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 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Disable16Bit - 16-bit operations typically have a larger encoding than
61 // corresponding 32-bit instructions, and 16-bit code is slow on some
62 // processors. This is an experimental flag to disable 16-bit operations
63 // (which forces them to be Legalized to 32-bit operations).
65 Disable16Bit("disable-16bit", cl::Hidden,
66 cl::desc("Disable use of 16-bit instructions"));
68 // Forward declarations.
69 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
72 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
73 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
74 default: llvm_unreachable("unknown subtarget type");
75 case X86Subtarget::isDarwin:
76 if (TM.getSubtarget<X86Subtarget>().is64Bit())
77 return new X8664_MachoTargetObjectFile();
78 return new TargetLoweringObjectFileMachO();
79 case X86Subtarget::isELF:
80 if (TM.getSubtarget<X86Subtarget>().is64Bit())
81 return new X8664_ELFTargetObjectFile(TM);
82 return new X8632_ELFTargetObjectFile(TM);
83 case X86Subtarget::isMingw:
84 case X86Subtarget::isCygwin:
85 case X86Subtarget::isWindows:
86 return new TargetLoweringObjectFileCOFF();
90 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
91 : TargetLowering(TM, createTLOF(TM)) {
92 Subtarget = &TM.getSubtarget<X86Subtarget>();
93 X86ScalarSSEf64 = Subtarget->hasSSE2();
94 X86ScalarSSEf32 = Subtarget->hasSSE1();
95 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
97 RegInfo = TM.getRegisterInfo();
100 // Set up the TargetLowering object.
102 // X86 is weird, it always uses i8 for shift amounts and setcc results.
103 setShiftAmountType(MVT::i8);
104 setBooleanContents(ZeroOrOneBooleanContent);
105 setSchedulingPreference(SchedulingForRegPressure);
106 setStackPointerRegisterToSaveRestore(X86StackPtr);
108 if (Subtarget->isTargetDarwin()) {
109 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
110 setUseUnderscoreSetJmp(false);
111 setUseUnderscoreLongJmp(false);
112 } else if (Subtarget->isTargetMingw()) {
113 // MS runtime is weird: it exports _setjmp, but longjmp!
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(false);
117 setUseUnderscoreSetJmp(true);
118 setUseUnderscoreLongJmp(true);
121 // Set up the register classes.
122 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
124 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
125 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
126 if (Subtarget->is64Bit())
127 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
129 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
131 // We don't accept any truncstore of integer registers.
132 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
134 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
135 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
137 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
138 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
139 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
141 // SETOEQ and SETUNE require checking two conditions.
142 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
145 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
146 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
147 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
149 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
151 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
152 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
153 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
155 if (Subtarget->is64Bit()) {
156 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
157 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
158 } else if (!UseSoftFloat) {
159 if (X86ScalarSSEf64) {
160 // We have an impenetrably clever algorithm for ui64->double only.
161 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
163 // We have an algorithm for SSE2, and we turn this into a 64-bit
164 // FILD for other targets.
165 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
168 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
170 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
174 // SSE has no i16 to fp conversion, only i32
175 if (X86ScalarSSEf32) {
176 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
177 // f32 and f64 cases are Legal, f80 case is not
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
188 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
189 // are Legal, f80 is custom lowered.
190 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
191 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
193 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
195 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
198 if (X86ScalarSSEf32) {
199 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
200 // f32 and f64 cases are Legal, f80 case is not
201 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
203 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
204 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
207 // Handle FP_TO_UINT by promoting the destination to a larger signed
209 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
210 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
211 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
213 if (Subtarget->is64Bit()) {
214 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
215 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
216 } else if (!UseSoftFloat) {
217 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
218 // Expand FP_TO_UINT into a select.
219 // FIXME: We would like to use a Custom expander here eventually to do
220 // the optimal thing for SSE vs. the default expansion in the legalizer.
221 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
223 // With SSE3 we can use fisttpll to convert to a signed i64; without
224 // SSE, we're stuck with a fistpll.
225 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
228 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
229 if (!X86ScalarSSEf64) {
230 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
231 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
234 // Scalar integer divide and remainder are lowered to use operations that
235 // produce two results, to match the available instructions. This exposes
236 // the two-result form to trivial CSE, which is able to combine x/y and x%y
237 // into a single instruction.
239 // Scalar integer multiply-high is also lowered to use two-result
240 // operations, to match the available instructions. However, plain multiply
241 // (low) operations are left as Legal, as there are single-result
242 // instructions for this in x86. Using the two-result multiply instructions
243 // when both high and low results are needed must be arranged by dagcombine.
244 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
248 setOperationAction(ISD::SREM , MVT::i8 , Expand);
249 setOperationAction(ISD::UREM , MVT::i8 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
254 setOperationAction(ISD::SREM , MVT::i16 , Expand);
255 setOperationAction(ISD::UREM , MVT::i16 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
260 setOperationAction(ISD::SREM , MVT::i32 , Expand);
261 setOperationAction(ISD::UREM , MVT::i32 , Expand);
262 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
263 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
264 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
265 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
266 setOperationAction(ISD::SREM , MVT::i64 , Expand);
267 setOperationAction(ISD::UREM , MVT::i64 , Expand);
269 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
270 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
271 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
272 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
273 if (Subtarget->is64Bit())
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
275 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
276 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
277 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
278 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
279 setOperationAction(ISD::FREM , MVT::f32 , Expand);
280 setOperationAction(ISD::FREM , MVT::f64 , Expand);
281 setOperationAction(ISD::FREM , MVT::f80 , Expand);
282 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
287 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
290 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
292 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
293 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
295 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
296 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
297 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
298 if (Subtarget->is64Bit()) {
299 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
300 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
301 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
304 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
305 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
307 // These should be promoted to a larger select which is supported.
308 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
309 // X86 wants to expand cmov itself.
310 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
312 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
314 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
315 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
316 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
317 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
318 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
323 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
324 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
325 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
326 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
327 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
328 if (Subtarget->is64Bit()) {
329 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
330 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
332 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
335 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
336 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
337 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
338 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
339 if (Subtarget->is64Bit())
340 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
341 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
342 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
345 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
346 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
347 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
348 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
350 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
351 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
352 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
353 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
354 if (Subtarget->is64Bit()) {
355 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
356 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
357 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
360 if (Subtarget->hasSSE1())
361 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
363 if (!Subtarget->hasSSE2())
364 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
366 // Expand certain atomics
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
368 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
369 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
370 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
377 if (!Subtarget->is64Bit()) {
378 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
383 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
384 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
387 // FIXME - use subtarget debug flags
388 if (!Subtarget->isTargetDarwin() &&
389 !Subtarget->isTargetELF() &&
390 !Subtarget->isTargetCygMing()) {
391 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
394 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
395 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
396 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
397 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
398 if (Subtarget->is64Bit()) {
399 setExceptionPointerRegister(X86::RAX);
400 setExceptionSelectorRegister(X86::RDX);
402 setExceptionPointerRegister(X86::EAX);
403 setExceptionSelectorRegister(X86::EDX);
405 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
406 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
408 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
410 setOperationAction(ISD::TRAP, MVT::Other, Legal);
412 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
413 setOperationAction(ISD::VASTART , MVT::Other, Custom);
414 setOperationAction(ISD::VAEND , MVT::Other, Expand);
415 if (Subtarget->is64Bit()) {
416 setOperationAction(ISD::VAARG , MVT::Other, Custom);
417 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
419 setOperationAction(ISD::VAARG , MVT::Other, Expand);
420 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
423 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
424 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
425 if (Subtarget->is64Bit())
426 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
427 if (Subtarget->isTargetCygMing())
428 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
430 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
432 if (!UseSoftFloat && X86ScalarSSEf64) {
433 // f32 and f64 use SSE.
434 // Set up the FP register classes.
435 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
436 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
438 // Use ANDPD to simulate FABS.
439 setOperationAction(ISD::FABS , MVT::f64, Custom);
440 setOperationAction(ISD::FABS , MVT::f32, Custom);
442 // Use XORP to simulate FNEG.
443 setOperationAction(ISD::FNEG , MVT::f64, Custom);
444 setOperationAction(ISD::FNEG , MVT::f32, Custom);
446 // Use ANDPD and ORPD to simulate FCOPYSIGN.
447 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
448 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
450 // We don't support sin/cos/fmod
451 setOperationAction(ISD::FSIN , MVT::f64, Expand);
452 setOperationAction(ISD::FCOS , MVT::f64, Expand);
453 setOperationAction(ISD::FSIN , MVT::f32, Expand);
454 setOperationAction(ISD::FCOS , MVT::f32, Expand);
456 // Expand FP immediates into loads from the stack, except for the special
458 addLegalFPImmediate(APFloat(+0.0)); // xorpd
459 addLegalFPImmediate(APFloat(+0.0f)); // xorps
460 } else if (!UseSoftFloat && X86ScalarSSEf32) {
461 // Use SSE for f32, x87 for f64.
462 // Set up the FP register classes.
463 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
466 // Use ANDPS to simulate FABS.
467 setOperationAction(ISD::FABS , MVT::f32, Custom);
469 // Use XORP to simulate FNEG.
470 setOperationAction(ISD::FNEG , MVT::f32, Custom);
472 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
474 // Use ANDPS and ORPS to simulate FCOPYSIGN.
475 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
476 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
478 // We don't support sin/cos/fmod
479 setOperationAction(ISD::FSIN , MVT::f32, Expand);
480 setOperationAction(ISD::FCOS , MVT::f32, Expand);
482 // Special cases we handle for FP constants.
483 addLegalFPImmediate(APFloat(+0.0f)); // xorps
484 addLegalFPImmediate(APFloat(+0.0)); // FLD0
485 addLegalFPImmediate(APFloat(+1.0)); // FLD1
486 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
487 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 } else if (!UseSoftFloat) {
494 // f32 and f64 in x87.
495 // Set up the FP register classes.
496 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
497 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
499 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
500 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
501 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
502 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
505 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
506 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
508 addLegalFPImmediate(APFloat(+0.0)); // FLD0
509 addLegalFPImmediate(APFloat(+1.0)); // FLD1
510 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
511 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
512 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
513 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
514 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
515 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
518 // Long double always uses X87.
520 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
521 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
522 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
525 APFloat TmpFlt(+0.0);
526 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
528 addLegalFPImmediate(TmpFlt); // FLD0
530 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
531 APFloat TmpFlt2(+1.0);
532 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
534 addLegalFPImmediate(TmpFlt2); // FLD1
535 TmpFlt2.changeSign();
536 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
540 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
541 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
545 // Always use a library call for pow.
546 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
547 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
548 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
550 setOperationAction(ISD::FLOG, MVT::f80, Expand);
551 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
552 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
553 setOperationAction(ISD::FEXP, MVT::f80, Expand);
554 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
556 // First set operation action for all vector types to either promote
557 // (for widening) or expand (for scalarization). Then we will selectively
558 // turn on ones that can be effectively codegen'd.
559 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
560 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
561 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
576 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
577 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
610 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
613 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
614 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
615 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
616 setTruncStoreAction((MVT::SimpleValueType)VT,
617 (MVT::SimpleValueType)InnerVT, Expand);
618 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
619 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
620 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
623 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
624 // with -msoft-float, disable use of MMX as well.
625 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
626 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
628 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
629 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
630 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
632 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
633 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
634 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
635 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
637 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
638 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
639 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
640 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
642 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
643 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
645 setOperationAction(ISD::AND, MVT::v8i8, Promote);
646 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
647 setOperationAction(ISD::AND, MVT::v4i16, Promote);
648 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
649 setOperationAction(ISD::AND, MVT::v2i32, Promote);
650 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
651 setOperationAction(ISD::AND, MVT::v1i64, Legal);
653 setOperationAction(ISD::OR, MVT::v8i8, Promote);
654 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
655 setOperationAction(ISD::OR, MVT::v4i16, Promote);
656 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
657 setOperationAction(ISD::OR, MVT::v2i32, Promote);
658 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
659 setOperationAction(ISD::OR, MVT::v1i64, Legal);
661 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
662 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
663 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
664 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
665 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
666 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
667 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
669 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
670 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
671 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
672 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
673 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
674 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
675 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
676 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
677 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
681 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
686 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
687 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
688 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
692 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
695 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
697 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
698 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
699 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
700 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
703 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
706 if (!UseSoftFloat && Subtarget->hasSSE1()) {
707 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
709 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
711 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
712 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
713 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
714 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
715 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
716 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
717 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
718 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
719 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
720 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
723 if (!UseSoftFloat && Subtarget->hasSSE2()) {
724 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
726 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
727 // registers cannot be used even for integer operations.
728 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
729 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
730 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
731 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
733 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
734 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
735 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
736 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
737 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
738 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
739 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
740 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
741 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
742 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
743 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
745 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
746 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
747 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
748 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
751 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
752 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
753 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
755 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
756 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
758 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
759 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
763 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
764 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
765 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
767 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
768 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
769 EVT VT = (MVT::SimpleValueType)i;
770 // Do not attempt to custom lower non-power-of-2 vectors
771 if (!isPowerOf2_32(VT.getVectorNumElements()))
773 // Do not attempt to custom lower non-128-bit vectors
774 if (!VT.is128BitVector())
776 setOperationAction(ISD::BUILD_VECTOR,
777 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::VECTOR_SHUFFLE,
779 VT.getSimpleVT().SimpleTy, Custom);
780 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
781 VT.getSimpleVT().SimpleTy, Custom);
784 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
785 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
786 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
787 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
788 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
791 if (Subtarget->is64Bit()) {
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
793 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
796 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
797 // FIXME: This produces lots of inefficiencies in isel since
798 // we then need notice that most of our operands have been implicitly
799 // converted to v2i64.
800 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
801 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
804 // Do not attempt to promote non-128-bit vectors
805 if (!VT.is128BitVector()) {
809 setOperationAction(ISD::AND, SVT, Promote);
810 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
811 setOperationAction(ISD::OR, SVT, Promote);
812 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
813 setOperationAction(ISD::XOR, SVT, Promote);
814 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
815 setOperationAction(ISD::LOAD, SVT, Promote);
816 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
817 setOperationAction(ISD::SELECT, SVT, Promote);
818 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
821 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
823 // Custom lower v2i64 and v2f64 selects.
824 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
825 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
826 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
827 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
829 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
830 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
831 if (!DisableMMX && Subtarget->hasMMX()) {
832 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
833 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
837 if (Subtarget->hasSSE41()) {
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // i8 and i16 vectors are custom , because the source register and source
842 // source memory operand types are not the same width. f32 vectors are
843 // custom since the immediate controlling the insert encodes additional
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
855 if (Subtarget->is64Bit()) {
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
861 if (Subtarget->hasSSE42()) {
862 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
865 if (!UseSoftFloat && Subtarget->hasAVX()) {
866 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
867 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
871 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
872 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
873 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
874 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
875 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
876 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
877 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
878 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
879 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
880 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
881 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
882 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
883 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
884 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
885 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
887 // Operations to consider commented out -v16i16 v32i8
888 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
889 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
890 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
891 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
892 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
893 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
894 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
895 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
896 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
897 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
898 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
899 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
900 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
901 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
903 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
904 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
905 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
906 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
908 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
909 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
910 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
914 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
917 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
919 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
922 // Not sure we want to do this since there are no 256-bit integer
925 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
926 // This includes 256-bit vectors
927 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
928 EVT VT = (MVT::SimpleValueType)i;
930 // Do not attempt to custom lower non-power-of-2 vectors
931 if (!isPowerOf2_32(VT.getVectorNumElements()))
934 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
935 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
936 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
939 if (Subtarget->is64Bit()) {
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
946 // Not sure we want to do this since there are no 256-bit integer
949 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
950 // Including 256-bit vectors
951 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
952 EVT VT = (MVT::SimpleValueType)i;
954 if (!VT.is256BitVector()) {
957 setOperationAction(ISD::AND, VT, Promote);
958 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
959 setOperationAction(ISD::OR, VT, Promote);
960 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
961 setOperationAction(ISD::XOR, VT, Promote);
962 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
963 setOperationAction(ISD::LOAD, VT, Promote);
964 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
965 setOperationAction(ISD::SELECT, VT, Promote);
966 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
969 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
973 // We want to custom lower some of our intrinsics.
974 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
976 // Add/Sub/Mul with overflow operations are custom lowered.
977 setOperationAction(ISD::SADDO, MVT::i32, Custom);
978 setOperationAction(ISD::SADDO, MVT::i64, Custom);
979 setOperationAction(ISD::UADDO, MVT::i32, Custom);
980 setOperationAction(ISD::UADDO, MVT::i64, Custom);
981 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
982 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
983 setOperationAction(ISD::USUBO, MVT::i32, Custom);
984 setOperationAction(ISD::USUBO, MVT::i64, Custom);
985 setOperationAction(ISD::SMULO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i64, Custom);
988 if (!Subtarget->is64Bit()) {
989 // These libcalls are not available in 32-bit.
990 setLibcallName(RTLIB::SHL_I128, 0);
991 setLibcallName(RTLIB::SRL_I128, 0);
992 setLibcallName(RTLIB::SRA_I128, 0);
995 // We have target-specific dag combine patterns for the following nodes:
996 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
997 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
998 setTargetDAGCombine(ISD::BUILD_VECTOR);
999 setTargetDAGCombine(ISD::SELECT);
1000 setTargetDAGCombine(ISD::SHL);
1001 setTargetDAGCombine(ISD::SRA);
1002 setTargetDAGCombine(ISD::SRL);
1003 setTargetDAGCombine(ISD::OR);
1004 setTargetDAGCombine(ISD::STORE);
1005 setTargetDAGCombine(ISD::MEMBARRIER);
1006 setTargetDAGCombine(ISD::ZERO_EXTEND);
1007 if (Subtarget->is64Bit())
1008 setTargetDAGCombine(ISD::MUL);
1010 computeRegisterProperties();
1012 // FIXME: These should be based on subtarget info. Plus, the values should
1013 // be smaller when we are in optimizing for size mode.
1014 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1015 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1016 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1017 setPrefLoopAlignment(16);
1018 benefitFromCodePlacementOpt = true;
1022 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1027 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1028 /// the desired ByVal argument alignment.
1029 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1032 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1033 if (VTy->getBitWidth() == 128)
1035 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1036 unsigned EltAlign = 0;
1037 getMaxByValAlign(ATy->getElementType(), EltAlign);
1038 if (EltAlign > MaxAlign)
1039 MaxAlign = EltAlign;
1040 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1041 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1042 unsigned EltAlign = 0;
1043 getMaxByValAlign(STy->getElementType(i), EltAlign);
1044 if (EltAlign > MaxAlign)
1045 MaxAlign = EltAlign;
1053 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1054 /// function arguments in the caller parameter area. For X86, aggregates
1055 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1056 /// are at 4-byte boundaries.
1057 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1058 if (Subtarget->is64Bit()) {
1059 // Max of 8 and alignment of type.
1060 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1067 if (Subtarget->hasSSE1())
1068 getMaxByValAlign(Ty, Align);
1072 /// getOptimalMemOpType - Returns the target specific optimal type for load
1073 /// and store operations as a result of memset, memcpy, and memmove lowering.
1074 /// If DstAlign is zero that means it's safe to destination alignment can
1075 /// satisfy any constraint. Similarly if SrcAlign is zero it means there
1076 /// isn't a need to check it against alignment requirement, probably because
1077 /// the source does not need to be loaded. It returns EVT::Other if
1078 /// SelectionDAG should be responsible for determining it.
1080 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1081 unsigned DstAlign, unsigned SrcAlign,
1083 SelectionDAG &DAG) const {
1084 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1085 // linux. This is because the stack realignment code can't handle certain
1086 // cases like PR2962. This should be removed when PR2962 is fixed.
1087 const Function *F = DAG.getMachineFunction().getFunction();
1088 if (!F->hasFnAttr(Attribute::NoImplicitFloat)) {
1090 (Subtarget->isUnalignedMemAccessFast() ||
1091 ((DstAlign == 0 || DstAlign >= 16) &&
1092 (SrcAlign == 0 || SrcAlign >= 16))) &&
1093 Subtarget->getStackAlignment() >= 16) {
1094 if (Subtarget->hasSSE2())
1096 if (SafeToUseFP && Subtarget->hasSSE1())
1098 } else if (SafeToUseFP &&
1100 !Subtarget->is64Bit() &&
1101 Subtarget->getStackAlignment() >= 8 &&
1102 Subtarget->hasSSE2())
1105 if (Subtarget->is64Bit() && Size >= 8)
1110 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1111 /// current function. The returned value is a member of the
1112 /// MachineJumpTableInfo::JTEntryKind enum.
1113 unsigned X86TargetLowering::getJumpTableEncoding() const {
1114 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1116 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1117 Subtarget->isPICStyleGOT())
1118 return MachineJumpTableInfo::EK_Custom32;
1120 // Otherwise, use the normal jump table encoding heuristics.
1121 return TargetLowering::getJumpTableEncoding();
1124 /// getPICBaseSymbol - Return the X86-32 PIC base.
1126 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1127 MCContext &Ctx) const {
1128 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1129 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1130 Twine(MF->getFunctionNumber())+"$pb");
1135 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1136 const MachineBasicBlock *MBB,
1137 unsigned uid,MCContext &Ctx) const{
1138 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT());
1140 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1142 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1143 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1146 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1148 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1149 SelectionDAG &DAG) const {
1150 if (!Subtarget->is64Bit())
1151 // This doesn't have DebugLoc associated with it, but is not really the
1152 // same as a Register.
1153 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1158 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1159 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1161 const MCExpr *X86TargetLowering::
1162 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1163 MCContext &Ctx) const {
1164 // X86-64 uses RIP relative addressing based on the jump table label.
1165 if (Subtarget->isPICStyleRIPRel())
1166 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1168 // Otherwise, the reference is relative to the PIC base.
1169 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1172 /// getFunctionAlignment - Return the Log2 alignment of this function.
1173 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1174 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1177 //===----------------------------------------------------------------------===//
1178 // Return Value Calling Convention Implementation
1179 //===----------------------------------------------------------------------===//
1181 #include "X86GenCallingConv.inc"
1184 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1185 const SmallVectorImpl<EVT> &OutTys,
1186 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1187 SelectionDAG &DAG) {
1188 SmallVector<CCValAssign, 16> RVLocs;
1189 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1190 RVLocs, *DAG.getContext());
1191 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1195 X86TargetLowering::LowerReturn(SDValue Chain,
1196 CallingConv::ID CallConv, bool isVarArg,
1197 const SmallVectorImpl<ISD::OutputArg> &Outs,
1198 DebugLoc dl, SelectionDAG &DAG) {
1200 SmallVector<CCValAssign, 16> RVLocs;
1201 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1202 RVLocs, *DAG.getContext());
1203 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1205 // Add the regs to the liveout set for the function.
1206 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1207 for (unsigned i = 0; i != RVLocs.size(); ++i)
1208 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1209 MRI.addLiveOut(RVLocs[i].getLocReg());
1213 SmallVector<SDValue, 6> RetOps;
1214 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1215 // Operand #1 = Bytes To Pop
1216 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1218 // Copy the result values into the output registers.
1219 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1220 CCValAssign &VA = RVLocs[i];
1221 assert(VA.isRegLoc() && "Can only return in registers!");
1222 SDValue ValToCopy = Outs[i].Val;
1224 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1225 // the RET instruction and handled by the FP Stackifier.
1226 if (VA.getLocReg() == X86::ST0 ||
1227 VA.getLocReg() == X86::ST1) {
1228 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1229 // change the value to the FP stack register class.
1230 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1231 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1232 RetOps.push_back(ValToCopy);
1233 // Don't emit a copytoreg.
1237 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1238 // which is returned in RAX / RDX.
1239 if (Subtarget->is64Bit()) {
1240 EVT ValVT = ValToCopy.getValueType();
1241 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1242 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1243 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1244 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1248 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1249 Flag = Chain.getValue(1);
1252 // The x86-64 ABI for returning structs by value requires that we copy
1253 // the sret argument into %rax for the return. We saved the argument into
1254 // a virtual register in the entry block, so now we copy the value out
1256 if (Subtarget->is64Bit() &&
1257 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1258 MachineFunction &MF = DAG.getMachineFunction();
1259 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1260 unsigned Reg = FuncInfo->getSRetReturnReg();
1262 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1263 FuncInfo->setSRetReturnReg(Reg);
1265 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1267 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1268 Flag = Chain.getValue(1);
1270 // RAX now acts like a return value.
1271 MRI.addLiveOut(X86::RAX);
1274 RetOps[0] = Chain; // Update chain.
1276 // Add the flag if we have it.
1278 RetOps.push_back(Flag);
1280 return DAG.getNode(X86ISD::RET_FLAG, dl,
1281 MVT::Other, &RetOps[0], RetOps.size());
1284 /// LowerCallResult - Lower the result values of a call into the
1285 /// appropriate copies out of appropriate physical registers.
1288 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1289 CallingConv::ID CallConv, bool isVarArg,
1290 const SmallVectorImpl<ISD::InputArg> &Ins,
1291 DebugLoc dl, SelectionDAG &DAG,
1292 SmallVectorImpl<SDValue> &InVals) {
1294 // Assign locations to each value returned by this call.
1295 SmallVector<CCValAssign, 16> RVLocs;
1296 bool Is64Bit = Subtarget->is64Bit();
1297 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1298 RVLocs, *DAG.getContext());
1299 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1301 // Copy all of the result registers out of their specified physreg.
1302 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1303 CCValAssign &VA = RVLocs[i];
1304 EVT CopyVT = VA.getValVT();
1306 // If this is x86-64, and we disabled SSE, we can't return FP values
1307 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1308 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1309 llvm_report_error("SSE register return with SSE disabled");
1312 // If this is a call to a function that returns an fp value on the floating
1313 // point stack, but where we prefer to use the value in xmm registers, copy
1314 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1315 if ((VA.getLocReg() == X86::ST0 ||
1316 VA.getLocReg() == X86::ST1) &&
1317 isScalarFPTypeInSSEReg(VA.getValVT())) {
1322 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1323 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1324 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1325 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1326 MVT::v2i64, InFlag).getValue(1);
1327 Val = Chain.getValue(0);
1328 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1329 Val, DAG.getConstant(0, MVT::i64));
1331 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1332 MVT::i64, InFlag).getValue(1);
1333 Val = Chain.getValue(0);
1335 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1337 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1338 CopyVT, InFlag).getValue(1);
1339 Val = Chain.getValue(0);
1341 InFlag = Chain.getValue(2);
1343 if (CopyVT != VA.getValVT()) {
1344 // Round the F80 the right size, which also moves to the appropriate xmm
1346 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1347 // This truncation won't change the value.
1348 DAG.getIntPtrConstant(1));
1351 InVals.push_back(Val);
1358 //===----------------------------------------------------------------------===//
1359 // C & StdCall & Fast Calling Convention implementation
1360 //===----------------------------------------------------------------------===//
1361 // StdCall calling convention seems to be standard for many Windows' API
1362 // routines and around. It differs from C calling convention just a little:
1363 // callee should clean up the stack, not caller. Symbols should be also
1364 // decorated in some fancy way :) It doesn't support any vector arguments.
1365 // For info on fast calling convention see Fast Calling Convention (tail call)
1366 // implementation LowerX86_32FastCCCallTo.
1368 /// CallIsStructReturn - Determines whether a call uses struct return
1370 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1374 return Outs[0].Flags.isSRet();
1377 /// ArgsAreStructReturn - Determines whether a function uses struct
1378 /// return semantics.
1380 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1384 return Ins[0].Flags.isSRet();
1387 /// IsCalleePop - Determines whether the callee is required to pop its
1388 /// own arguments. Callee pop is necessary to support tail calls.
1389 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1393 switch (CallingConv) {
1396 case CallingConv::X86_StdCall:
1397 return !Subtarget->is64Bit();
1398 case CallingConv::X86_FastCall:
1399 return !Subtarget->is64Bit();
1400 case CallingConv::Fast:
1401 return GuaranteedTailCallOpt;
1402 case CallingConv::GHC:
1403 return GuaranteedTailCallOpt;
1407 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1408 /// given CallingConvention value.
1409 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1410 if (Subtarget->is64Bit()) {
1411 if (CC == CallingConv::GHC)
1412 return CC_X86_64_GHC;
1413 else if (Subtarget->isTargetWin64())
1414 return CC_X86_Win64_C;
1419 if (CC == CallingConv::X86_FastCall)
1420 return CC_X86_32_FastCall;
1421 else if (CC == CallingConv::Fast)
1422 return CC_X86_32_FastCC;
1423 else if (CC == CallingConv::GHC)
1424 return CC_X86_32_GHC;
1429 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1430 /// by "Src" to address "Dst" with size and alignment information specified by
1431 /// the specific parameter attribute. The copy will be passed as a byval
1432 /// function parameter.
1434 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1435 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1437 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1438 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1439 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1442 /// IsTailCallConvention - Return true if the calling convention is one that
1443 /// supports tail call optimization.
1444 static bool IsTailCallConvention(CallingConv::ID CC) {
1445 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1448 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1449 /// a tailcall target by changing its ABI.
1450 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1451 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1455 X86TargetLowering::LowerMemArgument(SDValue Chain,
1456 CallingConv::ID CallConv,
1457 const SmallVectorImpl<ISD::InputArg> &Ins,
1458 DebugLoc dl, SelectionDAG &DAG,
1459 const CCValAssign &VA,
1460 MachineFrameInfo *MFI,
1462 // Create the nodes corresponding to a load from this parameter slot.
1463 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1464 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1465 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1468 // If value is passed by pointer we have address passed instead of the value
1470 if (VA.getLocInfo() == CCValAssign::Indirect)
1471 ValVT = VA.getLocVT();
1473 ValVT = VA.getValVT();
1475 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1476 // changed with more analysis.
1477 // In case of tail call optimization mark all arguments mutable. Since they
1478 // could be overwritten by lowering of arguments in case of a tail call.
1479 if (Flags.isByVal()) {
1480 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1481 VA.getLocMemOffset(), isImmutable, false);
1482 return DAG.getFrameIndex(FI, getPointerTy());
1484 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1485 VA.getLocMemOffset(), isImmutable, false);
1486 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1487 return DAG.getLoad(ValVT, dl, Chain, FIN,
1488 PseudoSourceValue::getFixedStack(FI), 0,
1494 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1495 CallingConv::ID CallConv,
1497 const SmallVectorImpl<ISD::InputArg> &Ins,
1500 SmallVectorImpl<SDValue> &InVals) {
1501 MachineFunction &MF = DAG.getMachineFunction();
1502 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1504 const Function* Fn = MF.getFunction();
1505 if (Fn->hasExternalLinkage() &&
1506 Subtarget->isTargetCygMing() &&
1507 Fn->getName() == "main")
1508 FuncInfo->setForceFramePointer(true);
1510 MachineFrameInfo *MFI = MF.getFrameInfo();
1511 bool Is64Bit = Subtarget->is64Bit();
1512 bool IsWin64 = Subtarget->isTargetWin64();
1514 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1515 "Var args not supported with calling convention fastcc or ghc");
1517 // Assign locations to all of the incoming arguments.
1518 SmallVector<CCValAssign, 16> ArgLocs;
1519 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1520 ArgLocs, *DAG.getContext());
1521 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1523 unsigned LastVal = ~0U;
1525 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1526 CCValAssign &VA = ArgLocs[i];
1527 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1529 assert(VA.getValNo() != LastVal &&
1530 "Don't support value assigned to multiple locs yet");
1531 LastVal = VA.getValNo();
1533 if (VA.isRegLoc()) {
1534 EVT RegVT = VA.getLocVT();
1535 TargetRegisterClass *RC = NULL;
1536 if (RegVT == MVT::i32)
1537 RC = X86::GR32RegisterClass;
1538 else if (Is64Bit && RegVT == MVT::i64)
1539 RC = X86::GR64RegisterClass;
1540 else if (RegVT == MVT::f32)
1541 RC = X86::FR32RegisterClass;
1542 else if (RegVT == MVT::f64)
1543 RC = X86::FR64RegisterClass;
1544 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1545 RC = X86::VR128RegisterClass;
1546 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1547 RC = X86::VR64RegisterClass;
1549 llvm_unreachable("Unknown argument type!");
1551 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1552 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1554 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1555 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1557 if (VA.getLocInfo() == CCValAssign::SExt)
1558 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1559 DAG.getValueType(VA.getValVT()));
1560 else if (VA.getLocInfo() == CCValAssign::ZExt)
1561 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1562 DAG.getValueType(VA.getValVT()));
1563 else if (VA.getLocInfo() == CCValAssign::BCvt)
1564 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1566 if (VA.isExtInLoc()) {
1567 // Handle MMX values passed in XMM regs.
1568 if (RegVT.isVector()) {
1569 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1570 ArgValue, DAG.getConstant(0, MVT::i64));
1571 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1573 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1576 assert(VA.isMemLoc());
1577 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1580 // If value is passed via pointer - do a load.
1581 if (VA.getLocInfo() == CCValAssign::Indirect)
1582 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1585 InVals.push_back(ArgValue);
1588 // The x86-64 ABI for returning structs by value requires that we copy
1589 // the sret argument into %rax for the return. Save the argument into
1590 // a virtual register so that we can access it from the return points.
1591 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1592 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1593 unsigned Reg = FuncInfo->getSRetReturnReg();
1595 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1596 FuncInfo->setSRetReturnReg(Reg);
1598 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1599 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1602 unsigned StackSize = CCInfo.getNextStackOffset();
1603 // Align stack specially for tail calls.
1604 if (FuncIsMadeTailCallSafe(CallConv))
1605 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1607 // If the function takes variable number of arguments, make a frame index for
1608 // the start of the first vararg value... for expansion of llvm.va_start.
1610 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1611 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1614 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1616 // FIXME: We should really autogenerate these arrays
1617 static const unsigned GPR64ArgRegsWin64[] = {
1618 X86::RCX, X86::RDX, X86::R8, X86::R9
1620 static const unsigned XMMArgRegsWin64[] = {
1621 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1623 static const unsigned GPR64ArgRegs64Bit[] = {
1624 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1626 static const unsigned XMMArgRegs64Bit[] = {
1627 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1628 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1630 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1633 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1634 GPR64ArgRegs = GPR64ArgRegsWin64;
1635 XMMArgRegs = XMMArgRegsWin64;
1637 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1638 GPR64ArgRegs = GPR64ArgRegs64Bit;
1639 XMMArgRegs = XMMArgRegs64Bit;
1641 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1643 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1646 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1647 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1648 "SSE register cannot be used when SSE is disabled!");
1649 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1650 "SSE register cannot be used when SSE is disabled!");
1651 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1652 // Kernel mode asks for SSE to be disabled, so don't push them
1654 TotalNumXMMRegs = 0;
1656 // For X86-64, if there are vararg parameters that are passed via
1657 // registers, then we must store them to their spots on the stack so they
1658 // may be loaded by deferencing the result of va_next.
1659 VarArgsGPOffset = NumIntRegs * 8;
1660 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1661 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1662 TotalNumXMMRegs * 16, 16,
1665 // Store the integer parameter registers.
1666 SmallVector<SDValue, 8> MemOps;
1667 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1668 unsigned Offset = VarArgsGPOffset;
1669 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1670 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1671 DAG.getIntPtrConstant(Offset));
1672 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1673 X86::GR64RegisterClass);
1674 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1676 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1677 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1678 Offset, false, false, 0);
1679 MemOps.push_back(Store);
1683 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1684 // Now store the XMM (fp + vector) parameter registers.
1685 SmallVector<SDValue, 11> SaveXMMOps;
1686 SaveXMMOps.push_back(Chain);
1688 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1689 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1690 SaveXMMOps.push_back(ALVal);
1692 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1693 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1695 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1696 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1697 X86::VR128RegisterClass);
1698 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1699 SaveXMMOps.push_back(Val);
1701 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1703 &SaveXMMOps[0], SaveXMMOps.size()));
1706 if (!MemOps.empty())
1707 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1708 &MemOps[0], MemOps.size());
1712 // Some CCs need callee pop.
1713 if (IsCalleePop(isVarArg, CallConv)) {
1714 BytesToPopOnReturn = StackSize; // Callee pops everything.
1716 BytesToPopOnReturn = 0; // Callee pops nothing.
1717 // If this is an sret function, the return should pop the hidden pointer.
1718 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1719 BytesToPopOnReturn = 4;
1723 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1724 if (CallConv == CallingConv::X86_FastCall)
1725 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1728 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1734 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1735 SDValue StackPtr, SDValue Arg,
1736 DebugLoc dl, SelectionDAG &DAG,
1737 const CCValAssign &VA,
1738 ISD::ArgFlagsTy Flags) {
1739 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1740 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1741 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1742 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1743 if (Flags.isByVal()) {
1744 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1746 return DAG.getStore(Chain, dl, Arg, PtrOff,
1747 PseudoSourceValue::getStack(), LocMemOffset,
1751 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1752 /// optimization is performed and it is required.
1754 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1755 SDValue &OutRetAddr, SDValue Chain,
1756 bool IsTailCall, bool Is64Bit,
1757 int FPDiff, DebugLoc dl) {
1758 // Adjust the Return address stack slot.
1759 EVT VT = getPointerTy();
1760 OutRetAddr = getReturnAddressFrameIndex(DAG);
1762 // Load the "old" Return address.
1763 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1764 return SDValue(OutRetAddr.getNode(), 1);
1767 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1768 /// optimization is performed and it is required (FPDiff!=0).
1770 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1771 SDValue Chain, SDValue RetAddrFrIdx,
1772 bool Is64Bit, int FPDiff, DebugLoc dl) {
1773 // Store the return address to the appropriate stack slot.
1774 if (!FPDiff) return Chain;
1775 // Calculate the new stack slot for the return address.
1776 int SlotSize = Is64Bit ? 8 : 4;
1777 int NewReturnAddrFI =
1778 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1779 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1780 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1781 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1782 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1788 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1789 CallingConv::ID CallConv, bool isVarArg,
1791 const SmallVectorImpl<ISD::OutputArg> &Outs,
1792 const SmallVectorImpl<ISD::InputArg> &Ins,
1793 DebugLoc dl, SelectionDAG &DAG,
1794 SmallVectorImpl<SDValue> &InVals) {
1795 MachineFunction &MF = DAG.getMachineFunction();
1796 bool Is64Bit = Subtarget->is64Bit();
1797 bool IsStructRet = CallIsStructReturn(Outs);
1798 bool IsSibcall = false;
1801 // Check if it's really possible to do a tail call.
1802 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1803 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1806 // Sibcalls are automatically detected tailcalls which do not require
1808 if (!GuaranteedTailCallOpt && isTailCall)
1815 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1816 "Var args not supported with calling convention fastcc or ghc");
1818 // Analyze operands of the call, assigning locations to each operand.
1819 SmallVector<CCValAssign, 16> ArgLocs;
1820 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1821 ArgLocs, *DAG.getContext());
1822 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1824 // Get a count of how many bytes are to be pushed on the stack.
1825 unsigned NumBytes = CCInfo.getNextStackOffset();
1827 // This is a sibcall. The memory operands are available in caller's
1828 // own caller's stack.
1830 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1831 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1834 if (isTailCall && !IsSibcall) {
1835 // Lower arguments at fp - stackoffset + fpdiff.
1836 unsigned NumBytesCallerPushed =
1837 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1838 FPDiff = NumBytesCallerPushed - NumBytes;
1840 // Set the delta of movement of the returnaddr stackslot.
1841 // But only set if delta is greater than previous delta.
1842 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1843 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1847 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1849 SDValue RetAddrFrIdx;
1850 // Load return adress for tail calls.
1851 if (isTailCall && FPDiff)
1852 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1853 Is64Bit, FPDiff, dl);
1855 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1856 SmallVector<SDValue, 8> MemOpChains;
1859 // Walk the register/memloc assignments, inserting copies/loads. In the case
1860 // of tail call optimization arguments are handle later.
1861 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1862 CCValAssign &VA = ArgLocs[i];
1863 EVT RegVT = VA.getLocVT();
1864 SDValue Arg = Outs[i].Val;
1865 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1866 bool isByVal = Flags.isByVal();
1868 // Promote the value if needed.
1869 switch (VA.getLocInfo()) {
1870 default: llvm_unreachable("Unknown loc info!");
1871 case CCValAssign::Full: break;
1872 case CCValAssign::SExt:
1873 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1875 case CCValAssign::ZExt:
1876 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1878 case CCValAssign::AExt:
1879 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1880 // Special case: passing MMX values in XMM registers.
1881 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1882 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1883 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1885 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1887 case CCValAssign::BCvt:
1888 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1890 case CCValAssign::Indirect: {
1891 // Store the argument.
1892 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1893 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1894 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1895 PseudoSourceValue::getFixedStack(FI), 0,
1902 if (VA.isRegLoc()) {
1903 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1904 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1905 assert(VA.isMemLoc());
1906 if (StackPtr.getNode() == 0)
1907 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1908 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1909 dl, DAG, VA, Flags));
1913 if (!MemOpChains.empty())
1914 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1915 &MemOpChains[0], MemOpChains.size());
1917 // Build a sequence of copy-to-reg nodes chained together with token chain
1918 // and flag operands which copy the outgoing args into registers.
1920 // Tail call byval lowering might overwrite argument registers so in case of
1921 // tail call optimization the copies to registers are lowered later.
1923 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1924 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1925 RegsToPass[i].second, InFlag);
1926 InFlag = Chain.getValue(1);
1929 if (Subtarget->isPICStyleGOT()) {
1930 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1933 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1934 DAG.getNode(X86ISD::GlobalBaseReg,
1935 DebugLoc::getUnknownLoc(),
1938 InFlag = Chain.getValue(1);
1940 // If we are tail calling and generating PIC/GOT style code load the
1941 // address of the callee into ECX. The value in ecx is used as target of
1942 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1943 // for tail calls on PIC/GOT architectures. Normally we would just put the
1944 // address of GOT into ebx and then call target@PLT. But for tail calls
1945 // ebx would be restored (since ebx is callee saved) before jumping to the
1948 // Note: The actual moving to ECX is done further down.
1949 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1950 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1951 !G->getGlobal()->hasProtectedVisibility())
1952 Callee = LowerGlobalAddress(Callee, DAG);
1953 else if (isa<ExternalSymbolSDNode>(Callee))
1954 Callee = LowerExternalSymbol(Callee, DAG);
1958 if (Is64Bit && isVarArg) {
1959 // From AMD64 ABI document:
1960 // For calls that may call functions that use varargs or stdargs
1961 // (prototype-less calls or calls to functions containing ellipsis (...) in
1962 // the declaration) %al is used as hidden argument to specify the number
1963 // of SSE registers used. The contents of %al do not need to match exactly
1964 // the number of registers, but must be an ubound on the number of SSE
1965 // registers used and is in the range 0 - 8 inclusive.
1967 // FIXME: Verify this on Win64
1968 // Count the number of XMM registers allocated.
1969 static const unsigned XMMArgRegs[] = {
1970 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1971 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1973 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1974 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1975 && "SSE registers cannot be used when SSE is disabled");
1977 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1978 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1979 InFlag = Chain.getValue(1);
1983 // For tail calls lower the arguments to the 'real' stack slot.
1985 // Force all the incoming stack arguments to be loaded from the stack
1986 // before any new outgoing arguments are stored to the stack, because the
1987 // outgoing stack slots may alias the incoming argument stack slots, and
1988 // the alias isn't otherwise explicit. This is slightly more conservative
1989 // than necessary, because it means that each store effectively depends
1990 // on every argument instead of just those arguments it would clobber.
1991 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1993 SmallVector<SDValue, 8> MemOpChains2;
1996 // Do not flag preceeding copytoreg stuff together with the following stuff.
1998 if (GuaranteedTailCallOpt) {
1999 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2000 CCValAssign &VA = ArgLocs[i];
2003 assert(VA.isMemLoc());
2004 SDValue Arg = Outs[i].Val;
2005 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2006 // Create frame index.
2007 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2008 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2009 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2010 FIN = DAG.getFrameIndex(FI, getPointerTy());
2012 if (Flags.isByVal()) {
2013 // Copy relative to framepointer.
2014 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2015 if (StackPtr.getNode() == 0)
2016 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2018 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2020 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2024 // Store relative to framepointer.
2025 MemOpChains2.push_back(
2026 DAG.getStore(ArgChain, dl, Arg, FIN,
2027 PseudoSourceValue::getFixedStack(FI), 0,
2033 if (!MemOpChains2.empty())
2034 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2035 &MemOpChains2[0], MemOpChains2.size());
2037 // Copy arguments to their registers.
2038 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2039 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2040 RegsToPass[i].second, InFlag);
2041 InFlag = Chain.getValue(1);
2045 // Store the return address to the appropriate stack slot.
2046 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2050 bool WasGlobalOrExternal = false;
2051 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2052 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2053 // In the 64-bit large code model, we have to make all calls
2054 // through a register, since the call instruction's 32-bit
2055 // pc-relative offset may not be large enough to hold the whole
2057 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2058 WasGlobalOrExternal = true;
2059 // If the callee is a GlobalAddress node (quite common, every direct call
2060 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2063 // We should use extra load for direct calls to dllimported functions in
2065 GlobalValue *GV = G->getGlobal();
2066 if (!GV->hasDLLImportLinkage()) {
2067 unsigned char OpFlags = 0;
2069 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2070 // external symbols most go through the PLT in PIC mode. If the symbol
2071 // has hidden or protected visibility, or if it is static or local, then
2072 // we don't need to use the PLT - we can directly call it.
2073 if (Subtarget->isTargetELF() &&
2074 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2075 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2076 OpFlags = X86II::MO_PLT;
2077 } else if (Subtarget->isPICStyleStubAny() &&
2078 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2079 Subtarget->getDarwinVers() < 9) {
2080 // PC-relative references to external symbols should go through $stub,
2081 // unless we're building with the leopard linker or later, which
2082 // automatically synthesizes these stubs.
2083 OpFlags = X86II::MO_DARWIN_STUB;
2086 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2087 G->getOffset(), OpFlags);
2089 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2090 WasGlobalOrExternal = true;
2091 unsigned char OpFlags = 0;
2093 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2094 // symbols should go through the PLT.
2095 if (Subtarget->isTargetELF() &&
2096 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2097 OpFlags = X86II::MO_PLT;
2098 } else if (Subtarget->isPICStyleStubAny() &&
2099 Subtarget->getDarwinVers() < 9) {
2100 // PC-relative references to external symbols should go through $stub,
2101 // unless we're building with the leopard linker or later, which
2102 // automatically synthesizes these stubs.
2103 OpFlags = X86II::MO_DARWIN_STUB;
2106 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2110 // Returns a chain & a flag for retval copy to use.
2111 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2112 SmallVector<SDValue, 8> Ops;
2114 if (!IsSibcall && isTailCall) {
2115 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2116 DAG.getIntPtrConstant(0, true), InFlag);
2117 InFlag = Chain.getValue(1);
2120 Ops.push_back(Chain);
2121 Ops.push_back(Callee);
2124 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2126 // Add argument registers to the end of the list so that they are known live
2128 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2129 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2130 RegsToPass[i].second.getValueType()));
2132 // Add an implicit use GOT pointer in EBX.
2133 if (!isTailCall && Subtarget->isPICStyleGOT())
2134 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2136 // Add an implicit use of AL for x86 vararg functions.
2137 if (Is64Bit && isVarArg)
2138 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2140 if (InFlag.getNode())
2141 Ops.push_back(InFlag);
2144 // If this is the first return lowered for this function, add the regs
2145 // to the liveout set for the function.
2146 if (MF.getRegInfo().liveout_empty()) {
2147 SmallVector<CCValAssign, 16> RVLocs;
2148 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2150 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2151 for (unsigned i = 0; i != RVLocs.size(); ++i)
2152 if (RVLocs[i].isRegLoc())
2153 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2155 return DAG.getNode(X86ISD::TC_RETURN, dl,
2156 NodeTys, &Ops[0], Ops.size());
2159 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2160 InFlag = Chain.getValue(1);
2162 // Create the CALLSEQ_END node.
2163 unsigned NumBytesForCalleeToPush;
2164 if (IsCalleePop(isVarArg, CallConv))
2165 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2166 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2167 // If this is a call to a struct-return function, the callee
2168 // pops the hidden struct pointer, so we have to push it back.
2169 // This is common for Darwin/X86, Linux & Mingw32 targets.
2170 NumBytesForCalleeToPush = 4;
2172 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2174 // Returns a flag for retval copy to use.
2176 Chain = DAG.getCALLSEQ_END(Chain,
2177 DAG.getIntPtrConstant(NumBytes, true),
2178 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2181 InFlag = Chain.getValue(1);
2184 // Handle result values, copying them out of physregs into vregs that we
2186 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2187 Ins, dl, DAG, InVals);
2191 //===----------------------------------------------------------------------===//
2192 // Fast Calling Convention (tail call) implementation
2193 //===----------------------------------------------------------------------===//
2195 // Like std call, callee cleans arguments, convention except that ECX is
2196 // reserved for storing the tail called function address. Only 2 registers are
2197 // free for argument passing (inreg). Tail call optimization is performed
2199 // * tailcallopt is enabled
2200 // * caller/callee are fastcc
2201 // On X86_64 architecture with GOT-style position independent code only local
2202 // (within module) calls are supported at the moment.
2203 // To keep the stack aligned according to platform abi the function
2204 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2205 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2206 // If a tail called function callee has more arguments than the caller the
2207 // caller needs to make sure that there is room to move the RETADDR to. This is
2208 // achieved by reserving an area the size of the argument delta right after the
2209 // original REtADDR, but before the saved framepointer or the spilled registers
2210 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2222 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2223 /// for a 16 byte align requirement.
2224 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2225 SelectionDAG& DAG) {
2226 MachineFunction &MF = DAG.getMachineFunction();
2227 const TargetMachine &TM = MF.getTarget();
2228 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2229 unsigned StackAlignment = TFI.getStackAlignment();
2230 uint64_t AlignMask = StackAlignment - 1;
2231 int64_t Offset = StackSize;
2232 uint64_t SlotSize = TD->getPointerSize();
2233 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2234 // Number smaller than 12 so just add the difference.
2235 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2237 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2238 Offset = ((~AlignMask) & Offset) + StackAlignment +
2239 (StackAlignment-SlotSize);
2244 /// MatchingStackOffset - Return true if the given stack call argument is
2245 /// already available in the same position (relatively) of the caller's
2246 /// incoming argument stack.
2248 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2249 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2250 const X86InstrInfo *TII) {
2251 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2253 if (Arg.getOpcode() == ISD::CopyFromReg) {
2254 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2255 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2257 MachineInstr *Def = MRI->getVRegDef(VR);
2260 if (!Flags.isByVal()) {
2261 if (!TII->isLoadFromStackSlot(Def, FI))
2264 unsigned Opcode = Def->getOpcode();
2265 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2266 Def->getOperand(1).isFI()) {
2267 FI = Def->getOperand(1).getIndex();
2268 Bytes = Flags.getByValSize();
2272 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2273 if (Flags.isByVal())
2274 // ByVal argument is passed in as a pointer but it's now being
2275 // dereferenced. e.g.
2276 // define @foo(%struct.X* %A) {
2277 // tail call @bar(%struct.X* byval %A)
2280 SDValue Ptr = Ld->getBasePtr();
2281 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2284 FI = FINode->getIndex();
2288 assert(FI != INT_MAX);
2289 if (!MFI->isFixedObjectIndex(FI))
2291 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2294 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2295 /// for tail call optimization. Targets which want to do tail call
2296 /// optimization should implement this function.
2298 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2299 CallingConv::ID CalleeCC,
2301 bool isCalleeStructRet,
2302 bool isCallerStructRet,
2303 const SmallVectorImpl<ISD::OutputArg> &Outs,
2304 const SmallVectorImpl<ISD::InputArg> &Ins,
2305 SelectionDAG& DAG) const {
2306 if (!IsTailCallConvention(CalleeCC) &&
2307 CalleeCC != CallingConv::C)
2310 // If -tailcallopt is specified, make fastcc functions tail-callable.
2311 const MachineFunction &MF = DAG.getMachineFunction();
2312 const Function *CallerF = DAG.getMachineFunction().getFunction();
2313 if (GuaranteedTailCallOpt) {
2314 if (IsTailCallConvention(CalleeCC) &&
2315 CallerF->getCallingConv() == CalleeCC)
2320 // Look for obvious safe cases to perform tail call optimization that does not
2321 // requite ABI changes. This is what gcc calls sibcall.
2323 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2324 // emit a special epilogue.
2325 if (RegInfo->needsStackRealignment(MF))
2328 // Do not sibcall optimize vararg calls unless the call site is not passing any
2330 if (isVarArg && !Outs.empty())
2333 // Also avoid sibcall optimization if either caller or callee uses struct
2334 // return semantics.
2335 if (isCalleeStructRet || isCallerStructRet)
2338 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2339 // Therefore if it's not used by the call it is not safe to optimize this into
2341 bool Unused = false;
2342 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2349 SmallVector<CCValAssign, 16> RVLocs;
2350 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2351 RVLocs, *DAG.getContext());
2352 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2353 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2354 CCValAssign &VA = RVLocs[i];
2355 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2360 // If the callee takes no arguments then go on to check the results of the
2362 if (!Outs.empty()) {
2363 // Check if stack adjustment is needed. For now, do not do this if any
2364 // argument is passed on the stack.
2365 SmallVector<CCValAssign, 16> ArgLocs;
2366 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2367 ArgLocs, *DAG.getContext());
2368 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2369 if (CCInfo.getNextStackOffset()) {
2370 MachineFunction &MF = DAG.getMachineFunction();
2371 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2373 if (Subtarget->isTargetWin64())
2374 // Win64 ABI has additional complications.
2377 // Check if the arguments are already laid out in the right way as
2378 // the caller's fixed stack objects.
2379 MachineFrameInfo *MFI = MF.getFrameInfo();
2380 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2381 const X86InstrInfo *TII =
2382 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2383 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2384 CCValAssign &VA = ArgLocs[i];
2385 EVT RegVT = VA.getLocVT();
2386 SDValue Arg = Outs[i].Val;
2387 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2388 if (VA.getLocInfo() == CCValAssign::Indirect)
2390 if (!VA.isRegLoc()) {
2391 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2403 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2405 DenseMap<const Value *, unsigned> &vm,
2406 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2407 DenseMap<const AllocaInst *, int> &am
2409 , SmallSet<Instruction*, 8> &cil
2412 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2420 //===----------------------------------------------------------------------===//
2421 // Other Lowering Hooks
2422 //===----------------------------------------------------------------------===//
2425 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2426 MachineFunction &MF = DAG.getMachineFunction();
2427 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2428 int ReturnAddrIndex = FuncInfo->getRAIndex();
2430 if (ReturnAddrIndex == 0) {
2431 // Set up a frame object for the return address.
2432 uint64_t SlotSize = TD->getPointerSize();
2433 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2435 FuncInfo->setRAIndex(ReturnAddrIndex);
2438 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2442 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2443 bool hasSymbolicDisplacement) {
2444 // Offset should fit into 32 bit immediate field.
2445 if (!isInt<32>(Offset))
2448 // If we don't have a symbolic displacement - we don't have any extra
2450 if (!hasSymbolicDisplacement)
2453 // FIXME: Some tweaks might be needed for medium code model.
2454 if (M != CodeModel::Small && M != CodeModel::Kernel)
2457 // For small code model we assume that latest object is 16MB before end of 31
2458 // bits boundary. We may also accept pretty large negative constants knowing
2459 // that all objects are in the positive half of address space.
2460 if (M == CodeModel::Small && Offset < 16*1024*1024)
2463 // For kernel code model we know that all object resist in the negative half
2464 // of 32bits address space. We may not accept negative offsets, since they may
2465 // be just off and we may accept pretty large positive ones.
2466 if (M == CodeModel::Kernel && Offset > 0)
2472 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2473 /// specific condition code, returning the condition code and the LHS/RHS of the
2474 /// comparison to make.
2475 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2476 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2478 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2479 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2480 // X > -1 -> X == 0, jump !sign.
2481 RHS = DAG.getConstant(0, RHS.getValueType());
2482 return X86::COND_NS;
2483 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2484 // X < 0 -> X == 0, jump on sign.
2486 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2488 RHS = DAG.getConstant(0, RHS.getValueType());
2489 return X86::COND_LE;
2493 switch (SetCCOpcode) {
2494 default: llvm_unreachable("Invalid integer condition!");
2495 case ISD::SETEQ: return X86::COND_E;
2496 case ISD::SETGT: return X86::COND_G;
2497 case ISD::SETGE: return X86::COND_GE;
2498 case ISD::SETLT: return X86::COND_L;
2499 case ISD::SETLE: return X86::COND_LE;
2500 case ISD::SETNE: return X86::COND_NE;
2501 case ISD::SETULT: return X86::COND_B;
2502 case ISD::SETUGT: return X86::COND_A;
2503 case ISD::SETULE: return X86::COND_BE;
2504 case ISD::SETUGE: return X86::COND_AE;
2508 // First determine if it is required or is profitable to flip the operands.
2510 // If LHS is a foldable load, but RHS is not, flip the condition.
2511 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2512 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2513 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2514 std::swap(LHS, RHS);
2517 switch (SetCCOpcode) {
2523 std::swap(LHS, RHS);
2527 // On a floating point condition, the flags are set as follows:
2529 // 0 | 0 | 0 | X > Y
2530 // 0 | 0 | 1 | X < Y
2531 // 1 | 0 | 0 | X == Y
2532 // 1 | 1 | 1 | unordered
2533 switch (SetCCOpcode) {
2534 default: llvm_unreachable("Condcode should be pre-legalized away");
2536 case ISD::SETEQ: return X86::COND_E;
2537 case ISD::SETOLT: // flipped
2539 case ISD::SETGT: return X86::COND_A;
2540 case ISD::SETOLE: // flipped
2542 case ISD::SETGE: return X86::COND_AE;
2543 case ISD::SETUGT: // flipped
2545 case ISD::SETLT: return X86::COND_B;
2546 case ISD::SETUGE: // flipped
2548 case ISD::SETLE: return X86::COND_BE;
2550 case ISD::SETNE: return X86::COND_NE;
2551 case ISD::SETUO: return X86::COND_P;
2552 case ISD::SETO: return X86::COND_NP;
2554 case ISD::SETUNE: return X86::COND_INVALID;
2558 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2559 /// code. Current x86 isa includes the following FP cmov instructions:
2560 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2561 static bool hasFPCMov(unsigned X86CC) {
2577 /// isFPImmLegal - Returns true if the target can instruction select the
2578 /// specified FP immediate natively. If false, the legalizer will
2579 /// materialize the FP immediate as a load from a constant pool.
2580 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2581 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2582 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2588 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2589 /// the specified range (L, H].
2590 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2591 return (Val < 0) || (Val >= Low && Val < Hi);
2594 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2595 /// specified value.
2596 static bool isUndefOrEqual(int Val, int CmpVal) {
2597 if (Val < 0 || Val == CmpVal)
2602 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2603 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2604 /// the second operand.
2605 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2606 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2607 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2608 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2609 return (Mask[0] < 2 && Mask[1] < 2);
2613 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2614 SmallVector<int, 8> M;
2616 return ::isPSHUFDMask(M, N->getValueType(0));
2619 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2620 /// is suitable for input to PSHUFHW.
2621 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2622 if (VT != MVT::v8i16)
2625 // Lower quadword copied in order or undef.
2626 for (int i = 0; i != 4; ++i)
2627 if (Mask[i] >= 0 && Mask[i] != i)
2630 // Upper quadword shuffled.
2631 for (int i = 4; i != 8; ++i)
2632 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2638 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2639 SmallVector<int, 8> M;
2641 return ::isPSHUFHWMask(M, N->getValueType(0));
2644 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2645 /// is suitable for input to PSHUFLW.
2646 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2647 if (VT != MVT::v8i16)
2650 // Upper quadword copied in order.
2651 for (int i = 4; i != 8; ++i)
2652 if (Mask[i] >= 0 && Mask[i] != i)
2655 // Lower quadword shuffled.
2656 for (int i = 0; i != 4; ++i)
2663 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2664 SmallVector<int, 8> M;
2666 return ::isPSHUFLWMask(M, N->getValueType(0));
2669 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2670 /// is suitable for input to PALIGNR.
2671 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2673 int i, e = VT.getVectorNumElements();
2675 // Do not handle v2i64 / v2f64 shuffles with palignr.
2676 if (e < 4 || !hasSSSE3)
2679 for (i = 0; i != e; ++i)
2683 // All undef, not a palignr.
2687 // Determine if it's ok to perform a palignr with only the LHS, since we
2688 // don't have access to the actual shuffle elements to see if RHS is undef.
2689 bool Unary = Mask[i] < (int)e;
2690 bool NeedsUnary = false;
2692 int s = Mask[i] - i;
2694 // Check the rest of the elements to see if they are consecutive.
2695 for (++i; i != e; ++i) {
2700 Unary = Unary && (m < (int)e);
2701 NeedsUnary = NeedsUnary || (m < s);
2703 if (NeedsUnary && !Unary)
2705 if (Unary && m != ((s+i) & (e-1)))
2707 if (!Unary && m != (s+i))
2713 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2714 SmallVector<int, 8> M;
2716 return ::isPALIGNRMask(M, N->getValueType(0), true);
2719 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2720 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2721 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2722 int NumElems = VT.getVectorNumElements();
2723 if (NumElems != 2 && NumElems != 4)
2726 int Half = NumElems / 2;
2727 for (int i = 0; i < Half; ++i)
2728 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2730 for (int i = Half; i < NumElems; ++i)
2731 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2737 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2738 SmallVector<int, 8> M;
2740 return ::isSHUFPMask(M, N->getValueType(0));
2743 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2744 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2745 /// half elements to come from vector 1 (which would equal the dest.) and
2746 /// the upper half to come from vector 2.
2747 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2748 int NumElems = VT.getVectorNumElements();
2750 if (NumElems != 2 && NumElems != 4)
2753 int Half = NumElems / 2;
2754 for (int i = 0; i < Half; ++i)
2755 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2757 for (int i = Half; i < NumElems; ++i)
2758 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2763 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2764 SmallVector<int, 8> M;
2766 return isCommutedSHUFPMask(M, N->getValueType(0));
2769 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2770 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2771 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2772 if (N->getValueType(0).getVectorNumElements() != 4)
2775 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2776 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2777 isUndefOrEqual(N->getMaskElt(1), 7) &&
2778 isUndefOrEqual(N->getMaskElt(2), 2) &&
2779 isUndefOrEqual(N->getMaskElt(3), 3);
2782 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2783 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2785 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2786 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2791 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2792 isUndefOrEqual(N->getMaskElt(1), 3) &&
2793 isUndefOrEqual(N->getMaskElt(2), 2) &&
2794 isUndefOrEqual(N->getMaskElt(3), 3);
2797 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2798 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2799 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2800 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2802 if (NumElems != 2 && NumElems != 4)
2805 for (unsigned i = 0; i < NumElems/2; ++i)
2806 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2809 for (unsigned i = NumElems/2; i < NumElems; ++i)
2810 if (!isUndefOrEqual(N->getMaskElt(i), i))
2816 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2817 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2818 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2819 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2821 if (NumElems != 2 && NumElems != 4)
2824 for (unsigned i = 0; i < NumElems/2; ++i)
2825 if (!isUndefOrEqual(N->getMaskElt(i), i))
2828 for (unsigned i = 0; i < NumElems/2; ++i)
2829 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2835 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2836 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2837 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2838 bool V2IsSplat = false) {
2839 int NumElts = VT.getVectorNumElements();
2840 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2843 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2845 int BitI1 = Mask[i+1];
2846 if (!isUndefOrEqual(BitI, j))
2849 if (!isUndefOrEqual(BitI1, NumElts))
2852 if (!isUndefOrEqual(BitI1, j + NumElts))
2859 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2860 SmallVector<int, 8> M;
2862 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2865 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2866 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2867 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2868 bool V2IsSplat = false) {
2869 int NumElts = VT.getVectorNumElements();
2870 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2873 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2875 int BitI1 = Mask[i+1];
2876 if (!isUndefOrEqual(BitI, j + NumElts/2))
2879 if (isUndefOrEqual(BitI1, NumElts))
2882 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2889 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2890 SmallVector<int, 8> M;
2892 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2895 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2896 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2898 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2899 int NumElems = VT.getVectorNumElements();
2900 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2903 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2905 int BitI1 = Mask[i+1];
2906 if (!isUndefOrEqual(BitI, j))
2908 if (!isUndefOrEqual(BitI1, j))
2914 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2915 SmallVector<int, 8> M;
2917 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2920 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2921 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2923 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2924 int NumElems = VT.getVectorNumElements();
2925 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2928 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2930 int BitI1 = Mask[i+1];
2931 if (!isUndefOrEqual(BitI, j))
2933 if (!isUndefOrEqual(BitI1, j))
2939 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2940 SmallVector<int, 8> M;
2942 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2945 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2946 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2947 /// MOVSD, and MOVD, i.e. setting the lowest element.
2948 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2949 if (VT.getVectorElementType().getSizeInBits() < 32)
2952 int NumElts = VT.getVectorNumElements();
2954 if (!isUndefOrEqual(Mask[0], NumElts))
2957 for (int i = 1; i < NumElts; ++i)
2958 if (!isUndefOrEqual(Mask[i], i))
2964 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2965 SmallVector<int, 8> M;
2967 return ::isMOVLMask(M, N->getValueType(0));
2970 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2971 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2972 /// element of vector 2 and the other elements to come from vector 1 in order.
2973 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2974 bool V2IsSplat = false, bool V2IsUndef = false) {
2975 int NumOps = VT.getVectorNumElements();
2976 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2979 if (!isUndefOrEqual(Mask[0], 0))
2982 for (int i = 1; i < NumOps; ++i)
2983 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2984 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2985 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2991 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2992 bool V2IsUndef = false) {
2993 SmallVector<int, 8> M;
2995 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2998 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2999 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3000 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3001 if (N->getValueType(0).getVectorNumElements() != 4)
3004 // Expect 1, 1, 3, 3
3005 for (unsigned i = 0; i < 2; ++i) {
3006 int Elt = N->getMaskElt(i);
3007 if (Elt >= 0 && Elt != 1)
3012 for (unsigned i = 2; i < 4; ++i) {
3013 int Elt = N->getMaskElt(i);
3014 if (Elt >= 0 && Elt != 3)
3019 // Don't use movshdup if it can be done with a shufps.
3020 // FIXME: verify that matching u, u, 3, 3 is what we want.
3024 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3025 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3026 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3027 if (N->getValueType(0).getVectorNumElements() != 4)
3030 // Expect 0, 0, 2, 2
3031 for (unsigned i = 0; i < 2; ++i)
3032 if (N->getMaskElt(i) > 0)
3036 for (unsigned i = 2; i < 4; ++i) {
3037 int Elt = N->getMaskElt(i);
3038 if (Elt >= 0 && Elt != 2)
3043 // Don't use movsldup if it can be done with a shufps.
3047 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3048 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3049 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3050 int e = N->getValueType(0).getVectorNumElements() / 2;
3052 for (int i = 0; i < e; ++i)
3053 if (!isUndefOrEqual(N->getMaskElt(i), i))
3055 for (int i = 0; i < e; ++i)
3056 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3061 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3062 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3063 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3064 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3065 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3067 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3069 for (int i = 0; i < NumOperands; ++i) {
3070 int Val = SVOp->getMaskElt(NumOperands-i-1);
3071 if (Val < 0) Val = 0;
3072 if (Val >= NumOperands) Val -= NumOperands;
3074 if (i != NumOperands - 1)
3080 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3081 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3082 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3083 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3085 // 8 nodes, but we only care about the last 4.
3086 for (unsigned i = 7; i >= 4; --i) {
3087 int Val = SVOp->getMaskElt(i);
3096 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3097 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3098 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3099 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3101 // 8 nodes, but we only care about the first 4.
3102 for (int i = 3; i >= 0; --i) {
3103 int Val = SVOp->getMaskElt(i);
3112 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3113 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3114 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3115 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3116 EVT VVT = N->getValueType(0);
3117 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3121 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3122 Val = SVOp->getMaskElt(i);
3126 return (Val - i) * EltSize;
3129 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3131 bool X86::isZeroNode(SDValue Elt) {
3132 return ((isa<ConstantSDNode>(Elt) &&
3133 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3134 (isa<ConstantFPSDNode>(Elt) &&
3135 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3138 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3139 /// their permute mask.
3140 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3141 SelectionDAG &DAG) {
3142 EVT VT = SVOp->getValueType(0);
3143 unsigned NumElems = VT.getVectorNumElements();
3144 SmallVector<int, 8> MaskVec;
3146 for (unsigned i = 0; i != NumElems; ++i) {
3147 int idx = SVOp->getMaskElt(i);
3149 MaskVec.push_back(idx);
3150 else if (idx < (int)NumElems)
3151 MaskVec.push_back(idx + NumElems);
3153 MaskVec.push_back(idx - NumElems);
3155 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3156 SVOp->getOperand(0), &MaskVec[0]);
3159 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3160 /// the two vector operands have swapped position.
3161 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3162 unsigned NumElems = VT.getVectorNumElements();
3163 for (unsigned i = 0; i != NumElems; ++i) {
3167 else if (idx < (int)NumElems)
3168 Mask[i] = idx + NumElems;
3170 Mask[i] = idx - NumElems;
3174 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3175 /// match movhlps. The lower half elements should come from upper half of
3176 /// V1 (and in order), and the upper half elements should come from the upper
3177 /// half of V2 (and in order).
3178 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3179 if (Op->getValueType(0).getVectorNumElements() != 4)
3181 for (unsigned i = 0, e = 2; i != e; ++i)
3182 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3184 for (unsigned i = 2; i != 4; ++i)
3185 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3190 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3191 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3193 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3194 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3196 N = N->getOperand(0).getNode();
3197 if (!ISD::isNON_EXTLoad(N))
3200 *LD = cast<LoadSDNode>(N);
3204 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3205 /// match movlp{s|d}. The lower half elements should come from lower half of
3206 /// V1 (and in order), and the upper half elements should come from the upper
3207 /// half of V2 (and in order). And since V1 will become the source of the
3208 /// MOVLP, it must be either a vector load or a scalar load to vector.
3209 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3210 ShuffleVectorSDNode *Op) {
3211 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3213 // Is V2 is a vector load, don't do this transformation. We will try to use
3214 // load folding shufps op.
3215 if (ISD::isNON_EXTLoad(V2))
3218 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3220 if (NumElems != 2 && NumElems != 4)
3222 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3223 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3225 for (unsigned i = NumElems/2; i != NumElems; ++i)
3226 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3231 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3233 static bool isSplatVector(SDNode *N) {
3234 if (N->getOpcode() != ISD::BUILD_VECTOR)
3237 SDValue SplatValue = N->getOperand(0);
3238 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3239 if (N->getOperand(i) != SplatValue)
3244 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3245 /// to an zero vector.
3246 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3247 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3248 SDValue V1 = N->getOperand(0);
3249 SDValue V2 = N->getOperand(1);
3250 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3251 for (unsigned i = 0; i != NumElems; ++i) {
3252 int Idx = N->getMaskElt(i);
3253 if (Idx >= (int)NumElems) {
3254 unsigned Opc = V2.getOpcode();
3255 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3257 if (Opc != ISD::BUILD_VECTOR ||
3258 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3260 } else if (Idx >= 0) {
3261 unsigned Opc = V1.getOpcode();
3262 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3264 if (Opc != ISD::BUILD_VECTOR ||
3265 !X86::isZeroNode(V1.getOperand(Idx)))
3272 /// getZeroVector - Returns a vector of specified type with all zero elements.
3274 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3276 assert(VT.isVector() && "Expected a vector type");
3278 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3279 // type. This ensures they get CSE'd.
3281 if (VT.getSizeInBits() == 64) { // MMX
3282 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3283 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3284 } else if (HasSSE2) { // SSE2
3285 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3286 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3288 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3289 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3291 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3294 /// getOnesVector - Returns a vector of specified type with all bits set.
3296 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3297 assert(VT.isVector() && "Expected a vector type");
3299 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3300 // type. This ensures they get CSE'd.
3301 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3303 if (VT.getSizeInBits() == 64) // MMX
3304 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3306 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3307 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3311 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3312 /// that point to V2 points to its first element.
3313 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3314 EVT VT = SVOp->getValueType(0);
3315 unsigned NumElems = VT.getVectorNumElements();
3317 bool Changed = false;
3318 SmallVector<int, 8> MaskVec;
3319 SVOp->getMask(MaskVec);
3321 for (unsigned i = 0; i != NumElems; ++i) {
3322 if (MaskVec[i] > (int)NumElems) {
3323 MaskVec[i] = NumElems;
3328 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3329 SVOp->getOperand(1), &MaskVec[0]);
3330 return SDValue(SVOp, 0);
3333 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3334 /// operation of specified width.
3335 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3337 unsigned NumElems = VT.getVectorNumElements();
3338 SmallVector<int, 8> Mask;
3339 Mask.push_back(NumElems);
3340 for (unsigned i = 1; i != NumElems; ++i)
3342 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3345 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3346 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3348 unsigned NumElems = VT.getVectorNumElements();
3349 SmallVector<int, 8> Mask;
3350 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3352 Mask.push_back(i + NumElems);
3354 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3357 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3358 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3360 unsigned NumElems = VT.getVectorNumElements();
3361 unsigned Half = NumElems/2;
3362 SmallVector<int, 8> Mask;
3363 for (unsigned i = 0; i != Half; ++i) {
3364 Mask.push_back(i + Half);
3365 Mask.push_back(i + NumElems + Half);
3367 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3370 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3371 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3373 if (SV->getValueType(0).getVectorNumElements() <= 4)
3374 return SDValue(SV, 0);
3376 EVT PVT = MVT::v4f32;
3377 EVT VT = SV->getValueType(0);
3378 DebugLoc dl = SV->getDebugLoc();
3379 SDValue V1 = SV->getOperand(0);
3380 int NumElems = VT.getVectorNumElements();
3381 int EltNo = SV->getSplatIndex();
3383 // unpack elements to the correct location
3384 while (NumElems > 4) {
3385 if (EltNo < NumElems/2) {
3386 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3388 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3389 EltNo -= NumElems/2;
3394 // Perform the splat.
3395 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3396 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3397 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3398 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3401 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3402 /// vector of zero or undef vector. This produces a shuffle where the low
3403 /// element of V2 is swizzled into the zero/undef vector, landing at element
3404 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3405 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3406 bool isZero, bool HasSSE2,
3407 SelectionDAG &DAG) {
3408 EVT VT = V2.getValueType();
3410 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3411 unsigned NumElems = VT.getVectorNumElements();
3412 SmallVector<int, 16> MaskVec;
3413 for (unsigned i = 0; i != NumElems; ++i)
3414 // If this is the insertion idx, put the low elt of V2 here.
3415 MaskVec.push_back(i == Idx ? NumElems : i);
3416 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3419 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3420 /// a shuffle that is zero.
3422 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3423 bool Low, SelectionDAG &DAG) {
3424 unsigned NumZeros = 0;
3425 for (int i = 0; i < NumElems; ++i) {
3426 unsigned Index = Low ? i : NumElems-i-1;
3427 int Idx = SVOp->getMaskElt(Index);
3432 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3433 if (Elt.getNode() && X86::isZeroNode(Elt))
3441 /// isVectorShift - Returns true if the shuffle can be implemented as a
3442 /// logical left or right shift of a vector.
3443 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3444 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3445 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3446 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3449 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3452 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3456 bool SeenV1 = false;
3457 bool SeenV2 = false;
3458 for (int i = NumZeros; i < NumElems; ++i) {
3459 int Val = isLeft ? (i - NumZeros) : i;
3460 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3472 if (SeenV1 && SeenV2)
3475 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3481 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3483 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3484 unsigned NumNonZero, unsigned NumZero,
3485 SelectionDAG &DAG, TargetLowering &TLI) {
3489 DebugLoc dl = Op.getDebugLoc();
3492 for (unsigned i = 0; i < 16; ++i) {
3493 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3494 if (ThisIsNonZero && First) {
3496 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3498 V = DAG.getUNDEF(MVT::v8i16);
3503 SDValue ThisElt(0, 0), LastElt(0, 0);
3504 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3505 if (LastIsNonZero) {
3506 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3507 MVT::i16, Op.getOperand(i-1));
3509 if (ThisIsNonZero) {
3510 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3511 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3512 ThisElt, DAG.getConstant(8, MVT::i8));
3514 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3518 if (ThisElt.getNode())
3519 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3520 DAG.getIntPtrConstant(i/2));
3524 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3527 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3529 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3530 unsigned NumNonZero, unsigned NumZero,
3531 SelectionDAG &DAG, TargetLowering &TLI) {
3535 DebugLoc dl = Op.getDebugLoc();
3538 for (unsigned i = 0; i < 8; ++i) {
3539 bool isNonZero = (NonZeros & (1 << i)) != 0;
3543 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3545 V = DAG.getUNDEF(MVT::v8i16);
3548 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3549 MVT::v8i16, V, Op.getOperand(i),
3550 DAG.getIntPtrConstant(i));
3557 /// getVShift - Return a vector logical shift node.
3559 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3560 unsigned NumBits, SelectionDAG &DAG,
3561 const TargetLowering &TLI, DebugLoc dl) {
3562 bool isMMX = VT.getSizeInBits() == 64;
3563 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3564 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3565 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3566 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3567 DAG.getNode(Opc, dl, ShVT, SrcOp,
3568 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3572 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3573 SelectionDAG &DAG) {
3575 // Check if the scalar load can be widened into a vector load. And if
3576 // the address is "base + cst" see if the cst can be "absorbed" into
3577 // the shuffle mask.
3578 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3579 SDValue Ptr = LD->getBasePtr();
3580 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3582 EVT PVT = LD->getValueType(0);
3583 if (PVT != MVT::i32 && PVT != MVT::f32)
3588 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3589 FI = FINode->getIndex();
3591 } else if (Ptr.getOpcode() == ISD::ADD &&
3592 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3593 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3594 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3595 Offset = Ptr.getConstantOperandVal(1);
3596 Ptr = Ptr.getOperand(0);
3601 SDValue Chain = LD->getChain();
3602 // Make sure the stack object alignment is at least 16.
3603 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3604 if (DAG.InferPtrAlignment(Ptr) < 16) {
3605 if (MFI->isFixedObjectIndex(FI)) {
3606 // Can't change the alignment. FIXME: It's possible to compute
3607 // the exact stack offset and reference FI + adjust offset instead.
3608 // If someone *really* cares about this. That's the way to implement it.
3611 MFI->setObjectAlignment(FI, 16);
3615 // (Offset % 16) must be multiple of 4. Then address is then
3616 // Ptr + (Offset & ~15).
3619 if ((Offset % 16) & 3)
3621 int64_t StartOffset = Offset & ~15;
3623 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3624 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3626 int EltNo = (Offset - StartOffset) >> 2;
3627 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3628 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3629 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3631 // Canonicalize it to a v4i32 shuffle.
3632 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3633 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3634 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3635 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3641 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3642 /// vector of type 'VT', see if the elements can be replaced by a single large
3643 /// load which has the same value as a build_vector whose operands are 'elts'.
3645 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3647 /// FIXME: we'd also like to handle the case where the last elements are zero
3648 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3649 /// There's even a handy isZeroNode for that purpose.
3650 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3651 DebugLoc &dl, SelectionDAG &DAG) {
3652 EVT EltVT = VT.getVectorElementType();
3653 unsigned NumElems = Elts.size();
3655 LoadSDNode *LDBase = NULL;
3656 unsigned LastLoadedElt = -1U;
3658 // For each element in the initializer, see if we've found a load or an undef.
3659 // If we don't find an initial load element, or later load elements are
3660 // non-consecutive, bail out.
3661 for (unsigned i = 0; i < NumElems; ++i) {
3662 SDValue Elt = Elts[i];
3664 if (!Elt.getNode() ||
3665 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3668 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3670 LDBase = cast<LoadSDNode>(Elt.getNode());
3674 if (Elt.getOpcode() == ISD::UNDEF)
3677 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3678 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3683 // If we have found an entire vector of loads and undefs, then return a large
3684 // load of the entire vector width starting at the base pointer. If we found
3685 // consecutive loads for the low half, generate a vzext_load node.
3686 if (LastLoadedElt == NumElems - 1) {
3687 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3688 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3689 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3690 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3691 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3692 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3693 LDBase->isVolatile(), LDBase->isNonTemporal(),
3694 LDBase->getAlignment());
3695 } else if (NumElems == 4 && LastLoadedElt == 1) {
3696 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3697 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3698 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3699 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3705 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3706 DebugLoc dl = Op.getDebugLoc();
3707 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3708 if (ISD::isBuildVectorAllZeros(Op.getNode())
3709 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3710 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3711 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3712 // eliminated on x86-32 hosts.
3713 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3716 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3717 return getOnesVector(Op.getValueType(), DAG, dl);
3718 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3721 EVT VT = Op.getValueType();
3722 EVT ExtVT = VT.getVectorElementType();
3723 unsigned EVTBits = ExtVT.getSizeInBits();
3725 unsigned NumElems = Op.getNumOperands();
3726 unsigned NumZero = 0;
3727 unsigned NumNonZero = 0;
3728 unsigned NonZeros = 0;
3729 bool IsAllConstants = true;
3730 SmallSet<SDValue, 8> Values;
3731 for (unsigned i = 0; i < NumElems; ++i) {
3732 SDValue Elt = Op.getOperand(i);
3733 if (Elt.getOpcode() == ISD::UNDEF)
3736 if (Elt.getOpcode() != ISD::Constant &&
3737 Elt.getOpcode() != ISD::ConstantFP)
3738 IsAllConstants = false;
3739 if (X86::isZeroNode(Elt))
3742 NonZeros |= (1 << i);
3747 if (NumNonZero == 0) {
3748 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3749 return DAG.getUNDEF(VT);
3752 // Special case for single non-zero, non-undef, element.
3753 if (NumNonZero == 1) {
3754 unsigned Idx = CountTrailingZeros_32(NonZeros);
3755 SDValue Item = Op.getOperand(Idx);
3757 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3758 // the value are obviously zero, truncate the value to i32 and do the
3759 // insertion that way. Only do this if the value is non-constant or if the
3760 // value is a constant being inserted into element 0. It is cheaper to do
3761 // a constant pool load than it is to do a movd + shuffle.
3762 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3763 (!IsAllConstants || Idx == 0)) {
3764 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3765 // Handle MMX and SSE both.
3766 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3767 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3769 // Truncate the value (which may itself be a constant) to i32, and
3770 // convert it to a vector with movd (S2V+shuffle to zero extend).
3771 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3772 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3773 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3774 Subtarget->hasSSE2(), DAG);
3776 // Now we have our 32-bit value zero extended in the low element of
3777 // a vector. If Idx != 0, swizzle it into place.
3779 SmallVector<int, 4> Mask;
3780 Mask.push_back(Idx);
3781 for (unsigned i = 1; i != VecElts; ++i)
3783 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3784 DAG.getUNDEF(Item.getValueType()),
3787 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3791 // If we have a constant or non-constant insertion into the low element of
3792 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3793 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3794 // depending on what the source datatype is.
3797 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3798 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3799 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3800 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3801 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3802 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3804 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3805 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3806 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3807 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3808 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3809 Subtarget->hasSSE2(), DAG);
3810 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3814 // Is it a vector logical left shift?
3815 if (NumElems == 2 && Idx == 1 &&
3816 X86::isZeroNode(Op.getOperand(0)) &&
3817 !X86::isZeroNode(Op.getOperand(1))) {
3818 unsigned NumBits = VT.getSizeInBits();
3819 return getVShift(true, VT,
3820 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3821 VT, Op.getOperand(1)),
3822 NumBits/2, DAG, *this, dl);
3825 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3828 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3829 // is a non-constant being inserted into an element other than the low one,
3830 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3831 // movd/movss) to move this into the low element, then shuffle it into
3833 if (EVTBits == 32) {
3834 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3836 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3837 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3838 Subtarget->hasSSE2(), DAG);
3839 SmallVector<int, 8> MaskVec;
3840 for (unsigned i = 0; i < NumElems; i++)
3841 MaskVec.push_back(i == Idx ? 0 : 1);
3842 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3846 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3847 if (Values.size() == 1) {
3848 if (EVTBits == 32) {
3849 // Instead of a shuffle like this:
3850 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3851 // Check if it's possible to issue this instead.
3852 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3853 unsigned Idx = CountTrailingZeros_32(NonZeros);
3854 SDValue Item = Op.getOperand(Idx);
3855 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3856 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3861 // A vector full of immediates; various special cases are already
3862 // handled, so this is best done with a single constant-pool load.
3866 // Let legalizer expand 2-wide build_vectors.
3867 if (EVTBits == 64) {
3868 if (NumNonZero == 1) {
3869 // One half is zero or undef.
3870 unsigned Idx = CountTrailingZeros_32(NonZeros);
3871 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3872 Op.getOperand(Idx));
3873 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3874 Subtarget->hasSSE2(), DAG);
3879 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3880 if (EVTBits == 8 && NumElems == 16) {
3881 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3883 if (V.getNode()) return V;
3886 if (EVTBits == 16 && NumElems == 8) {
3887 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3889 if (V.getNode()) return V;
3892 // If element VT is == 32 bits, turn it into a number of shuffles.
3893 SmallVector<SDValue, 8> V;
3895 if (NumElems == 4 && NumZero > 0) {
3896 for (unsigned i = 0; i < 4; ++i) {
3897 bool isZero = !(NonZeros & (1 << i));
3899 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3901 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3904 for (unsigned i = 0; i < 2; ++i) {
3905 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3908 V[i] = V[i*2]; // Must be a zero vector.
3911 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3914 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3917 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3922 SmallVector<int, 8> MaskVec;
3923 bool Reverse = (NonZeros & 0x3) == 2;
3924 for (unsigned i = 0; i < 2; ++i)
3925 MaskVec.push_back(Reverse ? 1-i : i);
3926 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3927 for (unsigned i = 0; i < 2; ++i)
3928 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3929 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3932 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3933 // Check for a build vector of consecutive loads.
3934 for (unsigned i = 0; i < NumElems; ++i)
3935 V[i] = Op.getOperand(i);
3937 // Check for elements which are consecutive loads.
3938 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3942 // For SSE 4.1, use inserts into undef.
3943 if (getSubtarget()->hasSSE41()) {
3944 V[0] = DAG.getUNDEF(VT);
3945 for (unsigned i = 0; i < NumElems; ++i)
3946 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3947 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3948 Op.getOperand(i), DAG.getIntPtrConstant(i));
3952 // Otherwise, expand into a number of unpckl*
3954 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3955 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3956 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3957 for (unsigned i = 0; i < NumElems; ++i)
3958 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3960 while (NumElems != 0) {
3961 for (unsigned i = 0; i < NumElems; ++i)
3962 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3971 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3972 // We support concatenate two MMX registers and place them in a MMX
3973 // register. This is better than doing a stack convert.
3974 DebugLoc dl = Op.getDebugLoc();
3975 EVT ResVT = Op.getValueType();
3976 assert(Op.getNumOperands() == 2);
3977 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3978 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3980 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3981 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3982 InVec = Op.getOperand(1);
3983 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3984 unsigned NumElts = ResVT.getVectorNumElements();
3985 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3986 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3987 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3989 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3990 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3991 Mask[0] = 0; Mask[1] = 2;
3992 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3994 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3997 // v8i16 shuffles - Prefer shuffles in the following order:
3998 // 1. [all] pshuflw, pshufhw, optional move
3999 // 2. [ssse3] 1 x pshufb
4000 // 3. [ssse3] 2 x pshufb + 1 x por
4001 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4003 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4004 SelectionDAG &DAG, X86TargetLowering &TLI) {
4005 SDValue V1 = SVOp->getOperand(0);
4006 SDValue V2 = SVOp->getOperand(1);
4007 DebugLoc dl = SVOp->getDebugLoc();
4008 SmallVector<int, 8> MaskVals;
4010 // Determine if more than 1 of the words in each of the low and high quadwords
4011 // of the result come from the same quadword of one of the two inputs. Undef
4012 // mask values count as coming from any quadword, for better codegen.
4013 SmallVector<unsigned, 4> LoQuad(4);
4014 SmallVector<unsigned, 4> HiQuad(4);
4015 BitVector InputQuads(4);
4016 for (unsigned i = 0; i < 8; ++i) {
4017 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4018 int EltIdx = SVOp->getMaskElt(i);
4019 MaskVals.push_back(EltIdx);
4028 InputQuads.set(EltIdx / 4);
4031 int BestLoQuad = -1;
4032 unsigned MaxQuad = 1;
4033 for (unsigned i = 0; i < 4; ++i) {
4034 if (LoQuad[i] > MaxQuad) {
4036 MaxQuad = LoQuad[i];
4040 int BestHiQuad = -1;
4042 for (unsigned i = 0; i < 4; ++i) {
4043 if (HiQuad[i] > MaxQuad) {
4045 MaxQuad = HiQuad[i];
4049 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4050 // of the two input vectors, shuffle them into one input vector so only a
4051 // single pshufb instruction is necessary. If There are more than 2 input
4052 // quads, disable the next transformation since it does not help SSSE3.
4053 bool V1Used = InputQuads[0] || InputQuads[1];
4054 bool V2Used = InputQuads[2] || InputQuads[3];
4055 if (TLI.getSubtarget()->hasSSSE3()) {
4056 if (InputQuads.count() == 2 && V1Used && V2Used) {
4057 BestLoQuad = InputQuads.find_first();
4058 BestHiQuad = InputQuads.find_next(BestLoQuad);
4060 if (InputQuads.count() > 2) {
4066 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4067 // the shuffle mask. If a quad is scored as -1, that means that it contains
4068 // words from all 4 input quadwords.
4070 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4071 SmallVector<int, 8> MaskV;
4072 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4073 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4074 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4075 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4076 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4077 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4079 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4080 // source words for the shuffle, to aid later transformations.
4081 bool AllWordsInNewV = true;
4082 bool InOrder[2] = { true, true };
4083 for (unsigned i = 0; i != 8; ++i) {
4084 int idx = MaskVals[i];
4086 InOrder[i/4] = false;
4087 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4089 AllWordsInNewV = false;
4093 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4094 if (AllWordsInNewV) {
4095 for (int i = 0; i != 8; ++i) {
4096 int idx = MaskVals[i];
4099 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4100 if ((idx != i) && idx < 4)
4102 if ((idx != i) && idx > 3)
4111 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4112 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4113 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4114 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4115 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4119 // If we have SSSE3, and all words of the result are from 1 input vector,
4120 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4121 // is present, fall back to case 4.
4122 if (TLI.getSubtarget()->hasSSSE3()) {
4123 SmallVector<SDValue,16> pshufbMask;
4125 // If we have elements from both input vectors, set the high bit of the
4126 // shuffle mask element to zero out elements that come from V2 in the V1
4127 // mask, and elements that come from V1 in the V2 mask, so that the two
4128 // results can be OR'd together.
4129 bool TwoInputs = V1Used && V2Used;
4130 for (unsigned i = 0; i != 8; ++i) {
4131 int EltIdx = MaskVals[i] * 2;
4132 if (TwoInputs && (EltIdx >= 16)) {
4133 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4134 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4137 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4138 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4140 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4141 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4142 DAG.getNode(ISD::BUILD_VECTOR, dl,
4143 MVT::v16i8, &pshufbMask[0], 16));
4145 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4147 // Calculate the shuffle mask for the second input, shuffle it, and
4148 // OR it with the first shuffled input.
4150 for (unsigned i = 0; i != 8; ++i) {
4151 int EltIdx = MaskVals[i] * 2;
4153 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4154 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4157 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4158 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4160 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4161 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4162 DAG.getNode(ISD::BUILD_VECTOR, dl,
4163 MVT::v16i8, &pshufbMask[0], 16));
4164 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4165 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4168 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4169 // and update MaskVals with new element order.
4170 BitVector InOrder(8);
4171 if (BestLoQuad >= 0) {
4172 SmallVector<int, 8> MaskV;
4173 for (int i = 0; i != 4; ++i) {
4174 int idx = MaskVals[i];
4176 MaskV.push_back(-1);
4178 } else if ((idx / 4) == BestLoQuad) {
4179 MaskV.push_back(idx & 3);
4182 MaskV.push_back(-1);
4185 for (unsigned i = 4; i != 8; ++i)
4187 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4191 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4192 // and update MaskVals with the new element order.
4193 if (BestHiQuad >= 0) {
4194 SmallVector<int, 8> MaskV;
4195 for (unsigned i = 0; i != 4; ++i)
4197 for (unsigned i = 4; i != 8; ++i) {
4198 int idx = MaskVals[i];
4200 MaskV.push_back(-1);
4202 } else if ((idx / 4) == BestHiQuad) {
4203 MaskV.push_back((idx & 3) + 4);
4206 MaskV.push_back(-1);
4209 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4213 // In case BestHi & BestLo were both -1, which means each quadword has a word
4214 // from each of the four input quadwords, calculate the InOrder bitvector now
4215 // before falling through to the insert/extract cleanup.
4216 if (BestLoQuad == -1 && BestHiQuad == -1) {
4218 for (int i = 0; i != 8; ++i)
4219 if (MaskVals[i] < 0 || MaskVals[i] == i)
4223 // The other elements are put in the right place using pextrw and pinsrw.
4224 for (unsigned i = 0; i != 8; ++i) {
4227 int EltIdx = MaskVals[i];
4230 SDValue ExtOp = (EltIdx < 8)
4231 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4232 DAG.getIntPtrConstant(EltIdx))
4233 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4234 DAG.getIntPtrConstant(EltIdx - 8));
4235 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4236 DAG.getIntPtrConstant(i));
4241 // v16i8 shuffles - Prefer shuffles in the following order:
4242 // 1. [ssse3] 1 x pshufb
4243 // 2. [ssse3] 2 x pshufb + 1 x por
4244 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4246 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4247 SelectionDAG &DAG, X86TargetLowering &TLI) {
4248 SDValue V1 = SVOp->getOperand(0);
4249 SDValue V2 = SVOp->getOperand(1);
4250 DebugLoc dl = SVOp->getDebugLoc();
4251 SmallVector<int, 16> MaskVals;
4252 SVOp->getMask(MaskVals);
4254 // If we have SSSE3, case 1 is generated when all result bytes come from
4255 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4256 // present, fall back to case 3.
4257 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4260 for (unsigned i = 0; i < 16; ++i) {
4261 int EltIdx = MaskVals[i];
4270 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4271 if (TLI.getSubtarget()->hasSSSE3()) {
4272 SmallVector<SDValue,16> pshufbMask;
4274 // If all result elements are from one input vector, then only translate
4275 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4277 // Otherwise, we have elements from both input vectors, and must zero out
4278 // elements that come from V2 in the first mask, and V1 in the second mask
4279 // so that we can OR them together.
4280 bool TwoInputs = !(V1Only || V2Only);
4281 for (unsigned i = 0; i != 16; ++i) {
4282 int EltIdx = MaskVals[i];
4283 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4284 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4287 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4289 // If all the elements are from V2, assign it to V1 and return after
4290 // building the first pshufb.
4293 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4294 DAG.getNode(ISD::BUILD_VECTOR, dl,
4295 MVT::v16i8, &pshufbMask[0], 16));
4299 // Calculate the shuffle mask for the second input, shuffle it, and
4300 // OR it with the first shuffled input.
4302 for (unsigned i = 0; i != 16; ++i) {
4303 int EltIdx = MaskVals[i];
4305 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4308 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4310 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4311 DAG.getNode(ISD::BUILD_VECTOR, dl,
4312 MVT::v16i8, &pshufbMask[0], 16));
4313 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4316 // No SSSE3 - Calculate in place words and then fix all out of place words
4317 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4318 // the 16 different words that comprise the two doublequadword input vectors.
4319 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4320 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4321 SDValue NewV = V2Only ? V2 : V1;
4322 for (int i = 0; i != 8; ++i) {
4323 int Elt0 = MaskVals[i*2];
4324 int Elt1 = MaskVals[i*2+1];
4326 // This word of the result is all undef, skip it.
4327 if (Elt0 < 0 && Elt1 < 0)
4330 // This word of the result is already in the correct place, skip it.
4331 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4333 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4336 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4337 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4340 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4341 // using a single extract together, load it and store it.
4342 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4343 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4344 DAG.getIntPtrConstant(Elt1 / 2));
4345 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4346 DAG.getIntPtrConstant(i));
4350 // If Elt1 is defined, extract it from the appropriate source. If the
4351 // source byte is not also odd, shift the extracted word left 8 bits
4352 // otherwise clear the bottom 8 bits if we need to do an or.
4354 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4355 DAG.getIntPtrConstant(Elt1 / 2));
4356 if ((Elt1 & 1) == 0)
4357 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4358 DAG.getConstant(8, TLI.getShiftAmountTy()));
4360 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4361 DAG.getConstant(0xFF00, MVT::i16));
4363 // If Elt0 is defined, extract it from the appropriate source. If the
4364 // source byte is not also even, shift the extracted word right 8 bits. If
4365 // Elt1 was also defined, OR the extracted values together before
4366 // inserting them in the result.
4368 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4369 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4370 if ((Elt0 & 1) != 0)
4371 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4372 DAG.getConstant(8, TLI.getShiftAmountTy()));
4374 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4375 DAG.getConstant(0x00FF, MVT::i16));
4376 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4379 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4380 DAG.getIntPtrConstant(i));
4382 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4385 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4386 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4387 /// done when every pair / quad of shuffle mask elements point to elements in
4388 /// the right sequence. e.g.
4389 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4391 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4393 TargetLowering &TLI, DebugLoc dl) {
4394 EVT VT = SVOp->getValueType(0);
4395 SDValue V1 = SVOp->getOperand(0);
4396 SDValue V2 = SVOp->getOperand(1);
4397 unsigned NumElems = VT.getVectorNumElements();
4398 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4399 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4400 EVT MaskEltVT = MaskVT.getVectorElementType();
4402 switch (VT.getSimpleVT().SimpleTy) {
4403 default: assert(false && "Unexpected!");
4404 case MVT::v4f32: NewVT = MVT::v2f64; break;
4405 case MVT::v4i32: NewVT = MVT::v2i64; break;
4406 case MVT::v8i16: NewVT = MVT::v4i32; break;
4407 case MVT::v16i8: NewVT = MVT::v4i32; break;
4410 if (NewWidth == 2) {
4416 int Scale = NumElems / NewWidth;
4417 SmallVector<int, 8> MaskVec;
4418 for (unsigned i = 0; i < NumElems; i += Scale) {
4420 for (int j = 0; j < Scale; ++j) {
4421 int EltIdx = SVOp->getMaskElt(i+j);
4425 StartIdx = EltIdx - (EltIdx % Scale);
4426 if (EltIdx != StartIdx + j)
4430 MaskVec.push_back(-1);
4432 MaskVec.push_back(StartIdx / Scale);
4435 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4436 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4437 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4440 /// getVZextMovL - Return a zero-extending vector move low node.
4442 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4443 SDValue SrcOp, SelectionDAG &DAG,
4444 const X86Subtarget *Subtarget, DebugLoc dl) {
4445 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4446 LoadSDNode *LD = NULL;
4447 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4448 LD = dyn_cast<LoadSDNode>(SrcOp);
4450 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4452 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4453 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4454 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4455 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4456 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4458 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4459 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4460 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4461 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4469 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4470 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4471 DAG.getNode(ISD::BIT_CONVERT, dl,
4475 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4478 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4479 SDValue V1 = SVOp->getOperand(0);
4480 SDValue V2 = SVOp->getOperand(1);
4481 DebugLoc dl = SVOp->getDebugLoc();
4482 EVT VT = SVOp->getValueType(0);
4484 SmallVector<std::pair<int, int>, 8> Locs;
4486 SmallVector<int, 8> Mask1(4U, -1);
4487 SmallVector<int, 8> PermMask;
4488 SVOp->getMask(PermMask);
4492 for (unsigned i = 0; i != 4; ++i) {
4493 int Idx = PermMask[i];
4495 Locs[i] = std::make_pair(-1, -1);
4497 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4499 Locs[i] = std::make_pair(0, NumLo);
4503 Locs[i] = std::make_pair(1, NumHi);
4505 Mask1[2+NumHi] = Idx;
4511 if (NumLo <= 2 && NumHi <= 2) {
4512 // If no more than two elements come from either vector. This can be
4513 // implemented with two shuffles. First shuffle gather the elements.
4514 // The second shuffle, which takes the first shuffle as both of its
4515 // vector operands, put the elements into the right order.
4516 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4518 SmallVector<int, 8> Mask2(4U, -1);
4520 for (unsigned i = 0; i != 4; ++i) {
4521 if (Locs[i].first == -1)
4524 unsigned Idx = (i < 2) ? 0 : 4;
4525 Idx += Locs[i].first * 2 + Locs[i].second;
4530 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4531 } else if (NumLo == 3 || NumHi == 3) {
4532 // Otherwise, we must have three elements from one vector, call it X, and
4533 // one element from the other, call it Y. First, use a shufps to build an
4534 // intermediate vector with the one element from Y and the element from X
4535 // that will be in the same half in the final destination (the indexes don't
4536 // matter). Then, use a shufps to build the final vector, taking the half
4537 // containing the element from Y from the intermediate, and the other half
4540 // Normalize it so the 3 elements come from V1.
4541 CommuteVectorShuffleMask(PermMask, VT);
4545 // Find the element from V2.
4547 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4548 int Val = PermMask[HiIndex];
4555 Mask1[0] = PermMask[HiIndex];
4557 Mask1[2] = PermMask[HiIndex^1];
4559 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4562 Mask1[0] = PermMask[0];
4563 Mask1[1] = PermMask[1];
4564 Mask1[2] = HiIndex & 1 ? 6 : 4;
4565 Mask1[3] = HiIndex & 1 ? 4 : 6;
4566 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4568 Mask1[0] = HiIndex & 1 ? 2 : 0;
4569 Mask1[1] = HiIndex & 1 ? 0 : 2;
4570 Mask1[2] = PermMask[2];
4571 Mask1[3] = PermMask[3];
4576 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4580 // Break it into (shuffle shuffle_hi, shuffle_lo).
4582 SmallVector<int,8> LoMask(4U, -1);
4583 SmallVector<int,8> HiMask(4U, -1);
4585 SmallVector<int,8> *MaskPtr = &LoMask;
4586 unsigned MaskIdx = 0;
4589 for (unsigned i = 0; i != 4; ++i) {
4596 int Idx = PermMask[i];
4598 Locs[i] = std::make_pair(-1, -1);
4599 } else if (Idx < 4) {
4600 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4601 (*MaskPtr)[LoIdx] = Idx;
4604 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4605 (*MaskPtr)[HiIdx] = Idx;
4610 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4611 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4612 SmallVector<int, 8> MaskOps;
4613 for (unsigned i = 0; i != 4; ++i) {
4614 if (Locs[i].first == -1) {
4615 MaskOps.push_back(-1);
4617 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4618 MaskOps.push_back(Idx);
4621 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4625 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4626 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4627 SDValue V1 = Op.getOperand(0);
4628 SDValue V2 = Op.getOperand(1);
4629 EVT VT = Op.getValueType();
4630 DebugLoc dl = Op.getDebugLoc();
4631 unsigned NumElems = VT.getVectorNumElements();
4632 bool isMMX = VT.getSizeInBits() == 64;
4633 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4634 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4635 bool V1IsSplat = false;
4636 bool V2IsSplat = false;
4638 if (isZeroShuffle(SVOp))
4639 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4641 // Promote splats to v4f32.
4642 if (SVOp->isSplat()) {
4643 if (isMMX || NumElems < 4)
4645 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4648 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4650 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4651 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4652 if (NewOp.getNode())
4653 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4654 LowerVECTOR_SHUFFLE(NewOp, DAG));
4655 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4656 // FIXME: Figure out a cleaner way to do this.
4657 // Try to make use of movq to zero out the top part.
4658 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4659 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4660 if (NewOp.getNode()) {
4661 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4662 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4663 DAG, Subtarget, dl);
4665 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4666 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4667 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4668 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4669 DAG, Subtarget, dl);
4673 if (X86::isPSHUFDMask(SVOp))
4676 // Check if this can be converted into a logical shift.
4677 bool isLeft = false;
4680 bool isShift = getSubtarget()->hasSSE2() &&
4681 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4682 if (isShift && ShVal.hasOneUse()) {
4683 // If the shifted value has multiple uses, it may be cheaper to use
4684 // v_set0 + movlhps or movhlps, etc.
4685 EVT EltVT = VT.getVectorElementType();
4686 ShAmt *= EltVT.getSizeInBits();
4687 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4690 if (X86::isMOVLMask(SVOp)) {
4693 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4694 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4699 // FIXME: fold these into legal mask.
4700 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4701 X86::isMOVSLDUPMask(SVOp) ||
4702 X86::isMOVHLPSMask(SVOp) ||
4703 X86::isMOVLHPSMask(SVOp) ||
4704 X86::isMOVLPMask(SVOp)))
4707 if (ShouldXformToMOVHLPS(SVOp) ||
4708 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4709 return CommuteVectorShuffle(SVOp, DAG);
4712 // No better options. Use a vshl / vsrl.
4713 EVT EltVT = VT.getVectorElementType();
4714 ShAmt *= EltVT.getSizeInBits();
4715 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4718 bool Commuted = false;
4719 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4720 // 1,1,1,1 -> v8i16 though.
4721 V1IsSplat = isSplatVector(V1.getNode());
4722 V2IsSplat = isSplatVector(V2.getNode());
4724 // Canonicalize the splat or undef, if present, to be on the RHS.
4725 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4726 Op = CommuteVectorShuffle(SVOp, DAG);
4727 SVOp = cast<ShuffleVectorSDNode>(Op);
4728 V1 = SVOp->getOperand(0);
4729 V2 = SVOp->getOperand(1);
4730 std::swap(V1IsSplat, V2IsSplat);
4731 std::swap(V1IsUndef, V2IsUndef);
4735 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4736 // Shuffling low element of v1 into undef, just return v1.
4739 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4740 // the instruction selector will not match, so get a canonical MOVL with
4741 // swapped operands to undo the commute.
4742 return getMOVL(DAG, dl, VT, V2, V1);
4745 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4746 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4747 X86::isUNPCKLMask(SVOp) ||
4748 X86::isUNPCKHMask(SVOp))
4752 // Normalize mask so all entries that point to V2 points to its first
4753 // element then try to match unpck{h|l} again. If match, return a
4754 // new vector_shuffle with the corrected mask.
4755 SDValue NewMask = NormalizeMask(SVOp, DAG);
4756 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4757 if (NSVOp != SVOp) {
4758 if (X86::isUNPCKLMask(NSVOp, true)) {
4760 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4767 // Commute is back and try unpck* again.
4768 // FIXME: this seems wrong.
4769 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4770 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4771 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4772 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4773 X86::isUNPCKLMask(NewSVOp) ||
4774 X86::isUNPCKHMask(NewSVOp))
4778 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4780 // Normalize the node to match x86 shuffle ops if needed
4781 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4782 return CommuteVectorShuffle(SVOp, DAG);
4784 // Check for legal shuffle and return?
4785 SmallVector<int, 16> PermMask;
4786 SVOp->getMask(PermMask);
4787 if (isShuffleMaskLegal(PermMask, VT))
4790 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4791 if (VT == MVT::v8i16) {
4792 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4793 if (NewOp.getNode())
4797 if (VT == MVT::v16i8) {
4798 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4799 if (NewOp.getNode())
4803 // Handle all 4 wide cases with a number of shuffles except for MMX.
4804 if (NumElems == 4 && !isMMX)
4805 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4811 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4812 SelectionDAG &DAG) {
4813 EVT VT = Op.getValueType();
4814 DebugLoc dl = Op.getDebugLoc();
4815 if (VT.getSizeInBits() == 8) {
4816 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4817 Op.getOperand(0), Op.getOperand(1));
4818 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4819 DAG.getValueType(VT));
4820 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4821 } else if (VT.getSizeInBits() == 16) {
4822 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4823 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4825 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4826 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4827 DAG.getNode(ISD::BIT_CONVERT, dl,
4831 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4832 Op.getOperand(0), Op.getOperand(1));
4833 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4834 DAG.getValueType(VT));
4835 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4836 } else if (VT == MVT::f32) {
4837 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4838 // the result back to FR32 register. It's only worth matching if the
4839 // result has a single use which is a store or a bitcast to i32. And in
4840 // the case of a store, it's not worth it if the index is a constant 0,
4841 // because a MOVSSmr can be used instead, which is smaller and faster.
4842 if (!Op.hasOneUse())
4844 SDNode *User = *Op.getNode()->use_begin();
4845 if ((User->getOpcode() != ISD::STORE ||
4846 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4847 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4848 (User->getOpcode() != ISD::BIT_CONVERT ||
4849 User->getValueType(0) != MVT::i32))
4851 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4852 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4855 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4856 } else if (VT == MVT::i32) {
4857 // ExtractPS works with constant index.
4858 if (isa<ConstantSDNode>(Op.getOperand(1)))
4866 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4867 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4870 if (Subtarget->hasSSE41()) {
4871 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4876 EVT VT = Op.getValueType();
4877 DebugLoc dl = Op.getDebugLoc();
4878 // TODO: handle v16i8.
4879 if (VT.getSizeInBits() == 16) {
4880 SDValue Vec = Op.getOperand(0);
4881 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4883 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4884 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4885 DAG.getNode(ISD::BIT_CONVERT, dl,
4888 // Transform it so it match pextrw which produces a 32-bit result.
4889 EVT EltVT = MVT::i32;
4890 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4891 Op.getOperand(0), Op.getOperand(1));
4892 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4893 DAG.getValueType(VT));
4894 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4895 } else if (VT.getSizeInBits() == 32) {
4896 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4900 // SHUFPS the element to the lowest double word, then movss.
4901 int Mask[4] = { Idx, -1, -1, -1 };
4902 EVT VVT = Op.getOperand(0).getValueType();
4903 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4904 DAG.getUNDEF(VVT), Mask);
4905 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4906 DAG.getIntPtrConstant(0));
4907 } else if (VT.getSizeInBits() == 64) {
4908 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4909 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4910 // to match extract_elt for f64.
4911 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4915 // UNPCKHPD the element to the lowest double word, then movsd.
4916 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4917 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4918 int Mask[2] = { 1, -1 };
4919 EVT VVT = Op.getOperand(0).getValueType();
4920 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4921 DAG.getUNDEF(VVT), Mask);
4922 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4923 DAG.getIntPtrConstant(0));
4930 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4931 EVT VT = Op.getValueType();
4932 EVT EltVT = VT.getVectorElementType();
4933 DebugLoc dl = Op.getDebugLoc();
4935 SDValue N0 = Op.getOperand(0);
4936 SDValue N1 = Op.getOperand(1);
4937 SDValue N2 = Op.getOperand(2);
4939 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4940 isa<ConstantSDNode>(N2)) {
4942 if (VT == MVT::v8i16)
4943 Opc = X86ISD::PINSRW;
4944 else if (VT == MVT::v4i16)
4945 Opc = X86ISD::MMX_PINSRW;
4946 else if (VT == MVT::v16i8)
4947 Opc = X86ISD::PINSRB;
4949 Opc = X86ISD::PINSRB;
4951 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4953 if (N1.getValueType() != MVT::i32)
4954 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4955 if (N2.getValueType() != MVT::i32)
4956 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4957 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4958 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4959 // Bits [7:6] of the constant are the source select. This will always be
4960 // zero here. The DAG Combiner may combine an extract_elt index into these
4961 // bits. For example (insert (extract, 3), 2) could be matched by putting
4962 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4963 // Bits [5:4] of the constant are the destination select. This is the
4964 // value of the incoming immediate.
4965 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4966 // combine either bitwise AND or insert of float 0.0 to set these bits.
4967 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4968 // Create this as a scalar to vector..
4969 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4970 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4971 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4972 // PINSR* works with constant index.
4979 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4980 EVT VT = Op.getValueType();
4981 EVT EltVT = VT.getVectorElementType();
4983 if (Subtarget->hasSSE41())
4984 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4986 if (EltVT == MVT::i8)
4989 DebugLoc dl = Op.getDebugLoc();
4990 SDValue N0 = Op.getOperand(0);
4991 SDValue N1 = Op.getOperand(1);
4992 SDValue N2 = Op.getOperand(2);
4994 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4995 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4996 // as its second argument.
4997 if (N1.getValueType() != MVT::i32)
4998 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4999 if (N2.getValueType() != MVT::i32)
5000 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5001 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5002 dl, VT, N0, N1, N2);
5008 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
5009 DebugLoc dl = Op.getDebugLoc();
5010 if (Op.getValueType() == MVT::v2f32)
5011 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5012 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5013 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5014 Op.getOperand(0))));
5016 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5017 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5019 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5020 EVT VT = MVT::v2i32;
5021 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5028 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5029 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5032 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5033 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5034 // one of the above mentioned nodes. It has to be wrapped because otherwise
5035 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5036 // be used to form addressing mode. These wrapped nodes will be selected
5039 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
5040 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5042 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5044 unsigned char OpFlag = 0;
5045 unsigned WrapperKind = X86ISD::Wrapper;
5046 CodeModel::Model M = getTargetMachine().getCodeModel();
5048 if (Subtarget->isPICStyleRIPRel() &&
5049 (M == CodeModel::Small || M == CodeModel::Kernel))
5050 WrapperKind = X86ISD::WrapperRIP;
5051 else if (Subtarget->isPICStyleGOT())
5052 OpFlag = X86II::MO_GOTOFF;
5053 else if (Subtarget->isPICStyleStubPIC())
5054 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5056 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5058 CP->getOffset(), OpFlag);
5059 DebugLoc DL = CP->getDebugLoc();
5060 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5061 // With PIC, the address is actually $g + Offset.
5063 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5064 DAG.getNode(X86ISD::GlobalBaseReg,
5065 DebugLoc::getUnknownLoc(), getPointerTy()),
5072 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
5073 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5075 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5077 unsigned char OpFlag = 0;
5078 unsigned WrapperKind = X86ISD::Wrapper;
5079 CodeModel::Model M = getTargetMachine().getCodeModel();
5081 if (Subtarget->isPICStyleRIPRel() &&
5082 (M == CodeModel::Small || M == CodeModel::Kernel))
5083 WrapperKind = X86ISD::WrapperRIP;
5084 else if (Subtarget->isPICStyleGOT())
5085 OpFlag = X86II::MO_GOTOFF;
5086 else if (Subtarget->isPICStyleStubPIC())
5087 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5089 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5091 DebugLoc DL = JT->getDebugLoc();
5092 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5094 // With PIC, the address is actually $g + Offset.
5096 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5097 DAG.getNode(X86ISD::GlobalBaseReg,
5098 DebugLoc::getUnknownLoc(), getPointerTy()),
5106 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
5107 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5109 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5111 unsigned char OpFlag = 0;
5112 unsigned WrapperKind = X86ISD::Wrapper;
5113 CodeModel::Model M = getTargetMachine().getCodeModel();
5115 if (Subtarget->isPICStyleRIPRel() &&
5116 (M == CodeModel::Small || M == CodeModel::Kernel))
5117 WrapperKind = X86ISD::WrapperRIP;
5118 else if (Subtarget->isPICStyleGOT())
5119 OpFlag = X86II::MO_GOTOFF;
5120 else if (Subtarget->isPICStyleStubPIC())
5121 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5123 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5125 DebugLoc DL = Op.getDebugLoc();
5126 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5129 // With PIC, the address is actually $g + Offset.
5130 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5131 !Subtarget->is64Bit()) {
5132 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5133 DAG.getNode(X86ISD::GlobalBaseReg,
5134 DebugLoc::getUnknownLoc(),
5143 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5144 // Create the TargetBlockAddressAddress node.
5145 unsigned char OpFlags =
5146 Subtarget->ClassifyBlockAddressReference();
5147 CodeModel::Model M = getTargetMachine().getCodeModel();
5148 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5149 DebugLoc dl = Op.getDebugLoc();
5150 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5151 /*isTarget=*/true, OpFlags);
5153 if (Subtarget->isPICStyleRIPRel() &&
5154 (M == CodeModel::Small || M == CodeModel::Kernel))
5155 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5157 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5159 // With PIC, the address is actually $g + Offset.
5160 if (isGlobalRelativeToPICBase(OpFlags)) {
5161 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5162 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5170 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5172 SelectionDAG &DAG) const {
5173 // Create the TargetGlobalAddress node, folding in the constant
5174 // offset if it is legal.
5175 unsigned char OpFlags =
5176 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5177 CodeModel::Model M = getTargetMachine().getCodeModel();
5179 if (OpFlags == X86II::MO_NO_FLAG &&
5180 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5181 // A direct static reference to a global.
5182 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5185 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5188 if (Subtarget->isPICStyleRIPRel() &&
5189 (M == CodeModel::Small || M == CodeModel::Kernel))
5190 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5192 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5194 // With PIC, the address is actually $g + Offset.
5195 if (isGlobalRelativeToPICBase(OpFlags)) {
5196 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5197 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5201 // For globals that require a load from a stub to get the address, emit the
5203 if (isGlobalStubReference(OpFlags))
5204 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5205 PseudoSourceValue::getGOT(), 0, false, false, 0);
5207 // If there was a non-zero offset that we didn't fold, create an explicit
5210 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5211 DAG.getConstant(Offset, getPointerTy()));
5217 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5218 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5219 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5220 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5224 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5225 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5226 unsigned char OperandFlags) {
5227 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5228 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5229 DebugLoc dl = GA->getDebugLoc();
5230 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5231 GA->getValueType(0),
5235 SDValue Ops[] = { Chain, TGA, *InFlag };
5236 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5238 SDValue Ops[] = { Chain, TGA };
5239 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5242 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5243 MFI->setHasCalls(true);
5245 SDValue Flag = Chain.getValue(1);
5246 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5249 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5251 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5254 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5255 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5256 DAG.getNode(X86ISD::GlobalBaseReg,
5257 DebugLoc::getUnknownLoc(),
5259 InFlag = Chain.getValue(1);
5261 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5264 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5266 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5268 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5269 X86::RAX, X86II::MO_TLSGD);
5272 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5273 // "local exec" model.
5274 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5275 const EVT PtrVT, TLSModel::Model model,
5277 DebugLoc dl = GA->getDebugLoc();
5278 // Get the Thread Pointer
5279 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5280 DebugLoc::getUnknownLoc(), PtrVT,
5281 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5284 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5285 NULL, 0, false, false, 0);
5287 unsigned char OperandFlags = 0;
5288 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5290 unsigned WrapperKind = X86ISD::Wrapper;
5291 if (model == TLSModel::LocalExec) {
5292 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5293 } else if (is64Bit) {
5294 assert(model == TLSModel::InitialExec);
5295 OperandFlags = X86II::MO_GOTTPOFF;
5296 WrapperKind = X86ISD::WrapperRIP;
5298 assert(model == TLSModel::InitialExec);
5299 OperandFlags = X86II::MO_INDNTPOFF;
5302 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5304 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5305 GA->getOffset(), OperandFlags);
5306 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5308 if (model == TLSModel::InitialExec)
5309 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5310 PseudoSourceValue::getGOT(), 0, false, false, 0);
5312 // The address of the thread local variable is the add of the thread
5313 // pointer with the offset of the variable.
5314 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5318 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5319 // TODO: implement the "local dynamic" model
5320 // TODO: implement the "initial exec"model for pic executables
5321 assert(Subtarget->isTargetELF() &&
5322 "TLS not implemented for non-ELF targets");
5323 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5324 const GlobalValue *GV = GA->getGlobal();
5326 // If GV is an alias then use the aliasee for determining
5327 // thread-localness.
5328 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5329 GV = GA->resolveAliasedGlobal(false);
5331 TLSModel::Model model = getTLSModel(GV,
5332 getTargetMachine().getRelocationModel());
5335 case TLSModel::GeneralDynamic:
5336 case TLSModel::LocalDynamic: // not implemented
5337 if (Subtarget->is64Bit())
5338 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5339 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5341 case TLSModel::InitialExec:
5342 case TLSModel::LocalExec:
5343 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5344 Subtarget->is64Bit());
5347 llvm_unreachable("Unreachable");
5352 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5353 /// take a 2 x i32 value to shift plus a shift amount.
5354 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5355 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5356 EVT VT = Op.getValueType();
5357 unsigned VTBits = VT.getSizeInBits();
5358 DebugLoc dl = Op.getDebugLoc();
5359 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5360 SDValue ShOpLo = Op.getOperand(0);
5361 SDValue ShOpHi = Op.getOperand(1);
5362 SDValue ShAmt = Op.getOperand(2);
5363 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5364 DAG.getConstant(VTBits - 1, MVT::i8))
5365 : DAG.getConstant(0, VT);
5368 if (Op.getOpcode() == ISD::SHL_PARTS) {
5369 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5370 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5372 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5373 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5376 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5377 DAG.getConstant(VTBits, MVT::i8));
5378 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5379 AndNode, DAG.getConstant(0, MVT::i8));
5382 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5383 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5384 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5386 if (Op.getOpcode() == ISD::SHL_PARTS) {
5387 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5388 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5390 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5391 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5394 SDValue Ops[2] = { Lo, Hi };
5395 return DAG.getMergeValues(Ops, 2, dl);
5398 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5399 EVT SrcVT = Op.getOperand(0).getValueType();
5401 if (SrcVT.isVector()) {
5402 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5408 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5409 "Unknown SINT_TO_FP to lower!");
5411 // These are really Legal; return the operand so the caller accepts it as
5413 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5415 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5416 Subtarget->is64Bit()) {
5420 DebugLoc dl = Op.getDebugLoc();
5421 unsigned Size = SrcVT.getSizeInBits()/8;
5422 MachineFunction &MF = DAG.getMachineFunction();
5423 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5424 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5425 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5427 PseudoSourceValue::getFixedStack(SSFI), 0,
5429 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5432 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5434 SelectionDAG &DAG) {
5436 DebugLoc dl = Op.getDebugLoc();
5438 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5440 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5442 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5443 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5444 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5445 Tys, Ops, array_lengthof(Ops));
5448 Chain = Result.getValue(1);
5449 SDValue InFlag = Result.getValue(2);
5451 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5452 // shouldn't be necessary except that RFP cannot be live across
5453 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5454 MachineFunction &MF = DAG.getMachineFunction();
5455 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5456 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5457 Tys = DAG.getVTList(MVT::Other);
5459 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5461 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5462 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5463 PseudoSourceValue::getFixedStack(SSFI), 0,
5470 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5471 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5472 // This algorithm is not obvious. Here it is in C code, more or less:
5474 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5475 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5476 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5478 // Copy ints to xmm registers.
5479 __m128i xh = _mm_cvtsi32_si128( hi );
5480 __m128i xl = _mm_cvtsi32_si128( lo );
5482 // Combine into low half of a single xmm register.
5483 __m128i x = _mm_unpacklo_epi32( xh, xl );
5487 // Merge in appropriate exponents to give the integer bits the right
5489 x = _mm_unpacklo_epi32( x, exp );
5491 // Subtract away the biases to deal with the IEEE-754 double precision
5493 d = _mm_sub_pd( (__m128d) x, bias );
5495 // All conversions up to here are exact. The correctly rounded result is
5496 // calculated using the current rounding mode using the following
5498 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5499 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5500 // store doesn't really need to be here (except
5501 // maybe to zero the other double)
5506 DebugLoc dl = Op.getDebugLoc();
5507 LLVMContext *Context = DAG.getContext();
5509 // Build some magic constants.
5510 std::vector<Constant*> CV0;
5511 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5512 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5513 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5514 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5515 Constant *C0 = ConstantVector::get(CV0);
5516 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5518 std::vector<Constant*> CV1;
5520 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5522 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5523 Constant *C1 = ConstantVector::get(CV1);
5524 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5526 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5527 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5529 DAG.getIntPtrConstant(1)));
5530 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5531 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5533 DAG.getIntPtrConstant(0)));
5534 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5535 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5536 PseudoSourceValue::getConstantPool(), 0,
5538 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5539 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5540 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5541 PseudoSourceValue::getConstantPool(), 0,
5543 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5545 // Add the halves; easiest way is to swap them into another reg first.
5546 int ShufMask[2] = { 1, -1 };
5547 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5548 DAG.getUNDEF(MVT::v2f64), ShufMask);
5549 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5550 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5551 DAG.getIntPtrConstant(0));
5554 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5555 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5556 DebugLoc dl = Op.getDebugLoc();
5557 // FP constant to bias correct the final result.
5558 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5561 // Load the 32-bit value into an XMM register.
5562 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5563 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5565 DAG.getIntPtrConstant(0)));
5567 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5568 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5569 DAG.getIntPtrConstant(0));
5571 // Or the load with the bias.
5572 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5573 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5574 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5576 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5577 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5578 MVT::v2f64, Bias)));
5579 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5580 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5581 DAG.getIntPtrConstant(0));
5583 // Subtract the bias.
5584 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5586 // Handle final rounding.
5587 EVT DestVT = Op.getValueType();
5589 if (DestVT.bitsLT(MVT::f64)) {
5590 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5591 DAG.getIntPtrConstant(0));
5592 } else if (DestVT.bitsGT(MVT::f64)) {
5593 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5596 // Handle final rounding.
5600 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5601 SDValue N0 = Op.getOperand(0);
5602 DebugLoc dl = Op.getDebugLoc();
5604 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5605 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5606 // the optimization here.
5607 if (DAG.SignBitIsZero(N0))
5608 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5610 EVT SrcVT = N0.getValueType();
5611 if (SrcVT == MVT::i64) {
5612 // We only handle SSE2 f64 target here; caller can expand the rest.
5613 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5616 return LowerUINT_TO_FP_i64(Op, DAG);
5617 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5618 return LowerUINT_TO_FP_i32(Op, DAG);
5621 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5623 // Make a 64-bit buffer, and use it to build an FILD.
5624 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5625 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5626 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5627 getPointerTy(), StackSlot, WordOff);
5628 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5629 StackSlot, NULL, 0, false, false, 0);
5630 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5631 OffsetSlot, NULL, 0, false, false, 0);
5632 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5635 std::pair<SDValue,SDValue> X86TargetLowering::
5636 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5637 DebugLoc dl = Op.getDebugLoc();
5639 EVT DstTy = Op.getValueType();
5642 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5646 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5647 DstTy.getSimpleVT() >= MVT::i16 &&
5648 "Unknown FP_TO_SINT to lower!");
5650 // These are really Legal.
5651 if (DstTy == MVT::i32 &&
5652 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5653 return std::make_pair(SDValue(), SDValue());
5654 if (Subtarget->is64Bit() &&
5655 DstTy == MVT::i64 &&
5656 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5657 return std::make_pair(SDValue(), SDValue());
5659 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5661 MachineFunction &MF = DAG.getMachineFunction();
5662 unsigned MemSize = DstTy.getSizeInBits()/8;
5663 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5664 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5667 switch (DstTy.getSimpleVT().SimpleTy) {
5668 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5669 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5670 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5671 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5674 SDValue Chain = DAG.getEntryNode();
5675 SDValue Value = Op.getOperand(0);
5676 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5677 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5678 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5679 PseudoSourceValue::getFixedStack(SSFI), 0,
5681 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5683 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5685 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5686 Chain = Value.getValue(1);
5687 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5688 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5691 // Build the FP_TO_INT*_IN_MEM
5692 SDValue Ops[] = { Chain, Value, StackSlot };
5693 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5695 return std::make_pair(FIST, StackSlot);
5698 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5699 if (Op.getValueType().isVector()) {
5700 if (Op.getValueType() == MVT::v2i32 &&
5701 Op.getOperand(0).getValueType() == MVT::v2f64) {
5707 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5708 SDValue FIST = Vals.first, StackSlot = Vals.second;
5709 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5710 if (FIST.getNode() == 0) return Op;
5713 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5714 FIST, StackSlot, NULL, 0, false, false, 0);
5717 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5718 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5719 SDValue FIST = Vals.first, StackSlot = Vals.second;
5720 assert(FIST.getNode() && "Unexpected failure");
5723 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5724 FIST, StackSlot, NULL, 0, false, false, 0);
5727 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5728 LLVMContext *Context = DAG.getContext();
5729 DebugLoc dl = Op.getDebugLoc();
5730 EVT VT = Op.getValueType();
5733 EltVT = VT.getVectorElementType();
5734 std::vector<Constant*> CV;
5735 if (EltVT == MVT::f64) {
5736 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5740 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5746 Constant *C = ConstantVector::get(CV);
5747 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5748 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5749 PseudoSourceValue::getConstantPool(), 0,
5751 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5754 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5755 LLVMContext *Context = DAG.getContext();
5756 DebugLoc dl = Op.getDebugLoc();
5757 EVT VT = Op.getValueType();
5760 EltVT = VT.getVectorElementType();
5761 std::vector<Constant*> CV;
5762 if (EltVT == MVT::f64) {
5763 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5767 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5773 Constant *C = ConstantVector::get(CV);
5774 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5775 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5776 PseudoSourceValue::getConstantPool(), 0,
5778 if (VT.isVector()) {
5779 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5780 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5781 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5783 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5785 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5789 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5790 LLVMContext *Context = DAG.getContext();
5791 SDValue Op0 = Op.getOperand(0);
5792 SDValue Op1 = Op.getOperand(1);
5793 DebugLoc dl = Op.getDebugLoc();
5794 EVT VT = Op.getValueType();
5795 EVT SrcVT = Op1.getValueType();
5797 // If second operand is smaller, extend it first.
5798 if (SrcVT.bitsLT(VT)) {
5799 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5802 // And if it is bigger, shrink it first.
5803 if (SrcVT.bitsGT(VT)) {
5804 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5808 // At this point the operands and the result should have the same
5809 // type, and that won't be f80 since that is not custom lowered.
5811 // First get the sign bit of second operand.
5812 std::vector<Constant*> CV;
5813 if (SrcVT == MVT::f64) {
5814 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5815 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5817 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5818 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5819 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5820 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5822 Constant *C = ConstantVector::get(CV);
5823 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5824 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5825 PseudoSourceValue::getConstantPool(), 0,
5827 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5829 // Shift sign bit right or left if the two operands have different types.
5830 if (SrcVT.bitsGT(VT)) {
5831 // Op0 is MVT::f32, Op1 is MVT::f64.
5832 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5833 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5834 DAG.getConstant(32, MVT::i32));
5835 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5836 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5837 DAG.getIntPtrConstant(0));
5840 // Clear first operand sign bit.
5842 if (VT == MVT::f64) {
5843 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5844 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5846 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5847 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5848 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5849 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5851 C = ConstantVector::get(CV);
5852 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5853 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5854 PseudoSourceValue::getConstantPool(), 0,
5856 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5858 // Or the value with the sign bit.
5859 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5862 /// Emit nodes that will be selected as "test Op0,Op0", or something
5864 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5865 SelectionDAG &DAG) {
5866 DebugLoc dl = Op.getDebugLoc();
5868 // CF and OF aren't always set the way we want. Determine which
5869 // of these we need.
5870 bool NeedCF = false;
5871 bool NeedOF = false;
5873 case X86::COND_A: case X86::COND_AE:
5874 case X86::COND_B: case X86::COND_BE:
5877 case X86::COND_G: case X86::COND_GE:
5878 case X86::COND_L: case X86::COND_LE:
5879 case X86::COND_O: case X86::COND_NO:
5885 // See if we can use the EFLAGS value from the operand instead of
5886 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5887 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5888 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5889 unsigned Opcode = 0;
5890 unsigned NumOperands = 0;
5891 switch (Op.getNode()->getOpcode()) {
5893 // Due to an isel shortcoming, be conservative if this add is likely to
5894 // be selected as part of a load-modify-store instruction. When the root
5895 // node in a match is a store, isel doesn't know how to remap non-chain
5896 // non-flag uses of other nodes in the match, such as the ADD in this
5897 // case. This leads to the ADD being left around and reselected, with
5898 // the result being two adds in the output.
5899 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5900 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5901 if (UI->getOpcode() == ISD::STORE)
5903 if (ConstantSDNode *C =
5904 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5905 // An add of one will be selected as an INC.
5906 if (C->getAPIntValue() == 1) {
5907 Opcode = X86ISD::INC;
5911 // An add of negative one (subtract of one) will be selected as a DEC.
5912 if (C->getAPIntValue().isAllOnesValue()) {
5913 Opcode = X86ISD::DEC;
5918 // Otherwise use a regular EFLAGS-setting add.
5919 Opcode = X86ISD::ADD;
5923 // If the primary and result isn't used, don't bother using X86ISD::AND,
5924 // because a TEST instruction will be better.
5925 bool NonFlagUse = false;
5926 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5927 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5929 unsigned UOpNo = UI.getOperandNo();
5930 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5931 // Look pass truncate.
5932 UOpNo = User->use_begin().getOperandNo();
5933 User = *User->use_begin();
5935 if (User->getOpcode() != ISD::BRCOND &&
5936 User->getOpcode() != ISD::SETCC &&
5937 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5949 // Due to the ISEL shortcoming noted above, be conservative if this op is
5950 // likely to be selected as part of a load-modify-store instruction.
5951 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5952 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5953 if (UI->getOpcode() == ISD::STORE)
5955 // Otherwise use a regular EFLAGS-setting instruction.
5956 switch (Op.getNode()->getOpcode()) {
5957 case ISD::SUB: Opcode = X86ISD::SUB; break;
5958 case ISD::OR: Opcode = X86ISD::OR; break;
5959 case ISD::XOR: Opcode = X86ISD::XOR; break;
5960 case ISD::AND: Opcode = X86ISD::AND; break;
5961 default: llvm_unreachable("unexpected operator!");
5972 return SDValue(Op.getNode(), 1);
5978 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5979 SmallVector<SDValue, 4> Ops;
5980 for (unsigned i = 0; i != NumOperands; ++i)
5981 Ops.push_back(Op.getOperand(i));
5982 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5983 DAG.ReplaceAllUsesWith(Op, New);
5984 return SDValue(New.getNode(), 1);
5988 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5989 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5990 DAG.getConstant(0, Op.getValueType()));
5993 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5995 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5996 SelectionDAG &DAG) {
5997 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5998 if (C->getAPIntValue() == 0)
5999 return EmitTest(Op0, X86CC, DAG);
6001 DebugLoc dl = Op0.getDebugLoc();
6002 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6005 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6006 /// if it's possible.
6007 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
6008 DebugLoc dl, SelectionDAG &DAG) {
6009 SDValue Op0 = And.getOperand(0);
6010 SDValue Op1 = And.getOperand(1);
6011 if (Op0.getOpcode() == ISD::TRUNCATE)
6012 Op0 = Op0.getOperand(0);
6013 if (Op1.getOpcode() == ISD::TRUNCATE)
6014 Op1 = Op1.getOperand(0);
6017 if (Op1.getOpcode() == ISD::SHL) {
6018 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6019 if (And10C->getZExtValue() == 1) {
6021 RHS = Op1.getOperand(1);
6023 } else if (Op0.getOpcode() == ISD::SHL) {
6024 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6025 if (And00C->getZExtValue() == 1) {
6027 RHS = Op0.getOperand(1);
6029 } else if (Op1.getOpcode() == ISD::Constant) {
6030 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6031 SDValue AndLHS = Op0;
6032 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6033 LHS = AndLHS.getOperand(0);
6034 RHS = AndLHS.getOperand(1);
6038 if (LHS.getNode()) {
6039 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
6040 // instruction. Since the shift amount is in-range-or-undefined, we know
6041 // that doing a bittest on the i16 value is ok. We extend to i32 because
6042 // the encoding for the i16 version is larger than the i32 version.
6043 if (LHS.getValueType() == MVT::i8)
6044 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6046 // If the operand types disagree, extend the shift amount to match. Since
6047 // BT ignores high bits (like shifts) we can use anyextend.
6048 if (LHS.getValueType() != RHS.getValueType())
6049 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6051 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6052 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6053 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6054 DAG.getConstant(Cond, MVT::i8), BT);
6060 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
6061 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6062 SDValue Op0 = Op.getOperand(0);
6063 SDValue Op1 = Op.getOperand(1);
6064 DebugLoc dl = Op.getDebugLoc();
6065 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6067 // Optimize to BT if possible.
6068 // Lower (X & (1 << N)) == 0 to BT(X, N).
6069 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6070 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6071 if (Op0.getOpcode() == ISD::AND &&
6073 Op1.getOpcode() == ISD::Constant &&
6074 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6075 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6076 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6077 if (NewSetCC.getNode())
6081 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6082 if (Op0.getOpcode() == X86ISD::SETCC &&
6083 Op1.getOpcode() == ISD::Constant &&
6084 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6085 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6086 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6087 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6088 bool Invert = (CC == ISD::SETNE) ^
6089 cast<ConstantSDNode>(Op1)->isNullValue();
6091 CCode = X86::GetOppositeBranchCondition(CCode);
6092 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6093 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6096 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6097 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6098 if (X86CC == X86::COND_INVALID)
6101 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6103 // Use sbb x, x to materialize carry bit into a GPR.
6104 if (X86CC == X86::COND_B)
6105 return DAG.getNode(ISD::AND, dl, MVT::i8,
6106 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6107 DAG.getConstant(X86CC, MVT::i8), Cond),
6108 DAG.getConstant(1, MVT::i8));
6110 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6111 DAG.getConstant(X86CC, MVT::i8), Cond);
6114 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6116 SDValue Op0 = Op.getOperand(0);
6117 SDValue Op1 = Op.getOperand(1);
6118 SDValue CC = Op.getOperand(2);
6119 EVT VT = Op.getValueType();
6120 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6121 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6122 DebugLoc dl = Op.getDebugLoc();
6126 EVT VT0 = Op0.getValueType();
6127 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6128 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6131 switch (SetCCOpcode) {
6134 case ISD::SETEQ: SSECC = 0; break;
6136 case ISD::SETGT: Swap = true; // Fallthrough
6138 case ISD::SETOLT: SSECC = 1; break;
6140 case ISD::SETGE: Swap = true; // Fallthrough
6142 case ISD::SETOLE: SSECC = 2; break;
6143 case ISD::SETUO: SSECC = 3; break;
6145 case ISD::SETNE: SSECC = 4; break;
6146 case ISD::SETULE: Swap = true;
6147 case ISD::SETUGE: SSECC = 5; break;
6148 case ISD::SETULT: Swap = true;
6149 case ISD::SETUGT: SSECC = 6; break;
6150 case ISD::SETO: SSECC = 7; break;
6153 std::swap(Op0, Op1);
6155 // In the two special cases we can't handle, emit two comparisons.
6157 if (SetCCOpcode == ISD::SETUEQ) {
6159 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6160 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6161 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6163 else if (SetCCOpcode == ISD::SETONE) {
6165 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6166 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6167 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6169 llvm_unreachable("Illegal FP comparison");
6171 // Handle all other FP comparisons here.
6172 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6175 // We are handling one of the integer comparisons here. Since SSE only has
6176 // GT and EQ comparisons for integer, swapping operands and multiple
6177 // operations may be required for some comparisons.
6178 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6179 bool Swap = false, Invert = false, FlipSigns = false;
6181 switch (VT.getSimpleVT().SimpleTy) {
6184 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6186 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6188 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6189 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6192 switch (SetCCOpcode) {
6194 case ISD::SETNE: Invert = true;
6195 case ISD::SETEQ: Opc = EQOpc; break;
6196 case ISD::SETLT: Swap = true;
6197 case ISD::SETGT: Opc = GTOpc; break;
6198 case ISD::SETGE: Swap = true;
6199 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6200 case ISD::SETULT: Swap = true;
6201 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6202 case ISD::SETUGE: Swap = true;
6203 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6206 std::swap(Op0, Op1);
6208 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6209 // bits of the inputs before performing those operations.
6211 EVT EltVT = VT.getVectorElementType();
6212 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6214 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6215 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6217 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6218 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6221 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6223 // If the logical-not of the result is required, perform that now.
6225 Result = DAG.getNOT(dl, Result, VT);
6230 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6231 static bool isX86LogicalCmp(SDValue Op) {
6232 unsigned Opc = Op.getNode()->getOpcode();
6233 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6235 if (Op.getResNo() == 1 &&
6236 (Opc == X86ISD::ADD ||
6237 Opc == X86ISD::SUB ||
6238 Opc == X86ISD::SMUL ||
6239 Opc == X86ISD::UMUL ||
6240 Opc == X86ISD::INC ||
6241 Opc == X86ISD::DEC ||
6242 Opc == X86ISD::OR ||
6243 Opc == X86ISD::XOR ||
6244 Opc == X86ISD::AND))
6250 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6251 bool addTest = true;
6252 SDValue Cond = Op.getOperand(0);
6253 DebugLoc dl = Op.getDebugLoc();
6256 if (Cond.getOpcode() == ISD::SETCC) {
6257 SDValue NewCond = LowerSETCC(Cond, DAG);
6258 if (NewCond.getNode())
6262 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6263 SDValue Op1 = Op.getOperand(1);
6264 SDValue Op2 = Op.getOperand(2);
6265 if (Cond.getOpcode() == X86ISD::SETCC &&
6266 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6267 SDValue Cmp = Cond.getOperand(1);
6268 if (Cmp.getOpcode() == X86ISD::CMP) {
6269 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6270 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6271 ConstantSDNode *RHSC =
6272 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6273 if (N1C && N1C->isAllOnesValue() &&
6274 N2C && N2C->isNullValue() &&
6275 RHSC && RHSC->isNullValue()) {
6276 SDValue CmpOp0 = Cmp.getOperand(0);
6277 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6278 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6279 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6280 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6285 // Look pass (and (setcc_carry (cmp ...)), 1).
6286 if (Cond.getOpcode() == ISD::AND &&
6287 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6288 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6289 if (C && C->getAPIntValue() == 1)
6290 Cond = Cond.getOperand(0);
6293 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6294 // setting operand in place of the X86ISD::SETCC.
6295 if (Cond.getOpcode() == X86ISD::SETCC ||
6296 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6297 CC = Cond.getOperand(0);
6299 SDValue Cmp = Cond.getOperand(1);
6300 unsigned Opc = Cmp.getOpcode();
6301 EVT VT = Op.getValueType();
6303 bool IllegalFPCMov = false;
6304 if (VT.isFloatingPoint() && !VT.isVector() &&
6305 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6306 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6308 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6309 Opc == X86ISD::BT) { // FIXME
6316 // Look pass the truncate.
6317 if (Cond.getOpcode() == ISD::TRUNCATE)
6318 Cond = Cond.getOperand(0);
6320 // We know the result of AND is compared against zero. Try to match
6322 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6323 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6324 if (NewSetCC.getNode()) {
6325 CC = NewSetCC.getOperand(0);
6326 Cond = NewSetCC.getOperand(1);
6333 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6334 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6337 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6338 // condition is true.
6339 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6340 SDValue Ops[] = { Op2, Op1, CC, Cond };
6341 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6344 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6345 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6346 // from the AND / OR.
6347 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6348 Opc = Op.getOpcode();
6349 if (Opc != ISD::OR && Opc != ISD::AND)
6351 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6352 Op.getOperand(0).hasOneUse() &&
6353 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6354 Op.getOperand(1).hasOneUse());
6357 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6358 // 1 and that the SETCC node has a single use.
6359 static bool isXor1OfSetCC(SDValue Op) {
6360 if (Op.getOpcode() != ISD::XOR)
6362 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6363 if (N1C && N1C->getAPIntValue() == 1) {
6364 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6365 Op.getOperand(0).hasOneUse();
6370 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6371 bool addTest = true;
6372 SDValue Chain = Op.getOperand(0);
6373 SDValue Cond = Op.getOperand(1);
6374 SDValue Dest = Op.getOperand(2);
6375 DebugLoc dl = Op.getDebugLoc();
6378 if (Cond.getOpcode() == ISD::SETCC) {
6379 SDValue NewCond = LowerSETCC(Cond, DAG);
6380 if (NewCond.getNode())
6384 // FIXME: LowerXALUO doesn't handle these!!
6385 else if (Cond.getOpcode() == X86ISD::ADD ||
6386 Cond.getOpcode() == X86ISD::SUB ||
6387 Cond.getOpcode() == X86ISD::SMUL ||
6388 Cond.getOpcode() == X86ISD::UMUL)
6389 Cond = LowerXALUO(Cond, DAG);
6392 // Look pass (and (setcc_carry (cmp ...)), 1).
6393 if (Cond.getOpcode() == ISD::AND &&
6394 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6395 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6396 if (C && C->getAPIntValue() == 1)
6397 Cond = Cond.getOperand(0);
6400 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6401 // setting operand in place of the X86ISD::SETCC.
6402 if (Cond.getOpcode() == X86ISD::SETCC ||
6403 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6404 CC = Cond.getOperand(0);
6406 SDValue Cmp = Cond.getOperand(1);
6407 unsigned Opc = Cmp.getOpcode();
6408 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6409 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6413 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6417 // These can only come from an arithmetic instruction with overflow,
6418 // e.g. SADDO, UADDO.
6419 Cond = Cond.getNode()->getOperand(1);
6426 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6427 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6428 if (CondOpc == ISD::OR) {
6429 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6430 // two branches instead of an explicit OR instruction with a
6432 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6433 isX86LogicalCmp(Cmp)) {
6434 CC = Cond.getOperand(0).getOperand(0);
6435 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6436 Chain, Dest, CC, Cmp);
6437 CC = Cond.getOperand(1).getOperand(0);
6441 } else { // ISD::AND
6442 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6443 // two branches instead of an explicit AND instruction with a
6444 // separate test. However, we only do this if this block doesn't
6445 // have a fall-through edge, because this requires an explicit
6446 // jmp when the condition is false.
6447 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6448 isX86LogicalCmp(Cmp) &&
6449 Op.getNode()->hasOneUse()) {
6450 X86::CondCode CCode =
6451 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6452 CCode = X86::GetOppositeBranchCondition(CCode);
6453 CC = DAG.getConstant(CCode, MVT::i8);
6454 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6455 // Look for an unconditional branch following this conditional branch.
6456 // We need this because we need to reverse the successors in order
6457 // to implement FCMP_OEQ.
6458 if (User.getOpcode() == ISD::BR) {
6459 SDValue FalseBB = User.getOperand(1);
6461 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6462 assert(NewBR == User);
6465 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6466 Chain, Dest, CC, Cmp);
6467 X86::CondCode CCode =
6468 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6469 CCode = X86::GetOppositeBranchCondition(CCode);
6470 CC = DAG.getConstant(CCode, MVT::i8);
6476 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6477 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6478 // It should be transformed during dag combiner except when the condition
6479 // is set by a arithmetics with overflow node.
6480 X86::CondCode CCode =
6481 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6482 CCode = X86::GetOppositeBranchCondition(CCode);
6483 CC = DAG.getConstant(CCode, MVT::i8);
6484 Cond = Cond.getOperand(0).getOperand(1);
6490 // Look pass the truncate.
6491 if (Cond.getOpcode() == ISD::TRUNCATE)
6492 Cond = Cond.getOperand(0);
6494 // We know the result of AND is compared against zero. Try to match
6496 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6497 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6498 if (NewSetCC.getNode()) {
6499 CC = NewSetCC.getOperand(0);
6500 Cond = NewSetCC.getOperand(1);
6507 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6508 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6510 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6511 Chain, Dest, CC, Cond);
6515 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6516 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6517 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6518 // that the guard pages used by the OS virtual memory manager are allocated in
6519 // correct sequence.
6521 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6522 SelectionDAG &DAG) {
6523 assert(Subtarget->isTargetCygMing() &&
6524 "This should be used only on Cygwin/Mingw targets");
6525 DebugLoc dl = Op.getDebugLoc();
6528 SDValue Chain = Op.getOperand(0);
6529 SDValue Size = Op.getOperand(1);
6530 // FIXME: Ensure alignment here
6534 EVT IntPtr = getPointerTy();
6535 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6537 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6538 Flag = Chain.getValue(1);
6540 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6542 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6543 Flag = Chain.getValue(1);
6545 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6547 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6548 return DAG.getMergeValues(Ops1, 2, dl);
6552 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6554 SDValue Dst, SDValue Src,
6555 SDValue Size, unsigned Align,
6557 uint64_t DstSVOff) {
6558 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6560 // If not DWORD aligned or size is more than the threshold, call the library.
6561 // The libc version is likely to be faster for these cases. It can use the
6562 // address value and run time information about the CPU.
6563 if ((Align & 3) != 0 ||
6565 ConstantSize->getZExtValue() >
6566 getSubtarget()->getMaxInlineSizeThreshold()) {
6567 SDValue InFlag(0, 0);
6569 // Check to see if there is a specialized entry-point for memory zeroing.
6570 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6572 if (const char *bzeroEntry = V &&
6573 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6574 EVT IntPtr = getPointerTy();
6575 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6576 TargetLowering::ArgListTy Args;
6577 TargetLowering::ArgListEntry Entry;
6579 Entry.Ty = IntPtrTy;
6580 Args.push_back(Entry);
6582 Args.push_back(Entry);
6583 std::pair<SDValue,SDValue> CallResult =
6584 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6585 false, false, false, false,
6586 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6587 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6588 return CallResult.second;
6591 // Otherwise have the target-independent code call memset.
6595 uint64_t SizeVal = ConstantSize->getZExtValue();
6596 SDValue InFlag(0, 0);
6599 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6600 unsigned BytesLeft = 0;
6601 bool TwoRepStos = false;
6604 uint64_t Val = ValC->getZExtValue() & 255;
6606 // If the value is a constant, then we can potentially use larger sets.
6607 switch (Align & 3) {
6608 case 2: // WORD aligned
6611 Val = (Val << 8) | Val;
6613 case 0: // DWORD aligned
6616 Val = (Val << 8) | Val;
6617 Val = (Val << 16) | Val;
6618 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6621 Val = (Val << 32) | Val;
6624 default: // Byte aligned
6627 Count = DAG.getIntPtrConstant(SizeVal);
6631 if (AVT.bitsGT(MVT::i8)) {
6632 unsigned UBytes = AVT.getSizeInBits() / 8;
6633 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6634 BytesLeft = SizeVal % UBytes;
6637 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6639 InFlag = Chain.getValue(1);
6642 Count = DAG.getIntPtrConstant(SizeVal);
6643 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6644 InFlag = Chain.getValue(1);
6647 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6650 InFlag = Chain.getValue(1);
6651 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6654 InFlag = Chain.getValue(1);
6656 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6657 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6658 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6661 InFlag = Chain.getValue(1);
6663 EVT CVT = Count.getValueType();
6664 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6665 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6666 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6669 InFlag = Chain.getValue(1);
6670 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6671 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6672 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6673 } else if (BytesLeft) {
6674 // Handle the last 1 - 7 bytes.
6675 unsigned Offset = SizeVal - BytesLeft;
6676 EVT AddrVT = Dst.getValueType();
6677 EVT SizeVT = Size.getValueType();
6679 Chain = DAG.getMemset(Chain, dl,
6680 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6681 DAG.getConstant(Offset, AddrVT)),
6683 DAG.getConstant(BytesLeft, SizeVT),
6684 Align, DstSV, DstSVOff + Offset);
6687 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6692 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6693 SDValue Chain, SDValue Dst, SDValue Src,
6694 SDValue Size, unsigned Align,
6696 const Value *DstSV, uint64_t DstSVOff,
6697 const Value *SrcSV, uint64_t SrcSVOff) {
6698 // This requires the copy size to be a constant, preferrably
6699 // within a subtarget-specific limit.
6700 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6703 uint64_t SizeVal = ConstantSize->getZExtValue();
6704 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6707 /// If not DWORD aligned, call the library.
6708 if ((Align & 3) != 0)
6713 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6716 unsigned UBytes = AVT.getSizeInBits() / 8;
6717 unsigned CountVal = SizeVal / UBytes;
6718 SDValue Count = DAG.getIntPtrConstant(CountVal);
6719 unsigned BytesLeft = SizeVal % UBytes;
6721 SDValue InFlag(0, 0);
6722 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6725 InFlag = Chain.getValue(1);
6726 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6729 InFlag = Chain.getValue(1);
6730 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6733 InFlag = Chain.getValue(1);
6735 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6736 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6737 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6738 array_lengthof(Ops));
6740 SmallVector<SDValue, 4> Results;
6741 Results.push_back(RepMovs);
6743 // Handle the last 1 - 7 bytes.
6744 unsigned Offset = SizeVal - BytesLeft;
6745 EVT DstVT = Dst.getValueType();
6746 EVT SrcVT = Src.getValueType();
6747 EVT SizeVT = Size.getValueType();
6748 Results.push_back(DAG.getMemcpy(Chain, dl,
6749 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6750 DAG.getConstant(Offset, DstVT)),
6751 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6752 DAG.getConstant(Offset, SrcVT)),
6753 DAG.getConstant(BytesLeft, SizeVT),
6754 Align, AlwaysInline,
6755 DstSV, DstSVOff + Offset,
6756 SrcSV, SrcSVOff + Offset));
6759 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6760 &Results[0], Results.size());
6763 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6764 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6765 DebugLoc dl = Op.getDebugLoc();
6767 if (!Subtarget->is64Bit()) {
6768 // vastart just stores the address of the VarArgsFrameIndex slot into the
6769 // memory location argument.
6770 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6771 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6776 // gp_offset (0 - 6 * 8)
6777 // fp_offset (48 - 48 + 8 * 16)
6778 // overflow_arg_area (point to parameters coming in memory).
6780 SmallVector<SDValue, 8> MemOps;
6781 SDValue FIN = Op.getOperand(1);
6783 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6784 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6785 FIN, SV, 0, false, false, 0);
6786 MemOps.push_back(Store);
6789 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6790 FIN, DAG.getIntPtrConstant(4));
6791 Store = DAG.getStore(Op.getOperand(0), dl,
6792 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6793 FIN, SV, 0, false, false, 0);
6794 MemOps.push_back(Store);
6796 // Store ptr to overflow_arg_area
6797 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6798 FIN, DAG.getIntPtrConstant(4));
6799 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6800 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6802 MemOps.push_back(Store);
6804 // Store ptr to reg_save_area.
6805 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6806 FIN, DAG.getIntPtrConstant(8));
6807 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6808 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6810 MemOps.push_back(Store);
6811 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6812 &MemOps[0], MemOps.size());
6815 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6816 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6817 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6818 SDValue Chain = Op.getOperand(0);
6819 SDValue SrcPtr = Op.getOperand(1);
6820 SDValue SrcSV = Op.getOperand(2);
6822 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6826 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6827 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6828 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6829 SDValue Chain = Op.getOperand(0);
6830 SDValue DstPtr = Op.getOperand(1);
6831 SDValue SrcPtr = Op.getOperand(2);
6832 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6833 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6834 DebugLoc dl = Op.getDebugLoc();
6836 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6837 DAG.getIntPtrConstant(24), 8, false,
6838 DstSV, 0, SrcSV, 0);
6842 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6843 DebugLoc dl = Op.getDebugLoc();
6844 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6846 default: return SDValue(); // Don't custom lower most intrinsics.
6847 // Comparison intrinsics.
6848 case Intrinsic::x86_sse_comieq_ss:
6849 case Intrinsic::x86_sse_comilt_ss:
6850 case Intrinsic::x86_sse_comile_ss:
6851 case Intrinsic::x86_sse_comigt_ss:
6852 case Intrinsic::x86_sse_comige_ss:
6853 case Intrinsic::x86_sse_comineq_ss:
6854 case Intrinsic::x86_sse_ucomieq_ss:
6855 case Intrinsic::x86_sse_ucomilt_ss:
6856 case Intrinsic::x86_sse_ucomile_ss:
6857 case Intrinsic::x86_sse_ucomigt_ss:
6858 case Intrinsic::x86_sse_ucomige_ss:
6859 case Intrinsic::x86_sse_ucomineq_ss:
6860 case Intrinsic::x86_sse2_comieq_sd:
6861 case Intrinsic::x86_sse2_comilt_sd:
6862 case Intrinsic::x86_sse2_comile_sd:
6863 case Intrinsic::x86_sse2_comigt_sd:
6864 case Intrinsic::x86_sse2_comige_sd:
6865 case Intrinsic::x86_sse2_comineq_sd:
6866 case Intrinsic::x86_sse2_ucomieq_sd:
6867 case Intrinsic::x86_sse2_ucomilt_sd:
6868 case Intrinsic::x86_sse2_ucomile_sd:
6869 case Intrinsic::x86_sse2_ucomigt_sd:
6870 case Intrinsic::x86_sse2_ucomige_sd:
6871 case Intrinsic::x86_sse2_ucomineq_sd: {
6873 ISD::CondCode CC = ISD::SETCC_INVALID;
6876 case Intrinsic::x86_sse_comieq_ss:
6877 case Intrinsic::x86_sse2_comieq_sd:
6881 case Intrinsic::x86_sse_comilt_ss:
6882 case Intrinsic::x86_sse2_comilt_sd:
6886 case Intrinsic::x86_sse_comile_ss:
6887 case Intrinsic::x86_sse2_comile_sd:
6891 case Intrinsic::x86_sse_comigt_ss:
6892 case Intrinsic::x86_sse2_comigt_sd:
6896 case Intrinsic::x86_sse_comige_ss:
6897 case Intrinsic::x86_sse2_comige_sd:
6901 case Intrinsic::x86_sse_comineq_ss:
6902 case Intrinsic::x86_sse2_comineq_sd:
6906 case Intrinsic::x86_sse_ucomieq_ss:
6907 case Intrinsic::x86_sse2_ucomieq_sd:
6908 Opc = X86ISD::UCOMI;
6911 case Intrinsic::x86_sse_ucomilt_ss:
6912 case Intrinsic::x86_sse2_ucomilt_sd:
6913 Opc = X86ISD::UCOMI;
6916 case Intrinsic::x86_sse_ucomile_ss:
6917 case Intrinsic::x86_sse2_ucomile_sd:
6918 Opc = X86ISD::UCOMI;
6921 case Intrinsic::x86_sse_ucomigt_ss:
6922 case Intrinsic::x86_sse2_ucomigt_sd:
6923 Opc = X86ISD::UCOMI;
6926 case Intrinsic::x86_sse_ucomige_ss:
6927 case Intrinsic::x86_sse2_ucomige_sd:
6928 Opc = X86ISD::UCOMI;
6931 case Intrinsic::x86_sse_ucomineq_ss:
6932 case Intrinsic::x86_sse2_ucomineq_sd:
6933 Opc = X86ISD::UCOMI;
6938 SDValue LHS = Op.getOperand(1);
6939 SDValue RHS = Op.getOperand(2);
6940 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6941 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6942 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6943 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6944 DAG.getConstant(X86CC, MVT::i8), Cond);
6945 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6947 // ptest intrinsics. The intrinsic these come from are designed to return
6948 // an integer value, not just an instruction so lower it to the ptest
6949 // pattern and a setcc for the result.
6950 case Intrinsic::x86_sse41_ptestz:
6951 case Intrinsic::x86_sse41_ptestc:
6952 case Intrinsic::x86_sse41_ptestnzc:{
6955 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6956 case Intrinsic::x86_sse41_ptestz:
6958 X86CC = X86::COND_E;
6960 case Intrinsic::x86_sse41_ptestc:
6962 X86CC = X86::COND_B;
6964 case Intrinsic::x86_sse41_ptestnzc:
6966 X86CC = X86::COND_A;
6970 SDValue LHS = Op.getOperand(1);
6971 SDValue RHS = Op.getOperand(2);
6972 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6973 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6974 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6975 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6978 // Fix vector shift instructions where the last operand is a non-immediate
6980 case Intrinsic::x86_sse2_pslli_w:
6981 case Intrinsic::x86_sse2_pslli_d:
6982 case Intrinsic::x86_sse2_pslli_q:
6983 case Intrinsic::x86_sse2_psrli_w:
6984 case Intrinsic::x86_sse2_psrli_d:
6985 case Intrinsic::x86_sse2_psrli_q:
6986 case Intrinsic::x86_sse2_psrai_w:
6987 case Intrinsic::x86_sse2_psrai_d:
6988 case Intrinsic::x86_mmx_pslli_w:
6989 case Intrinsic::x86_mmx_pslli_d:
6990 case Intrinsic::x86_mmx_pslli_q:
6991 case Intrinsic::x86_mmx_psrli_w:
6992 case Intrinsic::x86_mmx_psrli_d:
6993 case Intrinsic::x86_mmx_psrli_q:
6994 case Intrinsic::x86_mmx_psrai_w:
6995 case Intrinsic::x86_mmx_psrai_d: {
6996 SDValue ShAmt = Op.getOperand(2);
6997 if (isa<ConstantSDNode>(ShAmt))
7000 unsigned NewIntNo = 0;
7001 EVT ShAmtVT = MVT::v4i32;
7003 case Intrinsic::x86_sse2_pslli_w:
7004 NewIntNo = Intrinsic::x86_sse2_psll_w;
7006 case Intrinsic::x86_sse2_pslli_d:
7007 NewIntNo = Intrinsic::x86_sse2_psll_d;
7009 case Intrinsic::x86_sse2_pslli_q:
7010 NewIntNo = Intrinsic::x86_sse2_psll_q;
7012 case Intrinsic::x86_sse2_psrli_w:
7013 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7015 case Intrinsic::x86_sse2_psrli_d:
7016 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7018 case Intrinsic::x86_sse2_psrli_q:
7019 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7021 case Intrinsic::x86_sse2_psrai_w:
7022 NewIntNo = Intrinsic::x86_sse2_psra_w;
7024 case Intrinsic::x86_sse2_psrai_d:
7025 NewIntNo = Intrinsic::x86_sse2_psra_d;
7028 ShAmtVT = MVT::v2i32;
7030 case Intrinsic::x86_mmx_pslli_w:
7031 NewIntNo = Intrinsic::x86_mmx_psll_w;
7033 case Intrinsic::x86_mmx_pslli_d:
7034 NewIntNo = Intrinsic::x86_mmx_psll_d;
7036 case Intrinsic::x86_mmx_pslli_q:
7037 NewIntNo = Intrinsic::x86_mmx_psll_q;
7039 case Intrinsic::x86_mmx_psrli_w:
7040 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7042 case Intrinsic::x86_mmx_psrli_d:
7043 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7045 case Intrinsic::x86_mmx_psrli_q:
7046 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7048 case Intrinsic::x86_mmx_psrai_w:
7049 NewIntNo = Intrinsic::x86_mmx_psra_w;
7051 case Intrinsic::x86_mmx_psrai_d:
7052 NewIntNo = Intrinsic::x86_mmx_psra_d;
7054 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7060 // The vector shift intrinsics with scalars uses 32b shift amounts but
7061 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7065 ShOps[1] = DAG.getConstant(0, MVT::i32);
7066 if (ShAmtVT == MVT::v4i32) {
7067 ShOps[2] = DAG.getUNDEF(MVT::i32);
7068 ShOps[3] = DAG.getUNDEF(MVT::i32);
7069 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7071 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7074 EVT VT = Op.getValueType();
7075 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7076 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7077 DAG.getConstant(NewIntNo, MVT::i32),
7078 Op.getOperand(1), ShAmt);
7083 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
7084 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7085 DebugLoc dl = Op.getDebugLoc();
7088 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7090 DAG.getConstant(TD->getPointerSize(),
7091 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7092 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7093 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7095 NULL, 0, false, false, 0);
7098 // Just load the return address.
7099 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7100 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7101 RetAddrFI, NULL, 0, false, false, 0);
7104 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
7105 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7106 MFI->setFrameAddressIsTaken(true);
7107 EVT VT = Op.getValueType();
7108 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7109 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7110 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7111 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7113 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7118 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7119 SelectionDAG &DAG) {
7120 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7123 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7125 MachineFunction &MF = DAG.getMachineFunction();
7126 SDValue Chain = Op.getOperand(0);
7127 SDValue Offset = Op.getOperand(1);
7128 SDValue Handler = Op.getOperand(2);
7129 DebugLoc dl = Op.getDebugLoc();
7131 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7133 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7135 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7136 DAG.getIntPtrConstant(-TD->getPointerSize()));
7137 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7138 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7139 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7140 MF.getRegInfo().addLiveOut(StoreAddrReg);
7142 return DAG.getNode(X86ISD::EH_RETURN, dl,
7144 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7147 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7148 SelectionDAG &DAG) {
7149 SDValue Root = Op.getOperand(0);
7150 SDValue Trmp = Op.getOperand(1); // trampoline
7151 SDValue FPtr = Op.getOperand(2); // nested function
7152 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7153 DebugLoc dl = Op.getDebugLoc();
7155 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7157 if (Subtarget->is64Bit()) {
7158 SDValue OutChains[6];
7160 // Large code-model.
7161 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7162 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7164 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7165 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7167 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7169 // Load the pointer to the nested function into R11.
7170 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7171 SDValue Addr = Trmp;
7172 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7173 Addr, TrmpAddr, 0, false, false, 0);
7175 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7176 DAG.getConstant(2, MVT::i64));
7177 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7180 // Load the 'nest' parameter value into R10.
7181 // R10 is specified in X86CallingConv.td
7182 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7183 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7184 DAG.getConstant(10, MVT::i64));
7185 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7186 Addr, TrmpAddr, 10, false, false, 0);
7188 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7189 DAG.getConstant(12, MVT::i64));
7190 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7193 // Jump to the nested function.
7194 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7195 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7196 DAG.getConstant(20, MVT::i64));
7197 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7198 Addr, TrmpAddr, 20, false, false, 0);
7200 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7201 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7202 DAG.getConstant(22, MVT::i64));
7203 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7204 TrmpAddr, 22, false, false, 0);
7207 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7208 return DAG.getMergeValues(Ops, 2, dl);
7210 const Function *Func =
7211 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7212 CallingConv::ID CC = Func->getCallingConv();
7217 llvm_unreachable("Unsupported calling convention");
7218 case CallingConv::C:
7219 case CallingConv::X86_StdCall: {
7220 // Pass 'nest' parameter in ECX.
7221 // Must be kept in sync with X86CallingConv.td
7224 // Check that ECX wasn't needed by an 'inreg' parameter.
7225 const FunctionType *FTy = Func->getFunctionType();
7226 const AttrListPtr &Attrs = Func->getAttributes();
7228 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7229 unsigned InRegCount = 0;
7232 for (FunctionType::param_iterator I = FTy->param_begin(),
7233 E = FTy->param_end(); I != E; ++I, ++Idx)
7234 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7235 // FIXME: should only count parameters that are lowered to integers.
7236 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7238 if (InRegCount > 2) {
7239 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7244 case CallingConv::X86_FastCall:
7245 case CallingConv::Fast:
7246 // Pass 'nest' parameter in EAX.
7247 // Must be kept in sync with X86CallingConv.td
7252 SDValue OutChains[4];
7255 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7256 DAG.getConstant(10, MVT::i32));
7257 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7259 // This is storing the opcode for MOV32ri.
7260 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7261 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7262 OutChains[0] = DAG.getStore(Root, dl,
7263 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7264 Trmp, TrmpAddr, 0, false, false, 0);
7266 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7267 DAG.getConstant(1, MVT::i32));
7268 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7271 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7272 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7273 DAG.getConstant(5, MVT::i32));
7274 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7275 TrmpAddr, 5, false, false, 1);
7277 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7278 DAG.getConstant(6, MVT::i32));
7279 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7283 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7284 return DAG.getMergeValues(Ops, 2, dl);
7288 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7290 The rounding mode is in bits 11:10 of FPSR, and has the following
7297 FLT_ROUNDS, on the other hand, expects the following:
7304 To perform the conversion, we do:
7305 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7308 MachineFunction &MF = DAG.getMachineFunction();
7309 const TargetMachine &TM = MF.getTarget();
7310 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7311 unsigned StackAlignment = TFI.getStackAlignment();
7312 EVT VT = Op.getValueType();
7313 DebugLoc dl = Op.getDebugLoc();
7315 // Save FP Control Word to stack slot
7316 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7317 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7319 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7320 DAG.getEntryNode(), StackSlot);
7322 // Load FP Control Word from stack slot
7323 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7326 // Transform as necessary
7328 DAG.getNode(ISD::SRL, dl, MVT::i16,
7329 DAG.getNode(ISD::AND, dl, MVT::i16,
7330 CWD, DAG.getConstant(0x800, MVT::i16)),
7331 DAG.getConstant(11, MVT::i8));
7333 DAG.getNode(ISD::SRL, dl, MVT::i16,
7334 DAG.getNode(ISD::AND, dl, MVT::i16,
7335 CWD, DAG.getConstant(0x400, MVT::i16)),
7336 DAG.getConstant(9, MVT::i8));
7339 DAG.getNode(ISD::AND, dl, MVT::i16,
7340 DAG.getNode(ISD::ADD, dl, MVT::i16,
7341 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7342 DAG.getConstant(1, MVT::i16)),
7343 DAG.getConstant(3, MVT::i16));
7346 return DAG.getNode((VT.getSizeInBits() < 16 ?
7347 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7350 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7351 EVT VT = Op.getValueType();
7353 unsigned NumBits = VT.getSizeInBits();
7354 DebugLoc dl = Op.getDebugLoc();
7356 Op = Op.getOperand(0);
7357 if (VT == MVT::i8) {
7358 // Zero extend to i32 since there is not an i8 bsr.
7360 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7363 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7364 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7365 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7367 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7370 DAG.getConstant(NumBits+NumBits-1, OpVT),
7371 DAG.getConstant(X86::COND_E, MVT::i8),
7374 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7376 // Finally xor with NumBits-1.
7377 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7380 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7384 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7385 EVT VT = Op.getValueType();
7387 unsigned NumBits = VT.getSizeInBits();
7388 DebugLoc dl = Op.getDebugLoc();
7390 Op = Op.getOperand(0);
7391 if (VT == MVT::i8) {
7393 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7396 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7397 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7398 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7400 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7403 DAG.getConstant(NumBits, OpVT),
7404 DAG.getConstant(X86::COND_E, MVT::i8),
7407 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7410 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7414 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7415 EVT VT = Op.getValueType();
7416 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7417 DebugLoc dl = Op.getDebugLoc();
7419 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7420 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7421 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7422 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7423 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7425 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7426 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7427 // return AloBlo + AloBhi + AhiBlo;
7429 SDValue A = Op.getOperand(0);
7430 SDValue B = Op.getOperand(1);
7432 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7433 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7434 A, DAG.getConstant(32, MVT::i32));
7435 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7436 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7437 B, DAG.getConstant(32, MVT::i32));
7438 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7439 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7441 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7442 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7444 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7445 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7447 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7448 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7449 AloBhi, DAG.getConstant(32, MVT::i32));
7450 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7451 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7452 AhiBlo, DAG.getConstant(32, MVT::i32));
7453 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7454 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7459 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7460 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7461 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7462 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7463 // has only one use.
7464 SDNode *N = Op.getNode();
7465 SDValue LHS = N->getOperand(0);
7466 SDValue RHS = N->getOperand(1);
7467 unsigned BaseOp = 0;
7469 DebugLoc dl = Op.getDebugLoc();
7471 switch (Op.getOpcode()) {
7472 default: llvm_unreachable("Unknown ovf instruction!");
7474 // A subtract of one will be selected as a INC. Note that INC doesn't
7475 // set CF, so we can't do this for UADDO.
7476 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7477 if (C->getAPIntValue() == 1) {
7478 BaseOp = X86ISD::INC;
7482 BaseOp = X86ISD::ADD;
7486 BaseOp = X86ISD::ADD;
7490 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7491 // set CF, so we can't do this for USUBO.
7492 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7493 if (C->getAPIntValue() == 1) {
7494 BaseOp = X86ISD::DEC;
7498 BaseOp = X86ISD::SUB;
7502 BaseOp = X86ISD::SUB;
7506 BaseOp = X86ISD::SMUL;
7510 BaseOp = X86ISD::UMUL;
7515 // Also sets EFLAGS.
7516 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7517 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7520 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7521 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7523 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7527 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7528 EVT T = Op.getValueType();
7529 DebugLoc dl = Op.getDebugLoc();
7532 switch(T.getSimpleVT().SimpleTy) {
7534 assert(false && "Invalid value type!");
7535 case MVT::i8: Reg = X86::AL; size = 1; break;
7536 case MVT::i16: Reg = X86::AX; size = 2; break;
7537 case MVT::i32: Reg = X86::EAX; size = 4; break;
7539 assert(Subtarget->is64Bit() && "Node not type legal!");
7540 Reg = X86::RAX; size = 8;
7543 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7544 Op.getOperand(2), SDValue());
7545 SDValue Ops[] = { cpIn.getValue(0),
7548 DAG.getTargetConstant(size, MVT::i8),
7550 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7551 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7553 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7557 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7558 SelectionDAG &DAG) {
7559 assert(Subtarget->is64Bit() && "Result not type legalized?");
7560 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7561 SDValue TheChain = Op.getOperand(0);
7562 DebugLoc dl = Op.getDebugLoc();
7563 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7564 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7565 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7567 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7568 DAG.getConstant(32, MVT::i8));
7570 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7573 return DAG.getMergeValues(Ops, 2, dl);
7576 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7577 SDNode *Node = Op.getNode();
7578 DebugLoc dl = Node->getDebugLoc();
7579 EVT T = Node->getValueType(0);
7580 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7581 DAG.getConstant(0, T), Node->getOperand(2));
7582 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7583 cast<AtomicSDNode>(Node)->getMemoryVT(),
7584 Node->getOperand(0),
7585 Node->getOperand(1), negOp,
7586 cast<AtomicSDNode>(Node)->getSrcValue(),
7587 cast<AtomicSDNode>(Node)->getAlignment());
7590 /// LowerOperation - Provide custom lowering hooks for some operations.
7592 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7593 switch (Op.getOpcode()) {
7594 default: llvm_unreachable("Should not custom lower this!");
7595 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7596 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7597 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7598 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7599 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7600 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7601 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7602 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7603 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7604 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7605 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7606 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7607 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7608 case ISD::SHL_PARTS:
7609 case ISD::SRA_PARTS:
7610 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7611 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7612 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7613 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7614 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7615 case ISD::FABS: return LowerFABS(Op, DAG);
7616 case ISD::FNEG: return LowerFNEG(Op, DAG);
7617 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7618 case ISD::SETCC: return LowerSETCC(Op, DAG);
7619 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7620 case ISD::SELECT: return LowerSELECT(Op, DAG);
7621 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7622 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7623 case ISD::VASTART: return LowerVASTART(Op, DAG);
7624 case ISD::VAARG: return LowerVAARG(Op, DAG);
7625 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7626 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7627 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7628 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7629 case ISD::FRAME_TO_ARGS_OFFSET:
7630 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7631 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7632 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7633 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7634 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7635 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7636 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7637 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7643 case ISD::UMULO: return LowerXALUO(Op, DAG);
7644 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7648 void X86TargetLowering::
7649 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7650 SelectionDAG &DAG, unsigned NewOp) {
7651 EVT T = Node->getValueType(0);
7652 DebugLoc dl = Node->getDebugLoc();
7653 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7655 SDValue Chain = Node->getOperand(0);
7656 SDValue In1 = Node->getOperand(1);
7657 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7658 Node->getOperand(2), DAG.getIntPtrConstant(0));
7659 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7660 Node->getOperand(2), DAG.getIntPtrConstant(1));
7661 SDValue Ops[] = { Chain, In1, In2L, In2H };
7662 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7664 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7665 cast<MemSDNode>(Node)->getMemOperand());
7666 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7667 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7668 Results.push_back(Result.getValue(2));
7671 /// ReplaceNodeResults - Replace a node with an illegal result type
7672 /// with a new node built out of custom code.
7673 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7674 SmallVectorImpl<SDValue>&Results,
7675 SelectionDAG &DAG) {
7676 DebugLoc dl = N->getDebugLoc();
7677 switch (N->getOpcode()) {
7679 assert(false && "Do not know how to custom type legalize this operation!");
7681 case ISD::FP_TO_SINT: {
7682 std::pair<SDValue,SDValue> Vals =
7683 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7684 SDValue FIST = Vals.first, StackSlot = Vals.second;
7685 if (FIST.getNode() != 0) {
7686 EVT VT = N->getValueType(0);
7687 // Return a load from the stack slot.
7688 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7693 case ISD::READCYCLECOUNTER: {
7694 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7695 SDValue TheChain = N->getOperand(0);
7696 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7697 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7699 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7701 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7702 SDValue Ops[] = { eax, edx };
7703 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7704 Results.push_back(edx.getValue(1));
7707 case ISD::ATOMIC_CMP_SWAP: {
7708 EVT T = N->getValueType(0);
7709 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7710 SDValue cpInL, cpInH;
7711 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7712 DAG.getConstant(0, MVT::i32));
7713 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7714 DAG.getConstant(1, MVT::i32));
7715 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7716 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7718 SDValue swapInL, swapInH;
7719 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7720 DAG.getConstant(0, MVT::i32));
7721 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7722 DAG.getConstant(1, MVT::i32));
7723 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7725 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7726 swapInL.getValue(1));
7727 SDValue Ops[] = { swapInH.getValue(0),
7729 swapInH.getValue(1) };
7730 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7731 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7732 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7733 MVT::i32, Result.getValue(1));
7734 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7735 MVT::i32, cpOutL.getValue(2));
7736 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7737 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7738 Results.push_back(cpOutH.getValue(1));
7741 case ISD::ATOMIC_LOAD_ADD:
7742 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7744 case ISD::ATOMIC_LOAD_AND:
7745 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7747 case ISD::ATOMIC_LOAD_NAND:
7748 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7750 case ISD::ATOMIC_LOAD_OR:
7751 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7753 case ISD::ATOMIC_LOAD_SUB:
7754 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7756 case ISD::ATOMIC_LOAD_XOR:
7757 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7759 case ISD::ATOMIC_SWAP:
7760 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7765 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7767 default: return NULL;
7768 case X86ISD::BSF: return "X86ISD::BSF";
7769 case X86ISD::BSR: return "X86ISD::BSR";
7770 case X86ISD::SHLD: return "X86ISD::SHLD";
7771 case X86ISD::SHRD: return "X86ISD::SHRD";
7772 case X86ISD::FAND: return "X86ISD::FAND";
7773 case X86ISD::FOR: return "X86ISD::FOR";
7774 case X86ISD::FXOR: return "X86ISD::FXOR";
7775 case X86ISD::FSRL: return "X86ISD::FSRL";
7776 case X86ISD::FILD: return "X86ISD::FILD";
7777 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7778 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7779 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7780 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7781 case X86ISD::FLD: return "X86ISD::FLD";
7782 case X86ISD::FST: return "X86ISD::FST";
7783 case X86ISD::CALL: return "X86ISD::CALL";
7784 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7785 case X86ISD::BT: return "X86ISD::BT";
7786 case X86ISD::CMP: return "X86ISD::CMP";
7787 case X86ISD::COMI: return "X86ISD::COMI";
7788 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7789 case X86ISD::SETCC: return "X86ISD::SETCC";
7790 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7791 case X86ISD::CMOV: return "X86ISD::CMOV";
7792 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7793 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7794 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7795 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7796 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7797 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7798 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7799 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7800 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7801 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7802 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7803 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7804 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7805 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7806 case X86ISD::FMAX: return "X86ISD::FMAX";
7807 case X86ISD::FMIN: return "X86ISD::FMIN";
7808 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7809 case X86ISD::FRCP: return "X86ISD::FRCP";
7810 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7811 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7812 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7813 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7814 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7815 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7816 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7817 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7818 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7819 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7820 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7821 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7822 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7823 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7824 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7825 case X86ISD::VSHL: return "X86ISD::VSHL";
7826 case X86ISD::VSRL: return "X86ISD::VSRL";
7827 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7828 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7829 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7830 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7831 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7832 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7833 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7834 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7835 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7836 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7837 case X86ISD::ADD: return "X86ISD::ADD";
7838 case X86ISD::SUB: return "X86ISD::SUB";
7839 case X86ISD::SMUL: return "X86ISD::SMUL";
7840 case X86ISD::UMUL: return "X86ISD::UMUL";
7841 case X86ISD::INC: return "X86ISD::INC";
7842 case X86ISD::DEC: return "X86ISD::DEC";
7843 case X86ISD::OR: return "X86ISD::OR";
7844 case X86ISD::XOR: return "X86ISD::XOR";
7845 case X86ISD::AND: return "X86ISD::AND";
7846 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7847 case X86ISD::PTEST: return "X86ISD::PTEST";
7848 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7849 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7853 // isLegalAddressingMode - Return true if the addressing mode represented
7854 // by AM is legal for this target, for a load/store of the specified type.
7855 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7856 const Type *Ty) const {
7857 // X86 supports extremely general addressing modes.
7858 CodeModel::Model M = getTargetMachine().getCodeModel();
7860 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7861 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7866 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7868 // If a reference to this global requires an extra load, we can't fold it.
7869 if (isGlobalStubReference(GVFlags))
7872 // If BaseGV requires a register for the PIC base, we cannot also have a
7873 // BaseReg specified.
7874 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7877 // If lower 4G is not available, then we must use rip-relative addressing.
7878 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7888 // These scales always work.
7893 // These scales are formed with basereg+scalereg. Only accept if there is
7898 default: // Other stuff never works.
7906 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7907 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7909 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7910 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7911 if (NumBits1 <= NumBits2)
7916 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7917 if (!VT1.isInteger() || !VT2.isInteger())
7919 unsigned NumBits1 = VT1.getSizeInBits();
7920 unsigned NumBits2 = VT2.getSizeInBits();
7921 if (NumBits1 <= NumBits2)
7926 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7927 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7928 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7931 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7932 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7933 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7936 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7937 // i16 instructions are longer (0x66 prefix) and potentially slower.
7938 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7941 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7942 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7943 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7944 /// are assumed to be legal.
7946 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7948 // Only do shuffles on 128-bit vector types for now.
7949 if (VT.getSizeInBits() == 64)
7952 // FIXME: pshufb, blends, shifts.
7953 return (VT.getVectorNumElements() == 2 ||
7954 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7955 isMOVLMask(M, VT) ||
7956 isSHUFPMask(M, VT) ||
7957 isPSHUFDMask(M, VT) ||
7958 isPSHUFHWMask(M, VT) ||
7959 isPSHUFLWMask(M, VT) ||
7960 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7961 isUNPCKLMask(M, VT) ||
7962 isUNPCKHMask(M, VT) ||
7963 isUNPCKL_v_undef_Mask(M, VT) ||
7964 isUNPCKH_v_undef_Mask(M, VT));
7968 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7970 unsigned NumElts = VT.getVectorNumElements();
7971 // FIXME: This collection of masks seems suspect.
7974 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7975 return (isMOVLMask(Mask, VT) ||
7976 isCommutedMOVLMask(Mask, VT, true) ||
7977 isSHUFPMask(Mask, VT) ||
7978 isCommutedSHUFPMask(Mask, VT));
7983 //===----------------------------------------------------------------------===//
7984 // X86 Scheduler Hooks
7985 //===----------------------------------------------------------------------===//
7987 // private utility function
7989 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7990 MachineBasicBlock *MBB,
7998 TargetRegisterClass *RC,
7999 bool invSrc) const {
8000 // For the atomic bitwise operator, we generate
8003 // ld t1 = [bitinstr.addr]
8004 // op t2 = t1, [bitinstr.val]
8006 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8008 // fallthrough -->nextMBB
8009 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8010 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8011 MachineFunction::iterator MBBIter = MBB;
8014 /// First build the CFG
8015 MachineFunction *F = MBB->getParent();
8016 MachineBasicBlock *thisMBB = MBB;
8017 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8018 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8019 F->insert(MBBIter, newMBB);
8020 F->insert(MBBIter, nextMBB);
8022 // Move all successors to thisMBB to nextMBB
8023 nextMBB->transferSuccessors(thisMBB);
8025 // Update thisMBB to fall through to newMBB
8026 thisMBB->addSuccessor(newMBB);
8028 // newMBB jumps to itself and fall through to nextMBB
8029 newMBB->addSuccessor(nextMBB);
8030 newMBB->addSuccessor(newMBB);
8032 // Insert instructions into newMBB based on incoming instruction
8033 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8034 "unexpected number of operands");
8035 DebugLoc dl = bInstr->getDebugLoc();
8036 MachineOperand& destOper = bInstr->getOperand(0);
8037 MachineOperand* argOpers[2 + X86AddrNumOperands];
8038 int numArgs = bInstr->getNumOperands() - 1;
8039 for (int i=0; i < numArgs; ++i)
8040 argOpers[i] = &bInstr->getOperand(i+1);
8042 // x86 address has 4 operands: base, index, scale, and displacement
8043 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8044 int valArgIndx = lastAddrIndx + 1;
8046 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8047 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8048 for (int i=0; i <= lastAddrIndx; ++i)
8049 (*MIB).addOperand(*argOpers[i]);
8051 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8053 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8058 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8059 assert((argOpers[valArgIndx]->isReg() ||
8060 argOpers[valArgIndx]->isImm()) &&
8062 if (argOpers[valArgIndx]->isReg())
8063 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8065 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8067 (*MIB).addOperand(*argOpers[valArgIndx]);
8069 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8072 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8073 for (int i=0; i <= lastAddrIndx; ++i)
8074 (*MIB).addOperand(*argOpers[i]);
8076 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8077 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8078 bInstr->memoperands_end());
8080 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8084 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8086 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8090 // private utility function: 64 bit atomics on 32 bit host.
8092 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8093 MachineBasicBlock *MBB,
8098 bool invSrc) const {
8099 // For the atomic bitwise operator, we generate
8100 // thisMBB (instructions are in pairs, except cmpxchg8b)
8101 // ld t1,t2 = [bitinstr.addr]
8103 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8104 // op t5, t6 <- out1, out2, [bitinstr.val]
8105 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8106 // mov ECX, EBX <- t5, t6
8107 // mov EAX, EDX <- t1, t2
8108 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8109 // mov t3, t4 <- EAX, EDX
8111 // result in out1, out2
8112 // fallthrough -->nextMBB
8114 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8115 const unsigned LoadOpc = X86::MOV32rm;
8116 const unsigned copyOpc = X86::MOV32rr;
8117 const unsigned NotOpc = X86::NOT32r;
8118 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8119 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8120 MachineFunction::iterator MBBIter = MBB;
8123 /// First build the CFG
8124 MachineFunction *F = MBB->getParent();
8125 MachineBasicBlock *thisMBB = MBB;
8126 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8127 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8128 F->insert(MBBIter, newMBB);
8129 F->insert(MBBIter, nextMBB);
8131 // Move all successors to thisMBB to nextMBB
8132 nextMBB->transferSuccessors(thisMBB);
8134 // Update thisMBB to fall through to newMBB
8135 thisMBB->addSuccessor(newMBB);
8137 // newMBB jumps to itself and fall through to nextMBB
8138 newMBB->addSuccessor(nextMBB);
8139 newMBB->addSuccessor(newMBB);
8141 DebugLoc dl = bInstr->getDebugLoc();
8142 // Insert instructions into newMBB based on incoming instruction
8143 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8144 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8145 "unexpected number of operands");
8146 MachineOperand& dest1Oper = bInstr->getOperand(0);
8147 MachineOperand& dest2Oper = bInstr->getOperand(1);
8148 MachineOperand* argOpers[2 + X86AddrNumOperands];
8149 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8150 argOpers[i] = &bInstr->getOperand(i+2);
8152 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8153 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8155 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8156 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8157 for (int i=0; i <= lastAddrIndx; ++i)
8158 (*MIB).addOperand(*argOpers[i]);
8159 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8160 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8161 // add 4 to displacement.
8162 for (int i=0; i <= lastAddrIndx-2; ++i)
8163 (*MIB).addOperand(*argOpers[i]);
8164 MachineOperand newOp3 = *(argOpers[3]);
8166 newOp3.setImm(newOp3.getImm()+4);
8168 newOp3.setOffset(newOp3.getOffset()+4);
8169 (*MIB).addOperand(newOp3);
8170 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8172 // t3/4 are defined later, at the bottom of the loop
8173 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8174 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8175 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8176 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8177 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8178 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8180 // The subsequent operations should be using the destination registers of
8181 //the PHI instructions.
8183 t1 = F->getRegInfo().createVirtualRegister(RC);
8184 t2 = F->getRegInfo().createVirtualRegister(RC);
8185 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8186 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8188 t1 = dest1Oper.getReg();
8189 t2 = dest2Oper.getReg();
8192 int valArgIndx = lastAddrIndx + 1;
8193 assert((argOpers[valArgIndx]->isReg() ||
8194 argOpers[valArgIndx]->isImm()) &&
8196 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8197 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8198 if (argOpers[valArgIndx]->isReg())
8199 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8201 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8202 if (regOpcL != X86::MOV32rr)
8204 (*MIB).addOperand(*argOpers[valArgIndx]);
8205 assert(argOpers[valArgIndx + 1]->isReg() ==
8206 argOpers[valArgIndx]->isReg());
8207 assert(argOpers[valArgIndx + 1]->isImm() ==
8208 argOpers[valArgIndx]->isImm());
8209 if (argOpers[valArgIndx + 1]->isReg())
8210 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8212 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8213 if (regOpcH != X86::MOV32rr)
8215 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8217 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8219 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8222 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8224 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8227 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8228 for (int i=0; i <= lastAddrIndx; ++i)
8229 (*MIB).addOperand(*argOpers[i]);
8231 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8232 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8233 bInstr->memoperands_end());
8235 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8236 MIB.addReg(X86::EAX);
8237 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8238 MIB.addReg(X86::EDX);
8241 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8243 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8247 // private utility function
8249 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8250 MachineBasicBlock *MBB,
8251 unsigned cmovOpc) const {
8252 // For the atomic min/max operator, we generate
8255 // ld t1 = [min/max.addr]
8256 // mov t2 = [min/max.val]
8258 // cmov[cond] t2 = t1
8260 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8262 // fallthrough -->nextMBB
8264 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8265 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8266 MachineFunction::iterator MBBIter = MBB;
8269 /// First build the CFG
8270 MachineFunction *F = MBB->getParent();
8271 MachineBasicBlock *thisMBB = MBB;
8272 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8273 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8274 F->insert(MBBIter, newMBB);
8275 F->insert(MBBIter, nextMBB);
8277 // Move all successors of thisMBB to nextMBB
8278 nextMBB->transferSuccessors(thisMBB);
8280 // Update thisMBB to fall through to newMBB
8281 thisMBB->addSuccessor(newMBB);
8283 // newMBB jumps to newMBB and fall through to nextMBB
8284 newMBB->addSuccessor(nextMBB);
8285 newMBB->addSuccessor(newMBB);
8287 DebugLoc dl = mInstr->getDebugLoc();
8288 // Insert instructions into newMBB based on incoming instruction
8289 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8290 "unexpected number of operands");
8291 MachineOperand& destOper = mInstr->getOperand(0);
8292 MachineOperand* argOpers[2 + X86AddrNumOperands];
8293 int numArgs = mInstr->getNumOperands() - 1;
8294 for (int i=0; i < numArgs; ++i)
8295 argOpers[i] = &mInstr->getOperand(i+1);
8297 // x86 address has 4 operands: base, index, scale, and displacement
8298 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8299 int valArgIndx = lastAddrIndx + 1;
8301 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8302 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8303 for (int i=0; i <= lastAddrIndx; ++i)
8304 (*MIB).addOperand(*argOpers[i]);
8306 // We only support register and immediate values
8307 assert((argOpers[valArgIndx]->isReg() ||
8308 argOpers[valArgIndx]->isImm()) &&
8311 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8312 if (argOpers[valArgIndx]->isReg())
8313 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8315 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8316 (*MIB).addOperand(*argOpers[valArgIndx]);
8318 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8321 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8326 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8327 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8331 // Cmp and exchange if none has modified the memory location
8332 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8333 for (int i=0; i <= lastAddrIndx; ++i)
8334 (*MIB).addOperand(*argOpers[i]);
8336 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8337 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8338 mInstr->memoperands_end());
8340 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8341 MIB.addReg(X86::EAX);
8344 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8346 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8350 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8351 // all of this code can be replaced with that in the .td file.
8353 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8354 unsigned numArgs, bool memArg) const {
8356 MachineFunction *F = BB->getParent();
8357 DebugLoc dl = MI->getDebugLoc();
8358 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8362 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8364 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8366 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8368 for (unsigned i = 0; i < numArgs; ++i) {
8369 MachineOperand &Op = MI->getOperand(i+1);
8371 if (!(Op.isReg() && Op.isImplicit()))
8375 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8378 F->DeleteMachineInstr(MI);
8384 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8386 MachineBasicBlock *MBB) const {
8387 // Emit code to save XMM registers to the stack. The ABI says that the
8388 // number of registers to save is given in %al, so it's theoretically
8389 // possible to do an indirect jump trick to avoid saving all of them,
8390 // however this code takes a simpler approach and just executes all
8391 // of the stores if %al is non-zero. It's less code, and it's probably
8392 // easier on the hardware branch predictor, and stores aren't all that
8393 // expensive anyway.
8395 // Create the new basic blocks. One block contains all the XMM stores,
8396 // and one block is the final destination regardless of whether any
8397 // stores were performed.
8398 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8399 MachineFunction *F = MBB->getParent();
8400 MachineFunction::iterator MBBIter = MBB;
8402 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8403 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8404 F->insert(MBBIter, XMMSaveMBB);
8405 F->insert(MBBIter, EndMBB);
8408 // Move any original successors of MBB to the end block.
8409 EndMBB->transferSuccessors(MBB);
8410 // The original block will now fall through to the XMM save block.
8411 MBB->addSuccessor(XMMSaveMBB);
8412 // The XMMSaveMBB will fall through to the end block.
8413 XMMSaveMBB->addSuccessor(EndMBB);
8415 // Now add the instructions.
8416 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8417 DebugLoc DL = MI->getDebugLoc();
8419 unsigned CountReg = MI->getOperand(0).getReg();
8420 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8421 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8423 if (!Subtarget->isTargetWin64()) {
8424 // If %al is 0, branch around the XMM save block.
8425 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8426 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8427 MBB->addSuccessor(EndMBB);
8430 // In the XMM save block, save all the XMM argument registers.
8431 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8432 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8433 MachineMemOperand *MMO =
8434 F->getMachineMemOperand(
8435 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8436 MachineMemOperand::MOStore, Offset,
8437 /*Size=*/16, /*Align=*/16);
8438 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8439 .addFrameIndex(RegSaveFrameIndex)
8440 .addImm(/*Scale=*/1)
8441 .addReg(/*IndexReg=*/0)
8442 .addImm(/*Disp=*/Offset)
8443 .addReg(/*Segment=*/0)
8444 .addReg(MI->getOperand(i).getReg())
8445 .addMemOperand(MMO);
8448 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8454 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8455 MachineBasicBlock *BB,
8456 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8457 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8458 DebugLoc DL = MI->getDebugLoc();
8460 // To "insert" a SELECT_CC instruction, we actually have to insert the
8461 // diamond control-flow pattern. The incoming instruction knows the
8462 // destination vreg to set, the condition code register to branch on, the
8463 // true/false values to select between, and a branch opcode to use.
8464 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8465 MachineFunction::iterator It = BB;
8471 // cmpTY ccX, r1, r2
8473 // fallthrough --> copy0MBB
8474 MachineBasicBlock *thisMBB = BB;
8475 MachineFunction *F = BB->getParent();
8476 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8477 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8479 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8480 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8481 F->insert(It, copy0MBB);
8482 F->insert(It, sinkMBB);
8483 // Update machine-CFG edges by first adding all successors of the current
8484 // block to the new block which will contain the Phi node for the select.
8485 // Also inform sdisel of the edge changes.
8486 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8487 E = BB->succ_end(); I != E; ++I) {
8488 EM->insert(std::make_pair(*I, sinkMBB));
8489 sinkMBB->addSuccessor(*I);
8491 // Next, remove all successors of the current block, and add the true
8492 // and fallthrough blocks as its successors.
8493 while (!BB->succ_empty())
8494 BB->removeSuccessor(BB->succ_begin());
8495 // Add the true and fallthrough blocks as its successors.
8496 BB->addSuccessor(copy0MBB);
8497 BB->addSuccessor(sinkMBB);
8500 // %FalseValue = ...
8501 // # fallthrough to sinkMBB
8504 // Update machine-CFG edges
8505 BB->addSuccessor(sinkMBB);
8508 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8511 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8512 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8513 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8515 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8520 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8521 MachineBasicBlock *BB,
8522 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8523 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8524 DebugLoc DL = MI->getDebugLoc();
8525 MachineFunction *F = BB->getParent();
8527 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8528 // non-trivial part is impdef of ESP.
8529 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8532 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8533 .addExternalSymbol("_alloca")
8534 .addReg(X86::EAX, RegState::Implicit)
8535 .addReg(X86::ESP, RegState::Implicit)
8536 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8537 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8539 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8544 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8545 MachineBasicBlock *BB,
8546 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8547 switch (MI->getOpcode()) {
8548 default: assert(false && "Unexpected instr type to insert");
8549 case X86::MINGW_ALLOCA:
8550 return EmitLoweredMingwAlloca(MI, BB, EM);
8552 case X86::CMOV_V1I64:
8553 case X86::CMOV_FR32:
8554 case X86::CMOV_FR64:
8555 case X86::CMOV_V4F32:
8556 case X86::CMOV_V2F64:
8557 case X86::CMOV_V2I64:
8558 case X86::CMOV_GR16:
8559 case X86::CMOV_GR32:
8560 case X86::CMOV_RFP32:
8561 case X86::CMOV_RFP64:
8562 case X86::CMOV_RFP80:
8563 return EmitLoweredSelect(MI, BB, EM);
8565 case X86::FP32_TO_INT16_IN_MEM:
8566 case X86::FP32_TO_INT32_IN_MEM:
8567 case X86::FP32_TO_INT64_IN_MEM:
8568 case X86::FP64_TO_INT16_IN_MEM:
8569 case X86::FP64_TO_INT32_IN_MEM:
8570 case X86::FP64_TO_INT64_IN_MEM:
8571 case X86::FP80_TO_INT16_IN_MEM:
8572 case X86::FP80_TO_INT32_IN_MEM:
8573 case X86::FP80_TO_INT64_IN_MEM: {
8574 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8575 DebugLoc DL = MI->getDebugLoc();
8577 // Change the floating point control register to use "round towards zero"
8578 // mode when truncating to an integer value.
8579 MachineFunction *F = BB->getParent();
8580 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8581 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8583 // Load the old value of the high byte of the control word...
8585 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8586 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8589 // Set the high part to be round to zero...
8590 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8593 // Reload the modified control word now...
8594 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8596 // Restore the memory image of control word to original value
8597 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8600 // Get the X86 opcode to use.
8602 switch (MI->getOpcode()) {
8603 default: llvm_unreachable("illegal opcode!");
8604 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8605 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8606 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8607 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8608 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8609 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8610 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8611 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8612 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8616 MachineOperand &Op = MI->getOperand(0);
8618 AM.BaseType = X86AddressMode::RegBase;
8619 AM.Base.Reg = Op.getReg();
8621 AM.BaseType = X86AddressMode::FrameIndexBase;
8622 AM.Base.FrameIndex = Op.getIndex();
8624 Op = MI->getOperand(1);
8626 AM.Scale = Op.getImm();
8627 Op = MI->getOperand(2);
8629 AM.IndexReg = Op.getImm();
8630 Op = MI->getOperand(3);
8631 if (Op.isGlobal()) {
8632 AM.GV = Op.getGlobal();
8634 AM.Disp = Op.getImm();
8636 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8637 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8639 // Reload the original control word now.
8640 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8642 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8645 // DBG_VALUE. Only the frame index case is done here.
8646 case X86::DBG_VALUE: {
8647 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8648 DebugLoc DL = MI->getDebugLoc();
8650 MachineFunction *F = BB->getParent();
8651 AM.BaseType = X86AddressMode::FrameIndexBase;
8652 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8653 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8654 addImm(MI->getOperand(1).getImm()).
8655 addMetadata(MI->getOperand(2).getMetadata());
8656 F->DeleteMachineInstr(MI); // Remove pseudo.
8660 // String/text processing lowering.
8661 case X86::PCMPISTRM128REG:
8662 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8663 case X86::PCMPISTRM128MEM:
8664 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8665 case X86::PCMPESTRM128REG:
8666 return EmitPCMP(MI, BB, 5, false /* in mem */);
8667 case X86::PCMPESTRM128MEM:
8668 return EmitPCMP(MI, BB, 5, true /* in mem */);
8671 case X86::ATOMAND32:
8672 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8673 X86::AND32ri, X86::MOV32rm,
8674 X86::LCMPXCHG32, X86::MOV32rr,
8675 X86::NOT32r, X86::EAX,
8676 X86::GR32RegisterClass);
8678 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8679 X86::OR32ri, X86::MOV32rm,
8680 X86::LCMPXCHG32, X86::MOV32rr,
8681 X86::NOT32r, X86::EAX,
8682 X86::GR32RegisterClass);
8683 case X86::ATOMXOR32:
8684 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8685 X86::XOR32ri, X86::MOV32rm,
8686 X86::LCMPXCHG32, X86::MOV32rr,
8687 X86::NOT32r, X86::EAX,
8688 X86::GR32RegisterClass);
8689 case X86::ATOMNAND32:
8690 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8691 X86::AND32ri, X86::MOV32rm,
8692 X86::LCMPXCHG32, X86::MOV32rr,
8693 X86::NOT32r, X86::EAX,
8694 X86::GR32RegisterClass, true);
8695 case X86::ATOMMIN32:
8696 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8697 case X86::ATOMMAX32:
8698 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8699 case X86::ATOMUMIN32:
8700 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8701 case X86::ATOMUMAX32:
8702 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8704 case X86::ATOMAND16:
8705 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8706 X86::AND16ri, X86::MOV16rm,
8707 X86::LCMPXCHG16, X86::MOV16rr,
8708 X86::NOT16r, X86::AX,
8709 X86::GR16RegisterClass);
8711 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8712 X86::OR16ri, X86::MOV16rm,
8713 X86::LCMPXCHG16, X86::MOV16rr,
8714 X86::NOT16r, X86::AX,
8715 X86::GR16RegisterClass);
8716 case X86::ATOMXOR16:
8717 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8718 X86::XOR16ri, X86::MOV16rm,
8719 X86::LCMPXCHG16, X86::MOV16rr,
8720 X86::NOT16r, X86::AX,
8721 X86::GR16RegisterClass);
8722 case X86::ATOMNAND16:
8723 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8724 X86::AND16ri, X86::MOV16rm,
8725 X86::LCMPXCHG16, X86::MOV16rr,
8726 X86::NOT16r, X86::AX,
8727 X86::GR16RegisterClass, true);
8728 case X86::ATOMMIN16:
8729 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8730 case X86::ATOMMAX16:
8731 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8732 case X86::ATOMUMIN16:
8733 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8734 case X86::ATOMUMAX16:
8735 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8738 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8739 X86::AND8ri, X86::MOV8rm,
8740 X86::LCMPXCHG8, X86::MOV8rr,
8741 X86::NOT8r, X86::AL,
8742 X86::GR8RegisterClass);
8744 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8745 X86::OR8ri, X86::MOV8rm,
8746 X86::LCMPXCHG8, X86::MOV8rr,
8747 X86::NOT8r, X86::AL,
8748 X86::GR8RegisterClass);
8750 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8751 X86::XOR8ri, X86::MOV8rm,
8752 X86::LCMPXCHG8, X86::MOV8rr,
8753 X86::NOT8r, X86::AL,
8754 X86::GR8RegisterClass);
8755 case X86::ATOMNAND8:
8756 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8757 X86::AND8ri, X86::MOV8rm,
8758 X86::LCMPXCHG8, X86::MOV8rr,
8759 X86::NOT8r, X86::AL,
8760 X86::GR8RegisterClass, true);
8761 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8762 // This group is for 64-bit host.
8763 case X86::ATOMAND64:
8764 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8765 X86::AND64ri32, X86::MOV64rm,
8766 X86::LCMPXCHG64, X86::MOV64rr,
8767 X86::NOT64r, X86::RAX,
8768 X86::GR64RegisterClass);
8770 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8771 X86::OR64ri32, X86::MOV64rm,
8772 X86::LCMPXCHG64, X86::MOV64rr,
8773 X86::NOT64r, X86::RAX,
8774 X86::GR64RegisterClass);
8775 case X86::ATOMXOR64:
8776 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8777 X86::XOR64ri32, X86::MOV64rm,
8778 X86::LCMPXCHG64, X86::MOV64rr,
8779 X86::NOT64r, X86::RAX,
8780 X86::GR64RegisterClass);
8781 case X86::ATOMNAND64:
8782 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8783 X86::AND64ri32, X86::MOV64rm,
8784 X86::LCMPXCHG64, X86::MOV64rr,
8785 X86::NOT64r, X86::RAX,
8786 X86::GR64RegisterClass, true);
8787 case X86::ATOMMIN64:
8788 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8789 case X86::ATOMMAX64:
8790 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8791 case X86::ATOMUMIN64:
8792 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8793 case X86::ATOMUMAX64:
8794 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8796 // This group does 64-bit operations on a 32-bit host.
8797 case X86::ATOMAND6432:
8798 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8799 X86::AND32rr, X86::AND32rr,
8800 X86::AND32ri, X86::AND32ri,
8802 case X86::ATOMOR6432:
8803 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8804 X86::OR32rr, X86::OR32rr,
8805 X86::OR32ri, X86::OR32ri,
8807 case X86::ATOMXOR6432:
8808 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8809 X86::XOR32rr, X86::XOR32rr,
8810 X86::XOR32ri, X86::XOR32ri,
8812 case X86::ATOMNAND6432:
8813 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8814 X86::AND32rr, X86::AND32rr,
8815 X86::AND32ri, X86::AND32ri,
8817 case X86::ATOMADD6432:
8818 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8819 X86::ADD32rr, X86::ADC32rr,
8820 X86::ADD32ri, X86::ADC32ri,
8822 case X86::ATOMSUB6432:
8823 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8824 X86::SUB32rr, X86::SBB32rr,
8825 X86::SUB32ri, X86::SBB32ri,
8827 case X86::ATOMSWAP6432:
8828 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8829 X86::MOV32rr, X86::MOV32rr,
8830 X86::MOV32ri, X86::MOV32ri,
8832 case X86::VASTART_SAVE_XMM_REGS:
8833 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8837 //===----------------------------------------------------------------------===//
8838 // X86 Optimization Hooks
8839 //===----------------------------------------------------------------------===//
8841 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8845 const SelectionDAG &DAG,
8846 unsigned Depth) const {
8847 unsigned Opc = Op.getOpcode();
8848 assert((Opc >= ISD::BUILTIN_OP_END ||
8849 Opc == ISD::INTRINSIC_WO_CHAIN ||
8850 Opc == ISD::INTRINSIC_W_CHAIN ||
8851 Opc == ISD::INTRINSIC_VOID) &&
8852 "Should use MaskedValueIsZero if you don't know whether Op"
8853 " is a target node!");
8855 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8867 // These nodes' second result is a boolean.
8868 if (Op.getResNo() == 0)
8872 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8873 Mask.getBitWidth() - 1);
8878 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8879 /// node is a GlobalAddress + offset.
8880 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8881 GlobalValue* &GA, int64_t &Offset) const{
8882 if (N->getOpcode() == X86ISD::Wrapper) {
8883 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8884 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8885 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8889 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8892 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8893 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8894 /// if the load addresses are consecutive, non-overlapping, and in the right
8896 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8897 const TargetLowering &TLI) {
8898 DebugLoc dl = N->getDebugLoc();
8899 EVT VT = N->getValueType(0);
8900 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8902 if (VT.getSizeInBits() != 128)
8905 SmallVector<SDValue, 16> Elts;
8906 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8907 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8909 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8912 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8913 /// and convert it from being a bunch of shuffles and extracts to a simple
8914 /// store and scalar loads to extract the elements.
8915 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8916 const TargetLowering &TLI) {
8917 SDValue InputVector = N->getOperand(0);
8919 // Only operate on vectors of 4 elements, where the alternative shuffling
8920 // gets to be more expensive.
8921 if (InputVector.getValueType() != MVT::v4i32)
8924 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8925 // single use which is a sign-extend or zero-extend, and all elements are
8927 SmallVector<SDNode *, 4> Uses;
8928 unsigned ExtractedElements = 0;
8929 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8930 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8931 if (UI.getUse().getResNo() != InputVector.getResNo())
8934 SDNode *Extract = *UI;
8935 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8938 if (Extract->getValueType(0) != MVT::i32)
8940 if (!Extract->hasOneUse())
8942 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8943 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8945 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8948 // Record which element was extracted.
8949 ExtractedElements |=
8950 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8952 Uses.push_back(Extract);
8955 // If not all the elements were used, this may not be worthwhile.
8956 if (ExtractedElements != 15)
8959 // Ok, we've now decided to do the transformation.
8960 DebugLoc dl = InputVector.getDebugLoc();
8962 // Store the value to a temporary stack slot.
8963 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8964 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8967 // Replace each use (extract) with a load of the appropriate element.
8968 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8969 UE = Uses.end(); UI != UE; ++UI) {
8970 SDNode *Extract = *UI;
8972 // Compute the element's address.
8973 SDValue Idx = Extract->getOperand(1);
8975 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8976 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8977 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8979 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8982 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8983 NULL, 0, false, false, 0);
8985 // Replace the exact with the load.
8986 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8989 // The replacement was made in place; don't return anything.
8993 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8994 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8995 const X86Subtarget *Subtarget) {
8996 DebugLoc DL = N->getDebugLoc();
8997 SDValue Cond = N->getOperand(0);
8998 // Get the LHS/RHS of the select.
8999 SDValue LHS = N->getOperand(1);
9000 SDValue RHS = N->getOperand(2);
9002 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9003 // instructions match the semantics of the common C idiom x<y?x:y but not
9004 // x<=y?x:y, because of how they handle negative zero (which can be
9005 // ignored in unsafe-math mode).
9006 if (Subtarget->hasSSE2() &&
9007 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9008 Cond.getOpcode() == ISD::SETCC) {
9009 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9011 unsigned Opcode = 0;
9012 // Check for x CC y ? x : y.
9013 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9014 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9018 // Converting this to a min would handle NaNs incorrectly, and swapping
9019 // the operands would cause it to handle comparisons between positive
9020 // and negative zero incorrectly.
9021 if (!FiniteOnlyFPMath() &&
9022 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9023 if (!UnsafeFPMath &&
9024 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9026 std::swap(LHS, RHS);
9028 Opcode = X86ISD::FMIN;
9031 // Converting this to a min would handle comparisons between positive
9032 // and negative zero incorrectly.
9033 if (!UnsafeFPMath &&
9034 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9036 Opcode = X86ISD::FMIN;
9039 // Converting this to a min would handle both negative zeros and NaNs
9040 // incorrectly, but we can swap the operands to fix both.
9041 std::swap(LHS, RHS);
9045 Opcode = X86ISD::FMIN;
9049 // Converting this to a max would handle comparisons between positive
9050 // and negative zero incorrectly.
9051 if (!UnsafeFPMath &&
9052 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9054 Opcode = X86ISD::FMAX;
9057 // Converting this to a max would handle NaNs incorrectly, and swapping
9058 // the operands would cause it to handle comparisons between positive
9059 // and negative zero incorrectly.
9060 if (!FiniteOnlyFPMath() &&
9061 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9062 if (!UnsafeFPMath &&
9063 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9065 std::swap(LHS, RHS);
9067 Opcode = X86ISD::FMAX;
9070 // Converting this to a max would handle both negative zeros and NaNs
9071 // incorrectly, but we can swap the operands to fix both.
9072 std::swap(LHS, RHS);
9076 Opcode = X86ISD::FMAX;
9079 // Check for x CC y ? y : x -- a min/max with reversed arms.
9080 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9081 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9085 // Converting this to a min would handle comparisons between positive
9086 // and negative zero incorrectly, and swapping the operands would
9087 // cause it to handle NaNs incorrectly.
9088 if (!UnsafeFPMath &&
9089 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9090 if (!FiniteOnlyFPMath() &&
9091 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9093 std::swap(LHS, RHS);
9095 Opcode = X86ISD::FMIN;
9098 // Converting this to a min would handle NaNs incorrectly.
9099 if (!UnsafeFPMath &&
9100 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9102 Opcode = X86ISD::FMIN;
9105 // Converting this to a min would handle both negative zeros and NaNs
9106 // incorrectly, but we can swap the operands to fix both.
9107 std::swap(LHS, RHS);
9111 Opcode = X86ISD::FMIN;
9115 // Converting this to a max would handle NaNs incorrectly.
9116 if (!FiniteOnlyFPMath() &&
9117 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9119 Opcode = X86ISD::FMAX;
9122 // Converting this to a max would handle comparisons between positive
9123 // and negative zero incorrectly, and swapping the operands would
9124 // cause it to handle NaNs incorrectly.
9125 if (!UnsafeFPMath &&
9126 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9127 if (!FiniteOnlyFPMath() &&
9128 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9130 std::swap(LHS, RHS);
9132 Opcode = X86ISD::FMAX;
9135 // Converting this to a max would handle both negative zeros and NaNs
9136 // incorrectly, but we can swap the operands to fix both.
9137 std::swap(LHS, RHS);
9141 Opcode = X86ISD::FMAX;
9147 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9150 // If this is a select between two integer constants, try to do some
9152 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9153 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9154 // Don't do this for crazy integer types.
9155 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9156 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9157 // so that TrueC (the true value) is larger than FalseC.
9158 bool NeedsCondInvert = false;
9160 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9161 // Efficiently invertible.
9162 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9163 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9164 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9165 NeedsCondInvert = true;
9166 std::swap(TrueC, FalseC);
9169 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9170 if (FalseC->getAPIntValue() == 0 &&
9171 TrueC->getAPIntValue().isPowerOf2()) {
9172 if (NeedsCondInvert) // Invert the condition if needed.
9173 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9174 DAG.getConstant(1, Cond.getValueType()));
9176 // Zero extend the condition if needed.
9177 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9179 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9180 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9181 DAG.getConstant(ShAmt, MVT::i8));
9184 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9185 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9186 if (NeedsCondInvert) // Invert the condition if needed.
9187 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9188 DAG.getConstant(1, Cond.getValueType()));
9190 // Zero extend the condition if needed.
9191 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9192 FalseC->getValueType(0), Cond);
9193 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9194 SDValue(FalseC, 0));
9197 // Optimize cases that will turn into an LEA instruction. This requires
9198 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9199 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9200 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9201 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9203 bool isFastMultiplier = false;
9205 switch ((unsigned char)Diff) {
9207 case 1: // result = add base, cond
9208 case 2: // result = lea base( , cond*2)
9209 case 3: // result = lea base(cond, cond*2)
9210 case 4: // result = lea base( , cond*4)
9211 case 5: // result = lea base(cond, cond*4)
9212 case 8: // result = lea base( , cond*8)
9213 case 9: // result = lea base(cond, cond*8)
9214 isFastMultiplier = true;
9219 if (isFastMultiplier) {
9220 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9221 if (NeedsCondInvert) // Invert the condition if needed.
9222 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9223 DAG.getConstant(1, Cond.getValueType()));
9225 // Zero extend the condition if needed.
9226 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9228 // Scale the condition by the difference.
9230 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9231 DAG.getConstant(Diff, Cond.getValueType()));
9233 // Add the base if non-zero.
9234 if (FalseC->getAPIntValue() != 0)
9235 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9236 SDValue(FalseC, 0));
9246 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9247 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9248 TargetLowering::DAGCombinerInfo &DCI) {
9249 DebugLoc DL = N->getDebugLoc();
9251 // If the flag operand isn't dead, don't touch this CMOV.
9252 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9255 // If this is a select between two integer constants, try to do some
9256 // optimizations. Note that the operands are ordered the opposite of SELECT
9258 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9259 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9260 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9261 // larger than FalseC (the false value).
9262 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9264 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9265 CC = X86::GetOppositeBranchCondition(CC);
9266 std::swap(TrueC, FalseC);
9269 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9270 // This is efficient for any integer data type (including i8/i16) and
9272 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9273 SDValue Cond = N->getOperand(3);
9274 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9275 DAG.getConstant(CC, MVT::i8), Cond);
9277 // Zero extend the condition if needed.
9278 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9280 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9281 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9282 DAG.getConstant(ShAmt, MVT::i8));
9283 if (N->getNumValues() == 2) // Dead flag value?
9284 return DCI.CombineTo(N, Cond, SDValue());
9288 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9289 // for any integer data type, including i8/i16.
9290 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9291 SDValue Cond = N->getOperand(3);
9292 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9293 DAG.getConstant(CC, MVT::i8), Cond);
9295 // Zero extend the condition if needed.
9296 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9297 FalseC->getValueType(0), Cond);
9298 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9299 SDValue(FalseC, 0));
9301 if (N->getNumValues() == 2) // Dead flag value?
9302 return DCI.CombineTo(N, Cond, SDValue());
9306 // Optimize cases that will turn into an LEA instruction. This requires
9307 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9308 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9309 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9310 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9312 bool isFastMultiplier = false;
9314 switch ((unsigned char)Diff) {
9316 case 1: // result = add base, cond
9317 case 2: // result = lea base( , cond*2)
9318 case 3: // result = lea base(cond, cond*2)
9319 case 4: // result = lea base( , cond*4)
9320 case 5: // result = lea base(cond, cond*4)
9321 case 8: // result = lea base( , cond*8)
9322 case 9: // result = lea base(cond, cond*8)
9323 isFastMultiplier = true;
9328 if (isFastMultiplier) {
9329 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9330 SDValue Cond = N->getOperand(3);
9331 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9332 DAG.getConstant(CC, MVT::i8), Cond);
9333 // Zero extend the condition if needed.
9334 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9336 // Scale the condition by the difference.
9338 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9339 DAG.getConstant(Diff, Cond.getValueType()));
9341 // Add the base if non-zero.
9342 if (FalseC->getAPIntValue() != 0)
9343 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9344 SDValue(FalseC, 0));
9345 if (N->getNumValues() == 2) // Dead flag value?
9346 return DCI.CombineTo(N, Cond, SDValue());
9356 /// PerformMulCombine - Optimize a single multiply with constant into two
9357 /// in order to implement it with two cheaper instructions, e.g.
9358 /// LEA + SHL, LEA + LEA.
9359 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9360 TargetLowering::DAGCombinerInfo &DCI) {
9361 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9364 EVT VT = N->getValueType(0);
9368 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9371 uint64_t MulAmt = C->getZExtValue();
9372 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9375 uint64_t MulAmt1 = 0;
9376 uint64_t MulAmt2 = 0;
9377 if ((MulAmt % 9) == 0) {
9379 MulAmt2 = MulAmt / 9;
9380 } else if ((MulAmt % 5) == 0) {
9382 MulAmt2 = MulAmt / 5;
9383 } else if ((MulAmt % 3) == 0) {
9385 MulAmt2 = MulAmt / 3;
9388 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9389 DebugLoc DL = N->getDebugLoc();
9391 if (isPowerOf2_64(MulAmt2) &&
9392 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9393 // If second multiplifer is pow2, issue it first. We want the multiply by
9394 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9396 std::swap(MulAmt1, MulAmt2);
9399 if (isPowerOf2_64(MulAmt1))
9400 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9401 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9403 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9404 DAG.getConstant(MulAmt1, VT));
9406 if (isPowerOf2_64(MulAmt2))
9407 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9408 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9410 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9411 DAG.getConstant(MulAmt2, VT));
9413 // Do not add new nodes to DAG combiner worklist.
9414 DCI.CombineTo(N, NewMul, false);
9419 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9420 SDValue N0 = N->getOperand(0);
9421 SDValue N1 = N->getOperand(1);
9422 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9423 EVT VT = N0.getValueType();
9425 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9426 // since the result of setcc_c is all zero's or all ones.
9427 if (N1C && N0.getOpcode() == ISD::AND &&
9428 N0.getOperand(1).getOpcode() == ISD::Constant) {
9429 SDValue N00 = N0.getOperand(0);
9430 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9431 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9432 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9433 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9434 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9435 APInt ShAmt = N1C->getAPIntValue();
9436 Mask = Mask.shl(ShAmt);
9438 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9439 N00, DAG.getConstant(Mask, VT));
9446 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9448 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9449 const X86Subtarget *Subtarget) {
9450 EVT VT = N->getValueType(0);
9451 if (!VT.isVector() && VT.isInteger() &&
9452 N->getOpcode() == ISD::SHL)
9453 return PerformSHLCombine(N, DAG);
9455 // On X86 with SSE2 support, we can transform this to a vector shift if
9456 // all elements are shifted by the same amount. We can't do this in legalize
9457 // because the a constant vector is typically transformed to a constant pool
9458 // so we have no knowledge of the shift amount.
9459 if (!Subtarget->hasSSE2())
9462 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9465 SDValue ShAmtOp = N->getOperand(1);
9466 EVT EltVT = VT.getVectorElementType();
9467 DebugLoc DL = N->getDebugLoc();
9468 SDValue BaseShAmt = SDValue();
9469 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9470 unsigned NumElts = VT.getVectorNumElements();
9472 for (; i != NumElts; ++i) {
9473 SDValue Arg = ShAmtOp.getOperand(i);
9474 if (Arg.getOpcode() == ISD::UNDEF) continue;
9478 for (; i != NumElts; ++i) {
9479 SDValue Arg = ShAmtOp.getOperand(i);
9480 if (Arg.getOpcode() == ISD::UNDEF) continue;
9481 if (Arg != BaseShAmt) {
9485 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9486 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9487 SDValue InVec = ShAmtOp.getOperand(0);
9488 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9489 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9491 for (; i != NumElts; ++i) {
9492 SDValue Arg = InVec.getOperand(i);
9493 if (Arg.getOpcode() == ISD::UNDEF) continue;
9497 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9498 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9499 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9500 if (C->getZExtValue() == SplatIdx)
9501 BaseShAmt = InVec.getOperand(1);
9504 if (BaseShAmt.getNode() == 0)
9505 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9506 DAG.getIntPtrConstant(0));
9510 // The shift amount is an i32.
9511 if (EltVT.bitsGT(MVT::i32))
9512 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9513 else if (EltVT.bitsLT(MVT::i32))
9514 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9516 // The shift amount is identical so we can do a vector shift.
9517 SDValue ValOp = N->getOperand(0);
9518 switch (N->getOpcode()) {
9520 llvm_unreachable("Unknown shift opcode!");
9523 if (VT == MVT::v2i64)
9524 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9525 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9527 if (VT == MVT::v4i32)
9528 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9529 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9531 if (VT == MVT::v8i16)
9532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9533 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9537 if (VT == MVT::v4i32)
9538 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9539 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9541 if (VT == MVT::v8i16)
9542 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9543 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9547 if (VT == MVT::v2i64)
9548 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9549 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9551 if (VT == MVT::v4i32)
9552 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9553 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9555 if (VT == MVT::v8i16)
9556 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9557 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9564 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9565 const X86Subtarget *Subtarget) {
9566 EVT VT = N->getValueType(0);
9567 if (VT != MVT::i64 || !Subtarget->is64Bit())
9570 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9571 SDValue N0 = N->getOperand(0);
9572 SDValue N1 = N->getOperand(1);
9573 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9575 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9578 SDValue ShAmt0 = N0.getOperand(1);
9579 if (ShAmt0.getValueType() != MVT::i8)
9581 SDValue ShAmt1 = N1.getOperand(1);
9582 if (ShAmt1.getValueType() != MVT::i8)
9584 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9585 ShAmt0 = ShAmt0.getOperand(0);
9586 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9587 ShAmt1 = ShAmt1.getOperand(0);
9589 DebugLoc DL = N->getDebugLoc();
9590 unsigned Opc = X86ISD::SHLD;
9591 SDValue Op0 = N0.getOperand(0);
9592 SDValue Op1 = N1.getOperand(0);
9593 if (ShAmt0.getOpcode() == ISD::SUB) {
9595 std::swap(Op0, Op1);
9596 std::swap(ShAmt0, ShAmt1);
9599 if (ShAmt1.getOpcode() == ISD::SUB) {
9600 SDValue Sum = ShAmt1.getOperand(0);
9601 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9602 if (SumC->getSExtValue() == 64 &&
9603 ShAmt1.getOperand(1) == ShAmt0)
9604 return DAG.getNode(Opc, DL, VT,
9606 DAG.getNode(ISD::TRUNCATE, DL,
9609 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9610 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9612 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9613 return DAG.getNode(Opc, DL, VT,
9614 N0.getOperand(0), N1.getOperand(0),
9615 DAG.getNode(ISD::TRUNCATE, DL,
9622 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9623 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9624 const X86Subtarget *Subtarget) {
9625 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9626 // the FP state in cases where an emms may be missing.
9627 // A preferable solution to the general problem is to figure out the right
9628 // places to insert EMMS. This qualifies as a quick hack.
9630 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9631 StoreSDNode *St = cast<StoreSDNode>(N);
9632 EVT VT = St->getValue().getValueType();
9633 if (VT.getSizeInBits() != 64)
9636 const Function *F = DAG.getMachineFunction().getFunction();
9637 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9638 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9639 && Subtarget->hasSSE2();
9640 if ((VT.isVector() ||
9641 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9642 isa<LoadSDNode>(St->getValue()) &&
9643 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9644 St->getChain().hasOneUse() && !St->isVolatile()) {
9645 SDNode* LdVal = St->getValue().getNode();
9647 int TokenFactorIndex = -1;
9648 SmallVector<SDValue, 8> Ops;
9649 SDNode* ChainVal = St->getChain().getNode();
9650 // Must be a store of a load. We currently handle two cases: the load
9651 // is a direct child, and it's under an intervening TokenFactor. It is
9652 // possible to dig deeper under nested TokenFactors.
9653 if (ChainVal == LdVal)
9654 Ld = cast<LoadSDNode>(St->getChain());
9655 else if (St->getValue().hasOneUse() &&
9656 ChainVal->getOpcode() == ISD::TokenFactor) {
9657 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9658 if (ChainVal->getOperand(i).getNode() == LdVal) {
9659 TokenFactorIndex = i;
9660 Ld = cast<LoadSDNode>(St->getValue());
9662 Ops.push_back(ChainVal->getOperand(i));
9666 if (!Ld || !ISD::isNormalLoad(Ld))
9669 // If this is not the MMX case, i.e. we are just turning i64 load/store
9670 // into f64 load/store, avoid the transformation if there are multiple
9671 // uses of the loaded value.
9672 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9675 DebugLoc LdDL = Ld->getDebugLoc();
9676 DebugLoc StDL = N->getDebugLoc();
9677 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9678 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9680 if (Subtarget->is64Bit() || F64IsLegal) {
9681 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9682 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9683 Ld->getBasePtr(), Ld->getSrcValue(),
9684 Ld->getSrcValueOffset(), Ld->isVolatile(),
9685 Ld->isNonTemporal(), Ld->getAlignment());
9686 SDValue NewChain = NewLd.getValue(1);
9687 if (TokenFactorIndex != -1) {
9688 Ops.push_back(NewChain);
9689 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9692 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9693 St->getSrcValue(), St->getSrcValueOffset(),
9694 St->isVolatile(), St->isNonTemporal(),
9695 St->getAlignment());
9698 // Otherwise, lower to two pairs of 32-bit loads / stores.
9699 SDValue LoAddr = Ld->getBasePtr();
9700 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9701 DAG.getConstant(4, MVT::i32));
9703 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9704 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9705 Ld->isVolatile(), Ld->isNonTemporal(),
9706 Ld->getAlignment());
9707 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9708 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9709 Ld->isVolatile(), Ld->isNonTemporal(),
9710 MinAlign(Ld->getAlignment(), 4));
9712 SDValue NewChain = LoLd.getValue(1);
9713 if (TokenFactorIndex != -1) {
9714 Ops.push_back(LoLd);
9715 Ops.push_back(HiLd);
9716 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9720 LoAddr = St->getBasePtr();
9721 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9722 DAG.getConstant(4, MVT::i32));
9724 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9725 St->getSrcValue(), St->getSrcValueOffset(),
9726 St->isVolatile(), St->isNonTemporal(),
9727 St->getAlignment());
9728 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9730 St->getSrcValueOffset() + 4,
9732 St->isNonTemporal(),
9733 MinAlign(St->getAlignment(), 4));
9734 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9739 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9740 /// X86ISD::FXOR nodes.
9741 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9742 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9743 // F[X]OR(0.0, x) -> x
9744 // F[X]OR(x, 0.0) -> x
9745 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9746 if (C->getValueAPF().isPosZero())
9747 return N->getOperand(1);
9748 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9749 if (C->getValueAPF().isPosZero())
9750 return N->getOperand(0);
9754 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9755 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9756 // FAND(0.0, x) -> 0.0
9757 // FAND(x, 0.0) -> 0.0
9758 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9759 if (C->getValueAPF().isPosZero())
9760 return N->getOperand(0);
9761 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9762 if (C->getValueAPF().isPosZero())
9763 return N->getOperand(1);
9767 static SDValue PerformBTCombine(SDNode *N,
9769 TargetLowering::DAGCombinerInfo &DCI) {
9770 // BT ignores high bits in the bit index operand.
9771 SDValue Op1 = N->getOperand(1);
9772 if (Op1.hasOneUse()) {
9773 unsigned BitWidth = Op1.getValueSizeInBits();
9774 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9775 APInt KnownZero, KnownOne;
9776 TargetLowering::TargetLoweringOpt TLO(DAG);
9777 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9778 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9779 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9780 DCI.CommitTargetLoweringOpt(TLO);
9785 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9786 SDValue Op = N->getOperand(0);
9787 if (Op.getOpcode() == ISD::BIT_CONVERT)
9788 Op = Op.getOperand(0);
9789 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9790 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9791 VT.getVectorElementType().getSizeInBits() ==
9792 OpVT.getVectorElementType().getSizeInBits()) {
9793 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9798 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9799 // Locked instructions, in turn, have implicit fence semantics (all memory
9800 // operations are flushed before issuing the locked instruction, and the
9801 // are not buffered), so we can fold away the common pattern of
9802 // fence-atomic-fence.
9803 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9804 SDValue atomic = N->getOperand(0);
9805 switch (atomic.getOpcode()) {
9806 case ISD::ATOMIC_CMP_SWAP:
9807 case ISD::ATOMIC_SWAP:
9808 case ISD::ATOMIC_LOAD_ADD:
9809 case ISD::ATOMIC_LOAD_SUB:
9810 case ISD::ATOMIC_LOAD_AND:
9811 case ISD::ATOMIC_LOAD_OR:
9812 case ISD::ATOMIC_LOAD_XOR:
9813 case ISD::ATOMIC_LOAD_NAND:
9814 case ISD::ATOMIC_LOAD_MIN:
9815 case ISD::ATOMIC_LOAD_MAX:
9816 case ISD::ATOMIC_LOAD_UMIN:
9817 case ISD::ATOMIC_LOAD_UMAX:
9823 SDValue fence = atomic.getOperand(0);
9824 if (fence.getOpcode() != ISD::MEMBARRIER)
9827 switch (atomic.getOpcode()) {
9828 case ISD::ATOMIC_CMP_SWAP:
9829 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9830 atomic.getOperand(1), atomic.getOperand(2),
9831 atomic.getOperand(3));
9832 case ISD::ATOMIC_SWAP:
9833 case ISD::ATOMIC_LOAD_ADD:
9834 case ISD::ATOMIC_LOAD_SUB:
9835 case ISD::ATOMIC_LOAD_AND:
9836 case ISD::ATOMIC_LOAD_OR:
9837 case ISD::ATOMIC_LOAD_XOR:
9838 case ISD::ATOMIC_LOAD_NAND:
9839 case ISD::ATOMIC_LOAD_MIN:
9840 case ISD::ATOMIC_LOAD_MAX:
9841 case ISD::ATOMIC_LOAD_UMIN:
9842 case ISD::ATOMIC_LOAD_UMAX:
9843 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9844 atomic.getOperand(1), atomic.getOperand(2));
9850 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9851 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9852 // (and (i32 x86isd::setcc_carry), 1)
9853 // This eliminates the zext. This transformation is necessary because
9854 // ISD::SETCC is always legalized to i8.
9855 DebugLoc dl = N->getDebugLoc();
9856 SDValue N0 = N->getOperand(0);
9857 EVT VT = N->getValueType(0);
9858 if (N0.getOpcode() == ISD::AND &&
9860 N0.getOperand(0).hasOneUse()) {
9861 SDValue N00 = N0.getOperand(0);
9862 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9864 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9865 if (!C || C->getZExtValue() != 1)
9867 return DAG.getNode(ISD::AND, dl, VT,
9868 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9869 N00.getOperand(0), N00.getOperand(1)),
9870 DAG.getConstant(1, VT));
9876 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9877 DAGCombinerInfo &DCI) const {
9878 SelectionDAG &DAG = DCI.DAG;
9879 switch (N->getOpcode()) {
9881 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9882 case ISD::EXTRACT_VECTOR_ELT:
9883 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9884 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9885 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9886 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9889 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9890 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9891 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9893 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9894 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9895 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9896 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9897 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9898 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9904 //===----------------------------------------------------------------------===//
9905 // X86 Inline Assembly Support
9906 //===----------------------------------------------------------------------===//
9908 static bool LowerToBSwap(CallInst *CI) {
9909 // FIXME: this should verify that we are targetting a 486 or better. If not,
9910 // we will turn this bswap into something that will be lowered to logical ops
9911 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9912 // so don't worry about this.
9914 // Verify this is a simple bswap.
9915 if (CI->getNumOperands() != 2 ||
9916 CI->getType() != CI->getOperand(1)->getType() ||
9917 !CI->getType()->isIntegerTy())
9920 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9921 if (!Ty || Ty->getBitWidth() % 16 != 0)
9924 // Okay, we can do this xform, do so now.
9925 const Type *Tys[] = { Ty };
9926 Module *M = CI->getParent()->getParent()->getParent();
9927 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9929 Value *Op = CI->getOperand(1);
9930 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9932 CI->replaceAllUsesWith(Op);
9933 CI->eraseFromParent();
9937 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9938 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9939 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9941 std::string AsmStr = IA->getAsmString();
9943 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9944 SmallVector<StringRef, 4> AsmPieces;
9945 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9947 switch (AsmPieces.size()) {
9948 default: return false;
9950 AsmStr = AsmPieces[0];
9952 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9955 if (AsmPieces.size() == 2 &&
9956 (AsmPieces[0] == "bswap" ||
9957 AsmPieces[0] == "bswapq" ||
9958 AsmPieces[0] == "bswapl") &&
9959 (AsmPieces[1] == "$0" ||
9960 AsmPieces[1] == "${0:q}")) {
9961 // No need to check constraints, nothing other than the equivalent of
9962 // "=r,0" would be valid here.
9963 return LowerToBSwap(CI);
9965 // rorw $$8, ${0:w} --> llvm.bswap.i16
9966 if (CI->getType()->isIntegerTy(16) &&
9967 AsmPieces.size() == 3 &&
9968 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9969 AsmPieces[1] == "$$8," &&
9970 AsmPieces[2] == "${0:w}" &&
9971 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9973 const std::string &Constraints = IA->getConstraintString();
9974 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
9975 std::sort(AsmPieces.begin(), AsmPieces.end());
9976 if (AsmPieces.size() == 4 &&
9977 AsmPieces[0] == "~{cc}" &&
9978 AsmPieces[1] == "~{dirflag}" &&
9979 AsmPieces[2] == "~{flags}" &&
9980 AsmPieces[3] == "~{fpsr}") {
9981 return LowerToBSwap(CI);
9986 if (CI->getType()->isIntegerTy(64) &&
9987 Constraints.size() >= 2 &&
9988 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9989 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9990 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9991 SmallVector<StringRef, 4> Words;
9992 SplitString(AsmPieces[0], Words, " \t");
9993 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9995 SplitString(AsmPieces[1], Words, " \t");
9996 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9998 SplitString(AsmPieces[2], Words, " \t,");
9999 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10000 Words[2] == "%edx") {
10001 return LowerToBSwap(CI);
10013 /// getConstraintType - Given a constraint letter, return the type of
10014 /// constraint it is for this target.
10015 X86TargetLowering::ConstraintType
10016 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10017 if (Constraint.size() == 1) {
10018 switch (Constraint[0]) {
10030 return C_RegisterClass;
10038 return TargetLowering::getConstraintType(Constraint);
10041 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10042 /// with another that has more specific requirements based on the type of the
10043 /// corresponding operand.
10044 const char *X86TargetLowering::
10045 LowerXConstraint(EVT ConstraintVT) const {
10046 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10047 // 'f' like normal targets.
10048 if (ConstraintVT.isFloatingPoint()) {
10049 if (Subtarget->hasSSE2())
10051 if (Subtarget->hasSSE1())
10055 return TargetLowering::LowerXConstraint(ConstraintVT);
10058 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10059 /// vector. If it is invalid, don't add anything to Ops.
10060 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10063 std::vector<SDValue>&Ops,
10064 SelectionDAG &DAG) const {
10065 SDValue Result(0, 0);
10067 switch (Constraint) {
10070 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10071 if (C->getZExtValue() <= 31) {
10072 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10078 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10079 if (C->getZExtValue() <= 63) {
10080 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10086 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10087 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10088 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10094 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10095 if (C->getZExtValue() <= 255) {
10096 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10102 // 32-bit signed value
10103 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10104 const ConstantInt *CI = C->getConstantIntValue();
10105 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10106 C->getSExtValue())) {
10107 // Widen to 64 bits here to get it sign extended.
10108 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10111 // FIXME gcc accepts some relocatable values here too, but only in certain
10112 // memory models; it's complicated.
10117 // 32-bit unsigned value
10118 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10119 const ConstantInt *CI = C->getConstantIntValue();
10120 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10121 C->getZExtValue())) {
10122 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10126 // FIXME gcc accepts some relocatable values here too, but only in certain
10127 // memory models; it's complicated.
10131 // Literal immediates are always ok.
10132 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10133 // Widen to 64 bits here to get it sign extended.
10134 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10138 // If we are in non-pic codegen mode, we allow the address of a global (with
10139 // an optional displacement) to be used with 'i'.
10140 GlobalAddressSDNode *GA = 0;
10141 int64_t Offset = 0;
10143 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10145 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10146 Offset += GA->getOffset();
10148 } else if (Op.getOpcode() == ISD::ADD) {
10149 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10150 Offset += C->getZExtValue();
10151 Op = Op.getOperand(0);
10154 } else if (Op.getOpcode() == ISD::SUB) {
10155 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10156 Offset += -C->getZExtValue();
10157 Op = Op.getOperand(0);
10162 // Otherwise, this isn't something we can handle, reject it.
10166 GlobalValue *GV = GA->getGlobal();
10167 // If we require an extra load to get this address, as in PIC mode, we
10168 // can't accept it.
10169 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10170 getTargetMachine())))
10174 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10176 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10182 if (Result.getNode()) {
10183 Ops.push_back(Result);
10186 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10190 std::vector<unsigned> X86TargetLowering::
10191 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10193 if (Constraint.size() == 1) {
10194 // FIXME: not handling fp-stack yet!
10195 switch (Constraint[0]) { // GCC X86 Constraint Letters
10196 default: break; // Unknown constraint letter
10197 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10198 if (Subtarget->is64Bit()) {
10199 if (VT == MVT::i32)
10200 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10201 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10202 X86::R10D,X86::R11D,X86::R12D,
10203 X86::R13D,X86::R14D,X86::R15D,
10204 X86::EBP, X86::ESP, 0);
10205 else if (VT == MVT::i16)
10206 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10207 X86::SI, X86::DI, X86::R8W,X86::R9W,
10208 X86::R10W,X86::R11W,X86::R12W,
10209 X86::R13W,X86::R14W,X86::R15W,
10210 X86::BP, X86::SP, 0);
10211 else if (VT == MVT::i8)
10212 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10213 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10214 X86::R10B,X86::R11B,X86::R12B,
10215 X86::R13B,X86::R14B,X86::R15B,
10216 X86::BPL, X86::SPL, 0);
10218 else if (VT == MVT::i64)
10219 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10220 X86::RSI, X86::RDI, X86::R8, X86::R9,
10221 X86::R10, X86::R11, X86::R12,
10222 X86::R13, X86::R14, X86::R15,
10223 X86::RBP, X86::RSP, 0);
10227 // 32-bit fallthrough
10228 case 'Q': // Q_REGS
10229 if (VT == MVT::i32)
10230 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10231 else if (VT == MVT::i16)
10232 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10233 else if (VT == MVT::i8)
10234 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10235 else if (VT == MVT::i64)
10236 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10241 return std::vector<unsigned>();
10244 std::pair<unsigned, const TargetRegisterClass*>
10245 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10247 // First, see if this is a constraint that directly corresponds to an LLVM
10249 if (Constraint.size() == 1) {
10250 // GCC Constraint Letters
10251 switch (Constraint[0]) {
10253 case 'r': // GENERAL_REGS
10254 case 'l': // INDEX_REGS
10256 return std::make_pair(0U, X86::GR8RegisterClass);
10257 if (VT == MVT::i16)
10258 return std::make_pair(0U, X86::GR16RegisterClass);
10259 if (VT == MVT::i32 || !Subtarget->is64Bit())
10260 return std::make_pair(0U, X86::GR32RegisterClass);
10261 return std::make_pair(0U, X86::GR64RegisterClass);
10262 case 'R': // LEGACY_REGS
10264 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10265 if (VT == MVT::i16)
10266 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10267 if (VT == MVT::i32 || !Subtarget->is64Bit())
10268 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10269 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10270 case 'f': // FP Stack registers.
10271 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10272 // value to the correct fpstack register class.
10273 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10274 return std::make_pair(0U, X86::RFP32RegisterClass);
10275 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10276 return std::make_pair(0U, X86::RFP64RegisterClass);
10277 return std::make_pair(0U, X86::RFP80RegisterClass);
10278 case 'y': // MMX_REGS if MMX allowed.
10279 if (!Subtarget->hasMMX()) break;
10280 return std::make_pair(0U, X86::VR64RegisterClass);
10281 case 'Y': // SSE_REGS if SSE2 allowed
10282 if (!Subtarget->hasSSE2()) break;
10284 case 'x': // SSE_REGS if SSE1 allowed
10285 if (!Subtarget->hasSSE1()) break;
10287 switch (VT.getSimpleVT().SimpleTy) {
10289 // Scalar SSE types.
10292 return std::make_pair(0U, X86::FR32RegisterClass);
10295 return std::make_pair(0U, X86::FR64RegisterClass);
10303 return std::make_pair(0U, X86::VR128RegisterClass);
10309 // Use the default implementation in TargetLowering to convert the register
10310 // constraint into a member of a register class.
10311 std::pair<unsigned, const TargetRegisterClass*> Res;
10312 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10314 // Not found as a standard register?
10315 if (Res.second == 0) {
10316 // Map st(0) -> st(7) -> ST0
10317 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10318 tolower(Constraint[1]) == 's' &&
10319 tolower(Constraint[2]) == 't' &&
10320 Constraint[3] == '(' &&
10321 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10322 Constraint[5] == ')' &&
10323 Constraint[6] == '}') {
10325 Res.first = X86::ST0+Constraint[4]-'0';
10326 Res.second = X86::RFP80RegisterClass;
10330 // GCC allows "st(0)" to be called just plain "st".
10331 if (StringRef("{st}").equals_lower(Constraint)) {
10332 Res.first = X86::ST0;
10333 Res.second = X86::RFP80RegisterClass;
10338 if (StringRef("{flags}").equals_lower(Constraint)) {
10339 Res.first = X86::EFLAGS;
10340 Res.second = X86::CCRRegisterClass;
10344 // 'A' means EAX + EDX.
10345 if (Constraint == "A") {
10346 Res.first = X86::EAX;
10347 Res.second = X86::GR32_ADRegisterClass;
10353 // Otherwise, check to see if this is a register class of the wrong value
10354 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10355 // turn into {ax},{dx}.
10356 if (Res.second->hasType(VT))
10357 return Res; // Correct type already, nothing to do.
10359 // All of the single-register GCC register classes map their values onto
10360 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10361 // really want an 8-bit or 32-bit register, map to the appropriate register
10362 // class and return the appropriate register.
10363 if (Res.second == X86::GR16RegisterClass) {
10364 if (VT == MVT::i8) {
10365 unsigned DestReg = 0;
10366 switch (Res.first) {
10368 case X86::AX: DestReg = X86::AL; break;
10369 case X86::DX: DestReg = X86::DL; break;
10370 case X86::CX: DestReg = X86::CL; break;
10371 case X86::BX: DestReg = X86::BL; break;
10374 Res.first = DestReg;
10375 Res.second = X86::GR8RegisterClass;
10377 } else if (VT == MVT::i32) {
10378 unsigned DestReg = 0;
10379 switch (Res.first) {
10381 case X86::AX: DestReg = X86::EAX; break;
10382 case X86::DX: DestReg = X86::EDX; break;
10383 case X86::CX: DestReg = X86::ECX; break;
10384 case X86::BX: DestReg = X86::EBX; break;
10385 case X86::SI: DestReg = X86::ESI; break;
10386 case X86::DI: DestReg = X86::EDI; break;
10387 case X86::BP: DestReg = X86::EBP; break;
10388 case X86::SP: DestReg = X86::ESP; break;
10391 Res.first = DestReg;
10392 Res.second = X86::GR32RegisterClass;
10394 } else if (VT == MVT::i64) {
10395 unsigned DestReg = 0;
10396 switch (Res.first) {
10398 case X86::AX: DestReg = X86::RAX; break;
10399 case X86::DX: DestReg = X86::RDX; break;
10400 case X86::CX: DestReg = X86::RCX; break;
10401 case X86::BX: DestReg = X86::RBX; break;
10402 case X86::SI: DestReg = X86::RSI; break;
10403 case X86::DI: DestReg = X86::RDI; break;
10404 case X86::BP: DestReg = X86::RBP; break;
10405 case X86::SP: DestReg = X86::RSP; break;
10408 Res.first = DestReg;
10409 Res.second = X86::GR64RegisterClass;
10412 } else if (Res.second == X86::FR32RegisterClass ||
10413 Res.second == X86::FR64RegisterClass ||
10414 Res.second == X86::VR128RegisterClass) {
10415 // Handle references to XMM physical registers that got mapped into the
10416 // wrong class. This can happen with constraints like {xmm0} where the
10417 // target independent register mapper will just pick the first match it can
10418 // find, ignoring the required type.
10419 if (VT == MVT::f32)
10420 Res.second = X86::FR32RegisterClass;
10421 else if (VT == MVT::f64)
10422 Res.second = X86::FR64RegisterClass;
10423 else if (X86::VR128RegisterClass->hasType(VT))
10424 Res.second = X86::VR128RegisterClass;