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->isPICStyleRIPRel())
1025 // This doesn't have DebugLoc associated with it, but is not really the
1026 // same as a Register.
1027 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1032 /// getFunctionAlignment - Return the Log2 alignment of this function.
1033 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1034 return F->hasFnAttr(Attribute::OptimizeForSize) ? 1 : 4;
1037 //===----------------------------------------------------------------------===//
1038 // Return Value Calling Convention Implementation
1039 //===----------------------------------------------------------------------===//
1041 #include "X86GenCallingConv.inc"
1043 /// LowerRET - Lower an ISD::RET node.
1044 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1045 DebugLoc dl = Op.getDebugLoc();
1046 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
1048 SmallVector<CCValAssign, 16> RVLocs;
1049 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1050 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1051 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
1052 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
1054 // If this is the first return lowered for this function, add the regs to the
1055 // liveout set for the function.
1056 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1057 for (unsigned i = 0; i != RVLocs.size(); ++i)
1058 if (RVLocs[i].isRegLoc())
1059 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1061 SDValue Chain = Op.getOperand(0);
1063 // Handle tail call return.
1064 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
1065 if (Chain.getOpcode() == X86ISD::TAILCALL) {
1066 SDValue TailCall = Chain;
1067 SDValue TargetAddress = TailCall.getOperand(1);
1068 SDValue StackAdjustment = TailCall.getOperand(2);
1069 assert(((TargetAddress.getOpcode() == ISD::Register &&
1070 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
1071 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
1072 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
1073 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
1074 "Expecting an global address, external symbol, or register");
1075 assert(StackAdjustment.getOpcode() == ISD::Constant &&
1076 "Expecting a const value");
1078 SmallVector<SDValue,8> Operands;
1079 Operands.push_back(Chain.getOperand(0));
1080 Operands.push_back(TargetAddress);
1081 Operands.push_back(StackAdjustment);
1082 // Copy registers used by the call. Last operand is a flag so it is not
1084 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
1085 Operands.push_back(Chain.getOperand(i));
1087 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
1094 SmallVector<SDValue, 6> RetOps;
1095 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1096 // Operand #1 = Bytes To Pop
1097 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1099 // Copy the result values into the output registers.
1100 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1101 CCValAssign &VA = RVLocs[i];
1102 assert(VA.isRegLoc() && "Can only return in registers!");
1103 SDValue ValToCopy = Op.getOperand(i*2+1);
1105 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1106 // the RET instruction and handled by the FP Stackifier.
1107 if (VA.getLocReg() == X86::ST0 ||
1108 VA.getLocReg() == X86::ST1) {
1109 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1110 // change the value to the FP stack register class.
1111 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1112 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1113 RetOps.push_back(ValToCopy);
1114 // Don't emit a copytoreg.
1118 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1119 // which is returned in RAX / RDX.
1120 if (Subtarget->is64Bit()) {
1121 MVT ValVT = ValToCopy.getValueType();
1122 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1123 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1124 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1125 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1129 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1130 Flag = Chain.getValue(1);
1133 // The x86-64 ABI for returning structs by value requires that we copy
1134 // the sret argument into %rax for the return. We saved the argument into
1135 // a virtual register in the entry block, so now we copy the value out
1137 if (Subtarget->is64Bit() &&
1138 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1139 MachineFunction &MF = DAG.getMachineFunction();
1140 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1141 unsigned Reg = FuncInfo->getSRetReturnReg();
1143 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1144 FuncInfo->setSRetReturnReg(Reg);
1146 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1148 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1149 Flag = Chain.getValue(1);
1152 RetOps[0] = Chain; // Update chain.
1154 // Add the flag if we have it.
1156 RetOps.push_back(Flag);
1158 return DAG.getNode(X86ISD::RET_FLAG, dl,
1159 MVT::Other, &RetOps[0], RetOps.size());
1163 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1164 /// appropriate copies out of appropriate physical registers. This assumes that
1165 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1166 /// being lowered. The returns a SDNode with the same number of values as the
1168 SDNode *X86TargetLowering::
1169 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1170 unsigned CallingConv, SelectionDAG &DAG) {
1172 DebugLoc dl = TheCall->getDebugLoc();
1173 // Assign locations to each value returned by this call.
1174 SmallVector<CCValAssign, 16> RVLocs;
1175 bool isVarArg = TheCall->isVarArg();
1176 bool Is64Bit = Subtarget->is64Bit();
1177 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1178 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1180 SmallVector<SDValue, 8> ResultVals;
1182 // Copy all of the result registers out of their specified physreg.
1183 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1184 CCValAssign &VA = RVLocs[i];
1185 MVT CopyVT = VA.getValVT();
1187 // If this is x86-64, and we disabled SSE, we can't return FP values
1188 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1189 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1190 llvm_report_error("SSE register return with SSE disabled");
1193 // If this is a call to a function that returns an fp value on the floating
1194 // point stack, but where we prefer to use the value in xmm registers, copy
1195 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1196 if ((VA.getLocReg() == X86::ST0 ||
1197 VA.getLocReg() == X86::ST1) &&
1198 isScalarFPTypeInSSEReg(VA.getValVT())) {
1203 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1204 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1205 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1206 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1207 MVT::v2i64, InFlag).getValue(1);
1208 Val = Chain.getValue(0);
1209 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1210 Val, DAG.getConstant(0, MVT::i64));
1212 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1213 MVT::i64, InFlag).getValue(1);
1214 Val = Chain.getValue(0);
1216 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1218 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1219 CopyVT, InFlag).getValue(1);
1220 Val = Chain.getValue(0);
1222 InFlag = Chain.getValue(2);
1224 if (CopyVT != VA.getValVT()) {
1225 // Round the F80 the right size, which also moves to the appropriate xmm
1227 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1228 // This truncation won't change the value.
1229 DAG.getIntPtrConstant(1));
1232 ResultVals.push_back(Val);
1235 // Merge everything together with a MERGE_VALUES node.
1236 ResultVals.push_back(Chain);
1237 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1238 &ResultVals[0], ResultVals.size()).getNode();
1242 //===----------------------------------------------------------------------===//
1243 // C & StdCall & Fast Calling Convention implementation
1244 //===----------------------------------------------------------------------===//
1245 // StdCall calling convention seems to be standard for many Windows' API
1246 // routines and around. It differs from C calling convention just a little:
1247 // callee should clean up the stack, not caller. Symbols should be also
1248 // decorated in some fancy way :) It doesn't support any vector arguments.
1249 // For info on fast calling convention see Fast Calling Convention (tail call)
1250 // implementation LowerX86_32FastCCCallTo.
1252 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1254 static bool CallIsStructReturn(CallSDNode *TheCall) {
1255 unsigned NumOps = TheCall->getNumArgs();
1259 return TheCall->getArgFlags(0).isSRet();
1262 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1263 /// return semantics.
1264 static bool ArgsAreStructReturn(SDValue Op) {
1265 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1269 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1272 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1273 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1275 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1279 switch (CallingConv) {
1282 case CallingConv::X86_StdCall:
1283 return !Subtarget->is64Bit();
1284 case CallingConv::X86_FastCall:
1285 return !Subtarget->is64Bit();
1286 case CallingConv::Fast:
1287 return PerformTailCallOpt;
1291 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1292 /// given CallingConvention value.
1293 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1294 if (Subtarget->is64Bit()) {
1295 if (Subtarget->isTargetWin64())
1296 return CC_X86_Win64_C;
1301 if (CC == CallingConv::X86_FastCall)
1302 return CC_X86_32_FastCall;
1303 else if (CC == CallingConv::Fast)
1304 return CC_X86_32_FastCC;
1309 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1310 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1312 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1313 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1314 if (CC == CallingConv::X86_FastCall)
1316 else if (CC == CallingConv::X86_StdCall)
1322 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1323 /// in a register before calling.
1324 static bool CallRequiresGOTPtrInReg(const TargetMachine &TM,
1326 const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
1328 return !IsTailCall && !Subtarget.is64Bit() &&
1329 TM.getRelocationModel() == Reloc::PIC_ &&
1330 Subtarget.isPICStyleGOT();
1333 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1334 /// address to be loaded in a register.
1335 static bool CallRequiresFnAddressInReg(const TargetMachine &TM,
1337 const X86Subtarget &Subtarget = TM.getSubtarget<X86Subtarget>();
1338 return !Subtarget.is64Bit() && IsTailCall &&
1339 TM.getRelocationModel() == Reloc::PIC_ &&
1340 Subtarget.isPICStyleGOT();
1343 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1344 /// by "Src" to address "Dst" with size and alignment information specified by
1345 /// the specific parameter attribute. The copy will be passed as a byval
1346 /// function parameter.
1348 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1349 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1351 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1352 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1353 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1356 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1357 const CCValAssign &VA,
1358 MachineFrameInfo *MFI,
1360 SDValue Root, unsigned i) {
1361 // Create the nodes corresponding to a load from this parameter slot.
1362 ISD::ArgFlagsTy Flags =
1363 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1364 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1365 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1367 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1368 // changed with more analysis.
1369 // In case of tail call optimization mark all arguments mutable. Since they
1370 // could be overwritten by lowering of arguments in case of a tail call.
1371 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1372 VA.getLocMemOffset(), isImmutable);
1373 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1374 if (Flags.isByVal())
1376 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1377 PseudoSourceValue::getFixedStack(FI), 0);
1381 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1382 MachineFunction &MF = DAG.getMachineFunction();
1383 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1384 DebugLoc dl = Op.getDebugLoc();
1386 const Function* Fn = MF.getFunction();
1387 if (Fn->hasExternalLinkage() &&
1388 Subtarget->isTargetCygMing() &&
1389 Fn->getName() == "main")
1390 FuncInfo->setForceFramePointer(true);
1392 // Decorate the function name.
1393 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1395 MachineFrameInfo *MFI = MF.getFrameInfo();
1396 SDValue Root = Op.getOperand(0);
1397 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1398 unsigned CC = MF.getFunction()->getCallingConv();
1399 bool Is64Bit = Subtarget->is64Bit();
1400 bool IsWin64 = Subtarget->isTargetWin64();
1402 assert(!(isVarArg && CC == CallingConv::Fast) &&
1403 "Var args not supported with calling convention fastcc");
1405 // Assign locations to all of the incoming arguments.
1406 SmallVector<CCValAssign, 16> ArgLocs;
1407 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1408 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1410 SmallVector<SDValue, 8> ArgValues;
1411 unsigned LastVal = ~0U;
1412 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1413 CCValAssign &VA = ArgLocs[i];
1414 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1416 assert(VA.getValNo() != LastVal &&
1417 "Don't support value assigned to multiple locs yet");
1418 LastVal = VA.getValNo();
1420 if (VA.isRegLoc()) {
1421 MVT RegVT = VA.getLocVT();
1422 TargetRegisterClass *RC = NULL;
1423 if (RegVT == MVT::i32)
1424 RC = X86::GR32RegisterClass;
1425 else if (Is64Bit && RegVT == MVT::i64)
1426 RC = X86::GR64RegisterClass;
1427 else if (RegVT == MVT::f32)
1428 RC = X86::FR32RegisterClass;
1429 else if (RegVT == MVT::f64)
1430 RC = X86::FR64RegisterClass;
1431 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1432 RC = X86::VR128RegisterClass;
1433 else if (RegVT.isVector()) {
1434 assert(RegVT.getSizeInBits() == 64);
1436 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1438 // Darwin calling convention passes MMX values in either GPRs or
1439 // XMMs in x86-64. Other targets pass them in memory.
1440 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1441 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1444 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1449 assert(0 && "Unknown argument type!");
1452 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1453 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1455 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1456 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1458 if (VA.getLocInfo() == CCValAssign::SExt)
1459 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1460 DAG.getValueType(VA.getValVT()));
1461 else if (VA.getLocInfo() == CCValAssign::ZExt)
1462 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1463 DAG.getValueType(VA.getValVT()));
1465 if (VA.getLocInfo() != CCValAssign::Full)
1466 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1468 // Handle MMX values passed in GPRs.
1469 if (Is64Bit && RegVT != VA.getLocVT()) {
1470 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1471 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1472 else if (RC == X86::VR128RegisterClass) {
1473 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1474 ArgValue, DAG.getConstant(0, MVT::i64));
1475 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1479 ArgValues.push_back(ArgValue);
1481 assert(VA.isMemLoc());
1482 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1486 // The x86-64 ABI for returning structs by value requires that we copy
1487 // the sret argument into %rax for the return. Save the argument into
1488 // a virtual register so that we can access it from the return points.
1489 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1490 MachineFunction &MF = DAG.getMachineFunction();
1491 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1492 unsigned Reg = FuncInfo->getSRetReturnReg();
1494 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1495 FuncInfo->setSRetReturnReg(Reg);
1497 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1498 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1501 unsigned StackSize = CCInfo.getNextStackOffset();
1502 // align stack specially for tail calls
1503 if (PerformTailCallOpt && CC == CallingConv::Fast)
1504 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1506 // If the function takes variable number of arguments, make a frame index for
1507 // the start of the first vararg value... for expansion of llvm.va_start.
1509 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1510 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1513 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1515 // FIXME: We should really autogenerate these arrays
1516 static const unsigned GPR64ArgRegsWin64[] = {
1517 X86::RCX, X86::RDX, X86::R8, X86::R9
1519 static const unsigned XMMArgRegsWin64[] = {
1520 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1522 static const unsigned GPR64ArgRegs64Bit[] = {
1523 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1525 static const unsigned XMMArgRegs64Bit[] = {
1526 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1527 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1529 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1532 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1533 GPR64ArgRegs = GPR64ArgRegsWin64;
1534 XMMArgRegs = XMMArgRegsWin64;
1536 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1537 GPR64ArgRegs = GPR64ArgRegs64Bit;
1538 XMMArgRegs = XMMArgRegs64Bit;
1540 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1542 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1545 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1546 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1547 "SSE register cannot be used when SSE is disabled!");
1548 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1549 "SSE register cannot be used when SSE is disabled!");
1550 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1551 // Kernel mode asks for SSE to be disabled, so don't push them
1553 TotalNumXMMRegs = 0;
1555 // For X86-64, if there are vararg parameters that are passed via
1556 // registers, then we must store them to their spots on the stack so they
1557 // may be loaded by deferencing the result of va_next.
1558 VarArgsGPOffset = NumIntRegs * 8;
1559 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1560 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1561 TotalNumXMMRegs * 16, 16);
1563 // Store the integer parameter registers.
1564 SmallVector<SDValue, 8> MemOps;
1565 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1566 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1567 DAG.getIntPtrConstant(VarArgsGPOffset));
1568 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1569 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1570 X86::GR64RegisterClass);
1571 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1573 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1574 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1575 MemOps.push_back(Store);
1576 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1577 DAG.getIntPtrConstant(8));
1580 // Now store the XMM (fp + vector) parameter registers.
1581 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1582 DAG.getIntPtrConstant(VarArgsFPOffset));
1583 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1584 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1585 X86::VR128RegisterClass);
1586 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1588 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1589 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1590 MemOps.push_back(Store);
1591 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1592 DAG.getIntPtrConstant(16));
1594 if (!MemOps.empty())
1595 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1596 &MemOps[0], MemOps.size());
1600 ArgValues.push_back(Root);
1602 // Some CCs need callee pop.
1603 if (IsCalleePop(isVarArg, CC)) {
1604 BytesToPopOnReturn = StackSize; // Callee pops everything.
1605 BytesCallerReserves = 0;
1607 BytesToPopOnReturn = 0; // Callee pops nothing.
1608 // If this is an sret function, the return should pop the hidden pointer.
1609 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1610 BytesToPopOnReturn = 4;
1611 BytesCallerReserves = StackSize;
1615 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1616 if (CC == CallingConv::X86_FastCall)
1617 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1620 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1622 // Return the new list of results.
1623 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1624 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1628 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1629 const SDValue &StackPtr,
1630 const CCValAssign &VA,
1632 SDValue Arg, ISD::ArgFlagsTy Flags) {
1633 DebugLoc dl = TheCall->getDebugLoc();
1634 unsigned LocMemOffset = VA.getLocMemOffset();
1635 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1636 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1637 if (Flags.isByVal()) {
1638 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1640 return DAG.getStore(Chain, dl, Arg, PtrOff,
1641 PseudoSourceValue::getStack(), LocMemOffset);
1644 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1645 /// optimization is performed and it is required.
1647 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1648 SDValue &OutRetAddr,
1654 if (!IsTailCall || FPDiff==0) return Chain;
1656 // Adjust the Return address stack slot.
1657 MVT VT = getPointerTy();
1658 OutRetAddr = getReturnAddressFrameIndex(DAG);
1660 // Load the "old" Return address.
1661 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1662 return SDValue(OutRetAddr.getNode(), 1);
1665 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1666 /// optimization is performed and it is required (FPDiff!=0).
1668 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1669 SDValue Chain, SDValue RetAddrFrIdx,
1670 bool Is64Bit, int FPDiff, DebugLoc dl) {
1671 // Store the return address to the appropriate stack slot.
1672 if (!FPDiff) return Chain;
1673 // Calculate the new stack slot for the return address.
1674 int SlotSize = Is64Bit ? 8 : 4;
1675 int NewReturnAddrFI =
1676 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1677 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1678 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1679 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1680 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1684 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1685 MachineFunction &MF = DAG.getMachineFunction();
1686 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1687 SDValue Chain = TheCall->getChain();
1688 unsigned CC = TheCall->getCallingConv();
1689 bool isVarArg = TheCall->isVarArg();
1690 bool IsTailCall = TheCall->isTailCall() &&
1691 CC == CallingConv::Fast && PerformTailCallOpt;
1692 SDValue Callee = TheCall->getCallee();
1693 bool Is64Bit = Subtarget->is64Bit();
1694 bool IsStructRet = CallIsStructReturn(TheCall);
1695 DebugLoc dl = TheCall->getDebugLoc();
1697 assert(!(isVarArg && CC == CallingConv::Fast) &&
1698 "Var args not supported with calling convention fastcc");
1700 // Analyze operands of the call, assigning locations to each operand.
1701 SmallVector<CCValAssign, 16> ArgLocs;
1702 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1703 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1705 // Get a count of how many bytes are to be pushed on the stack.
1706 unsigned NumBytes = CCInfo.getNextStackOffset();
1707 if (PerformTailCallOpt && CC == CallingConv::Fast)
1708 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1712 // Lower arguments at fp - stackoffset + fpdiff.
1713 unsigned NumBytesCallerPushed =
1714 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1715 FPDiff = NumBytesCallerPushed - NumBytes;
1717 // Set the delta of movement of the returnaddr stackslot.
1718 // But only set if delta is greater than previous delta.
1719 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1720 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1723 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1725 SDValue RetAddrFrIdx;
1726 // Load return adress for tail calls.
1727 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1730 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1731 SmallVector<SDValue, 8> MemOpChains;
1734 // Walk the register/memloc assignments, inserting copies/loads. In the case
1735 // of tail call optimization arguments are handle later.
1736 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1737 CCValAssign &VA = ArgLocs[i];
1738 SDValue Arg = TheCall->getArg(i);
1739 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1740 bool isByVal = Flags.isByVal();
1742 // Promote the value if needed.
1743 switch (VA.getLocInfo()) {
1744 default: assert(0 && "Unknown loc info!");
1745 case CCValAssign::Full: break;
1746 case CCValAssign::SExt:
1747 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1749 case CCValAssign::ZExt:
1750 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1752 case CCValAssign::AExt:
1753 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1757 if (VA.isRegLoc()) {
1759 MVT RegVT = VA.getLocVT();
1760 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1761 switch (VA.getLocReg()) {
1764 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1766 // Special case: passing MMX values in GPR registers.
1767 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1770 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1771 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1772 // Special case: passing MMX values in XMM registers.
1773 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1774 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1775 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1780 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1782 if (!IsTailCall || (IsTailCall && isByVal)) {
1783 assert(VA.isMemLoc());
1784 if (StackPtr.getNode() == 0)
1785 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1787 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1788 Chain, Arg, Flags));
1793 if (!MemOpChains.empty())
1794 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1795 &MemOpChains[0], MemOpChains.size());
1797 // Build a sequence of copy-to-reg nodes chained together with token chain
1798 // and flag operands which copy the outgoing args into registers.
1800 // Tail call byval lowering might overwrite argument registers so in case of
1801 // tail call optimization the copies to registers are lowered later.
1803 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1804 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1805 RegsToPass[i].second, InFlag);
1806 InFlag = Chain.getValue(1);
1809 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1811 if (CallRequiresGOTPtrInReg(getTargetMachine(), IsTailCall)) {
1812 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1813 DAG.getNode(X86ISD::GlobalBaseReg,
1814 DebugLoc::getUnknownLoc(),
1817 InFlag = Chain.getValue(1);
1820 // If we are tail calling and generating PIC/GOT style code load the address
1821 // of the callee into ecx. The value in ecx is used as target of the tail
1822 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1823 // calls on PIC/GOT architectures. Normally we would just put the address of
1824 // GOT into ebx and then call target@PLT. But for tail calls ebx would be
1825 // restored (since ebx is callee saved) before jumping to the target@PLT.
1826 if (CallRequiresFnAddressInReg(getTargetMachine(), IsTailCall)) {
1827 // Note: The actual moving to ecx is done further down.
1828 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1829 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1830 !G->getGlobal()->hasProtectedVisibility())
1831 Callee = LowerGlobalAddress(Callee, DAG);
1832 else if (isa<ExternalSymbolSDNode>(Callee))
1833 Callee = LowerExternalSymbol(Callee,DAG);
1836 if (Is64Bit && isVarArg) {
1837 // From AMD64 ABI document:
1838 // For calls that may call functions that use varargs or stdargs
1839 // (prototype-less calls or calls to functions containing ellipsis (...) in
1840 // the declaration) %al is used as hidden argument to specify the number
1841 // of SSE registers used. The contents of %al do not need to match exactly
1842 // the number of registers, but must be an ubound on the number of SSE
1843 // registers used and is in the range 0 - 8 inclusive.
1845 // FIXME: Verify this on Win64
1846 // Count the number of XMM registers allocated.
1847 static const unsigned XMMArgRegs[] = {
1848 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1849 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1851 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1852 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1853 && "SSE registers cannot be used when SSE is disabled");
1855 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1856 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1857 InFlag = Chain.getValue(1);
1861 // For tail calls lower the arguments to the 'real' stack slot.
1863 SmallVector<SDValue, 8> MemOpChains2;
1866 // Do not flag preceeding copytoreg stuff together with the following stuff.
1868 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1869 CCValAssign &VA = ArgLocs[i];
1870 if (!VA.isRegLoc()) {
1871 assert(VA.isMemLoc());
1872 SDValue Arg = TheCall->getArg(i);
1873 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1874 // Create frame index.
1875 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1876 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1877 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1878 FIN = DAG.getFrameIndex(FI, getPointerTy());
1880 if (Flags.isByVal()) {
1881 // Copy relative to framepointer.
1882 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1883 if (StackPtr.getNode() == 0)
1884 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1886 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1888 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1891 // Store relative to framepointer.
1892 MemOpChains2.push_back(
1893 DAG.getStore(Chain, dl, Arg, FIN,
1894 PseudoSourceValue::getFixedStack(FI), 0));
1899 if (!MemOpChains2.empty())
1900 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1901 &MemOpChains2[0], MemOpChains2.size());
1903 // Copy arguments to their registers.
1904 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1905 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1906 RegsToPass[i].second, InFlag);
1907 InFlag = Chain.getValue(1);
1911 // Store the return address to the appropriate stack slot.
1912 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1916 // If the callee is a GlobalAddress node (quite common, every direct call is)
1917 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1918 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1919 // We should use extra load for direct calls to dllimported functions in
1921 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1922 getTargetMachine(), true))
1923 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1925 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1926 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1927 } else if (IsTailCall) {
1928 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1930 Chain = DAG.getCopyToReg(Chain, dl,
1931 DAG.getRegister(Opc, getPointerTy()),
1933 Callee = DAG.getRegister(Opc, getPointerTy());
1934 // Add register as live out.
1935 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1938 // Returns a chain & a flag for retval copy to use.
1939 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1940 SmallVector<SDValue, 8> Ops;
1943 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1944 DAG.getIntPtrConstant(0, true), InFlag);
1945 InFlag = Chain.getValue(1);
1947 // Returns a chain & a flag for retval copy to use.
1948 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1952 Ops.push_back(Chain);
1953 Ops.push_back(Callee);
1956 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1958 // Add argument registers to the end of the list so that they are known live
1960 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1961 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1962 RegsToPass[i].second.getValueType()));
1964 // Add an implicit use GOT pointer in EBX.
1965 if (!IsTailCall && !Is64Bit &&
1966 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1967 Subtarget->isPICStyleGOT())
1968 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1970 // Add an implicit use of AL for x86 vararg functions.
1971 if (Is64Bit && isVarArg)
1972 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1974 if (InFlag.getNode())
1975 Ops.push_back(InFlag);
1978 assert(InFlag.getNode() &&
1979 "Flag must be set. Depend on flag being set in LowerRET");
1980 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1981 TheCall->getVTList(), &Ops[0], Ops.size());
1983 return SDValue(Chain.getNode(), Op.getResNo());
1986 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1987 InFlag = Chain.getValue(1);
1989 // Create the CALLSEQ_END node.
1990 unsigned NumBytesForCalleeToPush;
1991 if (IsCalleePop(isVarArg, CC))
1992 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1993 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1994 // If this is is a call to a struct-return function, the callee
1995 // pops the hidden struct pointer, so we have to push it back.
1996 // This is common for Darwin/X86, Linux & Mingw32 targets.
1997 NumBytesForCalleeToPush = 4;
1999 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2001 // Returns a flag for retval copy to use.
2002 Chain = DAG.getCALLSEQ_END(Chain,
2003 DAG.getIntPtrConstant(NumBytes, true),
2004 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2007 InFlag = Chain.getValue(1);
2009 // Handle result values, copying them out of physregs into vregs that we
2011 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
2016 //===----------------------------------------------------------------------===//
2017 // Fast Calling Convention (tail call) implementation
2018 //===----------------------------------------------------------------------===//
2020 // Like std call, callee cleans arguments, convention except that ECX is
2021 // reserved for storing the tail called function address. Only 2 registers are
2022 // free for argument passing (inreg). Tail call optimization is performed
2024 // * tailcallopt is enabled
2025 // * caller/callee are fastcc
2026 // On X86_64 architecture with GOT-style position independent code only local
2027 // (within module) calls are supported at the moment.
2028 // To keep the stack aligned according to platform abi the function
2029 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2030 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2031 // If a tail called function callee has more arguments than the caller the
2032 // caller needs to make sure that there is room to move the RETADDR to. This is
2033 // achieved by reserving an area the size of the argument delta right after the
2034 // original REtADDR, but before the saved framepointer or the spilled registers
2035 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2047 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2048 /// for a 16 byte align requirement.
2049 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2050 SelectionDAG& DAG) {
2051 MachineFunction &MF = DAG.getMachineFunction();
2052 const TargetMachine &TM = MF.getTarget();
2053 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2054 unsigned StackAlignment = TFI.getStackAlignment();
2055 uint64_t AlignMask = StackAlignment - 1;
2056 int64_t Offset = StackSize;
2057 uint64_t SlotSize = TD->getPointerSize();
2058 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2059 // Number smaller than 12 so just add the difference.
2060 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2062 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2063 Offset = ((~AlignMask) & Offset) + StackAlignment +
2064 (StackAlignment-SlotSize);
2069 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2070 /// following the call is a return. A function is eligible if caller/callee
2071 /// calling conventions match, currently only fastcc supports tail calls, and
2072 /// the function CALL is immediatly followed by a RET.
2073 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2075 SelectionDAG& DAG) const {
2076 if (!PerformTailCallOpt)
2079 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2080 MachineFunction &MF = DAG.getMachineFunction();
2081 unsigned CallerCC = MF.getFunction()->getCallingConv();
2082 unsigned CalleeCC= TheCall->getCallingConv();
2083 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2084 SDValue Callee = TheCall->getCallee();
2085 // On x86/32Bit PIC/GOT tail calls are supported.
2086 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
2087 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
2090 // Can only do local tail calls (in same module, hidden or protected) on
2091 // x86_64 PIC/GOT at the moment.
2092 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2093 return G->getGlobal()->hasHiddenVisibility()
2094 || G->getGlobal()->hasProtectedVisibility();
2102 X86TargetLowering::createFastISel(MachineFunction &mf,
2103 MachineModuleInfo *mmo,
2105 DenseMap<const Value *, unsigned> &vm,
2106 DenseMap<const BasicBlock *,
2107 MachineBasicBlock *> &bm,
2108 DenseMap<const AllocaInst *, int> &am
2110 , SmallSet<Instruction*, 8> &cil
2113 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2121 //===----------------------------------------------------------------------===//
2122 // Other Lowering Hooks
2123 //===----------------------------------------------------------------------===//
2126 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2127 MachineFunction &MF = DAG.getMachineFunction();
2128 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2129 int ReturnAddrIndex = FuncInfo->getRAIndex();
2131 if (ReturnAddrIndex == 0) {
2132 // Set up a frame object for the return address.
2133 uint64_t SlotSize = TD->getPointerSize();
2134 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2135 FuncInfo->setRAIndex(ReturnAddrIndex);
2138 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2142 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2143 /// specific condition code, returning the condition code and the LHS/RHS of the
2144 /// comparison to make.
2145 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2146 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2148 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2149 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2150 // X > -1 -> X == 0, jump !sign.
2151 RHS = DAG.getConstant(0, RHS.getValueType());
2152 return X86::COND_NS;
2153 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2154 // X < 0 -> X == 0, jump on sign.
2156 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2158 RHS = DAG.getConstant(0, RHS.getValueType());
2159 return X86::COND_LE;
2163 switch (SetCCOpcode) {
2164 default: assert(0 && "Invalid integer condition!");
2165 case ISD::SETEQ: return X86::COND_E;
2166 case ISD::SETGT: return X86::COND_G;
2167 case ISD::SETGE: return X86::COND_GE;
2168 case ISD::SETLT: return X86::COND_L;
2169 case ISD::SETLE: return X86::COND_LE;
2170 case ISD::SETNE: return X86::COND_NE;
2171 case ISD::SETULT: return X86::COND_B;
2172 case ISD::SETUGT: return X86::COND_A;
2173 case ISD::SETULE: return X86::COND_BE;
2174 case ISD::SETUGE: return X86::COND_AE;
2178 // First determine if it is required or is profitable to flip the operands.
2180 // If LHS is a foldable load, but RHS is not, flip the condition.
2181 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2182 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2183 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2184 std::swap(LHS, RHS);
2187 switch (SetCCOpcode) {
2193 std::swap(LHS, RHS);
2197 // On a floating point condition, the flags are set as follows:
2199 // 0 | 0 | 0 | X > Y
2200 // 0 | 0 | 1 | X < Y
2201 // 1 | 0 | 0 | X == Y
2202 // 1 | 1 | 1 | unordered
2203 switch (SetCCOpcode) {
2204 default: assert(0 && "Condcode should be pre-legalized away");
2206 case ISD::SETEQ: return X86::COND_E;
2207 case ISD::SETOLT: // flipped
2209 case ISD::SETGT: return X86::COND_A;
2210 case ISD::SETOLE: // flipped
2212 case ISD::SETGE: return X86::COND_AE;
2213 case ISD::SETUGT: // flipped
2215 case ISD::SETLT: return X86::COND_B;
2216 case ISD::SETUGE: // flipped
2218 case ISD::SETLE: return X86::COND_BE;
2220 case ISD::SETNE: return X86::COND_NE;
2221 case ISD::SETUO: return X86::COND_P;
2222 case ISD::SETO: return X86::COND_NP;
2226 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2227 /// code. Current x86 isa includes the following FP cmov instructions:
2228 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2229 static bool hasFPCMov(unsigned X86CC) {
2245 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2246 /// the specified range (L, H].
2247 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2248 return (Val < 0) || (Val >= Low && Val < Hi);
2251 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2252 /// specified value.
2253 static bool isUndefOrEqual(int Val, int CmpVal) {
2254 if (Val < 0 || Val == CmpVal)
2259 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2260 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2261 /// the second operand.
2262 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2263 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2264 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2265 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2266 return (Mask[0] < 2 && Mask[1] < 2);
2270 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2271 SmallVector<int, 8> M;
2273 return ::isPSHUFDMask(M, N->getValueType(0));
2276 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2277 /// is suitable for input to PSHUFHW.
2278 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2279 if (VT != MVT::v8i16)
2282 // Lower quadword copied in order or undef.
2283 for (int i = 0; i != 4; ++i)
2284 if (Mask[i] >= 0 && Mask[i] != i)
2287 // Upper quadword shuffled.
2288 for (int i = 4; i != 8; ++i)
2289 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2295 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2296 SmallVector<int, 8> M;
2298 return ::isPSHUFHWMask(M, N->getValueType(0));
2301 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2302 /// is suitable for input to PSHUFLW.
2303 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2304 if (VT != MVT::v8i16)
2307 // Upper quadword copied in order.
2308 for (int i = 4; i != 8; ++i)
2309 if (Mask[i] >= 0 && Mask[i] != i)
2312 // Lower quadword shuffled.
2313 for (int i = 0; i != 4; ++i)
2320 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2321 SmallVector<int, 8> M;
2323 return ::isPSHUFLWMask(M, N->getValueType(0));
2326 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2327 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2328 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2329 int NumElems = VT.getVectorNumElements();
2330 if (NumElems != 2 && NumElems != 4)
2333 int Half = NumElems / 2;
2334 for (int i = 0; i < Half; ++i)
2335 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2337 for (int i = Half; i < NumElems; ++i)
2338 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2344 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2345 SmallVector<int, 8> M;
2347 return ::isSHUFPMask(M, N->getValueType(0));
2350 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2351 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2352 /// half elements to come from vector 1 (which would equal the dest.) and
2353 /// the upper half to come from vector 2.
2354 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2355 int NumElems = VT.getVectorNumElements();
2357 if (NumElems != 2 && NumElems != 4)
2360 int Half = NumElems / 2;
2361 for (int i = 0; i < Half; ++i)
2362 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2364 for (int i = Half; i < NumElems; ++i)
2365 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2370 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2371 SmallVector<int, 8> M;
2373 return isCommutedSHUFPMask(M, N->getValueType(0));
2376 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2377 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2378 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2379 if (N->getValueType(0).getVectorNumElements() != 4)
2382 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2383 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2384 isUndefOrEqual(N->getMaskElt(1), 7) &&
2385 isUndefOrEqual(N->getMaskElt(2), 2) &&
2386 isUndefOrEqual(N->getMaskElt(3), 3);
2389 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2390 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2391 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2392 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2394 if (NumElems != 2 && NumElems != 4)
2397 for (unsigned i = 0; i < NumElems/2; ++i)
2398 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2401 for (unsigned i = NumElems/2; i < NumElems; ++i)
2402 if (!isUndefOrEqual(N->getMaskElt(i), i))
2408 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2409 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2411 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2412 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2414 if (NumElems != 2 && NumElems != 4)
2417 for (unsigned i = 0; i < NumElems/2; ++i)
2418 if (!isUndefOrEqual(N->getMaskElt(i), i))
2421 for (unsigned i = 0; i < NumElems/2; ++i)
2422 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2428 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2429 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2431 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2432 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2437 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2438 isUndefOrEqual(N->getMaskElt(1), 3) &&
2439 isUndefOrEqual(N->getMaskElt(2), 2) &&
2440 isUndefOrEqual(N->getMaskElt(3), 3);
2443 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2444 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2445 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2446 bool V2IsSplat = false) {
2447 int NumElts = VT.getVectorNumElements();
2448 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2451 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2453 int BitI1 = Mask[i+1];
2454 if (!isUndefOrEqual(BitI, j))
2457 if (!isUndefOrEqual(BitI1, NumElts))
2460 if (!isUndefOrEqual(BitI1, j + NumElts))
2467 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2468 SmallVector<int, 8> M;
2470 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2473 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2474 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2475 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2476 bool V2IsSplat = false) {
2477 int NumElts = VT.getVectorNumElements();
2478 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2481 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2483 int BitI1 = Mask[i+1];
2484 if (!isUndefOrEqual(BitI, j + NumElts/2))
2487 if (isUndefOrEqual(BitI1, NumElts))
2490 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2497 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2498 SmallVector<int, 8> M;
2500 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2503 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2504 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2506 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2507 int NumElems = VT.getVectorNumElements();
2508 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2511 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2513 int BitI1 = Mask[i+1];
2514 if (!isUndefOrEqual(BitI, j))
2516 if (!isUndefOrEqual(BitI1, j))
2522 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2523 SmallVector<int, 8> M;
2525 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2528 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2529 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2531 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2532 int NumElems = VT.getVectorNumElements();
2533 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2536 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2538 int BitI1 = Mask[i+1];
2539 if (!isUndefOrEqual(BitI, j))
2541 if (!isUndefOrEqual(BitI1, j))
2547 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2548 SmallVector<int, 8> M;
2550 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2553 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2554 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2555 /// MOVSD, and MOVD, i.e. setting the lowest element.
2556 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2557 if (VT.getVectorElementType().getSizeInBits() < 32)
2560 int NumElts = VT.getVectorNumElements();
2562 if (!isUndefOrEqual(Mask[0], NumElts))
2565 for (int i = 1; i < NumElts; ++i)
2566 if (!isUndefOrEqual(Mask[i], i))
2572 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2573 SmallVector<int, 8> M;
2575 return ::isMOVLMask(M, N->getValueType(0));
2578 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2579 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2580 /// element of vector 2 and the other elements to come from vector 1 in order.
2581 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2582 bool V2IsSplat = false, bool V2IsUndef = false) {
2583 int NumOps = VT.getVectorNumElements();
2584 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2587 if (!isUndefOrEqual(Mask[0], 0))
2590 for (int i = 1; i < NumOps; ++i)
2591 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2592 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2593 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2599 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2600 bool V2IsUndef = false) {
2601 SmallVector<int, 8> M;
2603 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2606 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2607 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2608 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2609 if (N->getValueType(0).getVectorNumElements() != 4)
2612 // Expect 1, 1, 3, 3
2613 for (unsigned i = 0; i < 2; ++i) {
2614 int Elt = N->getMaskElt(i);
2615 if (Elt >= 0 && Elt != 1)
2620 for (unsigned i = 2; i < 4; ++i) {
2621 int Elt = N->getMaskElt(i);
2622 if (Elt >= 0 && Elt != 3)
2627 // Don't use movshdup if it can be done with a shufps.
2628 // FIXME: verify that matching u, u, 3, 3 is what we want.
2632 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2633 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2634 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2635 if (N->getValueType(0).getVectorNumElements() != 4)
2638 // Expect 0, 0, 2, 2
2639 for (unsigned i = 0; i < 2; ++i)
2640 if (N->getMaskElt(i) > 0)
2644 for (unsigned i = 2; i < 4; ++i) {
2645 int Elt = N->getMaskElt(i);
2646 if (Elt >= 0 && Elt != 2)
2651 // Don't use movsldup if it can be done with a shufps.
2655 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2656 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2657 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2658 int e = N->getValueType(0).getVectorNumElements() / 2;
2660 for (int i = 0; i < e; ++i)
2661 if (!isUndefOrEqual(N->getMaskElt(i), i))
2663 for (int i = 0; i < e; ++i)
2664 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2669 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2670 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2672 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2673 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2674 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2676 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2678 for (int i = 0; i < NumOperands; ++i) {
2679 int Val = SVOp->getMaskElt(NumOperands-i-1);
2680 if (Val < 0) Val = 0;
2681 if (Val >= NumOperands) Val -= NumOperands;
2683 if (i != NumOperands - 1)
2689 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2690 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2692 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2693 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2695 // 8 nodes, but we only care about the last 4.
2696 for (unsigned i = 7; i >= 4; --i) {
2697 int Val = SVOp->getMaskElt(i);
2706 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2707 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2709 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2710 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2712 // 8 nodes, but we only care about the first 4.
2713 for (int i = 3; i >= 0; --i) {
2714 int Val = SVOp->getMaskElt(i);
2723 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2724 /// their permute mask.
2725 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2726 SelectionDAG &DAG) {
2727 MVT VT = SVOp->getValueType(0);
2728 unsigned NumElems = VT.getVectorNumElements();
2729 SmallVector<int, 8> MaskVec;
2731 for (unsigned i = 0; i != NumElems; ++i) {
2732 int idx = SVOp->getMaskElt(i);
2734 MaskVec.push_back(idx);
2735 else if (idx < (int)NumElems)
2736 MaskVec.push_back(idx + NumElems);
2738 MaskVec.push_back(idx - NumElems);
2740 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2741 SVOp->getOperand(0), &MaskVec[0]);
2744 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2745 /// the two vector operands have swapped position.
2746 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2747 unsigned NumElems = VT.getVectorNumElements();
2748 for (unsigned i = 0; i != NumElems; ++i) {
2752 else if (idx < (int)NumElems)
2753 Mask[i] = idx + NumElems;
2755 Mask[i] = idx - NumElems;
2759 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2760 /// match movhlps. The lower half elements should come from upper half of
2761 /// V1 (and in order), and the upper half elements should come from the upper
2762 /// half of V2 (and in order).
2763 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2764 if (Op->getValueType(0).getVectorNumElements() != 4)
2766 for (unsigned i = 0, e = 2; i != e; ++i)
2767 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2769 for (unsigned i = 2; i != 4; ++i)
2770 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2775 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2776 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2778 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2779 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2781 N = N->getOperand(0).getNode();
2782 if (!ISD::isNON_EXTLoad(N))
2785 *LD = cast<LoadSDNode>(N);
2789 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2790 /// match movlp{s|d}. The lower half elements should come from lower half of
2791 /// V1 (and in order), and the upper half elements should come from the upper
2792 /// half of V2 (and in order). And since V1 will become the source of the
2793 /// MOVLP, it must be either a vector load or a scalar load to vector.
2794 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2795 ShuffleVectorSDNode *Op) {
2796 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2798 // Is V2 is a vector load, don't do this transformation. We will try to use
2799 // load folding shufps op.
2800 if (ISD::isNON_EXTLoad(V2))
2803 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2805 if (NumElems != 2 && NumElems != 4)
2807 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2808 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2810 for (unsigned i = NumElems/2; i != NumElems; ++i)
2811 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2816 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2818 static bool isSplatVector(SDNode *N) {
2819 if (N->getOpcode() != ISD::BUILD_VECTOR)
2822 SDValue SplatValue = N->getOperand(0);
2823 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2824 if (N->getOperand(i) != SplatValue)
2829 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2831 static inline bool isZeroNode(SDValue Elt) {
2832 return ((isa<ConstantSDNode>(Elt) &&
2833 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2834 (isa<ConstantFPSDNode>(Elt) &&
2835 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2838 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2839 /// to an zero vector.
2840 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2841 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2842 SDValue V1 = N->getOperand(0);
2843 SDValue V2 = N->getOperand(1);
2844 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2845 for (unsigned i = 0; i != NumElems; ++i) {
2846 int Idx = N->getMaskElt(i);
2847 if (Idx >= (int)NumElems) {
2848 unsigned Opc = V2.getOpcode();
2849 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2851 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2853 } else if (Idx >= 0) {
2854 unsigned Opc = V1.getOpcode();
2855 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2857 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2864 /// getZeroVector - Returns a vector of specified type with all zero elements.
2866 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2868 assert(VT.isVector() && "Expected a vector type");
2870 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2871 // type. This ensures they get CSE'd.
2873 if (VT.getSizeInBits() == 64) { // MMX
2874 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2875 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2876 } else if (HasSSE2) { // SSE2
2877 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2878 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2880 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2881 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2883 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2886 /// getOnesVector - Returns a vector of specified type with all bits set.
2888 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2889 assert(VT.isVector() && "Expected a vector type");
2891 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2892 // type. This ensures they get CSE'd.
2893 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2895 if (VT.getSizeInBits() == 64) // MMX
2896 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2898 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2899 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2903 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2904 /// that point to V2 points to its first element.
2905 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2906 MVT VT = SVOp->getValueType(0);
2907 unsigned NumElems = VT.getVectorNumElements();
2909 bool Changed = false;
2910 SmallVector<int, 8> MaskVec;
2911 SVOp->getMask(MaskVec);
2913 for (unsigned i = 0; i != NumElems; ++i) {
2914 if (MaskVec[i] > (int)NumElems) {
2915 MaskVec[i] = NumElems;
2920 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2921 SVOp->getOperand(1), &MaskVec[0]);
2922 return SDValue(SVOp, 0);
2925 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2926 /// operation of specified width.
2927 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2929 unsigned NumElems = VT.getVectorNumElements();
2930 SmallVector<int, 8> Mask;
2931 Mask.push_back(NumElems);
2932 for (unsigned i = 1; i != NumElems; ++i)
2934 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2937 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2938 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2940 unsigned NumElems = VT.getVectorNumElements();
2941 SmallVector<int, 8> Mask;
2942 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2944 Mask.push_back(i + NumElems);
2946 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2949 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2950 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2952 unsigned NumElems = VT.getVectorNumElements();
2953 unsigned Half = NumElems/2;
2954 SmallVector<int, 8> Mask;
2955 for (unsigned i = 0; i != Half; ++i) {
2956 Mask.push_back(i + Half);
2957 Mask.push_back(i + NumElems + Half);
2959 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2962 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2963 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2965 if (SV->getValueType(0).getVectorNumElements() <= 4)
2966 return SDValue(SV, 0);
2968 MVT PVT = MVT::v4f32;
2969 MVT VT = SV->getValueType(0);
2970 DebugLoc dl = SV->getDebugLoc();
2971 SDValue V1 = SV->getOperand(0);
2972 int NumElems = VT.getVectorNumElements();
2973 int EltNo = SV->getSplatIndex();
2975 // unpack elements to the correct location
2976 while (NumElems > 4) {
2977 if (EltNo < NumElems/2) {
2978 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2980 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2981 EltNo -= NumElems/2;
2986 // Perform the splat.
2987 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2988 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2989 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2990 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2993 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2994 /// vector of zero or undef vector. This produces a shuffle where the low
2995 /// element of V2 is swizzled into the zero/undef vector, landing at element
2996 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2997 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2998 bool isZero, bool HasSSE2,
2999 SelectionDAG &DAG) {
3000 MVT VT = V2.getValueType();
3002 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3003 unsigned NumElems = VT.getVectorNumElements();
3004 SmallVector<int, 16> MaskVec;
3005 for (unsigned i = 0; i != NumElems; ++i)
3006 // If this is the insertion idx, put the low elt of V2 here.
3007 MaskVec.push_back(i == Idx ? NumElems : i);
3008 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3011 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3012 /// a shuffle that is zero.
3014 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3015 bool Low, SelectionDAG &DAG) {
3016 unsigned NumZeros = 0;
3017 for (int i = 0; i < NumElems; ++i) {
3018 unsigned Index = Low ? i : NumElems-i-1;
3019 int Idx = SVOp->getMaskElt(Index);
3024 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3025 if (Elt.getNode() && isZeroNode(Elt))
3033 /// isVectorShift - Returns true if the shuffle can be implemented as a
3034 /// logical left or right shift of a vector.
3035 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3036 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3037 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3038 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3041 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3044 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3048 bool SeenV1 = false;
3049 bool SeenV2 = false;
3050 for (int i = NumZeros; i < NumElems; ++i) {
3051 int Val = isLeft ? (i - NumZeros) : i;
3052 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3064 if (SeenV1 && SeenV2)
3067 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3073 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3075 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3076 unsigned NumNonZero, unsigned NumZero,
3077 SelectionDAG &DAG, TargetLowering &TLI) {
3081 DebugLoc dl = Op.getDebugLoc();
3084 for (unsigned i = 0; i < 16; ++i) {
3085 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3086 if (ThisIsNonZero && First) {
3088 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3090 V = DAG.getUNDEF(MVT::v8i16);
3095 SDValue ThisElt(0, 0), LastElt(0, 0);
3096 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3097 if (LastIsNonZero) {
3098 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3099 MVT::i16, Op.getOperand(i-1));
3101 if (ThisIsNonZero) {
3102 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3103 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3104 ThisElt, DAG.getConstant(8, MVT::i8));
3106 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3110 if (ThisElt.getNode())
3111 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3112 DAG.getIntPtrConstant(i/2));
3116 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3119 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3121 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3122 unsigned NumNonZero, unsigned NumZero,
3123 SelectionDAG &DAG, TargetLowering &TLI) {
3127 DebugLoc dl = Op.getDebugLoc();
3130 for (unsigned i = 0; i < 8; ++i) {
3131 bool isNonZero = (NonZeros & (1 << i)) != 0;
3135 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3137 V = DAG.getUNDEF(MVT::v8i16);
3140 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3141 MVT::v8i16, V, Op.getOperand(i),
3142 DAG.getIntPtrConstant(i));
3149 /// getVShift - Return a vector logical shift node.
3151 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3152 unsigned NumBits, SelectionDAG &DAG,
3153 const TargetLowering &TLI, DebugLoc dl) {
3154 bool isMMX = VT.getSizeInBits() == 64;
3155 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3156 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3157 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3158 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3159 DAG.getNode(Opc, dl, ShVT, SrcOp,
3160 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3164 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3165 DebugLoc dl = Op.getDebugLoc();
3166 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3167 if (ISD::isBuildVectorAllZeros(Op.getNode())
3168 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3169 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3170 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3171 // eliminated on x86-32 hosts.
3172 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3175 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3176 return getOnesVector(Op.getValueType(), DAG, dl);
3177 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3180 MVT VT = Op.getValueType();
3181 MVT EVT = VT.getVectorElementType();
3182 unsigned EVTBits = EVT.getSizeInBits();
3184 unsigned NumElems = Op.getNumOperands();
3185 unsigned NumZero = 0;
3186 unsigned NumNonZero = 0;
3187 unsigned NonZeros = 0;
3188 bool IsAllConstants = true;
3189 SmallSet<SDValue, 8> Values;
3190 for (unsigned i = 0; i < NumElems; ++i) {
3191 SDValue Elt = Op.getOperand(i);
3192 if (Elt.getOpcode() == ISD::UNDEF)
3195 if (Elt.getOpcode() != ISD::Constant &&
3196 Elt.getOpcode() != ISD::ConstantFP)
3197 IsAllConstants = false;
3198 if (isZeroNode(Elt))
3201 NonZeros |= (1 << i);
3206 if (NumNonZero == 0) {
3207 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3208 return DAG.getUNDEF(VT);
3211 // Special case for single non-zero, non-undef, element.
3212 if (NumNonZero == 1) {
3213 unsigned Idx = CountTrailingZeros_32(NonZeros);
3214 SDValue Item = Op.getOperand(Idx);
3216 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3217 // the value are obviously zero, truncate the value to i32 and do the
3218 // insertion that way. Only do this if the value is non-constant or if the
3219 // value is a constant being inserted into element 0. It is cheaper to do
3220 // a constant pool load than it is to do a movd + shuffle.
3221 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3222 (!IsAllConstants || Idx == 0)) {
3223 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3224 // Handle MMX and SSE both.
3225 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3226 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3228 // Truncate the value (which may itself be a constant) to i32, and
3229 // convert it to a vector with movd (S2V+shuffle to zero extend).
3230 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3231 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3232 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3233 Subtarget->hasSSE2(), DAG);
3235 // Now we have our 32-bit value zero extended in the low element of
3236 // a vector. If Idx != 0, swizzle it into place.
3238 SmallVector<int, 4> Mask;
3239 Mask.push_back(Idx);
3240 for (unsigned i = 1; i != VecElts; ++i)
3242 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3243 DAG.getUNDEF(Item.getValueType()),
3246 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3250 // If we have a constant or non-constant insertion into the low element of
3251 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3252 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3253 // depending on what the source datatype is.
3256 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3257 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3258 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3259 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3260 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3261 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3263 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3264 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3265 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3266 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3267 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3268 Subtarget->hasSSE2(), DAG);
3269 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3273 // Is it a vector logical left shift?
3274 if (NumElems == 2 && Idx == 1 &&
3275 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3276 unsigned NumBits = VT.getSizeInBits();
3277 return getVShift(true, VT,
3278 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3279 VT, Op.getOperand(1)),
3280 NumBits/2, DAG, *this, dl);
3283 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3286 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3287 // is a non-constant being inserted into an element other than the low one,
3288 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3289 // movd/movss) to move this into the low element, then shuffle it into
3291 if (EVTBits == 32) {
3292 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3294 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3295 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3296 Subtarget->hasSSE2(), DAG);
3297 SmallVector<int, 8> MaskVec;
3298 for (unsigned i = 0; i < NumElems; i++)
3299 MaskVec.push_back(i == Idx ? 0 : 1);
3300 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3304 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3305 if (Values.size() == 1)
3308 // A vector full of immediates; various special cases are already
3309 // handled, so this is best done with a single constant-pool load.
3313 // Let legalizer expand 2-wide build_vectors.
3314 if (EVTBits == 64) {
3315 if (NumNonZero == 1) {
3316 // One half is zero or undef.
3317 unsigned Idx = CountTrailingZeros_32(NonZeros);
3318 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3319 Op.getOperand(Idx));
3320 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3321 Subtarget->hasSSE2(), DAG);
3326 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3327 if (EVTBits == 8 && NumElems == 16) {
3328 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3330 if (V.getNode()) return V;
3333 if (EVTBits == 16 && NumElems == 8) {
3334 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3336 if (V.getNode()) return V;
3339 // If element VT is == 32 bits, turn it into a number of shuffles.
3340 SmallVector<SDValue, 8> V;
3342 if (NumElems == 4 && NumZero > 0) {
3343 for (unsigned i = 0; i < 4; ++i) {
3344 bool isZero = !(NonZeros & (1 << i));
3346 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3348 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3351 for (unsigned i = 0; i < 2; ++i) {
3352 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3355 V[i] = V[i*2]; // Must be a zero vector.
3358 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3361 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3364 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3369 SmallVector<int, 8> MaskVec;
3370 bool Reverse = (NonZeros & 0x3) == 2;
3371 for (unsigned i = 0; i < 2; ++i)
3372 MaskVec.push_back(Reverse ? 1-i : i);
3373 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3374 for (unsigned i = 0; i < 2; ++i)
3375 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3376 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3379 if (Values.size() > 2) {
3380 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3381 // values to be inserted is equal to the number of elements, in which case
3382 // use the unpack code below in the hopes of matching the consecutive elts
3383 // load merge pattern for shuffles.
3384 // FIXME: We could probably just check that here directly.
3385 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3386 getSubtarget()->hasSSE41()) {
3387 V[0] = DAG.getUNDEF(VT);
3388 for (unsigned i = 0; i < NumElems; ++i)
3389 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3390 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3391 Op.getOperand(i), DAG.getIntPtrConstant(i));
3394 // Expand into a number of unpckl*.
3396 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3397 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3398 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3399 for (unsigned i = 0; i < NumElems; ++i)
3400 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3402 while (NumElems != 0) {
3403 for (unsigned i = 0; i < NumElems; ++i)
3404 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3413 // v8i16 shuffles - Prefer shuffles in the following order:
3414 // 1. [all] pshuflw, pshufhw, optional move
3415 // 2. [ssse3] 1 x pshufb
3416 // 3. [ssse3] 2 x pshufb + 1 x por
3417 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3419 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3420 SelectionDAG &DAG, X86TargetLowering &TLI) {
3421 SDValue V1 = SVOp->getOperand(0);
3422 SDValue V2 = SVOp->getOperand(1);
3423 DebugLoc dl = SVOp->getDebugLoc();
3424 SmallVector<int, 8> MaskVals;
3426 // Determine if more than 1 of the words in each of the low and high quadwords
3427 // of the result come from the same quadword of one of the two inputs. Undef
3428 // mask values count as coming from any quadword, for better codegen.
3429 SmallVector<unsigned, 4> LoQuad(4);
3430 SmallVector<unsigned, 4> HiQuad(4);
3431 BitVector InputQuads(4);
3432 for (unsigned i = 0; i < 8; ++i) {
3433 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3434 int EltIdx = SVOp->getMaskElt(i);
3435 MaskVals.push_back(EltIdx);
3444 InputQuads.set(EltIdx / 4);
3447 int BestLoQuad = -1;
3448 unsigned MaxQuad = 1;
3449 for (unsigned i = 0; i < 4; ++i) {
3450 if (LoQuad[i] > MaxQuad) {
3452 MaxQuad = LoQuad[i];
3456 int BestHiQuad = -1;
3458 for (unsigned i = 0; i < 4; ++i) {
3459 if (HiQuad[i] > MaxQuad) {
3461 MaxQuad = HiQuad[i];
3465 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3466 // of the two input vectors, shuffle them into one input vector so only a
3467 // single pshufb instruction is necessary. If There are more than 2 input
3468 // quads, disable the next transformation since it does not help SSSE3.
3469 bool V1Used = InputQuads[0] || InputQuads[1];
3470 bool V2Used = InputQuads[2] || InputQuads[3];
3471 if (TLI.getSubtarget()->hasSSSE3()) {
3472 if (InputQuads.count() == 2 && V1Used && V2Used) {
3473 BestLoQuad = InputQuads.find_first();
3474 BestHiQuad = InputQuads.find_next(BestLoQuad);
3476 if (InputQuads.count() > 2) {
3482 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3483 // the shuffle mask. If a quad is scored as -1, that means that it contains
3484 // words from all 4 input quadwords.
3486 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3487 SmallVector<int, 8> MaskV;
3488 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3489 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3490 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3491 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3492 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3493 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3495 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3496 // source words for the shuffle, to aid later transformations.
3497 bool AllWordsInNewV = true;
3498 bool InOrder[2] = { true, true };
3499 for (unsigned i = 0; i != 8; ++i) {
3500 int idx = MaskVals[i];
3502 InOrder[i/4] = false;
3503 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3505 AllWordsInNewV = false;
3509 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3510 if (AllWordsInNewV) {
3511 for (int i = 0; i != 8; ++i) {
3512 int idx = MaskVals[i];
3515 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3516 if ((idx != i) && idx < 4)
3518 if ((idx != i) && idx > 3)
3527 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3528 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3529 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3530 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3531 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3535 // If we have SSSE3, and all words of the result are from 1 input vector,
3536 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3537 // is present, fall back to case 4.
3538 if (TLI.getSubtarget()->hasSSSE3()) {
3539 SmallVector<SDValue,16> pshufbMask;
3541 // If we have elements from both input vectors, set the high bit of the
3542 // shuffle mask element to zero out elements that come from V2 in the V1
3543 // mask, and elements that come from V1 in the V2 mask, so that the two
3544 // results can be OR'd together.
3545 bool TwoInputs = V1Used && V2Used;
3546 for (unsigned i = 0; i != 8; ++i) {
3547 int EltIdx = MaskVals[i] * 2;
3548 if (TwoInputs && (EltIdx >= 16)) {
3549 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3550 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3553 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3554 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3556 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3557 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3558 DAG.getNode(ISD::BUILD_VECTOR, dl,
3559 MVT::v16i8, &pshufbMask[0], 16));
3561 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3563 // Calculate the shuffle mask for the second input, shuffle it, and
3564 // OR it with the first shuffled input.
3566 for (unsigned i = 0; i != 8; ++i) {
3567 int EltIdx = MaskVals[i] * 2;
3569 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3570 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3573 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3574 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3576 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3577 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3578 DAG.getNode(ISD::BUILD_VECTOR, dl,
3579 MVT::v16i8, &pshufbMask[0], 16));
3580 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3581 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3584 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3585 // and update MaskVals with new element order.
3586 BitVector InOrder(8);
3587 if (BestLoQuad >= 0) {
3588 SmallVector<int, 8> MaskV;
3589 for (int i = 0; i != 4; ++i) {
3590 int idx = MaskVals[i];
3592 MaskV.push_back(-1);
3594 } else if ((idx / 4) == BestLoQuad) {
3595 MaskV.push_back(idx & 3);
3598 MaskV.push_back(-1);
3601 for (unsigned i = 4; i != 8; ++i)
3603 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3607 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3608 // and update MaskVals with the new element order.
3609 if (BestHiQuad >= 0) {
3610 SmallVector<int, 8> MaskV;
3611 for (unsigned i = 0; i != 4; ++i)
3613 for (unsigned i = 4; i != 8; ++i) {
3614 int idx = MaskVals[i];
3616 MaskV.push_back(-1);
3618 } else if ((idx / 4) == BestHiQuad) {
3619 MaskV.push_back((idx & 3) + 4);
3622 MaskV.push_back(-1);
3625 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3629 // In case BestHi & BestLo were both -1, which means each quadword has a word
3630 // from each of the four input quadwords, calculate the InOrder bitvector now
3631 // before falling through to the insert/extract cleanup.
3632 if (BestLoQuad == -1 && BestHiQuad == -1) {
3634 for (int i = 0; i != 8; ++i)
3635 if (MaskVals[i] < 0 || MaskVals[i] == i)
3639 // The other elements are put in the right place using pextrw and pinsrw.
3640 for (unsigned i = 0; i != 8; ++i) {
3643 int EltIdx = MaskVals[i];
3646 SDValue ExtOp = (EltIdx < 8)
3647 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3648 DAG.getIntPtrConstant(EltIdx))
3649 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3650 DAG.getIntPtrConstant(EltIdx - 8));
3651 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3652 DAG.getIntPtrConstant(i));
3657 // v16i8 shuffles - Prefer shuffles in the following order:
3658 // 1. [ssse3] 1 x pshufb
3659 // 2. [ssse3] 2 x pshufb + 1 x por
3660 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3662 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3663 SelectionDAG &DAG, X86TargetLowering &TLI) {
3664 SDValue V1 = SVOp->getOperand(0);
3665 SDValue V2 = SVOp->getOperand(1);
3666 DebugLoc dl = SVOp->getDebugLoc();
3667 SmallVector<int, 16> MaskVals;
3668 SVOp->getMask(MaskVals);
3670 // If we have SSSE3, case 1 is generated when all result bytes come from
3671 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3672 // present, fall back to case 3.
3673 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3676 for (unsigned i = 0; i < 16; ++i) {
3677 int EltIdx = MaskVals[i];
3686 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3687 if (TLI.getSubtarget()->hasSSSE3()) {
3688 SmallVector<SDValue,16> pshufbMask;
3690 // If all result elements are from one input vector, then only translate
3691 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3693 // Otherwise, we have elements from both input vectors, and must zero out
3694 // elements that come from V2 in the first mask, and V1 in the second mask
3695 // so that we can OR them together.
3696 bool TwoInputs = !(V1Only || V2Only);
3697 for (unsigned i = 0; i != 16; ++i) {
3698 int EltIdx = MaskVals[i];
3699 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3700 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3703 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3705 // If all the elements are from V2, assign it to V1 and return after
3706 // building the first pshufb.
3709 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3710 DAG.getNode(ISD::BUILD_VECTOR, dl,
3711 MVT::v16i8, &pshufbMask[0], 16));
3715 // Calculate the shuffle mask for the second input, shuffle it, and
3716 // OR it with the first shuffled input.
3718 for (unsigned i = 0; i != 16; ++i) {
3719 int EltIdx = MaskVals[i];
3721 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3724 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3726 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3727 DAG.getNode(ISD::BUILD_VECTOR, dl,
3728 MVT::v16i8, &pshufbMask[0], 16));
3729 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3732 // No SSSE3 - Calculate in place words and then fix all out of place words
3733 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3734 // the 16 different words that comprise the two doublequadword input vectors.
3735 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3736 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3737 SDValue NewV = V2Only ? V2 : V1;
3738 for (int i = 0; i != 8; ++i) {
3739 int Elt0 = MaskVals[i*2];
3740 int Elt1 = MaskVals[i*2+1];
3742 // This word of the result is all undef, skip it.
3743 if (Elt0 < 0 && Elt1 < 0)
3746 // This word of the result is already in the correct place, skip it.
3747 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3749 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3752 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3753 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3756 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3757 // using a single extract together, load it and store it.
3758 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3759 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3760 DAG.getIntPtrConstant(Elt1 / 2));
3761 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3762 DAG.getIntPtrConstant(i));
3766 // If Elt1 is defined, extract it from the appropriate source. If the
3767 // source byte is not also odd, shift the extracted word left 8 bits
3768 // otherwise clear the bottom 8 bits if we need to do an or.
3770 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3771 DAG.getIntPtrConstant(Elt1 / 2));
3772 if ((Elt1 & 1) == 0)
3773 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3774 DAG.getConstant(8, TLI.getShiftAmountTy()));
3776 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3777 DAG.getConstant(0xFF00, MVT::i16));
3779 // If Elt0 is defined, extract it from the appropriate source. If the
3780 // source byte is not also even, shift the extracted word right 8 bits. If
3781 // Elt1 was also defined, OR the extracted values together before
3782 // inserting them in the result.
3784 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3785 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3786 if ((Elt0 & 1) != 0)
3787 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3788 DAG.getConstant(8, TLI.getShiftAmountTy()));
3790 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3791 DAG.getConstant(0x00FF, MVT::i16));
3792 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3795 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3796 DAG.getIntPtrConstant(i));
3798 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3801 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3802 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3803 /// done when every pair / quad of shuffle mask elements point to elements in
3804 /// the right sequence. e.g.
3805 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3807 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3809 TargetLowering &TLI, DebugLoc dl) {
3810 MVT VT = SVOp->getValueType(0);
3811 SDValue V1 = SVOp->getOperand(0);
3812 SDValue V2 = SVOp->getOperand(1);
3813 unsigned NumElems = VT.getVectorNumElements();
3814 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3815 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3816 MVT MaskEltVT = MaskVT.getVectorElementType();
3818 switch (VT.getSimpleVT()) {
3819 default: assert(false && "Unexpected!");
3820 case MVT::v4f32: NewVT = MVT::v2f64; break;
3821 case MVT::v4i32: NewVT = MVT::v2i64; break;
3822 case MVT::v8i16: NewVT = MVT::v4i32; break;
3823 case MVT::v16i8: NewVT = MVT::v4i32; break;
3826 if (NewWidth == 2) {
3832 int Scale = NumElems / NewWidth;
3833 SmallVector<int, 8> MaskVec;
3834 for (unsigned i = 0; i < NumElems; i += Scale) {
3836 for (int j = 0; j < Scale; ++j) {
3837 int EltIdx = SVOp->getMaskElt(i+j);
3841 StartIdx = EltIdx - (EltIdx % Scale);
3842 if (EltIdx != StartIdx + j)
3846 MaskVec.push_back(-1);
3848 MaskVec.push_back(StartIdx / Scale);
3851 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3852 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3853 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3856 /// getVZextMovL - Return a zero-extending vector move low node.
3858 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3859 SDValue SrcOp, SelectionDAG &DAG,
3860 const X86Subtarget *Subtarget, DebugLoc dl) {
3861 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3862 LoadSDNode *LD = NULL;
3863 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3864 LD = dyn_cast<LoadSDNode>(SrcOp);
3866 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3868 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3869 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3870 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3871 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3872 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3874 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3875 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3876 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3877 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3885 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3886 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3887 DAG.getNode(ISD::BIT_CONVERT, dl,
3891 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3894 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3895 SDValue V1 = SVOp->getOperand(0);
3896 SDValue V2 = SVOp->getOperand(1);
3897 DebugLoc dl = SVOp->getDebugLoc();
3898 MVT VT = SVOp->getValueType(0);
3900 SmallVector<std::pair<int, int>, 8> Locs;
3902 SmallVector<int, 8> Mask1(4U, -1);
3903 SmallVector<int, 8> PermMask;
3904 SVOp->getMask(PermMask);
3908 for (unsigned i = 0; i != 4; ++i) {
3909 int Idx = PermMask[i];
3911 Locs[i] = std::make_pair(-1, -1);
3913 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3915 Locs[i] = std::make_pair(0, NumLo);
3919 Locs[i] = std::make_pair(1, NumHi);
3921 Mask1[2+NumHi] = Idx;
3927 if (NumLo <= 2 && NumHi <= 2) {
3928 // If no more than two elements come from either vector. This can be
3929 // implemented with two shuffles. First shuffle gather the elements.
3930 // The second shuffle, which takes the first shuffle as both of its
3931 // vector operands, put the elements into the right order.
3932 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3934 SmallVector<int, 8> Mask2(4U, -1);
3936 for (unsigned i = 0; i != 4; ++i) {
3937 if (Locs[i].first == -1)
3940 unsigned Idx = (i < 2) ? 0 : 4;
3941 Idx += Locs[i].first * 2 + Locs[i].second;
3946 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3947 } else if (NumLo == 3 || NumHi == 3) {
3948 // Otherwise, we must have three elements from one vector, call it X, and
3949 // one element from the other, call it Y. First, use a shufps to build an
3950 // intermediate vector with the one element from Y and the element from X
3951 // that will be in the same half in the final destination (the indexes don't
3952 // matter). Then, use a shufps to build the final vector, taking the half
3953 // containing the element from Y from the intermediate, and the other half
3956 // Normalize it so the 3 elements come from V1.
3957 CommuteVectorShuffleMask(PermMask, VT);
3961 // Find the element from V2.
3963 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3964 int Val = PermMask[HiIndex];
3971 Mask1[0] = PermMask[HiIndex];
3973 Mask1[2] = PermMask[HiIndex^1];
3975 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3978 Mask1[0] = PermMask[0];
3979 Mask1[1] = PermMask[1];
3980 Mask1[2] = HiIndex & 1 ? 6 : 4;
3981 Mask1[3] = HiIndex & 1 ? 4 : 6;
3982 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3984 Mask1[0] = HiIndex & 1 ? 2 : 0;
3985 Mask1[1] = HiIndex & 1 ? 0 : 2;
3986 Mask1[2] = PermMask[2];
3987 Mask1[3] = PermMask[3];
3992 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3996 // Break it into (shuffle shuffle_hi, shuffle_lo).
3998 SmallVector<int,8> LoMask(4U, -1);
3999 SmallVector<int,8> HiMask(4U, -1);
4001 SmallVector<int,8> *MaskPtr = &LoMask;
4002 unsigned MaskIdx = 0;
4005 for (unsigned i = 0; i != 4; ++i) {
4012 int Idx = PermMask[i];
4014 Locs[i] = std::make_pair(-1, -1);
4015 } else if (Idx < 4) {
4016 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4017 (*MaskPtr)[LoIdx] = Idx;
4020 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4021 (*MaskPtr)[HiIdx] = Idx;
4026 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4027 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4028 SmallVector<int, 8> MaskOps;
4029 for (unsigned i = 0; i != 4; ++i) {
4030 if (Locs[i].first == -1) {
4031 MaskOps.push_back(-1);
4033 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4034 MaskOps.push_back(Idx);
4037 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4041 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4042 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4043 SDValue V1 = Op.getOperand(0);
4044 SDValue V2 = Op.getOperand(1);
4045 MVT VT = Op.getValueType();
4046 DebugLoc dl = Op.getDebugLoc();
4047 unsigned NumElems = VT.getVectorNumElements();
4048 bool isMMX = VT.getSizeInBits() == 64;
4049 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4050 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4051 bool V1IsSplat = false;
4052 bool V2IsSplat = false;
4054 if (isZeroShuffle(SVOp))
4055 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4057 // Promote splats to v4f32.
4058 if (SVOp->isSplat()) {
4059 if (isMMX || NumElems < 4)
4061 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4064 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4066 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4067 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4068 if (NewOp.getNode())
4069 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4070 LowerVECTOR_SHUFFLE(NewOp, DAG));
4071 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4072 // FIXME: Figure out a cleaner way to do this.
4073 // Try to make use of movq to zero out the top part.
4074 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4075 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4076 if (NewOp.getNode()) {
4077 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4078 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4079 DAG, Subtarget, dl);
4081 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4082 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4083 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4084 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4085 DAG, Subtarget, dl);
4089 if (X86::isPSHUFDMask(SVOp))
4092 // Check if this can be converted into a logical shift.
4093 bool isLeft = false;
4096 bool isShift = getSubtarget()->hasSSE2() &&
4097 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4098 if (isShift && ShVal.hasOneUse()) {
4099 // If the shifted value has multiple uses, it may be cheaper to use
4100 // v_set0 + movlhps or movhlps, etc.
4101 MVT EVT = VT.getVectorElementType();
4102 ShAmt *= EVT.getSizeInBits();
4103 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4106 if (X86::isMOVLMask(SVOp)) {
4109 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4110 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4115 // FIXME: fold these into legal mask.
4116 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4117 X86::isMOVSLDUPMask(SVOp) ||
4118 X86::isMOVHLPSMask(SVOp) ||
4119 X86::isMOVHPMask(SVOp) ||
4120 X86::isMOVLPMask(SVOp)))
4123 if (ShouldXformToMOVHLPS(SVOp) ||
4124 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4125 return CommuteVectorShuffle(SVOp, DAG);
4128 // No better options. Use a vshl / vsrl.
4129 MVT EVT = VT.getVectorElementType();
4130 ShAmt *= EVT.getSizeInBits();
4131 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4134 bool Commuted = false;
4135 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4136 // 1,1,1,1 -> v8i16 though.
4137 V1IsSplat = isSplatVector(V1.getNode());
4138 V2IsSplat = isSplatVector(V2.getNode());
4140 // Canonicalize the splat or undef, if present, to be on the RHS.
4141 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4142 Op = CommuteVectorShuffle(SVOp, DAG);
4143 SVOp = cast<ShuffleVectorSDNode>(Op);
4144 V1 = SVOp->getOperand(0);
4145 V2 = SVOp->getOperand(1);
4146 std::swap(V1IsSplat, V2IsSplat);
4147 std::swap(V1IsUndef, V2IsUndef);
4151 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4152 // Shuffling low element of v1 into undef, just return v1.
4155 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4156 // the instruction selector will not match, so get a canonical MOVL with
4157 // swapped operands to undo the commute.
4158 return getMOVL(DAG, dl, VT, V2, V1);
4161 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4162 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4163 X86::isUNPCKLMask(SVOp) ||
4164 X86::isUNPCKHMask(SVOp))
4168 // Normalize mask so all entries that point to V2 points to its first
4169 // element then try to match unpck{h|l} again. If match, return a
4170 // new vector_shuffle with the corrected mask.
4171 SDValue NewMask = NormalizeMask(SVOp, DAG);
4172 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4173 if (NSVOp != SVOp) {
4174 if (X86::isUNPCKLMask(NSVOp, true)) {
4176 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4183 // Commute is back and try unpck* again.
4184 // FIXME: this seems wrong.
4185 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4186 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4187 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4188 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4189 X86::isUNPCKLMask(NewSVOp) ||
4190 X86::isUNPCKHMask(NewSVOp))
4194 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4196 // Normalize the node to match x86 shuffle ops if needed
4197 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4198 return CommuteVectorShuffle(SVOp, DAG);
4200 // Check for legal shuffle and return?
4201 SmallVector<int, 16> PermMask;
4202 SVOp->getMask(PermMask);
4203 if (isShuffleMaskLegal(PermMask, VT))
4206 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4207 if (VT == MVT::v8i16) {
4208 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4209 if (NewOp.getNode())
4213 if (VT == MVT::v16i8) {
4214 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4215 if (NewOp.getNode())
4219 // Handle all 4 wide cases with a number of shuffles except for MMX.
4220 if (NumElems == 4 && !isMMX)
4221 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4227 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4228 SelectionDAG &DAG) {
4229 MVT VT = Op.getValueType();
4230 DebugLoc dl = Op.getDebugLoc();
4231 if (VT.getSizeInBits() == 8) {
4232 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4233 Op.getOperand(0), Op.getOperand(1));
4234 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4235 DAG.getValueType(VT));
4236 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4237 } else if (VT.getSizeInBits() == 16) {
4238 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4239 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4241 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4242 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4243 DAG.getNode(ISD::BIT_CONVERT, dl,
4247 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4248 Op.getOperand(0), Op.getOperand(1));
4249 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4250 DAG.getValueType(VT));
4251 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4252 } else if (VT == MVT::f32) {
4253 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4254 // the result back to FR32 register. It's only worth matching if the
4255 // result has a single use which is a store or a bitcast to i32. And in
4256 // the case of a store, it's not worth it if the index is a constant 0,
4257 // because a MOVSSmr can be used instead, which is smaller and faster.
4258 if (!Op.hasOneUse())
4260 SDNode *User = *Op.getNode()->use_begin();
4261 if ((User->getOpcode() != ISD::STORE ||
4262 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4263 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4264 (User->getOpcode() != ISD::BIT_CONVERT ||
4265 User->getValueType(0) != MVT::i32))
4267 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4268 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4271 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4272 } else if (VT == MVT::i32) {
4273 // ExtractPS works with constant index.
4274 if (isa<ConstantSDNode>(Op.getOperand(1)))
4282 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4283 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4286 if (Subtarget->hasSSE41()) {
4287 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4292 MVT VT = Op.getValueType();
4293 DebugLoc dl = Op.getDebugLoc();
4294 // TODO: handle v16i8.
4295 if (VT.getSizeInBits() == 16) {
4296 SDValue Vec = Op.getOperand(0);
4297 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4299 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4300 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4301 DAG.getNode(ISD::BIT_CONVERT, dl,
4304 // Transform it so it match pextrw which produces a 32-bit result.
4305 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4306 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4307 Op.getOperand(0), Op.getOperand(1));
4308 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4309 DAG.getValueType(VT));
4310 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4311 } else if (VT.getSizeInBits() == 32) {
4312 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4316 // SHUFPS the element to the lowest double word, then movss.
4317 int Mask[4] = { Idx, -1, -1, -1 };
4318 MVT VVT = Op.getOperand(0).getValueType();
4319 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4320 DAG.getUNDEF(VVT), Mask);
4321 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4322 DAG.getIntPtrConstant(0));
4323 } else if (VT.getSizeInBits() == 64) {
4324 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4325 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4326 // to match extract_elt for f64.
4327 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4331 // UNPCKHPD the element to the lowest double word, then movsd.
4332 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4333 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4334 int Mask[2] = { 1, -1 };
4335 MVT VVT = Op.getOperand(0).getValueType();
4336 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4337 DAG.getUNDEF(VVT), Mask);
4338 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4339 DAG.getIntPtrConstant(0));
4346 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4347 MVT VT = Op.getValueType();
4348 MVT EVT = VT.getVectorElementType();
4349 DebugLoc dl = Op.getDebugLoc();
4351 SDValue N0 = Op.getOperand(0);
4352 SDValue N1 = Op.getOperand(1);
4353 SDValue N2 = Op.getOperand(2);
4355 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4356 isa<ConstantSDNode>(N2)) {
4357 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4359 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4361 if (N1.getValueType() != MVT::i32)
4362 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4363 if (N2.getValueType() != MVT::i32)
4364 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4365 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4366 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4367 // Bits [7:6] of the constant are the source select. This will always be
4368 // zero here. The DAG Combiner may combine an extract_elt index into these
4369 // bits. For example (insert (extract, 3), 2) could be matched by putting
4370 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4371 // Bits [5:4] of the constant are the destination select. This is the
4372 // value of the incoming immediate.
4373 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4374 // combine either bitwise AND or insert of float 0.0 to set these bits.
4375 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4376 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4377 } else if (EVT == MVT::i32) {
4378 // InsertPS works with constant index.
4379 if (isa<ConstantSDNode>(N2))
4386 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4387 MVT VT = Op.getValueType();
4388 MVT EVT = VT.getVectorElementType();
4390 if (Subtarget->hasSSE41())
4391 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4396 DebugLoc dl = Op.getDebugLoc();
4397 SDValue N0 = Op.getOperand(0);
4398 SDValue N1 = Op.getOperand(1);
4399 SDValue N2 = Op.getOperand(2);
4401 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4402 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4403 // as its second argument.
4404 if (N1.getValueType() != MVT::i32)
4405 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4406 if (N2.getValueType() != MVT::i32)
4407 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4408 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4414 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4415 DebugLoc dl = Op.getDebugLoc();
4416 if (Op.getValueType() == MVT::v2f32)
4417 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4418 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4419 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4420 Op.getOperand(0))));
4422 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4423 MVT VT = MVT::v2i32;
4424 switch (Op.getValueType().getSimpleVT()) {
4431 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4432 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4435 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4436 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4437 // one of the above mentioned nodes. It has to be wrapped because otherwise
4438 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4439 // be used to form addressing mode. These wrapped nodes will be selected
4442 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4443 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4445 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4447 unsigned char OpFlag = 0;
4448 unsigned WrapperKind = X86ISD::Wrapper;
4449 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4450 if (Subtarget->isPICStyleStub())
4451 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4452 else if (Subtarget->isPICStyleGOT())
4453 OpFlag = X86II::MO_GOTOFF;
4454 else if (Subtarget->isPICStyleRIPRel() &&
4455 getTargetMachine().getCodeModel() == CodeModel::Small)
4456 WrapperKind = X86ISD::WrapperRIP;
4459 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4461 CP->getOffset(), OpFlag);
4462 DebugLoc DL = CP->getDebugLoc();
4463 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4464 // With PIC, the address is actually $g + Offset.
4466 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4467 DAG.getNode(X86ISD::GlobalBaseReg,
4468 DebugLoc::getUnknownLoc(), getPointerTy()),
4475 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4476 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4478 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4480 unsigned char OpFlag = 0;
4481 unsigned WrapperKind = X86ISD::Wrapper;
4482 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4483 if (Subtarget->isPICStyleStub())
4484 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4485 else if (Subtarget->isPICStyleGOT())
4486 OpFlag = X86II::MO_GOTOFF;
4487 else if (Subtarget->isPICStyleRIPRel())
4488 WrapperKind = X86ISD::WrapperRIP;
4491 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4493 DebugLoc DL = JT->getDebugLoc();
4494 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4496 // With PIC, the address is actually $g + Offset.
4498 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4499 DAG.getNode(X86ISD::GlobalBaseReg,
4500 DebugLoc::getUnknownLoc(), getPointerTy()),
4508 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4509 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4511 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4513 unsigned char OpFlag = 0;
4514 unsigned WrapperKind = X86ISD::Wrapper;
4515 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4516 if (Subtarget->isPICStyleStub())
4517 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4518 else if (Subtarget->isPICStyleGOT())
4519 OpFlag = X86II::MO_GOTOFF;
4520 else if (Subtarget->isPICStyleRIPRel())
4521 WrapperKind = X86ISD::WrapperRIP;
4524 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4526 DebugLoc DL = Op.getDebugLoc();
4527 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4530 // With PIC, the address is actually $g + Offset.
4531 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4532 !Subtarget->isPICStyleRIPRel()) {
4533 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4534 DAG.getNode(X86ISD::GlobalBaseReg,
4535 DebugLoc::getUnknownLoc(),
4544 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4546 SelectionDAG &DAG) const {
4547 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4548 bool ExtraLoadRequired =
4549 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4551 // Create the TargetGlobalAddress node, folding in the constant
4552 // offset if it is legal.
4554 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4555 // A direct static reference to a global.
4556 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4559 unsigned char OpFlags = 0;
4561 if (GV->hasDLLImportLinkage())
4562 OpFlags = X86II::MO_DLLIMPORT;
4563 else if (Subtarget->isPICStyleRIPRel() &&
4564 getTargetMachine().getRelocationModel() != Reloc::Static) {
4565 if (ExtraLoadRequired)
4566 OpFlags = X86II::MO_GOTPCREL;
4567 } else if (Subtarget->isPICStyleGOT() &&
4568 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
4569 if (ExtraLoadRequired)
4570 OpFlags = X86II::MO_GOT;
4572 OpFlags = X86II::MO_GOTOFF;
4575 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4578 if (Subtarget->isPICStyleRIPRel() &&
4579 getTargetMachine().getCodeModel() == CodeModel::Small)
4580 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4582 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4584 // With PIC, the address is actually $g + Offset.
4585 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4586 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4587 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4591 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4592 // load the value at address GV, not the value of GV itself. This means that
4593 // the GlobalAddress must be in the base or index register of the address, not
4594 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4595 // The same applies for external symbols during PIC codegen
4596 if (ExtraLoadRequired)
4597 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4598 PseudoSourceValue::getGOT(), 0);
4600 // If there was a non-zero offset that we didn't fold, create an explicit
4603 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4604 DAG.getConstant(Offset, getPointerTy()));
4610 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4611 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4612 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4613 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4617 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4618 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4619 unsigned char OperandFlags) {
4620 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4621 DebugLoc dl = GA->getDebugLoc();
4622 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4623 GA->getValueType(0),
4627 SDValue Ops[] = { Chain, TGA, *InFlag };
4628 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4630 SDValue Ops[] = { Chain, TGA };
4631 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4633 SDValue Flag = Chain.getValue(1);
4634 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4637 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4639 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4642 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4643 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4644 DAG.getNode(X86ISD::GlobalBaseReg,
4645 DebugLoc::getUnknownLoc(),
4647 InFlag = Chain.getValue(1);
4649 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4652 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4654 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4656 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4657 X86::RAX, X86II::MO_TLSGD);
4660 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4661 // "local exec" model.
4662 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4663 const MVT PtrVT, TLSModel::Model model,
4665 DebugLoc dl = GA->getDebugLoc();
4666 // Get the Thread Pointer
4667 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4668 DebugLoc::getUnknownLoc(), PtrVT,
4669 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4672 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4675 unsigned char OperandFlags = 0;
4676 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4678 unsigned WrapperKind = X86ISD::Wrapper;
4679 if (model == TLSModel::LocalExec) {
4680 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4681 } else if (is64Bit) {
4682 assert(model == TLSModel::InitialExec);
4683 OperandFlags = X86II::MO_GOTTPOFF;
4684 WrapperKind = X86ISD::WrapperRIP;
4686 assert(model == TLSModel::InitialExec);
4687 OperandFlags = X86II::MO_INDNTPOFF;
4690 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4692 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4693 GA->getOffset(), OperandFlags);
4694 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4696 if (model == TLSModel::InitialExec)
4697 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4698 PseudoSourceValue::getGOT(), 0);
4700 // The address of the thread local variable is the add of the thread
4701 // pointer with the offset of the variable.
4702 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4706 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4707 // TODO: implement the "local dynamic" model
4708 // TODO: implement the "initial exec"model for pic executables
4709 assert(Subtarget->isTargetELF() &&
4710 "TLS not implemented for non-ELF targets");
4711 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4712 const GlobalValue *GV = GA->getGlobal();
4714 // If GV is an alias then use the aliasee for determining
4715 // thread-localness.
4716 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4717 GV = GA->resolveAliasedGlobal(false);
4719 TLSModel::Model model = getTLSModel(GV,
4720 getTargetMachine().getRelocationModel());
4723 case TLSModel::GeneralDynamic:
4724 case TLSModel::LocalDynamic: // not implemented
4725 if (Subtarget->is64Bit())
4726 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4727 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4729 case TLSModel::InitialExec:
4730 case TLSModel::LocalExec:
4731 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4732 Subtarget->is64Bit());
4735 assert(0 && "Unreachable");
4740 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4741 /// take a 2 x i32 value to shift plus a shift amount.
4742 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4743 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4744 MVT VT = Op.getValueType();
4745 unsigned VTBits = VT.getSizeInBits();
4746 DebugLoc dl = Op.getDebugLoc();
4747 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4748 SDValue ShOpLo = Op.getOperand(0);
4749 SDValue ShOpHi = Op.getOperand(1);
4750 SDValue ShAmt = Op.getOperand(2);
4751 SDValue Tmp1 = isSRA ?
4752 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4753 DAG.getConstant(VTBits - 1, MVT::i8)) :
4754 DAG.getConstant(0, VT);
4757 if (Op.getOpcode() == ISD::SHL_PARTS) {
4758 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4759 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4761 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4762 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4765 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4766 DAG.getConstant(VTBits, MVT::i8));
4767 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4768 AndNode, DAG.getConstant(0, MVT::i8));
4771 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4772 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4773 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4775 if (Op.getOpcode() == ISD::SHL_PARTS) {
4776 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4777 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4779 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4780 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4783 SDValue Ops[2] = { Lo, Hi };
4784 return DAG.getMergeValues(Ops, 2, dl);
4787 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4788 MVT SrcVT = Op.getOperand(0).getValueType();
4790 if (SrcVT.isVector()) {
4791 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4797 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4798 "Unknown SINT_TO_FP to lower!");
4800 // These are really Legal; return the operand so the caller accepts it as
4802 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4804 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4805 Subtarget->is64Bit()) {
4809 DebugLoc dl = Op.getDebugLoc();
4810 unsigned Size = SrcVT.getSizeInBits()/8;
4811 MachineFunction &MF = DAG.getMachineFunction();
4812 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4813 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4814 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4816 PseudoSourceValue::getFixedStack(SSFI), 0);
4817 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4820 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4822 SelectionDAG &DAG) {
4824 DebugLoc dl = Op.getDebugLoc();
4826 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4828 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4830 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4831 SmallVector<SDValue, 8> Ops;
4832 Ops.push_back(Chain);
4833 Ops.push_back(StackSlot);
4834 Ops.push_back(DAG.getValueType(SrcVT));
4835 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4836 Tys, &Ops[0], Ops.size());
4839 Chain = Result.getValue(1);
4840 SDValue InFlag = Result.getValue(2);
4842 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4843 // shouldn't be necessary except that RFP cannot be live across
4844 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4845 MachineFunction &MF = DAG.getMachineFunction();
4846 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4847 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4848 Tys = DAG.getVTList(MVT::Other);
4849 SmallVector<SDValue, 8> Ops;
4850 Ops.push_back(Chain);
4851 Ops.push_back(Result);
4852 Ops.push_back(StackSlot);
4853 Ops.push_back(DAG.getValueType(Op.getValueType()));
4854 Ops.push_back(InFlag);
4855 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4856 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4857 PseudoSourceValue::getFixedStack(SSFI), 0);
4863 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4864 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4865 // This algorithm is not obvious. Here it is in C code, more or less:
4867 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4868 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4869 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4871 // Copy ints to xmm registers.
4872 __m128i xh = _mm_cvtsi32_si128( hi );
4873 __m128i xl = _mm_cvtsi32_si128( lo );
4875 // Combine into low half of a single xmm register.
4876 __m128i x = _mm_unpacklo_epi32( xh, xl );
4880 // Merge in appropriate exponents to give the integer bits the right
4882 x = _mm_unpacklo_epi32( x, exp );
4884 // Subtract away the biases to deal with the IEEE-754 double precision
4886 d = _mm_sub_pd( (__m128d) x, bias );
4888 // All conversions up to here are exact. The correctly rounded result is
4889 // calculated using the current rounding mode using the following
4891 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4892 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4893 // store doesn't really need to be here (except
4894 // maybe to zero the other double)
4899 DebugLoc dl = Op.getDebugLoc();
4901 // Build some magic constants.
4902 std::vector<Constant*> CV0;
4903 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4904 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4905 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4906 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4907 Constant *C0 = ConstantVector::get(CV0);
4908 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4910 std::vector<Constant*> CV1;
4911 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4912 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4913 Constant *C1 = ConstantVector::get(CV1);
4914 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4916 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4917 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4919 DAG.getIntPtrConstant(1)));
4920 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4921 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4923 DAG.getIntPtrConstant(0)));
4924 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4925 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4926 PseudoSourceValue::getConstantPool(), 0,
4928 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4929 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4930 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4931 PseudoSourceValue::getConstantPool(), 0,
4933 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4935 // Add the halves; easiest way is to swap them into another reg first.
4936 int ShufMask[2] = { 1, -1 };
4937 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4938 DAG.getUNDEF(MVT::v2f64), ShufMask);
4939 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4940 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4941 DAG.getIntPtrConstant(0));
4944 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4945 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4946 DebugLoc dl = Op.getDebugLoc();
4947 // FP constant to bias correct the final result.
4948 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4951 // Load the 32-bit value into an XMM register.
4952 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4953 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4955 DAG.getIntPtrConstant(0)));
4957 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4958 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4959 DAG.getIntPtrConstant(0));
4961 // Or the load with the bias.
4962 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4963 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4964 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4966 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4967 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4968 MVT::v2f64, Bias)));
4969 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4970 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4971 DAG.getIntPtrConstant(0));
4973 // Subtract the bias.
4974 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4976 // Handle final rounding.
4977 MVT DestVT = Op.getValueType();
4979 if (DestVT.bitsLT(MVT::f64)) {
4980 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4981 DAG.getIntPtrConstant(0));
4982 } else if (DestVT.bitsGT(MVT::f64)) {
4983 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4986 // Handle final rounding.
4990 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4991 SDValue N0 = Op.getOperand(0);
4992 DebugLoc dl = Op.getDebugLoc();
4994 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4995 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4996 // the optimization here.
4997 if (DAG.SignBitIsZero(N0))
4998 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5000 MVT SrcVT = N0.getValueType();
5001 if (SrcVT == MVT::i64) {
5002 // We only handle SSE2 f64 target here; caller can expand the rest.
5003 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5006 return LowerUINT_TO_FP_i64(Op, DAG);
5007 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5008 return LowerUINT_TO_FP_i32(Op, DAG);
5011 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5013 // Make a 64-bit buffer, and use it to build an FILD.
5014 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5015 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5016 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5017 getPointerTy(), StackSlot, WordOff);
5018 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5019 StackSlot, NULL, 0);
5020 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5021 OffsetSlot, NULL, 0);
5022 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5025 std::pair<SDValue,SDValue> X86TargetLowering::
5026 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5027 DebugLoc dl = Op.getDebugLoc();
5029 MVT DstTy = Op.getValueType();
5032 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5036 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5037 DstTy.getSimpleVT() >= MVT::i16 &&
5038 "Unknown FP_TO_SINT to lower!");
5040 // These are really Legal.
5041 if (DstTy == MVT::i32 &&
5042 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5043 return std::make_pair(SDValue(), SDValue());
5044 if (Subtarget->is64Bit() &&
5045 DstTy == MVT::i64 &&
5046 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5047 return std::make_pair(SDValue(), SDValue());
5049 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5051 MachineFunction &MF = DAG.getMachineFunction();
5052 unsigned MemSize = DstTy.getSizeInBits()/8;
5053 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5054 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5057 switch (DstTy.getSimpleVT()) {
5058 default: assert(0 && "Invalid FP_TO_SINT to lower!");
5059 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5060 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5061 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5064 SDValue Chain = DAG.getEntryNode();
5065 SDValue Value = Op.getOperand(0);
5066 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5067 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5068 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5069 PseudoSourceValue::getFixedStack(SSFI), 0);
5070 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5072 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5074 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5075 Chain = Value.getValue(1);
5076 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5077 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5080 // Build the FP_TO_INT*_IN_MEM
5081 SDValue Ops[] = { Chain, Value, StackSlot };
5082 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5084 return std::make_pair(FIST, StackSlot);
5087 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5088 if (Op.getValueType().isVector()) {
5089 if (Op.getValueType() == MVT::v2i32 &&
5090 Op.getOperand(0).getValueType() == MVT::v2f64) {
5096 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5097 SDValue FIST = Vals.first, StackSlot = Vals.second;
5098 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5099 if (FIST.getNode() == 0) return Op;
5102 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5103 FIST, StackSlot, NULL, 0);
5106 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5107 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5108 SDValue FIST = Vals.first, StackSlot = Vals.second;
5109 assert(FIST.getNode() && "Unexpected failure");
5112 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5113 FIST, StackSlot, NULL, 0);
5116 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5117 DebugLoc dl = Op.getDebugLoc();
5118 MVT VT = Op.getValueType();
5121 EltVT = VT.getVectorElementType();
5122 std::vector<Constant*> CV;
5123 if (EltVT == MVT::f64) {
5124 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
5128 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
5134 Constant *C = ConstantVector::get(CV);
5135 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5136 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5137 PseudoSourceValue::getConstantPool(), 0,
5139 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5142 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5143 DebugLoc dl = Op.getDebugLoc();
5144 MVT VT = Op.getValueType();
5146 unsigned EltNum = 1;
5147 if (VT.isVector()) {
5148 EltVT = VT.getVectorElementType();
5149 EltNum = VT.getVectorNumElements();
5151 std::vector<Constant*> CV;
5152 if (EltVT == MVT::f64) {
5153 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
5157 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
5163 Constant *C = ConstantVector::get(CV);
5164 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5165 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5166 PseudoSourceValue::getConstantPool(), 0,
5168 if (VT.isVector()) {
5169 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5170 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5171 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5173 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5175 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5179 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5180 SDValue Op0 = Op.getOperand(0);
5181 SDValue Op1 = Op.getOperand(1);
5182 DebugLoc dl = Op.getDebugLoc();
5183 MVT VT = Op.getValueType();
5184 MVT SrcVT = Op1.getValueType();
5186 // If second operand is smaller, extend it first.
5187 if (SrcVT.bitsLT(VT)) {
5188 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5191 // And if it is bigger, shrink it first.
5192 if (SrcVT.bitsGT(VT)) {
5193 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5197 // At this point the operands and the result should have the same
5198 // type, and that won't be f80 since that is not custom lowered.
5200 // First get the sign bit of second operand.
5201 std::vector<Constant*> CV;
5202 if (SrcVT == MVT::f64) {
5203 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
5204 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5206 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
5207 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5208 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5209 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5211 Constant *C = ConstantVector::get(CV);
5212 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5213 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5214 PseudoSourceValue::getConstantPool(), 0,
5216 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5218 // Shift sign bit right or left if the two operands have different types.
5219 if (SrcVT.bitsGT(VT)) {
5220 // Op0 is MVT::f32, Op1 is MVT::f64.
5221 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5222 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5223 DAG.getConstant(32, MVT::i32));
5224 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5225 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5226 DAG.getIntPtrConstant(0));
5229 // Clear first operand sign bit.
5231 if (VT == MVT::f64) {
5232 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5233 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5235 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5236 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5237 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5238 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5240 C = ConstantVector::get(CV);
5241 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5242 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5243 PseudoSourceValue::getConstantPool(), 0,
5245 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5247 // Or the value with the sign bit.
5248 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5251 /// Emit nodes that will be selected as "test Op0,Op0", or something
5253 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5254 SelectionDAG &DAG) {
5255 DebugLoc dl = Op.getDebugLoc();
5257 // CF and OF aren't always set the way we want. Determine which
5258 // of these we need.
5259 bool NeedCF = false;
5260 bool NeedOF = false;
5262 case X86::COND_A: case X86::COND_AE:
5263 case X86::COND_B: case X86::COND_BE:
5266 case X86::COND_G: case X86::COND_GE:
5267 case X86::COND_L: case X86::COND_LE:
5268 case X86::COND_O: case X86::COND_NO:
5274 // See if we can use the EFLAGS value from the operand instead of
5275 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5276 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5277 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5278 unsigned Opcode = 0;
5279 unsigned NumOperands = 0;
5280 switch (Op.getNode()->getOpcode()) {
5282 // Due to an isel shortcoming, be conservative if this add is likely to
5283 // be selected as part of a load-modify-store instruction. When the root
5284 // node in a match is a store, isel doesn't know how to remap non-chain
5285 // non-flag uses of other nodes in the match, such as the ADD in this
5286 // case. This leads to the ADD being left around and reselected, with
5287 // the result being two adds in the output.
5288 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5289 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5290 if (UI->getOpcode() == ISD::STORE)
5292 if (ConstantSDNode *C =
5293 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5294 // An add of one will be selected as an INC.
5295 if (C->getAPIntValue() == 1) {
5296 Opcode = X86ISD::INC;
5300 // An add of negative one (subtract of one) will be selected as a DEC.
5301 if (C->getAPIntValue().isAllOnesValue()) {
5302 Opcode = X86ISD::DEC;
5307 // Otherwise use a regular EFLAGS-setting add.
5308 Opcode = X86ISD::ADD;
5312 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5313 // likely to be selected as part of a load-modify-store instruction.
5314 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5315 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5316 if (UI->getOpcode() == ISD::STORE)
5318 // Otherwise use a regular EFLAGS-setting sub.
5319 Opcode = X86ISD::SUB;
5326 return SDValue(Op.getNode(), 1);
5332 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5333 SmallVector<SDValue, 4> Ops;
5334 for (unsigned i = 0; i != NumOperands; ++i)
5335 Ops.push_back(Op.getOperand(i));
5336 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5337 DAG.ReplaceAllUsesWith(Op, New);
5338 return SDValue(New.getNode(), 1);
5342 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5343 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5344 DAG.getConstant(0, Op.getValueType()));
5347 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5349 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5350 SelectionDAG &DAG) {
5351 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5352 if (C->getAPIntValue() == 0)
5353 return EmitTest(Op0, X86CC, DAG);
5355 DebugLoc dl = Op0.getDebugLoc();
5356 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5359 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5360 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5361 SDValue Op0 = Op.getOperand(0);
5362 SDValue Op1 = Op.getOperand(1);
5363 DebugLoc dl = Op.getDebugLoc();
5364 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5366 // Lower (X & (1 << N)) == 0 to BT(X, N).
5367 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5368 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5369 if (Op0.getOpcode() == ISD::AND &&
5371 Op1.getOpcode() == ISD::Constant &&
5372 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5373 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5375 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5376 if (ConstantSDNode *Op010C =
5377 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5378 if (Op010C->getZExtValue() == 1) {
5379 LHS = Op0.getOperand(0);
5380 RHS = Op0.getOperand(1).getOperand(1);
5382 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5383 if (ConstantSDNode *Op000C =
5384 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5385 if (Op000C->getZExtValue() == 1) {
5386 LHS = Op0.getOperand(1);
5387 RHS = Op0.getOperand(0).getOperand(1);
5389 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5390 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5391 SDValue AndLHS = Op0.getOperand(0);
5392 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5393 LHS = AndLHS.getOperand(0);
5394 RHS = AndLHS.getOperand(1);
5398 if (LHS.getNode()) {
5399 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5400 // instruction. Since the shift amount is in-range-or-undefined, we know
5401 // that doing a bittest on the i16 value is ok. We extend to i32 because
5402 // the encoding for the i16 version is larger than the i32 version.
5403 if (LHS.getValueType() == MVT::i8)
5404 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5406 // If the operand types disagree, extend the shift amount to match. Since
5407 // BT ignores high bits (like shifts) we can use anyextend.
5408 if (LHS.getValueType() != RHS.getValueType())
5409 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5411 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5412 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5413 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5414 DAG.getConstant(Cond, MVT::i8), BT);
5418 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5419 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5421 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5422 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5423 DAG.getConstant(X86CC, MVT::i8), Cond);
5426 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5428 SDValue Op0 = Op.getOperand(0);
5429 SDValue Op1 = Op.getOperand(1);
5430 SDValue CC = Op.getOperand(2);
5431 MVT VT = Op.getValueType();
5432 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5433 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5434 DebugLoc dl = Op.getDebugLoc();
5438 MVT VT0 = Op0.getValueType();
5439 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5440 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5443 switch (SetCCOpcode) {
5446 case ISD::SETEQ: SSECC = 0; break;
5448 case ISD::SETGT: Swap = true; // Fallthrough
5450 case ISD::SETOLT: SSECC = 1; break;
5452 case ISD::SETGE: Swap = true; // Fallthrough
5454 case ISD::SETOLE: SSECC = 2; break;
5455 case ISD::SETUO: SSECC = 3; break;
5457 case ISD::SETNE: SSECC = 4; break;
5458 case ISD::SETULE: Swap = true;
5459 case ISD::SETUGE: SSECC = 5; break;
5460 case ISD::SETULT: Swap = true;
5461 case ISD::SETUGT: SSECC = 6; break;
5462 case ISD::SETO: SSECC = 7; break;
5465 std::swap(Op0, Op1);
5467 // In the two special cases we can't handle, emit two comparisons.
5469 if (SetCCOpcode == ISD::SETUEQ) {
5471 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5472 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5473 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5475 else if (SetCCOpcode == ISD::SETONE) {
5477 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5478 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5479 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5481 assert(0 && "Illegal FP comparison");
5483 // Handle all other FP comparisons here.
5484 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5487 // We are handling one of the integer comparisons here. Since SSE only has
5488 // GT and EQ comparisons for integer, swapping operands and multiple
5489 // operations may be required for some comparisons.
5490 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5491 bool Swap = false, Invert = false, FlipSigns = false;
5493 switch (VT.getSimpleVT()) {
5495 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5496 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5497 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5498 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5501 switch (SetCCOpcode) {
5503 case ISD::SETNE: Invert = true;
5504 case ISD::SETEQ: Opc = EQOpc; break;
5505 case ISD::SETLT: Swap = true;
5506 case ISD::SETGT: Opc = GTOpc; break;
5507 case ISD::SETGE: Swap = true;
5508 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5509 case ISD::SETULT: Swap = true;
5510 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5511 case ISD::SETUGE: Swap = true;
5512 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5515 std::swap(Op0, Op1);
5517 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5518 // bits of the inputs before performing those operations.
5520 MVT EltVT = VT.getVectorElementType();
5521 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5523 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5524 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5526 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5527 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5530 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5532 // If the logical-not of the result is required, perform that now.
5534 Result = DAG.getNOT(dl, Result, VT);
5539 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5540 static bool isX86LogicalCmp(SDValue Op) {
5541 unsigned Opc = Op.getNode()->getOpcode();
5542 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5544 if (Op.getResNo() == 1 &&
5545 (Opc == X86ISD::ADD ||
5546 Opc == X86ISD::SUB ||
5547 Opc == X86ISD::SMUL ||
5548 Opc == X86ISD::UMUL ||
5549 Opc == X86ISD::INC ||
5550 Opc == X86ISD::DEC))
5556 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5557 bool addTest = true;
5558 SDValue Cond = Op.getOperand(0);
5559 DebugLoc dl = Op.getDebugLoc();
5562 if (Cond.getOpcode() == ISD::SETCC)
5563 Cond = LowerSETCC(Cond, DAG);
5565 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5566 // setting operand in place of the X86ISD::SETCC.
5567 if (Cond.getOpcode() == X86ISD::SETCC) {
5568 CC = Cond.getOperand(0);
5570 SDValue Cmp = Cond.getOperand(1);
5571 unsigned Opc = Cmp.getOpcode();
5572 MVT VT = Op.getValueType();
5574 bool IllegalFPCMov = false;
5575 if (VT.isFloatingPoint() && !VT.isVector() &&
5576 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5577 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5579 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5580 Opc == X86ISD::BT) { // FIXME
5587 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5588 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5591 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5592 SmallVector<SDValue, 4> Ops;
5593 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5594 // condition is true.
5595 Ops.push_back(Op.getOperand(2));
5596 Ops.push_back(Op.getOperand(1));
5598 Ops.push_back(Cond);
5599 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5602 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5603 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5604 // from the AND / OR.
5605 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5606 Opc = Op.getOpcode();
5607 if (Opc != ISD::OR && Opc != ISD::AND)
5609 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5610 Op.getOperand(0).hasOneUse() &&
5611 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5612 Op.getOperand(1).hasOneUse());
5615 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5616 // 1 and that the SETCC node has a single use.
5617 static bool isXor1OfSetCC(SDValue Op) {
5618 if (Op.getOpcode() != ISD::XOR)
5620 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5621 if (N1C && N1C->getAPIntValue() == 1) {
5622 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5623 Op.getOperand(0).hasOneUse();
5628 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5629 bool addTest = true;
5630 SDValue Chain = Op.getOperand(0);
5631 SDValue Cond = Op.getOperand(1);
5632 SDValue Dest = Op.getOperand(2);
5633 DebugLoc dl = Op.getDebugLoc();
5636 if (Cond.getOpcode() == ISD::SETCC)
5637 Cond = LowerSETCC(Cond, DAG);
5639 // FIXME: LowerXALUO doesn't handle these!!
5640 else if (Cond.getOpcode() == X86ISD::ADD ||
5641 Cond.getOpcode() == X86ISD::SUB ||
5642 Cond.getOpcode() == X86ISD::SMUL ||
5643 Cond.getOpcode() == X86ISD::UMUL)
5644 Cond = LowerXALUO(Cond, DAG);
5647 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5648 // setting operand in place of the X86ISD::SETCC.
5649 if (Cond.getOpcode() == X86ISD::SETCC) {
5650 CC = Cond.getOperand(0);
5652 SDValue Cmp = Cond.getOperand(1);
5653 unsigned Opc = Cmp.getOpcode();
5654 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5655 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5659 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5663 // These can only come from an arithmetic instruction with overflow,
5664 // e.g. SADDO, UADDO.
5665 Cond = Cond.getNode()->getOperand(1);
5672 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5673 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5674 if (CondOpc == ISD::OR) {
5675 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5676 // two branches instead of an explicit OR instruction with a
5678 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5679 isX86LogicalCmp(Cmp)) {
5680 CC = Cond.getOperand(0).getOperand(0);
5681 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5682 Chain, Dest, CC, Cmp);
5683 CC = Cond.getOperand(1).getOperand(0);
5687 } else { // ISD::AND
5688 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5689 // two branches instead of an explicit AND instruction with a
5690 // separate test. However, we only do this if this block doesn't
5691 // have a fall-through edge, because this requires an explicit
5692 // jmp when the condition is false.
5693 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5694 isX86LogicalCmp(Cmp) &&
5695 Op.getNode()->hasOneUse()) {
5696 X86::CondCode CCode =
5697 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5698 CCode = X86::GetOppositeBranchCondition(CCode);
5699 CC = DAG.getConstant(CCode, MVT::i8);
5700 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5701 // Look for an unconditional branch following this conditional branch.
5702 // We need this because we need to reverse the successors in order
5703 // to implement FCMP_OEQ.
5704 if (User.getOpcode() == ISD::BR) {
5705 SDValue FalseBB = User.getOperand(1);
5707 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5708 assert(NewBR == User);
5711 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5712 Chain, Dest, CC, Cmp);
5713 X86::CondCode CCode =
5714 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5715 CCode = X86::GetOppositeBranchCondition(CCode);
5716 CC = DAG.getConstant(CCode, MVT::i8);
5722 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5723 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5724 // It should be transformed during dag combiner except when the condition
5725 // is set by a arithmetics with overflow node.
5726 X86::CondCode CCode =
5727 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5728 CCode = X86::GetOppositeBranchCondition(CCode);
5729 CC = DAG.getConstant(CCode, MVT::i8);
5730 Cond = Cond.getOperand(0).getOperand(1);
5736 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5737 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5739 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5740 Chain, Dest, CC, Cond);
5744 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5745 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5746 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5747 // that the guard pages used by the OS virtual memory manager are allocated in
5748 // correct sequence.
5750 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5751 SelectionDAG &DAG) {
5752 assert(Subtarget->isTargetCygMing() &&
5753 "This should be used only on Cygwin/Mingw targets");
5754 DebugLoc dl = Op.getDebugLoc();
5757 SDValue Chain = Op.getOperand(0);
5758 SDValue Size = Op.getOperand(1);
5759 // FIXME: Ensure alignment here
5763 MVT IntPtr = getPointerTy();
5764 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5766 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5768 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5769 Flag = Chain.getValue(1);
5771 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5772 SDValue Ops[] = { Chain,
5773 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5774 DAG.getRegister(X86::EAX, IntPtr),
5775 DAG.getRegister(X86StackPtr, SPTy),
5777 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5778 Flag = Chain.getValue(1);
5780 Chain = DAG.getCALLSEQ_END(Chain,
5781 DAG.getIntPtrConstant(0, true),
5782 DAG.getIntPtrConstant(0, true),
5785 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5787 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5788 return DAG.getMergeValues(Ops1, 2, dl);
5792 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5794 SDValue Dst, SDValue Src,
5795 SDValue Size, unsigned Align,
5797 uint64_t DstSVOff) {
5798 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5800 // If not DWORD aligned or size is more than the threshold, call the library.
5801 // The libc version is likely to be faster for these cases. It can use the
5802 // address value and run time information about the CPU.
5803 if ((Align & 3) != 0 ||
5805 ConstantSize->getZExtValue() >
5806 getSubtarget()->getMaxInlineSizeThreshold()) {
5807 SDValue InFlag(0, 0);
5809 // Check to see if there is a specialized entry-point for memory zeroing.
5810 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5812 if (const char *bzeroEntry = V &&
5813 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5814 MVT IntPtr = getPointerTy();
5815 const Type *IntPtrTy = TD->getIntPtrType();
5816 TargetLowering::ArgListTy Args;
5817 TargetLowering::ArgListEntry Entry;
5819 Entry.Ty = IntPtrTy;
5820 Args.push_back(Entry);
5822 Args.push_back(Entry);
5823 std::pair<SDValue,SDValue> CallResult =
5824 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5825 0, CallingConv::C, false,
5826 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5827 return CallResult.second;
5830 // Otherwise have the target-independent code call memset.
5834 uint64_t SizeVal = ConstantSize->getZExtValue();
5835 SDValue InFlag(0, 0);
5838 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5839 unsigned BytesLeft = 0;
5840 bool TwoRepStos = false;
5843 uint64_t Val = ValC->getZExtValue() & 255;
5845 // If the value is a constant, then we can potentially use larger sets.
5846 switch (Align & 3) {
5847 case 2: // WORD aligned
5850 Val = (Val << 8) | Val;
5852 case 0: // DWORD aligned
5855 Val = (Val << 8) | Val;
5856 Val = (Val << 16) | Val;
5857 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5860 Val = (Val << 32) | Val;
5863 default: // Byte aligned
5866 Count = DAG.getIntPtrConstant(SizeVal);
5870 if (AVT.bitsGT(MVT::i8)) {
5871 unsigned UBytes = AVT.getSizeInBits() / 8;
5872 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5873 BytesLeft = SizeVal % UBytes;
5876 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5878 InFlag = Chain.getValue(1);
5881 Count = DAG.getIntPtrConstant(SizeVal);
5882 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5883 InFlag = Chain.getValue(1);
5886 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5889 InFlag = Chain.getValue(1);
5890 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5893 InFlag = Chain.getValue(1);
5895 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5896 SmallVector<SDValue, 8> Ops;
5897 Ops.push_back(Chain);
5898 Ops.push_back(DAG.getValueType(AVT));
5899 Ops.push_back(InFlag);
5900 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5903 InFlag = Chain.getValue(1);
5905 MVT CVT = Count.getValueType();
5906 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5907 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5908 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5911 InFlag = Chain.getValue(1);
5912 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5914 Ops.push_back(Chain);
5915 Ops.push_back(DAG.getValueType(MVT::i8));
5916 Ops.push_back(InFlag);
5917 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5918 } else if (BytesLeft) {
5919 // Handle the last 1 - 7 bytes.
5920 unsigned Offset = SizeVal - BytesLeft;
5921 MVT AddrVT = Dst.getValueType();
5922 MVT SizeVT = Size.getValueType();
5924 Chain = DAG.getMemset(Chain, dl,
5925 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5926 DAG.getConstant(Offset, AddrVT)),
5928 DAG.getConstant(BytesLeft, SizeVT),
5929 Align, DstSV, DstSVOff + Offset);
5932 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5937 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5938 SDValue Chain, SDValue Dst, SDValue Src,
5939 SDValue Size, unsigned Align,
5941 const Value *DstSV, uint64_t DstSVOff,
5942 const Value *SrcSV, uint64_t SrcSVOff) {
5943 // This requires the copy size to be a constant, preferrably
5944 // within a subtarget-specific limit.
5945 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5948 uint64_t SizeVal = ConstantSize->getZExtValue();
5949 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5952 /// If not DWORD aligned, call the library.
5953 if ((Align & 3) != 0)
5958 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5961 unsigned UBytes = AVT.getSizeInBits() / 8;
5962 unsigned CountVal = SizeVal / UBytes;
5963 SDValue Count = DAG.getIntPtrConstant(CountVal);
5964 unsigned BytesLeft = SizeVal % UBytes;
5966 SDValue InFlag(0, 0);
5967 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5970 InFlag = Chain.getValue(1);
5971 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5974 InFlag = Chain.getValue(1);
5975 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5978 InFlag = Chain.getValue(1);
5980 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5981 SmallVector<SDValue, 8> Ops;
5982 Ops.push_back(Chain);
5983 Ops.push_back(DAG.getValueType(AVT));
5984 Ops.push_back(InFlag);
5985 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5987 SmallVector<SDValue, 4> Results;
5988 Results.push_back(RepMovs);
5990 // Handle the last 1 - 7 bytes.
5991 unsigned Offset = SizeVal - BytesLeft;
5992 MVT DstVT = Dst.getValueType();
5993 MVT SrcVT = Src.getValueType();
5994 MVT SizeVT = Size.getValueType();
5995 Results.push_back(DAG.getMemcpy(Chain, dl,
5996 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5997 DAG.getConstant(Offset, DstVT)),
5998 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5999 DAG.getConstant(Offset, SrcVT)),
6000 DAG.getConstant(BytesLeft, SizeVT),
6001 Align, AlwaysInline,
6002 DstSV, DstSVOff + Offset,
6003 SrcSV, SrcSVOff + Offset));
6006 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6007 &Results[0], Results.size());
6010 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6011 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6012 DebugLoc dl = Op.getDebugLoc();
6014 if (!Subtarget->is64Bit()) {
6015 // vastart just stores the address of the VarArgsFrameIndex slot into the
6016 // memory location argument.
6017 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6018 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6022 // gp_offset (0 - 6 * 8)
6023 // fp_offset (48 - 48 + 8 * 16)
6024 // overflow_arg_area (point to parameters coming in memory).
6026 SmallVector<SDValue, 8> MemOps;
6027 SDValue FIN = Op.getOperand(1);
6029 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6030 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6032 MemOps.push_back(Store);
6035 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6036 FIN, DAG.getIntPtrConstant(4));
6037 Store = DAG.getStore(Op.getOperand(0), dl,
6038 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6040 MemOps.push_back(Store);
6042 // Store ptr to overflow_arg_area
6043 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6044 FIN, DAG.getIntPtrConstant(4));
6045 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6046 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6047 MemOps.push_back(Store);
6049 // Store ptr to reg_save_area.
6050 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6051 FIN, DAG.getIntPtrConstant(8));
6052 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6053 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6054 MemOps.push_back(Store);
6055 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6056 &MemOps[0], MemOps.size());
6059 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6060 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6061 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6062 SDValue Chain = Op.getOperand(0);
6063 SDValue SrcPtr = Op.getOperand(1);
6064 SDValue SrcSV = Op.getOperand(2);
6066 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6070 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6071 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6072 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6073 SDValue Chain = Op.getOperand(0);
6074 SDValue DstPtr = Op.getOperand(1);
6075 SDValue SrcPtr = Op.getOperand(2);
6076 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6077 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6078 DebugLoc dl = Op.getDebugLoc();
6080 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6081 DAG.getIntPtrConstant(24), 8, false,
6082 DstSV, 0, SrcSV, 0);
6086 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6087 DebugLoc dl = Op.getDebugLoc();
6088 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6090 default: return SDValue(); // Don't custom lower most intrinsics.
6091 // Comparison intrinsics.
6092 case Intrinsic::x86_sse_comieq_ss:
6093 case Intrinsic::x86_sse_comilt_ss:
6094 case Intrinsic::x86_sse_comile_ss:
6095 case Intrinsic::x86_sse_comigt_ss:
6096 case Intrinsic::x86_sse_comige_ss:
6097 case Intrinsic::x86_sse_comineq_ss:
6098 case Intrinsic::x86_sse_ucomieq_ss:
6099 case Intrinsic::x86_sse_ucomilt_ss:
6100 case Intrinsic::x86_sse_ucomile_ss:
6101 case Intrinsic::x86_sse_ucomigt_ss:
6102 case Intrinsic::x86_sse_ucomige_ss:
6103 case Intrinsic::x86_sse_ucomineq_ss:
6104 case Intrinsic::x86_sse2_comieq_sd:
6105 case Intrinsic::x86_sse2_comilt_sd:
6106 case Intrinsic::x86_sse2_comile_sd:
6107 case Intrinsic::x86_sse2_comigt_sd:
6108 case Intrinsic::x86_sse2_comige_sd:
6109 case Intrinsic::x86_sse2_comineq_sd:
6110 case Intrinsic::x86_sse2_ucomieq_sd:
6111 case Intrinsic::x86_sse2_ucomilt_sd:
6112 case Intrinsic::x86_sse2_ucomile_sd:
6113 case Intrinsic::x86_sse2_ucomigt_sd:
6114 case Intrinsic::x86_sse2_ucomige_sd:
6115 case Intrinsic::x86_sse2_ucomineq_sd: {
6117 ISD::CondCode CC = ISD::SETCC_INVALID;
6120 case Intrinsic::x86_sse_comieq_ss:
6121 case Intrinsic::x86_sse2_comieq_sd:
6125 case Intrinsic::x86_sse_comilt_ss:
6126 case Intrinsic::x86_sse2_comilt_sd:
6130 case Intrinsic::x86_sse_comile_ss:
6131 case Intrinsic::x86_sse2_comile_sd:
6135 case Intrinsic::x86_sse_comigt_ss:
6136 case Intrinsic::x86_sse2_comigt_sd:
6140 case Intrinsic::x86_sse_comige_ss:
6141 case Intrinsic::x86_sse2_comige_sd:
6145 case Intrinsic::x86_sse_comineq_ss:
6146 case Intrinsic::x86_sse2_comineq_sd:
6150 case Intrinsic::x86_sse_ucomieq_ss:
6151 case Intrinsic::x86_sse2_ucomieq_sd:
6152 Opc = X86ISD::UCOMI;
6155 case Intrinsic::x86_sse_ucomilt_ss:
6156 case Intrinsic::x86_sse2_ucomilt_sd:
6157 Opc = X86ISD::UCOMI;
6160 case Intrinsic::x86_sse_ucomile_ss:
6161 case Intrinsic::x86_sse2_ucomile_sd:
6162 Opc = X86ISD::UCOMI;
6165 case Intrinsic::x86_sse_ucomigt_ss:
6166 case Intrinsic::x86_sse2_ucomigt_sd:
6167 Opc = X86ISD::UCOMI;
6170 case Intrinsic::x86_sse_ucomige_ss:
6171 case Intrinsic::x86_sse2_ucomige_sd:
6172 Opc = X86ISD::UCOMI;
6175 case Intrinsic::x86_sse_ucomineq_ss:
6176 case Intrinsic::x86_sse2_ucomineq_sd:
6177 Opc = X86ISD::UCOMI;
6182 SDValue LHS = Op.getOperand(1);
6183 SDValue RHS = Op.getOperand(2);
6184 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6185 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6186 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6187 DAG.getConstant(X86CC, MVT::i8), Cond);
6188 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6191 // Fix vector shift instructions where the last operand is a non-immediate
6193 case Intrinsic::x86_sse2_pslli_w:
6194 case Intrinsic::x86_sse2_pslli_d:
6195 case Intrinsic::x86_sse2_pslli_q:
6196 case Intrinsic::x86_sse2_psrli_w:
6197 case Intrinsic::x86_sse2_psrli_d:
6198 case Intrinsic::x86_sse2_psrli_q:
6199 case Intrinsic::x86_sse2_psrai_w:
6200 case Intrinsic::x86_sse2_psrai_d:
6201 case Intrinsic::x86_mmx_pslli_w:
6202 case Intrinsic::x86_mmx_pslli_d:
6203 case Intrinsic::x86_mmx_pslli_q:
6204 case Intrinsic::x86_mmx_psrli_w:
6205 case Intrinsic::x86_mmx_psrli_d:
6206 case Intrinsic::x86_mmx_psrli_q:
6207 case Intrinsic::x86_mmx_psrai_w:
6208 case Intrinsic::x86_mmx_psrai_d: {
6209 SDValue ShAmt = Op.getOperand(2);
6210 if (isa<ConstantSDNode>(ShAmt))
6213 unsigned NewIntNo = 0;
6214 MVT ShAmtVT = MVT::v4i32;
6216 case Intrinsic::x86_sse2_pslli_w:
6217 NewIntNo = Intrinsic::x86_sse2_psll_w;
6219 case Intrinsic::x86_sse2_pslli_d:
6220 NewIntNo = Intrinsic::x86_sse2_psll_d;
6222 case Intrinsic::x86_sse2_pslli_q:
6223 NewIntNo = Intrinsic::x86_sse2_psll_q;
6225 case Intrinsic::x86_sse2_psrli_w:
6226 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6228 case Intrinsic::x86_sse2_psrli_d:
6229 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6231 case Intrinsic::x86_sse2_psrli_q:
6232 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6234 case Intrinsic::x86_sse2_psrai_w:
6235 NewIntNo = Intrinsic::x86_sse2_psra_w;
6237 case Intrinsic::x86_sse2_psrai_d:
6238 NewIntNo = Intrinsic::x86_sse2_psra_d;
6241 ShAmtVT = MVT::v2i32;
6243 case Intrinsic::x86_mmx_pslli_w:
6244 NewIntNo = Intrinsic::x86_mmx_psll_w;
6246 case Intrinsic::x86_mmx_pslli_d:
6247 NewIntNo = Intrinsic::x86_mmx_psll_d;
6249 case Intrinsic::x86_mmx_pslli_q:
6250 NewIntNo = Intrinsic::x86_mmx_psll_q;
6252 case Intrinsic::x86_mmx_psrli_w:
6253 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6255 case Intrinsic::x86_mmx_psrli_d:
6256 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6258 case Intrinsic::x86_mmx_psrli_q:
6259 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6261 case Intrinsic::x86_mmx_psrai_w:
6262 NewIntNo = Intrinsic::x86_mmx_psra_w;
6264 case Intrinsic::x86_mmx_psrai_d:
6265 NewIntNo = Intrinsic::x86_mmx_psra_d;
6267 default: LLVM_UNREACHABLE("Impossible intrinsic"); // Can't reach here.
6272 MVT VT = Op.getValueType();
6273 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6274 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6275 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6276 DAG.getConstant(NewIntNo, MVT::i32),
6277 Op.getOperand(1), ShAmt);
6282 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6283 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6284 DebugLoc dl = Op.getDebugLoc();
6287 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6289 DAG.getConstant(TD->getPointerSize(),
6290 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6291 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6292 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6297 // Just load the return address.
6298 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6299 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6300 RetAddrFI, NULL, 0);
6303 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6304 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6305 MFI->setFrameAddressIsTaken(true);
6306 MVT VT = Op.getValueType();
6307 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6308 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6309 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6310 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6312 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6316 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6317 SelectionDAG &DAG) {
6318 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6321 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6323 MachineFunction &MF = DAG.getMachineFunction();
6324 SDValue Chain = Op.getOperand(0);
6325 SDValue Offset = Op.getOperand(1);
6326 SDValue Handler = Op.getOperand(2);
6327 DebugLoc dl = Op.getDebugLoc();
6329 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6331 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6333 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6334 DAG.getIntPtrConstant(-TD->getPointerSize()));
6335 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6336 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6337 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6338 MF.getRegInfo().addLiveOut(StoreAddrReg);
6340 return DAG.getNode(X86ISD::EH_RETURN, dl,
6342 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6345 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6346 SelectionDAG &DAG) {
6347 SDValue Root = Op.getOperand(0);
6348 SDValue Trmp = Op.getOperand(1); // trampoline
6349 SDValue FPtr = Op.getOperand(2); // nested function
6350 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6351 DebugLoc dl = Op.getDebugLoc();
6353 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6355 const X86InstrInfo *TII =
6356 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6358 if (Subtarget->is64Bit()) {
6359 SDValue OutChains[6];
6361 // Large code-model.
6363 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6364 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6366 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6367 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6369 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6371 // Load the pointer to the nested function into R11.
6372 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6373 SDValue Addr = Trmp;
6374 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6377 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6378 DAG.getConstant(2, MVT::i64));
6379 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6381 // Load the 'nest' parameter value into R10.
6382 // R10 is specified in X86CallingConv.td
6383 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6384 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6385 DAG.getConstant(10, MVT::i64));
6386 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6387 Addr, TrmpAddr, 10);
6389 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6390 DAG.getConstant(12, MVT::i64));
6391 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6393 // Jump to the nested function.
6394 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6395 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6396 DAG.getConstant(20, MVT::i64));
6397 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6398 Addr, TrmpAddr, 20);
6400 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6401 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6402 DAG.getConstant(22, MVT::i64));
6403 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6407 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6408 return DAG.getMergeValues(Ops, 2, dl);
6410 const Function *Func =
6411 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6412 unsigned CC = Func->getCallingConv();
6417 assert(0 && "Unsupported calling convention");
6418 case CallingConv::C:
6419 case CallingConv::X86_StdCall: {
6420 // Pass 'nest' parameter in ECX.
6421 // Must be kept in sync with X86CallingConv.td
6424 // Check that ECX wasn't needed by an 'inreg' parameter.
6425 const FunctionType *FTy = Func->getFunctionType();
6426 const AttrListPtr &Attrs = Func->getAttributes();
6428 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6429 unsigned InRegCount = 0;
6432 for (FunctionType::param_iterator I = FTy->param_begin(),
6433 E = FTy->param_end(); I != E; ++I, ++Idx)
6434 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6435 // FIXME: should only count parameters that are lowered to integers.
6436 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6438 if (InRegCount > 2) {
6439 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6444 case CallingConv::X86_FastCall:
6445 case CallingConv::Fast:
6446 // Pass 'nest' parameter in EAX.
6447 // Must be kept in sync with X86CallingConv.td
6452 SDValue OutChains[4];
6455 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6456 DAG.getConstant(10, MVT::i32));
6457 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6459 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6460 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6461 OutChains[0] = DAG.getStore(Root, dl,
6462 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6465 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6466 DAG.getConstant(1, MVT::i32));
6467 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6469 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6470 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6471 DAG.getConstant(5, MVT::i32));
6472 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6473 TrmpAddr, 5, false, 1);
6475 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6476 DAG.getConstant(6, MVT::i32));
6477 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6480 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6481 return DAG.getMergeValues(Ops, 2, dl);
6485 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6487 The rounding mode is in bits 11:10 of FPSR, and has the following
6494 FLT_ROUNDS, on the other hand, expects the following:
6501 To perform the conversion, we do:
6502 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6505 MachineFunction &MF = DAG.getMachineFunction();
6506 const TargetMachine &TM = MF.getTarget();
6507 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6508 unsigned StackAlignment = TFI.getStackAlignment();
6509 MVT VT = Op.getValueType();
6510 DebugLoc dl = Op.getDebugLoc();
6512 // Save FP Control Word to stack slot
6513 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6514 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6516 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6517 DAG.getEntryNode(), StackSlot);
6519 // Load FP Control Word from stack slot
6520 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6522 // Transform as necessary
6524 DAG.getNode(ISD::SRL, dl, MVT::i16,
6525 DAG.getNode(ISD::AND, dl, MVT::i16,
6526 CWD, DAG.getConstant(0x800, MVT::i16)),
6527 DAG.getConstant(11, MVT::i8));
6529 DAG.getNode(ISD::SRL, dl, MVT::i16,
6530 DAG.getNode(ISD::AND, dl, MVT::i16,
6531 CWD, DAG.getConstant(0x400, MVT::i16)),
6532 DAG.getConstant(9, MVT::i8));
6535 DAG.getNode(ISD::AND, dl, MVT::i16,
6536 DAG.getNode(ISD::ADD, dl, MVT::i16,
6537 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6538 DAG.getConstant(1, MVT::i16)),
6539 DAG.getConstant(3, MVT::i16));
6542 return DAG.getNode((VT.getSizeInBits() < 16 ?
6543 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6546 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6547 MVT VT = Op.getValueType();
6549 unsigned NumBits = VT.getSizeInBits();
6550 DebugLoc dl = Op.getDebugLoc();
6552 Op = Op.getOperand(0);
6553 if (VT == MVT::i8) {
6554 // Zero extend to i32 since there is not an i8 bsr.
6556 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6559 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6560 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6561 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6563 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6564 SmallVector<SDValue, 4> Ops;
6566 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6567 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6568 Ops.push_back(Op.getValue(1));
6569 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6571 // Finally xor with NumBits-1.
6572 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6575 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6579 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6580 MVT VT = Op.getValueType();
6582 unsigned NumBits = VT.getSizeInBits();
6583 DebugLoc dl = Op.getDebugLoc();
6585 Op = Op.getOperand(0);
6586 if (VT == MVT::i8) {
6588 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6591 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6592 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6593 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6595 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6596 SmallVector<SDValue, 4> Ops;
6598 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6599 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6600 Ops.push_back(Op.getValue(1));
6601 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6604 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6608 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6609 MVT VT = Op.getValueType();
6610 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6611 DebugLoc dl = Op.getDebugLoc();
6613 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6614 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6615 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6616 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6617 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6619 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6620 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6621 // return AloBlo + AloBhi + AhiBlo;
6623 SDValue A = Op.getOperand(0);
6624 SDValue B = Op.getOperand(1);
6626 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6627 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6628 A, DAG.getConstant(32, MVT::i32));
6629 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6630 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6631 B, DAG.getConstant(32, MVT::i32));
6632 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6633 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6635 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6636 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6638 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6639 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6641 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6642 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6643 AloBhi, DAG.getConstant(32, MVT::i32));
6644 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6645 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6646 AhiBlo, DAG.getConstant(32, MVT::i32));
6647 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6648 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6653 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6654 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6655 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6656 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6657 // has only one use.
6658 SDNode *N = Op.getNode();
6659 SDValue LHS = N->getOperand(0);
6660 SDValue RHS = N->getOperand(1);
6661 unsigned BaseOp = 0;
6663 DebugLoc dl = Op.getDebugLoc();
6665 switch (Op.getOpcode()) {
6666 default: assert(0 && "Unknown ovf instruction!");
6668 // A subtract of one will be selected as a INC. Note that INC doesn't
6669 // set CF, so we can't do this for UADDO.
6670 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6671 if (C->getAPIntValue() == 1) {
6672 BaseOp = X86ISD::INC;
6676 BaseOp = X86ISD::ADD;
6680 BaseOp = X86ISD::ADD;
6684 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6685 // set CF, so we can't do this for USUBO.
6686 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6687 if (C->getAPIntValue() == 1) {
6688 BaseOp = X86ISD::DEC;
6692 BaseOp = X86ISD::SUB;
6696 BaseOp = X86ISD::SUB;
6700 BaseOp = X86ISD::SMUL;
6704 BaseOp = X86ISD::UMUL;
6709 // Also sets EFLAGS.
6710 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6711 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6714 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6715 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6717 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6721 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6722 MVT T = Op.getValueType();
6723 DebugLoc dl = Op.getDebugLoc();
6726 switch(T.getSimpleVT()) {
6728 assert(false && "Invalid value type!");
6729 case MVT::i8: Reg = X86::AL; size = 1; break;
6730 case MVT::i16: Reg = X86::AX; size = 2; break;
6731 case MVT::i32: Reg = X86::EAX; size = 4; break;
6733 assert(Subtarget->is64Bit() && "Node not type legal!");
6734 Reg = X86::RAX; size = 8;
6737 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6738 Op.getOperand(2), SDValue());
6739 SDValue Ops[] = { cpIn.getValue(0),
6742 DAG.getTargetConstant(size, MVT::i8),
6744 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6745 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6747 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6751 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6752 SelectionDAG &DAG) {
6753 assert(Subtarget->is64Bit() && "Result not type legalized?");
6754 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6755 SDValue TheChain = Op.getOperand(0);
6756 DebugLoc dl = Op.getDebugLoc();
6757 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6758 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6759 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6761 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6762 DAG.getConstant(32, MVT::i8));
6764 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6767 return DAG.getMergeValues(Ops, 2, dl);
6770 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6771 SDNode *Node = Op.getNode();
6772 DebugLoc dl = Node->getDebugLoc();
6773 MVT T = Node->getValueType(0);
6774 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6775 DAG.getConstant(0, T), Node->getOperand(2));
6776 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6777 cast<AtomicSDNode>(Node)->getMemoryVT(),
6778 Node->getOperand(0),
6779 Node->getOperand(1), negOp,
6780 cast<AtomicSDNode>(Node)->getSrcValue(),
6781 cast<AtomicSDNode>(Node)->getAlignment());
6784 /// LowerOperation - Provide custom lowering hooks for some operations.
6786 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6787 switch (Op.getOpcode()) {
6788 default: assert(0 && "Should not custom lower this!");
6789 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6790 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6791 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6792 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6793 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6794 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6795 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6796 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6797 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6798 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6799 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6800 case ISD::SHL_PARTS:
6801 case ISD::SRA_PARTS:
6802 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6803 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6804 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6805 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6806 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6807 case ISD::FABS: return LowerFABS(Op, DAG);
6808 case ISD::FNEG: return LowerFNEG(Op, DAG);
6809 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6810 case ISD::SETCC: return LowerSETCC(Op, DAG);
6811 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6812 case ISD::SELECT: return LowerSELECT(Op, DAG);
6813 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6814 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6815 case ISD::CALL: return LowerCALL(Op, DAG);
6816 case ISD::RET: return LowerRET(Op, DAG);
6817 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6818 case ISD::VASTART: return LowerVASTART(Op, DAG);
6819 case ISD::VAARG: return LowerVAARG(Op, DAG);
6820 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6821 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6822 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6823 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6824 case ISD::FRAME_TO_ARGS_OFFSET:
6825 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6826 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6827 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6828 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6829 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6830 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6831 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6832 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6838 case ISD::UMULO: return LowerXALUO(Op, DAG);
6839 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6843 void X86TargetLowering::
6844 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6845 SelectionDAG &DAG, unsigned NewOp) {
6846 MVT T = Node->getValueType(0);
6847 DebugLoc dl = Node->getDebugLoc();
6848 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6850 SDValue Chain = Node->getOperand(0);
6851 SDValue In1 = Node->getOperand(1);
6852 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6853 Node->getOperand(2), DAG.getIntPtrConstant(0));
6854 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6855 Node->getOperand(2), DAG.getIntPtrConstant(1));
6856 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6857 // have a MemOperand. Pass the info through as a normal operand.
6858 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6859 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6860 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6861 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6862 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6863 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6864 Results.push_back(Result.getValue(2));
6867 /// ReplaceNodeResults - Replace a node with an illegal result type
6868 /// with a new node built out of custom code.
6869 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6870 SmallVectorImpl<SDValue>&Results,
6871 SelectionDAG &DAG) {
6872 DebugLoc dl = N->getDebugLoc();
6873 switch (N->getOpcode()) {
6875 assert(false && "Do not know how to custom type legalize this operation!");
6877 case ISD::FP_TO_SINT: {
6878 std::pair<SDValue,SDValue> Vals =
6879 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6880 SDValue FIST = Vals.first, StackSlot = Vals.second;
6881 if (FIST.getNode() != 0) {
6882 MVT VT = N->getValueType(0);
6883 // Return a load from the stack slot.
6884 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6888 case ISD::READCYCLECOUNTER: {
6889 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6890 SDValue TheChain = N->getOperand(0);
6891 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6892 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6894 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6896 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6897 SDValue Ops[] = { eax, edx };
6898 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6899 Results.push_back(edx.getValue(1));
6902 case ISD::ATOMIC_CMP_SWAP: {
6903 MVT T = N->getValueType(0);
6904 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6905 SDValue cpInL, cpInH;
6906 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6907 DAG.getConstant(0, MVT::i32));
6908 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6909 DAG.getConstant(1, MVT::i32));
6910 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6911 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6913 SDValue swapInL, swapInH;
6914 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6915 DAG.getConstant(0, MVT::i32));
6916 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6917 DAG.getConstant(1, MVT::i32));
6918 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6920 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6921 swapInL.getValue(1));
6922 SDValue Ops[] = { swapInH.getValue(0),
6924 swapInH.getValue(1) };
6925 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6926 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6927 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6928 MVT::i32, Result.getValue(1));
6929 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6930 MVT::i32, cpOutL.getValue(2));
6931 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6932 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6933 Results.push_back(cpOutH.getValue(1));
6936 case ISD::ATOMIC_LOAD_ADD:
6937 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6939 case ISD::ATOMIC_LOAD_AND:
6940 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6942 case ISD::ATOMIC_LOAD_NAND:
6943 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6945 case ISD::ATOMIC_LOAD_OR:
6946 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6948 case ISD::ATOMIC_LOAD_SUB:
6949 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6951 case ISD::ATOMIC_LOAD_XOR:
6952 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6954 case ISD::ATOMIC_SWAP:
6955 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6960 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6962 default: return NULL;
6963 case X86ISD::BSF: return "X86ISD::BSF";
6964 case X86ISD::BSR: return "X86ISD::BSR";
6965 case X86ISD::SHLD: return "X86ISD::SHLD";
6966 case X86ISD::SHRD: return "X86ISD::SHRD";
6967 case X86ISD::FAND: return "X86ISD::FAND";
6968 case X86ISD::FOR: return "X86ISD::FOR";
6969 case X86ISD::FXOR: return "X86ISD::FXOR";
6970 case X86ISD::FSRL: return "X86ISD::FSRL";
6971 case X86ISD::FILD: return "X86ISD::FILD";
6972 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6973 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6974 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6975 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6976 case X86ISD::FLD: return "X86ISD::FLD";
6977 case X86ISD::FST: return "X86ISD::FST";
6978 case X86ISD::CALL: return "X86ISD::CALL";
6979 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6980 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6981 case X86ISD::BT: return "X86ISD::BT";
6982 case X86ISD::CMP: return "X86ISD::CMP";
6983 case X86ISD::COMI: return "X86ISD::COMI";
6984 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6985 case X86ISD::SETCC: return "X86ISD::SETCC";
6986 case X86ISD::CMOV: return "X86ISD::CMOV";
6987 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6988 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6989 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6990 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6991 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6992 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6993 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
6994 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6995 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6996 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6997 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6998 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6999 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7000 case X86ISD::FMAX: return "X86ISD::FMAX";
7001 case X86ISD::FMIN: return "X86ISD::FMIN";
7002 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7003 case X86ISD::FRCP: return "X86ISD::FRCP";
7004 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7005 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7006 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7007 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7008 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7009 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7010 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7011 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7012 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7013 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7014 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7015 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7016 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7017 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7018 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7019 case X86ISD::VSHL: return "X86ISD::VSHL";
7020 case X86ISD::VSRL: return "X86ISD::VSRL";
7021 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7022 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7023 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7024 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7025 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7026 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7027 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7028 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7029 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7030 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7031 case X86ISD::ADD: return "X86ISD::ADD";
7032 case X86ISD::SUB: return "X86ISD::SUB";
7033 case X86ISD::SMUL: return "X86ISD::SMUL";
7034 case X86ISD::UMUL: return "X86ISD::UMUL";
7035 case X86ISD::INC: return "X86ISD::INC";
7036 case X86ISD::DEC: return "X86ISD::DEC";
7037 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7041 // isLegalAddressingMode - Return true if the addressing mode represented
7042 // by AM is legal for this target, for a load/store of the specified type.
7043 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7044 const Type *Ty) const {
7045 // X86 supports extremely general addressing modes.
7047 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7048 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7052 // We can only fold this if we don't need an extra load.
7053 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
7055 // If BaseGV requires a register, we cannot also have a BaseReg.
7056 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
7060 // X86-64 only supports addr of globals in small code model.
7061 if (Subtarget->is64Bit()) {
7062 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7064 // If lower 4G is not available, then we must use rip-relative addressing.
7065 if (AM.BaseOffs || AM.Scale > 1)
7076 // These scales always work.
7081 // These scales are formed with basereg+scalereg. Only accept if there is
7086 default: // Other stuff never works.
7094 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7095 if (!Ty1->isInteger() || !Ty2->isInteger())
7097 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7098 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7099 if (NumBits1 <= NumBits2)
7101 return Subtarget->is64Bit() || NumBits1 < 64;
7104 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7105 if (!VT1.isInteger() || !VT2.isInteger())
7107 unsigned NumBits1 = VT1.getSizeInBits();
7108 unsigned NumBits2 = VT2.getSizeInBits();
7109 if (NumBits1 <= NumBits2)
7111 return Subtarget->is64Bit() || NumBits1 < 64;
7114 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7115 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7116 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7119 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7120 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7121 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7124 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
7125 // i16 instructions are longer (0x66 prefix) and potentially slower.
7126 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7129 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7130 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7131 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7132 /// are assumed to be legal.
7134 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7136 // Only do shuffles on 128-bit vector types for now.
7137 if (VT.getSizeInBits() == 64)
7140 // FIXME: pshufb, blends, palignr, shifts.
7141 return (VT.getVectorNumElements() == 2 ||
7142 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7143 isMOVLMask(M, VT) ||
7144 isSHUFPMask(M, VT) ||
7145 isPSHUFDMask(M, VT) ||
7146 isPSHUFHWMask(M, VT) ||
7147 isPSHUFLWMask(M, VT) ||
7148 isUNPCKLMask(M, VT) ||
7149 isUNPCKHMask(M, VT) ||
7150 isUNPCKL_v_undef_Mask(M, VT) ||
7151 isUNPCKH_v_undef_Mask(M, VT));
7155 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7157 unsigned NumElts = VT.getVectorNumElements();
7158 // FIXME: This collection of masks seems suspect.
7161 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7162 return (isMOVLMask(Mask, VT) ||
7163 isCommutedMOVLMask(Mask, VT, true) ||
7164 isSHUFPMask(Mask, VT) ||
7165 isCommutedSHUFPMask(Mask, VT));
7170 //===----------------------------------------------------------------------===//
7171 // X86 Scheduler Hooks
7172 //===----------------------------------------------------------------------===//
7174 // private utility function
7176 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7177 MachineBasicBlock *MBB,
7185 TargetRegisterClass *RC,
7186 bool invSrc) const {
7187 // For the atomic bitwise operator, we generate
7190 // ld t1 = [bitinstr.addr]
7191 // op t2 = t1, [bitinstr.val]
7193 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7195 // fallthrough -->nextMBB
7196 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7197 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7198 MachineFunction::iterator MBBIter = MBB;
7201 /// First build the CFG
7202 MachineFunction *F = MBB->getParent();
7203 MachineBasicBlock *thisMBB = MBB;
7204 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7205 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7206 F->insert(MBBIter, newMBB);
7207 F->insert(MBBIter, nextMBB);
7209 // Move all successors to thisMBB to nextMBB
7210 nextMBB->transferSuccessors(thisMBB);
7212 // Update thisMBB to fall through to newMBB
7213 thisMBB->addSuccessor(newMBB);
7215 // newMBB jumps to itself and fall through to nextMBB
7216 newMBB->addSuccessor(nextMBB);
7217 newMBB->addSuccessor(newMBB);
7219 // Insert instructions into newMBB based on incoming instruction
7220 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7221 "unexpected number of operands");
7222 DebugLoc dl = bInstr->getDebugLoc();
7223 MachineOperand& destOper = bInstr->getOperand(0);
7224 MachineOperand* argOpers[2 + X86AddrNumOperands];
7225 int numArgs = bInstr->getNumOperands() - 1;
7226 for (int i=0; i < numArgs; ++i)
7227 argOpers[i] = &bInstr->getOperand(i+1);
7229 // x86 address has 4 operands: base, index, scale, and displacement
7230 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7231 int valArgIndx = lastAddrIndx + 1;
7233 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7234 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7235 for (int i=0; i <= lastAddrIndx; ++i)
7236 (*MIB).addOperand(*argOpers[i]);
7238 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7240 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7245 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7246 assert((argOpers[valArgIndx]->isReg() ||
7247 argOpers[valArgIndx]->isImm()) &&
7249 if (argOpers[valArgIndx]->isReg())
7250 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7252 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7254 (*MIB).addOperand(*argOpers[valArgIndx]);
7256 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7259 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7260 for (int i=0; i <= lastAddrIndx; ++i)
7261 (*MIB).addOperand(*argOpers[i]);
7263 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7264 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7266 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7270 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7272 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7276 // private utility function: 64 bit atomics on 32 bit host.
7278 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7279 MachineBasicBlock *MBB,
7284 bool invSrc) const {
7285 // For the atomic bitwise operator, we generate
7286 // thisMBB (instructions are in pairs, except cmpxchg8b)
7287 // ld t1,t2 = [bitinstr.addr]
7289 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7290 // op t5, t6 <- out1, out2, [bitinstr.val]
7291 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7292 // mov ECX, EBX <- t5, t6
7293 // mov EAX, EDX <- t1, t2
7294 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7295 // mov t3, t4 <- EAX, EDX
7297 // result in out1, out2
7298 // fallthrough -->nextMBB
7300 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7301 const unsigned LoadOpc = X86::MOV32rm;
7302 const unsigned copyOpc = X86::MOV32rr;
7303 const unsigned NotOpc = X86::NOT32r;
7304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7305 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7306 MachineFunction::iterator MBBIter = MBB;
7309 /// First build the CFG
7310 MachineFunction *F = MBB->getParent();
7311 MachineBasicBlock *thisMBB = MBB;
7312 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7313 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7314 F->insert(MBBIter, newMBB);
7315 F->insert(MBBIter, nextMBB);
7317 // Move all successors to thisMBB to nextMBB
7318 nextMBB->transferSuccessors(thisMBB);
7320 // Update thisMBB to fall through to newMBB
7321 thisMBB->addSuccessor(newMBB);
7323 // newMBB jumps to itself and fall through to nextMBB
7324 newMBB->addSuccessor(nextMBB);
7325 newMBB->addSuccessor(newMBB);
7327 DebugLoc dl = bInstr->getDebugLoc();
7328 // Insert instructions into newMBB based on incoming instruction
7329 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7330 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7331 "unexpected number of operands");
7332 MachineOperand& dest1Oper = bInstr->getOperand(0);
7333 MachineOperand& dest2Oper = bInstr->getOperand(1);
7334 MachineOperand* argOpers[2 + X86AddrNumOperands];
7335 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7336 argOpers[i] = &bInstr->getOperand(i+2);
7338 // x86 address has 4 operands: base, index, scale, and displacement
7339 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7341 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7342 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7343 for (int i=0; i <= lastAddrIndx; ++i)
7344 (*MIB).addOperand(*argOpers[i]);
7345 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7346 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7347 // add 4 to displacement.
7348 for (int i=0; i <= lastAddrIndx-2; ++i)
7349 (*MIB).addOperand(*argOpers[i]);
7350 MachineOperand newOp3 = *(argOpers[3]);
7352 newOp3.setImm(newOp3.getImm()+4);
7354 newOp3.setOffset(newOp3.getOffset()+4);
7355 (*MIB).addOperand(newOp3);
7356 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7358 // t3/4 are defined later, at the bottom of the loop
7359 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7360 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7361 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7362 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7363 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7364 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7366 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7367 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7369 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7370 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7376 int valArgIndx = lastAddrIndx + 1;
7377 assert((argOpers[valArgIndx]->isReg() ||
7378 argOpers[valArgIndx]->isImm()) &&
7380 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7381 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7382 if (argOpers[valArgIndx]->isReg())
7383 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7385 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7386 if (regOpcL != X86::MOV32rr)
7388 (*MIB).addOperand(*argOpers[valArgIndx]);
7389 assert(argOpers[valArgIndx + 1]->isReg() ==
7390 argOpers[valArgIndx]->isReg());
7391 assert(argOpers[valArgIndx + 1]->isImm() ==
7392 argOpers[valArgIndx]->isImm());
7393 if (argOpers[valArgIndx + 1]->isReg())
7394 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7396 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7397 if (regOpcH != X86::MOV32rr)
7399 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7401 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7403 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7406 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7408 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7411 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7412 for (int i=0; i <= lastAddrIndx; ++i)
7413 (*MIB).addOperand(*argOpers[i]);
7415 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7416 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7418 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7419 MIB.addReg(X86::EAX);
7420 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7421 MIB.addReg(X86::EDX);
7424 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7426 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7430 // private utility function
7432 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7433 MachineBasicBlock *MBB,
7434 unsigned cmovOpc) const {
7435 // For the atomic min/max operator, we generate
7438 // ld t1 = [min/max.addr]
7439 // mov t2 = [min/max.val]
7441 // cmov[cond] t2 = t1
7443 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7445 // fallthrough -->nextMBB
7447 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7448 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7449 MachineFunction::iterator MBBIter = MBB;
7452 /// First build the CFG
7453 MachineFunction *F = MBB->getParent();
7454 MachineBasicBlock *thisMBB = MBB;
7455 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7456 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7457 F->insert(MBBIter, newMBB);
7458 F->insert(MBBIter, nextMBB);
7460 // Move all successors to thisMBB to nextMBB
7461 nextMBB->transferSuccessors(thisMBB);
7463 // Update thisMBB to fall through to newMBB
7464 thisMBB->addSuccessor(newMBB);
7466 // newMBB jumps to newMBB and fall through to nextMBB
7467 newMBB->addSuccessor(nextMBB);
7468 newMBB->addSuccessor(newMBB);
7470 DebugLoc dl = mInstr->getDebugLoc();
7471 // Insert instructions into newMBB based on incoming instruction
7472 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7473 "unexpected number of operands");
7474 MachineOperand& destOper = mInstr->getOperand(0);
7475 MachineOperand* argOpers[2 + X86AddrNumOperands];
7476 int numArgs = mInstr->getNumOperands() - 1;
7477 for (int i=0; i < numArgs; ++i)
7478 argOpers[i] = &mInstr->getOperand(i+1);
7480 // x86 address has 4 operands: base, index, scale, and displacement
7481 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7482 int valArgIndx = lastAddrIndx + 1;
7484 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7485 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7486 for (int i=0; i <= lastAddrIndx; ++i)
7487 (*MIB).addOperand(*argOpers[i]);
7489 // We only support register and immediate values
7490 assert((argOpers[valArgIndx]->isReg() ||
7491 argOpers[valArgIndx]->isImm()) &&
7494 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7495 if (argOpers[valArgIndx]->isReg())
7496 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7498 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7499 (*MIB).addOperand(*argOpers[valArgIndx]);
7501 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7504 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7509 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7510 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7514 // Cmp and exchange if none has modified the memory location
7515 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7516 for (int i=0; i <= lastAddrIndx; ++i)
7517 (*MIB).addOperand(*argOpers[i]);
7519 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7520 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7522 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7523 MIB.addReg(X86::EAX);
7526 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7528 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7534 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7535 MachineBasicBlock *BB) const {
7536 DebugLoc dl = MI->getDebugLoc();
7537 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7538 switch (MI->getOpcode()) {
7539 default: assert(false && "Unexpected instr type to insert");
7540 case X86::CMOV_V1I64:
7541 case X86::CMOV_FR32:
7542 case X86::CMOV_FR64:
7543 case X86::CMOV_V4F32:
7544 case X86::CMOV_V2F64:
7545 case X86::CMOV_V2I64: {
7546 // To "insert" a SELECT_CC instruction, we actually have to insert the
7547 // diamond control-flow pattern. The incoming instruction knows the
7548 // destination vreg to set, the condition code register to branch on, the
7549 // true/false values to select between, and a branch opcode to use.
7550 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7551 MachineFunction::iterator It = BB;
7557 // cmpTY ccX, r1, r2
7559 // fallthrough --> copy0MBB
7560 MachineBasicBlock *thisMBB = BB;
7561 MachineFunction *F = BB->getParent();
7562 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7563 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7565 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7566 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7567 F->insert(It, copy0MBB);
7568 F->insert(It, sinkMBB);
7569 // Update machine-CFG edges by transferring all successors of the current
7570 // block to the new block which will contain the Phi node for the select.
7571 sinkMBB->transferSuccessors(BB);
7573 // Add the true and fallthrough blocks as its successors.
7574 BB->addSuccessor(copy0MBB);
7575 BB->addSuccessor(sinkMBB);
7578 // %FalseValue = ...
7579 // # fallthrough to sinkMBB
7582 // Update machine-CFG edges
7583 BB->addSuccessor(sinkMBB);
7586 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7589 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7590 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7591 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7593 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7597 case X86::FP32_TO_INT16_IN_MEM:
7598 case X86::FP32_TO_INT32_IN_MEM:
7599 case X86::FP32_TO_INT64_IN_MEM:
7600 case X86::FP64_TO_INT16_IN_MEM:
7601 case X86::FP64_TO_INT32_IN_MEM:
7602 case X86::FP64_TO_INT64_IN_MEM:
7603 case X86::FP80_TO_INT16_IN_MEM:
7604 case X86::FP80_TO_INT32_IN_MEM:
7605 case X86::FP80_TO_INT64_IN_MEM: {
7606 // Change the floating point control register to use "round towards zero"
7607 // mode when truncating to an integer value.
7608 MachineFunction *F = BB->getParent();
7609 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7610 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7612 // Load the old value of the high byte of the control word...
7614 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7615 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7618 // Set the high part to be round to zero...
7619 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7622 // Reload the modified control word now...
7623 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7625 // Restore the memory image of control word to original value
7626 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7629 // Get the X86 opcode to use.
7631 switch (MI->getOpcode()) {
7632 default: assert(0 && "illegal opcode!");
7633 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7634 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7635 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7636 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7637 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7638 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7639 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7640 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7641 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7645 MachineOperand &Op = MI->getOperand(0);
7647 AM.BaseType = X86AddressMode::RegBase;
7648 AM.Base.Reg = Op.getReg();
7650 AM.BaseType = X86AddressMode::FrameIndexBase;
7651 AM.Base.FrameIndex = Op.getIndex();
7653 Op = MI->getOperand(1);
7655 AM.Scale = Op.getImm();
7656 Op = MI->getOperand(2);
7658 AM.IndexReg = Op.getImm();
7659 Op = MI->getOperand(3);
7660 if (Op.isGlobal()) {
7661 AM.GV = Op.getGlobal();
7663 AM.Disp = Op.getImm();
7665 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7666 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7668 // Reload the original control word now.
7669 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7671 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7674 case X86::ATOMAND32:
7675 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7676 X86::AND32ri, X86::MOV32rm,
7677 X86::LCMPXCHG32, X86::MOV32rr,
7678 X86::NOT32r, X86::EAX,
7679 X86::GR32RegisterClass);
7681 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7682 X86::OR32ri, X86::MOV32rm,
7683 X86::LCMPXCHG32, X86::MOV32rr,
7684 X86::NOT32r, X86::EAX,
7685 X86::GR32RegisterClass);
7686 case X86::ATOMXOR32:
7687 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7688 X86::XOR32ri, X86::MOV32rm,
7689 X86::LCMPXCHG32, X86::MOV32rr,
7690 X86::NOT32r, X86::EAX,
7691 X86::GR32RegisterClass);
7692 case X86::ATOMNAND32:
7693 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7694 X86::AND32ri, X86::MOV32rm,
7695 X86::LCMPXCHG32, X86::MOV32rr,
7696 X86::NOT32r, X86::EAX,
7697 X86::GR32RegisterClass, true);
7698 case X86::ATOMMIN32:
7699 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7700 case X86::ATOMMAX32:
7701 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7702 case X86::ATOMUMIN32:
7703 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7704 case X86::ATOMUMAX32:
7705 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7707 case X86::ATOMAND16:
7708 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7709 X86::AND16ri, X86::MOV16rm,
7710 X86::LCMPXCHG16, X86::MOV16rr,
7711 X86::NOT16r, X86::AX,
7712 X86::GR16RegisterClass);
7714 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7715 X86::OR16ri, X86::MOV16rm,
7716 X86::LCMPXCHG16, X86::MOV16rr,
7717 X86::NOT16r, X86::AX,
7718 X86::GR16RegisterClass);
7719 case X86::ATOMXOR16:
7720 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7721 X86::XOR16ri, X86::MOV16rm,
7722 X86::LCMPXCHG16, X86::MOV16rr,
7723 X86::NOT16r, X86::AX,
7724 X86::GR16RegisterClass);
7725 case X86::ATOMNAND16:
7726 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7727 X86::AND16ri, X86::MOV16rm,
7728 X86::LCMPXCHG16, X86::MOV16rr,
7729 X86::NOT16r, X86::AX,
7730 X86::GR16RegisterClass, true);
7731 case X86::ATOMMIN16:
7732 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7733 case X86::ATOMMAX16:
7734 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7735 case X86::ATOMUMIN16:
7736 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7737 case X86::ATOMUMAX16:
7738 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7741 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7742 X86::AND8ri, X86::MOV8rm,
7743 X86::LCMPXCHG8, X86::MOV8rr,
7744 X86::NOT8r, X86::AL,
7745 X86::GR8RegisterClass);
7747 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7748 X86::OR8ri, X86::MOV8rm,
7749 X86::LCMPXCHG8, X86::MOV8rr,
7750 X86::NOT8r, X86::AL,
7751 X86::GR8RegisterClass);
7753 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7754 X86::XOR8ri, X86::MOV8rm,
7755 X86::LCMPXCHG8, X86::MOV8rr,
7756 X86::NOT8r, X86::AL,
7757 X86::GR8RegisterClass);
7758 case X86::ATOMNAND8:
7759 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7760 X86::AND8ri, X86::MOV8rm,
7761 X86::LCMPXCHG8, X86::MOV8rr,
7762 X86::NOT8r, X86::AL,
7763 X86::GR8RegisterClass, true);
7764 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7765 // This group is for 64-bit host.
7766 case X86::ATOMAND64:
7767 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7768 X86::AND64ri32, X86::MOV64rm,
7769 X86::LCMPXCHG64, X86::MOV64rr,
7770 X86::NOT64r, X86::RAX,
7771 X86::GR64RegisterClass);
7773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7774 X86::OR64ri32, X86::MOV64rm,
7775 X86::LCMPXCHG64, X86::MOV64rr,
7776 X86::NOT64r, X86::RAX,
7777 X86::GR64RegisterClass);
7778 case X86::ATOMXOR64:
7779 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7780 X86::XOR64ri32, X86::MOV64rm,
7781 X86::LCMPXCHG64, X86::MOV64rr,
7782 X86::NOT64r, X86::RAX,
7783 X86::GR64RegisterClass);
7784 case X86::ATOMNAND64:
7785 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7786 X86::AND64ri32, X86::MOV64rm,
7787 X86::LCMPXCHG64, X86::MOV64rr,
7788 X86::NOT64r, X86::RAX,
7789 X86::GR64RegisterClass, true);
7790 case X86::ATOMMIN64:
7791 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7792 case X86::ATOMMAX64:
7793 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7794 case X86::ATOMUMIN64:
7795 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7796 case X86::ATOMUMAX64:
7797 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7799 // This group does 64-bit operations on a 32-bit host.
7800 case X86::ATOMAND6432:
7801 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7802 X86::AND32rr, X86::AND32rr,
7803 X86::AND32ri, X86::AND32ri,
7805 case X86::ATOMOR6432:
7806 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7807 X86::OR32rr, X86::OR32rr,
7808 X86::OR32ri, X86::OR32ri,
7810 case X86::ATOMXOR6432:
7811 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7812 X86::XOR32rr, X86::XOR32rr,
7813 X86::XOR32ri, X86::XOR32ri,
7815 case X86::ATOMNAND6432:
7816 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7817 X86::AND32rr, X86::AND32rr,
7818 X86::AND32ri, X86::AND32ri,
7820 case X86::ATOMADD6432:
7821 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7822 X86::ADD32rr, X86::ADC32rr,
7823 X86::ADD32ri, X86::ADC32ri,
7825 case X86::ATOMSUB6432:
7826 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7827 X86::SUB32rr, X86::SBB32rr,
7828 X86::SUB32ri, X86::SBB32ri,
7830 case X86::ATOMSWAP6432:
7831 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7832 X86::MOV32rr, X86::MOV32rr,
7833 X86::MOV32ri, X86::MOV32ri,
7838 //===----------------------------------------------------------------------===//
7839 // X86 Optimization Hooks
7840 //===----------------------------------------------------------------------===//
7842 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7846 const SelectionDAG &DAG,
7847 unsigned Depth) const {
7848 unsigned Opc = Op.getOpcode();
7849 assert((Opc >= ISD::BUILTIN_OP_END ||
7850 Opc == ISD::INTRINSIC_WO_CHAIN ||
7851 Opc == ISD::INTRINSIC_W_CHAIN ||
7852 Opc == ISD::INTRINSIC_VOID) &&
7853 "Should use MaskedValueIsZero if you don't know whether Op"
7854 " is a target node!");
7856 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7865 // These nodes' second result is a boolean.
7866 if (Op.getResNo() == 0)
7870 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7871 Mask.getBitWidth() - 1);
7876 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7877 /// node is a GlobalAddress + offset.
7878 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7879 GlobalValue* &GA, int64_t &Offset) const{
7880 if (N->getOpcode() == X86ISD::Wrapper) {
7881 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7882 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7883 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7887 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7890 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7891 const TargetLowering &TLI) {
7894 if (TLI.isGAPlusOffset(Base, GV, Offset))
7895 return (GV->getAlignment() >= N && (Offset % N) == 0);
7896 // DAG combine handles the stack object case.
7900 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7901 MVT EVT, LoadSDNode *&LDBase,
7902 unsigned &LastLoadedElt,
7903 SelectionDAG &DAG, MachineFrameInfo *MFI,
7904 const TargetLowering &TLI) {
7906 LastLoadedElt = -1U;
7907 for (unsigned i = 0; i < NumElems; ++i) {
7908 if (N->getMaskElt(i) < 0) {
7914 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7915 if (!Elt.getNode() ||
7916 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7919 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7921 LDBase = cast<LoadSDNode>(Elt.getNode());
7925 if (Elt.getOpcode() == ISD::UNDEF)
7928 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7929 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7936 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7937 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7938 /// if the load addresses are consecutive, non-overlapping, and in the right
7939 /// order. In the case of v2i64, it will see if it can rewrite the
7940 /// shuffle to be an appropriate build vector so it can take advantage of
7941 // performBuildVectorCombine.
7942 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7943 const TargetLowering &TLI) {
7944 DebugLoc dl = N->getDebugLoc();
7945 MVT VT = N->getValueType(0);
7946 MVT EVT = VT.getVectorElementType();
7947 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7948 unsigned NumElems = VT.getVectorNumElements();
7950 if (VT.getSizeInBits() != 128)
7953 // Try to combine a vector_shuffle into a 128-bit load.
7954 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7955 LoadSDNode *LD = NULL;
7956 unsigned LastLoadedElt;
7957 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7961 if (LastLoadedElt == NumElems - 1) {
7962 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7963 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7964 LD->getSrcValue(), LD->getSrcValueOffset(),
7966 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7967 LD->getSrcValue(), LD->getSrcValueOffset(),
7968 LD->isVolatile(), LD->getAlignment());
7969 } else if (NumElems == 4 && LastLoadedElt == 1) {
7970 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7971 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7972 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7973 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
7978 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7979 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7980 const X86Subtarget *Subtarget) {
7981 DebugLoc DL = N->getDebugLoc();
7982 SDValue Cond = N->getOperand(0);
7983 // Get the LHS/RHS of the select.
7984 SDValue LHS = N->getOperand(1);
7985 SDValue RHS = N->getOperand(2);
7987 // If we have SSE[12] support, try to form min/max nodes.
7988 if (Subtarget->hasSSE2() &&
7989 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7990 Cond.getOpcode() == ISD::SETCC) {
7991 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7993 unsigned Opcode = 0;
7994 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7997 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8000 if (!UnsafeFPMath) break;
8002 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8004 Opcode = X86ISD::FMIN;
8007 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8010 if (!UnsafeFPMath) break;
8012 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8014 Opcode = X86ISD::FMAX;
8017 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8020 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8023 if (!UnsafeFPMath) break;
8025 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8027 Opcode = X86ISD::FMIN;
8030 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8033 if (!UnsafeFPMath) break;
8035 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8037 Opcode = X86ISD::FMAX;
8043 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8046 // If this is a select between two integer constants, try to do some
8048 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8049 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8050 // Don't do this for crazy integer types.
8051 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8052 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8053 // so that TrueC (the true value) is larger than FalseC.
8054 bool NeedsCondInvert = false;
8056 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8057 // Efficiently invertible.
8058 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8059 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8060 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8061 NeedsCondInvert = true;
8062 std::swap(TrueC, FalseC);
8065 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8066 if (FalseC->getAPIntValue() == 0 &&
8067 TrueC->getAPIntValue().isPowerOf2()) {
8068 if (NeedsCondInvert) // Invert the condition if needed.
8069 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8070 DAG.getConstant(1, Cond.getValueType()));
8072 // Zero extend the condition if needed.
8073 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8075 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8076 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8077 DAG.getConstant(ShAmt, MVT::i8));
8080 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8081 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8082 if (NeedsCondInvert) // Invert the condition if needed.
8083 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8084 DAG.getConstant(1, Cond.getValueType()));
8086 // Zero extend the condition if needed.
8087 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8088 FalseC->getValueType(0), Cond);
8089 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8090 SDValue(FalseC, 0));
8093 // Optimize cases that will turn into an LEA instruction. This requires
8094 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8095 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8096 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8097 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8099 bool isFastMultiplier = false;
8101 switch ((unsigned char)Diff) {
8103 case 1: // result = add base, cond
8104 case 2: // result = lea base( , cond*2)
8105 case 3: // result = lea base(cond, cond*2)
8106 case 4: // result = lea base( , cond*4)
8107 case 5: // result = lea base(cond, cond*4)
8108 case 8: // result = lea base( , cond*8)
8109 case 9: // result = lea base(cond, cond*8)
8110 isFastMultiplier = true;
8115 if (isFastMultiplier) {
8116 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8117 if (NeedsCondInvert) // Invert the condition if needed.
8118 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8119 DAG.getConstant(1, Cond.getValueType()));
8121 // Zero extend the condition if needed.
8122 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8124 // Scale the condition by the difference.
8126 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8127 DAG.getConstant(Diff, Cond.getValueType()));
8129 // Add the base if non-zero.
8130 if (FalseC->getAPIntValue() != 0)
8131 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8132 SDValue(FalseC, 0));
8142 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8143 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8144 TargetLowering::DAGCombinerInfo &DCI) {
8145 DebugLoc DL = N->getDebugLoc();
8147 // If the flag operand isn't dead, don't touch this CMOV.
8148 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8151 // If this is a select between two integer constants, try to do some
8152 // optimizations. Note that the operands are ordered the opposite of SELECT
8154 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8155 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8156 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8157 // larger than FalseC (the false value).
8158 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8160 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8161 CC = X86::GetOppositeBranchCondition(CC);
8162 std::swap(TrueC, FalseC);
8165 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8166 // This is efficient for any integer data type (including i8/i16) and
8168 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8169 SDValue Cond = N->getOperand(3);
8170 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8171 DAG.getConstant(CC, MVT::i8), Cond);
8173 // Zero extend the condition if needed.
8174 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8176 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8177 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8178 DAG.getConstant(ShAmt, MVT::i8));
8179 if (N->getNumValues() == 2) // Dead flag value?
8180 return DCI.CombineTo(N, Cond, SDValue());
8184 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8185 // for any integer data type, including i8/i16.
8186 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8187 SDValue Cond = N->getOperand(3);
8188 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8189 DAG.getConstant(CC, MVT::i8), Cond);
8191 // Zero extend the condition if needed.
8192 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8193 FalseC->getValueType(0), Cond);
8194 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8195 SDValue(FalseC, 0));
8197 if (N->getNumValues() == 2) // Dead flag value?
8198 return DCI.CombineTo(N, Cond, SDValue());
8202 // Optimize cases that will turn into an LEA instruction. This requires
8203 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8204 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8205 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8206 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8208 bool isFastMultiplier = false;
8210 switch ((unsigned char)Diff) {
8212 case 1: // result = add base, cond
8213 case 2: // result = lea base( , cond*2)
8214 case 3: // result = lea base(cond, cond*2)
8215 case 4: // result = lea base( , cond*4)
8216 case 5: // result = lea base(cond, cond*4)
8217 case 8: // result = lea base( , cond*8)
8218 case 9: // result = lea base(cond, cond*8)
8219 isFastMultiplier = true;
8224 if (isFastMultiplier) {
8225 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8226 SDValue Cond = N->getOperand(3);
8227 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8228 DAG.getConstant(CC, MVT::i8), Cond);
8229 // Zero extend the condition if needed.
8230 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8232 // Scale the condition by the difference.
8234 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8235 DAG.getConstant(Diff, Cond.getValueType()));
8237 // Add the base if non-zero.
8238 if (FalseC->getAPIntValue() != 0)
8239 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8240 SDValue(FalseC, 0));
8241 if (N->getNumValues() == 2) // Dead flag value?
8242 return DCI.CombineTo(N, Cond, SDValue());
8252 /// PerformMulCombine - Optimize a single multiply with constant into two
8253 /// in order to implement it with two cheaper instructions, e.g.
8254 /// LEA + SHL, LEA + LEA.
8255 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8256 TargetLowering::DAGCombinerInfo &DCI) {
8257 if (DAG.getMachineFunction().
8258 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8261 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8264 MVT VT = N->getValueType(0);
8268 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8271 uint64_t MulAmt = C->getZExtValue();
8272 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8275 uint64_t MulAmt1 = 0;
8276 uint64_t MulAmt2 = 0;
8277 if ((MulAmt % 9) == 0) {
8279 MulAmt2 = MulAmt / 9;
8280 } else if ((MulAmt % 5) == 0) {
8282 MulAmt2 = MulAmt / 5;
8283 } else if ((MulAmt % 3) == 0) {
8285 MulAmt2 = MulAmt / 3;
8288 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8289 DebugLoc DL = N->getDebugLoc();
8291 if (isPowerOf2_64(MulAmt2) &&
8292 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8293 // If second multiplifer is pow2, issue it first. We want the multiply by
8294 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8296 std::swap(MulAmt1, MulAmt2);
8299 if (isPowerOf2_64(MulAmt1))
8300 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8301 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8303 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8304 DAG.getConstant(MulAmt1, VT));
8306 if (isPowerOf2_64(MulAmt2))
8307 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8308 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8310 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8311 DAG.getConstant(MulAmt2, VT));
8313 // Do not add new nodes to DAG combiner worklist.
8314 DCI.CombineTo(N, NewMul, false);
8320 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8322 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8323 const X86Subtarget *Subtarget) {
8324 // On X86 with SSE2 support, we can transform this to a vector shift if
8325 // all elements are shifted by the same amount. We can't do this in legalize
8326 // because the a constant vector is typically transformed to a constant pool
8327 // so we have no knowledge of the shift amount.
8328 if (!Subtarget->hasSSE2())
8331 MVT VT = N->getValueType(0);
8332 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8335 SDValue ShAmtOp = N->getOperand(1);
8336 MVT EltVT = VT.getVectorElementType();
8337 DebugLoc DL = N->getDebugLoc();
8339 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8340 unsigned NumElts = VT.getVectorNumElements();
8342 for (; i != NumElts; ++i) {
8343 SDValue Arg = ShAmtOp.getOperand(i);
8344 if (Arg.getOpcode() == ISD::UNDEF) continue;
8348 for (; i != NumElts; ++i) {
8349 SDValue Arg = ShAmtOp.getOperand(i);
8350 if (Arg.getOpcode() == ISD::UNDEF) continue;
8351 if (Arg != BaseShAmt) {
8355 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8356 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8357 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8358 DAG.getIntPtrConstant(0));
8362 if (EltVT.bitsGT(MVT::i32))
8363 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8364 else if (EltVT.bitsLT(MVT::i32))
8365 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8367 // The shift amount is identical so we can do a vector shift.
8368 SDValue ValOp = N->getOperand(0);
8369 switch (N->getOpcode()) {
8371 assert(0 && "Unknown shift opcode!");
8374 if (VT == MVT::v2i64)
8375 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8376 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8378 if (VT == MVT::v4i32)
8379 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8380 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8382 if (VT == MVT::v8i16)
8383 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8384 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8388 if (VT == MVT::v4i32)
8389 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8390 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8392 if (VT == MVT::v8i16)
8393 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8394 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8398 if (VT == MVT::v2i64)
8399 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8400 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8402 if (VT == MVT::v4i32)
8403 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8404 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8406 if (VT == MVT::v8i16)
8407 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8408 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8415 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8416 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8417 const X86Subtarget *Subtarget) {
8418 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8419 // the FP state in cases where an emms may be missing.
8420 // A preferable solution to the general problem is to figure out the right
8421 // places to insert EMMS. This qualifies as a quick hack.
8423 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8424 StoreSDNode *St = cast<StoreSDNode>(N);
8425 MVT VT = St->getValue().getValueType();
8426 if (VT.getSizeInBits() != 64)
8429 const Function *F = DAG.getMachineFunction().getFunction();
8430 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8431 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8432 && Subtarget->hasSSE2();
8433 if ((VT.isVector() ||
8434 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8435 isa<LoadSDNode>(St->getValue()) &&
8436 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8437 St->getChain().hasOneUse() && !St->isVolatile()) {
8438 SDNode* LdVal = St->getValue().getNode();
8440 int TokenFactorIndex = -1;
8441 SmallVector<SDValue, 8> Ops;
8442 SDNode* ChainVal = St->getChain().getNode();
8443 // Must be a store of a load. We currently handle two cases: the load
8444 // is a direct child, and it's under an intervening TokenFactor. It is
8445 // possible to dig deeper under nested TokenFactors.
8446 if (ChainVal == LdVal)
8447 Ld = cast<LoadSDNode>(St->getChain());
8448 else if (St->getValue().hasOneUse() &&
8449 ChainVal->getOpcode() == ISD::TokenFactor) {
8450 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8451 if (ChainVal->getOperand(i).getNode() == LdVal) {
8452 TokenFactorIndex = i;
8453 Ld = cast<LoadSDNode>(St->getValue());
8455 Ops.push_back(ChainVal->getOperand(i));
8459 if (!Ld || !ISD::isNormalLoad(Ld))
8462 // If this is not the MMX case, i.e. we are just turning i64 load/store
8463 // into f64 load/store, avoid the transformation if there are multiple
8464 // uses of the loaded value.
8465 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8468 DebugLoc LdDL = Ld->getDebugLoc();
8469 DebugLoc StDL = N->getDebugLoc();
8470 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8471 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8473 if (Subtarget->is64Bit() || F64IsLegal) {
8474 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8475 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8476 Ld->getBasePtr(), Ld->getSrcValue(),
8477 Ld->getSrcValueOffset(), Ld->isVolatile(),
8478 Ld->getAlignment());
8479 SDValue NewChain = NewLd.getValue(1);
8480 if (TokenFactorIndex != -1) {
8481 Ops.push_back(NewChain);
8482 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8485 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8486 St->getSrcValue(), St->getSrcValueOffset(),
8487 St->isVolatile(), St->getAlignment());
8490 // Otherwise, lower to two pairs of 32-bit loads / stores.
8491 SDValue LoAddr = Ld->getBasePtr();
8492 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8493 DAG.getConstant(4, MVT::i32));
8495 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8496 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8497 Ld->isVolatile(), Ld->getAlignment());
8498 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8499 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8501 MinAlign(Ld->getAlignment(), 4));
8503 SDValue NewChain = LoLd.getValue(1);
8504 if (TokenFactorIndex != -1) {
8505 Ops.push_back(LoLd);
8506 Ops.push_back(HiLd);
8507 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8511 LoAddr = St->getBasePtr();
8512 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8513 DAG.getConstant(4, MVT::i32));
8515 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8516 St->getSrcValue(), St->getSrcValueOffset(),
8517 St->isVolatile(), St->getAlignment());
8518 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8520 St->getSrcValueOffset() + 4,
8522 MinAlign(St->getAlignment(), 4));
8523 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8528 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8529 /// X86ISD::FXOR nodes.
8530 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8531 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8532 // F[X]OR(0.0, x) -> x
8533 // F[X]OR(x, 0.0) -> x
8534 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8535 if (C->getValueAPF().isPosZero())
8536 return N->getOperand(1);
8537 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8538 if (C->getValueAPF().isPosZero())
8539 return N->getOperand(0);
8543 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8544 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8545 // FAND(0.0, x) -> 0.0
8546 // FAND(x, 0.0) -> 0.0
8547 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8548 if (C->getValueAPF().isPosZero())
8549 return N->getOperand(0);
8550 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8551 if (C->getValueAPF().isPosZero())
8552 return N->getOperand(1);
8556 static SDValue PerformBTCombine(SDNode *N,
8558 TargetLowering::DAGCombinerInfo &DCI) {
8559 // BT ignores high bits in the bit index operand.
8560 SDValue Op1 = N->getOperand(1);
8561 if (Op1.hasOneUse()) {
8562 unsigned BitWidth = Op1.getValueSizeInBits();
8563 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8564 APInt KnownZero, KnownOne;
8565 TargetLowering::TargetLoweringOpt TLO(DAG);
8566 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8567 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8568 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8569 DCI.CommitTargetLoweringOpt(TLO);
8574 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8575 SDValue Op = N->getOperand(0);
8576 if (Op.getOpcode() == ISD::BIT_CONVERT)
8577 Op = Op.getOperand(0);
8578 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8579 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8580 VT.getVectorElementType().getSizeInBits() ==
8581 OpVT.getVectorElementType().getSizeInBits()) {
8582 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8587 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8588 // Locked instructions, in turn, have implicit fence semantics (all memory
8589 // operations are flushed before issuing the locked instruction, and the
8590 // are not buffered), so we can fold away the common pattern of
8591 // fence-atomic-fence.
8592 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8593 SDValue atomic = N->getOperand(0);
8594 switch (atomic.getOpcode()) {
8595 case ISD::ATOMIC_CMP_SWAP:
8596 case ISD::ATOMIC_SWAP:
8597 case ISD::ATOMIC_LOAD_ADD:
8598 case ISD::ATOMIC_LOAD_SUB:
8599 case ISD::ATOMIC_LOAD_AND:
8600 case ISD::ATOMIC_LOAD_OR:
8601 case ISD::ATOMIC_LOAD_XOR:
8602 case ISD::ATOMIC_LOAD_NAND:
8603 case ISD::ATOMIC_LOAD_MIN:
8604 case ISD::ATOMIC_LOAD_MAX:
8605 case ISD::ATOMIC_LOAD_UMIN:
8606 case ISD::ATOMIC_LOAD_UMAX:
8612 SDValue fence = atomic.getOperand(0);
8613 if (fence.getOpcode() != ISD::MEMBARRIER)
8616 switch (atomic.getOpcode()) {
8617 case ISD::ATOMIC_CMP_SWAP:
8618 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8619 atomic.getOperand(1), atomic.getOperand(2),
8620 atomic.getOperand(3));
8621 case ISD::ATOMIC_SWAP:
8622 case ISD::ATOMIC_LOAD_ADD:
8623 case ISD::ATOMIC_LOAD_SUB:
8624 case ISD::ATOMIC_LOAD_AND:
8625 case ISD::ATOMIC_LOAD_OR:
8626 case ISD::ATOMIC_LOAD_XOR:
8627 case ISD::ATOMIC_LOAD_NAND:
8628 case ISD::ATOMIC_LOAD_MIN:
8629 case ISD::ATOMIC_LOAD_MAX:
8630 case ISD::ATOMIC_LOAD_UMIN:
8631 case ISD::ATOMIC_LOAD_UMAX:
8632 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8633 atomic.getOperand(1), atomic.getOperand(2));
8639 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8640 DAGCombinerInfo &DCI) const {
8641 SelectionDAG &DAG = DCI.DAG;
8642 switch (N->getOpcode()) {
8644 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8645 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8646 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8647 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8650 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8651 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8653 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8654 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8655 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8656 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8657 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8663 //===----------------------------------------------------------------------===//
8664 // X86 Inline Assembly Support
8665 //===----------------------------------------------------------------------===//
8667 /// getConstraintType - Given a constraint letter, return the type of
8668 /// constraint it is for this target.
8669 X86TargetLowering::ConstraintType
8670 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8671 if (Constraint.size() == 1) {
8672 switch (Constraint[0]) {
8684 return C_RegisterClass;
8692 return TargetLowering::getConstraintType(Constraint);
8695 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8696 /// with another that has more specific requirements based on the type of the
8697 /// corresponding operand.
8698 const char *X86TargetLowering::
8699 LowerXConstraint(MVT ConstraintVT) const {
8700 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8701 // 'f' like normal targets.
8702 if (ConstraintVT.isFloatingPoint()) {
8703 if (Subtarget->hasSSE2())
8705 if (Subtarget->hasSSE1())
8709 return TargetLowering::LowerXConstraint(ConstraintVT);
8712 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8713 /// vector. If it is invalid, don't add anything to Ops.
8714 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8717 std::vector<SDValue>&Ops,
8718 SelectionDAG &DAG) const {
8719 SDValue Result(0, 0);
8721 switch (Constraint) {
8724 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8725 if (C->getZExtValue() <= 31) {
8726 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8732 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8733 if (C->getZExtValue() <= 63) {
8734 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8740 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8741 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8742 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8748 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8749 if (C->getZExtValue() <= 255) {
8750 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8756 // 32-bit signed value
8757 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8758 const ConstantInt *CI = C->getConstantIntValue();
8759 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8760 // Widen to 64 bits here to get it sign extended.
8761 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8764 // FIXME gcc accepts some relocatable values here too, but only in certain
8765 // memory models; it's complicated.
8770 // 32-bit unsigned value
8771 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8772 const ConstantInt *CI = C->getConstantIntValue();
8773 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8774 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8778 // FIXME gcc accepts some relocatable values here too, but only in certain
8779 // memory models; it's complicated.
8783 // Literal immediates are always ok.
8784 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8785 // Widen to 64 bits here to get it sign extended.
8786 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8790 // If we are in non-pic codegen mode, we allow the address of a global (with
8791 // an optional displacement) to be used with 'i'.
8792 GlobalAddressSDNode *GA = 0;
8795 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8797 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8798 Offset += GA->getOffset();
8800 } else if (Op.getOpcode() == ISD::ADD) {
8801 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8802 Offset += C->getZExtValue();
8803 Op = Op.getOperand(0);
8806 } else if (Op.getOpcode() == ISD::SUB) {
8807 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8808 Offset += -C->getZExtValue();
8809 Op = Op.getOperand(0);
8814 // Otherwise, this isn't something we can handle, reject it.
8817 // If we require an extra load to get this address, as in PIC mode, we
8819 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(),
8820 getTargetMachine(), false))
8824 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
8826 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8833 if (Result.getNode()) {
8834 Ops.push_back(Result);
8837 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8841 std::vector<unsigned> X86TargetLowering::
8842 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8844 if (Constraint.size() == 1) {
8845 // FIXME: not handling fp-stack yet!
8846 switch (Constraint[0]) { // GCC X86 Constraint Letters
8847 default: break; // Unknown constraint letter
8848 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8851 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8852 else if (VT == MVT::i16)
8853 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8854 else if (VT == MVT::i8)
8855 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8856 else if (VT == MVT::i64)
8857 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8862 return std::vector<unsigned>();
8865 std::pair<unsigned, const TargetRegisterClass*>
8866 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8868 // First, see if this is a constraint that directly corresponds to an LLVM
8870 if (Constraint.size() == 1) {
8871 // GCC Constraint Letters
8872 switch (Constraint[0]) {
8874 case 'r': // GENERAL_REGS
8875 case 'R': // LEGACY_REGS
8876 case 'l': // INDEX_REGS
8878 return std::make_pair(0U, X86::GR8RegisterClass);
8880 return std::make_pair(0U, X86::GR16RegisterClass);
8881 if (VT == MVT::i32 || !Subtarget->is64Bit())
8882 return std::make_pair(0U, X86::GR32RegisterClass);
8883 return std::make_pair(0U, X86::GR64RegisterClass);
8884 case 'f': // FP Stack registers.
8885 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8886 // value to the correct fpstack register class.
8887 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8888 return std::make_pair(0U, X86::RFP32RegisterClass);
8889 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8890 return std::make_pair(0U, X86::RFP64RegisterClass);
8891 return std::make_pair(0U, X86::RFP80RegisterClass);
8892 case 'y': // MMX_REGS if MMX allowed.
8893 if (!Subtarget->hasMMX()) break;
8894 return std::make_pair(0U, X86::VR64RegisterClass);
8895 case 'Y': // SSE_REGS if SSE2 allowed
8896 if (!Subtarget->hasSSE2()) break;
8898 case 'x': // SSE_REGS if SSE1 allowed
8899 if (!Subtarget->hasSSE1()) break;
8901 switch (VT.getSimpleVT()) {
8903 // Scalar SSE types.
8906 return std::make_pair(0U, X86::FR32RegisterClass);
8909 return std::make_pair(0U, X86::FR64RegisterClass);
8917 return std::make_pair(0U, X86::VR128RegisterClass);
8923 // Use the default implementation in TargetLowering to convert the register
8924 // constraint into a member of a register class.
8925 std::pair<unsigned, const TargetRegisterClass*> Res;
8926 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8928 // Not found as a standard register?
8929 if (Res.second == 0) {
8930 // GCC calls "st(0)" just plain "st".
8931 if (StringsEqualNoCase("{st}", Constraint)) {
8932 Res.first = X86::ST0;
8933 Res.second = X86::RFP80RegisterClass;
8935 // 'A' means EAX + EDX.
8936 if (Constraint == "A") {
8937 Res.first = X86::EAX;
8938 Res.second = X86::GRADRegisterClass;
8943 // Otherwise, check to see if this is a register class of the wrong value
8944 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8945 // turn into {ax},{dx}.
8946 if (Res.second->hasType(VT))
8947 return Res; // Correct type already, nothing to do.
8949 // All of the single-register GCC register classes map their values onto
8950 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8951 // really want an 8-bit or 32-bit register, map to the appropriate register
8952 // class and return the appropriate register.
8953 if (Res.second == X86::GR16RegisterClass) {
8954 if (VT == MVT::i8) {
8955 unsigned DestReg = 0;
8956 switch (Res.first) {
8958 case X86::AX: DestReg = X86::AL; break;
8959 case X86::DX: DestReg = X86::DL; break;
8960 case X86::CX: DestReg = X86::CL; break;
8961 case X86::BX: DestReg = X86::BL; break;
8964 Res.first = DestReg;
8965 Res.second = X86::GR8RegisterClass;
8967 } else if (VT == MVT::i32) {
8968 unsigned DestReg = 0;
8969 switch (Res.first) {
8971 case X86::AX: DestReg = X86::EAX; break;
8972 case X86::DX: DestReg = X86::EDX; break;
8973 case X86::CX: DestReg = X86::ECX; break;
8974 case X86::BX: DestReg = X86::EBX; break;
8975 case X86::SI: DestReg = X86::ESI; break;
8976 case X86::DI: DestReg = X86::EDI; break;
8977 case X86::BP: DestReg = X86::EBP; break;
8978 case X86::SP: DestReg = X86::ESP; break;
8981 Res.first = DestReg;
8982 Res.second = X86::GR32RegisterClass;
8984 } else if (VT == MVT::i64) {
8985 unsigned DestReg = 0;
8986 switch (Res.first) {
8988 case X86::AX: DestReg = X86::RAX; break;
8989 case X86::DX: DestReg = X86::RDX; break;
8990 case X86::CX: DestReg = X86::RCX; break;
8991 case X86::BX: DestReg = X86::RBX; break;
8992 case X86::SI: DestReg = X86::RSI; break;
8993 case X86::DI: DestReg = X86::RDI; break;
8994 case X86::BP: DestReg = X86::RBP; break;
8995 case X86::SP: DestReg = X86::RSP; break;
8998 Res.first = DestReg;
8999 Res.second = X86::GR64RegisterClass;
9002 } else if (Res.second == X86::FR32RegisterClass ||
9003 Res.second == X86::FR64RegisterClass ||
9004 Res.second == X86::VR128RegisterClass) {
9005 // Handle references to XMM physical registers that got mapped into the
9006 // wrong class. This can happen with constraints like {xmm0} where the
9007 // target independent register mapper will just pick the first match it can
9008 // find, ignoring the required type.
9010 Res.second = X86::FR32RegisterClass;
9011 else if (VT == MVT::f64)
9012 Res.second = X86::FR64RegisterClass;
9013 else if (X86::VR128RegisterClass->hasType(VT))
9014 Res.second = X86::VR128RegisterClass;
9020 //===----------------------------------------------------------------------===//
9021 // X86 Widen vector type
9022 //===----------------------------------------------------------------------===//
9024 /// getWidenVectorType: given a vector type, returns the type to widen
9025 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9026 /// If there is no vector type that we want to widen to, returns MVT::Other
9027 /// When and where to widen is target dependent based on the cost of
9028 /// scalarizing vs using the wider vector type.
9030 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9031 assert(VT.isVector());
9032 if (isTypeLegal(VT))
9035 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9036 // type based on element type. This would speed up our search (though
9037 // it may not be worth it since the size of the list is relatively
9039 MVT EltVT = VT.getVectorElementType();
9040 unsigned NElts = VT.getVectorNumElements();
9042 // On X86, it make sense to widen any vector wider than 1
9046 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9047 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9048 MVT SVT = (MVT::SimpleValueType)nVT;
9050 if (isTypeLegal(SVT) &&
9051 SVT.getVectorElementType() == EltVT &&
9052 SVT.getVectorNumElements() > NElts)