elements that come from V2 in the V1
3549 // mask, and elements that come from V1 in the V2 mask, so that the two
3550 // results can be OR'd together.
3551 bool TwoInputs = V1Used && V2Used;
3552 for (unsigned i = 0; i != 8; ++i) {
3553 int EltIdx = MaskVals[i] * 2;
3554 if (TwoInputs && (EltIdx >= 16)) {
3555 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3556 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3559 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3560 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3562 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3563 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3564 DAG.getNode(ISD::BUILD_VECTOR, dl,
3565 MVT::v16i8, &pshufbMask[0], 16));
3567 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3569 // Calculate the shuffle mask for the second input, shuffle it, and
3570 // OR it with the first shuffled input.
3572 for (unsigned i = 0; i != 8; ++i) {
3573 int EltIdx = MaskVals[i] * 2;
3575 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3576 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3579 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3580 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3582 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3583 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3584 DAG.getNode(ISD::BUILD_VECTOR, dl,
3585 MVT::v16i8, &pshufbMask[0], 16));
3586 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3587 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3590 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3591 // and update MaskVals with new element order.
3592 BitVector InOrder(8);
3593 if (BestLoQuad >= 0) {
3594 SmallVector<int, 8> MaskV;
3595 for (int i = 0; i != 4; ++i) {
3596 int idx = MaskVals[i];
3598 MaskV.push_back(-1);
3600 } else if ((idx / 4) == BestLoQuad) {
3601 MaskV.push_back(idx & 3);
3604 MaskV.push_back(-1);
3607 for (unsigned i = 4; i != 8; ++i)
3609 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3613 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3614 // and update MaskVals with the new element order.
3615 if (BestHiQuad >= 0) {
3616 SmallVector<int, 8> MaskV;
3617 for (unsigned i = 0; i != 4; ++i)
3619 for (unsigned i = 4; i != 8; ++i) {
3620 int idx = MaskVals[i];
3622 MaskV.push_back(-1);
3624 } else if ((idx / 4) == BestHiQuad) {
3625 MaskV.push_back((idx & 3) + 4);
3628 MaskV.push_back(-1);
3631 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3635 // In case BestHi & BestLo were both -1, which means each quadword has a word
3636 // from each of the four input quadwords, calculate the InOrder bitvector now
3637 // before falling through to the insert/extract cleanup.
3638 if (BestLoQuad == -1 && BestHiQuad == -1) {
3640 for (int i = 0; i != 8; ++i)
3641 if (MaskVals[i] < 0 || MaskVals[i] == i)
3645 // The other elements are put in the right place using pextrw and pinsrw.
3646 for (unsigned i = 0; i != 8; ++i) {
3649 int EltIdx = MaskVals[i];
3652 SDValue ExtOp = (EltIdx < 8)
3653 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3654 DAG.getIntPtrConstant(EltIdx))
3655 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3656 DAG.getIntPtrConstant(EltIdx - 8));
3657 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3658 DAG.getIntPtrConstant(i));
3663 // v16i8 shuffles - Prefer shuffles in the following order:
3664 // 1. [ssse3] 1 x pshufb
3665 // 2. [ssse3] 2 x pshufb + 1 x por
3666 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3668 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3669 SelectionDAG &DAG, X86TargetLowering &TLI) {
3670 SDValue V1 = SVOp->getOperand(0);
3671 SDValue V2 = SVOp->getOperand(1);
3672 DebugLoc dl = SVOp->getDebugLoc();
3673 SmallVector<int, 16> MaskVals;
3674 SVOp->getMask(MaskVals);
3676 // If we have SSSE3, case 1 is generated when all result bytes come from
3677 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3678 // present, fall back to case 3.
3679 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3682 for (unsigned i = 0; i < 16; ++i) {
3683 int EltIdx = MaskVals[i];
3692 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3693 if (TLI.getSubtarget()->hasSSSE3()) {
3694 SmallVector<SDValue,16> pshufbMask;
3696 // If all result elements are from one input vector, then only translate
3697 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3699 // Otherwise, we have elements from both input vectors, and must zero out
3700 // elements that come from V2 in the first mask, and V1 in the second mask
3701 // so that we can OR them together.
3702 bool TwoInputs = !(V1Only || V2Only);
3703 for (unsigned i = 0; i != 16; ++i) {
3704 int EltIdx = MaskVals[i];
3705 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3706 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3709 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3711 // If all the elements are from V2, assign it to V1 and return after
3712 // building the first pshufb.
3715 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3716 DAG.getNode(ISD::BUILD_VECTOR, dl,
3717 MVT::v16i8, &pshufbMask[0], 16));
3721 // Calculate the shuffle mask for the second input, shuffle it, and
3722 // OR it with the first shuffled input.
3724 for (unsigned i = 0; i != 16; ++i) {
3725 int EltIdx = MaskVals[i];
3727 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3730 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3732 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3733 DAG.getNode(ISD::BUILD_VECTOR, dl,
3734 MVT::v16i8, &pshufbMask[0], 16));
3735 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3738 // No SSSE3 - Calculate in place words and then fix all out of place words
3739 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3740 // the 16 different words that comprise the two doublequadword input vectors.
3741 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3742 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3743 SDValue NewV = V2Only ? V2 : V1;
3744 for (int i = 0; i != 8; ++i) {
3745 int Elt0 = MaskVals[i*2];
3746 int Elt1 = MaskVals[i*2+1];
3748 // This word of the result is all undef, skip it.
3749 if (Elt0 < 0 && Elt1 < 0)
3752 // This word of the result is already in the correct place, skip it.
3753 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3755 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3758 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3759 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3762 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3763 // using a single extract together, load it and store it.
3764 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3765 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3766 DAG.getIntPtrConstant(Elt1 / 2));
3767 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3768 DAG.getIntPtrConstant(i));
3772 // If Elt1 is defined, extract it from the appropriate source. If the
3773 // source byte is not also odd, shift the extracted word left 8 bits
3774 // otherwise clear the bottom 8 bits if we need to do an or.
3776 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3777 DAG.getIntPtrConstant(Elt1 / 2));
3778 if ((Elt1 & 1) == 0)
3779 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3780 DAG.getConstant(8, TLI.getShiftAmountTy()));
3782 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3783 DAG.getConstant(0xFF00, MVT::i16));
3785 // If Elt0 is defined, extract it from the appropriate source. If the
3786 // source byte is not also even, shift the extracted word right 8 bits. If
3787 // Elt1 was also defined, OR the extracted values together before
3788 // inserting them in the result.
3790 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3791 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3792 if ((Elt0 & 1) != 0)
3793 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3794 DAG.getConstant(8, TLI.getShiftAmountTy()));
3796 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3797 DAG.getConstant(0x00FF, MVT::i16));
3798 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3801 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3802 DAG.getIntPtrConstant(i));
3804 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3807 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3808 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3809 /// done when every pair / quad of shuffle mask elements point to elements in
3810 /// the right sequence. e.g.
3811 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3813 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3815 TargetLowering &TLI, DebugLoc dl) {
3816 MVT VT = SVOp->getValueType(0);
3817 SDValue V1 = SVOp->getOperand(0);
3818 SDValue V2 = SVOp->getOperand(1);
3819 unsigned NumElems = VT.getVectorNumElements();
3820 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3821 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3822 MVT MaskEltVT = MaskVT.getVectorElementType();
3824 switch (VT.getSimpleVT()) {
3825 default: assert(false && "Unexpected!");
3826 case MVT::v4f32: NewVT = MVT::v2f64; break;
3827 case MVT::v4i32: NewVT = MVT::v2i64; break;
3828 case MVT::v8i16: NewVT = MVT::v4i32; break;
3829 case MVT::v16i8: NewVT = MVT::v4i32; break;
3832 if (NewWidth == 2) {
3838 int Scale = NumElems / NewWidth;
3839 SmallVector<int, 8> MaskVec;
3840 for (unsigned i = 0; i < NumElems; i += Scale) {
3842 for (int j = 0; j < Scale; ++j) {
3843 int EltIdx = SVOp->getMaskElt(i+j);
3847 StartIdx = EltIdx - (EltIdx % Scale);
3848 if (EltIdx != StartIdx + j)
3852 MaskVec.push_back(-1);
3854 MaskVec.push_back(StartIdx / Scale);
3857 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3858 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3859 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3862 /// getVZextMovL - Return a zero-extending vector move low node.
3864 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3865 SDValue SrcOp, SelectionDAG &DAG,
3866 const X86Subtarget *Subtarget, DebugLoc dl) {
3867 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3868 LoadSDNode *LD = NULL;
3869 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3870 LD = dyn_cast<LoadSDNode>(SrcOp);
3872 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3874 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3875 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3876 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3877 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3878 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3880 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3881 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3882 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3883 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3891 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3892 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3893 DAG.getNode(ISD::BIT_CONVERT, dl,
3897 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3900 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3901 SDValue V1 = SVOp->getOperand(0);
3902 SDValue V2 = SVOp->getOperand(1);
3903 DebugLoc dl = SVOp->getDebugLoc();
3904 MVT VT = SVOp->getValueType(0);
3906 SmallVector<std::pair<int, int>, 8> Locs;
3908 SmallVector<int, 8> Mask1(4U, -1);
3909 SmallVector<int, 8> PermMask;
3910 SVOp->getMask(PermMask);
3914 for (unsigned i = 0; i != 4; ++i) {
3915 int Idx = PermMask[i];
3917 Locs[i] = std::make_pair(-1, -1);
3919 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3921 Locs[i] = std::make_pair(0, NumLo);
3925 Locs[i] = std::make_pair(1, NumHi);
3927 Mask1[2+NumHi] = Idx;
3933 if (NumLo <= 2 && NumHi <= 2) {
3934 // If no more than two elements come from either vector. This can be
3935 // implemented with two shuffles. First shuffle gather the elements.
3936 // The second shuffle, which takes the first shuffle as both of its
3937 // vector operands, put the elements into the right order.
3938 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3940 SmallVector<int, 8> Mask2(4U, -1);
3942 for (unsigned i = 0; i != 4; ++i) {
3943 if (Locs[i].first == -1)
3946 unsigned Idx = (i < 2) ? 0 : 4;
3947 Idx += Locs[i].first * 2 + Locs[i].second;
3952 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3953 } else if (NumLo == 3 || NumHi == 3) {
3954 // Otherwise, we must have three elements from one vector, call it X, and
3955 // one element from the other, call it Y. First, use a shufps to build an
3956 // intermediate vector with the one element from Y and the element from X
3957 // that will be in the same half in the final destination (the indexes don't
3958 // matter). Then, use a shufps to build the final vector, taking the half
3959 // containing the element from Y from the intermediate, and the other half
3962 // Normalize it so the 3 elements come from V1.
3963 CommuteVectorShuffleMask(PermMask, VT);
3967 // Find the element from V2.
3969 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3970 int Val = PermMask[HiIndex];
3977 Mask1[0] = PermMask[HiIndex];
3979 Mask1[2] = PermMask[HiIndex^1];
3981 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3984 Mask1[0] = PermMask[0];
3985 Mask1[1] = PermMask[1];
3986 Mask1[2] = HiIndex & 1 ? 6 : 4;
3987 Mask1[3] = HiIndex & 1 ? 4 : 6;
3988 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3990 Mask1[0] = HiIndex & 1 ? 2 : 0;
3991 Mask1[1] = HiIndex & 1 ? 0 : 2;
3992 Mask1[2] = PermMask[2];
3993 Mask1[3] = PermMask[3];
3998 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4002 // Break it into (shuffle shuffle_hi, shuffle_lo).
4004 SmallVector<int,8> LoMask(4U, -1);
4005 SmallVector<int,8> HiMask(4U, -1);
4007 SmallVector<int,8> *MaskPtr = &LoMask;
4008 unsigned MaskIdx = 0;
4011 for (unsigned i = 0; i != 4; ++i) {
4018 int Idx = PermMask[i];
4020 Locs[i] = std::make_pair(-1, -1);
4021 } else if (Idx < 4) {
4022 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4023 (*MaskPtr)[LoIdx] = Idx;
4026 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4027 (*MaskPtr)[HiIdx] = Idx;
4032 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4033 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4034 SmallVector<int, 8> MaskOps;
4035 for (unsigned i = 0; i != 4; ++i) {
4036 if (Locs[i].first == -1) {
4037 MaskOps.push_back(-1);
4039 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4040 MaskOps.push_back(Idx);
4043 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4047 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4048 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4049 SDValue V1 = Op.getOperand(0);
4050 SDValue V2 = Op.getOperand(1);
4051 MVT VT = Op.getValueType();
4052 DebugLoc dl = Op.getDebugLoc();
4053 unsigned NumElems = VT.getVectorNumElements();
4054 bool isMMX = VT.getSizeInBits() == 64;
4055 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4056 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4057 bool V1IsSplat = false;
4058 bool V2IsSplat = false;
4060 if (isZeroShuffle(SVOp))
4061 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4063 // Promote splats to v4f32.
4064 if (SVOp->isSplat()) {
4065 if (isMMX || NumElems < 4)
4067 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4070 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4072 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4073 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4074 if (NewOp.getNode())
4075 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4076 LowerVECTOR_SHUFFLE(NewOp, DAG));
4077 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4078 // FIXME: Figure out a cleaner way to do this.
4079 // Try to make use of movq to zero out the top part.
4080 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4081 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4082 if (NewOp.getNode()) {
4083 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4084 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4085 DAG, Subtarget, dl);
4087 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4088 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4089 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4090 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4091 DAG, Subtarget, dl);
4095 if (X86::isPSHUFDMask(SVOp))
4098 // Check if this can be converted into a logical shift.
4099 bool isLeft = false;
4102 bool isShift = getSubtarget()->hasSSE2() &&
4103 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4104 if (isShift && ShVal.hasOneUse()) {
4105 // If the shifted value has multiple uses, it may be cheaper to use
4106 // v_set0 + movlhps or movhlps, etc.
4107 MVT EVT = VT.getVectorElementType();
4108 ShAmt *= EVT.getSizeInBits();
4109 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4112 if (X86::isMOVLMask(SVOp)) {
4115 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4116 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4121 // FIXME: fold these into legal mask.
4122 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4123 X86::isMOVSLDUPMask(SVOp) ||
4124 X86::isMOVHLPSMask(SVOp) ||
4125 X86::isMOVHPMask(SVOp) ||
4126 X86::isMOVLPMask(SVOp)))
4129 if (ShouldXformToMOVHLPS(SVOp) ||
4130 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4131 return CommuteVectorShuffle(SVOp, DAG);
4134 // No better options. Use a vshl / vsrl.
4135 MVT EVT = VT.getVectorElementType();
4136 ShAmt *= EVT.getSizeInBits();
4137 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4140 bool Commuted = false;
4141 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4142 // 1,1,1,1 -> v8i16 though.
4143 V1IsSplat = isSplatVector(V1.getNode());
4144 V2IsSplat = isSplatVector(V2.getNode());
4146 // Canonicalize the splat or undef, if present, to be on the RHS.
4147 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4148 Op = CommuteVectorShuffle(SVOp, DAG);
4149 SVOp = cast<ShuffleVectorSDNode>(Op);
4150 V1 = SVOp->getOperand(0);
4151 V2 = SVOp->getOperand(1);
4152 std::swap(V1IsSplat, V2IsSplat);
4153 std::swap(V1IsUndef, V2IsUndef);
4157 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4158 // Shuffling low element of v1 into undef, just return v1.
4161 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4162 // the instruction selector will not match, so get a canonical MOVL with
4163 // swapped operands to undo the commute.
4164 return getMOVL(DAG, dl, VT, V2, V1);
4167 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4168 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4169 X86::isUNPCKLMask(SVOp) ||
4170 X86::isUNPCKHMask(SVOp))
4174 // Normalize mask so all entries that point to V2 points to its first
4175 // element then try to match unpck{h|l} again. If match, return a
4176 // new vector_shuffle with the corrected mask.
4177 SDValue NewMask = NormalizeMask(SVOp, DAG);
4178 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4179 if (NSVOp != SVOp) {
4180 if (X86::isUNPCKLMask(NSVOp, true)) {
4182 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4189 // Commute is back and try unpck* again.
4190 // FIXME: this seems wrong.
4191 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4192 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4193 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4194 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4195 X86::isUNPCKLMask(NewSVOp) ||
4196 X86::isUNPCKHMask(NewSVOp))
4200 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4202 // Normalize the node to match x86 shuffle ops if needed
4203 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4204 return CommuteVectorShuffle(SVOp, DAG);
4206 // Check for legal shuffle and return?
4207 SmallVector<int, 16> PermMask;
4208 SVOp->getMask(PermMask);
4209 if (isShuffleMaskLegal(PermMask, VT))
4212 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4213 if (VT == MVT::v8i16) {
4214 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4215 if (NewOp.getNode())
4219 if (VT == MVT::v16i8) {
4220 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4221 if (NewOp.getNode())
4225 // Handle all 4 wide cases with a number of shuffles except for MMX.
4226 if (NumElems == 4 && !isMMX)
4227 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4233 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4234 SelectionDAG &DAG) {
4235 MVT VT = Op.getValueType();
4236 DebugLoc dl = Op.getDebugLoc();
4237 if (VT.getSizeInBits() == 8) {
4238 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4239 Op.getOperand(0), Op.getOperand(1));
4240 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4241 DAG.getValueType(VT));
4242 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4243 } else if (VT.getSizeInBits() == 16) {
4244 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4245 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4247 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4248 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4249 DAG.getNode(ISD::BIT_CONVERT, dl,
4253 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4254 Op.getOperand(0), Op.getOperand(1));
4255 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4256 DAG.getValueType(VT));
4257 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4258 } else if (VT == MVT::f32) {
4259 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4260 // the result back to FR32 register. It's only worth matching if the
4261 // result has a single use which is a store or a bitcast to i32. And in
4262 // the case of a store, it's not worth it if the index is a constant 0,
4263 // because a MOVSSmr can be used instead, which is smaller and faster.
4264 if (!Op.hasOneUse())
4266 SDNode *User = *Op.getNode()->use_begin();
4267 if ((User->getOpcode() != ISD::STORE ||
4268 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4269 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4270 (User->getOpcode() != ISD::BIT_CONVERT ||
4271 User->getValueType(0) != MVT::i32))
4273 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4274 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4277 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4278 } else if (VT == MVT::i32) {
4279 // ExtractPS works with constant index.
4280 if (isa<ConstantSDNode>(Op.getOperand(1)))
4288 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4289 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4292 if (Subtarget->hasSSE41()) {
4293 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4298 MVT VT = Op.getValueType();
4299 DebugLoc dl = Op.getDebugLoc();
4300 // TODO: handle v16i8.
4301 if (VT.getSizeInBits() == 16) {
4302 SDValue Vec = Op.getOperand(0);
4303 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4305 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4306 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4307 DAG.getNode(ISD::BIT_CONVERT, dl,
4310 // Transform it so it match pextrw which produces a 32-bit result.
4311 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4312 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4313 Op.getOperand(0), Op.getOperand(1));
4314 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4315 DAG.getValueType(VT));
4316 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4317 } else if (VT.getSizeInBits() == 32) {
4318 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4322 // SHUFPS the element to the lowest double word, then movss.
4323 int Mask[4] = { Idx, -1, -1, -1 };
4324 MVT VVT = Op.getOperand(0).getValueType();
4325 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4326 DAG.getUNDEF(VVT), Mask);
4327 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4328 DAG.getIntPtrConstant(0));
4329 } else if (VT.getSizeInBits() == 64) {
4330 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4331 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4332 // to match extract_elt for f64.
4333 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4337 // UNPCKHPD the element to the lowest double word, then movsd.
4338 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4339 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4340 int Mask[2] = { 1, -1 };
4341 MVT VVT = Op.getOperand(0).getValueType();
4342 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4343 DAG.getUNDEF(VVT), Mask);
4344 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4345 DAG.getIntPtrConstant(0));
4352 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4353 MVT VT = Op.getValueType();
4354 MVT EVT = VT.getVectorElementType();
4355 DebugLoc dl = Op.getDebugLoc();
4357 SDValue N0 = Op.getOperand(0);
4358 SDValue N1 = Op.getOperand(1);
4359 SDValue N2 = Op.getOperand(2);
4361 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4362 isa<ConstantSDNode>(N2)) {
4363 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4365 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4367 if (N1.getValueType() != MVT::i32)
4368 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4369 if (N2.getValueType() != MVT::i32)
4370 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4371 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4372 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4373 // Bits [7:6] of the constant are the source select. This will always be
4374 // zero here. The DAG Combiner may combine an extract_elt index into these
4375 // bits. For example (insert (extract, 3), 2) could be matched by putting
4376 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4377 // Bits [5:4] of the constant are the destination select. This is the
4378 // value of the incoming immediate.
4379 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4380 // combine either bitwise AND or insert of float 0.0 to set these bits.
4381 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4382 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4383 } else if (EVT == MVT::i32) {
4384 // InsertPS works with constant index.
4385 if (isa<ConstantSDNode>(N2))
4392 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4393 MVT VT = Op.getValueType();
4394 MVT EVT = VT.getVectorElementType();
4396 if (Subtarget->hasSSE41())
4397 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4402 DebugLoc dl = Op.getDebugLoc();
4403 SDValue N0 = Op.getOperand(0);
4404 SDValue N1 = Op.getOperand(1);
4405 SDValue N2 = Op.getOperand(2);
4407 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4408 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4409 // as its second argument.
4410 if (N1.getValueType() != MVT::i32)
4411 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4412 if (N2.getValueType() != MVT::i32)
4413 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4414 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4420 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4421 DebugLoc dl = Op.getDebugLoc();
4422 if (Op.getValueType() == MVT::v2f32)
4423 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4424 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4425 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4426 Op.getOperand(0))));
4428 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4429 MVT VT = MVT::v2i32;
4430 switch (Op.getValueType().getSimpleVT()) {
4437 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4438 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4441 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4442 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4443 // one of the above mentioned nodes. It has to be wrapped because otherwise
4444 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4445 // be used to form addressing mode. These wrapped nodes will be selected
4448 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4449 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4451 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4453 unsigned char OpFlag = 0;
4454 unsigned WrapperKind = X86ISD::Wrapper;
4456 if (Subtarget->is64Bit() &&
4457 getTargetMachine().getCodeModel() == CodeModel::Small) {
4458 WrapperKind = X86ISD::WrapperRIP;
4459 } else if (Subtarget->isPICStyleGOT()) {
4460 OpFlag = X86II::MO_GOTOFF;
4461 } else if (Subtarget->isPICStyleStub() &&
4462 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4463 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4466 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4468 CP->getOffset(), OpFlag);
4469 DebugLoc DL = CP->getDebugLoc();
4470 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4471 // With PIC, the address is actually $g + Offset.
4473 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4474 DAG.getNode(X86ISD::GlobalBaseReg,
4475 DebugLoc::getUnknownLoc(), getPointerTy()),
4482 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4483 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4485 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4487 unsigned char OpFlag = 0;
4488 unsigned WrapperKind = X86ISD::Wrapper;
4490 if (Subtarget->is64Bit()) {
4491 WrapperKind = X86ISD::WrapperRIP;
4492 } else if (Subtarget->isPICStyleGOT()) {
4493 OpFlag = X86II::MO_GOTOFF;
4494 } else if (Subtarget->isPICStyleStub() &&
4495 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4496 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4499 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4501 DebugLoc DL = JT->getDebugLoc();
4502 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4504 // With PIC, the address is actually $g + Offset.
4506 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4507 DAG.getNode(X86ISD::GlobalBaseReg,
4508 DebugLoc::getUnknownLoc(), getPointerTy()),
4516 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4517 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4519 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4521 unsigned char OpFlag = 0;
4522 unsigned WrapperKind = X86ISD::Wrapper;
4523 if (Subtarget->is64Bit()) {
4524 WrapperKind = X86ISD::WrapperRIP;
4525 } else if (Subtarget->isPICStyleGOT()) {
4526 OpFlag = X86II::MO_GOTOFF;
4527 } else if (Subtarget->isPICStyleStub() &&
4528 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4529 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4532 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4534 DebugLoc DL = Op.getDebugLoc();
4535 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4538 // With PIC, the address is actually $g + Offset.
4539 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4540 !Subtarget->is64Bit()) {
4541 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4542 DAG.getNode(X86ISD::GlobalBaseReg,
4543 DebugLoc::getUnknownLoc(),
4552 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4554 SelectionDAG &DAG) const {
4555 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4557 // Create the TargetGlobalAddress node, folding in the constant
4558 // offset if it is legal.
4559 unsigned char OpFlags =
4560 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4561 bool ExtraLoadRequired = isGlobalStubReference(OpFlags);
4565 if (!IsPIC && !ExtraLoadRequired && isInt32(Offset)) {
4566 // A direct static reference to a global.
4567 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4570 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4573 if (Subtarget->is64Bit() &&
4574 getTargetMachine().getCodeModel() == CodeModel::Small)
4575 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4577 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4579 // With PIC, the address is actually $g + Offset.
4580 if (IsPIC && !Subtarget->is64Bit()) {
4581 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4582 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4586 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4587 // load the value at address GV, not the value of GV itself. This means that
4588 // the GlobalAddress must be in the base or index register of the address, not
4589 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4590 // The same applies for external symbols during PIC codegen
4591 if (ExtraLoadRequired)
4592 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4593 PseudoSourceValue::getGOT(), 0);
4595 // If there was a non-zero offset that we didn't fold, create an explicit
4598 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4599 DAG.getConstant(Offset, getPointerTy()));
4605 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4606 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4607 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4608 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4612 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4613 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4614 unsigned char OperandFlags) {
4615 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4616 DebugLoc dl = GA->getDebugLoc();
4617 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4618 GA->getValueType(0),
4622 SDValue Ops[] = { Chain, TGA, *InFlag };
4623 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4625 SDValue Ops[] = { Chain, TGA };
4626 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4628 SDValue Flag = Chain.getValue(1);
4629 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4632 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4634 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4637 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4638 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4639 DAG.getNode(X86ISD::GlobalBaseReg,
4640 DebugLoc::getUnknownLoc(),
4642 InFlag = Chain.getValue(1);
4644 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4647 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4649 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4651 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4652 X86::RAX, X86II::MO_TLSGD);
4655 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4656 // "local exec" model.
4657 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4658 const MVT PtrVT, TLSModel::Model model,
4660 DebugLoc dl = GA->getDebugLoc();
4661 // Get the Thread Pointer
4662 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4663 DebugLoc::getUnknownLoc(), PtrVT,
4664 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4667 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4670 unsigned char OperandFlags = 0;
4671 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4673 unsigned WrapperKind = X86ISD::Wrapper;
4674 if (model == TLSModel::LocalExec) {
4675 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4676 } else if (is64Bit) {
4677 assert(model == TLSModel::InitialExec);
4678 OperandFlags = X86II::MO_GOTTPOFF;
4679 WrapperKind = X86ISD::WrapperRIP;
4681 assert(model == TLSModel::InitialExec);
4682 OperandFlags = X86II::MO_INDNTPOFF;
4685 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4687 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4688 GA->getOffset(), OperandFlags);
4689 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4691 if (model == TLSModel::InitialExec)
4692 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4693 PseudoSourceValue::getGOT(), 0);
4695 // The address of the thread local variable is the add of the thread
4696 // pointer with the offset of the variable.
4697 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4701 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4702 // TODO: implement the "local dynamic" model
4703 // TODO: implement the "initial exec"model for pic executables
4704 assert(Subtarget->isTargetELF() &&
4705 "TLS not implemented for non-ELF targets");
4706 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4707 const GlobalValue *GV = GA->getGlobal();
4709 // If GV is an alias then use the aliasee for determining
4710 // thread-localness.
4711 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4712 GV = GA->resolveAliasedGlobal(false);
4714 TLSModel::Model model = getTLSModel(GV,
4715 getTargetMachine().getRelocationModel());
4718 case TLSModel::GeneralDynamic:
4719 case TLSModel::LocalDynamic: // not implemented
4720 if (Subtarget->is64Bit())
4721 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4722 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4724 case TLSModel::InitialExec:
4725 case TLSModel::LocalExec:
4726 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4727 Subtarget->is64Bit());
4730 assert(0 && "Unreachable");
4735 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4736 /// take a 2 x i32 value to shift plus a shift amount.
4737 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4738 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4739 MVT VT = Op.getValueType();
4740 unsigned VTBits = VT.getSizeInBits();
4741 DebugLoc dl = Op.getDebugLoc();
4742 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4743 SDValue ShOpLo = Op.getOperand(0);
4744 SDValue ShOpHi = Op.getOperand(1);
4745 SDValue ShAmt = Op.getOperand(2);
4746 SDValue Tmp1 = isSRA ?
4747 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4748 DAG.getConstant(VTBits - 1, MVT::i8)) :
4749 DAG.getConstant(0, VT);
4752 if (Op.getOpcode() == ISD::SHL_PARTS) {
4753 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4754 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4756 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4757 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4760 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4761 DAG.getConstant(VTBits, MVT::i8));
4762 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4763 AndNode, DAG.getConstant(0, MVT::i8));
4766 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4767 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4768 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4770 if (Op.getOpcode() == ISD::SHL_PARTS) {
4771 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4772 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4774 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4775 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4778 SDValue Ops[2] = { Lo, Hi };
4779 return DAG.getMergeValues(Ops, 2, dl);
4782 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4783 MVT SrcVT = Op.getOperand(0).getValueType();
4785 if (SrcVT.isVector()) {
4786 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4792 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4793 "Unknown SINT_TO_FP to lower!");
4795 // These are really Legal; return the operand so the caller accepts it as
4797 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4799 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4800 Subtarget->is64Bit()) {
4804 DebugLoc dl = Op.getDebugLoc();
4805 unsigned Size = SrcVT.getSizeInBits()/8;
4806 MachineFunction &MF = DAG.getMachineFunction();
4807 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4808 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4809 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4811 PseudoSourceValue::getFixedStack(SSFI), 0);
4812 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4815 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4817 SelectionDAG &DAG) {
4819 DebugLoc dl = Op.getDebugLoc();
4821 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4823 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4825 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4826 SmallVector<SDValue, 8> Ops;
4827 Ops.push_back(Chain);
4828 Ops.push_back(StackSlot);
4829 Ops.push_back(DAG.getValueType(SrcVT));
4830 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4831 Tys, &Ops[0], Ops.size());
4834 Chain = Result.getValue(1);
4835 SDValue InFlag = Result.getValue(2);
4837 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4838 // shouldn't be necessary except that RFP cannot be live across
4839 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4840 MachineFunction &MF = DAG.getMachineFunction();
4841 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4842 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4843 Tys = DAG.getVTList(MVT::Other);
4844 SmallVector<SDValue, 8> Ops;
4845 Ops.push_back(Chain);
4846 Ops.push_back(Result);
4847 Ops.push_back(StackSlot);
4848 Ops.push_back(DAG.getValueType(Op.getValueType()));
4849 Ops.push_back(InFlag);
4850 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4851 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4852 PseudoSourceValue::getFixedStack(SSFI), 0);
4858 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4859 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4860 // This algorithm is not obvious. Here it is in C code, more or less:
4862 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4863 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4864 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4866 // Copy ints to xmm registers.
4867 __m128i xh = _mm_cvtsi32_si128( hi );
4868 __m128i xl = _mm_cvtsi32_si128( lo );
4870 // Combine into low half of a single xmm register.
4871 __m128i x = _mm_unpacklo_epi32( xh, xl );
4875 // Merge in appropriate exponents to give the integer bits the right
4877 x = _mm_unpacklo_epi32( x, exp );
4879 // Subtract away the biases to deal with the IEEE-754 double precision
4881 d = _mm_sub_pd( (__m128d) x, bias );
4883 // All conversions up to here are exact. The correctly rounded result is
4884 // calculated using the current rounding mode using the following
4886 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4887 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4888 // store doesn't really need to be here (except
4889 // maybe to zero the other double)
4894 DebugLoc dl = Op.getDebugLoc();
4896 // Build some magic constants.
4897 std::vector<Constant*> CV0;
4898 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4899 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4900 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4901 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4902 Constant *C0 = ConstantVector::get(CV0);
4903 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4905 std::vector<Constant*> CV1;
4906 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4907 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4908 Constant *C1 = ConstantVector::get(CV1);
4909 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4911 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4912 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4914 DAG.getIntPtrConstant(1)));
4915 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4916 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4918 DAG.getIntPtrConstant(0)));
4919 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4920 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4921 PseudoSourceValue::getConstantPool(), 0,
4923 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4924 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4925 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4926 PseudoSourceValue::getConstantPool(), 0,
4928 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4930 // Add the halves; easiest way is to swap them into another reg first.
4931 int ShufMask[2] = { 1, -1 };
4932 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4933 DAG.getUNDEF(MVT::v2f64), ShufMask);
4934 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4935 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4936 DAG.getIntPtrConstant(0));
4939 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4940 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4941 DebugLoc dl = Op.getDebugLoc();
4942 // FP constant to bias correct the final result.
4943 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4946 // Load the 32-bit value into an XMM register.
4947 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4948 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4950 DAG.getIntPtrConstant(0)));
4952 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4953 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4954 DAG.getIntPtrConstant(0));
4956 // Or the load with the bias.
4957 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4958 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4959 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4961 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4962 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4963 MVT::v2f64, Bias)));
4964 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4965 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4966 DAG.getIntPtrConstant(0));
4968 // Subtract the bias.
4969 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4971 // Handle final rounding.
4972 MVT DestVT = Op.getValueType();
4974 if (DestVT.bitsLT(MVT::f64)) {
4975 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4976 DAG.getIntPtrConstant(0));
4977 } else if (DestVT.bitsGT(MVT::f64)) {
4978 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4981 // Handle final rounding.
4985 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4986 SDValue N0 = Op.getOperand(0);
4987 DebugLoc dl = Op.getDebugLoc();
4989 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4990 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4991 // the optimization here.
4992 if (DAG.SignBitIsZero(N0))
4993 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4995 MVT SrcVT = N0.getValueType();
4996 if (SrcVT == MVT::i64) {
4997 // We only handle SSE2 f64 target here; caller can expand the rest.
4998 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5001 return LowerUINT_TO_FP_i64(Op, DAG);
5002 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5003 return LowerUINT_TO_FP_i32(Op, DAG);
5006 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5008 // Make a 64-bit buffer, and use it to build an FILD.
5009 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5010 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5011 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5012 getPointerTy(), StackSlot, WordOff);
5013 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5014 StackSlot, NULL, 0);
5015 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5016 OffsetSlot, NULL, 0);
5017 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5020 std::pair<SDValue,SDValue> X86TargetLowering::
5021 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5022 DebugLoc dl = Op.getDebugLoc();
5024 MVT DstTy = Op.getValueType();
5027 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5031 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5032 DstTy.getSimpleVT() >= MVT::i16 &&
5033 "Unknown FP_TO_SINT to lower!");
5035 // These are really Legal.
5036 if (DstTy == MVT::i32 &&
5037 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5038 return std::make_pair(SDValue(), SDValue());
5039 if (Subtarget->is64Bit() &&
5040 DstTy == MVT::i64 &&
5041 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5042 return std::make_pair(SDValue(), SDValue());
5044 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5046 MachineFunction &MF = DAG.getMachineFunction();
5047 unsigned MemSize = DstTy.getSizeInBits()/8;
5048 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5049 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5052 switch (DstTy.getSimpleVT()) {
5053 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5054 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5055 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5056 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5059 SDValue Chain = DAG.getEntryNode();
5060 SDValue Value = Op.getOperand(0);
5061 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5062 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5063 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5064 PseudoSourceValue::getFixedStack(SSFI), 0);
5065 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5067 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5069 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5070 Chain = Value.getValue(1);
5071 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5072 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5075 // Build the FP_TO_INT*_IN_MEM
5076 SDValue Ops[] = { Chain, Value, StackSlot };
5077 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5079 return std::make_pair(FIST, StackSlot);
5082 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5083 if (Op.getValueType().isVector()) {
5084 if (Op.getValueType() == MVT::v2i32 &&
5085 Op.getOperand(0).getValueType() == MVT::v2f64) {
5091 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5092 SDValue FIST = Vals.first, StackSlot = Vals.second;
5093 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5094 if (FIST.getNode() == 0) return Op;
5097 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5098 FIST, StackSlot, NULL, 0);
5101 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5102 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5103 SDValue FIST = Vals.first, StackSlot = Vals.second;
5104 assert(FIST.getNode() && "Unexpected failure");
5107 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5108 FIST, StackSlot, NULL, 0);
5111 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5112 DebugLoc dl = Op.getDebugLoc();
5113 MVT VT = Op.getValueType();
5116 EltVT = VT.getVectorElementType();
5117 std::vector<Constant*> CV;
5118 if (EltVT == MVT::f64) {
5119 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5123 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5129 Constant *C = ConstantVector::get(CV);
5130 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5131 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5132 PseudoSourceValue::getConstantPool(), 0,
5134 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5137 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5138 DebugLoc dl = Op.getDebugLoc();
5139 MVT VT = Op.getValueType();
5141 unsigned EltNum = 1;
5142 if (VT.isVector()) {
5143 EltVT = VT.getVectorElementType();
5144 EltNum = VT.getVectorNumElements();
5146 std::vector<Constant*> CV;
5147 if (EltVT == MVT::f64) {
5148 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5152 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5158 Constant *C = ConstantVector::get(CV);
5159 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5160 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5161 PseudoSourceValue::getConstantPool(), 0,
5163 if (VT.isVector()) {
5164 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5165 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5166 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5168 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5170 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5174 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5175 SDValue Op0 = Op.getOperand(0);
5176 SDValue Op1 = Op.getOperand(1);
5177 DebugLoc dl = Op.getDebugLoc();
5178 MVT VT = Op.getValueType();
5179 MVT SrcVT = Op1.getValueType();
5181 // If second operand is smaller, extend it first.
5182 if (SrcVT.bitsLT(VT)) {
5183 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5186 // And if it is bigger, shrink it first.
5187 if (SrcVT.bitsGT(VT)) {
5188 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5192 // At this point the operands and the result should have the same
5193 // type, and that won't be f80 since that is not custom lowered.
5195 // First get the sign bit of second operand.
5196 std::vector<Constant*> CV;
5197 if (SrcVT == MVT::f64) {
5198 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5199 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5201 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5202 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5203 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5204 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5206 Constant *C = ConstantVector::get(CV);
5207 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5208 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5209 PseudoSourceValue::getConstantPool(), 0,
5211 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5213 // Shift sign bit right or left if the two operands have different types.
5214 if (SrcVT.bitsGT(VT)) {
5215 // Op0 is MVT::f32, Op1 is MVT::f64.
5216 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5217 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5218 DAG.getConstant(32, MVT::i32));
5219 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5220 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5221 DAG.getIntPtrConstant(0));
5224 // Clear first operand sign bit.
5226 if (VT == MVT::f64) {
5227 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5228 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5230 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5231 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5232 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5233 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5235 C = ConstantVector::get(CV);
5236 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5237 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5238 PseudoSourceValue::getConstantPool(), 0,
5240 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5242 // Or the value with the sign bit.
5243 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5246 /// Emit nodes that will be selected as "test Op0,Op0", or something
5248 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5249 SelectionDAG &DAG) {
5250 DebugLoc dl = Op.getDebugLoc();
5252 // CF and OF aren't always set the way we want. Determine which
5253 // of these we need.
5254 bool NeedCF = false;
5255 bool NeedOF = false;
5257 case X86::COND_A: case X86::COND_AE:
5258 case X86::COND_B: case X86::COND_BE:
5261 case X86::COND_G: case X86::COND_GE:
5262 case X86::COND_L: case X86::COND_LE:
5263 case X86::COND_O: case X86::COND_NO:
5269 // See if we can use the EFLAGS value from the operand instead of
5270 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5271 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5272 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5273 unsigned Opcode = 0;
5274 unsigned NumOperands = 0;
5275 switch (Op.getNode()->getOpcode()) {
5277 // Due to an isel shortcoming, be conservative if this add is likely to
5278 // be selected as part of a load-modify-store instruction. When the root
5279 // node in a match is a store, isel doesn't know how to remap non-chain
5280 // non-flag uses of other nodes in the match, such as the ADD in this
5281 // case. This leads to the ADD being left around and reselected, with
5282 // the result being two adds in the output.
5283 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5284 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5285 if (UI->getOpcode() == ISD::STORE)
5287 if (ConstantSDNode *C =
5288 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5289 // An add of one will be selected as an INC.
5290 if (C->getAPIntValue() == 1) {
5291 Opcode = X86ISD::INC;
5295 // An add of negative one (subtract of one) will be selected as a DEC.
5296 if (C->getAPIntValue().isAllOnesValue()) {
5297 Opcode = X86ISD::DEC;
5302 // Otherwise use a regular EFLAGS-setting add.
5303 Opcode = X86ISD::ADD;
5307 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5308 // likely to be selected as part of a load-modify-store instruction.
5309 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5310 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5311 if (UI->getOpcode() == ISD::STORE)
5313 // Otherwise use a regular EFLAGS-setting sub.
5314 Opcode = X86ISD::SUB;
5321 return SDValue(Op.getNode(), 1);
5327 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5328 SmallVector<SDValue, 4> Ops;
5329 for (unsigned i = 0; i != NumOperands; ++i)
5330 Ops.push_back(Op.getOperand(i));
5331 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5332 DAG.ReplaceAllUsesWith(Op, New);
5333 return SDValue(New.getNode(), 1);
5337 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5338 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5339 DAG.getConstant(0, Op.getValueType()));
5342 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5344 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5345 SelectionDAG &DAG) {
5346 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5347 if (C->getAPIntValue() == 0)
5348 return EmitTest(Op0, X86CC, DAG);
5350 DebugLoc dl = Op0.getDebugLoc();
5351 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5354 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5355 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5356 SDValue Op0 = Op.getOperand(0);
5357 SDValue Op1 = Op.getOperand(1);
5358 DebugLoc dl = Op.getDebugLoc();
5359 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5361 // Lower (X & (1 << N)) == 0 to BT(X, N).
5362 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5363 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5364 if (Op0.getOpcode() == ISD::AND &&
5366 Op1.getOpcode() == ISD::Constant &&
5367 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5368 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5370 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5371 if (ConstantSDNode *Op010C =
5372 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5373 if (Op010C->getZExtValue() == 1) {
5374 LHS = Op0.getOperand(0);
5375 RHS = Op0.getOperand(1).getOperand(1);
5377 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5378 if (ConstantSDNode *Op000C =
5379 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5380 if (Op000C->getZExtValue() == 1) {
5381 LHS = Op0.getOperand(1);
5382 RHS = Op0.getOperand(0).getOperand(1);
5384 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5385 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5386 SDValue AndLHS = Op0.getOperand(0);
5387 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5388 LHS = AndLHS.getOperand(0);
5389 RHS = AndLHS.getOperand(1);
5393 if (LHS.getNode()) {
5394 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5395 // instruction. Since the shift amount is in-range-or-undefined, we know
5396 // that doing a bittest on the i16 value is ok. We extend to i32 because
5397 // the encoding for the i16 version is larger than the i32 version.
5398 if (LHS.getValueType() == MVT::i8)
5399 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5401 // If the operand types disagree, extend the shift amount to match. Since
5402 // BT ignores high bits (like shifts) we can use anyextend.
5403 if (LHS.getValueType() != RHS.getValueType())
5404 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5406 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5407 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5408 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5409 DAG.getConstant(Cond, MVT::i8), BT);
5413 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5414 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5416 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5417 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5418 DAG.getConstant(X86CC, MVT::i8), Cond);
5421 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5423 SDValue Op0 = Op.getOperand(0);
5424 SDValue Op1 = Op.getOperand(1);
5425 SDValue CC = Op.getOperand(2);
5426 MVT VT = Op.getValueType();
5427 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5428 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5429 DebugLoc dl = Op.getDebugLoc();
5433 MVT VT0 = Op0.getValueType();
5434 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5435 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5438 switch (SetCCOpcode) {
5441 case ISD::SETEQ: SSECC = 0; break;
5443 case ISD::SETGT: Swap = true; // Fallthrough
5445 case ISD::SETOLT: SSECC = 1; break;
5447 case ISD::SETGE: Swap = true; // Fallthrough
5449 case ISD::SETOLE: SSECC = 2; break;
5450 case ISD::SETUO: SSECC = 3; break;
5452 case ISD::SETNE: SSECC = 4; break;
5453 case ISD::SETULE: Swap = true;
5454 case ISD::SETUGE: SSECC = 5; break;
5455 case ISD::SETULT: Swap = true;
5456 case ISD::SETUGT: SSECC = 6; break;
5457 case ISD::SETO: SSECC = 7; break;
5460 std::swap(Op0, Op1);
5462 // In the two special cases we can't handle, emit two comparisons.
5464 if (SetCCOpcode == ISD::SETUEQ) {
5466 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5467 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5468 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5470 else if (SetCCOpcode == ISD::SETONE) {
5472 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5473 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5474 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5476 assert(0 && "Illegal FP comparison");
5478 // Handle all other FP comparisons here.
5479 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5482 // We are handling one of the integer comparisons here. Since SSE only has
5483 // GT and EQ comparisons for integer, swapping operands and multiple
5484 // operations may be required for some comparisons.
5485 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5486 bool Swap = false, Invert = false, FlipSigns = false;
5488 switch (VT.getSimpleVT()) {
5490 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5491 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5492 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5493 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5496 switch (SetCCOpcode) {
5498 case ISD::SETNE: Invert = true;
5499 case ISD::SETEQ: Opc = EQOpc; break;
5500 case ISD::SETLT: Swap = true;
5501 case ISD::SETGT: Opc = GTOpc; break;
5502 case ISD::SETGE: Swap = true;
5503 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5504 case ISD::SETULT: Swap = true;
5505 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5506 case ISD::SETUGE: Swap = true;
5507 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5510 std::swap(Op0, Op1);
5512 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5513 // bits of the inputs before performing those operations.
5515 MVT EltVT = VT.getVectorElementType();
5516 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5518 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5519 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5521 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5522 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5525 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5527 // If the logical-not of the result is required, perform that now.
5529 Result = DAG.getNOT(dl, Result, VT);
5534 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5535 static bool isX86LogicalCmp(SDValue Op) {
5536 unsigned Opc = Op.getNode()->getOpcode();
5537 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5539 if (Op.getResNo() == 1 &&
5540 (Opc == X86ISD::ADD ||
5541 Opc == X86ISD::SUB ||
5542 Opc == X86ISD::SMUL ||
5543 Opc == X86ISD::UMUL ||
5544 Opc == X86ISD::INC ||
5545 Opc == X86ISD::DEC))
5551 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5552 bool addTest = true;
5553 SDValue Cond = Op.getOperand(0);
5554 DebugLoc dl = Op.getDebugLoc();
5557 if (Cond.getOpcode() == ISD::SETCC)
5558 Cond = LowerSETCC(Cond, DAG);
5560 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5561 // setting operand in place of the X86ISD::SETCC.
5562 if (Cond.getOpcode() == X86ISD::SETCC) {
5563 CC = Cond.getOperand(0);
5565 SDValue Cmp = Cond.getOperand(1);
5566 unsigned Opc = Cmp.getOpcode();
5567 MVT VT = Op.getValueType();
5569 bool IllegalFPCMov = false;
5570 if (VT.isFloatingPoint() && !VT.isVector() &&
5571 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5572 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5574 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5575 Opc == X86ISD::BT) { // FIXME
5582 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5583 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5586 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5587 SmallVector<SDValue, 4> Ops;
5588 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5589 // condition is true.
5590 Ops.push_back(Op.getOperand(2));
5591 Ops.push_back(Op.getOperand(1));
5593 Ops.push_back(Cond);
5594 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5597 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5598 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5599 // from the AND / OR.
5600 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5601 Opc = Op.getOpcode();
5602 if (Opc != ISD::OR && Opc != ISD::AND)
5604 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5605 Op.getOperand(0).hasOneUse() &&
5606 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5607 Op.getOperand(1).hasOneUse());
5610 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5611 // 1 and that the SETCC node has a single use.
5612 static bool isXor1OfSetCC(SDValue Op) {
5613 if (Op.getOpcode() != ISD::XOR)
5615 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5616 if (N1C && N1C->getAPIntValue() == 1) {
5617 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5618 Op.getOperand(0).hasOneUse();
5623 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5624 bool addTest = true;
5625 SDValue Chain = Op.getOperand(0);
5626 SDValue Cond = Op.getOperand(1);
5627 SDValue Dest = Op.getOperand(2);
5628 DebugLoc dl = Op.getDebugLoc();
5631 if (Cond.getOpcode() == ISD::SETCC)
5632 Cond = LowerSETCC(Cond, DAG);
5634 // FIXME: LowerXALUO doesn't handle these!!
5635 else if (Cond.getOpcode() == X86ISD::ADD ||
5636 Cond.getOpcode() == X86ISD::SUB ||
5637 Cond.getOpcode() == X86ISD::SMUL ||
5638 Cond.getOpcode() == X86ISD::UMUL)
5639 Cond = LowerXALUO(Cond, DAG);
5642 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5643 // setting operand in place of the X86ISD::SETCC.
5644 if (Cond.getOpcode() == X86ISD::SETCC) {
5645 CC = Cond.getOperand(0);
5647 SDValue Cmp = Cond.getOperand(1);
5648 unsigned Opc = Cmp.getOpcode();
5649 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5650 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5654 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5658 // These can only come from an arithmetic instruction with overflow,
5659 // e.g. SADDO, UADDO.
5660 Cond = Cond.getNode()->getOperand(1);
5667 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5668 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5669 if (CondOpc == ISD::OR) {
5670 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5671 // two branches instead of an explicit OR instruction with a
5673 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5674 isX86LogicalCmp(Cmp)) {
5675 CC = Cond.getOperand(0).getOperand(0);
5676 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5677 Chain, Dest, CC, Cmp);
5678 CC = Cond.getOperand(1).getOperand(0);
5682 } else { // ISD::AND
5683 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5684 // two branches instead of an explicit AND instruction with a
5685 // separate test. However, we only do this if this block doesn't
5686 // have a fall-through edge, because this requires an explicit
5687 // jmp when the condition is false.
5688 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5689 isX86LogicalCmp(Cmp) &&
5690 Op.getNode()->hasOneUse()) {
5691 X86::CondCode CCode =
5692 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5693 CCode = X86::GetOppositeBranchCondition(CCode);
5694 CC = DAG.getConstant(CCode, MVT::i8);
5695 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5696 // Look for an unconditional branch following this conditional branch.
5697 // We need this because we need to reverse the successors in order
5698 // to implement FCMP_OEQ.
5699 if (User.getOpcode() == ISD::BR) {
5700 SDValue FalseBB = User.getOperand(1);
5702 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5703 assert(NewBR == User);
5706 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5707 Chain, Dest, CC, Cmp);
5708 X86::CondCode CCode =
5709 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5710 CCode = X86::GetOppositeBranchCondition(CCode);
5711 CC = DAG.getConstant(CCode, MVT::i8);
5717 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5718 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5719 // It should be transformed during dag combiner except when the condition
5720 // is set by a arithmetics with overflow node.
5721 X86::CondCode CCode =
5722 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5723 CCode = X86::GetOppositeBranchCondition(CCode);
5724 CC = DAG.getConstant(CCode, MVT::i8);
5725 Cond = Cond.getOperand(0).getOperand(1);
5731 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5732 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5734 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5735 Chain, Dest, CC, Cond);
5739 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5740 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5741 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5742 // that the guard pages used by the OS virtual memory manager are allocated in
5743 // correct sequence.
5745 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5746 SelectionDAG &DAG) {
5747 assert(Subtarget->isTargetCygMing() &&
5748 "This should be used only on Cygwin/Mingw targets");
5749 DebugLoc dl = Op.getDebugLoc();
5752 SDValue Chain = Op.getOperand(0);
5753 SDValue Size = Op.getOperand(1);
5754 // FIXME: Ensure alignment here
5758 MVT IntPtr = getPointerTy();
5759 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5761 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5763 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5764 Flag = Chain.getValue(1);
5766 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5767 SDValue Ops[] = { Chain,
5768 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5769 DAG.getRegister(X86::EAX, IntPtr),
5770 DAG.getRegister(X86StackPtr, SPTy),
5772 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5773 Flag = Chain.getValue(1);
5775 Chain = DAG.getCALLSEQ_END(Chain,
5776 DAG.getIntPtrConstant(0, true),
5777 DAG.getIntPtrConstant(0, true),
5780 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5782 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5783 return DAG.getMergeValues(Ops1, 2, dl);
5787 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5789 SDValue Dst, SDValue Src,
5790 SDValue Size, unsigned Align,
5792 uint64_t DstSVOff) {
5793 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5795 // If not DWORD aligned or size is more than the threshold, call the library.
5796 // The libc version is likely to be faster for these cases. It can use the
5797 // address value and run time information about the CPU.
5798 if ((Align & 3) != 0 ||
5800 ConstantSize->getZExtValue() >
5801 getSubtarget()->getMaxInlineSizeThreshold()) {
5802 SDValue InFlag(0, 0);
5804 // Check to see if there is a specialized entry-point for memory zeroing.
5805 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5807 if (const char *bzeroEntry = V &&
5808 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5809 MVT IntPtr = getPointerTy();
5810 const Type *IntPtrTy = TD->getIntPtrType();
5811 TargetLowering::ArgListTy Args;
5812 TargetLowering::ArgListEntry Entry;
5814 Entry.Ty = IntPtrTy;
5815 Args.push_back(Entry);
5817 Args.push_back(Entry);
5818 std::pair<SDValue,SDValue> CallResult =
5819 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5820 0, CallingConv::C, false,
5821 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5822 return CallResult.second;
5825 // Otherwise have the target-independent code call memset.
5829 uint64_t SizeVal = ConstantSize->getZExtValue();
5830 SDValue InFlag(0, 0);
5833 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5834 unsigned BytesLeft = 0;
5835 bool TwoRepStos = false;
5838 uint64_t Val = ValC->getZExtValue() & 255;
5840 // If the value is a constant, then we can potentially use larger sets.
5841 switch (Align & 3) {
5842 case 2: // WORD aligned
5845 Val = (Val << 8) | Val;
5847 case 0: // DWORD aligned
5850 Val = (Val << 8) | Val;
5851 Val = (Val << 16) | Val;
5852 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5855 Val = (Val << 32) | Val;
5858 default: // Byte aligned
5861 Count = DAG.getIntPtrConstant(SizeVal);
5865 if (AVT.bitsGT(MVT::i8)) {
5866 unsigned UBytes = AVT.getSizeInBits() / 8;
5867 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5868 BytesLeft = SizeVal % UBytes;
5871 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5873 InFlag = Chain.getValue(1);
5876 Count = DAG.getIntPtrConstant(SizeVal);
5877 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5878 InFlag = Chain.getValue(1);
5881 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5884 InFlag = Chain.getValue(1);
5885 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5888 InFlag = Chain.getValue(1);
5890 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5891 SmallVector<SDValue, 8> Ops;
5892 Ops.push_back(Chain);
5893 Ops.push_back(DAG.getValueType(AVT));
5894 Ops.push_back(InFlag);
5895 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5898 InFlag = Chain.getValue(1);
5900 MVT CVT = Count.getValueType();
5901 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5902 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5903 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5906 InFlag = Chain.getValue(1);
5907 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5909 Ops.push_back(Chain);
5910 Ops.push_back(DAG.getValueType(MVT::i8));
5911 Ops.push_back(InFlag);
5912 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5913 } else if (BytesLeft) {
5914 // Handle the last 1 - 7 bytes.
5915 unsigned Offset = SizeVal - BytesLeft;
5916 MVT AddrVT = Dst.getValueType();
5917 MVT SizeVT = Size.getValueType();
5919 Chain = DAG.getMemset(Chain, dl,
5920 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5921 DAG.getConstant(Offset, AddrVT)),
5923 DAG.getConstant(BytesLeft, SizeVT),
5924 Align, DstSV, DstSVOff + Offset);
5927 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5932 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5933 SDValue Chain, SDValue Dst, SDValue Src,
5934 SDValue Size, unsigned Align,
5936 const Value *DstSV, uint64_t DstSVOff,
5937 const Value *SrcSV, uint64_t SrcSVOff) {
5938 // This requires the copy size to be a constant, preferrably
5939 // within a subtarget-specific limit.
5940 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5943 uint64_t SizeVal = ConstantSize->getZExtValue();
5944 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5947 /// If not DWORD aligned, call the library.
5948 if ((Align & 3) != 0)
5953 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5956 unsigned UBytes = AVT.getSizeInBits() / 8;
5957 unsigned CountVal = SizeVal / UBytes;
5958 SDValue Count = DAG.getIntPtrConstant(CountVal);
5959 unsigned BytesLeft = SizeVal % UBytes;
5961 SDValue InFlag(0, 0);
5962 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5965 InFlag = Chain.getValue(1);
5966 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5969 InFlag = Chain.getValue(1);
5970 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5973 InFlag = Chain.getValue(1);
5975 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5976 SmallVector<SDValue, 8> Ops;
5977 Ops.push_back(Chain);
5978 Ops.push_back(DAG.getValueType(AVT));
5979 Ops.push_back(InFlag);
5980 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5982 SmallVector<SDValue, 4> Results;
5983 Results.push_back(RepMovs);
5985 // Handle the last 1 - 7 bytes.
5986 unsigned Offset = SizeVal - BytesLeft;
5987 MVT DstVT = Dst.getValueType();
5988 MVT SrcVT = Src.getValueType();
5989 MVT SizeVT = Size.getValueType();
5990 Results.push_back(DAG.getMemcpy(Chain, dl,
5991 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5992 DAG.getConstant(Offset, DstVT)),
5993 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5994 DAG.getConstant(Offset, SrcVT)),
5995 DAG.getConstant(BytesLeft, SizeVT),
5996 Align, AlwaysInline,
5997 DstSV, DstSVOff + Offset,
5998 SrcSV, SrcSVOff + Offset));
6001 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6002 &Results[0], Results.size());
6005 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6006 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6007 DebugLoc dl = Op.getDebugLoc();
6009 if (!Subtarget->is64Bit()) {
6010 // vastart just stores the address of the VarArgsFrameIndex slot into the
6011 // memory location argument.
6012 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6013 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6017 // gp_offset (0 - 6 * 8)
6018 // fp_offset (48 - 48 + 8 * 16)
6019 // overflow_arg_area (point to parameters coming in memory).
6021 SmallVector<SDValue, 8> MemOps;
6022 SDValue FIN = Op.getOperand(1);
6024 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6025 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6027 MemOps.push_back(Store);
6030 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6031 FIN, DAG.getIntPtrConstant(4));
6032 Store = DAG.getStore(Op.getOperand(0), dl,
6033 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6035 MemOps.push_back(Store);
6037 // Store ptr to overflow_arg_area
6038 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6039 FIN, DAG.getIntPtrConstant(4));
6040 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6041 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6042 MemOps.push_back(Store);
6044 // Store ptr to reg_save_area.
6045 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6046 FIN, DAG.getIntPtrConstant(8));
6047 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6048 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6049 MemOps.push_back(Store);
6050 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6051 &MemOps[0], MemOps.size());
6054 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6055 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6056 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6057 SDValue Chain = Op.getOperand(0);
6058 SDValue SrcPtr = Op.getOperand(1);
6059 SDValue SrcSV = Op.getOperand(2);
6061 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6065 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6066 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6067 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6068 SDValue Chain = Op.getOperand(0);
6069 SDValue DstPtr = Op.getOperand(1);
6070 SDValue SrcPtr = Op.getOperand(2);
6071 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6072 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6073 DebugLoc dl = Op.getDebugLoc();
6075 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6076 DAG.getIntPtrConstant(24), 8, false,
6077 DstSV, 0, SrcSV, 0);
6081 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6082 DebugLoc dl = Op.getDebugLoc();
6083 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6085 default: return SDValue(); // Don't custom lower most intrinsics.
6086 // Comparison intrinsics.
6087 case Intrinsic::x86_sse_comieq_ss:
6088 case Intrinsic::x86_sse_comilt_ss:
6089 case Intrinsic::x86_sse_comile_ss:
6090 case Intrinsic::x86_sse_comigt_ss:
6091 case Intrinsic::x86_sse_comige_ss:
6092 case Intrinsic::x86_sse_comineq_ss:
6093 case Intrinsic::x86_sse_ucomieq_ss:
6094 case Intrinsic::x86_sse_ucomilt_ss:
6095 case Intrinsic::x86_sse_ucomile_ss:
6096 case Intrinsic::x86_sse_ucomigt_ss:
6097 case Intrinsic::x86_sse_ucomige_ss:
6098 case Intrinsic::x86_sse_ucomineq_ss:
6099 case Intrinsic::x86_sse2_comieq_sd:
6100 case Intrinsic::x86_sse2_comilt_sd:
6101 case Intrinsic::x86_sse2_comile_sd:
6102 case Intrinsic::x86_sse2_comigt_sd:
6103 case Intrinsic::x86_sse2_comige_sd:
6104 case Intrinsic::x86_sse2_comineq_sd:
6105 case Intrinsic::x86_sse2_ucomieq_sd:
6106 case Intrinsic::x86_sse2_ucomilt_sd:
6107 case Intrinsic::x86_sse2_ucomile_sd:
6108 case Intrinsic::x86_sse2_ucomigt_sd:
6109 case Intrinsic::x86_sse2_ucomige_sd:
6110 case Intrinsic::x86_sse2_ucomineq_sd: {
6112 ISD::CondCode CC = ISD::SETCC_INVALID;
6115 case Intrinsic::x86_sse_comieq_ss:
6116 case Intrinsic::x86_sse2_comieq_sd:
6120 case Intrinsic::x86_sse_comilt_ss:
6121 case Intrinsic::x86_sse2_comilt_sd:
6125 case Intrinsic::x86_sse_comile_ss:
6126 case Intrinsic::x86_sse2_comile_sd:
6130 case Intrinsic::x86_sse_comigt_ss:
6131 case Intrinsic::x86_sse2_comigt_sd:
6135 case Intrinsic::x86_sse_comige_ss:
6136 case Intrinsic::x86_sse2_comige_sd:
6140 case Intrinsic::x86_sse_comineq_ss:
6141 case Intrinsic::x86_sse2_comineq_sd:
6145 case Intrinsic::x86_sse_ucomieq_ss:
6146 case Intrinsic::x86_sse2_ucomieq_sd:
6147 Opc = X86ISD::UCOMI;
6150 case Intrinsic::x86_sse_ucomilt_ss:
6151 case Intrinsic::x86_sse2_ucomilt_sd:
6152 Opc = X86ISD::UCOMI;
6155 case Intrinsic::x86_sse_ucomile_ss:
6156 case Intrinsic::x86_sse2_ucomile_sd:
6157 Opc = X86ISD::UCOMI;
6160 case Intrinsic::x86_sse_ucomigt_ss:
6161 case Intrinsic::x86_sse2_ucomigt_sd:
6162 Opc = X86ISD::UCOMI;
6165 case Intrinsic::x86_sse_ucomige_ss:
6166 case Intrinsic::x86_sse2_ucomige_sd:
6167 Opc = X86ISD::UCOMI;
6170 case Intrinsic::x86_sse_ucomineq_ss:
6171 case Intrinsic::x86_sse2_ucomineq_sd:
6172 Opc = X86ISD::UCOMI;
6177 SDValue LHS = Op.getOperand(1);
6178 SDValue RHS = Op.getOperand(2);
6179 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6180 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6181 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6182 DAG.getConstant(X86CC, MVT::i8), Cond);
6183 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6186 // Fix vector shift instructions where the last operand is a non-immediate
6188 case Intrinsic::x86_sse2_pslli_w:
6189 case Intrinsic::x86_sse2_pslli_d:
6190 case Intrinsic::x86_sse2_pslli_q:
6191 case Intrinsic::x86_sse2_psrli_w:
6192 case Intrinsic::x86_sse2_psrli_d:
6193 case Intrinsic::x86_sse2_psrli_q:
6194 case Intrinsic::x86_sse2_psrai_w:
6195 case Intrinsic::x86_sse2_psrai_d:
6196 case Intrinsic::x86_mmx_pslli_w:
6197 case Intrinsic::x86_mmx_pslli_d:
6198 case Intrinsic::x86_mmx_pslli_q:
6199 case Intrinsic::x86_mmx_psrli_w:
6200 case Intrinsic::x86_mmx_psrli_d:
6201 case Intrinsic::x86_mmx_psrli_q:
6202 case Intrinsic::x86_mmx_psrai_w:
6203 case Intrinsic::x86_mmx_psrai_d: {
6204 SDValue ShAmt = Op.getOperand(2);
6205 if (isa<ConstantSDNode>(ShAmt))
6208 unsigned NewIntNo = 0;
6209 MVT ShAmtVT = MVT::v4i32;
6211 case Intrinsic::x86_sse2_pslli_w:
6212 NewIntNo = Intrinsic::x86_sse2_psll_w;
6214 case Intrinsic::x86_sse2_pslli_d:
6215 NewIntNo = Intrinsic::x86_sse2_psll_d;
6217 case Intrinsic::x86_sse2_pslli_q:
6218 NewIntNo = Intrinsic::x86_sse2_psll_q;
6220 case Intrinsic::x86_sse2_psrli_w:
6221 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6223 case Intrinsic::x86_sse2_psrli_d:
6224 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6226 case Intrinsic::x86_sse2_psrli_q:
6227 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6229 case Intrinsic::x86_sse2_psrai_w:
6230 NewIntNo = Intrinsic::x86_sse2_psra_w;
6232 case Intrinsic::x86_sse2_psrai_d:
6233 NewIntNo = Intrinsic::x86_sse2_psra_d;
6236 ShAmtVT = MVT::v2i32;
6238 case Intrinsic::x86_mmx_pslli_w:
6239 NewIntNo = Intrinsic::x86_mmx_psll_w;
6241 case Intrinsic::x86_mmx_pslli_d:
6242 NewIntNo = Intrinsic::x86_mmx_psll_d;
6244 case Intrinsic::x86_mmx_pslli_q:
6245 NewIntNo = Intrinsic::x86_mmx_psll_q;
6247 case Intrinsic::x86_mmx_psrli_w:
6248 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6250 case Intrinsic::x86_mmx_psrli_d:
6251 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6253 case Intrinsic::x86_mmx_psrli_q:
6254 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6256 case Intrinsic::x86_mmx_psrai_w:
6257 NewIntNo = Intrinsic::x86_mmx_psra_w;
6259 case Intrinsic::x86_mmx_psrai_d:
6260 NewIntNo = Intrinsic::x86_mmx_psra_d;
6262 default: LLVM_UNREACHABLE("Impossible intrinsic"); // Can't reach here.
6267 MVT VT = Op.getValueType();
6268 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6269 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6270 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6271 DAG.getConstant(NewIntNo, MVT::i32),
6272 Op.getOperand(1), ShAmt);
6277 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6278 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6279 DebugLoc dl = Op.getDebugLoc();
6282 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6284 DAG.getConstant(TD->getPointerSize(),
6285 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6286 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6287 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6292 // Just load the return address.
6293 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6294 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6295 RetAddrFI, NULL, 0);
6298 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6299 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6300 MFI->setFrameAddressIsTaken(true);
6301 MVT VT = Op.getValueType();
6302 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6303 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6304 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6305 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6307 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6311 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6312 SelectionDAG &DAG) {
6313 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6316 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6318 MachineFunction &MF = DAG.getMachineFunction();
6319 SDValue Chain = Op.getOperand(0);
6320 SDValue Offset = Op.getOperand(1);
6321 SDValue Handler = Op.getOperand(2);
6322 DebugLoc dl = Op.getDebugLoc();
6324 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6326 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6328 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6329 DAG.getIntPtrConstant(-TD->getPointerSize()));
6330 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6331 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6332 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6333 MF.getRegInfo().addLiveOut(StoreAddrReg);
6335 return DAG.getNode(X86ISD::EH_RETURN, dl,
6337 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6340 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6341 SelectionDAG &DAG) {
6342 SDValue Root = Op.getOperand(0);
6343 SDValue Trmp = Op.getOperand(1); // trampoline
6344 SDValue FPtr = Op.getOperand(2); // nested function
6345 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6346 DebugLoc dl = Op.getDebugLoc();
6348 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6350 const X86InstrInfo *TII =
6351 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6353 if (Subtarget->is64Bit()) {
6354 SDValue OutChains[6];
6356 // Large code-model.
6358 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6359 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6361 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6362 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6364 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6366 // Load the pointer to the nested function into R11.
6367 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6368 SDValue Addr = Trmp;
6369 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6372 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6373 DAG.getConstant(2, MVT::i64));
6374 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6376 // Load the 'nest' parameter value into R10.
6377 // R10 is specified in X86CallingConv.td
6378 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6379 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6380 DAG.getConstant(10, MVT::i64));
6381 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6382 Addr, TrmpAddr, 10);
6384 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6385 DAG.getConstant(12, MVT::i64));
6386 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6388 // Jump to the nested function.
6389 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6390 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6391 DAG.getConstant(20, MVT::i64));
6392 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6393 Addr, TrmpAddr, 20);
6395 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6396 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6397 DAG.getConstant(22, MVT::i64));
6398 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6402 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6403 return DAG.getMergeValues(Ops, 2, dl);
6405 const Function *Func =
6406 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6407 unsigned CC = Func->getCallingConv();
6412 assert(0 && "Unsupported calling convention");
6413 case CallingConv::C:
6414 case CallingConv::X86_StdCall: {
6415 // Pass 'nest' parameter in ECX.
6416 // Must be kept in sync with X86CallingConv.td
6419 // Check that ECX wasn't needed by an 'inreg' parameter.
6420 const FunctionType *FTy = Func->getFunctionType();
6421 const AttrListPtr &Attrs = Func->getAttributes();
6423 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6424 unsigned InRegCount = 0;
6427 for (FunctionType::param_iterator I = FTy->param_begin(),
6428 E = FTy->param_end(); I != E; ++I, ++Idx)
6429 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6430 // FIXME: should only count parameters that are lowered to integers.
6431 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6433 if (InRegCount > 2) {
6434 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6439 case CallingConv::X86_FastCall:
6440 case CallingConv::Fast:
6441 // Pass 'nest' parameter in EAX.
6442 // Must be kept in sync with X86CallingConv.td
6447 SDValue OutChains[4];
6450 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6451 DAG.getConstant(10, MVT::i32));
6452 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6454 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6455 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6456 OutChains[0] = DAG.getStore(Root, dl,
6457 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6460 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6461 DAG.getConstant(1, MVT::i32));
6462 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6464 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6465 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6466 DAG.getConstant(5, MVT::i32));
6467 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6468 TrmpAddr, 5, false, 1);
6470 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6471 DAG.getConstant(6, MVT::i32));
6472 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6475 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6476 return DAG.getMergeValues(Ops, 2, dl);
6480 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6482 The rounding mode is in bits 11:10 of FPSR, and has the following
6489 FLT_ROUNDS, on the other hand, expects the following:
6496 To perform the conversion, we do:
6497 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6500 MachineFunction &MF = DAG.getMachineFunction();
6501 const TargetMachine &TM = MF.getTarget();
6502 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6503 unsigned StackAlignment = TFI.getStackAlignment();
6504 MVT VT = Op.getValueType();
6505 DebugLoc dl = Op.getDebugLoc();
6507 // Save FP Control Word to stack slot
6508 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6509 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6511 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6512 DAG.getEntryNode(), StackSlot);
6514 // Load FP Control Word from stack slot
6515 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6517 // Transform as necessary
6519 DAG.getNode(ISD::SRL, dl, MVT::i16,
6520 DAG.getNode(ISD::AND, dl, MVT::i16,
6521 CWD, DAG.getConstant(0x800, MVT::i16)),
6522 DAG.getConstant(11, MVT::i8));
6524 DAG.getNode(ISD::SRL, dl, MVT::i16,
6525 DAG.getNode(ISD::AND, dl, MVT::i16,
6526 CWD, DAG.getConstant(0x400, MVT::i16)),
6527 DAG.getConstant(9, MVT::i8));
6530 DAG.getNode(ISD::AND, dl, MVT::i16,
6531 DAG.getNode(ISD::ADD, dl, MVT::i16,
6532 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6533 DAG.getConstant(1, MVT::i16)),
6534 DAG.getConstant(3, MVT::i16));
6537 return DAG.getNode((VT.getSizeInBits() < 16 ?
6538 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6541 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6542 MVT VT = Op.getValueType();
6544 unsigned NumBits = VT.getSizeInBits();
6545 DebugLoc dl = Op.getDebugLoc();
6547 Op = Op.getOperand(0);
6548 if (VT == MVT::i8) {
6549 // Zero extend to i32 since there is not an i8 bsr.
6551 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6554 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6555 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6556 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6558 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6559 SmallVector<SDValue, 4> Ops;
6561 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6562 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6563 Ops.push_back(Op.getValue(1));
6564 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6566 // Finally xor with NumBits-1.
6567 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6570 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6574 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6575 MVT VT = Op.getValueType();
6577 unsigned NumBits = VT.getSizeInBits();
6578 DebugLoc dl = Op.getDebugLoc();
6580 Op = Op.getOperand(0);
6581 if (VT == MVT::i8) {
6583 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6586 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6587 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6588 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6590 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6591 SmallVector<SDValue, 4> Ops;
6593 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6594 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6595 Ops.push_back(Op.getValue(1));
6596 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6599 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6603 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6604 MVT VT = Op.getValueType();
6605 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6606 DebugLoc dl = Op.getDebugLoc();
6608 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6609 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6610 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6611 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6612 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6614 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6615 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6616 // return AloBlo + AloBhi + AhiBlo;
6618 SDValue A = Op.getOperand(0);
6619 SDValue B = Op.getOperand(1);
6621 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6622 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6623 A, DAG.getConstant(32, MVT::i32));
6624 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6625 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6626 B, DAG.getConstant(32, MVT::i32));
6627 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6628 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6630 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6631 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6633 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6634 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6636 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6637 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6638 AloBhi, DAG.getConstant(32, MVT::i32));
6639 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6640 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6641 AhiBlo, DAG.getConstant(32, MVT::i32));
6642 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6643 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6648 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6649 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6650 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6651 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6652 // has only one use.
6653 SDNode *N = Op.getNode();
6654 SDValue LHS = N->getOperand(0);
6655 SDValue RHS = N->getOperand(1);
6656 unsigned BaseOp = 0;
6658 DebugLoc dl = Op.getDebugLoc();
6660 switch (Op.getOpcode()) {
6661 default: assert(0 && "Unknown ovf instruction!");
6663 // A subtract of one will be selected as a INC. Note that INC doesn't
6664 // set CF, so we can't do this for UADDO.
6665 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6666 if (C->getAPIntValue() == 1) {
6667 BaseOp = X86ISD::INC;
6671 BaseOp = X86ISD::ADD;
6675 BaseOp = X86ISD::ADD;
6679 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6680 // set CF, so we can't do this for USUBO.
6681 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6682 if (C->getAPIntValue() == 1) {
6683 BaseOp = X86ISD::DEC;
6687 BaseOp = X86ISD::SUB;
6691 BaseOp = X86ISD::SUB;
6695 BaseOp = X86ISD::SMUL;
6699 BaseOp = X86ISD::UMUL;
6704 // Also sets EFLAGS.
6705 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6706 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6709 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6710 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6712 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6716 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6717 MVT T = Op.getValueType();
6718 DebugLoc dl = Op.getDebugLoc();
6721 switch(T.getSimpleVT()) {
6723 assert(false && "Invalid value type!");
6724 case MVT::i8: Reg = X86::AL; size = 1; break;
6725 case MVT::i16: Reg = X86::AX; size = 2; break;
6726 case MVT::i32: Reg = X86::EAX; size = 4; break;
6728 assert(Subtarget->is64Bit() && "Node not type legal!");
6729 Reg = X86::RAX; size = 8;
6732 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6733 Op.getOperand(2), SDValue());
6734 SDValue Ops[] = { cpIn.getValue(0),
6737 DAG.getTargetConstant(size, MVT::i8),
6739 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6740 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6742 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6746 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6747 SelectionDAG &DAG) {
6748 assert(Subtarget->is64Bit() && "Result not type legalized?");
6749 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6750 SDValue TheChain = Op.getOperand(0);
6751 DebugLoc dl = Op.getDebugLoc();
6752 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6753 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6754 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6756 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6757 DAG.getConstant(32, MVT::i8));
6759 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6762 return DAG.getMergeValues(Ops, 2, dl);
6765 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6766 SDNode *Node = Op.getNode();
6767 DebugLoc dl = Node->getDebugLoc();
6768 MVT T = Node->getValueType(0);
6769 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6770 DAG.getConstant(0, T), Node->getOperand(2));
6771 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6772 cast<AtomicSDNode>(Node)->getMemoryVT(),
6773 Node->getOperand(0),
6774 Node->getOperand(1), negOp,
6775 cast<AtomicSDNode>(Node)->getSrcValue(),
6776 cast<AtomicSDNode>(Node)->getAlignment());
6779 /// LowerOperation - Provide custom lowering hooks for some operations.
6781 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6782 switch (Op.getOpcode()) {
6783 default: assert(0 && "Should not custom lower this!");
6784 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6785 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6786 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6787 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6788 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6789 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6790 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6791 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6792 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6793 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6794 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6795 case ISD::SHL_PARTS:
6796 case ISD::SRA_PARTS:
6797 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6798 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6799 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6800 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6801 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6802 case ISD::FABS: return LowerFABS(Op, DAG);
6803 case ISD::FNEG: return LowerFNEG(Op, DAG);
6804 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6805 case ISD::SETCC: return LowerSETCC(Op, DAG);
6806 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6807 case ISD::SELECT: return LowerSELECT(Op, DAG);
6808 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6809 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6810 case ISD::CALL: return LowerCALL(Op, DAG);
6811 case ISD::RET: return LowerRET(Op, DAG);
6812 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6813 case ISD::VASTART: return LowerVASTART(Op, DAG);
6814 case ISD::VAARG: return LowerVAARG(Op, DAG);
6815 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6816 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6817 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6818 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6819 case ISD::FRAME_TO_ARGS_OFFSET:
6820 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6821 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6822 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6823 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6824 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6825 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6826 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6827 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6833 case ISD::UMULO: return LowerXALUO(Op, DAG);
6834 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6838 void X86TargetLowering::
6839 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6840 SelectionDAG &DAG, unsigned NewOp) {
6841 MVT T = Node->getValueType(0);
6842 DebugLoc dl = Node->getDebugLoc();
6843 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6845 SDValue Chain = Node->getOperand(0);
6846 SDValue In1 = Node->getOperand(1);
6847 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6848 Node->getOperand(2), DAG.getIntPtrConstant(0));
6849 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6850 Node->getOperand(2), DAG.getIntPtrConstant(1));
6851 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6852 // have a MemOperand. Pass the info through as a normal operand.
6853 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6854 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6855 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6856 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6857 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6858 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6859 Results.push_back(Result.getValue(2));
6862 /// ReplaceNodeResults - Replace a node with an illegal result type
6863 /// with a new node built out of custom code.
6864 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6865 SmallVectorImpl<SDValue>&Results,
6866 SelectionDAG &DAG) {
6867 DebugLoc dl = N->getDebugLoc();
6868 switch (N->getOpcode()) {
6870 assert(false && "Do not know how to custom type legalize this operation!");
6872 case ISD::FP_TO_SINT: {
6873 std::pair<SDValue,SDValue> Vals =
6874 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6875 SDValue FIST = Vals.first, StackSlot = Vals.second;
6876 if (FIST.getNode() != 0) {
6877 MVT VT = N->getValueType(0);
6878 // Return a load from the stack slot.
6879 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6883 case ISD::READCYCLECOUNTER: {
6884 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6885 SDValue TheChain = N->getOperand(0);
6886 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6887 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6889 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6891 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6892 SDValue Ops[] = { eax, edx };
6893 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6894 Results.push_back(edx.getValue(1));
6897 case ISD::ATOMIC_CMP_SWAP: {
6898 MVT T = N->getValueType(0);
6899 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6900 SDValue cpInL, cpInH;
6901 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6902 DAG.getConstant(0, MVT::i32));
6903 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6904 DAG.getConstant(1, MVT::i32));
6905 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6906 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6908 SDValue swapInL, swapInH;
6909 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6910 DAG.getConstant(0, MVT::i32));
6911 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6912 DAG.getConstant(1, MVT::i32));
6913 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6915 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6916 swapInL.getValue(1));
6917 SDValue Ops[] = { swapInH.getValue(0),
6919 swapInH.getValue(1) };
6920 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6921 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6922 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6923 MVT::i32, Result.getValue(1));
6924 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6925 MVT::i32, cpOutL.getValue(2));
6926 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6927 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6928 Results.push_back(cpOutH.getValue(1));
6931 case ISD::ATOMIC_LOAD_ADD:
6932 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6934 case ISD::ATOMIC_LOAD_AND:
6935 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6937 case ISD::ATOMIC_LOAD_NAND:
6938 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6940 case ISD::ATOMIC_LOAD_OR:
6941 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6943 case ISD::ATOMIC_LOAD_SUB:
6944 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6946 case ISD::ATOMIC_LOAD_XOR:
6947 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6949 case ISD::ATOMIC_SWAP:
6950 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6955 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6957 default: return NULL;
6958 case X86ISD::BSF: return "X86ISD::BSF";
6959 case X86ISD::BSR: return "X86ISD::BSR";
6960 case X86ISD::SHLD: return "X86ISD::SHLD";
6961 case X86ISD::SHRD: return "X86ISD::SHRD";
6962 case X86ISD::FAND: return "X86ISD::FAND";
6963 case X86ISD::FOR: return "X86ISD::FOR";
6964 case X86ISD::FXOR: return "X86ISD::FXOR";
6965 case X86ISD::FSRL: return "X86ISD::FSRL";
6966 case X86ISD::FILD: return "X86ISD::FILD";
6967 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6968 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6969 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6970 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6971 case X86ISD::FLD: return "X86ISD::FLD";
6972 case X86ISD::FST: return "X86ISD::FST";
6973 case X86ISD::CALL: return "X86ISD::CALL";
6974 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6975 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6976 case X86ISD::BT: return "X86ISD::BT";
6977 case X86ISD::CMP: return "X86ISD::CMP";
6978 case X86ISD::COMI: return "X86ISD::COMI";
6979 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6980 case X86ISD::SETCC: return "X86ISD::SETCC";
6981 case X86ISD::CMOV: return "X86ISD::CMOV";
6982 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6983 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6984 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6985 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6986 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6987 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6988 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
6989 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6990 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6991 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6992 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6993 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6994 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6995 case X86ISD::FMAX: return "X86ISD::FMAX";
6996 case X86ISD::FMIN: return "X86ISD::FMIN";
6997 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6998 case X86ISD::FRCP: return "X86ISD::FRCP";
6999 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7000 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7001 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7002 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7003 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7004 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7005 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7006 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7007 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7008 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7009 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7010 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7011 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7012 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7013 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7014 case X86ISD::VSHL: return "X86ISD::VSHL";
7015 case X86ISD::VSRL: return "X86ISD::VSRL";
7016 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7017 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7018 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7019 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7020 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7021 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7022 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7023 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7024 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7025 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7026 case X86ISD::ADD: return "X86ISD::ADD";
7027 case X86ISD::SUB: return "X86ISD::SUB";
7028 case X86ISD::SMUL: return "X86ISD::SMUL";
7029 case X86ISD::UMUL: return "X86ISD::UMUL";
7030 case X86ISD::INC: return "X86ISD::INC";
7031 case X86ISD::DEC: return "X86ISD::DEC";
7032 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7036 // isLegalAddressingMode - Return true if the addressing mode represented
7037 // by AM is legal for this target, for a load/store of the specified type.
7038 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7039 const Type *Ty) const {
7040 // X86 supports extremely general addressing modes.
7042 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7043 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7047 // We can only fold this if we don't need an extra load.
7048 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine()))
7050 // If BaseGV requires a register, we cannot also have a BaseReg.
7051 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine()) &&
7055 // X86-64 only supports addr of globals in small code model.
7056 if (Subtarget->is64Bit()) {
7057 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7059 // If lower 4G is not available, then we must use rip-relative addressing.
7060 if (AM.BaseOffs || AM.Scale > 1)
7071 // These scales always work.
7076 // These scales are formed with basereg+scalereg. Only accept if there is
7081 default: // Other stuff never works.
7089 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7090 if (!Ty1->isInteger() || !Ty2->isInteger())
7092 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7093 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7094 if (NumBits1 <= NumBits2)
7096 return Subtarget->is64Bit() || NumBits1 < 64;
7099 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7100 if (!VT1.isInteger() || !VT2.isInteger())
7102 unsigned NumBits1 = VT1.getSizeInBits();
7103 unsigned NumBits2 = VT2.getSizeInBits();
7104 if (NumBits1 <= NumBits2)
7106 return Subtarget->is64Bit() || NumBits1 < 64;
7109 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7110 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7111 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7114 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7115 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7116 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7119 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
7120 // i16 instructions are longer (0x66 prefix) and potentially slower.
7121 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7124 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7125 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7126 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7127 /// are assumed to be legal.
7129 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7131 // Only do shuffles on 128-bit vector types for now.
7132 if (VT.getSizeInBits() == 64)
7135 // FIXME: pshufb, blends, palignr, shifts.
7136 return (VT.getVectorNumElements() == 2 ||
7137 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7138 isMOVLMask(M, VT) ||
7139 isSHUFPMask(M, VT) ||
7140 isPSHUFDMask(M, VT) ||
7141 isPSHUFHWMask(M, VT) ||
7142 isPSHUFLWMask(M, VT) ||
7143 isUNPCKLMask(M, VT) ||
7144 isUNPCKHMask(M, VT) ||
7145 isUNPCKL_v_undef_Mask(M, VT) ||
7146 isUNPCKH_v_undef_Mask(M, VT));
7150 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7152 unsigned NumElts = VT.getVectorNumElements();
7153 // FIXME: This collection of masks seems suspect.
7156 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7157 return (isMOVLMask(Mask, VT) ||
7158 isCommutedMOVLMask(Mask, VT, true) ||
7159 isSHUFPMask(Mask, VT) ||
7160 isCommutedSHUFPMask(Mask, VT));
7165 //===----------------------------------------------------------------------===//
7166 // X86 Scheduler Hooks
7167 //===----------------------------------------------------------------------===//
7169 // private utility function
7171 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7172 MachineBasicBlock *MBB,
7180 TargetRegisterClass *RC,
7181 bool invSrc) const {
7182 // For the atomic bitwise operator, we generate
7185 // ld t1 = [bitinstr.addr]
7186 // op t2 = t1, [bitinstr.val]
7188 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7190 // fallthrough -->nextMBB
7191 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7192 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7193 MachineFunction::iterator MBBIter = MBB;
7196 /// First build the CFG
7197 MachineFunction *F = MBB->getParent();
7198 MachineBasicBlock *thisMBB = MBB;
7199 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7200 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7201 F->insert(MBBIter, newMBB);
7202 F->insert(MBBIter, nextMBB);
7204 // Move all successors to thisMBB to nextMBB
7205 nextMBB->transferSuccessors(thisMBB);
7207 // Update thisMBB to fall through to newMBB
7208 thisMBB->addSuccessor(newMBB);
7210 // newMBB jumps to itself and fall through to nextMBB
7211 newMBB->addSuccessor(nextMBB);
7212 newMBB->addSuccessor(newMBB);
7214 // Insert instructions into newMBB based on incoming instruction
7215 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7216 "unexpected number of operands");
7217 DebugLoc dl = bInstr->getDebugLoc();
7218 MachineOperand& destOper = bInstr->getOperand(0);
7219 MachineOperand* argOpers[2 + X86AddrNumOperands];
7220 int numArgs = bInstr->getNumOperands() - 1;
7221 for (int i=0; i < numArgs; ++i)
7222 argOpers[i] = &bInstr->getOperand(i+1);
7224 // x86 address has 4 operands: base, index, scale, and displacement
7225 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7226 int valArgIndx = lastAddrIndx + 1;
7228 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7229 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7230 for (int i=0; i <= lastAddrIndx; ++i)
7231 (*MIB).addOperand(*argOpers[i]);
7233 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7235 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7240 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7241 assert((argOpers[valArgIndx]->isReg() ||
7242 argOpers[valArgIndx]->isImm()) &&
7244 if (argOpers[valArgIndx]->isReg())
7245 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7247 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7249 (*MIB).addOperand(*argOpers[valArgIndx]);
7251 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7254 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7255 for (int i=0; i <= lastAddrIndx; ++i)
7256 (*MIB).addOperand(*argOpers[i]);
7258 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7259 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7261 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7265 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7267 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7271 // private utility function: 64 bit atomics on 32 bit host.
7273 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7274 MachineBasicBlock *MBB,
7279 bool invSrc) const {
7280 // For the atomic bitwise operator, we generate
7281 // thisMBB (instructions are in pairs, except cmpxchg8b)
7282 // ld t1,t2 = [bitinstr.addr]
7284 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7285 // op t5, t6 <- out1, out2, [bitinstr.val]
7286 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7287 // mov ECX, EBX <- t5, t6
7288 // mov EAX, EDX <- t1, t2
7289 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7290 // mov t3, t4 <- EAX, EDX
7292 // result in out1, out2
7293 // fallthrough -->nextMBB
7295 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7296 const unsigned LoadOpc = X86::MOV32rm;
7297 const unsigned copyOpc = X86::MOV32rr;
7298 const unsigned NotOpc = X86::NOT32r;
7299 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7300 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7301 MachineFunction::iterator MBBIter = MBB;
7304 /// First build the CFG
7305 MachineFunction *F = MBB->getParent();
7306 MachineBasicBlock *thisMBB = MBB;
7307 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7308 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7309 F->insert(MBBIter, newMBB);
7310 F->insert(MBBIter, nextMBB);
7312 // Move all successors to thisMBB to nextMBB
7313 nextMBB->transferSuccessors(thisMBB);
7315 // Update thisMBB to fall through to newMBB
7316 thisMBB->addSuccessor(newMBB);
7318 // newMBB jumps to itself and fall through to nextMBB
7319 newMBB->addSuccessor(nextMBB);
7320 newMBB->addSuccessor(newMBB);
7322 DebugLoc dl = bInstr->getDebugLoc();
7323 // Insert instructions into newMBB based on incoming instruction
7324 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7325 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7326 "unexpected number of operands");
7327 MachineOperand& dest1Oper = bInstr->getOperand(0);
7328 MachineOperand& dest2Oper = bInstr->getOperand(1);
7329 MachineOperand* argOpers[2 + X86AddrNumOperands];
7330 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7331 argOpers[i] = &bInstr->getOperand(i+2);
7333 // x86 address has 4 operands: base, index, scale, and displacement
7334 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7336 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7337 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7338 for (int i=0; i <= lastAddrIndx; ++i)
7339 (*MIB).addOperand(*argOpers[i]);
7340 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7341 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7342 // add 4 to displacement.
7343 for (int i=0; i <= lastAddrIndx-2; ++i)
7344 (*MIB).addOperand(*argOpers[i]);
7345 MachineOperand newOp3 = *(argOpers[3]);
7347 newOp3.setImm(newOp3.getImm()+4);
7349 newOp3.setOffset(newOp3.getOffset()+4);
7350 (*MIB).addOperand(newOp3);
7351 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7353 // t3/4 are defined later, at the bottom of the loop
7354 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7355 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7356 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7357 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7358 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7359 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7361 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7362 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7364 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7365 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7371 int valArgIndx = lastAddrIndx + 1;
7372 assert((argOpers[valArgIndx]->isReg() ||
7373 argOpers[valArgIndx]->isImm()) &&
7375 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7376 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7377 if (argOpers[valArgIndx]->isReg())
7378 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7380 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7381 if (regOpcL != X86::MOV32rr)
7383 (*MIB).addOperand(*argOpers[valArgIndx]);
7384 assert(argOpers[valArgIndx + 1]->isReg() ==
7385 argOpers[valArgIndx]->isReg());
7386 assert(argOpers[valArgIndx + 1]->isImm() ==
7387 argOpers[valArgIndx]->isImm());
7388 if (argOpers[valArgIndx + 1]->isReg())
7389 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7391 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7392 if (regOpcH != X86::MOV32rr)
7394 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7396 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7398 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7401 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7403 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7406 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7407 for (int i=0; i <= lastAddrIndx; ++i)
7408 (*MIB).addOperand(*argOpers[i]);
7410 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7411 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7413 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7414 MIB.addReg(X86::EAX);
7415 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7416 MIB.addReg(X86::EDX);
7419 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7421 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7425 // private utility function
7427 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7428 MachineBasicBlock *MBB,
7429 unsigned cmovOpc) const {
7430 // For the atomic min/max operator, we generate
7433 // ld t1 = [min/max.addr]
7434 // mov t2 = [min/max.val]
7436 // cmov[cond] t2 = t1
7438 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7440 // fallthrough -->nextMBB
7442 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7443 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7444 MachineFunction::iterator MBBIter = MBB;
7447 /// First build the CFG
7448 MachineFunction *F = MBB->getParent();
7449 MachineBasicBlock *thisMBB = MBB;
7450 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7451 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7452 F->insert(MBBIter, newMBB);
7453 F->insert(MBBIter, nextMBB);
7455 // Move all successors to thisMBB to nextMBB
7456 nextMBB->transferSuccessors(thisMBB);
7458 // Update thisMBB to fall through to newMBB
7459 thisMBB->addSuccessor(newMBB);
7461 // newMBB jumps to newMBB and fall through to nextMBB
7462 newMBB->addSuccessor(nextMBB);
7463 newMBB->addSuccessor(newMBB);
7465 DebugLoc dl = mInstr->getDebugLoc();
7466 // Insert instructions into newMBB based on incoming instruction
7467 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7468 "unexpected number of operands");
7469 MachineOperand& destOper = mInstr->getOperand(0);
7470 MachineOperand* argOpers[2 + X86AddrNumOperands];
7471 int numArgs = mInstr->getNumOperands() - 1;
7472 for (int i=0; i < numArgs; ++i)
7473 argOpers[i] = &mInstr->getOperand(i+1);
7475 // x86 address has 4 operands: base, index, scale, and displacement
7476 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7477 int valArgIndx = lastAddrIndx + 1;
7479 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7480 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7481 for (int i=0; i <= lastAddrIndx; ++i)
7482 (*MIB).addOperand(*argOpers[i]);
7484 // We only support register and immediate values
7485 assert((argOpers[valArgIndx]->isReg() ||
7486 argOpers[valArgIndx]->isImm()) &&
7489 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7490 if (argOpers[valArgIndx]->isReg())
7491 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7493 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7494 (*MIB).addOperand(*argOpers[valArgIndx]);
7496 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7499 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7504 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7505 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7509 // Cmp and exchange if none has modified the memory location
7510 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7511 for (int i=0; i <= lastAddrIndx; ++i)
7512 (*MIB).addOperand(*argOpers[i]);
7514 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7515 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7517 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7518 MIB.addReg(X86::EAX);
7521 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7523 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7529 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7530 MachineBasicBlock *BB) const {
7531 DebugLoc dl = MI->getDebugLoc();
7532 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7533 switch (MI->getOpcode()) {
7534 default: assert(false && "Unexpected instr type to insert");
7535 case X86::CMOV_V1I64:
7536 case X86::CMOV_FR32:
7537 case X86::CMOV_FR64:
7538 case X86::CMOV_V4F32:
7539 case X86::CMOV_V2F64:
7540 case X86::CMOV_V2I64: {
7541 // To "insert" a SELECT_CC instruction, we actually have to insert the
7542 // diamond control-flow pattern. The incoming instruction knows the
7543 // destination vreg to set, the condition code register to branch on, the
7544 // true/false values to select between, and a branch opcode to use.
7545 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7546 MachineFunction::iterator It = BB;
7552 // cmpTY ccX, r1, r2
7554 // fallthrough --> copy0MBB
7555 MachineBasicBlock *thisMBB = BB;
7556 MachineFunction *F = BB->getParent();
7557 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7558 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7560 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7561 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7562 F->insert(It, copy0MBB);
7563 F->insert(It, sinkMBB);
7564 // Update machine-CFG edges by transferring all successors of the current
7565 // block to the new block which will contain the Phi node for the select.
7566 sinkMBB->transferSuccessors(BB);
7568 // Add the true and fallthrough blocks as its successors.
7569 BB->addSuccessor(copy0MBB);
7570 BB->addSuccessor(sinkMBB);
7573 // %FalseValue = ...
7574 // # fallthrough to sinkMBB
7577 // Update machine-CFG edges
7578 BB->addSuccessor(sinkMBB);
7581 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7584 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7585 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7586 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7588 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7592 case X86::FP32_TO_INT16_IN_MEM:
7593 case X86::FP32_TO_INT32_IN_MEM:
7594 case X86::FP32_TO_INT64_IN_MEM:
7595 case X86::FP64_TO_INT16_IN_MEM:
7596 case X86::FP64_TO_INT32_IN_MEM:
7597 case X86::FP64_TO_INT64_IN_MEM:
7598 case X86::FP80_TO_INT16_IN_MEM:
7599 case X86::FP80_TO_INT32_IN_MEM:
7600 case X86::FP80_TO_INT64_IN_MEM: {
7601 // Change the floating point control register to use "round towards zero"
7602 // mode when truncating to an integer value.
7603 MachineFunction *F = BB->getParent();
7604 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7605 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7607 // Load the old value of the high byte of the control word...
7609 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7610 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7613 // Set the high part to be round to zero...
7614 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7617 // Reload the modified control word now...
7618 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7620 // Restore the memory image of control word to original value
7621 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7624 // Get the X86 opcode to use.
7626 switch (MI->getOpcode()) {
7627 default: assert(0 && "illegal opcode!");
7628 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7629 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7630 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7631 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7632 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7633 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7634 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7635 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7636 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7640 MachineOperand &Op = MI->getOperand(0);
7642 AM.BaseType = X86AddressMode::RegBase;
7643 AM.Base.Reg = Op.getReg();
7645 AM.BaseType = X86AddressMode::FrameIndexBase;
7646 AM.Base.FrameIndex = Op.getIndex();
7648 Op = MI->getOperand(1);
7650 AM.Scale = Op.getImm();
7651 Op = MI->getOperand(2);
7653 AM.IndexReg = Op.getImm();
7654 Op = MI->getOperand(3);
7655 if (Op.isGlobal()) {
7656 AM.GV = Op.getGlobal();
7658 AM.Disp = Op.getImm();
7660 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7661 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7663 // Reload the original control word now.
7664 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7666 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7669 case X86::ATOMAND32:
7670 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7671 X86::AND32ri, X86::MOV32rm,
7672 X86::LCMPXCHG32, X86::MOV32rr,
7673 X86::NOT32r, X86::EAX,
7674 X86::GR32RegisterClass);
7676 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7677 X86::OR32ri, X86::MOV32rm,
7678 X86::LCMPXCHG32, X86::MOV32rr,
7679 X86::NOT32r, X86::EAX,
7680 X86::GR32RegisterClass);
7681 case X86::ATOMXOR32:
7682 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7683 X86::XOR32ri, X86::MOV32rm,
7684 X86::LCMPXCHG32, X86::MOV32rr,
7685 X86::NOT32r, X86::EAX,
7686 X86::GR32RegisterClass);
7687 case X86::ATOMNAND32:
7688 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7689 X86::AND32ri, X86::MOV32rm,
7690 X86::LCMPXCHG32, X86::MOV32rr,
7691 X86::NOT32r, X86::EAX,
7692 X86::GR32RegisterClass, true);
7693 case X86::ATOMMIN32:
7694 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7695 case X86::ATOMMAX32:
7696 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7697 case X86::ATOMUMIN32:
7698 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7699 case X86::ATOMUMAX32:
7700 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7702 case X86::ATOMAND16:
7703 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7704 X86::AND16ri, X86::MOV16rm,
7705 X86::LCMPXCHG16, X86::MOV16rr,
7706 X86::NOT16r, X86::AX,
7707 X86::GR16RegisterClass);
7709 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7710 X86::OR16ri, X86::MOV16rm,
7711 X86::LCMPXCHG16, X86::MOV16rr,
7712 X86::NOT16r, X86::AX,
7713 X86::GR16RegisterClass);
7714 case X86::ATOMXOR16:
7715 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7716 X86::XOR16ri, X86::MOV16rm,
7717 X86::LCMPXCHG16, X86::MOV16rr,
7718 X86::NOT16r, X86::AX,
7719 X86::GR16RegisterClass);
7720 case X86::ATOMNAND16:
7721 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7722 X86::AND16ri, X86::MOV16rm,
7723 X86::LCMPXCHG16, X86::MOV16rr,
7724 X86::NOT16r, X86::AX,
7725 X86::GR16RegisterClass, true);
7726 case X86::ATOMMIN16:
7727 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7728 case X86::ATOMMAX16:
7729 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7730 case X86::ATOMUMIN16:
7731 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7732 case X86::ATOMUMAX16:
7733 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7736 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7737 X86::AND8ri, X86::MOV8rm,
7738 X86::LCMPXCHG8, X86::MOV8rr,
7739 X86::NOT8r, X86::AL,
7740 X86::GR8RegisterClass);
7742 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7743 X86::OR8ri, X86::MOV8rm,
7744 X86::LCMPXCHG8, X86::MOV8rr,
7745 X86::NOT8r, X86::AL,
7746 X86::GR8RegisterClass);
7748 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7749 X86::XOR8ri, X86::MOV8rm,
7750 X86::LCMPXCHG8, X86::MOV8rr,
7751 X86::NOT8r, X86::AL,
7752 X86::GR8RegisterClass);
7753 case X86::ATOMNAND8:
7754 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7755 X86::AND8ri, X86::MOV8rm,
7756 X86::LCMPXCHG8, X86::MOV8rr,
7757 X86::NOT8r, X86::AL,
7758 X86::GR8RegisterClass, true);
7759 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7760 // This group is for 64-bit host.
7761 case X86::ATOMAND64:
7762 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7763 X86::AND64ri32, X86::MOV64rm,
7764 X86::LCMPXCHG64, X86::MOV64rr,
7765 X86::NOT64r, X86::RAX,
7766 X86::GR64RegisterClass);
7768 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7769 X86::OR64ri32, X86::MOV64rm,
7770 X86::LCMPXCHG64, X86::MOV64rr,
7771 X86::NOT64r, X86::RAX,
7772 X86::GR64RegisterClass);
7773 case X86::ATOMXOR64:
7774 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7775 X86::XOR64ri32, X86::MOV64rm,
7776 X86::LCMPXCHG64, X86::MOV64rr,
7777 X86::NOT64r, X86::RAX,
7778 X86::GR64RegisterClass);
7779 case X86::ATOMNAND64:
7780 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7781 X86::AND64ri32, X86::MOV64rm,
7782 X86::LCMPXCHG64, X86::MOV64rr,
7783 X86::NOT64r, X86::RAX,
7784 X86::GR64RegisterClass, true);
7785 case X86::ATOMMIN64:
7786 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7787 case X86::ATOMMAX64:
7788 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7789 case X86::ATOMUMIN64:
7790 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7791 case X86::ATOMUMAX64:
7792 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7794 // This group does 64-bit operations on a 32-bit host.
7795 case X86::ATOMAND6432:
7796 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7797 X86::AND32rr, X86::AND32rr,
7798 X86::AND32ri, X86::AND32ri,
7800 case X86::ATOMOR6432:
7801 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7802 X86::OR32rr, X86::OR32rr,
7803 X86::OR32ri, X86::OR32ri,
7805 case X86::ATOMXOR6432:
7806 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7807 X86::XOR32rr, X86::XOR32rr,
7808 X86::XOR32ri, X86::XOR32ri,
7810 case X86::ATOMNAND6432:
7811 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7812 X86::AND32rr, X86::AND32rr,
7813 X86::AND32ri, X86::AND32ri,
7815 case X86::ATOMADD6432:
7816 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7817 X86::ADD32rr, X86::ADC32rr,
7818 X86::ADD32ri, X86::ADC32ri,
7820 case X86::ATOMSUB6432:
7821 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7822 X86::SUB32rr, X86::SBB32rr,
7823 X86::SUB32ri, X86::SBB32ri,
7825 case X86::ATOMSWAP6432:
7826 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7827 X86::MOV32rr, X86::MOV32rr,
7828 X86::MOV32ri, X86::MOV32ri,
7833 //===----------------------------------------------------------------------===//
7834 // X86 Optimization Hooks
7835 //===----------------------------------------------------------------------===//
7837 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7841 const SelectionDAG &DAG,
7842 unsigned Depth) const {
7843 unsigned Opc = Op.getOpcode();
7844 assert((Opc >= ISD::BUILTIN_OP_END ||
7845 Opc == ISD::INTRINSIC_WO_CHAIN ||
7846 Opc == ISD::INTRINSIC_W_CHAIN ||
7847 Opc == ISD::INTRINSIC_VOID) &&
7848 "Should use MaskedValueIsZero if you don't know whether Op"
7849 " is a target node!");
7851 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7860 // These nodes' second result is a boolean.
7861 if (Op.getResNo() == 0)
7865 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7866 Mask.getBitWidth() - 1);
7871 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7872 /// node is a GlobalAddress + offset.
7873 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7874 GlobalValue* &GA, int64_t &Offset) const{
7875 if (N->getOpcode() == X86ISD::Wrapper) {
7876 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7877 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7878 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7882 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7885 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7886 const TargetLowering &TLI) {
7889 if (TLI.isGAPlusOffset(Base, GV, Offset))
7890 return (GV->getAlignment() >= N && (Offset % N) == 0);
7891 // DAG combine handles the stack object case.
7895 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7896 MVT EVT, LoadSDNode *&LDBase,
7897 unsigned &LastLoadedElt,
7898 SelectionDAG &DAG, MachineFrameInfo *MFI,
7899 const TargetLowering &TLI) {
7901 LastLoadedElt = -1U;
7902 for (unsigned i = 0; i < NumElems; ++i) {
7903 if (N->getMaskElt(i) < 0) {
7909 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7910 if (!Elt.getNode() ||
7911 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7914 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7916 LDBase = cast<LoadSDNode>(Elt.getNode());
7920 if (Elt.getOpcode() == ISD::UNDEF)
7923 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7924 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7931 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7932 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7933 /// if the load addresses are consecutive, non-overlapping, and in the right
7934 /// order. In the case of v2i64, it will see if it can rewrite the
7935 /// shuffle to be an appropriate build vector so it can take advantage of
7936 // performBuildVectorCombine.
7937 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7938 const TargetLowering &TLI) {
7939 DebugLoc dl = N->getDebugLoc();
7940 MVT VT = N->getValueType(0);
7941 MVT EVT = VT.getVectorElementType();
7942 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7943 unsigned NumElems = VT.getVectorNumElements();
7945 if (VT.getSizeInBits() != 128)
7948 // Try to combine a vector_shuffle into a 128-bit load.
7949 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7950 LoadSDNode *LD = NULL;
7951 unsigned LastLoadedElt;
7952 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7956 if (LastLoadedElt == NumElems - 1) {
7957 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7958 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7959 LD->getSrcValue(), LD->getSrcValueOffset(),
7961 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7962 LD->getSrcValue(), LD->getSrcValueOffset(),
7963 LD->isVolatile(), LD->getAlignment());
7964 } else if (NumElems == 4 && LastLoadedElt == 1) {
7965 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7966 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7967 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7968 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
7973 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7974 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7975 const X86Subtarget *Subtarget) {
7976 DebugLoc DL = N->getDebugLoc();
7977 SDValue Cond = N->getOperand(0);
7978 // Get the LHS/RHS of the select.
7979 SDValue LHS = N->getOperand(1);
7980 SDValue RHS = N->getOperand(2);
7982 // If we have SSE[12] support, try to form min/max nodes.
7983 if (Subtarget->hasSSE2() &&
7984 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7985 Cond.getOpcode() == ISD::SETCC) {
7986 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7988 unsigned Opcode = 0;
7989 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7992 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7995 if (!UnsafeFPMath) break;
7997 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7999 Opcode = X86ISD::FMIN;
8002 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8005 if (!UnsafeFPMath) break;
8007 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8009 Opcode = X86ISD::FMAX;
8012 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8015 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8018 if (!UnsafeFPMath) break;
8020 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8022 Opcode = X86ISD::FMIN;
8025 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8028 if (!UnsafeFPMath) break;
8030 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8032 Opcode = X86ISD::FMAX;
8038 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8041 // If this is a select between two integer constants, try to do some
8043 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8044 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8045 // Don't do this for crazy integer types.
8046 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8047 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8048 // so that TrueC (the true value) is larger than FalseC.
8049 bool NeedsCondInvert = false;
8051 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8052 // Efficiently invertible.
8053 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8054 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8055 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8056 NeedsCondInvert = true;
8057 std::swap(TrueC, FalseC);
8060 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8061 if (FalseC->getAPIntValue() == 0 &&
8062 TrueC->getAPIntValue().isPowerOf2()) {
8063 if (NeedsCondInvert) // Invert the condition if needed.
8064 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8065 DAG.getConstant(1, Cond.getValueType()));
8067 // Zero extend the condition if needed.
8068 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8070 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8071 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8072 DAG.getConstant(ShAmt, MVT::i8));
8075 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8076 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8077 if (NeedsCondInvert) // Invert the condition if needed.
8078 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8079 DAG.getConstant(1, Cond.getValueType()));
8081 // Zero extend the condition if needed.
8082 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8083 FalseC->getValueType(0), Cond);
8084 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8085 SDValue(FalseC, 0));
8088 // Optimize cases that will turn into an LEA instruction. This requires
8089 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8090 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8091 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8092 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8094 bool isFastMultiplier = false;
8096 switch ((unsigned char)Diff) {
8098 case 1: // result = add base, cond
8099 case 2: // result = lea base( , cond*2)
8100 case 3: // result = lea base(cond, cond*2)
8101 case 4: // result = lea base( , cond*4)
8102 case 5: // result = lea base(cond, cond*4)
8103 case 8: // result = lea base( , cond*8)
8104 case 9: // result = lea base(cond, cond*8)
8105 isFastMultiplier = true;
8110 if (isFastMultiplier) {
8111 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8112 if (NeedsCondInvert) // Invert the condition if needed.
8113 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8114 DAG.getConstant(1, Cond.getValueType()));
8116 // Zero extend the condition if needed.
8117 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8119 // Scale the condition by the difference.
8121 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8122 DAG.getConstant(Diff, Cond.getValueType()));
8124 // Add the base if non-zero.
8125 if (FalseC->getAPIntValue() != 0)
8126 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8127 SDValue(FalseC, 0));
8137 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8138 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8139 TargetLowering::DAGCombinerInfo &DCI) {
8140 DebugLoc DL = N->getDebugLoc();
8142 // If the flag operand isn't dead, don't touch this CMOV.
8143 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8146 // If this is a select between two integer constants, try to do some
8147 // optimizations. Note that the operands are ordered the opposite of SELECT
8149 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8150 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8151 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8152 // larger than FalseC (the false value).
8153 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8155 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8156 CC = X86::GetOppositeBranchCondition(CC);
8157 std::swap(TrueC, FalseC);
8160 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8161 // This is efficient for any integer data type (including i8/i16) and
8163 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8164 SDValue Cond = N->getOperand(3);
8165 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8166 DAG.getConstant(CC, MVT::i8), Cond);
8168 // Zero extend the condition if needed.
8169 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8171 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8172 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8173 DAG.getConstant(ShAmt, MVT::i8));
8174 if (N->getNumValues() == 2) // Dead flag value?
8175 return DCI.CombineTo(N, Cond, SDValue());
8179 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8180 // for any integer data type, including i8/i16.
8181 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8182 SDValue Cond = N->getOperand(3);
8183 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8184 DAG.getConstant(CC, MVT::i8), Cond);
8186 // Zero extend the condition if needed.
8187 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8188 FalseC->getValueType(0), Cond);
8189 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8190 SDValue(FalseC, 0));
8192 if (N->getNumValues() == 2) // Dead flag value?
8193 return DCI.CombineTo(N, Cond, SDValue());
8197 // Optimize cases that will turn into an LEA instruction. This requires
8198 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8199 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8200 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8201 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8203 bool isFastMultiplier = false;
8205 switch ((unsigned char)Diff) {
8207 case 1: // result = add base, cond
8208 case 2: // result = lea base( , cond*2)
8209 case 3: // result = lea base(cond, cond*2)
8210 case 4: // result = lea base( , cond*4)
8211 case 5: // result = lea base(cond, cond*4)
8212 case 8: // result = lea base( , cond*8)
8213 case 9: // result = lea base(cond, cond*8)
8214 isFastMultiplier = true;
8219 if (isFastMultiplier) {
8220 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8221 SDValue Cond = N->getOperand(3);
8222 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8223 DAG.getConstant(CC, MVT::i8), Cond);
8224 // Zero extend the condition if needed.
8225 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8227 // Scale the condition by the difference.
8229 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8230 DAG.getConstant(Diff, Cond.getValueType()));
8232 // Add the base if non-zero.
8233 if (FalseC->getAPIntValue() != 0)
8234 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8235 SDValue(FalseC, 0));
8236 if (N->getNumValues() == 2) // Dead flag value?
8237 return DCI.CombineTo(N, Cond, SDValue());
8247 /// PerformMulCombine - Optimize a single multiply with constant into two
8248 /// in order to implement it with two cheaper instructions, e.g.
8249 /// LEA + SHL, LEA + LEA.
8250 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8251 TargetLowering::DAGCombinerInfo &DCI) {
8252 if (DAG.getMachineFunction().
8253 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8256 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8259 MVT VT = N->getValueType(0);
8263 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8266 uint64_t MulAmt = C->getZExtValue();
8267 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8270 uint64_t MulAmt1 = 0;
8271 uint64_t MulAmt2 = 0;
8272 if ((MulAmt % 9) == 0) {
8274 MulAmt2 = MulAmt / 9;
8275 } else if ((MulAmt % 5) == 0) {
8277 MulAmt2 = MulAmt / 5;
8278 } else if ((MulAmt % 3) == 0) {
8280 MulAmt2 = MulAmt / 3;
8283 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8284 DebugLoc DL = N->getDebugLoc();
8286 if (isPowerOf2_64(MulAmt2) &&
8287 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8288 // If second multiplifer is pow2, issue it first. We want the multiply by
8289 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8291 std::swap(MulAmt1, MulAmt2);
8294 if (isPowerOf2_64(MulAmt1))
8295 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8296 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8298 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8299 DAG.getConstant(MulAmt1, VT));
8301 if (isPowerOf2_64(MulAmt2))
8302 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8303 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8305 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8306 DAG.getConstant(MulAmt2, VT));
8308 // Do not add new nodes to DAG combiner worklist.
8309 DCI.CombineTo(N, NewMul, false);
8315 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8317 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8318 const X86Subtarget *Subtarget) {
8319 // On X86 with SSE2 support, we can transform this to a vector shift if
8320 // all elements are shifted by the same amount. We can't do this in legalize
8321 // because the a constant vector is typically transformed to a constant pool
8322 // so we have no knowledge of the shift amount.
8323 if (!Subtarget->hasSSE2())
8326 MVT VT = N->getValueType(0);
8327 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8330 SDValue ShAmtOp = N->getOperand(1);
8331 MVT EltVT = VT.getVectorElementType();
8332 DebugLoc DL = N->getDebugLoc();
8334 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8335 unsigned NumElts = VT.getVectorNumElements();
8337 for (; i != NumElts; ++i) {
8338 SDValue Arg = ShAmtOp.getOperand(i);
8339 if (Arg.getOpcode() == ISD::UNDEF) continue;
8343 for (; i != NumElts; ++i) {
8344 SDValue Arg = ShAmtOp.getOperand(i);
8345 if (Arg.getOpcode() == ISD::UNDEF) continue;
8346 if (Arg != BaseShAmt) {
8350 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8351 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8352 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8353 DAG.getIntPtrConstant(0));
8357 if (EltVT.bitsGT(MVT::i32))
8358 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8359 else if (EltVT.bitsLT(MVT::i32))
8360 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8362 // The shift amount is identical so we can do a vector shift.
8363 SDValue ValOp = N->getOperand(0);
8364 switch (N->getOpcode()) {
8366 assert(0 && "Unknown shift opcode!");
8369 if (VT == MVT::v2i64)
8370 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8371 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8373 if (VT == MVT::v4i32)
8374 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8375 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8377 if (VT == MVT::v8i16)
8378 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8379 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8383 if (VT == MVT::v4i32)
8384 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8385 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8387 if (VT == MVT::v8i16)
8388 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8389 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8393 if (VT == MVT::v2i64)
8394 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8395 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8397 if (VT == MVT::v4i32)
8398 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8399 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8401 if (VT == MVT::v8i16)
8402 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8403 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8410 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8411 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8412 const X86Subtarget *Subtarget) {
8413 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8414 // the FP state in cases where an emms may be missing.
8415 // A preferable solution to the general problem is to figure out the right
8416 // places to insert EMMS. This qualifies as a quick hack.
8418 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8419 StoreSDNode *St = cast<StoreSDNode>(N);
8420 MVT VT = St->getValue().getValueType();
8421 if (VT.getSizeInBits() != 64)
8424 const Function *F = DAG.getMachineFunction().getFunction();
8425 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8426 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8427 && Subtarget->hasSSE2();
8428 if ((VT.isVector() ||
8429 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8430 isa<LoadSDNode>(St->getValue()) &&
8431 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8432 St->getChain().hasOneUse() && !St->isVolatile()) {
8433 SDNode* LdVal = St->getValue().getNode();
8435 int TokenFactorIndex = -1;
8436 SmallVector<SDValue, 8> Ops;
8437 SDNode* ChainVal = St->getChain().getNode();
8438 // Must be a store of a load. We currently handle two cases: the load
8439 // is a direct child, and it's under an intervening TokenFactor. It is
8440 // possible to dig deeper under nested TokenFactors.
8441 if (ChainVal == LdVal)
8442 Ld = cast<LoadSDNode>(St->getChain());
8443 else if (St->getValue().hasOneUse() &&
8444 ChainVal->getOpcode() == ISD::TokenFactor) {
8445 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8446 if (ChainVal->getOperand(i).getNode() == LdVal) {
8447 TokenFactorIndex = i;
8448 Ld = cast<LoadSDNode>(St->getValue());
8450 Ops.push_back(ChainVal->getOperand(i));
8454 if (!Ld || !ISD::isNormalLoad(Ld))
8457 // If this is not the MMX case, i.e. we are just turning i64 load/store
8458 // into f64 load/store, avoid the transformation if there are multiple
8459 // uses of the loaded value.
8460 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8463 DebugLoc LdDL = Ld->getDebugLoc();
8464 DebugLoc StDL = N->getDebugLoc();
8465 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8466 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8468 if (Subtarget->is64Bit() || F64IsLegal) {
8469 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8470 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8471 Ld->getBasePtr(), Ld->getSrcValue(),
8472 Ld->getSrcValueOffset(), Ld->isVolatile(),
8473 Ld->getAlignment());
8474 SDValue NewChain = NewLd.getValue(1);
8475 if (TokenFactorIndex != -1) {
8476 Ops.push_back(NewChain);
8477 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8480 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8481 St->getSrcValue(), St->getSrcValueOffset(),
8482 St->isVolatile(), St->getAlignment());
8485 // Otherwise, lower to two pairs of 32-bit loads / stores.
8486 SDValue LoAddr = Ld->getBasePtr();
8487 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8488 DAG.getConstant(4, MVT::i32));
8490 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8491 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8492 Ld->isVolatile(), Ld->getAlignment());
8493 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8494 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8496 MinAlign(Ld->getAlignment(), 4));
8498 SDValue NewChain = LoLd.getValue(1);
8499 if (TokenFactorIndex != -1) {
8500 Ops.push_back(LoLd);
8501 Ops.push_back(HiLd);
8502 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8506 LoAddr = St->getBasePtr();
8507 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8508 DAG.getConstant(4, MVT::i32));
8510 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8511 St->getSrcValue(), St->getSrcValueOffset(),
8512 St->isVolatile(), St->getAlignment());
8513 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8515 St->getSrcValueOffset() + 4,
8517 MinAlign(St->getAlignment(), 4));
8518 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8523 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8524 /// X86ISD::FXOR nodes.
8525 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8526 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8527 // F[X]OR(0.0, x) -> x
8528 // F[X]OR(x, 0.0) -> x
8529 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8530 if (C->getValueAPF().isPosZero())
8531 return N->getOperand(1);
8532 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8533 if (C->getValueAPF().isPosZero())
8534 return N->getOperand(0);
8538 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8539 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8540 // FAND(0.0, x) -> 0.0
8541 // FAND(x, 0.0) -> 0.0
8542 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8543 if (C->getValueAPF().isPosZero())
8544 return N->getOperand(0);
8545 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8546 if (C->getValueAPF().isPosZero())
8547 return N->getOperand(1);
8551 static SDValue PerformBTCombine(SDNode *N,
8553 TargetLowering::DAGCombinerInfo &DCI) {
8554 // BT ignores high bits in the bit index operand.
8555 SDValue Op1 = N->getOperand(1);
8556 if (Op1.hasOneUse()) {
8557 unsigned BitWidth = Op1.getValueSizeInBits();
8558 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8559 APInt KnownZero, KnownOne;
8560 TargetLowering::TargetLoweringOpt TLO(DAG);
8561 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8562 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8563 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8564 DCI.CommitTargetLoweringOpt(TLO);
8569 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8570 SDValue Op = N->getOperand(0);
8571 if (Op.getOpcode() == ISD::BIT_CONVERT)
8572 Op = Op.getOperand(0);
8573 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8574 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8575 VT.getVectorElementType().getSizeInBits() ==
8576 OpVT.getVectorElementType().getSizeInBits()) {
8577 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8582 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8583 // Locked instructions, in turn, have implicit fence semantics (all memory
8584 // operations are flushed before issuing the locked instruction, and the
8585 // are not buffered), so we can fold away the common pattern of
8586 // fence-atomic-fence.
8587 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8588 SDValue atomic = N->getOperand(0);
8589 switch (atomic.getOpcode()) {
8590 case ISD::ATOMIC_CMP_SWAP:
8591 case ISD::ATOMIC_SWAP:
8592 case ISD::ATOMIC_LOAD_ADD:
8593 case ISD::ATOMIC_LOAD_SUB:
8594 case ISD::ATOMIC_LOAD_AND:
8595 case ISD::ATOMIC_LOAD_OR:
8596 case ISD::ATOMIC_LOAD_XOR:
8597 case ISD::ATOMIC_LOAD_NAND:
8598 case ISD::ATOMIC_LOAD_MIN:
8599 case ISD::ATOMIC_LOAD_MAX:
8600 case ISD::ATOMIC_LOAD_UMIN:
8601 case ISD::ATOMIC_LOAD_UMAX:
8607 SDValue fence = atomic.getOperand(0);
8608 if (fence.getOpcode() != ISD::MEMBARRIER)
8611 switch (atomic.getOpcode()) {
8612 case ISD::ATOMIC_CMP_SWAP:
8613 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8614 atomic.getOperand(1), atomic.getOperand(2),
8615 atomic.getOperand(3));
8616 case ISD::ATOMIC_SWAP:
8617 case ISD::ATOMIC_LOAD_ADD:
8618 case ISD::ATOMIC_LOAD_SUB:
8619 case ISD::ATOMIC_LOAD_AND:
8620 case ISD::ATOMIC_LOAD_OR:
8621 case ISD::ATOMIC_LOAD_XOR:
8622 case ISD::ATOMIC_LOAD_NAND:
8623 case ISD::ATOMIC_LOAD_MIN:
8624 case ISD::ATOMIC_LOAD_MAX:
8625 case ISD::ATOMIC_LOAD_UMIN:
8626 case ISD::ATOMIC_LOAD_UMAX:
8627 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8628 atomic.getOperand(1), atomic.getOperand(2));
8634 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8635 DAGCombinerInfo &DCI) const {
8636 SelectionDAG &DAG = DCI.DAG;
8637 switch (N->getOpcode()) {
8639 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8640 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8641 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8642 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8645 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8646 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8648 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8649 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8650 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8651 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8652 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8658 //===----------------------------------------------------------------------===//
8659 // X86 Inline Assembly Support
8660 //===----------------------------------------------------------------------===//
8662 /// getConstraintType - Given a constraint letter, return the type of
8663 /// constraint it is for this target.
8664 X86TargetLowering::ConstraintType
8665 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8666 if (Constraint.size() == 1) {
8667 switch (Constraint[0]) {
8679 return C_RegisterClass;
8687 return TargetLowering::getConstraintType(Constraint);
8690 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8691 /// with another that has more specific requirements based on the type of the
8692 /// corresponding operand.
8693 const char *X86TargetLowering::
8694 LowerXConstraint(MVT ConstraintVT) const {
8695 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8696 // 'f' like normal targets.
8697 if (ConstraintVT.isFloatingPoint()) {
8698 if (Subtarget->hasSSE2())
8700 if (Subtarget->hasSSE1())
8704 return TargetLowering::LowerXConstraint(ConstraintVT);
8707 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8708 /// vector. If it is invalid, don't add anything to Ops.
8709 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8712 std::vector<SDValue>&Ops,
8713 SelectionDAG &DAG) const {
8714 SDValue Result(0, 0);
8716 switch (Constraint) {
8719 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8720 if (C->getZExtValue() <= 31) {
8721 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8727 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8728 if (C->getZExtValue() <= 63) {
8729 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8735 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8736 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8737 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8743 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8744 if (C->getZExtValue() <= 255) {
8745 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8751 // 32-bit signed value
8752 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8753 const ConstantInt *CI = C->getConstantIntValue();
8754 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8755 // Widen to 64 bits here to get it sign extended.
8756 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8759 // FIXME gcc accepts some relocatable values here too, but only in certain
8760 // memory models; it's complicated.
8765 // 32-bit unsigned value
8766 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8767 const ConstantInt *CI = C->getConstantIntValue();
8768 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8769 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8773 // FIXME gcc accepts some relocatable values here too, but only in certain
8774 // memory models; it's complicated.
8778 // Literal immediates are always ok.
8779 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8780 // Widen to 64 bits here to get it sign extended.
8781 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8785 // If we are in non-pic codegen mode, we allow the address of a global (with
8786 // an optional displacement) to be used with 'i'.
8787 GlobalAddressSDNode *GA = 0;
8790 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8792 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8793 Offset += GA->getOffset();
8795 } else if (Op.getOpcode() == ISD::ADD) {
8796 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8797 Offset += C->getZExtValue();
8798 Op = Op.getOperand(0);
8801 } else if (Op.getOpcode() == ISD::SUB) {
8802 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8803 Offset += -C->getZExtValue();
8804 Op = Op.getOperand(0);
8809 // Otherwise, this isn't something we can handle, reject it.
8813 // If we require an extra load to get this address, as in PIC mode, we
8815 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine()))
8819 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
8821 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8828 if (Result.getNode()) {
8829 Ops.push_back(Result);
8832 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8836 std::vector<unsigned> X86TargetLowering::
8837 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8839 if (Constraint.size() == 1) {
8840 // FIXME: not handling fp-stack yet!
8841 switch (Constraint[0]) { // GCC X86 Constraint Letters
8842 default: break; // Unknown constraint letter
8843 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8846 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8847 else if (VT == MVT::i16)
8848 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8849 else if (VT == MVT::i8)
8850 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8851 else if (VT == MVT::i64)
8852 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8857 return std::vector<unsigned>();
8860 std::pair<unsigned, const TargetRegisterClass*>
8861 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8863 // First, see if this is a constraint that directly corresponds to an LLVM
8865 if (Constraint.size() == 1) {
8866 // GCC Constraint Letters
8867 switch (Constraint[0]) {
8869 case 'r': // GENERAL_REGS
8870 case 'R': // LEGACY_REGS
8871 case 'l': // INDEX_REGS
8873 return std::make_pair(0U, X86::GR8RegisterClass);
8875 return std::make_pair(0U, X86::GR16RegisterClass);
8876 if (VT == MVT::i32 || !Subtarget->is64Bit())
8877 return std::make_pair(0U, X86::GR32RegisterClass);
8878 return std::make_pair(0U, X86::GR64RegisterClass);
8879 case 'f': // FP Stack registers.
8880 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8881 // value to the correct fpstack register class.
8882 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8883 return std::make_pair(0U, X86::RFP32RegisterClass);
8884 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8885 return std::make_pair(0U, X86::RFP64RegisterClass);
8886 return std::make_pair(0U, X86::RFP80RegisterClass);
8887 case 'y': // MMX_REGS if MMX allowed.
8888 if (!Subtarget->hasMMX()) break;
8889 return std::make_pair(0U, X86::VR64RegisterClass);
8890 case 'Y': // SSE_REGS if SSE2 allowed
8891 if (!Subtarget->hasSSE2()) break;
8893 case 'x': // SSE_REGS if SSE1 allowed
8894 if (!Subtarget->hasSSE1()) break;
8896 switch (VT.getSimpleVT()) {
8898 // Scalar SSE types.
8901 return std::make_pair(0U, X86::FR32RegisterClass);
8904 return std::make_pair(0U, X86::FR64RegisterClass);
8912 return std::make_pair(0U, X86::VR128RegisterClass);
8918 // Use the default implementation in TargetLowering to convert the register
8919 // constraint into a member of a register class.
8920 std::pair<unsigned, const TargetRegisterClass*> Res;
8921 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8923 // Not found as a standard register?
8924 if (Res.second == 0) {
8925 // GCC calls "st(0)" just plain "st".
8926 if (StringsEqualNoCase("{st}", Constraint)) {
8927 Res.first = X86::ST0;
8928 Res.second = X86::RFP80RegisterClass;
8930 // 'A' means EAX + EDX.
8931 if (Constraint == "A") {
8932 Res.first = X86::EAX;
8933 Res.second = X86::GRADRegisterClass;
8938 // Otherwise, check to see if this is a register class of the wrong value
8939 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8940 // turn into {ax},{dx}.
8941 if (Res.second->hasType(VT))
8942 return Res; // Correct type already, nothing to do.
8944 // All of the single-register GCC register classes map their values onto
8945 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8946 // really want an 8-bit or 32-bit register, map to the appropriate register
8947 // class and return the appropriate register.
8948 if (Res.second == X86::GR16RegisterClass) {
8949 if (VT == MVT::i8) {
8950 unsigned DestReg = 0;
8951 switch (Res.first) {
8953 case X86::AX: DestReg = X86::AL; break;
8954 case X86::DX: DestReg = X86::DL; break;
8955 case X86::CX: DestReg = X86::CL; break;
8956 case X86::BX: DestReg = X86::BL; break;
8959 Res.first = DestReg;
8960 Res.second = X86::GR8RegisterClass;
8962 } else if (VT == MVT::i32) {
8963 unsigned DestReg = 0;
8964 switch (Res.first) {
8966 case X86::AX: DestReg = X86::EAX; break;
8967 case X86::DX: DestReg = X86::EDX; break;
8968 case X86::CX: DestReg = X86::ECX; break;
8969 case X86::BX: DestReg = X86::EBX; break;
8970 case X86::SI: DestReg = X86::ESI; break;
8971 case X86::DI: DestReg = X86::EDI; break;
8972 case X86::BP: DestReg = X86::EBP; break;
8973 case X86::SP: DestReg = X86::ESP; break;
8976 Res.first = DestReg;
8977 Res.second = X86::GR32RegisterClass;
8979 } else if (VT == MVT::i64) {
8980 unsigned DestReg = 0;
8981 switch (Res.first) {
8983 case X86::AX: DestReg = X86::RAX; break;
8984 case X86::DX: DestReg = X86::RDX; break;
8985 case X86::CX: DestReg = X86::RCX; break;
8986 case X86::BX: DestReg = X86::RBX; break;
8987 case X86::SI: DestReg = X86::RSI; break;
8988 case X86::DI: DestReg = X86::RDI; break;
8989 case X86::BP: DestReg = X86::RBP; break;
8990 case X86::SP: DestReg = X86::RSP; break;
8993 Res.first = DestReg;
8994 Res.second = X86::GR64RegisterClass;
8997 } else if (Res.second == X86::FR32RegisterClass ||
8998 Res.second == X86::FR64RegisterClass ||
8999 Res.second == X86::VR128RegisterClass) {
9000 // Handle references to XMM physical registers that got mapped into the
9001 // wrong class. This can happen with constraints like {xmm0} where the
9002 // target independent register mapper will just pick the first match it can
9003 // find, ignoring the required type.
9005 Res.second = X86::FR32RegisterClass;
9006 else if (VT == MVT::f64)
9007 Res.second = X86::FR64RegisterClass;
9008 else if (X86::VR128RegisterClass->hasType(VT))
9009 Res.second = X86::VR128RegisterClass;
9015 //===----------------------------------------------------------------------===//
9016 // X86 Widen vector type
9017 //===----------------------------------------------------------------------===//
9019 /// getWidenVectorType: given a vector type, returns the type to widen
9020 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9021 /// If there is no vector type that we want to widen to, returns MVT::Other
9022 /// When and where to widen is target dependent based on the cost of
9023 /// scalarizing vs using the wider vector type.
9025 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9026 assert(VT.isVector());
9027 if (isTypeLegal(VT))
9030 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9031 // type based on element type. This would speed up our search (though
9032 // it may not be worth it since the size of the list is relatively
9034 MVT EltVT = VT.getVectorElementType();
9035 unsigned NElts = VT.getVectorNumElements();
9037 // On X86, it make sense to widen any vector wider than 1
9041 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9042 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9043 MVT SVT = (MVT::SimpleValueType)nVT;
9045 if (isTypeLegal(SVT) &&
9046 SVT.getVectorElementType() == EltVT &&
9047 SVT.getVectorNumElements() > NElts)