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 Promote16Bit("promote-16bit", cl::Hidden,
69 cl::desc("Promote 16-bit instructions"));
71 // Forward declarations.
72 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
75 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
76 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
77 default: llvm_unreachable("unknown subtarget type");
78 case X86Subtarget::isDarwin:
79 if (TM.getSubtarget<X86Subtarget>().is64Bit())
80 return new X8664_MachoTargetObjectFile();
81 return new TargetLoweringObjectFileMachO();
82 case X86Subtarget::isELF:
83 if (TM.getSubtarget<X86Subtarget>().is64Bit())
84 return new X8664_ELFTargetObjectFile(TM);
85 return new X8632_ELFTargetObjectFile(TM);
86 case X86Subtarget::isMingw:
87 case X86Subtarget::isCygwin:
88 case X86Subtarget::isWindows:
89 return new TargetLoweringObjectFileCOFF();
93 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
94 : TargetLowering(TM, createTLOF(TM)) {
95 Subtarget = &TM.getSubtarget<X86Subtarget>();
96 X86ScalarSSEf64 = Subtarget->hasSSE2();
97 X86ScalarSSEf32 = Subtarget->hasSSE1();
98 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
100 RegInfo = TM.getRegisterInfo();
101 TD = getTargetData();
103 // Set up the TargetLowering object.
105 // X86 is weird, it always uses i8 for shift amounts and setcc results.
106 setShiftAmountType(MVT::i8);
107 setBooleanContents(ZeroOrOneBooleanContent);
108 setSchedulingPreference(SchedulingForRegPressure);
109 setStackPointerRegisterToSaveRestore(X86StackPtr);
111 if (Subtarget->isTargetDarwin()) {
112 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
113 setUseUnderscoreSetJmp(false);
114 setUseUnderscoreLongJmp(false);
115 } else if (Subtarget->isTargetMingw()) {
116 // MS runtime is weird: it exports _setjmp, but longjmp!
117 setUseUnderscoreSetJmp(true);
118 setUseUnderscoreLongJmp(false);
120 setUseUnderscoreSetJmp(true);
121 setUseUnderscoreLongJmp(true);
124 // Set up the register classes.
125 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
127 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
128 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
129 if (Subtarget->is64Bit())
130 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
132 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
134 // We don't accept any truncstore of integer registers.
135 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
137 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
138 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
140 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
141 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
142 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
144 // SETOEQ and SETUNE require checking two conditions.
145 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
146 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
147 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
148 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
149 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
150 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
152 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
154 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
155 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
156 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
158 if (Subtarget->is64Bit()) {
159 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
160 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
161 } else if (!UseSoftFloat) {
162 if (X86ScalarSSEf64) {
163 // We have an impenetrably clever algorithm for ui64->double only.
164 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
166 // We have an algorithm for SSE2, and we turn this into a 64-bit
167 // FILD for other targets.
168 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
173 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
177 // SSE has no i16 to fp conversion, only i32
178 if (X86ScalarSSEf32) {
179 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
180 // f32 and f64 cases are Legal, f80 case is not
181 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
183 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
187 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
191 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
192 // are Legal, f80 is custom lowered.
193 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
196 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
198 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
201 if (X86ScalarSSEf32) {
202 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
203 // f32 and f64 cases are Legal, f80 case is not
204 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
207 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
210 // Handle FP_TO_UINT by promoting the destination to a larger signed
212 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
213 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
214 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
216 if (Subtarget->is64Bit()) {
217 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
219 } else if (!UseSoftFloat) {
220 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
221 // Expand FP_TO_UINT into a select.
222 // FIXME: We would like to use a Custom expander here eventually to do
223 // the optimal thing for SSE vs. the default expansion in the legalizer.
224 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
226 // With SSE3 we can use fisttpll to convert to a signed i64; without
227 // SSE, we're stuck with a fistpll.
228 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
231 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
232 if (!X86ScalarSSEf64) {
233 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
234 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
237 // Scalar integer divide and remainder are lowered to use operations that
238 // produce two results, to match the available instructions. This exposes
239 // the two-result form to trivial CSE, which is able to combine x/y and x%y
240 // into a single instruction.
242 // Scalar integer multiply-high is also lowered to use two-result
243 // operations, to match the available instructions. However, plain multiply
244 // (low) operations are left as Legal, as there are single-result
245 // instructions for this in x86. Using the two-result multiply instructions
246 // when both high and low results are needed must be arranged by dagcombine.
247 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
248 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
249 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
250 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
251 setOperationAction(ISD::SREM , MVT::i8 , Expand);
252 setOperationAction(ISD::UREM , MVT::i8 , Expand);
253 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
254 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
255 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
256 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
257 setOperationAction(ISD::SREM , MVT::i16 , Expand);
258 setOperationAction(ISD::UREM , MVT::i16 , Expand);
259 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
260 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
261 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
262 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
263 setOperationAction(ISD::SREM , MVT::i32 , Expand);
264 setOperationAction(ISD::UREM , MVT::i32 , Expand);
265 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
266 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
267 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
268 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
269 setOperationAction(ISD::SREM , MVT::i64 , Expand);
270 setOperationAction(ISD::UREM , MVT::i64 , Expand);
272 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
273 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
274 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
275 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
276 if (Subtarget->is64Bit())
277 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
278 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
279 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
280 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
281 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
282 setOperationAction(ISD::FREM , MVT::f32 , Expand);
283 setOperationAction(ISD::FREM , MVT::f64 , Expand);
284 setOperationAction(ISD::FREM , MVT::f80 , Expand);
285 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
287 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
288 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
289 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
290 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
292 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
293 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
295 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
296 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
298 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
299 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
300 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
301 if (Subtarget->is64Bit()) {
302 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
303 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
304 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
307 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
308 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
310 // These should be promoted to a larger select which is supported.
311 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
312 // X86 wants to expand cmov itself.
313 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
315 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
317 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
318 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
319 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
320 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
321 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
322 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
324 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
326 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
328 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
329 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
330 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
331 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
333 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
335 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
338 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
339 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
340 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
341 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
342 if (Subtarget->is64Bit())
343 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
344 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
345 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
346 if (Subtarget->is64Bit()) {
347 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
348 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
349 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
350 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
351 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
353 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
354 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
355 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
356 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
357 if (Subtarget->is64Bit()) {
358 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
359 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
360 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
363 if (Subtarget->hasSSE1())
364 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
366 if (!Subtarget->hasSSE2())
367 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
369 // Expand certain atomics
370 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
371 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
372 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
373 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
380 if (!Subtarget->is64Bit()) {
381 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
383 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
384 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
385 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
386 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
387 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
390 // FIXME - use subtarget debug flags
391 if (!Subtarget->isTargetDarwin() &&
392 !Subtarget->isTargetELF() &&
393 !Subtarget->isTargetCygMing()) {
394 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
397 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
398 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
399 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
400 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
401 if (Subtarget->is64Bit()) {
402 setExceptionPointerRegister(X86::RAX);
403 setExceptionSelectorRegister(X86::RDX);
405 setExceptionPointerRegister(X86::EAX);
406 setExceptionSelectorRegister(X86::EDX);
408 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
409 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
411 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
413 setOperationAction(ISD::TRAP, MVT::Other, Legal);
415 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
416 setOperationAction(ISD::VASTART , MVT::Other, Custom);
417 setOperationAction(ISD::VAEND , MVT::Other, Expand);
418 if (Subtarget->is64Bit()) {
419 setOperationAction(ISD::VAARG , MVT::Other, Custom);
420 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
422 setOperationAction(ISD::VAARG , MVT::Other, Expand);
423 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
426 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
427 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
430 if (Subtarget->isTargetCygMing())
431 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
433 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
435 if (!UseSoftFloat && X86ScalarSSEf64) {
436 // f32 and f64 use SSE.
437 // Set up the FP register classes.
438 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
439 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
441 // Use ANDPD to simulate FABS.
442 setOperationAction(ISD::FABS , MVT::f64, Custom);
443 setOperationAction(ISD::FABS , MVT::f32, Custom);
445 // Use XORP to simulate FNEG.
446 setOperationAction(ISD::FNEG , MVT::f64, Custom);
447 setOperationAction(ISD::FNEG , MVT::f32, Custom);
449 // Use ANDPD and ORPD to simulate FCOPYSIGN.
450 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
451 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
453 // We don't support sin/cos/fmod
454 setOperationAction(ISD::FSIN , MVT::f64, Expand);
455 setOperationAction(ISD::FCOS , MVT::f64, Expand);
456 setOperationAction(ISD::FSIN , MVT::f32, Expand);
457 setOperationAction(ISD::FCOS , MVT::f32, Expand);
459 // Expand FP immediates into loads from the stack, except for the special
461 addLegalFPImmediate(APFloat(+0.0)); // xorpd
462 addLegalFPImmediate(APFloat(+0.0f)); // xorps
463 } else if (!UseSoftFloat && X86ScalarSSEf32) {
464 // Use SSE for f32, x87 for f64.
465 // Set up the FP register classes.
466 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
467 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
469 // Use ANDPS to simulate FABS.
470 setOperationAction(ISD::FABS , MVT::f32, Custom);
472 // Use XORP to simulate FNEG.
473 setOperationAction(ISD::FNEG , MVT::f32, Custom);
475 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
477 // Use ANDPS and ORPS to simulate FCOPYSIGN.
478 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
479 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
481 // We don't support sin/cos/fmod
482 setOperationAction(ISD::FSIN , MVT::f32, Expand);
483 setOperationAction(ISD::FCOS , MVT::f32, Expand);
485 // Special cases we handle for FP constants.
486 addLegalFPImmediate(APFloat(+0.0f)); // xorps
487 addLegalFPImmediate(APFloat(+0.0)); // FLD0
488 addLegalFPImmediate(APFloat(+1.0)); // FLD1
489 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
490 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
493 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
494 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
496 } else if (!UseSoftFloat) {
497 // f32 and f64 in x87.
498 // Set up the FP register classes.
499 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
500 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
502 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
503 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
504 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
505 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
508 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
509 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
511 addLegalFPImmediate(APFloat(+0.0)); // FLD0
512 addLegalFPImmediate(APFloat(+1.0)); // FLD1
513 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
514 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
515 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
516 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
517 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
518 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
521 // Long double always uses X87.
523 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
524 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
525 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
528 APFloat TmpFlt(+0.0);
529 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
531 addLegalFPImmediate(TmpFlt); // FLD0
533 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
534 APFloat TmpFlt2(+1.0);
535 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
537 addLegalFPImmediate(TmpFlt2); // FLD1
538 TmpFlt2.changeSign();
539 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
543 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
544 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
548 // Always use a library call for pow.
549 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
550 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
551 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
553 setOperationAction(ISD::FLOG, MVT::f80, Expand);
554 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
555 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
556 setOperationAction(ISD::FEXP, MVT::f80, Expand);
557 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
559 // First set operation action for all vector types to either promote
560 // (for widening) or expand (for scalarization). Then we will selectively
561 // turn on ones that can be effectively codegen'd.
562 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
563 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
564 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
579 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
580 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
612 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
613 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
614 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
615 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
616 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
617 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
618 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
619 setTruncStoreAction((MVT::SimpleValueType)VT,
620 (MVT::SimpleValueType)InnerVT, Expand);
621 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
622 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
623 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
626 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
627 // with -msoft-float, disable use of MMX as well.
628 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
629 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
630 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
631 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
632 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
633 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
635 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
636 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
637 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
638 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
640 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
641 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
642 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
643 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
645 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
646 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
648 setOperationAction(ISD::AND, MVT::v8i8, Promote);
649 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
650 setOperationAction(ISD::AND, MVT::v4i16, Promote);
651 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
652 setOperationAction(ISD::AND, MVT::v2i32, Promote);
653 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
654 setOperationAction(ISD::AND, MVT::v1i64, Legal);
656 setOperationAction(ISD::OR, MVT::v8i8, Promote);
657 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
658 setOperationAction(ISD::OR, MVT::v4i16, Promote);
659 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
660 setOperationAction(ISD::OR, MVT::v2i32, Promote);
661 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
662 setOperationAction(ISD::OR, MVT::v1i64, Legal);
664 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
665 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
666 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
667 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
668 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
669 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
670 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
672 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
674 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
675 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
676 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
677 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
678 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
679 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
680 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
682 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
684 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
685 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
686 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
688 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
689 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
690 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
691 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
693 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
694 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
695 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
696 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
698 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
700 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
701 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
702 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
703 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
704 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
705 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
706 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
709 if (!UseSoftFloat && Subtarget->hasSSE1()) {
710 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
712 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
713 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
714 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
715 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
716 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
717 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
718 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
719 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
720 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
721 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
722 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
723 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
726 if (!UseSoftFloat && Subtarget->hasSSE2()) {
727 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
729 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
730 // registers cannot be used even for integer operations.
731 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
732 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
733 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
734 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
736 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
737 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
738 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
739 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
740 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
741 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
742 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
743 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
744 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
745 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
746 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
747 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
748 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
749 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
750 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
751 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
753 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
754 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
755 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
756 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
758 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
759 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
760 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
761 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
762 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
764 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
765 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
766 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
767 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
768 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
770 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
771 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
772 EVT VT = (MVT::SimpleValueType)i;
773 // Do not attempt to custom lower non-power-of-2 vectors
774 if (!isPowerOf2_32(VT.getVectorNumElements()))
776 // Do not attempt to custom lower non-128-bit vectors
777 if (!VT.is128BitVector())
779 setOperationAction(ISD::BUILD_VECTOR,
780 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::VECTOR_SHUFFLE,
782 VT.getSimpleVT().SimpleTy, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
784 VT.getSimpleVT().SimpleTy, Custom);
787 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
788 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
789 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
790 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
794 if (Subtarget->is64Bit()) {
795 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
799 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor 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
1074 /// lowering. If DstAlign is zero that means it's safe to destination
1075 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1076 /// means there isn't a need to check it against alignment requirement,
1077 /// probably because the source does not need to be loaded. If
1078 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1079 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1080 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1081 /// constant so it does not need to be loaded.
1082 /// It returns EVT::Other if the type should be determined using generic
1083 /// target-independent logic.
1085 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1086 unsigned DstAlign, unsigned SrcAlign,
1087 bool NonScalarIntSafe,
1089 MachineFunction &MF) const {
1090 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1091 // linux. This is because the stack realignment code can't handle certain
1092 // cases like PR2962. This should be removed when PR2962 is fixed.
1093 const Function *F = MF.getFunction();
1094 if (NonScalarIntSafe &&
1095 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1097 (Subtarget->isUnalignedMemAccessFast() ||
1098 ((DstAlign == 0 || DstAlign >= 16) &&
1099 (SrcAlign == 0 || SrcAlign >= 16))) &&
1100 Subtarget->getStackAlignment() >= 16) {
1101 if (Subtarget->hasSSE2())
1103 if (Subtarget->hasSSE1())
1105 } else if (!MemcpyStrSrc && Size >= 8 &&
1106 !Subtarget->is64Bit() &&
1107 Subtarget->getStackAlignment() >= 8 &&
1108 Subtarget->hasSSE2()) {
1109 // Do not use f64 to lower memcpy if source is string constant. It's
1110 // better to use i32 to avoid the loads.
1114 if (Subtarget->is64Bit() && Size >= 8)
1119 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1120 /// current function. The returned value is a member of the
1121 /// MachineJumpTableInfo::JTEntryKind enum.
1122 unsigned X86TargetLowering::getJumpTableEncoding() const {
1123 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1125 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1126 Subtarget->isPICStyleGOT())
1127 return MachineJumpTableInfo::EK_Custom32;
1129 // Otherwise, use the normal jump table encoding heuristics.
1130 return TargetLowering::getJumpTableEncoding();
1133 /// getPICBaseSymbol - Return the X86-32 PIC base.
1135 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1136 MCContext &Ctx) const {
1137 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1138 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1139 Twine(MF->getFunctionNumber())+"$pb");
1144 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1145 const MachineBasicBlock *MBB,
1146 unsigned uid,MCContext &Ctx) const{
1147 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1148 Subtarget->isPICStyleGOT());
1149 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1151 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1152 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1155 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1157 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1158 SelectionDAG &DAG) const {
1159 if (!Subtarget->is64Bit())
1160 // This doesn't have DebugLoc associated with it, but is not really the
1161 // same as a Register.
1162 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1166 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1167 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1169 const MCExpr *X86TargetLowering::
1170 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1171 MCContext &Ctx) const {
1172 // X86-64 uses RIP relative addressing based on the jump table label.
1173 if (Subtarget->isPICStyleRIPRel())
1174 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1176 // Otherwise, the reference is relative to the PIC base.
1177 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1180 /// getFunctionAlignment - Return the Log2 alignment of this function.
1181 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1182 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1185 //===----------------------------------------------------------------------===//
1186 // Return Value Calling Convention Implementation
1187 //===----------------------------------------------------------------------===//
1189 #include "X86GenCallingConv.inc"
1192 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1193 const SmallVectorImpl<EVT> &OutTys,
1194 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1195 SelectionDAG &DAG) const {
1196 SmallVector<CCValAssign, 16> RVLocs;
1197 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1198 RVLocs, *DAG.getContext());
1199 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1203 X86TargetLowering::LowerReturn(SDValue Chain,
1204 CallingConv::ID CallConv, bool isVarArg,
1205 const SmallVectorImpl<ISD::OutputArg> &Outs,
1206 DebugLoc dl, SelectionDAG &DAG) const {
1207 MachineFunction &MF = DAG.getMachineFunction();
1208 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1210 SmallVector<CCValAssign, 16> RVLocs;
1211 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1212 RVLocs, *DAG.getContext());
1213 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1215 // Add the regs to the liveout set for the function.
1216 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1217 for (unsigned i = 0; i != RVLocs.size(); ++i)
1218 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1219 MRI.addLiveOut(RVLocs[i].getLocReg());
1223 SmallVector<SDValue, 6> RetOps;
1224 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1225 // Operand #1 = Bytes To Pop
1226 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1229 // Copy the result values into the output registers.
1230 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1231 CCValAssign &VA = RVLocs[i];
1232 assert(VA.isRegLoc() && "Can only return in registers!");
1233 SDValue ValToCopy = Outs[i].Val;
1235 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1236 // the RET instruction and handled by the FP Stackifier.
1237 if (VA.getLocReg() == X86::ST0 ||
1238 VA.getLocReg() == X86::ST1) {
1239 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1240 // change the value to the FP stack register class.
1241 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1242 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1243 RetOps.push_back(ValToCopy);
1244 // Don't emit a copytoreg.
1248 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1249 // which is returned in RAX / RDX.
1250 if (Subtarget->is64Bit()) {
1251 EVT ValVT = ValToCopy.getValueType();
1252 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1253 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1254 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1255 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1259 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1260 Flag = Chain.getValue(1);
1263 // The x86-64 ABI for returning structs by value requires that we copy
1264 // the sret argument into %rax for the return. We saved the argument into
1265 // a virtual register in the entry block, so now we copy the value out
1267 if (Subtarget->is64Bit() &&
1268 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1269 MachineFunction &MF = DAG.getMachineFunction();
1270 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1271 unsigned Reg = FuncInfo->getSRetReturnReg();
1273 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1274 FuncInfo->setSRetReturnReg(Reg);
1276 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1278 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1279 Flag = Chain.getValue(1);
1281 // RAX now acts like a return value.
1282 MRI.addLiveOut(X86::RAX);
1285 RetOps[0] = Chain; // Update chain.
1287 // Add the flag if we have it.
1289 RetOps.push_back(Flag);
1291 return DAG.getNode(X86ISD::RET_FLAG, dl,
1292 MVT::Other, &RetOps[0], RetOps.size());
1295 /// LowerCallResult - Lower the result values of a call into the
1296 /// appropriate copies out of appropriate physical registers.
1299 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1300 CallingConv::ID CallConv, bool isVarArg,
1301 const SmallVectorImpl<ISD::InputArg> &Ins,
1302 DebugLoc dl, SelectionDAG &DAG,
1303 SmallVectorImpl<SDValue> &InVals) const {
1305 // Assign locations to each value returned by this call.
1306 SmallVector<CCValAssign, 16> RVLocs;
1307 bool Is64Bit = Subtarget->is64Bit();
1308 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1309 RVLocs, *DAG.getContext());
1310 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1312 // Copy all of the result registers out of their specified physreg.
1313 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1314 CCValAssign &VA = RVLocs[i];
1315 EVT CopyVT = VA.getValVT();
1317 // If this is x86-64, and we disabled SSE, we can't return FP values
1318 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1319 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1320 report_fatal_error("SSE register return with SSE disabled");
1323 // If this is a call to a function that returns an fp value on the floating
1324 // point stack, but where we prefer to use the value in xmm registers, copy
1325 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1326 if ((VA.getLocReg() == X86::ST0 ||
1327 VA.getLocReg() == X86::ST1) &&
1328 isScalarFPTypeInSSEReg(VA.getValVT())) {
1333 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1334 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1335 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1336 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1337 MVT::v2i64, InFlag).getValue(1);
1338 Val = Chain.getValue(0);
1339 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1340 Val, DAG.getConstant(0, MVT::i64));
1342 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1343 MVT::i64, InFlag).getValue(1);
1344 Val = Chain.getValue(0);
1346 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1348 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1349 CopyVT, InFlag).getValue(1);
1350 Val = Chain.getValue(0);
1352 InFlag = Chain.getValue(2);
1354 if (CopyVT != VA.getValVT()) {
1355 // Round the F80 the right size, which also moves to the appropriate xmm
1357 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1358 // This truncation won't change the value.
1359 DAG.getIntPtrConstant(1));
1362 InVals.push_back(Val);
1369 //===----------------------------------------------------------------------===//
1370 // C & StdCall & Fast Calling Convention implementation
1371 //===----------------------------------------------------------------------===//
1372 // StdCall calling convention seems to be standard for many Windows' API
1373 // routines and around. It differs from C calling convention just a little:
1374 // callee should clean up the stack, not caller. Symbols should be also
1375 // decorated in some fancy way :) It doesn't support any vector arguments.
1376 // For info on fast calling convention see Fast Calling Convention (tail call)
1377 // implementation LowerX86_32FastCCCallTo.
1379 /// CallIsStructReturn - Determines whether a call uses struct return
1381 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1385 return Outs[0].Flags.isSRet();
1388 /// ArgsAreStructReturn - Determines whether a function uses struct
1389 /// return semantics.
1391 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1395 return Ins[0].Flags.isSRet();
1398 /// IsCalleePop - Determines whether the callee is required to pop its
1399 /// own arguments. Callee pop is necessary to support tail calls.
1400 bool X86TargetLowering::IsCalleePop(bool IsVarArg,
1401 CallingConv::ID CallingConv) const {
1405 switch (CallingConv) {
1408 case CallingConv::X86_StdCall:
1409 return !Subtarget->is64Bit();
1410 case CallingConv::X86_FastCall:
1411 return !Subtarget->is64Bit();
1412 case CallingConv::Fast:
1413 return GuaranteedTailCallOpt;
1414 case CallingConv::GHC:
1415 return GuaranteedTailCallOpt;
1419 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1420 /// given CallingConvention value.
1421 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1422 if (Subtarget->is64Bit()) {
1423 if (CC == CallingConv::GHC)
1424 return CC_X86_64_GHC;
1425 else if (Subtarget->isTargetWin64())
1426 return CC_X86_Win64_C;
1431 if (CC == CallingConv::X86_FastCall)
1432 return CC_X86_32_FastCall;
1433 else if (CC == CallingConv::Fast)
1434 return CC_X86_32_FastCC;
1435 else if (CC == CallingConv::GHC)
1436 return CC_X86_32_GHC;
1441 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1442 /// by "Src" to address "Dst" with size and alignment information specified by
1443 /// the specific parameter attribute. The copy will be passed as a byval
1444 /// function parameter.
1446 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1447 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1449 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1450 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1451 /*isVolatile*/false, /*AlwaysInline=*/true,
1455 /// IsTailCallConvention - Return true if the calling convention is one that
1456 /// supports tail call optimization.
1457 static bool IsTailCallConvention(CallingConv::ID CC) {
1458 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1461 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1462 /// a tailcall target by changing its ABI.
1463 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1464 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1468 X86TargetLowering::LowerMemArgument(SDValue Chain,
1469 CallingConv::ID CallConv,
1470 const SmallVectorImpl<ISD::InputArg> &Ins,
1471 DebugLoc dl, SelectionDAG &DAG,
1472 const CCValAssign &VA,
1473 MachineFrameInfo *MFI,
1475 // Create the nodes corresponding to a load from this parameter slot.
1476 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1477 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1478 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1481 // If value is passed by pointer we have address passed instead of the value
1483 if (VA.getLocInfo() == CCValAssign::Indirect)
1484 ValVT = VA.getLocVT();
1486 ValVT = VA.getValVT();
1488 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1489 // changed with more analysis.
1490 // In case of tail call optimization mark all arguments mutable. Since they
1491 // could be overwritten by lowering of arguments in case of a tail call.
1492 if (Flags.isByVal()) {
1493 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1494 VA.getLocMemOffset(), isImmutable, false);
1495 return DAG.getFrameIndex(FI, getPointerTy());
1497 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1498 VA.getLocMemOffset(), isImmutable, false);
1499 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1500 return DAG.getLoad(ValVT, dl, Chain, FIN,
1501 PseudoSourceValue::getFixedStack(FI), 0,
1507 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1508 CallingConv::ID CallConv,
1510 const SmallVectorImpl<ISD::InputArg> &Ins,
1513 SmallVectorImpl<SDValue> &InVals)
1515 MachineFunction &MF = DAG.getMachineFunction();
1516 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1518 const Function* Fn = MF.getFunction();
1519 if (Fn->hasExternalLinkage() &&
1520 Subtarget->isTargetCygMing() &&
1521 Fn->getName() == "main")
1522 FuncInfo->setForceFramePointer(true);
1524 MachineFrameInfo *MFI = MF.getFrameInfo();
1525 bool Is64Bit = Subtarget->is64Bit();
1526 bool IsWin64 = Subtarget->isTargetWin64();
1528 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1529 "Var args not supported with calling convention fastcc or ghc");
1531 // Assign locations to all of the incoming arguments.
1532 SmallVector<CCValAssign, 16> ArgLocs;
1533 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1534 ArgLocs, *DAG.getContext());
1535 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1537 unsigned LastVal = ~0U;
1539 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1540 CCValAssign &VA = ArgLocs[i];
1541 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1543 assert(VA.getValNo() != LastVal &&
1544 "Don't support value assigned to multiple locs yet");
1545 LastVal = VA.getValNo();
1547 if (VA.isRegLoc()) {
1548 EVT RegVT = VA.getLocVT();
1549 TargetRegisterClass *RC = NULL;
1550 if (RegVT == MVT::i32)
1551 RC = X86::GR32RegisterClass;
1552 else if (Is64Bit && RegVT == MVT::i64)
1553 RC = X86::GR64RegisterClass;
1554 else if (RegVT == MVT::f32)
1555 RC = X86::FR32RegisterClass;
1556 else if (RegVT == MVT::f64)
1557 RC = X86::FR64RegisterClass;
1558 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1559 RC = X86::VR128RegisterClass;
1560 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1561 RC = X86::VR64RegisterClass;
1563 llvm_unreachable("Unknown argument type!");
1565 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1566 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1568 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1569 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1571 if (VA.getLocInfo() == CCValAssign::SExt)
1572 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1573 DAG.getValueType(VA.getValVT()));
1574 else if (VA.getLocInfo() == CCValAssign::ZExt)
1575 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1576 DAG.getValueType(VA.getValVT()));
1577 else if (VA.getLocInfo() == CCValAssign::BCvt)
1578 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1580 if (VA.isExtInLoc()) {
1581 // Handle MMX values passed in XMM regs.
1582 if (RegVT.isVector()) {
1583 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1584 ArgValue, DAG.getConstant(0, MVT::i64));
1585 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1587 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1590 assert(VA.isMemLoc());
1591 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1594 // If value is passed via pointer - do a load.
1595 if (VA.getLocInfo() == CCValAssign::Indirect)
1596 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1599 InVals.push_back(ArgValue);
1602 // The x86-64 ABI for returning structs by value requires that we copy
1603 // the sret argument into %rax for the return. Save the argument into
1604 // a virtual register so that we can access it from the return points.
1605 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1606 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1607 unsigned Reg = FuncInfo->getSRetReturnReg();
1609 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1610 FuncInfo->setSRetReturnReg(Reg);
1612 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1613 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1616 unsigned StackSize = CCInfo.getNextStackOffset();
1617 // Align stack specially for tail calls.
1618 if (FuncIsMadeTailCallSafe(CallConv))
1619 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1621 // If the function takes variable number of arguments, make a frame index for
1622 // the start of the first vararg value... for expansion of llvm.va_start.
1624 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1625 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
1629 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1631 // FIXME: We should really autogenerate these arrays
1632 static const unsigned GPR64ArgRegsWin64[] = {
1633 X86::RCX, X86::RDX, X86::R8, X86::R9
1635 static const unsigned XMMArgRegsWin64[] = {
1636 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1638 static const unsigned GPR64ArgRegs64Bit[] = {
1639 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1641 static const unsigned XMMArgRegs64Bit[] = {
1642 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1643 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1645 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1648 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1649 GPR64ArgRegs = GPR64ArgRegsWin64;
1650 XMMArgRegs = XMMArgRegsWin64;
1652 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1653 GPR64ArgRegs = GPR64ArgRegs64Bit;
1654 XMMArgRegs = XMMArgRegs64Bit;
1656 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1658 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1661 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1662 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1663 "SSE register cannot be used when SSE is disabled!");
1664 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1665 "SSE register cannot be used when SSE is disabled!");
1666 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1667 // Kernel mode asks for SSE to be disabled, so don't push them
1669 TotalNumXMMRegs = 0;
1671 // For X86-64, if there are vararg parameters that are passed via
1672 // registers, then we must store them to their spots on the stack so they
1673 // may be loaded by deferencing the result of va_next.
1674 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1675 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1676 FuncInfo->setRegSaveFrameIndex(
1677 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1680 // Store the integer parameter registers.
1681 SmallVector<SDValue, 8> MemOps;
1682 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1684 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1685 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1686 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1687 DAG.getIntPtrConstant(Offset));
1688 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1689 X86::GR64RegisterClass);
1690 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1692 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1693 PseudoSourceValue::getFixedStack(
1694 FuncInfo->getRegSaveFrameIndex()),
1695 Offset, false, false, 0);
1696 MemOps.push_back(Store);
1700 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1701 // Now store the XMM (fp + vector) parameter registers.
1702 SmallVector<SDValue, 11> SaveXMMOps;
1703 SaveXMMOps.push_back(Chain);
1705 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1706 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1707 SaveXMMOps.push_back(ALVal);
1709 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1710 FuncInfo->getRegSaveFrameIndex()));
1711 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1712 FuncInfo->getVarArgsFPOffset()));
1714 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1715 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1716 X86::VR128RegisterClass);
1717 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1718 SaveXMMOps.push_back(Val);
1720 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1722 &SaveXMMOps[0], SaveXMMOps.size()));
1725 if (!MemOps.empty())
1726 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1727 &MemOps[0], MemOps.size());
1731 // Some CCs need callee pop.
1732 if (IsCalleePop(isVarArg, CallConv)) {
1733 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1735 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1736 // If this is an sret function, the return should pop the hidden pointer.
1737 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1738 FuncInfo->setBytesToPopOnReturn(4);
1742 // RegSaveFrameIndex is X86-64 only.
1743 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1744 if (CallConv == CallingConv::X86_FastCall)
1745 // fastcc functions can't have varargs.
1746 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1753 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1754 SDValue StackPtr, SDValue Arg,
1755 DebugLoc dl, SelectionDAG &DAG,
1756 const CCValAssign &VA,
1757 ISD::ArgFlagsTy Flags) const {
1758 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1759 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1760 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1761 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1762 if (Flags.isByVal()) {
1763 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1765 return DAG.getStore(Chain, dl, Arg, PtrOff,
1766 PseudoSourceValue::getStack(), LocMemOffset,
1770 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1771 /// optimization is performed and it is required.
1773 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1774 SDValue &OutRetAddr, SDValue Chain,
1775 bool IsTailCall, bool Is64Bit,
1776 int FPDiff, DebugLoc dl) const {
1777 // Adjust the Return address stack slot.
1778 EVT VT = getPointerTy();
1779 OutRetAddr = getReturnAddressFrameIndex(DAG);
1781 // Load the "old" Return address.
1782 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1783 return SDValue(OutRetAddr.getNode(), 1);
1786 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1787 /// optimization is performed and it is required (FPDiff!=0).
1789 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1790 SDValue Chain, SDValue RetAddrFrIdx,
1791 bool Is64Bit, int FPDiff, DebugLoc dl) {
1792 // Store the return address to the appropriate stack slot.
1793 if (!FPDiff) return Chain;
1794 // Calculate the new stack slot for the return address.
1795 int SlotSize = Is64Bit ? 8 : 4;
1796 int NewReturnAddrFI =
1797 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1798 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1799 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1800 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1801 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1807 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1808 CallingConv::ID CallConv, bool isVarArg,
1810 const SmallVectorImpl<ISD::OutputArg> &Outs,
1811 const SmallVectorImpl<ISD::InputArg> &Ins,
1812 DebugLoc dl, SelectionDAG &DAG,
1813 SmallVectorImpl<SDValue> &InVals) const {
1814 MachineFunction &MF = DAG.getMachineFunction();
1815 bool Is64Bit = Subtarget->is64Bit();
1816 bool IsStructRet = CallIsStructReturn(Outs);
1817 bool IsSibcall = false;
1820 // Check if it's really possible to do a tail call.
1821 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1822 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1825 // Sibcalls are automatically detected tailcalls which do not require
1827 if (!GuaranteedTailCallOpt && isTailCall)
1834 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1835 "Var args not supported with calling convention fastcc or ghc");
1837 // Analyze operands of the call, assigning locations to each operand.
1838 SmallVector<CCValAssign, 16> ArgLocs;
1839 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1840 ArgLocs, *DAG.getContext());
1841 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1843 // Get a count of how many bytes are to be pushed on the stack.
1844 unsigned NumBytes = CCInfo.getNextStackOffset();
1846 // This is a sibcall. The memory operands are available in caller's
1847 // own caller's stack.
1849 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1850 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1853 if (isTailCall && !IsSibcall) {
1854 // Lower arguments at fp - stackoffset + fpdiff.
1855 unsigned NumBytesCallerPushed =
1856 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1857 FPDiff = NumBytesCallerPushed - NumBytes;
1859 // Set the delta of movement of the returnaddr stackslot.
1860 // But only set if delta is greater than previous delta.
1861 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1862 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1866 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1868 SDValue RetAddrFrIdx;
1869 // Load return adress for tail calls.
1870 if (isTailCall && FPDiff)
1871 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1872 Is64Bit, FPDiff, dl);
1874 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1875 SmallVector<SDValue, 8> MemOpChains;
1878 // Walk the register/memloc assignments, inserting copies/loads. In the case
1879 // of tail call optimization arguments are handle later.
1880 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1881 CCValAssign &VA = ArgLocs[i];
1882 EVT RegVT = VA.getLocVT();
1883 SDValue Arg = Outs[i].Val;
1884 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1885 bool isByVal = Flags.isByVal();
1887 // Promote the value if needed.
1888 switch (VA.getLocInfo()) {
1889 default: llvm_unreachable("Unknown loc info!");
1890 case CCValAssign::Full: break;
1891 case CCValAssign::SExt:
1892 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1894 case CCValAssign::ZExt:
1895 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1897 case CCValAssign::AExt:
1898 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1899 // Special case: passing MMX values in XMM registers.
1900 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1901 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1902 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1904 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1906 case CCValAssign::BCvt:
1907 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1909 case CCValAssign::Indirect: {
1910 // Store the argument.
1911 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1912 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1913 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1914 PseudoSourceValue::getFixedStack(FI), 0,
1921 if (VA.isRegLoc()) {
1922 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1923 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1924 assert(VA.isMemLoc());
1925 if (StackPtr.getNode() == 0)
1926 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1927 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1928 dl, DAG, VA, Flags));
1932 if (!MemOpChains.empty())
1933 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1934 &MemOpChains[0], MemOpChains.size());
1936 // Build a sequence of copy-to-reg nodes chained together with token chain
1937 // and flag operands which copy the outgoing args into registers.
1939 // Tail call byval lowering might overwrite argument registers so in case of
1940 // tail call optimization the copies to registers are lowered later.
1942 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1943 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1944 RegsToPass[i].second, InFlag);
1945 InFlag = Chain.getValue(1);
1948 if (Subtarget->isPICStyleGOT()) {
1949 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1952 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1953 DAG.getNode(X86ISD::GlobalBaseReg,
1954 DebugLoc(), getPointerTy()),
1956 InFlag = Chain.getValue(1);
1958 // If we are tail calling and generating PIC/GOT style code load the
1959 // address of the callee into ECX. The value in ecx is used as target of
1960 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1961 // for tail calls on PIC/GOT architectures. Normally we would just put the
1962 // address of GOT into ebx and then call target@PLT. But for tail calls
1963 // ebx would be restored (since ebx is callee saved) before jumping to the
1966 // Note: The actual moving to ECX is done further down.
1967 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1968 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1969 !G->getGlobal()->hasProtectedVisibility())
1970 Callee = LowerGlobalAddress(Callee, DAG);
1971 else if (isa<ExternalSymbolSDNode>(Callee))
1972 Callee = LowerExternalSymbol(Callee, DAG);
1976 if (Is64Bit && isVarArg) {
1977 // From AMD64 ABI document:
1978 // For calls that may call functions that use varargs or stdargs
1979 // (prototype-less calls or calls to functions containing ellipsis (...) in
1980 // the declaration) %al is used as hidden argument to specify the number
1981 // of SSE registers used. The contents of %al do not need to match exactly
1982 // the number of registers, but must be an ubound on the number of SSE
1983 // registers used and is in the range 0 - 8 inclusive.
1985 // FIXME: Verify this on Win64
1986 // Count the number of XMM registers allocated.
1987 static const unsigned XMMArgRegs[] = {
1988 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1989 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1991 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1992 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1993 && "SSE registers cannot be used when SSE is disabled");
1995 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1996 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1997 InFlag = Chain.getValue(1);
2001 // For tail calls lower the arguments to the 'real' stack slot.
2003 // Force all the incoming stack arguments to be loaded from the stack
2004 // before any new outgoing arguments are stored to the stack, because the
2005 // outgoing stack slots may alias the incoming argument stack slots, and
2006 // the alias isn't otherwise explicit. This is slightly more conservative
2007 // than necessary, because it means that each store effectively depends
2008 // on every argument instead of just those arguments it would clobber.
2009 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2011 SmallVector<SDValue, 8> MemOpChains2;
2014 // Do not flag preceeding copytoreg stuff together with the following stuff.
2016 if (GuaranteedTailCallOpt) {
2017 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2018 CCValAssign &VA = ArgLocs[i];
2021 assert(VA.isMemLoc());
2022 SDValue Arg = Outs[i].Val;
2023 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2024 // Create frame index.
2025 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2026 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2027 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2028 FIN = DAG.getFrameIndex(FI, getPointerTy());
2030 if (Flags.isByVal()) {
2031 // Copy relative to framepointer.
2032 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2033 if (StackPtr.getNode() == 0)
2034 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2036 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2038 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2042 // Store relative to framepointer.
2043 MemOpChains2.push_back(
2044 DAG.getStore(ArgChain, dl, Arg, FIN,
2045 PseudoSourceValue::getFixedStack(FI), 0,
2051 if (!MemOpChains2.empty())
2052 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2053 &MemOpChains2[0], MemOpChains2.size());
2055 // Copy arguments to their registers.
2056 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2057 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2058 RegsToPass[i].second, InFlag);
2059 InFlag = Chain.getValue(1);
2063 // Store the return address to the appropriate stack slot.
2064 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2068 bool WasGlobalOrExternal = false;
2069 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2070 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2071 // In the 64-bit large code model, we have to make all calls
2072 // through a register, since the call instruction's 32-bit
2073 // pc-relative offset may not be large enough to hold the whole
2075 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2076 WasGlobalOrExternal = true;
2077 // If the callee is a GlobalAddress node (quite common, every direct call
2078 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2081 // We should use extra load for direct calls to dllimported functions in
2083 const GlobalValue *GV = G->getGlobal();
2084 if (!GV->hasDLLImportLinkage()) {
2085 unsigned char OpFlags = 0;
2087 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2088 // external symbols most go through the PLT in PIC mode. If the symbol
2089 // has hidden or protected visibility, or if it is static or local, then
2090 // we don't need to use the PLT - we can directly call it.
2091 if (Subtarget->isTargetELF() &&
2092 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2093 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2094 OpFlags = X86II::MO_PLT;
2095 } else if (Subtarget->isPICStyleStubAny() &&
2096 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2097 Subtarget->getDarwinVers() < 9) {
2098 // PC-relative references to external symbols should go through $stub,
2099 // unless we're building with the leopard linker or later, which
2100 // automatically synthesizes these stubs.
2101 OpFlags = X86II::MO_DARWIN_STUB;
2104 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2105 G->getOffset(), OpFlags);
2107 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2108 WasGlobalOrExternal = true;
2109 unsigned char OpFlags = 0;
2111 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2112 // symbols should go through the PLT.
2113 if (Subtarget->isTargetELF() &&
2114 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2115 OpFlags = X86II::MO_PLT;
2116 } else if (Subtarget->isPICStyleStubAny() &&
2117 Subtarget->getDarwinVers() < 9) {
2118 // PC-relative references to external symbols should go through $stub,
2119 // unless we're building with the leopard linker or later, which
2120 // automatically synthesizes these stubs.
2121 OpFlags = X86II::MO_DARWIN_STUB;
2124 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2128 // Returns a chain & a flag for retval copy to use.
2129 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2130 SmallVector<SDValue, 8> Ops;
2132 if (!IsSibcall && isTailCall) {
2133 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2134 DAG.getIntPtrConstant(0, true), InFlag);
2135 InFlag = Chain.getValue(1);
2138 Ops.push_back(Chain);
2139 Ops.push_back(Callee);
2142 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2144 // Add argument registers to the end of the list so that they are known live
2146 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2147 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2148 RegsToPass[i].second.getValueType()));
2150 // Add an implicit use GOT pointer in EBX.
2151 if (!isTailCall && Subtarget->isPICStyleGOT())
2152 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2154 // Add an implicit use of AL for x86 vararg functions.
2155 if (Is64Bit && isVarArg)
2156 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2158 if (InFlag.getNode())
2159 Ops.push_back(InFlag);
2162 // If this is the first return lowered for this function, add the regs
2163 // to the liveout set for the function.
2164 if (MF.getRegInfo().liveout_empty()) {
2165 SmallVector<CCValAssign, 16> RVLocs;
2166 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2168 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2169 for (unsigned i = 0; i != RVLocs.size(); ++i)
2170 if (RVLocs[i].isRegLoc())
2171 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2173 return DAG.getNode(X86ISD::TC_RETURN, dl,
2174 NodeTys, &Ops[0], Ops.size());
2177 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2178 InFlag = Chain.getValue(1);
2180 // Create the CALLSEQ_END node.
2181 unsigned NumBytesForCalleeToPush;
2182 if (IsCalleePop(isVarArg, CallConv))
2183 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2184 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2185 // If this is a call to a struct-return function, the callee
2186 // pops the hidden struct pointer, so we have to push it back.
2187 // This is common for Darwin/X86, Linux & Mingw32 targets.
2188 NumBytesForCalleeToPush = 4;
2190 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2192 // Returns a flag for retval copy to use.
2194 Chain = DAG.getCALLSEQ_END(Chain,
2195 DAG.getIntPtrConstant(NumBytes, true),
2196 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2199 InFlag = Chain.getValue(1);
2202 // Handle result values, copying them out of physregs into vregs that we
2204 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2205 Ins, dl, DAG, InVals);
2209 //===----------------------------------------------------------------------===//
2210 // Fast Calling Convention (tail call) implementation
2211 //===----------------------------------------------------------------------===//
2213 // Like std call, callee cleans arguments, convention except that ECX is
2214 // reserved for storing the tail called function address. Only 2 registers are
2215 // free for argument passing (inreg). Tail call optimization is performed
2217 // * tailcallopt is enabled
2218 // * caller/callee are fastcc
2219 // On X86_64 architecture with GOT-style position independent code only local
2220 // (within module) calls are supported at the moment.
2221 // To keep the stack aligned according to platform abi the function
2222 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2223 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2224 // If a tail called function callee has more arguments than the caller the
2225 // caller needs to make sure that there is room to move the RETADDR to. This is
2226 // achieved by reserving an area the size of the argument delta right after the
2227 // original REtADDR, but before the saved framepointer or the spilled registers
2228 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2240 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2241 /// for a 16 byte align requirement.
2243 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2244 SelectionDAG& DAG) const {
2245 MachineFunction &MF = DAG.getMachineFunction();
2246 const TargetMachine &TM = MF.getTarget();
2247 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2248 unsigned StackAlignment = TFI.getStackAlignment();
2249 uint64_t AlignMask = StackAlignment - 1;
2250 int64_t Offset = StackSize;
2251 uint64_t SlotSize = TD->getPointerSize();
2252 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2253 // Number smaller than 12 so just add the difference.
2254 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2256 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2257 Offset = ((~AlignMask) & Offset) + StackAlignment +
2258 (StackAlignment-SlotSize);
2263 /// MatchingStackOffset - Return true if the given stack call argument is
2264 /// already available in the same position (relatively) of the caller's
2265 /// incoming argument stack.
2267 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2268 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2269 const X86InstrInfo *TII) {
2270 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2272 if (Arg.getOpcode() == ISD::CopyFromReg) {
2273 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2274 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2276 MachineInstr *Def = MRI->getVRegDef(VR);
2279 if (!Flags.isByVal()) {
2280 if (!TII->isLoadFromStackSlot(Def, FI))
2283 unsigned Opcode = Def->getOpcode();
2284 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2285 Def->getOperand(1).isFI()) {
2286 FI = Def->getOperand(1).getIndex();
2287 Bytes = Flags.getByValSize();
2291 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2292 if (Flags.isByVal())
2293 // ByVal argument is passed in as a pointer but it's now being
2294 // dereferenced. e.g.
2295 // define @foo(%struct.X* %A) {
2296 // tail call @bar(%struct.X* byval %A)
2299 SDValue Ptr = Ld->getBasePtr();
2300 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2303 FI = FINode->getIndex();
2307 assert(FI != INT_MAX);
2308 if (!MFI->isFixedObjectIndex(FI))
2310 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2313 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2314 /// for tail call optimization. Targets which want to do tail call
2315 /// optimization should implement this function.
2317 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2318 CallingConv::ID CalleeCC,
2320 bool isCalleeStructRet,
2321 bool isCallerStructRet,
2322 const SmallVectorImpl<ISD::OutputArg> &Outs,
2323 const SmallVectorImpl<ISD::InputArg> &Ins,
2324 SelectionDAG& DAG) const {
2325 if (!IsTailCallConvention(CalleeCC) &&
2326 CalleeCC != CallingConv::C)
2329 // If -tailcallopt is specified, make fastcc functions tail-callable.
2330 const MachineFunction &MF = DAG.getMachineFunction();
2331 const Function *CallerF = DAG.getMachineFunction().getFunction();
2332 if (GuaranteedTailCallOpt) {
2333 if (IsTailCallConvention(CalleeCC) &&
2334 CallerF->getCallingConv() == CalleeCC)
2339 // Look for obvious safe cases to perform tail call optimization that does not
2340 // requite ABI changes. This is what gcc calls sibcall.
2342 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2343 // emit a special epilogue.
2344 if (RegInfo->needsStackRealignment(MF))
2347 // Do not sibcall optimize vararg calls unless the call site is not passing any
2349 if (isVarArg && !Outs.empty())
2352 // Also avoid sibcall optimization if either caller or callee uses struct
2353 // return semantics.
2354 if (isCalleeStructRet || isCallerStructRet)
2357 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2358 // Therefore if it's not used by the call it is not safe to optimize this into
2360 bool Unused = false;
2361 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2368 SmallVector<CCValAssign, 16> RVLocs;
2369 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2370 RVLocs, *DAG.getContext());
2371 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2372 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2373 CCValAssign &VA = RVLocs[i];
2374 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2379 // If the callee takes no arguments then go on to check the results of the
2381 if (!Outs.empty()) {
2382 // Check if stack adjustment is needed. For now, do not do this if any
2383 // argument is passed on the stack.
2384 SmallVector<CCValAssign, 16> ArgLocs;
2385 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2386 ArgLocs, *DAG.getContext());
2387 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2388 if (CCInfo.getNextStackOffset()) {
2389 MachineFunction &MF = DAG.getMachineFunction();
2390 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2392 if (Subtarget->isTargetWin64())
2393 // Win64 ABI has additional complications.
2396 // Check if the arguments are already laid out in the right way as
2397 // the caller's fixed stack objects.
2398 MachineFrameInfo *MFI = MF.getFrameInfo();
2399 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2400 const X86InstrInfo *TII =
2401 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2402 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2403 CCValAssign &VA = ArgLocs[i];
2404 EVT RegVT = VA.getLocVT();
2405 SDValue Arg = Outs[i].Val;
2406 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2407 if (VA.getLocInfo() == CCValAssign::Indirect)
2409 if (!VA.isRegLoc()) {
2410 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2422 X86TargetLowering::createFastISel(MachineFunction &mf,
2423 DenseMap<const Value *, unsigned> &vm,
2424 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2425 DenseMap<const AllocaInst *, int> &am
2427 , SmallSet<const Instruction *, 8> &cil
2430 return X86::createFastISel(mf, vm, bm, am
2438 //===----------------------------------------------------------------------===//
2439 // Other Lowering Hooks
2440 //===----------------------------------------------------------------------===//
2443 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2444 MachineFunction &MF = DAG.getMachineFunction();
2445 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2446 int ReturnAddrIndex = FuncInfo->getRAIndex();
2448 if (ReturnAddrIndex == 0) {
2449 // Set up a frame object for the return address.
2450 uint64_t SlotSize = TD->getPointerSize();
2451 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2453 FuncInfo->setRAIndex(ReturnAddrIndex);
2456 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2460 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2461 bool hasSymbolicDisplacement) {
2462 // Offset should fit into 32 bit immediate field.
2463 if (!isInt<32>(Offset))
2466 // If we don't have a symbolic displacement - we don't have any extra
2468 if (!hasSymbolicDisplacement)
2471 // FIXME: Some tweaks might be needed for medium code model.
2472 if (M != CodeModel::Small && M != CodeModel::Kernel)
2475 // For small code model we assume that latest object is 16MB before end of 31
2476 // bits boundary. We may also accept pretty large negative constants knowing
2477 // that all objects are in the positive half of address space.
2478 if (M == CodeModel::Small && Offset < 16*1024*1024)
2481 // For kernel code model we know that all object resist in the negative half
2482 // of 32bits address space. We may not accept negative offsets, since they may
2483 // be just off and we may accept pretty large positive ones.
2484 if (M == CodeModel::Kernel && Offset > 0)
2490 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2491 /// specific condition code, returning the condition code and the LHS/RHS of the
2492 /// comparison to make.
2493 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2494 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2496 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2497 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2498 // X > -1 -> X == 0, jump !sign.
2499 RHS = DAG.getConstant(0, RHS.getValueType());
2500 return X86::COND_NS;
2501 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2502 // X < 0 -> X == 0, jump on sign.
2504 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2506 RHS = DAG.getConstant(0, RHS.getValueType());
2507 return X86::COND_LE;
2511 switch (SetCCOpcode) {
2512 default: llvm_unreachable("Invalid integer condition!");
2513 case ISD::SETEQ: return X86::COND_E;
2514 case ISD::SETGT: return X86::COND_G;
2515 case ISD::SETGE: return X86::COND_GE;
2516 case ISD::SETLT: return X86::COND_L;
2517 case ISD::SETLE: return X86::COND_LE;
2518 case ISD::SETNE: return X86::COND_NE;
2519 case ISD::SETULT: return X86::COND_B;
2520 case ISD::SETUGT: return X86::COND_A;
2521 case ISD::SETULE: return X86::COND_BE;
2522 case ISD::SETUGE: return X86::COND_AE;
2526 // First determine if it is required or is profitable to flip the operands.
2528 // If LHS is a foldable load, but RHS is not, flip the condition.
2529 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2530 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2531 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2532 std::swap(LHS, RHS);
2535 switch (SetCCOpcode) {
2541 std::swap(LHS, RHS);
2545 // On a floating point condition, the flags are set as follows:
2547 // 0 | 0 | 0 | X > Y
2548 // 0 | 0 | 1 | X < Y
2549 // 1 | 0 | 0 | X == Y
2550 // 1 | 1 | 1 | unordered
2551 switch (SetCCOpcode) {
2552 default: llvm_unreachable("Condcode should be pre-legalized away");
2554 case ISD::SETEQ: return X86::COND_E;
2555 case ISD::SETOLT: // flipped
2557 case ISD::SETGT: return X86::COND_A;
2558 case ISD::SETOLE: // flipped
2560 case ISD::SETGE: return X86::COND_AE;
2561 case ISD::SETUGT: // flipped
2563 case ISD::SETLT: return X86::COND_B;
2564 case ISD::SETUGE: // flipped
2566 case ISD::SETLE: return X86::COND_BE;
2568 case ISD::SETNE: return X86::COND_NE;
2569 case ISD::SETUO: return X86::COND_P;
2570 case ISD::SETO: return X86::COND_NP;
2572 case ISD::SETUNE: return X86::COND_INVALID;
2576 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2577 /// code. Current x86 isa includes the following FP cmov instructions:
2578 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2579 static bool hasFPCMov(unsigned X86CC) {
2595 /// isFPImmLegal - Returns true if the target can instruction select the
2596 /// specified FP immediate natively. If false, the legalizer will
2597 /// materialize the FP immediate as a load from a constant pool.
2598 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2599 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2600 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2606 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2607 /// the specified range (L, H].
2608 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2609 return (Val < 0) || (Val >= Low && Val < Hi);
2612 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2613 /// specified value.
2614 static bool isUndefOrEqual(int Val, int CmpVal) {
2615 if (Val < 0 || Val == CmpVal)
2620 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2621 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2622 /// the second operand.
2623 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2624 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2625 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2626 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2627 return (Mask[0] < 2 && Mask[1] < 2);
2631 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2632 SmallVector<int, 8> M;
2634 return ::isPSHUFDMask(M, N->getValueType(0));
2637 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2638 /// is suitable for input to PSHUFHW.
2639 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2640 if (VT != MVT::v8i16)
2643 // Lower quadword copied in order or undef.
2644 for (int i = 0; i != 4; ++i)
2645 if (Mask[i] >= 0 && Mask[i] != i)
2648 // Upper quadword shuffled.
2649 for (int i = 4; i != 8; ++i)
2650 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2656 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2657 SmallVector<int, 8> M;
2659 return ::isPSHUFHWMask(M, N->getValueType(0));
2662 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2663 /// is suitable for input to PSHUFLW.
2664 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2665 if (VT != MVT::v8i16)
2668 // Upper quadword copied in order.
2669 for (int i = 4; i != 8; ++i)
2670 if (Mask[i] >= 0 && Mask[i] != i)
2673 // Lower quadword shuffled.
2674 for (int i = 0; i != 4; ++i)
2681 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2682 SmallVector<int, 8> M;
2684 return ::isPSHUFLWMask(M, N->getValueType(0));
2687 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2688 /// is suitable for input to PALIGNR.
2689 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2691 int i, e = VT.getVectorNumElements();
2693 // Do not handle v2i64 / v2f64 shuffles with palignr.
2694 if (e < 4 || !hasSSSE3)
2697 for (i = 0; i != e; ++i)
2701 // All undef, not a palignr.
2705 // Determine if it's ok to perform a palignr with only the LHS, since we
2706 // don't have access to the actual shuffle elements to see if RHS is undef.
2707 bool Unary = Mask[i] < (int)e;
2708 bool NeedsUnary = false;
2710 int s = Mask[i] - i;
2712 // Check the rest of the elements to see if they are consecutive.
2713 for (++i; i != e; ++i) {
2718 Unary = Unary && (m < (int)e);
2719 NeedsUnary = NeedsUnary || (m < s);
2721 if (NeedsUnary && !Unary)
2723 if (Unary && m != ((s+i) & (e-1)))
2725 if (!Unary && m != (s+i))
2731 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2732 SmallVector<int, 8> M;
2734 return ::isPALIGNRMask(M, N->getValueType(0), true);
2737 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2738 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2739 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2740 int NumElems = VT.getVectorNumElements();
2741 if (NumElems != 2 && NumElems != 4)
2744 int Half = NumElems / 2;
2745 for (int i = 0; i < Half; ++i)
2746 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2748 for (int i = Half; i < NumElems; ++i)
2749 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2755 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2756 SmallVector<int, 8> M;
2758 return ::isSHUFPMask(M, N->getValueType(0));
2761 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2762 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2763 /// half elements to come from vector 1 (which would equal the dest.) and
2764 /// the upper half to come from vector 2.
2765 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2766 int NumElems = VT.getVectorNumElements();
2768 if (NumElems != 2 && NumElems != 4)
2771 int Half = NumElems / 2;
2772 for (int i = 0; i < Half; ++i)
2773 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2775 for (int i = Half; i < NumElems; ++i)
2776 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2781 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2782 SmallVector<int, 8> M;
2784 return isCommutedSHUFPMask(M, N->getValueType(0));
2787 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2788 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2789 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2790 if (N->getValueType(0).getVectorNumElements() != 4)
2793 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2794 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2795 isUndefOrEqual(N->getMaskElt(1), 7) &&
2796 isUndefOrEqual(N->getMaskElt(2), 2) &&
2797 isUndefOrEqual(N->getMaskElt(3), 3);
2800 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2801 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2803 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2804 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2809 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2810 isUndefOrEqual(N->getMaskElt(1), 3) &&
2811 isUndefOrEqual(N->getMaskElt(2), 2) &&
2812 isUndefOrEqual(N->getMaskElt(3), 3);
2815 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2816 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2817 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2818 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2820 if (NumElems != 2 && NumElems != 4)
2823 for (unsigned i = 0; i < NumElems/2; ++i)
2824 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2827 for (unsigned i = NumElems/2; i < NumElems; ++i)
2828 if (!isUndefOrEqual(N->getMaskElt(i), i))
2834 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2835 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2836 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2837 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2839 if (NumElems != 2 && NumElems != 4)
2842 for (unsigned i = 0; i < NumElems/2; ++i)
2843 if (!isUndefOrEqual(N->getMaskElt(i), i))
2846 for (unsigned i = 0; i < NumElems/2; ++i)
2847 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2853 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2854 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2855 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2856 bool V2IsSplat = false) {
2857 int NumElts = VT.getVectorNumElements();
2858 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2861 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2863 int BitI1 = Mask[i+1];
2864 if (!isUndefOrEqual(BitI, j))
2867 if (!isUndefOrEqual(BitI1, NumElts))
2870 if (!isUndefOrEqual(BitI1, j + NumElts))
2877 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2878 SmallVector<int, 8> M;
2880 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2883 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2884 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2885 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2886 bool V2IsSplat = false) {
2887 int NumElts = VT.getVectorNumElements();
2888 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2891 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2893 int BitI1 = Mask[i+1];
2894 if (!isUndefOrEqual(BitI, j + NumElts/2))
2897 if (isUndefOrEqual(BitI1, NumElts))
2900 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2907 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2908 SmallVector<int, 8> M;
2910 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2913 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2914 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2916 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2917 int NumElems = VT.getVectorNumElements();
2918 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2921 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2923 int BitI1 = Mask[i+1];
2924 if (!isUndefOrEqual(BitI, j))
2926 if (!isUndefOrEqual(BitI1, j))
2932 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2933 SmallVector<int, 8> M;
2935 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2938 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2939 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2941 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2942 int NumElems = VT.getVectorNumElements();
2943 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2946 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2948 int BitI1 = Mask[i+1];
2949 if (!isUndefOrEqual(BitI, j))
2951 if (!isUndefOrEqual(BitI1, j))
2957 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2958 SmallVector<int, 8> M;
2960 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2963 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2964 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2965 /// MOVSD, and MOVD, i.e. setting the lowest element.
2966 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2967 if (VT.getVectorElementType().getSizeInBits() < 32)
2970 int NumElts = VT.getVectorNumElements();
2972 if (!isUndefOrEqual(Mask[0], NumElts))
2975 for (int i = 1; i < NumElts; ++i)
2976 if (!isUndefOrEqual(Mask[i], i))
2982 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2983 SmallVector<int, 8> M;
2985 return ::isMOVLMask(M, N->getValueType(0));
2988 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2989 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2990 /// element of vector 2 and the other elements to come from vector 1 in order.
2991 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2992 bool V2IsSplat = false, bool V2IsUndef = false) {
2993 int NumOps = VT.getVectorNumElements();
2994 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2997 if (!isUndefOrEqual(Mask[0], 0))
3000 for (int i = 1; i < NumOps; ++i)
3001 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3002 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3003 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3009 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3010 bool V2IsUndef = false) {
3011 SmallVector<int, 8> M;
3013 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3016 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3017 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3018 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3019 if (N->getValueType(0).getVectorNumElements() != 4)
3022 // Expect 1, 1, 3, 3
3023 for (unsigned i = 0; i < 2; ++i) {
3024 int Elt = N->getMaskElt(i);
3025 if (Elt >= 0 && Elt != 1)
3030 for (unsigned i = 2; i < 4; ++i) {
3031 int Elt = N->getMaskElt(i);
3032 if (Elt >= 0 && Elt != 3)
3037 // Don't use movshdup if it can be done with a shufps.
3038 // FIXME: verify that matching u, u, 3, 3 is what we want.
3042 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3043 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3044 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3045 if (N->getValueType(0).getVectorNumElements() != 4)
3048 // Expect 0, 0, 2, 2
3049 for (unsigned i = 0; i < 2; ++i)
3050 if (N->getMaskElt(i) > 0)
3054 for (unsigned i = 2; i < 4; ++i) {
3055 int Elt = N->getMaskElt(i);
3056 if (Elt >= 0 && Elt != 2)
3061 // Don't use movsldup if it can be done with a shufps.
3065 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3066 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3067 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3068 int e = N->getValueType(0).getVectorNumElements() / 2;
3070 for (int i = 0; i < e; ++i)
3071 if (!isUndefOrEqual(N->getMaskElt(i), i))
3073 for (int i = 0; i < e; ++i)
3074 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3079 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3080 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3081 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3082 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3083 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3085 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3087 for (int i = 0; i < NumOperands; ++i) {
3088 int Val = SVOp->getMaskElt(NumOperands-i-1);
3089 if (Val < 0) Val = 0;
3090 if (Val >= NumOperands) Val -= NumOperands;
3092 if (i != NumOperands - 1)
3098 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3099 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3100 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3101 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3103 // 8 nodes, but we only care about the last 4.
3104 for (unsigned i = 7; i >= 4; --i) {
3105 int Val = SVOp->getMaskElt(i);
3114 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3115 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3116 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3117 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3119 // 8 nodes, but we only care about the first 4.
3120 for (int i = 3; i >= 0; --i) {
3121 int Val = SVOp->getMaskElt(i);
3130 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3131 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3132 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3133 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3134 EVT VVT = N->getValueType(0);
3135 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3139 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3140 Val = SVOp->getMaskElt(i);
3144 return (Val - i) * EltSize;
3147 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3149 bool X86::isZeroNode(SDValue Elt) {
3150 return ((isa<ConstantSDNode>(Elt) &&
3151 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3152 (isa<ConstantFPSDNode>(Elt) &&
3153 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3156 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3157 /// their permute mask.
3158 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3159 SelectionDAG &DAG) {
3160 EVT VT = SVOp->getValueType(0);
3161 unsigned NumElems = VT.getVectorNumElements();
3162 SmallVector<int, 8> MaskVec;
3164 for (unsigned i = 0; i != NumElems; ++i) {
3165 int idx = SVOp->getMaskElt(i);
3167 MaskVec.push_back(idx);
3168 else if (idx < (int)NumElems)
3169 MaskVec.push_back(idx + NumElems);
3171 MaskVec.push_back(idx - NumElems);
3173 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3174 SVOp->getOperand(0), &MaskVec[0]);
3177 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3178 /// the two vector operands have swapped position.
3179 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3180 unsigned NumElems = VT.getVectorNumElements();
3181 for (unsigned i = 0; i != NumElems; ++i) {
3185 else if (idx < (int)NumElems)
3186 Mask[i] = idx + NumElems;
3188 Mask[i] = idx - NumElems;
3192 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3193 /// match movhlps. The lower half elements should come from upper half of
3194 /// V1 (and in order), and the upper half elements should come from the upper
3195 /// half of V2 (and in order).
3196 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3197 if (Op->getValueType(0).getVectorNumElements() != 4)
3199 for (unsigned i = 0, e = 2; i != e; ++i)
3200 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3202 for (unsigned i = 2; i != 4; ++i)
3203 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3208 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3209 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3211 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3212 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3214 N = N->getOperand(0).getNode();
3215 if (!ISD::isNON_EXTLoad(N))
3218 *LD = cast<LoadSDNode>(N);
3222 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3223 /// match movlp{s|d}. The lower half elements should come from lower half of
3224 /// V1 (and in order), and the upper half elements should come from the upper
3225 /// half of V2 (and in order). And since V1 will become the source of the
3226 /// MOVLP, it must be either a vector load or a scalar load to vector.
3227 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3228 ShuffleVectorSDNode *Op) {
3229 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3231 // Is V2 is a vector load, don't do this transformation. We will try to use
3232 // load folding shufps op.
3233 if (ISD::isNON_EXTLoad(V2))
3236 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3238 if (NumElems != 2 && NumElems != 4)
3240 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3241 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3243 for (unsigned i = NumElems/2; i != NumElems; ++i)
3244 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3249 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3251 static bool isSplatVector(SDNode *N) {
3252 if (N->getOpcode() != ISD::BUILD_VECTOR)
3255 SDValue SplatValue = N->getOperand(0);
3256 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3257 if (N->getOperand(i) != SplatValue)
3262 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3263 /// to an zero vector.
3264 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3265 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3266 SDValue V1 = N->getOperand(0);
3267 SDValue V2 = N->getOperand(1);
3268 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3269 for (unsigned i = 0; i != NumElems; ++i) {
3270 int Idx = N->getMaskElt(i);
3271 if (Idx >= (int)NumElems) {
3272 unsigned Opc = V2.getOpcode();
3273 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3275 if (Opc != ISD::BUILD_VECTOR ||
3276 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3278 } else if (Idx >= 0) {
3279 unsigned Opc = V1.getOpcode();
3280 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3282 if (Opc != ISD::BUILD_VECTOR ||
3283 !X86::isZeroNode(V1.getOperand(Idx)))
3290 /// getZeroVector - Returns a vector of specified type with all zero elements.
3292 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3294 assert(VT.isVector() && "Expected a vector type");
3296 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3297 // type. This ensures they get CSE'd.
3299 if (VT.getSizeInBits() == 64) { // MMX
3300 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3301 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3302 } else if (HasSSE2) { // SSE2
3303 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3304 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3306 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3307 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3309 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3312 /// getOnesVector - Returns a vector of specified type with all bits set.
3314 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3315 assert(VT.isVector() && "Expected a vector type");
3317 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3318 // type. This ensures they get CSE'd.
3319 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3321 if (VT.getSizeInBits() == 64) // MMX
3322 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3324 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3325 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3329 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3330 /// that point to V2 points to its first element.
3331 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3332 EVT VT = SVOp->getValueType(0);
3333 unsigned NumElems = VT.getVectorNumElements();
3335 bool Changed = false;
3336 SmallVector<int, 8> MaskVec;
3337 SVOp->getMask(MaskVec);
3339 for (unsigned i = 0; i != NumElems; ++i) {
3340 if (MaskVec[i] > (int)NumElems) {
3341 MaskVec[i] = NumElems;
3346 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3347 SVOp->getOperand(1), &MaskVec[0]);
3348 return SDValue(SVOp, 0);
3351 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3352 /// operation of specified width.
3353 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3355 unsigned NumElems = VT.getVectorNumElements();
3356 SmallVector<int, 8> Mask;
3357 Mask.push_back(NumElems);
3358 for (unsigned i = 1; i != NumElems; ++i)
3360 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3363 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3364 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3366 unsigned NumElems = VT.getVectorNumElements();
3367 SmallVector<int, 8> Mask;
3368 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3370 Mask.push_back(i + NumElems);
3372 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3375 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3376 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3378 unsigned NumElems = VT.getVectorNumElements();
3379 unsigned Half = NumElems/2;
3380 SmallVector<int, 8> Mask;
3381 for (unsigned i = 0; i != Half; ++i) {
3382 Mask.push_back(i + Half);
3383 Mask.push_back(i + NumElems + Half);
3385 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3388 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3389 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3391 if (SV->getValueType(0).getVectorNumElements() <= 4)
3392 return SDValue(SV, 0);
3394 EVT PVT = MVT::v4f32;
3395 EVT VT = SV->getValueType(0);
3396 DebugLoc dl = SV->getDebugLoc();
3397 SDValue V1 = SV->getOperand(0);
3398 int NumElems = VT.getVectorNumElements();
3399 int EltNo = SV->getSplatIndex();
3401 // unpack elements to the correct location
3402 while (NumElems > 4) {
3403 if (EltNo < NumElems/2) {
3404 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3406 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3407 EltNo -= NumElems/2;
3412 // Perform the splat.
3413 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3414 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3415 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3416 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3419 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3420 /// vector of zero or undef vector. This produces a shuffle where the low
3421 /// element of V2 is swizzled into the zero/undef vector, landing at element
3422 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3423 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3424 bool isZero, bool HasSSE2,
3425 SelectionDAG &DAG) {
3426 EVT VT = V2.getValueType();
3428 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3429 unsigned NumElems = VT.getVectorNumElements();
3430 SmallVector<int, 16> MaskVec;
3431 for (unsigned i = 0; i != NumElems; ++i)
3432 // If this is the insertion idx, put the low elt of V2 here.
3433 MaskVec.push_back(i == Idx ? NumElems : i);
3434 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3437 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3438 /// a shuffle that is zero.
3440 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3441 bool Low, SelectionDAG &DAG) {
3442 unsigned NumZeros = 0;
3443 for (int i = 0; i < NumElems; ++i) {
3444 unsigned Index = Low ? i : NumElems-i-1;
3445 int Idx = SVOp->getMaskElt(Index);
3450 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3451 if (Elt.getNode() && X86::isZeroNode(Elt))
3459 /// isVectorShift - Returns true if the shuffle can be implemented as a
3460 /// logical left or right shift of a vector.
3461 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3462 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3463 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3464 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3467 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3470 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3474 bool SeenV1 = false;
3475 bool SeenV2 = false;
3476 for (unsigned i = NumZeros; i < NumElems; ++i) {
3477 unsigned Val = isLeft ? (i - NumZeros) : i;
3478 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3481 unsigned Idx = (unsigned) Idx_;
3491 if (SeenV1 && SeenV2)
3494 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3500 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3502 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3503 unsigned NumNonZero, unsigned NumZero,
3505 const TargetLowering &TLI) {
3509 DebugLoc dl = Op.getDebugLoc();
3512 for (unsigned i = 0; i < 16; ++i) {
3513 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3514 if (ThisIsNonZero && First) {
3516 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3518 V = DAG.getUNDEF(MVT::v8i16);
3523 SDValue ThisElt(0, 0), LastElt(0, 0);
3524 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3525 if (LastIsNonZero) {
3526 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3527 MVT::i16, Op.getOperand(i-1));
3529 if (ThisIsNonZero) {
3530 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3531 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3532 ThisElt, DAG.getConstant(8, MVT::i8));
3534 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3538 if (ThisElt.getNode())
3539 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3540 DAG.getIntPtrConstant(i/2));
3544 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3547 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3549 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3550 unsigned NumNonZero, unsigned NumZero,
3552 const TargetLowering &TLI) {
3556 DebugLoc dl = Op.getDebugLoc();
3559 for (unsigned i = 0; i < 8; ++i) {
3560 bool isNonZero = (NonZeros & (1 << i)) != 0;
3564 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3566 V = DAG.getUNDEF(MVT::v8i16);
3569 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3570 MVT::v8i16, V, Op.getOperand(i),
3571 DAG.getIntPtrConstant(i));
3578 /// getVShift - Return a vector logical shift node.
3580 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3581 unsigned NumBits, SelectionDAG &DAG,
3582 const TargetLowering &TLI, DebugLoc dl) {
3583 bool isMMX = VT.getSizeInBits() == 64;
3584 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3585 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3586 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3587 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3588 DAG.getNode(Opc, dl, ShVT, SrcOp,
3589 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3593 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3594 SelectionDAG &DAG) const {
3596 // Check if the scalar load can be widened into a vector load. And if
3597 // the address is "base + cst" see if the cst can be "absorbed" into
3598 // the shuffle mask.
3599 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3600 SDValue Ptr = LD->getBasePtr();
3601 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3603 EVT PVT = LD->getValueType(0);
3604 if (PVT != MVT::i32 && PVT != MVT::f32)
3609 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3610 FI = FINode->getIndex();
3612 } else if (Ptr.getOpcode() == ISD::ADD &&
3613 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3614 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3615 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3616 Offset = Ptr.getConstantOperandVal(1);
3617 Ptr = Ptr.getOperand(0);
3622 SDValue Chain = LD->getChain();
3623 // Make sure the stack object alignment is at least 16.
3624 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3625 if (DAG.InferPtrAlignment(Ptr) < 16) {
3626 if (MFI->isFixedObjectIndex(FI)) {
3627 // Can't change the alignment. FIXME: It's possible to compute
3628 // the exact stack offset and reference FI + adjust offset instead.
3629 // If someone *really* cares about this. That's the way to implement it.
3632 MFI->setObjectAlignment(FI, 16);
3636 // (Offset % 16) must be multiple of 4. Then address is then
3637 // Ptr + (Offset & ~15).
3640 if ((Offset % 16) & 3)
3642 int64_t StartOffset = Offset & ~15;
3644 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3645 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3647 int EltNo = (Offset - StartOffset) >> 2;
3648 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3649 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3650 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3652 // Canonicalize it to a v4i32 shuffle.
3653 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3654 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3655 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3656 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3662 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3663 /// vector of type 'VT', see if the elements can be replaced by a single large
3664 /// load which has the same value as a build_vector whose operands are 'elts'.
3666 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3668 /// FIXME: we'd also like to handle the case where the last elements are zero
3669 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3670 /// There's even a handy isZeroNode for that purpose.
3671 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3672 DebugLoc &dl, SelectionDAG &DAG) {
3673 EVT EltVT = VT.getVectorElementType();
3674 unsigned NumElems = Elts.size();
3676 LoadSDNode *LDBase = NULL;
3677 unsigned LastLoadedElt = -1U;
3679 // For each element in the initializer, see if we've found a load or an undef.
3680 // If we don't find an initial load element, or later load elements are
3681 // non-consecutive, bail out.
3682 for (unsigned i = 0; i < NumElems; ++i) {
3683 SDValue Elt = Elts[i];
3685 if (!Elt.getNode() ||
3686 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3689 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3691 LDBase = cast<LoadSDNode>(Elt.getNode());
3695 if (Elt.getOpcode() == ISD::UNDEF)
3698 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3699 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3704 // If we have found an entire vector of loads and undefs, then return a large
3705 // load of the entire vector width starting at the base pointer. If we found
3706 // consecutive loads for the low half, generate a vzext_load node.
3707 if (LastLoadedElt == NumElems - 1) {
3708 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3709 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3710 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3711 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3712 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3713 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3714 LDBase->isVolatile(), LDBase->isNonTemporal(),
3715 LDBase->getAlignment());
3716 } else if (NumElems == 4 && LastLoadedElt == 1) {
3717 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3718 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3719 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3720 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3726 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3727 DebugLoc dl = Op.getDebugLoc();
3728 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3729 if (ISD::isBuildVectorAllZeros(Op.getNode())
3730 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3731 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3732 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3733 // eliminated on x86-32 hosts.
3734 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3737 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3738 return getOnesVector(Op.getValueType(), DAG, dl);
3739 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3742 EVT VT = Op.getValueType();
3743 EVT ExtVT = VT.getVectorElementType();
3744 unsigned EVTBits = ExtVT.getSizeInBits();
3746 unsigned NumElems = Op.getNumOperands();
3747 unsigned NumZero = 0;
3748 unsigned NumNonZero = 0;
3749 unsigned NonZeros = 0;
3750 bool IsAllConstants = true;
3751 SmallSet<SDValue, 8> Values;
3752 for (unsigned i = 0; i < NumElems; ++i) {
3753 SDValue Elt = Op.getOperand(i);
3754 if (Elt.getOpcode() == ISD::UNDEF)
3757 if (Elt.getOpcode() != ISD::Constant &&
3758 Elt.getOpcode() != ISD::ConstantFP)
3759 IsAllConstants = false;
3760 if (X86::isZeroNode(Elt))
3763 NonZeros |= (1 << i);
3768 if (NumNonZero == 0) {
3769 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3770 return DAG.getUNDEF(VT);
3773 // Special case for single non-zero, non-undef, element.
3774 if (NumNonZero == 1) {
3775 unsigned Idx = CountTrailingZeros_32(NonZeros);
3776 SDValue Item = Op.getOperand(Idx);
3778 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3779 // the value are obviously zero, truncate the value to i32 and do the
3780 // insertion that way. Only do this if the value is non-constant or if the
3781 // value is a constant being inserted into element 0. It is cheaper to do
3782 // a constant pool load than it is to do a movd + shuffle.
3783 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3784 (!IsAllConstants || Idx == 0)) {
3785 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3786 // Handle MMX and SSE both.
3787 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3788 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3790 // Truncate the value (which may itself be a constant) to i32, and
3791 // convert it to a vector with movd (S2V+shuffle to zero extend).
3792 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3793 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3794 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3795 Subtarget->hasSSE2(), DAG);
3797 // Now we have our 32-bit value zero extended in the low element of
3798 // a vector. If Idx != 0, swizzle it into place.
3800 SmallVector<int, 4> Mask;
3801 Mask.push_back(Idx);
3802 for (unsigned i = 1; i != VecElts; ++i)
3804 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3805 DAG.getUNDEF(Item.getValueType()),
3808 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3812 // If we have a constant or non-constant insertion into the low element of
3813 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3814 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3815 // depending on what the source datatype is.
3818 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3819 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3820 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3821 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3822 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3823 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3825 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3826 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3827 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3828 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3829 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3830 Subtarget->hasSSE2(), DAG);
3831 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3835 // Is it a vector logical left shift?
3836 if (NumElems == 2 && Idx == 1 &&
3837 X86::isZeroNode(Op.getOperand(0)) &&
3838 !X86::isZeroNode(Op.getOperand(1))) {
3839 unsigned NumBits = VT.getSizeInBits();
3840 return getVShift(true, VT,
3841 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3842 VT, Op.getOperand(1)),
3843 NumBits/2, DAG, *this, dl);
3846 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3849 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3850 // is a non-constant being inserted into an element other than the low one,
3851 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3852 // movd/movss) to move this into the low element, then shuffle it into
3854 if (EVTBits == 32) {
3855 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3857 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3858 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3859 Subtarget->hasSSE2(), DAG);
3860 SmallVector<int, 8> MaskVec;
3861 for (unsigned i = 0; i < NumElems; i++)
3862 MaskVec.push_back(i == Idx ? 0 : 1);
3863 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3867 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3868 if (Values.size() == 1) {
3869 if (EVTBits == 32) {
3870 // Instead of a shuffle like this:
3871 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3872 // Check if it's possible to issue this instead.
3873 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3874 unsigned Idx = CountTrailingZeros_32(NonZeros);
3875 SDValue Item = Op.getOperand(Idx);
3876 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3877 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3882 // A vector full of immediates; various special cases are already
3883 // handled, so this is best done with a single constant-pool load.
3887 // Let legalizer expand 2-wide build_vectors.
3888 if (EVTBits == 64) {
3889 if (NumNonZero == 1) {
3890 // One half is zero or undef.
3891 unsigned Idx = CountTrailingZeros_32(NonZeros);
3892 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3893 Op.getOperand(Idx));
3894 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3895 Subtarget->hasSSE2(), DAG);
3900 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3901 if (EVTBits == 8 && NumElems == 16) {
3902 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3904 if (V.getNode()) return V;
3907 if (EVTBits == 16 && NumElems == 8) {
3908 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3910 if (V.getNode()) return V;
3913 // If element VT is == 32 bits, turn it into a number of shuffles.
3914 SmallVector<SDValue, 8> V;
3916 if (NumElems == 4 && NumZero > 0) {
3917 for (unsigned i = 0; i < 4; ++i) {
3918 bool isZero = !(NonZeros & (1 << i));
3920 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3922 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3925 for (unsigned i = 0; i < 2; ++i) {
3926 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3929 V[i] = V[i*2]; // Must be a zero vector.
3932 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3935 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3938 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3943 SmallVector<int, 8> MaskVec;
3944 bool Reverse = (NonZeros & 0x3) == 2;
3945 for (unsigned i = 0; i < 2; ++i)
3946 MaskVec.push_back(Reverse ? 1-i : i);
3947 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3948 for (unsigned i = 0; i < 2; ++i)
3949 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3950 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3953 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3954 // Check for a build vector of consecutive loads.
3955 for (unsigned i = 0; i < NumElems; ++i)
3956 V[i] = Op.getOperand(i);
3958 // Check for elements which are consecutive loads.
3959 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
3963 // For SSE 4.1, use inserts into undef.
3964 if (getSubtarget()->hasSSE41()) {
3965 V[0] = DAG.getUNDEF(VT);
3966 for (unsigned i = 0; i < NumElems; ++i)
3967 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3968 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3969 Op.getOperand(i), DAG.getIntPtrConstant(i));
3973 // Otherwise, expand into a number of unpckl*
3975 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3976 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3977 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3978 for (unsigned i = 0; i < NumElems; ++i)
3979 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3981 while (NumElems != 0) {
3982 for (unsigned i = 0; i < NumElems; ++i)
3983 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3992 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
3993 // We support concatenate two MMX registers and place them in a MMX
3994 // register. This is better than doing a stack convert.
3995 DebugLoc dl = Op.getDebugLoc();
3996 EVT ResVT = Op.getValueType();
3997 assert(Op.getNumOperands() == 2);
3998 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3999 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4001 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4002 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4003 InVec = Op.getOperand(1);
4004 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4005 unsigned NumElts = ResVT.getVectorNumElements();
4006 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4007 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4008 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4010 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4011 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4012 Mask[0] = 0; Mask[1] = 2;
4013 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4015 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4018 // v8i16 shuffles - Prefer shuffles in the following order:
4019 // 1. [all] pshuflw, pshufhw, optional move
4020 // 2. [ssse3] 1 x pshufb
4021 // 3. [ssse3] 2 x pshufb + 1 x por
4022 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4024 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4026 const X86TargetLowering &TLI) {
4027 SDValue V1 = SVOp->getOperand(0);
4028 SDValue V2 = SVOp->getOperand(1);
4029 DebugLoc dl = SVOp->getDebugLoc();
4030 SmallVector<int, 8> MaskVals;
4032 // Determine if more than 1 of the words in each of the low and high quadwords
4033 // of the result come from the same quadword of one of the two inputs. Undef
4034 // mask values count as coming from any quadword, for better codegen.
4035 SmallVector<unsigned, 4> LoQuad(4);
4036 SmallVector<unsigned, 4> HiQuad(4);
4037 BitVector InputQuads(4);
4038 for (unsigned i = 0; i < 8; ++i) {
4039 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4040 int EltIdx = SVOp->getMaskElt(i);
4041 MaskVals.push_back(EltIdx);
4050 InputQuads.set(EltIdx / 4);
4053 int BestLoQuad = -1;
4054 unsigned MaxQuad = 1;
4055 for (unsigned i = 0; i < 4; ++i) {
4056 if (LoQuad[i] > MaxQuad) {
4058 MaxQuad = LoQuad[i];
4062 int BestHiQuad = -1;
4064 for (unsigned i = 0; i < 4; ++i) {
4065 if (HiQuad[i] > MaxQuad) {
4067 MaxQuad = HiQuad[i];
4071 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4072 // of the two input vectors, shuffle them into one input vector so only a
4073 // single pshufb instruction is necessary. If There are more than 2 input
4074 // quads, disable the next transformation since it does not help SSSE3.
4075 bool V1Used = InputQuads[0] || InputQuads[1];
4076 bool V2Used = InputQuads[2] || InputQuads[3];
4077 if (TLI.getSubtarget()->hasSSSE3()) {
4078 if (InputQuads.count() == 2 && V1Used && V2Used) {
4079 BestLoQuad = InputQuads.find_first();
4080 BestHiQuad = InputQuads.find_next(BestLoQuad);
4082 if (InputQuads.count() > 2) {
4088 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4089 // the shuffle mask. If a quad is scored as -1, that means that it contains
4090 // words from all 4 input quadwords.
4092 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4093 SmallVector<int, 8> MaskV;
4094 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4095 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4096 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4097 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4098 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4099 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4101 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4102 // source words for the shuffle, to aid later transformations.
4103 bool AllWordsInNewV = true;
4104 bool InOrder[2] = { true, true };
4105 for (unsigned i = 0; i != 8; ++i) {
4106 int idx = MaskVals[i];
4108 InOrder[i/4] = false;
4109 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4111 AllWordsInNewV = false;
4115 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4116 if (AllWordsInNewV) {
4117 for (int i = 0; i != 8; ++i) {
4118 int idx = MaskVals[i];
4121 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4122 if ((idx != i) && idx < 4)
4124 if ((idx != i) && idx > 3)
4133 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4134 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4135 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4136 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4137 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4141 // If we have SSSE3, and all words of the result are from 1 input vector,
4142 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4143 // is present, fall back to case 4.
4144 if (TLI.getSubtarget()->hasSSSE3()) {
4145 SmallVector<SDValue,16> pshufbMask;
4147 // If we have elements from both input vectors, set the high bit of the
4148 // shuffle mask element to zero out elements that come from V2 in the V1
4149 // mask, and elements that come from V1 in the V2 mask, so that the two
4150 // results can be OR'd together.
4151 bool TwoInputs = V1Used && V2Used;
4152 for (unsigned i = 0; i != 8; ++i) {
4153 int EltIdx = MaskVals[i] * 2;
4154 if (TwoInputs && (EltIdx >= 16)) {
4155 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4156 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4159 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4160 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4162 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4163 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4164 DAG.getNode(ISD::BUILD_VECTOR, dl,
4165 MVT::v16i8, &pshufbMask[0], 16));
4167 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4169 // Calculate the shuffle mask for the second input, shuffle it, and
4170 // OR it with the first shuffled input.
4172 for (unsigned i = 0; i != 8; ++i) {
4173 int EltIdx = MaskVals[i] * 2;
4175 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4176 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4179 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4180 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4182 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4183 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4184 DAG.getNode(ISD::BUILD_VECTOR, dl,
4185 MVT::v16i8, &pshufbMask[0], 16));
4186 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4187 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4190 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4191 // and update MaskVals with new element order.
4192 BitVector InOrder(8);
4193 if (BestLoQuad >= 0) {
4194 SmallVector<int, 8> MaskV;
4195 for (int i = 0; i != 4; ++i) {
4196 int idx = MaskVals[i];
4198 MaskV.push_back(-1);
4200 } else if ((idx / 4) == BestLoQuad) {
4201 MaskV.push_back(idx & 3);
4204 MaskV.push_back(-1);
4207 for (unsigned i = 4; i != 8; ++i)
4209 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4213 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4214 // and update MaskVals with the new element order.
4215 if (BestHiQuad >= 0) {
4216 SmallVector<int, 8> MaskV;
4217 for (unsigned i = 0; i != 4; ++i)
4219 for (unsigned i = 4; i != 8; ++i) {
4220 int idx = MaskVals[i];
4222 MaskV.push_back(-1);
4224 } else if ((idx / 4) == BestHiQuad) {
4225 MaskV.push_back((idx & 3) + 4);
4228 MaskV.push_back(-1);
4231 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4235 // In case BestHi & BestLo were both -1, which means each quadword has a word
4236 // from each of the four input quadwords, calculate the InOrder bitvector now
4237 // before falling through to the insert/extract cleanup.
4238 if (BestLoQuad == -1 && BestHiQuad == -1) {
4240 for (int i = 0; i != 8; ++i)
4241 if (MaskVals[i] < 0 || MaskVals[i] == i)
4245 // The other elements are put in the right place using pextrw and pinsrw.
4246 for (unsigned i = 0; i != 8; ++i) {
4249 int EltIdx = MaskVals[i];
4252 SDValue ExtOp = (EltIdx < 8)
4253 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4254 DAG.getIntPtrConstant(EltIdx))
4255 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4256 DAG.getIntPtrConstant(EltIdx - 8));
4257 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4258 DAG.getIntPtrConstant(i));
4263 // v16i8 shuffles - Prefer shuffles in the following order:
4264 // 1. [ssse3] 1 x pshufb
4265 // 2. [ssse3] 2 x pshufb + 1 x por
4266 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4268 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4270 const X86TargetLowering &TLI) {
4271 SDValue V1 = SVOp->getOperand(0);
4272 SDValue V2 = SVOp->getOperand(1);
4273 DebugLoc dl = SVOp->getDebugLoc();
4274 SmallVector<int, 16> MaskVals;
4275 SVOp->getMask(MaskVals);
4277 // If we have SSSE3, case 1 is generated when all result bytes come from
4278 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4279 // present, fall back to case 3.
4280 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4283 for (unsigned i = 0; i < 16; ++i) {
4284 int EltIdx = MaskVals[i];
4293 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4294 if (TLI.getSubtarget()->hasSSSE3()) {
4295 SmallVector<SDValue,16> pshufbMask;
4297 // If all result elements are from one input vector, then only translate
4298 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4300 // Otherwise, we have elements from both input vectors, and must zero out
4301 // elements that come from V2 in the first mask, and V1 in the second mask
4302 // so that we can OR them together.
4303 bool TwoInputs = !(V1Only || V2Only);
4304 for (unsigned i = 0; i != 16; ++i) {
4305 int EltIdx = MaskVals[i];
4306 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4307 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4310 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4312 // If all the elements are from V2, assign it to V1 and return after
4313 // building the first pshufb.
4316 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4317 DAG.getNode(ISD::BUILD_VECTOR, dl,
4318 MVT::v16i8, &pshufbMask[0], 16));
4322 // Calculate the shuffle mask for the second input, shuffle it, and
4323 // OR it with the first shuffled input.
4325 for (unsigned i = 0; i != 16; ++i) {
4326 int EltIdx = MaskVals[i];
4328 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4331 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4333 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4334 DAG.getNode(ISD::BUILD_VECTOR, dl,
4335 MVT::v16i8, &pshufbMask[0], 16));
4336 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4339 // No SSSE3 - Calculate in place words and then fix all out of place words
4340 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4341 // the 16 different words that comprise the two doublequadword input vectors.
4342 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4343 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4344 SDValue NewV = V2Only ? V2 : V1;
4345 for (int i = 0; i != 8; ++i) {
4346 int Elt0 = MaskVals[i*2];
4347 int Elt1 = MaskVals[i*2+1];
4349 // This word of the result is all undef, skip it.
4350 if (Elt0 < 0 && Elt1 < 0)
4353 // This word of the result is already in the correct place, skip it.
4354 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4356 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4359 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4360 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4363 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4364 // using a single extract together, load it and store it.
4365 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4366 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4367 DAG.getIntPtrConstant(Elt1 / 2));
4368 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4369 DAG.getIntPtrConstant(i));
4373 // If Elt1 is defined, extract it from the appropriate source. If the
4374 // source byte is not also odd, shift the extracted word left 8 bits
4375 // otherwise clear the bottom 8 bits if we need to do an or.
4377 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4378 DAG.getIntPtrConstant(Elt1 / 2));
4379 if ((Elt1 & 1) == 0)
4380 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4381 DAG.getConstant(8, TLI.getShiftAmountTy()));
4383 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4384 DAG.getConstant(0xFF00, MVT::i16));
4386 // If Elt0 is defined, extract it from the appropriate source. If the
4387 // source byte is not also even, shift the extracted word right 8 bits. If
4388 // Elt1 was also defined, OR the extracted values together before
4389 // inserting them in the result.
4391 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4392 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4393 if ((Elt0 & 1) != 0)
4394 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4395 DAG.getConstant(8, TLI.getShiftAmountTy()));
4397 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4398 DAG.getConstant(0x00FF, MVT::i16));
4399 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4402 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4403 DAG.getIntPtrConstant(i));
4405 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4408 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4409 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4410 /// done when every pair / quad of shuffle mask elements point to elements in
4411 /// the right sequence. e.g.
4412 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4414 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4416 const TargetLowering &TLI, DebugLoc dl) {
4417 EVT VT = SVOp->getValueType(0);
4418 SDValue V1 = SVOp->getOperand(0);
4419 SDValue V2 = SVOp->getOperand(1);
4420 unsigned NumElems = VT.getVectorNumElements();
4421 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4422 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4423 EVT MaskEltVT = MaskVT.getVectorElementType();
4425 switch (VT.getSimpleVT().SimpleTy) {
4426 default: assert(false && "Unexpected!");
4427 case MVT::v4f32: NewVT = MVT::v2f64; break;
4428 case MVT::v4i32: NewVT = MVT::v2i64; break;
4429 case MVT::v8i16: NewVT = MVT::v4i32; break;
4430 case MVT::v16i8: NewVT = MVT::v4i32; break;
4433 if (NewWidth == 2) {
4439 int Scale = NumElems / NewWidth;
4440 SmallVector<int, 8> MaskVec;
4441 for (unsigned i = 0; i < NumElems; i += Scale) {
4443 for (int j = 0; j < Scale; ++j) {
4444 int EltIdx = SVOp->getMaskElt(i+j);
4448 StartIdx = EltIdx - (EltIdx % Scale);
4449 if (EltIdx != StartIdx + j)
4453 MaskVec.push_back(-1);
4455 MaskVec.push_back(StartIdx / Scale);
4458 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4459 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4460 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4463 /// getVZextMovL - Return a zero-extending vector move low node.
4465 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4466 SDValue SrcOp, SelectionDAG &DAG,
4467 const X86Subtarget *Subtarget, DebugLoc dl) {
4468 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4469 LoadSDNode *LD = NULL;
4470 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4471 LD = dyn_cast<LoadSDNode>(SrcOp);
4473 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4475 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4476 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4477 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4478 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4479 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4481 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4482 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4483 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4484 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4492 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4493 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4494 DAG.getNode(ISD::BIT_CONVERT, dl,
4498 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4501 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4502 SDValue V1 = SVOp->getOperand(0);
4503 SDValue V2 = SVOp->getOperand(1);
4504 DebugLoc dl = SVOp->getDebugLoc();
4505 EVT VT = SVOp->getValueType(0);
4507 SmallVector<std::pair<int, int>, 8> Locs;
4509 SmallVector<int, 8> Mask1(4U, -1);
4510 SmallVector<int, 8> PermMask;
4511 SVOp->getMask(PermMask);
4515 for (unsigned i = 0; i != 4; ++i) {
4516 int Idx = PermMask[i];
4518 Locs[i] = std::make_pair(-1, -1);
4520 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4522 Locs[i] = std::make_pair(0, NumLo);
4526 Locs[i] = std::make_pair(1, NumHi);
4528 Mask1[2+NumHi] = Idx;
4534 if (NumLo <= 2 && NumHi <= 2) {
4535 // If no more than two elements come from either vector. This can be
4536 // implemented with two shuffles. First shuffle gather the elements.
4537 // The second shuffle, which takes the first shuffle as both of its
4538 // vector operands, put the elements into the right order.
4539 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4541 SmallVector<int, 8> Mask2(4U, -1);
4543 for (unsigned i = 0; i != 4; ++i) {
4544 if (Locs[i].first == -1)
4547 unsigned Idx = (i < 2) ? 0 : 4;
4548 Idx += Locs[i].first * 2 + Locs[i].second;
4553 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4554 } else if (NumLo == 3 || NumHi == 3) {
4555 // Otherwise, we must have three elements from one vector, call it X, and
4556 // one element from the other, call it Y. First, use a shufps to build an
4557 // intermediate vector with the one element from Y and the element from X
4558 // that will be in the same half in the final destination (the indexes don't
4559 // matter). Then, use a shufps to build the final vector, taking the half
4560 // containing the element from Y from the intermediate, and the other half
4563 // Normalize it so the 3 elements come from V1.
4564 CommuteVectorShuffleMask(PermMask, VT);
4568 // Find the element from V2.
4570 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4571 int Val = PermMask[HiIndex];
4578 Mask1[0] = PermMask[HiIndex];
4580 Mask1[2] = PermMask[HiIndex^1];
4582 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4585 Mask1[0] = PermMask[0];
4586 Mask1[1] = PermMask[1];
4587 Mask1[2] = HiIndex & 1 ? 6 : 4;
4588 Mask1[3] = HiIndex & 1 ? 4 : 6;
4589 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4591 Mask1[0] = HiIndex & 1 ? 2 : 0;
4592 Mask1[1] = HiIndex & 1 ? 0 : 2;
4593 Mask1[2] = PermMask[2];
4594 Mask1[3] = PermMask[3];
4599 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4603 // Break it into (shuffle shuffle_hi, shuffle_lo).
4605 SmallVector<int,8> LoMask(4U, -1);
4606 SmallVector<int,8> HiMask(4U, -1);
4608 SmallVector<int,8> *MaskPtr = &LoMask;
4609 unsigned MaskIdx = 0;
4612 for (unsigned i = 0; i != 4; ++i) {
4619 int Idx = PermMask[i];
4621 Locs[i] = std::make_pair(-1, -1);
4622 } else if (Idx < 4) {
4623 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4624 (*MaskPtr)[LoIdx] = Idx;
4627 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4628 (*MaskPtr)[HiIdx] = Idx;
4633 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4634 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4635 SmallVector<int, 8> MaskOps;
4636 for (unsigned i = 0; i != 4; ++i) {
4637 if (Locs[i].first == -1) {
4638 MaskOps.push_back(-1);
4640 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4641 MaskOps.push_back(Idx);
4644 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4648 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4649 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4650 SDValue V1 = Op.getOperand(0);
4651 SDValue V2 = Op.getOperand(1);
4652 EVT VT = Op.getValueType();
4653 DebugLoc dl = Op.getDebugLoc();
4654 unsigned NumElems = VT.getVectorNumElements();
4655 bool isMMX = VT.getSizeInBits() == 64;
4656 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4657 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4658 bool V1IsSplat = false;
4659 bool V2IsSplat = false;
4661 if (isZeroShuffle(SVOp))
4662 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4664 // Promote splats to v4f32.
4665 if (SVOp->isSplat()) {
4666 if (isMMX || NumElems < 4)
4668 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4671 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4673 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4674 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4675 if (NewOp.getNode())
4676 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4677 LowerVECTOR_SHUFFLE(NewOp, DAG));
4678 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4679 // FIXME: Figure out a cleaner way to do this.
4680 // Try to make use of movq to zero out the top part.
4681 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4682 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4683 if (NewOp.getNode()) {
4684 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4685 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4686 DAG, Subtarget, dl);
4688 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4689 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4690 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4691 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4692 DAG, Subtarget, dl);
4696 if (X86::isPSHUFDMask(SVOp))
4699 // Check if this can be converted into a logical shift.
4700 bool isLeft = false;
4703 bool isShift = getSubtarget()->hasSSE2() &&
4704 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4705 if (isShift && ShVal.hasOneUse()) {
4706 // If the shifted value has multiple uses, it may be cheaper to use
4707 // v_set0 + movlhps or movhlps, etc.
4708 EVT EltVT = VT.getVectorElementType();
4709 ShAmt *= EltVT.getSizeInBits();
4710 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4713 if (X86::isMOVLMask(SVOp)) {
4716 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4717 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4722 // FIXME: fold these into legal mask.
4723 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4724 X86::isMOVSLDUPMask(SVOp) ||
4725 X86::isMOVHLPSMask(SVOp) ||
4726 X86::isMOVLHPSMask(SVOp) ||
4727 X86::isMOVLPMask(SVOp)))
4730 if (ShouldXformToMOVHLPS(SVOp) ||
4731 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4732 return CommuteVectorShuffle(SVOp, DAG);
4735 // No better options. Use a vshl / vsrl.
4736 EVT EltVT = VT.getVectorElementType();
4737 ShAmt *= EltVT.getSizeInBits();
4738 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4741 bool Commuted = false;
4742 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4743 // 1,1,1,1 -> v8i16 though.
4744 V1IsSplat = isSplatVector(V1.getNode());
4745 V2IsSplat = isSplatVector(V2.getNode());
4747 // Canonicalize the splat or undef, if present, to be on the RHS.
4748 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4749 Op = CommuteVectorShuffle(SVOp, DAG);
4750 SVOp = cast<ShuffleVectorSDNode>(Op);
4751 V1 = SVOp->getOperand(0);
4752 V2 = SVOp->getOperand(1);
4753 std::swap(V1IsSplat, V2IsSplat);
4754 std::swap(V1IsUndef, V2IsUndef);
4758 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4759 // Shuffling low element of v1 into undef, just return v1.
4762 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4763 // the instruction selector will not match, so get a canonical MOVL with
4764 // swapped operands to undo the commute.
4765 return getMOVL(DAG, dl, VT, V2, V1);
4768 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4769 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4770 X86::isUNPCKLMask(SVOp) ||
4771 X86::isUNPCKHMask(SVOp))
4775 // Normalize mask so all entries that point to V2 points to its first
4776 // element then try to match unpck{h|l} again. If match, return a
4777 // new vector_shuffle with the corrected mask.
4778 SDValue NewMask = NormalizeMask(SVOp, DAG);
4779 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4780 if (NSVOp != SVOp) {
4781 if (X86::isUNPCKLMask(NSVOp, true)) {
4783 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4790 // Commute is back and try unpck* again.
4791 // FIXME: this seems wrong.
4792 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4793 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4794 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4795 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4796 X86::isUNPCKLMask(NewSVOp) ||
4797 X86::isUNPCKHMask(NewSVOp))
4801 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4803 // Normalize the node to match x86 shuffle ops if needed
4804 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4805 return CommuteVectorShuffle(SVOp, DAG);
4807 // Check for legal shuffle and return?
4808 SmallVector<int, 16> PermMask;
4809 SVOp->getMask(PermMask);
4810 if (isShuffleMaskLegal(PermMask, VT))
4813 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4814 if (VT == MVT::v8i16) {
4815 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4816 if (NewOp.getNode())
4820 if (VT == MVT::v16i8) {
4821 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4822 if (NewOp.getNode())
4826 // Handle all 4 wide cases with a number of shuffles except for MMX.
4827 if (NumElems == 4 && !isMMX)
4828 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4834 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4835 SelectionDAG &DAG) const {
4836 EVT VT = Op.getValueType();
4837 DebugLoc dl = Op.getDebugLoc();
4838 if (VT.getSizeInBits() == 8) {
4839 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4840 Op.getOperand(0), Op.getOperand(1));
4841 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4842 DAG.getValueType(VT));
4843 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4844 } else if (VT.getSizeInBits() == 16) {
4845 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4846 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4848 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4849 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4850 DAG.getNode(ISD::BIT_CONVERT, dl,
4854 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4855 Op.getOperand(0), Op.getOperand(1));
4856 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4857 DAG.getValueType(VT));
4858 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4859 } else if (VT == MVT::f32) {
4860 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4861 // the result back to FR32 register. It's only worth matching if the
4862 // result has a single use which is a store or a bitcast to i32. And in
4863 // the case of a store, it's not worth it if the index is a constant 0,
4864 // because a MOVSSmr can be used instead, which is smaller and faster.
4865 if (!Op.hasOneUse())
4867 SDNode *User = *Op.getNode()->use_begin();
4868 if ((User->getOpcode() != ISD::STORE ||
4869 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4870 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4871 (User->getOpcode() != ISD::BIT_CONVERT ||
4872 User->getValueType(0) != MVT::i32))
4874 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4875 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4878 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4879 } else if (VT == MVT::i32) {
4880 // ExtractPS works with constant index.
4881 if (isa<ConstantSDNode>(Op.getOperand(1)))
4889 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4890 SelectionDAG &DAG) const {
4891 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4894 if (Subtarget->hasSSE41()) {
4895 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4900 EVT VT = Op.getValueType();
4901 DebugLoc dl = Op.getDebugLoc();
4902 // TODO: handle v16i8.
4903 if (VT.getSizeInBits() == 16) {
4904 SDValue Vec = Op.getOperand(0);
4905 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4907 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4908 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4909 DAG.getNode(ISD::BIT_CONVERT, dl,
4912 // Transform it so it match pextrw which produces a 32-bit result.
4913 EVT EltVT = MVT::i32;
4914 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4915 Op.getOperand(0), Op.getOperand(1));
4916 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4917 DAG.getValueType(VT));
4918 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4919 } else if (VT.getSizeInBits() == 32) {
4920 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4924 // SHUFPS the element to the lowest double word, then movss.
4925 int Mask[4] = { Idx, -1, -1, -1 };
4926 EVT VVT = Op.getOperand(0).getValueType();
4927 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4928 DAG.getUNDEF(VVT), Mask);
4929 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4930 DAG.getIntPtrConstant(0));
4931 } else if (VT.getSizeInBits() == 64) {
4932 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4933 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4934 // to match extract_elt for f64.
4935 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4939 // UNPCKHPD the element to the lowest double word, then movsd.
4940 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4941 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4942 int Mask[2] = { 1, -1 };
4943 EVT VVT = Op.getOperand(0).getValueType();
4944 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4945 DAG.getUNDEF(VVT), Mask);
4946 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4947 DAG.getIntPtrConstant(0));
4954 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4955 SelectionDAG &DAG) const {
4956 EVT VT = Op.getValueType();
4957 EVT EltVT = VT.getVectorElementType();
4958 DebugLoc dl = Op.getDebugLoc();
4960 SDValue N0 = Op.getOperand(0);
4961 SDValue N1 = Op.getOperand(1);
4962 SDValue N2 = Op.getOperand(2);
4964 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4965 isa<ConstantSDNode>(N2)) {
4967 if (VT == MVT::v8i16)
4968 Opc = X86ISD::PINSRW;
4969 else if (VT == MVT::v4i16)
4970 Opc = X86ISD::MMX_PINSRW;
4971 else if (VT == MVT::v16i8)
4972 Opc = X86ISD::PINSRB;
4974 Opc = X86ISD::PINSRB;
4976 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4978 if (N1.getValueType() != MVT::i32)
4979 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4980 if (N2.getValueType() != MVT::i32)
4981 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4982 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4983 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4984 // Bits [7:6] of the constant are the source select. This will always be
4985 // zero here. The DAG Combiner may combine an extract_elt index into these
4986 // bits. For example (insert (extract, 3), 2) could be matched by putting
4987 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4988 // Bits [5:4] of the constant are the destination select. This is the
4989 // value of the incoming immediate.
4990 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4991 // combine either bitwise AND or insert of float 0.0 to set these bits.
4992 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4993 // Create this as a scalar to vector..
4994 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4995 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4996 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4997 // PINSR* works with constant index.
5004 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5005 EVT VT = Op.getValueType();
5006 EVT EltVT = VT.getVectorElementType();
5008 if (Subtarget->hasSSE41())
5009 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5011 if (EltVT == MVT::i8)
5014 DebugLoc dl = Op.getDebugLoc();
5015 SDValue N0 = Op.getOperand(0);
5016 SDValue N1 = Op.getOperand(1);
5017 SDValue N2 = Op.getOperand(2);
5019 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5020 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5021 // as its second argument.
5022 if (N1.getValueType() != MVT::i32)
5023 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5024 if (N2.getValueType() != MVT::i32)
5025 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5026 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5027 dl, VT, N0, N1, N2);
5033 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5034 DebugLoc dl = Op.getDebugLoc();
5035 if (Op.getValueType() == MVT::v2f32)
5036 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5037 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5038 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5039 Op.getOperand(0))));
5041 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5042 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5044 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5045 EVT VT = MVT::v2i32;
5046 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5053 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5054 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5057 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5058 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5059 // one of the above mentioned nodes. It has to be wrapped because otherwise
5060 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5061 // be used to form addressing mode. These wrapped nodes will be selected
5064 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5065 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5067 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5069 unsigned char OpFlag = 0;
5070 unsigned WrapperKind = X86ISD::Wrapper;
5071 CodeModel::Model M = getTargetMachine().getCodeModel();
5073 if (Subtarget->isPICStyleRIPRel() &&
5074 (M == CodeModel::Small || M == CodeModel::Kernel))
5075 WrapperKind = X86ISD::WrapperRIP;
5076 else if (Subtarget->isPICStyleGOT())
5077 OpFlag = X86II::MO_GOTOFF;
5078 else if (Subtarget->isPICStyleStubPIC())
5079 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5081 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5083 CP->getOffset(), OpFlag);
5084 DebugLoc DL = CP->getDebugLoc();
5085 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5086 // With PIC, the address is actually $g + Offset.
5088 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5089 DAG.getNode(X86ISD::GlobalBaseReg,
5090 DebugLoc(), getPointerTy()),
5097 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5098 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5100 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5102 unsigned char OpFlag = 0;
5103 unsigned WrapperKind = X86ISD::Wrapper;
5104 CodeModel::Model M = getTargetMachine().getCodeModel();
5106 if (Subtarget->isPICStyleRIPRel() &&
5107 (M == CodeModel::Small || M == CodeModel::Kernel))
5108 WrapperKind = X86ISD::WrapperRIP;
5109 else if (Subtarget->isPICStyleGOT())
5110 OpFlag = X86II::MO_GOTOFF;
5111 else if (Subtarget->isPICStyleStubPIC())
5112 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5114 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5116 DebugLoc DL = JT->getDebugLoc();
5117 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5119 // With PIC, the address is actually $g + Offset.
5121 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5122 DAG.getNode(X86ISD::GlobalBaseReg,
5123 DebugLoc(), getPointerTy()),
5131 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5132 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5134 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5136 unsigned char OpFlag = 0;
5137 unsigned WrapperKind = X86ISD::Wrapper;
5138 CodeModel::Model M = getTargetMachine().getCodeModel();
5140 if (Subtarget->isPICStyleRIPRel() &&
5141 (M == CodeModel::Small || M == CodeModel::Kernel))
5142 WrapperKind = X86ISD::WrapperRIP;
5143 else if (Subtarget->isPICStyleGOT())
5144 OpFlag = X86II::MO_GOTOFF;
5145 else if (Subtarget->isPICStyleStubPIC())
5146 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5148 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5150 DebugLoc DL = Op.getDebugLoc();
5151 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5154 // With PIC, the address is actually $g + Offset.
5155 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5156 !Subtarget->is64Bit()) {
5157 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5158 DAG.getNode(X86ISD::GlobalBaseReg,
5159 DebugLoc(), getPointerTy()),
5167 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5168 // Create the TargetBlockAddressAddress node.
5169 unsigned char OpFlags =
5170 Subtarget->ClassifyBlockAddressReference();
5171 CodeModel::Model M = getTargetMachine().getCodeModel();
5172 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5173 DebugLoc dl = Op.getDebugLoc();
5174 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5175 /*isTarget=*/true, OpFlags);
5177 if (Subtarget->isPICStyleRIPRel() &&
5178 (M == CodeModel::Small || M == CodeModel::Kernel))
5179 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5181 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5183 // With PIC, the address is actually $g + Offset.
5184 if (isGlobalRelativeToPICBase(OpFlags)) {
5185 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5186 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5194 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5196 SelectionDAG &DAG) const {
5197 // Create the TargetGlobalAddress node, folding in the constant
5198 // offset if it is legal.
5199 unsigned char OpFlags =
5200 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5201 CodeModel::Model M = getTargetMachine().getCodeModel();
5203 if (OpFlags == X86II::MO_NO_FLAG &&
5204 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5205 // A direct static reference to a global.
5206 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5209 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5212 if (Subtarget->isPICStyleRIPRel() &&
5213 (M == CodeModel::Small || M == CodeModel::Kernel))
5214 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5216 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5218 // With PIC, the address is actually $g + Offset.
5219 if (isGlobalRelativeToPICBase(OpFlags)) {
5220 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5221 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5225 // For globals that require a load from a stub to get the address, emit the
5227 if (isGlobalStubReference(OpFlags))
5228 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5229 PseudoSourceValue::getGOT(), 0, false, false, 0);
5231 // If there was a non-zero offset that we didn't fold, create an explicit
5234 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5235 DAG.getConstant(Offset, getPointerTy()));
5241 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5242 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5243 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5244 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5248 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5249 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5250 unsigned char OperandFlags) {
5251 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5252 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5253 DebugLoc dl = GA->getDebugLoc();
5254 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5255 GA->getValueType(0),
5259 SDValue Ops[] = { Chain, TGA, *InFlag };
5260 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5262 SDValue Ops[] = { Chain, TGA };
5263 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5266 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5267 MFI->setHasCalls(true);
5269 SDValue Flag = Chain.getValue(1);
5270 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5273 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5275 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5278 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5279 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5280 DAG.getNode(X86ISD::GlobalBaseReg,
5281 DebugLoc(), PtrVT), InFlag);
5282 InFlag = Chain.getValue(1);
5284 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5287 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5289 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5291 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5292 X86::RAX, X86II::MO_TLSGD);
5295 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5296 // "local exec" model.
5297 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5298 const EVT PtrVT, TLSModel::Model model,
5300 DebugLoc dl = GA->getDebugLoc();
5301 // Get the Thread Pointer
5302 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5304 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5307 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5308 NULL, 0, false, false, 0);
5310 unsigned char OperandFlags = 0;
5311 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5313 unsigned WrapperKind = X86ISD::Wrapper;
5314 if (model == TLSModel::LocalExec) {
5315 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5316 } else if (is64Bit) {
5317 assert(model == TLSModel::InitialExec);
5318 OperandFlags = X86II::MO_GOTTPOFF;
5319 WrapperKind = X86ISD::WrapperRIP;
5321 assert(model == TLSModel::InitialExec);
5322 OperandFlags = X86II::MO_INDNTPOFF;
5325 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5327 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5328 GA->getOffset(), OperandFlags);
5329 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5331 if (model == TLSModel::InitialExec)
5332 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5333 PseudoSourceValue::getGOT(), 0, false, false, 0);
5335 // The address of the thread local variable is the add of the thread
5336 // pointer with the offset of the variable.
5337 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5341 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5342 // TODO: implement the "local dynamic" model
5343 // TODO: implement the "initial exec"model for pic executables
5344 assert(Subtarget->isTargetELF() &&
5345 "TLS not implemented for non-ELF targets");
5346 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5347 const GlobalValue *GV = GA->getGlobal();
5349 // If GV is an alias then use the aliasee for determining
5350 // thread-localness.
5351 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5352 GV = GA->resolveAliasedGlobal(false);
5354 TLSModel::Model model = getTLSModel(GV,
5355 getTargetMachine().getRelocationModel());
5358 case TLSModel::GeneralDynamic:
5359 case TLSModel::LocalDynamic: // not implemented
5360 if (Subtarget->is64Bit())
5361 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5362 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5364 case TLSModel::InitialExec:
5365 case TLSModel::LocalExec:
5366 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5367 Subtarget->is64Bit());
5370 llvm_unreachable("Unreachable");
5375 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5376 /// take a 2 x i32 value to shift plus a shift amount.
5377 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5378 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5379 EVT VT = Op.getValueType();
5380 unsigned VTBits = VT.getSizeInBits();
5381 DebugLoc dl = Op.getDebugLoc();
5382 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5383 SDValue ShOpLo = Op.getOperand(0);
5384 SDValue ShOpHi = Op.getOperand(1);
5385 SDValue ShAmt = Op.getOperand(2);
5386 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5387 DAG.getConstant(VTBits - 1, MVT::i8))
5388 : DAG.getConstant(0, VT);
5391 if (Op.getOpcode() == ISD::SHL_PARTS) {
5392 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5393 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5395 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5396 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5399 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5400 DAG.getConstant(VTBits, MVT::i8));
5401 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5402 AndNode, DAG.getConstant(0, MVT::i8));
5405 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5406 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5407 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5409 if (Op.getOpcode() == ISD::SHL_PARTS) {
5410 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5411 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5413 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5414 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5417 SDValue Ops[2] = { Lo, Hi };
5418 return DAG.getMergeValues(Ops, 2, dl);
5421 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5422 SelectionDAG &DAG) const {
5423 EVT SrcVT = Op.getOperand(0).getValueType();
5425 if (SrcVT.isVector()) {
5426 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5432 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5433 "Unknown SINT_TO_FP to lower!");
5435 // These are really Legal; return the operand so the caller accepts it as
5437 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5439 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5440 Subtarget->is64Bit()) {
5444 DebugLoc dl = Op.getDebugLoc();
5445 unsigned Size = SrcVT.getSizeInBits()/8;
5446 MachineFunction &MF = DAG.getMachineFunction();
5447 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5448 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5449 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5451 PseudoSourceValue::getFixedStack(SSFI), 0,
5453 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5456 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5458 SelectionDAG &DAG) const {
5460 DebugLoc dl = Op.getDebugLoc();
5462 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5464 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5466 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5467 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5468 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5469 Tys, Ops, array_lengthof(Ops));
5472 Chain = Result.getValue(1);
5473 SDValue InFlag = Result.getValue(2);
5475 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5476 // shouldn't be necessary except that RFP cannot be live across
5477 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5478 MachineFunction &MF = DAG.getMachineFunction();
5479 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5480 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5481 Tys = DAG.getVTList(MVT::Other);
5483 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5485 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5486 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5487 PseudoSourceValue::getFixedStack(SSFI), 0,
5494 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5495 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5496 SelectionDAG &DAG) const {
5497 // This algorithm is not obvious. Here it is in C code, more or less:
5499 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5500 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5501 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5503 // Copy ints to xmm registers.
5504 __m128i xh = _mm_cvtsi32_si128( hi );
5505 __m128i xl = _mm_cvtsi32_si128( lo );
5507 // Combine into low half of a single xmm register.
5508 __m128i x = _mm_unpacklo_epi32( xh, xl );
5512 // Merge in appropriate exponents to give the integer bits the right
5514 x = _mm_unpacklo_epi32( x, exp );
5516 // Subtract away the biases to deal with the IEEE-754 double precision
5518 d = _mm_sub_pd( (__m128d) x, bias );
5520 // All conversions up to here are exact. The correctly rounded result is
5521 // calculated using the current rounding mode using the following
5523 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5524 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5525 // store doesn't really need to be here (except
5526 // maybe to zero the other double)
5531 DebugLoc dl = Op.getDebugLoc();
5532 LLVMContext *Context = DAG.getContext();
5534 // Build some magic constants.
5535 std::vector<Constant*> CV0;
5536 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5537 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5538 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5539 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5540 Constant *C0 = ConstantVector::get(CV0);
5541 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5543 std::vector<Constant*> CV1;
5545 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5547 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5548 Constant *C1 = ConstantVector::get(CV1);
5549 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5551 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5552 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5554 DAG.getIntPtrConstant(1)));
5555 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5556 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5558 DAG.getIntPtrConstant(0)));
5559 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5560 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5561 PseudoSourceValue::getConstantPool(), 0,
5563 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5564 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5565 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5566 PseudoSourceValue::getConstantPool(), 0,
5568 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5570 // Add the halves; easiest way is to swap them into another reg first.
5571 int ShufMask[2] = { 1, -1 };
5572 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5573 DAG.getUNDEF(MVT::v2f64), ShufMask);
5574 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5575 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5576 DAG.getIntPtrConstant(0));
5579 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5580 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5581 SelectionDAG &DAG) const {
5582 DebugLoc dl = Op.getDebugLoc();
5583 // FP constant to bias correct the final result.
5584 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5587 // Load the 32-bit value into an XMM register.
5588 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5589 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5591 DAG.getIntPtrConstant(0)));
5593 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5594 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5595 DAG.getIntPtrConstant(0));
5597 // Or the load with the bias.
5598 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5599 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5600 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5602 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5603 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5604 MVT::v2f64, Bias)));
5605 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5606 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5607 DAG.getIntPtrConstant(0));
5609 // Subtract the bias.
5610 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5612 // Handle final rounding.
5613 EVT DestVT = Op.getValueType();
5615 if (DestVT.bitsLT(MVT::f64)) {
5616 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5617 DAG.getIntPtrConstant(0));
5618 } else if (DestVT.bitsGT(MVT::f64)) {
5619 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5622 // Handle final rounding.
5626 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5627 SelectionDAG &DAG) const {
5628 SDValue N0 = Op.getOperand(0);
5629 DebugLoc dl = Op.getDebugLoc();
5631 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5632 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5633 // the optimization here.
5634 if (DAG.SignBitIsZero(N0))
5635 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5637 EVT SrcVT = N0.getValueType();
5638 if (SrcVT == MVT::i64) {
5639 // We only handle SSE2 f64 target here; caller can expand the rest.
5640 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5643 return LowerUINT_TO_FP_i64(Op, DAG);
5644 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5645 return LowerUINT_TO_FP_i32(Op, DAG);
5648 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5650 // Make a 64-bit buffer, and use it to build an FILD.
5651 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5652 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5653 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5654 getPointerTy(), StackSlot, WordOff);
5655 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5656 StackSlot, NULL, 0, false, false, 0);
5657 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5658 OffsetSlot, NULL, 0, false, false, 0);
5659 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5662 std::pair<SDValue,SDValue> X86TargetLowering::
5663 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5664 DebugLoc dl = Op.getDebugLoc();
5666 EVT DstTy = Op.getValueType();
5669 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5673 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5674 DstTy.getSimpleVT() >= MVT::i16 &&
5675 "Unknown FP_TO_SINT to lower!");
5677 // These are really Legal.
5678 if (DstTy == MVT::i32 &&
5679 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5680 return std::make_pair(SDValue(), SDValue());
5681 if (Subtarget->is64Bit() &&
5682 DstTy == MVT::i64 &&
5683 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5684 return std::make_pair(SDValue(), SDValue());
5686 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5688 MachineFunction &MF = DAG.getMachineFunction();
5689 unsigned MemSize = DstTy.getSizeInBits()/8;
5690 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5691 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5694 switch (DstTy.getSimpleVT().SimpleTy) {
5695 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5696 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5697 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5698 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5701 SDValue Chain = DAG.getEntryNode();
5702 SDValue Value = Op.getOperand(0);
5703 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5704 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5705 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5706 PseudoSourceValue::getFixedStack(SSFI), 0,
5708 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5710 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5712 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5713 Chain = Value.getValue(1);
5714 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5715 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5718 // Build the FP_TO_INT*_IN_MEM
5719 SDValue Ops[] = { Chain, Value, StackSlot };
5720 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5722 return std::make_pair(FIST, StackSlot);
5725 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5726 SelectionDAG &DAG) const {
5727 if (Op.getValueType().isVector()) {
5728 if (Op.getValueType() == MVT::v2i32 &&
5729 Op.getOperand(0).getValueType() == MVT::v2f64) {
5735 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5736 SDValue FIST = Vals.first, StackSlot = Vals.second;
5737 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5738 if (FIST.getNode() == 0) return Op;
5741 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5742 FIST, StackSlot, NULL, 0, false, false, 0);
5745 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5746 SelectionDAG &DAG) const {
5747 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5748 SDValue FIST = Vals.first, StackSlot = Vals.second;
5749 assert(FIST.getNode() && "Unexpected failure");
5752 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5753 FIST, StackSlot, NULL, 0, false, false, 0);
5756 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5757 SelectionDAG &DAG) const {
5758 LLVMContext *Context = DAG.getContext();
5759 DebugLoc dl = Op.getDebugLoc();
5760 EVT VT = Op.getValueType();
5763 EltVT = VT.getVectorElementType();
5764 std::vector<Constant*> CV;
5765 if (EltVT == MVT::f64) {
5766 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5770 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5776 Constant *C = ConstantVector::get(CV);
5777 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5778 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5779 PseudoSourceValue::getConstantPool(), 0,
5781 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5784 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5785 LLVMContext *Context = DAG.getContext();
5786 DebugLoc dl = Op.getDebugLoc();
5787 EVT VT = Op.getValueType();
5790 EltVT = VT.getVectorElementType();
5791 std::vector<Constant*> CV;
5792 if (EltVT == MVT::f64) {
5793 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5797 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5803 Constant *C = ConstantVector::get(CV);
5804 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5805 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5806 PseudoSourceValue::getConstantPool(), 0,
5808 if (VT.isVector()) {
5809 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5810 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5811 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5813 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5815 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5819 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5820 LLVMContext *Context = DAG.getContext();
5821 SDValue Op0 = Op.getOperand(0);
5822 SDValue Op1 = Op.getOperand(1);
5823 DebugLoc dl = Op.getDebugLoc();
5824 EVT VT = Op.getValueType();
5825 EVT SrcVT = Op1.getValueType();
5827 // If second operand is smaller, extend it first.
5828 if (SrcVT.bitsLT(VT)) {
5829 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5832 // And if it is bigger, shrink it first.
5833 if (SrcVT.bitsGT(VT)) {
5834 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5838 // At this point the operands and the result should have the same
5839 // type, and that won't be f80 since that is not custom lowered.
5841 // First get the sign bit of second operand.
5842 std::vector<Constant*> CV;
5843 if (SrcVT == MVT::f64) {
5844 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5845 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5847 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5848 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5849 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5850 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5852 Constant *C = ConstantVector::get(CV);
5853 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5854 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5855 PseudoSourceValue::getConstantPool(), 0,
5857 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5859 // Shift sign bit right or left if the two operands have different types.
5860 if (SrcVT.bitsGT(VT)) {
5861 // Op0 is MVT::f32, Op1 is MVT::f64.
5862 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5863 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5864 DAG.getConstant(32, MVT::i32));
5865 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5866 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5867 DAG.getIntPtrConstant(0));
5870 // Clear first operand sign bit.
5872 if (VT == MVT::f64) {
5873 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5874 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5876 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5877 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5878 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5879 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5881 C = ConstantVector::get(CV);
5882 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5883 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5884 PseudoSourceValue::getConstantPool(), 0,
5886 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5888 // Or the value with the sign bit.
5889 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5892 /// Emit nodes that will be selected as "test Op0,Op0", or something
5894 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5895 SelectionDAG &DAG) const {
5896 DebugLoc dl = Op.getDebugLoc();
5898 // CF and OF aren't always set the way we want. Determine which
5899 // of these we need.
5900 bool NeedCF = false;
5901 bool NeedOF = false;
5903 case X86::COND_A: case X86::COND_AE:
5904 case X86::COND_B: case X86::COND_BE:
5907 case X86::COND_G: case X86::COND_GE:
5908 case X86::COND_L: case X86::COND_LE:
5909 case X86::COND_O: case X86::COND_NO:
5915 // See if we can use the EFLAGS value from the operand instead of
5916 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5917 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5918 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5919 unsigned Opcode = 0;
5920 unsigned NumOperands = 0;
5921 switch (Op.getNode()->getOpcode()) {
5923 // Due to an isel shortcoming, be conservative if this add is likely to
5924 // be selected as part of a load-modify-store instruction. When the root
5925 // node in a match is a store, isel doesn't know how to remap non-chain
5926 // non-flag uses of other nodes in the match, such as the ADD in this
5927 // case. This leads to the ADD being left around and reselected, with
5928 // the result being two adds in the output.
5929 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5930 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5931 if (UI->getOpcode() == ISD::STORE)
5933 if (ConstantSDNode *C =
5934 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5935 // An add of one will be selected as an INC.
5936 if (C->getAPIntValue() == 1) {
5937 Opcode = X86ISD::INC;
5941 // An add of negative one (subtract of one) will be selected as a DEC.
5942 if (C->getAPIntValue().isAllOnesValue()) {
5943 Opcode = X86ISD::DEC;
5948 // Otherwise use a regular EFLAGS-setting add.
5949 Opcode = X86ISD::ADD;
5953 // If the primary and result isn't used, don't bother using X86ISD::AND,
5954 // because a TEST instruction will be better.
5955 bool NonFlagUse = false;
5956 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5957 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5959 unsigned UOpNo = UI.getOperandNo();
5960 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5961 // Look pass truncate.
5962 UOpNo = User->use_begin().getOperandNo();
5963 User = *User->use_begin();
5965 if (User->getOpcode() != ISD::BRCOND &&
5966 User->getOpcode() != ISD::SETCC &&
5967 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5979 // Due to the ISEL shortcoming noted above, be conservative if this op is
5980 // likely to be selected as part of a load-modify-store instruction.
5981 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5982 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5983 if (UI->getOpcode() == ISD::STORE)
5985 // Otherwise use a regular EFLAGS-setting instruction.
5986 switch (Op.getNode()->getOpcode()) {
5987 case ISD::SUB: Opcode = X86ISD::SUB; break;
5988 case ISD::OR: Opcode = X86ISD::OR; break;
5989 case ISD::XOR: Opcode = X86ISD::XOR; break;
5990 case ISD::AND: Opcode = X86ISD::AND; break;
5991 default: llvm_unreachable("unexpected operator!");
6002 return SDValue(Op.getNode(), 1);
6008 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6009 SmallVector<SDValue, 4> Ops;
6010 for (unsigned i = 0; i != NumOperands; ++i)
6011 Ops.push_back(Op.getOperand(i));
6012 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6013 DAG.ReplaceAllUsesWith(Op, New);
6014 return SDValue(New.getNode(), 1);
6018 // Otherwise just emit a CMP with 0, which is the TEST pattern.
6019 if (Promote16Bit && Op.getValueType() == MVT::i16)
6020 Op = DAG.getNode(ISD::ANY_EXTEND, Op.getDebugLoc(), MVT::i32, Op);
6021 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6022 DAG.getConstant(0, Op.getValueType()));
6025 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6027 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6028 SelectionDAG &DAG) const {
6029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6030 if (C->getAPIntValue() == 0)
6031 return EmitTest(Op0, X86CC, DAG);
6033 DebugLoc dl = Op0.getDebugLoc();
6034 if (Promote16Bit && Op0.getValueType() == MVT::i16) {
6035 Op0 = DAG.getNode(ISD::ANY_EXTEND, Op0.getDebugLoc(), MVT::i32, Op0);
6036 Op1 = DAG.getNode(ISD::ANY_EXTEND, Op1.getDebugLoc(), MVT::i32, Op1);
6038 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6041 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6042 /// if it's possible.
6043 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
6044 DebugLoc dl, SelectionDAG &DAG) {
6045 SDValue Op0 = And.getOperand(0);
6046 SDValue Op1 = And.getOperand(1);
6047 if (Op0.getOpcode() == ISD::TRUNCATE)
6048 Op0 = Op0.getOperand(0);
6049 if (Op1.getOpcode() == ISD::TRUNCATE)
6050 Op1 = Op1.getOperand(0);
6053 if (Op1.getOpcode() == ISD::SHL) {
6054 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6055 if (And10C->getZExtValue() == 1) {
6057 RHS = Op1.getOperand(1);
6059 } else if (Op0.getOpcode() == ISD::SHL) {
6060 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6061 if (And00C->getZExtValue() == 1) {
6063 RHS = Op0.getOperand(1);
6065 } else if (Op1.getOpcode() == ISD::Constant) {
6066 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6067 SDValue AndLHS = Op0;
6068 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6069 LHS = AndLHS.getOperand(0);
6070 RHS = AndLHS.getOperand(1);
6074 if (LHS.getNode()) {
6075 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6076 // instruction. Since the shift amount is in-range-or-undefined, we know
6077 // that doing a bittest on the i32 value is ok. We extend to i32 because
6078 // the encoding for the i16 version is larger than the i32 version.
6079 // Also promote i16 to i32 for performance / code size reason.
6080 if (LHS.getValueType() == MVT::i8 ||
6081 (Promote16Bit && LHS.getValueType() == MVT::i16))
6082 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6084 // If the operand types disagree, extend the shift amount to match. Since
6085 // BT ignores high bits (like shifts) we can use anyextend.
6086 if (LHS.getValueType() != RHS.getValueType())
6087 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6089 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6090 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6091 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6092 DAG.getConstant(Cond, MVT::i8), BT);
6098 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6099 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6100 SDValue Op0 = Op.getOperand(0);
6101 SDValue Op1 = Op.getOperand(1);
6102 DebugLoc dl = Op.getDebugLoc();
6103 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6105 // Optimize to BT if possible.
6106 // Lower (X & (1 << N)) == 0 to BT(X, N).
6107 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6108 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6109 if (Op0.getOpcode() == ISD::AND &&
6111 Op1.getOpcode() == ISD::Constant &&
6112 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6113 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6114 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6115 if (NewSetCC.getNode())
6119 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6120 if (Op0.getOpcode() == X86ISD::SETCC &&
6121 Op1.getOpcode() == ISD::Constant &&
6122 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6123 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6124 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6125 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6126 bool Invert = (CC == ISD::SETNE) ^
6127 cast<ConstantSDNode>(Op1)->isNullValue();
6129 CCode = X86::GetOppositeBranchCondition(CCode);
6130 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6131 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6134 bool isFP = Op1.getValueType().isFloatingPoint();
6135 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6136 if (X86CC == X86::COND_INVALID)
6139 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6141 // Use sbb x, x to materialize carry bit into a GPR.
6142 if (X86CC == X86::COND_B)
6143 return DAG.getNode(ISD::AND, dl, MVT::i8,
6144 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6145 DAG.getConstant(X86CC, MVT::i8), Cond),
6146 DAG.getConstant(1, MVT::i8));
6148 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6149 DAG.getConstant(X86CC, MVT::i8), Cond);
6152 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6154 SDValue Op0 = Op.getOperand(0);
6155 SDValue Op1 = Op.getOperand(1);
6156 SDValue CC = Op.getOperand(2);
6157 EVT VT = Op.getValueType();
6158 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6159 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6160 DebugLoc dl = Op.getDebugLoc();
6164 EVT VT0 = Op0.getValueType();
6165 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6166 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6169 switch (SetCCOpcode) {
6172 case ISD::SETEQ: SSECC = 0; break;
6174 case ISD::SETGT: Swap = true; // Fallthrough
6176 case ISD::SETOLT: SSECC = 1; break;
6178 case ISD::SETGE: Swap = true; // Fallthrough
6180 case ISD::SETOLE: SSECC = 2; break;
6181 case ISD::SETUO: SSECC = 3; break;
6183 case ISD::SETNE: SSECC = 4; break;
6184 case ISD::SETULE: Swap = true;
6185 case ISD::SETUGE: SSECC = 5; break;
6186 case ISD::SETULT: Swap = true;
6187 case ISD::SETUGT: SSECC = 6; break;
6188 case ISD::SETO: SSECC = 7; break;
6191 std::swap(Op0, Op1);
6193 // In the two special cases we can't handle, emit two comparisons.
6195 if (SetCCOpcode == ISD::SETUEQ) {
6197 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6198 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6199 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6201 else if (SetCCOpcode == ISD::SETONE) {
6203 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6204 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6205 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6207 llvm_unreachable("Illegal FP comparison");
6209 // Handle all other FP comparisons here.
6210 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6213 // We are handling one of the integer comparisons here. Since SSE only has
6214 // GT and EQ comparisons for integer, swapping operands and multiple
6215 // operations may be required for some comparisons.
6216 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6217 bool Swap = false, Invert = false, FlipSigns = false;
6219 switch (VT.getSimpleVT().SimpleTy) {
6222 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6224 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6226 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6227 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6230 switch (SetCCOpcode) {
6232 case ISD::SETNE: Invert = true;
6233 case ISD::SETEQ: Opc = EQOpc; break;
6234 case ISD::SETLT: Swap = true;
6235 case ISD::SETGT: Opc = GTOpc; break;
6236 case ISD::SETGE: Swap = true;
6237 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6238 case ISD::SETULT: Swap = true;
6239 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6240 case ISD::SETUGE: Swap = true;
6241 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6244 std::swap(Op0, Op1);
6246 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6247 // bits of the inputs before performing those operations.
6249 EVT EltVT = VT.getVectorElementType();
6250 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6252 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6253 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6255 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6256 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6259 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6261 // If the logical-not of the result is required, perform that now.
6263 Result = DAG.getNOT(dl, Result, VT);
6268 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6269 static bool isX86LogicalCmp(SDValue Op) {
6270 unsigned Opc = Op.getNode()->getOpcode();
6271 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6273 if (Op.getResNo() == 1 &&
6274 (Opc == X86ISD::ADD ||
6275 Opc == X86ISD::SUB ||
6276 Opc == X86ISD::SMUL ||
6277 Opc == X86ISD::UMUL ||
6278 Opc == X86ISD::INC ||
6279 Opc == X86ISD::DEC ||
6280 Opc == X86ISD::OR ||
6281 Opc == X86ISD::XOR ||
6282 Opc == X86ISD::AND))
6288 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6289 bool addTest = true;
6290 SDValue Cond = Op.getOperand(0);
6291 DebugLoc dl = Op.getDebugLoc();
6294 if (Cond.getOpcode() == ISD::SETCC) {
6295 SDValue NewCond = LowerSETCC(Cond, DAG);
6296 if (NewCond.getNode())
6300 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6301 SDValue Op1 = Op.getOperand(1);
6302 SDValue Op2 = Op.getOperand(2);
6303 if (Cond.getOpcode() == X86ISD::SETCC &&
6304 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6305 SDValue Cmp = Cond.getOperand(1);
6306 if (Cmp.getOpcode() == X86ISD::CMP) {
6307 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6308 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6309 ConstantSDNode *RHSC =
6310 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6311 if (N1C && N1C->isAllOnesValue() &&
6312 N2C && N2C->isNullValue() &&
6313 RHSC && RHSC->isNullValue()) {
6314 SDValue CmpOp0 = Cmp.getOperand(0);
6315 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6316 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6317 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6318 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6323 // Look pass (and (setcc_carry (cmp ...)), 1).
6324 if (Cond.getOpcode() == ISD::AND &&
6325 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6326 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6327 if (C && C->getAPIntValue() == 1)
6328 Cond = Cond.getOperand(0);
6331 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6332 // setting operand in place of the X86ISD::SETCC.
6333 if (Cond.getOpcode() == X86ISD::SETCC ||
6334 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6335 CC = Cond.getOperand(0);
6337 SDValue Cmp = Cond.getOperand(1);
6338 unsigned Opc = Cmp.getOpcode();
6339 EVT VT = Op.getValueType();
6341 bool IllegalFPCMov = false;
6342 if (VT.isFloatingPoint() && !VT.isVector() &&
6343 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6344 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6346 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6347 Opc == X86ISD::BT) { // FIXME
6354 // Look pass the truncate.
6355 if (Cond.getOpcode() == ISD::TRUNCATE)
6356 Cond = Cond.getOperand(0);
6358 // We know the result of AND is compared against zero. Try to match
6360 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6361 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6362 if (NewSetCC.getNode()) {
6363 CC = NewSetCC.getOperand(0);
6364 Cond = NewSetCC.getOperand(1);
6371 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6372 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6375 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6376 // condition is true.
6377 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6378 SDValue Ops[] = { Op2, Op1, CC, Cond };
6379 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6382 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6383 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6384 // from the AND / OR.
6385 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6386 Opc = Op.getOpcode();
6387 if (Opc != ISD::OR && Opc != ISD::AND)
6389 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6390 Op.getOperand(0).hasOneUse() &&
6391 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6392 Op.getOperand(1).hasOneUse());
6395 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6396 // 1 and that the SETCC node has a single use.
6397 static bool isXor1OfSetCC(SDValue Op) {
6398 if (Op.getOpcode() != ISD::XOR)
6400 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6401 if (N1C && N1C->getAPIntValue() == 1) {
6402 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6403 Op.getOperand(0).hasOneUse();
6408 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6409 bool addTest = true;
6410 SDValue Chain = Op.getOperand(0);
6411 SDValue Cond = Op.getOperand(1);
6412 SDValue Dest = Op.getOperand(2);
6413 DebugLoc dl = Op.getDebugLoc();
6416 if (Cond.getOpcode() == ISD::SETCC) {
6417 SDValue NewCond = LowerSETCC(Cond, DAG);
6418 if (NewCond.getNode())
6422 // FIXME: LowerXALUO doesn't handle these!!
6423 else if (Cond.getOpcode() == X86ISD::ADD ||
6424 Cond.getOpcode() == X86ISD::SUB ||
6425 Cond.getOpcode() == X86ISD::SMUL ||
6426 Cond.getOpcode() == X86ISD::UMUL)
6427 Cond = LowerXALUO(Cond, DAG);
6430 // Look pass (and (setcc_carry (cmp ...)), 1).
6431 if (Cond.getOpcode() == ISD::AND &&
6432 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6433 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6434 if (C && C->getAPIntValue() == 1)
6435 Cond = Cond.getOperand(0);
6438 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6439 // setting operand in place of the X86ISD::SETCC.
6440 if (Cond.getOpcode() == X86ISD::SETCC ||
6441 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6442 CC = Cond.getOperand(0);
6444 SDValue Cmp = Cond.getOperand(1);
6445 unsigned Opc = Cmp.getOpcode();
6446 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6447 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6451 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6455 // These can only come from an arithmetic instruction with overflow,
6456 // e.g. SADDO, UADDO.
6457 Cond = Cond.getNode()->getOperand(1);
6464 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6465 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6466 if (CondOpc == ISD::OR) {
6467 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6468 // two branches instead of an explicit OR instruction with a
6470 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6471 isX86LogicalCmp(Cmp)) {
6472 CC = Cond.getOperand(0).getOperand(0);
6473 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6474 Chain, Dest, CC, Cmp);
6475 CC = Cond.getOperand(1).getOperand(0);
6479 } else { // ISD::AND
6480 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6481 // two branches instead of an explicit AND instruction with a
6482 // separate test. However, we only do this if this block doesn't
6483 // have a fall-through edge, because this requires an explicit
6484 // jmp when the condition is false.
6485 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6486 isX86LogicalCmp(Cmp) &&
6487 Op.getNode()->hasOneUse()) {
6488 X86::CondCode CCode =
6489 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6490 CCode = X86::GetOppositeBranchCondition(CCode);
6491 CC = DAG.getConstant(CCode, MVT::i8);
6492 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6493 // Look for an unconditional branch following this conditional branch.
6494 // We need this because we need to reverse the successors in order
6495 // to implement FCMP_OEQ.
6496 if (User.getOpcode() == ISD::BR) {
6497 SDValue FalseBB = User.getOperand(1);
6499 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6500 assert(NewBR == User);
6503 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6504 Chain, Dest, CC, Cmp);
6505 X86::CondCode CCode =
6506 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6507 CCode = X86::GetOppositeBranchCondition(CCode);
6508 CC = DAG.getConstant(CCode, MVT::i8);
6514 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6515 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6516 // It should be transformed during dag combiner except when the condition
6517 // is set by a arithmetics with overflow node.
6518 X86::CondCode CCode =
6519 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6520 CCode = X86::GetOppositeBranchCondition(CCode);
6521 CC = DAG.getConstant(CCode, MVT::i8);
6522 Cond = Cond.getOperand(0).getOperand(1);
6528 // Look pass the truncate.
6529 if (Cond.getOpcode() == ISD::TRUNCATE)
6530 Cond = Cond.getOperand(0);
6532 // We know the result of AND is compared against zero. Try to match
6534 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6535 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6536 if (NewSetCC.getNode()) {
6537 CC = NewSetCC.getOperand(0);
6538 Cond = NewSetCC.getOperand(1);
6545 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6546 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6548 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6549 Chain, Dest, CC, Cond);
6553 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6554 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6555 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6556 // that the guard pages used by the OS virtual memory manager are allocated in
6557 // correct sequence.
6559 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6560 SelectionDAG &DAG) const {
6561 assert(Subtarget->isTargetCygMing() &&
6562 "This should be used only on Cygwin/Mingw targets");
6563 DebugLoc dl = Op.getDebugLoc();
6566 SDValue Chain = Op.getOperand(0);
6567 SDValue Size = Op.getOperand(1);
6568 // FIXME: Ensure alignment here
6572 EVT IntPtr = getPointerTy();
6573 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6575 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6576 Flag = Chain.getValue(1);
6578 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6580 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6581 Flag = Chain.getValue(1);
6583 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6585 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6586 return DAG.getMergeValues(Ops1, 2, dl);
6590 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6592 SDValue Dst, SDValue Src,
6593 SDValue Size, unsigned Align,
6596 uint64_t DstSVOff) const {
6597 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6599 // If not DWORD aligned or size is more than the threshold, call the library.
6600 // The libc version is likely to be faster for these cases. It can use the
6601 // address value and run time information about the CPU.
6602 if ((Align & 3) != 0 ||
6604 ConstantSize->getZExtValue() >
6605 getSubtarget()->getMaxInlineSizeThreshold()) {
6606 SDValue InFlag(0, 0);
6608 // Check to see if there is a specialized entry-point for memory zeroing.
6609 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6611 if (const char *bzeroEntry = V &&
6612 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6613 EVT IntPtr = getPointerTy();
6614 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6615 TargetLowering::ArgListTy Args;
6616 TargetLowering::ArgListEntry Entry;
6618 Entry.Ty = IntPtrTy;
6619 Args.push_back(Entry);
6621 Args.push_back(Entry);
6622 std::pair<SDValue,SDValue> CallResult =
6623 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6624 false, false, false, false,
6625 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6626 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6627 return CallResult.second;
6630 // Otherwise have the target-independent code call memset.
6634 uint64_t SizeVal = ConstantSize->getZExtValue();
6635 SDValue InFlag(0, 0);
6638 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6639 unsigned BytesLeft = 0;
6640 bool TwoRepStos = false;
6643 uint64_t Val = ValC->getZExtValue() & 255;
6645 // If the value is a constant, then we can potentially use larger sets.
6646 switch (Align & 3) {
6647 case 2: // WORD aligned
6650 Val = (Val << 8) | Val;
6652 case 0: // DWORD aligned
6655 Val = (Val << 8) | Val;
6656 Val = (Val << 16) | Val;
6657 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6660 Val = (Val << 32) | Val;
6663 default: // Byte aligned
6666 Count = DAG.getIntPtrConstant(SizeVal);
6670 if (AVT.bitsGT(MVT::i8)) {
6671 unsigned UBytes = AVT.getSizeInBits() / 8;
6672 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6673 BytesLeft = SizeVal % UBytes;
6676 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6678 InFlag = Chain.getValue(1);
6681 Count = DAG.getIntPtrConstant(SizeVal);
6682 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6683 InFlag = Chain.getValue(1);
6686 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6689 InFlag = Chain.getValue(1);
6690 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6693 InFlag = Chain.getValue(1);
6695 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6696 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6697 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6700 InFlag = Chain.getValue(1);
6702 EVT CVT = Count.getValueType();
6703 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6704 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6705 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6708 InFlag = Chain.getValue(1);
6709 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6710 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6711 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6712 } else if (BytesLeft) {
6713 // Handle the last 1 - 7 bytes.
6714 unsigned Offset = SizeVal - BytesLeft;
6715 EVT AddrVT = Dst.getValueType();
6716 EVT SizeVT = Size.getValueType();
6718 Chain = DAG.getMemset(Chain, dl,
6719 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6720 DAG.getConstant(Offset, AddrVT)),
6722 DAG.getConstant(BytesLeft, SizeVT),
6723 Align, isVolatile, DstSV, DstSVOff + Offset);
6726 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6731 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6732 SDValue Chain, SDValue Dst, SDValue Src,
6733 SDValue Size, unsigned Align,
6734 bool isVolatile, bool AlwaysInline,
6738 uint64_t SrcSVOff) const {
6739 // This requires the copy size to be a constant, preferrably
6740 // within a subtarget-specific limit.
6741 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6744 uint64_t SizeVal = ConstantSize->getZExtValue();
6745 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6748 /// If not DWORD aligned, call the library.
6749 if ((Align & 3) != 0)
6754 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6757 unsigned UBytes = AVT.getSizeInBits() / 8;
6758 unsigned CountVal = SizeVal / UBytes;
6759 SDValue Count = DAG.getIntPtrConstant(CountVal);
6760 unsigned BytesLeft = SizeVal % UBytes;
6762 SDValue InFlag(0, 0);
6763 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6766 InFlag = Chain.getValue(1);
6767 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6770 InFlag = Chain.getValue(1);
6771 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6774 InFlag = Chain.getValue(1);
6776 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6777 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6778 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6779 array_lengthof(Ops));
6781 SmallVector<SDValue, 4> Results;
6782 Results.push_back(RepMovs);
6784 // Handle the last 1 - 7 bytes.
6785 unsigned Offset = SizeVal - BytesLeft;
6786 EVT DstVT = Dst.getValueType();
6787 EVT SrcVT = Src.getValueType();
6788 EVT SizeVT = Size.getValueType();
6789 Results.push_back(DAG.getMemcpy(Chain, dl,
6790 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6791 DAG.getConstant(Offset, DstVT)),
6792 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6793 DAG.getConstant(Offset, SrcVT)),
6794 DAG.getConstant(BytesLeft, SizeVT),
6795 Align, isVolatile, AlwaysInline,
6796 DstSV, DstSVOff + Offset,
6797 SrcSV, SrcSVOff + Offset));
6800 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6801 &Results[0], Results.size());
6804 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6805 MachineFunction &MF = DAG.getMachineFunction();
6806 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6808 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6809 DebugLoc dl = Op.getDebugLoc();
6811 if (!Subtarget->is64Bit()) {
6812 // vastart just stores the address of the VarArgsFrameIndex slot into the
6813 // memory location argument.
6814 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6816 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6821 // gp_offset (0 - 6 * 8)
6822 // fp_offset (48 - 48 + 8 * 16)
6823 // overflow_arg_area (point to parameters coming in memory).
6825 SmallVector<SDValue, 8> MemOps;
6826 SDValue FIN = Op.getOperand(1);
6828 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6829 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6831 FIN, SV, 0, false, false, 0);
6832 MemOps.push_back(Store);
6835 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6836 FIN, DAG.getIntPtrConstant(4));
6837 Store = DAG.getStore(Op.getOperand(0), dl,
6838 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6840 FIN, SV, 0, false, false, 0);
6841 MemOps.push_back(Store);
6843 // Store ptr to overflow_arg_area
6844 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6845 FIN, DAG.getIntPtrConstant(4));
6846 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6848 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6850 MemOps.push_back(Store);
6852 // Store ptr to reg_save_area.
6853 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6854 FIN, DAG.getIntPtrConstant(8));
6855 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6857 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6859 MemOps.push_back(Store);
6860 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6861 &MemOps[0], MemOps.size());
6864 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6865 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6866 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6867 SDValue Chain = Op.getOperand(0);
6868 SDValue SrcPtr = Op.getOperand(1);
6869 SDValue SrcSV = Op.getOperand(2);
6871 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6875 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6876 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6877 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6878 SDValue Chain = Op.getOperand(0);
6879 SDValue DstPtr = Op.getOperand(1);
6880 SDValue SrcPtr = Op.getOperand(2);
6881 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6882 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6883 DebugLoc dl = Op.getDebugLoc();
6885 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6886 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6887 false, DstSV, 0, SrcSV, 0);
6891 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6892 DebugLoc dl = Op.getDebugLoc();
6893 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6895 default: return SDValue(); // Don't custom lower most intrinsics.
6896 // Comparison intrinsics.
6897 case Intrinsic::x86_sse_comieq_ss:
6898 case Intrinsic::x86_sse_comilt_ss:
6899 case Intrinsic::x86_sse_comile_ss:
6900 case Intrinsic::x86_sse_comigt_ss:
6901 case Intrinsic::x86_sse_comige_ss:
6902 case Intrinsic::x86_sse_comineq_ss:
6903 case Intrinsic::x86_sse_ucomieq_ss:
6904 case Intrinsic::x86_sse_ucomilt_ss:
6905 case Intrinsic::x86_sse_ucomile_ss:
6906 case Intrinsic::x86_sse_ucomigt_ss:
6907 case Intrinsic::x86_sse_ucomige_ss:
6908 case Intrinsic::x86_sse_ucomineq_ss:
6909 case Intrinsic::x86_sse2_comieq_sd:
6910 case Intrinsic::x86_sse2_comilt_sd:
6911 case Intrinsic::x86_sse2_comile_sd:
6912 case Intrinsic::x86_sse2_comigt_sd:
6913 case Intrinsic::x86_sse2_comige_sd:
6914 case Intrinsic::x86_sse2_comineq_sd:
6915 case Intrinsic::x86_sse2_ucomieq_sd:
6916 case Intrinsic::x86_sse2_ucomilt_sd:
6917 case Intrinsic::x86_sse2_ucomile_sd:
6918 case Intrinsic::x86_sse2_ucomigt_sd:
6919 case Intrinsic::x86_sse2_ucomige_sd:
6920 case Intrinsic::x86_sse2_ucomineq_sd: {
6922 ISD::CondCode CC = ISD::SETCC_INVALID;
6925 case Intrinsic::x86_sse_comieq_ss:
6926 case Intrinsic::x86_sse2_comieq_sd:
6930 case Intrinsic::x86_sse_comilt_ss:
6931 case Intrinsic::x86_sse2_comilt_sd:
6935 case Intrinsic::x86_sse_comile_ss:
6936 case Intrinsic::x86_sse2_comile_sd:
6940 case Intrinsic::x86_sse_comigt_ss:
6941 case Intrinsic::x86_sse2_comigt_sd:
6945 case Intrinsic::x86_sse_comige_ss:
6946 case Intrinsic::x86_sse2_comige_sd:
6950 case Intrinsic::x86_sse_comineq_ss:
6951 case Intrinsic::x86_sse2_comineq_sd:
6955 case Intrinsic::x86_sse_ucomieq_ss:
6956 case Intrinsic::x86_sse2_ucomieq_sd:
6957 Opc = X86ISD::UCOMI;
6960 case Intrinsic::x86_sse_ucomilt_ss:
6961 case Intrinsic::x86_sse2_ucomilt_sd:
6962 Opc = X86ISD::UCOMI;
6965 case Intrinsic::x86_sse_ucomile_ss:
6966 case Intrinsic::x86_sse2_ucomile_sd:
6967 Opc = X86ISD::UCOMI;
6970 case Intrinsic::x86_sse_ucomigt_ss:
6971 case Intrinsic::x86_sse2_ucomigt_sd:
6972 Opc = X86ISD::UCOMI;
6975 case Intrinsic::x86_sse_ucomige_ss:
6976 case Intrinsic::x86_sse2_ucomige_sd:
6977 Opc = X86ISD::UCOMI;
6980 case Intrinsic::x86_sse_ucomineq_ss:
6981 case Intrinsic::x86_sse2_ucomineq_sd:
6982 Opc = X86ISD::UCOMI;
6987 SDValue LHS = Op.getOperand(1);
6988 SDValue RHS = Op.getOperand(2);
6989 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6990 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6991 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6992 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6993 DAG.getConstant(X86CC, MVT::i8), Cond);
6994 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6996 // ptest intrinsics. The intrinsic these come from are designed to return
6997 // an integer value, not just an instruction so lower it to the ptest
6998 // pattern and a setcc for the result.
6999 case Intrinsic::x86_sse41_ptestz:
7000 case Intrinsic::x86_sse41_ptestc:
7001 case Intrinsic::x86_sse41_ptestnzc:{
7004 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7005 case Intrinsic::x86_sse41_ptestz:
7007 X86CC = X86::COND_E;
7009 case Intrinsic::x86_sse41_ptestc:
7011 X86CC = X86::COND_B;
7013 case Intrinsic::x86_sse41_ptestnzc:
7015 X86CC = X86::COND_A;
7019 SDValue LHS = Op.getOperand(1);
7020 SDValue RHS = Op.getOperand(2);
7021 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
7022 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7023 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7024 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7027 // Fix vector shift instructions where the last operand is a non-immediate
7029 case Intrinsic::x86_sse2_pslli_w:
7030 case Intrinsic::x86_sse2_pslli_d:
7031 case Intrinsic::x86_sse2_pslli_q:
7032 case Intrinsic::x86_sse2_psrli_w:
7033 case Intrinsic::x86_sse2_psrli_d:
7034 case Intrinsic::x86_sse2_psrli_q:
7035 case Intrinsic::x86_sse2_psrai_w:
7036 case Intrinsic::x86_sse2_psrai_d:
7037 case Intrinsic::x86_mmx_pslli_w:
7038 case Intrinsic::x86_mmx_pslli_d:
7039 case Intrinsic::x86_mmx_pslli_q:
7040 case Intrinsic::x86_mmx_psrli_w:
7041 case Intrinsic::x86_mmx_psrli_d:
7042 case Intrinsic::x86_mmx_psrli_q:
7043 case Intrinsic::x86_mmx_psrai_w:
7044 case Intrinsic::x86_mmx_psrai_d: {
7045 SDValue ShAmt = Op.getOperand(2);
7046 if (isa<ConstantSDNode>(ShAmt))
7049 unsigned NewIntNo = 0;
7050 EVT ShAmtVT = MVT::v4i32;
7052 case Intrinsic::x86_sse2_pslli_w:
7053 NewIntNo = Intrinsic::x86_sse2_psll_w;
7055 case Intrinsic::x86_sse2_pslli_d:
7056 NewIntNo = Intrinsic::x86_sse2_psll_d;
7058 case Intrinsic::x86_sse2_pslli_q:
7059 NewIntNo = Intrinsic::x86_sse2_psll_q;
7061 case Intrinsic::x86_sse2_psrli_w:
7062 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7064 case Intrinsic::x86_sse2_psrli_d:
7065 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7067 case Intrinsic::x86_sse2_psrli_q:
7068 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7070 case Intrinsic::x86_sse2_psrai_w:
7071 NewIntNo = Intrinsic::x86_sse2_psra_w;
7073 case Intrinsic::x86_sse2_psrai_d:
7074 NewIntNo = Intrinsic::x86_sse2_psra_d;
7077 ShAmtVT = MVT::v2i32;
7079 case Intrinsic::x86_mmx_pslli_w:
7080 NewIntNo = Intrinsic::x86_mmx_psll_w;
7082 case Intrinsic::x86_mmx_pslli_d:
7083 NewIntNo = Intrinsic::x86_mmx_psll_d;
7085 case Intrinsic::x86_mmx_pslli_q:
7086 NewIntNo = Intrinsic::x86_mmx_psll_q;
7088 case Intrinsic::x86_mmx_psrli_w:
7089 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7091 case Intrinsic::x86_mmx_psrli_d:
7092 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7094 case Intrinsic::x86_mmx_psrli_q:
7095 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7097 case Intrinsic::x86_mmx_psrai_w:
7098 NewIntNo = Intrinsic::x86_mmx_psra_w;
7100 case Intrinsic::x86_mmx_psrai_d:
7101 NewIntNo = Intrinsic::x86_mmx_psra_d;
7103 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7109 // The vector shift intrinsics with scalars uses 32b shift amounts but
7110 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7114 ShOps[1] = DAG.getConstant(0, MVT::i32);
7115 if (ShAmtVT == MVT::v4i32) {
7116 ShOps[2] = DAG.getUNDEF(MVT::i32);
7117 ShOps[3] = DAG.getUNDEF(MVT::i32);
7118 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7120 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7123 EVT VT = Op.getValueType();
7124 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7125 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7126 DAG.getConstant(NewIntNo, MVT::i32),
7127 Op.getOperand(1), ShAmt);
7132 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7133 SelectionDAG &DAG) const {
7134 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7135 DebugLoc dl = Op.getDebugLoc();
7138 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7140 DAG.getConstant(TD->getPointerSize(),
7141 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7142 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7143 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7145 NULL, 0, false, false, 0);
7148 // Just load the return address.
7149 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7150 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7151 RetAddrFI, NULL, 0, false, false, 0);
7154 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7155 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7156 MFI->setFrameAddressIsTaken(true);
7157 EVT VT = Op.getValueType();
7158 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7159 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7160 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7161 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7163 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7168 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7169 SelectionDAG &DAG) const {
7170 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7173 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7174 MachineFunction &MF = DAG.getMachineFunction();
7175 SDValue Chain = Op.getOperand(0);
7176 SDValue Offset = Op.getOperand(1);
7177 SDValue Handler = Op.getOperand(2);
7178 DebugLoc dl = Op.getDebugLoc();
7180 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7182 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7184 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7185 DAG.getIntPtrConstant(-TD->getPointerSize()));
7186 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7187 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7188 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7189 MF.getRegInfo().addLiveOut(StoreAddrReg);
7191 return DAG.getNode(X86ISD::EH_RETURN, dl,
7193 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7196 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7197 SelectionDAG &DAG) const {
7198 SDValue Root = Op.getOperand(0);
7199 SDValue Trmp = Op.getOperand(1); // trampoline
7200 SDValue FPtr = Op.getOperand(2); // nested function
7201 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7202 DebugLoc dl = Op.getDebugLoc();
7204 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7206 if (Subtarget->is64Bit()) {
7207 SDValue OutChains[6];
7209 // Large code-model.
7210 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7211 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7213 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7214 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7216 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7218 // Load the pointer to the nested function into R11.
7219 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7220 SDValue Addr = Trmp;
7221 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7222 Addr, TrmpAddr, 0, false, false, 0);
7224 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7225 DAG.getConstant(2, MVT::i64));
7226 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7229 // Load the 'nest' parameter value into R10.
7230 // R10 is specified in X86CallingConv.td
7231 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7232 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7233 DAG.getConstant(10, MVT::i64));
7234 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7235 Addr, TrmpAddr, 10, false, false, 0);
7237 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7238 DAG.getConstant(12, MVT::i64));
7239 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7242 // Jump to the nested function.
7243 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7244 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7245 DAG.getConstant(20, MVT::i64));
7246 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7247 Addr, TrmpAddr, 20, false, false, 0);
7249 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7250 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7251 DAG.getConstant(22, MVT::i64));
7252 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7253 TrmpAddr, 22, false, false, 0);
7256 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7257 return DAG.getMergeValues(Ops, 2, dl);
7259 const Function *Func =
7260 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7261 CallingConv::ID CC = Func->getCallingConv();
7266 llvm_unreachable("Unsupported calling convention");
7267 case CallingConv::C:
7268 case CallingConv::X86_StdCall: {
7269 // Pass 'nest' parameter in ECX.
7270 // Must be kept in sync with X86CallingConv.td
7273 // Check that ECX wasn't needed by an 'inreg' parameter.
7274 const FunctionType *FTy = Func->getFunctionType();
7275 const AttrListPtr &Attrs = Func->getAttributes();
7277 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7278 unsigned InRegCount = 0;
7281 for (FunctionType::param_iterator I = FTy->param_begin(),
7282 E = FTy->param_end(); I != E; ++I, ++Idx)
7283 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7284 // FIXME: should only count parameters that are lowered to integers.
7285 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7287 if (InRegCount > 2) {
7288 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7293 case CallingConv::X86_FastCall:
7294 case CallingConv::Fast:
7295 // Pass 'nest' parameter in EAX.
7296 // Must be kept in sync with X86CallingConv.td
7301 SDValue OutChains[4];
7304 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7305 DAG.getConstant(10, MVT::i32));
7306 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7308 // This is storing the opcode for MOV32ri.
7309 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7310 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7311 OutChains[0] = DAG.getStore(Root, dl,
7312 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7313 Trmp, TrmpAddr, 0, false, false, 0);
7315 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7316 DAG.getConstant(1, MVT::i32));
7317 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7320 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7321 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7322 DAG.getConstant(5, MVT::i32));
7323 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7324 TrmpAddr, 5, false, false, 1);
7326 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7327 DAG.getConstant(6, MVT::i32));
7328 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7332 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7333 return DAG.getMergeValues(Ops, 2, dl);
7337 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7338 SelectionDAG &DAG) const {
7340 The rounding mode is in bits 11:10 of FPSR, and has the following
7347 FLT_ROUNDS, on the other hand, expects the following:
7354 To perform the conversion, we do:
7355 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7358 MachineFunction &MF = DAG.getMachineFunction();
7359 const TargetMachine &TM = MF.getTarget();
7360 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7361 unsigned StackAlignment = TFI.getStackAlignment();
7362 EVT VT = Op.getValueType();
7363 DebugLoc dl = Op.getDebugLoc();
7365 // Save FP Control Word to stack slot
7366 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7367 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7369 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7370 DAG.getEntryNode(), StackSlot);
7372 // Load FP Control Word from stack slot
7373 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7376 // Transform as necessary
7378 DAG.getNode(ISD::SRL, dl, MVT::i16,
7379 DAG.getNode(ISD::AND, dl, MVT::i16,
7380 CWD, DAG.getConstant(0x800, MVT::i16)),
7381 DAG.getConstant(11, MVT::i8));
7383 DAG.getNode(ISD::SRL, dl, MVT::i16,
7384 DAG.getNode(ISD::AND, dl, MVT::i16,
7385 CWD, DAG.getConstant(0x400, MVT::i16)),
7386 DAG.getConstant(9, MVT::i8));
7389 DAG.getNode(ISD::AND, dl, MVT::i16,
7390 DAG.getNode(ISD::ADD, dl, MVT::i16,
7391 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7392 DAG.getConstant(1, MVT::i16)),
7393 DAG.getConstant(3, MVT::i16));
7396 return DAG.getNode((VT.getSizeInBits() < 16 ?
7397 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7400 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7401 EVT VT = Op.getValueType();
7403 unsigned NumBits = VT.getSizeInBits();
7404 DebugLoc dl = Op.getDebugLoc();
7406 Op = Op.getOperand(0);
7407 if (VT == MVT::i8) {
7408 // Zero extend to i32 since there is not an i8 bsr.
7410 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7413 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7414 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7415 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7417 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7420 DAG.getConstant(NumBits+NumBits-1, OpVT),
7421 DAG.getConstant(X86::COND_E, MVT::i8),
7424 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7426 // Finally xor with NumBits-1.
7427 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7430 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7434 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7435 EVT VT = Op.getValueType();
7437 unsigned NumBits = VT.getSizeInBits();
7438 DebugLoc dl = Op.getDebugLoc();
7440 Op = Op.getOperand(0);
7441 if (VT == MVT::i8) {
7443 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7446 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7447 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7448 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7450 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7453 DAG.getConstant(NumBits, OpVT),
7454 DAG.getConstant(X86::COND_E, MVT::i8),
7457 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7460 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7464 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7465 EVT VT = Op.getValueType();
7466 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7467 DebugLoc dl = Op.getDebugLoc();
7469 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7470 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7471 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7472 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7473 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7475 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7476 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7477 // return AloBlo + AloBhi + AhiBlo;
7479 SDValue A = Op.getOperand(0);
7480 SDValue B = Op.getOperand(1);
7482 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7483 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7484 A, DAG.getConstant(32, MVT::i32));
7485 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7486 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7487 B, DAG.getConstant(32, MVT::i32));
7488 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7489 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7491 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7492 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7494 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7495 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7497 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7498 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7499 AloBhi, DAG.getConstant(32, MVT::i32));
7500 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7501 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7502 AhiBlo, DAG.getConstant(32, MVT::i32));
7503 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7504 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7509 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7510 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7511 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7512 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7513 // has only one use.
7514 SDNode *N = Op.getNode();
7515 SDValue LHS = N->getOperand(0);
7516 SDValue RHS = N->getOperand(1);
7517 unsigned BaseOp = 0;
7519 DebugLoc dl = Op.getDebugLoc();
7521 switch (Op.getOpcode()) {
7522 default: llvm_unreachable("Unknown ovf instruction!");
7524 // A subtract of one will be selected as a INC. Note that INC doesn't
7525 // set CF, so we can't do this for UADDO.
7526 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7527 if (C->getAPIntValue() == 1) {
7528 BaseOp = X86ISD::INC;
7532 BaseOp = X86ISD::ADD;
7536 BaseOp = X86ISD::ADD;
7540 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7541 // set CF, so we can't do this for USUBO.
7542 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7543 if (C->getAPIntValue() == 1) {
7544 BaseOp = X86ISD::DEC;
7548 BaseOp = X86ISD::SUB;
7552 BaseOp = X86ISD::SUB;
7556 BaseOp = X86ISD::SMUL;
7560 BaseOp = X86ISD::UMUL;
7565 // Also sets EFLAGS.
7566 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7567 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7570 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7571 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7573 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7577 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7578 EVT T = Op.getValueType();
7579 DebugLoc dl = Op.getDebugLoc();
7582 switch(T.getSimpleVT().SimpleTy) {
7584 assert(false && "Invalid value type!");
7585 case MVT::i8: Reg = X86::AL; size = 1; break;
7586 case MVT::i16: Reg = X86::AX; size = 2; break;
7587 case MVT::i32: Reg = X86::EAX; size = 4; break;
7589 assert(Subtarget->is64Bit() && "Node not type legal!");
7590 Reg = X86::RAX; size = 8;
7593 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7594 Op.getOperand(2), SDValue());
7595 SDValue Ops[] = { cpIn.getValue(0),
7598 DAG.getTargetConstant(size, MVT::i8),
7600 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7601 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7603 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7607 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7608 SelectionDAG &DAG) const {
7609 assert(Subtarget->is64Bit() && "Result not type legalized?");
7610 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7611 SDValue TheChain = Op.getOperand(0);
7612 DebugLoc dl = Op.getDebugLoc();
7613 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7614 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7615 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7617 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7618 DAG.getConstant(32, MVT::i8));
7620 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7623 return DAG.getMergeValues(Ops, 2, dl);
7626 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7627 SDNode *Node = Op.getNode();
7628 DebugLoc dl = Node->getDebugLoc();
7629 EVT T = Node->getValueType(0);
7630 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7631 DAG.getConstant(0, T), Node->getOperand(2));
7632 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7633 cast<AtomicSDNode>(Node)->getMemoryVT(),
7634 Node->getOperand(0),
7635 Node->getOperand(1), negOp,
7636 cast<AtomicSDNode>(Node)->getSrcValue(),
7637 cast<AtomicSDNode>(Node)->getAlignment());
7640 /// LowerOperation - Provide custom lowering hooks for some operations.
7642 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7643 switch (Op.getOpcode()) {
7644 default: llvm_unreachable("Should not custom lower this!");
7645 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7646 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7647 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7648 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7649 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7650 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7651 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7652 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7653 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7654 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7655 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7656 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7657 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7658 case ISD::SHL_PARTS:
7659 case ISD::SRA_PARTS:
7660 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7661 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7662 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7663 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7664 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7665 case ISD::FABS: return LowerFABS(Op, DAG);
7666 case ISD::FNEG: return LowerFNEG(Op, DAG);
7667 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7668 case ISD::SETCC: return LowerSETCC(Op, DAG);
7669 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7670 case ISD::SELECT: return LowerSELECT(Op, DAG);
7671 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7672 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7673 case ISD::VASTART: return LowerVASTART(Op, DAG);
7674 case ISD::VAARG: return LowerVAARG(Op, DAG);
7675 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7676 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7677 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7678 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7679 case ISD::FRAME_TO_ARGS_OFFSET:
7680 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7681 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7682 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7683 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7684 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7685 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7686 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7687 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7693 case ISD::UMULO: return LowerXALUO(Op, DAG);
7694 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7698 void X86TargetLowering::
7699 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7700 SelectionDAG &DAG, unsigned NewOp) const {
7701 EVT T = Node->getValueType(0);
7702 DebugLoc dl = Node->getDebugLoc();
7703 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7705 SDValue Chain = Node->getOperand(0);
7706 SDValue In1 = Node->getOperand(1);
7707 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7708 Node->getOperand(2), DAG.getIntPtrConstant(0));
7709 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7710 Node->getOperand(2), DAG.getIntPtrConstant(1));
7711 SDValue Ops[] = { Chain, In1, In2L, In2H };
7712 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7714 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7715 cast<MemSDNode>(Node)->getMemOperand());
7716 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7717 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7718 Results.push_back(Result.getValue(2));
7721 /// ReplaceNodeResults - Replace a node with an illegal result type
7722 /// with a new node built out of custom code.
7723 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7724 SmallVectorImpl<SDValue>&Results,
7725 SelectionDAG &DAG) const {
7726 DebugLoc dl = N->getDebugLoc();
7727 switch (N->getOpcode()) {
7729 assert(false && "Do not know how to custom type legalize this operation!");
7731 case ISD::FP_TO_SINT: {
7732 std::pair<SDValue,SDValue> Vals =
7733 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7734 SDValue FIST = Vals.first, StackSlot = Vals.second;
7735 if (FIST.getNode() != 0) {
7736 EVT VT = N->getValueType(0);
7737 // Return a load from the stack slot.
7738 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7743 case ISD::READCYCLECOUNTER: {
7744 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7745 SDValue TheChain = N->getOperand(0);
7746 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7747 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7749 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7751 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7752 SDValue Ops[] = { eax, edx };
7753 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7754 Results.push_back(edx.getValue(1));
7757 case ISD::ATOMIC_CMP_SWAP: {
7758 EVT T = N->getValueType(0);
7759 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7760 SDValue cpInL, cpInH;
7761 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7762 DAG.getConstant(0, MVT::i32));
7763 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7764 DAG.getConstant(1, MVT::i32));
7765 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7766 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7768 SDValue swapInL, swapInH;
7769 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7770 DAG.getConstant(0, MVT::i32));
7771 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7772 DAG.getConstant(1, MVT::i32));
7773 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7775 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7776 swapInL.getValue(1));
7777 SDValue Ops[] = { swapInH.getValue(0),
7779 swapInH.getValue(1) };
7780 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7781 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7782 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7783 MVT::i32, Result.getValue(1));
7784 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7785 MVT::i32, cpOutL.getValue(2));
7786 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7787 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7788 Results.push_back(cpOutH.getValue(1));
7791 case ISD::ATOMIC_LOAD_ADD:
7792 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7794 case ISD::ATOMIC_LOAD_AND:
7795 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7797 case ISD::ATOMIC_LOAD_NAND:
7798 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7800 case ISD::ATOMIC_LOAD_OR:
7801 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7803 case ISD::ATOMIC_LOAD_SUB:
7804 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7806 case ISD::ATOMIC_LOAD_XOR:
7807 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7809 case ISD::ATOMIC_SWAP:
7810 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7815 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7817 default: return NULL;
7818 case X86ISD::BSF: return "X86ISD::BSF";
7819 case X86ISD::BSR: return "X86ISD::BSR";
7820 case X86ISD::SHLD: return "X86ISD::SHLD";
7821 case X86ISD::SHRD: return "X86ISD::SHRD";
7822 case X86ISD::FAND: return "X86ISD::FAND";
7823 case X86ISD::FOR: return "X86ISD::FOR";
7824 case X86ISD::FXOR: return "X86ISD::FXOR";
7825 case X86ISD::FSRL: return "X86ISD::FSRL";
7826 case X86ISD::FILD: return "X86ISD::FILD";
7827 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7828 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7829 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7830 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7831 case X86ISD::FLD: return "X86ISD::FLD";
7832 case X86ISD::FST: return "X86ISD::FST";
7833 case X86ISD::CALL: return "X86ISD::CALL";
7834 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7835 case X86ISD::BT: return "X86ISD::BT";
7836 case X86ISD::CMP: return "X86ISD::CMP";
7837 case X86ISD::COMI: return "X86ISD::COMI";
7838 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7839 case X86ISD::SETCC: return "X86ISD::SETCC";
7840 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7841 case X86ISD::CMOV: return "X86ISD::CMOV";
7842 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7843 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7844 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7845 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7846 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7847 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7848 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7849 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7850 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7851 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7852 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7853 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7854 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7855 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7856 case X86ISD::FMAX: return "X86ISD::FMAX";
7857 case X86ISD::FMIN: return "X86ISD::FMIN";
7858 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7859 case X86ISD::FRCP: return "X86ISD::FRCP";
7860 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7861 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7862 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7863 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7864 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7865 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7866 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7867 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7868 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7869 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7870 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7871 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7872 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7873 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7874 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7875 case X86ISD::VSHL: return "X86ISD::VSHL";
7876 case X86ISD::VSRL: return "X86ISD::VSRL";
7877 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7878 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7879 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7880 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7881 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7882 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7883 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7884 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7885 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7886 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7887 case X86ISD::ADD: return "X86ISD::ADD";
7888 case X86ISD::SUB: return "X86ISD::SUB";
7889 case X86ISD::SMUL: return "X86ISD::SMUL";
7890 case X86ISD::UMUL: return "X86ISD::UMUL";
7891 case X86ISD::INC: return "X86ISD::INC";
7892 case X86ISD::DEC: return "X86ISD::DEC";
7893 case X86ISD::OR: return "X86ISD::OR";
7894 case X86ISD::XOR: return "X86ISD::XOR";
7895 case X86ISD::AND: return "X86ISD::AND";
7896 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7897 case X86ISD::PTEST: return "X86ISD::PTEST";
7898 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7899 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7903 // isLegalAddressingMode - Return true if the addressing mode represented
7904 // by AM is legal for this target, for a load/store of the specified type.
7905 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7906 const Type *Ty) const {
7907 // X86 supports extremely general addressing modes.
7908 CodeModel::Model M = getTargetMachine().getCodeModel();
7910 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7911 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7916 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7918 // If a reference to this global requires an extra load, we can't fold it.
7919 if (isGlobalStubReference(GVFlags))
7922 // If BaseGV requires a register for the PIC base, we cannot also have a
7923 // BaseReg specified.
7924 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7927 // If lower 4G is not available, then we must use rip-relative addressing.
7928 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7938 // These scales always work.
7943 // These scales are formed with basereg+scalereg. Only accept if there is
7948 default: // Other stuff never works.
7956 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7957 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7959 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7960 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7961 if (NumBits1 <= NumBits2)
7966 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7967 if (!VT1.isInteger() || !VT2.isInteger())
7969 unsigned NumBits1 = VT1.getSizeInBits();
7970 unsigned NumBits2 = VT2.getSizeInBits();
7971 if (NumBits1 <= NumBits2)
7976 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7977 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7978 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7981 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7982 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7983 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7986 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7987 // i16 instructions are longer (0x66 prefix) and potentially slower.
7988 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7991 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7992 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7993 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7994 /// are assumed to be legal.
7996 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7998 // Very little shuffling can be done for 64-bit vectors right now.
7999 if (VT.getSizeInBits() == 64)
8000 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8002 // FIXME: pshufb, blends, shifts.
8003 return (VT.getVectorNumElements() == 2 ||
8004 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8005 isMOVLMask(M, VT) ||
8006 isSHUFPMask(M, VT) ||
8007 isPSHUFDMask(M, VT) ||
8008 isPSHUFHWMask(M, VT) ||
8009 isPSHUFLWMask(M, VT) ||
8010 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8011 isUNPCKLMask(M, VT) ||
8012 isUNPCKHMask(M, VT) ||
8013 isUNPCKL_v_undef_Mask(M, VT) ||
8014 isUNPCKH_v_undef_Mask(M, VT));
8018 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8020 unsigned NumElts = VT.getVectorNumElements();
8021 // FIXME: This collection of masks seems suspect.
8024 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8025 return (isMOVLMask(Mask, VT) ||
8026 isCommutedMOVLMask(Mask, VT, true) ||
8027 isSHUFPMask(Mask, VT) ||
8028 isCommutedSHUFPMask(Mask, VT));
8033 //===----------------------------------------------------------------------===//
8034 // X86 Scheduler Hooks
8035 //===----------------------------------------------------------------------===//
8037 // private utility function
8039 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8040 MachineBasicBlock *MBB,
8048 TargetRegisterClass *RC,
8049 bool invSrc) const {
8050 // For the atomic bitwise operator, we generate
8053 // ld t1 = [bitinstr.addr]
8054 // op t2 = t1, [bitinstr.val]
8056 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8058 // fallthrough -->nextMBB
8059 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8060 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8061 MachineFunction::iterator MBBIter = MBB;
8064 /// First build the CFG
8065 MachineFunction *F = MBB->getParent();
8066 MachineBasicBlock *thisMBB = MBB;
8067 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8068 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8069 F->insert(MBBIter, newMBB);
8070 F->insert(MBBIter, nextMBB);
8072 // Move all successors to thisMBB to nextMBB
8073 nextMBB->transferSuccessors(thisMBB);
8075 // Update thisMBB to fall through to newMBB
8076 thisMBB->addSuccessor(newMBB);
8078 // newMBB jumps to itself and fall through to nextMBB
8079 newMBB->addSuccessor(nextMBB);
8080 newMBB->addSuccessor(newMBB);
8082 // Insert instructions into newMBB based on incoming instruction
8083 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8084 "unexpected number of operands");
8085 DebugLoc dl = bInstr->getDebugLoc();
8086 MachineOperand& destOper = bInstr->getOperand(0);
8087 MachineOperand* argOpers[2 + X86AddrNumOperands];
8088 int numArgs = bInstr->getNumOperands() - 1;
8089 for (int i=0; i < numArgs; ++i)
8090 argOpers[i] = &bInstr->getOperand(i+1);
8092 // x86 address has 4 operands: base, index, scale, and displacement
8093 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8094 int valArgIndx = lastAddrIndx + 1;
8096 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8097 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8098 for (int i=0; i <= lastAddrIndx; ++i)
8099 (*MIB).addOperand(*argOpers[i]);
8101 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8103 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8108 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8109 assert((argOpers[valArgIndx]->isReg() ||
8110 argOpers[valArgIndx]->isImm()) &&
8112 if (argOpers[valArgIndx]->isReg())
8113 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8115 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8117 (*MIB).addOperand(*argOpers[valArgIndx]);
8119 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8122 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8123 for (int i=0; i <= lastAddrIndx; ++i)
8124 (*MIB).addOperand(*argOpers[i]);
8126 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8127 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8128 bInstr->memoperands_end());
8130 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8134 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8136 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8140 // private utility function: 64 bit atomics on 32 bit host.
8142 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8143 MachineBasicBlock *MBB,
8148 bool invSrc) const {
8149 // For the atomic bitwise operator, we generate
8150 // thisMBB (instructions are in pairs, except cmpxchg8b)
8151 // ld t1,t2 = [bitinstr.addr]
8153 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8154 // op t5, t6 <- out1, out2, [bitinstr.val]
8155 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8156 // mov ECX, EBX <- t5, t6
8157 // mov EAX, EDX <- t1, t2
8158 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8159 // mov t3, t4 <- EAX, EDX
8161 // result in out1, out2
8162 // fallthrough -->nextMBB
8164 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8165 const unsigned LoadOpc = X86::MOV32rm;
8166 const unsigned copyOpc = X86::MOV32rr;
8167 const unsigned NotOpc = X86::NOT32r;
8168 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8169 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8170 MachineFunction::iterator MBBIter = MBB;
8173 /// First build the CFG
8174 MachineFunction *F = MBB->getParent();
8175 MachineBasicBlock *thisMBB = MBB;
8176 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8177 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8178 F->insert(MBBIter, newMBB);
8179 F->insert(MBBIter, nextMBB);
8181 // Move all successors to thisMBB to nextMBB
8182 nextMBB->transferSuccessors(thisMBB);
8184 // Update thisMBB to fall through to newMBB
8185 thisMBB->addSuccessor(newMBB);
8187 // newMBB jumps to itself and fall through to nextMBB
8188 newMBB->addSuccessor(nextMBB);
8189 newMBB->addSuccessor(newMBB);
8191 DebugLoc dl = bInstr->getDebugLoc();
8192 // Insert instructions into newMBB based on incoming instruction
8193 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8194 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8195 "unexpected number of operands");
8196 MachineOperand& dest1Oper = bInstr->getOperand(0);
8197 MachineOperand& dest2Oper = bInstr->getOperand(1);
8198 MachineOperand* argOpers[2 + X86AddrNumOperands];
8199 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8200 argOpers[i] = &bInstr->getOperand(i+2);
8202 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8203 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8205 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8206 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8207 for (int i=0; i <= lastAddrIndx; ++i)
8208 (*MIB).addOperand(*argOpers[i]);
8209 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8210 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8211 // add 4 to displacement.
8212 for (int i=0; i <= lastAddrIndx-2; ++i)
8213 (*MIB).addOperand(*argOpers[i]);
8214 MachineOperand newOp3 = *(argOpers[3]);
8216 newOp3.setImm(newOp3.getImm()+4);
8218 newOp3.setOffset(newOp3.getOffset()+4);
8219 (*MIB).addOperand(newOp3);
8220 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8222 // t3/4 are defined later, at the bottom of the loop
8223 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8224 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8225 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8226 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8227 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8228 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8230 // The subsequent operations should be using the destination registers of
8231 //the PHI instructions.
8233 t1 = F->getRegInfo().createVirtualRegister(RC);
8234 t2 = F->getRegInfo().createVirtualRegister(RC);
8235 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8236 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8238 t1 = dest1Oper.getReg();
8239 t2 = dest2Oper.getReg();
8242 int valArgIndx = lastAddrIndx + 1;
8243 assert((argOpers[valArgIndx]->isReg() ||
8244 argOpers[valArgIndx]->isImm()) &&
8246 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8247 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8248 if (argOpers[valArgIndx]->isReg())
8249 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8251 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8252 if (regOpcL != X86::MOV32rr)
8254 (*MIB).addOperand(*argOpers[valArgIndx]);
8255 assert(argOpers[valArgIndx + 1]->isReg() ==
8256 argOpers[valArgIndx]->isReg());
8257 assert(argOpers[valArgIndx + 1]->isImm() ==
8258 argOpers[valArgIndx]->isImm());
8259 if (argOpers[valArgIndx + 1]->isReg())
8260 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8262 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8263 if (regOpcH != X86::MOV32rr)
8265 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8267 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8269 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8272 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8274 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8277 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8278 for (int i=0; i <= lastAddrIndx; ++i)
8279 (*MIB).addOperand(*argOpers[i]);
8281 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8282 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8283 bInstr->memoperands_end());
8285 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8286 MIB.addReg(X86::EAX);
8287 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8288 MIB.addReg(X86::EDX);
8291 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8293 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8297 // private utility function
8299 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8300 MachineBasicBlock *MBB,
8301 unsigned cmovOpc) const {
8302 // For the atomic min/max operator, we generate
8305 // ld t1 = [min/max.addr]
8306 // mov t2 = [min/max.val]
8308 // cmov[cond] t2 = t1
8310 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8312 // fallthrough -->nextMBB
8314 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8315 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8316 MachineFunction::iterator MBBIter = MBB;
8319 /// First build the CFG
8320 MachineFunction *F = MBB->getParent();
8321 MachineBasicBlock *thisMBB = MBB;
8322 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8323 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8324 F->insert(MBBIter, newMBB);
8325 F->insert(MBBIter, nextMBB);
8327 // Move all successors of thisMBB to nextMBB
8328 nextMBB->transferSuccessors(thisMBB);
8330 // Update thisMBB to fall through to newMBB
8331 thisMBB->addSuccessor(newMBB);
8333 // newMBB jumps to newMBB and fall through to nextMBB
8334 newMBB->addSuccessor(nextMBB);
8335 newMBB->addSuccessor(newMBB);
8337 DebugLoc dl = mInstr->getDebugLoc();
8338 // Insert instructions into newMBB based on incoming instruction
8339 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8340 "unexpected number of operands");
8341 MachineOperand& destOper = mInstr->getOperand(0);
8342 MachineOperand* argOpers[2 + X86AddrNumOperands];
8343 int numArgs = mInstr->getNumOperands() - 1;
8344 for (int i=0; i < numArgs; ++i)
8345 argOpers[i] = &mInstr->getOperand(i+1);
8347 // x86 address has 4 operands: base, index, scale, and displacement
8348 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8349 int valArgIndx = lastAddrIndx + 1;
8351 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8352 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8353 for (int i=0; i <= lastAddrIndx; ++i)
8354 (*MIB).addOperand(*argOpers[i]);
8356 // We only support register and immediate values
8357 assert((argOpers[valArgIndx]->isReg() ||
8358 argOpers[valArgIndx]->isImm()) &&
8361 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8362 if (argOpers[valArgIndx]->isReg())
8363 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8365 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8366 (*MIB).addOperand(*argOpers[valArgIndx]);
8368 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8371 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8376 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8377 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8381 // Cmp and exchange if none has modified the memory location
8382 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8383 for (int i=0; i <= lastAddrIndx; ++i)
8384 (*MIB).addOperand(*argOpers[i]);
8386 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8387 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8388 mInstr->memoperands_end());
8390 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8391 MIB.addReg(X86::EAX);
8394 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8396 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8400 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8401 // all of this code can be replaced with that in the .td file.
8403 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8404 unsigned numArgs, bool memArg) const {
8406 MachineFunction *F = BB->getParent();
8407 DebugLoc dl = MI->getDebugLoc();
8408 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8412 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8414 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8416 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8418 for (unsigned i = 0; i < numArgs; ++i) {
8419 MachineOperand &Op = MI->getOperand(i+1);
8421 if (!(Op.isReg() && Op.isImplicit()))
8425 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8428 F->DeleteMachineInstr(MI);
8434 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8436 MachineBasicBlock *MBB) const {
8437 // Emit code to save XMM registers to the stack. The ABI says that the
8438 // number of registers to save is given in %al, so it's theoretically
8439 // possible to do an indirect jump trick to avoid saving all of them,
8440 // however this code takes a simpler approach and just executes all
8441 // of the stores if %al is non-zero. It's less code, and it's probably
8442 // easier on the hardware branch predictor, and stores aren't all that
8443 // expensive anyway.
8445 // Create the new basic blocks. One block contains all the XMM stores,
8446 // and one block is the final destination regardless of whether any
8447 // stores were performed.
8448 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8449 MachineFunction *F = MBB->getParent();
8450 MachineFunction::iterator MBBIter = MBB;
8452 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8453 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8454 F->insert(MBBIter, XMMSaveMBB);
8455 F->insert(MBBIter, EndMBB);
8458 // Move any original successors of MBB to the end block.
8459 EndMBB->transferSuccessors(MBB);
8460 // The original block will now fall through to the XMM save block.
8461 MBB->addSuccessor(XMMSaveMBB);
8462 // The XMMSaveMBB will fall through to the end block.
8463 XMMSaveMBB->addSuccessor(EndMBB);
8465 // Now add the instructions.
8466 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8467 DebugLoc DL = MI->getDebugLoc();
8469 unsigned CountReg = MI->getOperand(0).getReg();
8470 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8471 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8473 if (!Subtarget->isTargetWin64()) {
8474 // If %al is 0, branch around the XMM save block.
8475 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8476 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8477 MBB->addSuccessor(EndMBB);
8480 // In the XMM save block, save all the XMM argument registers.
8481 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8482 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8483 MachineMemOperand *MMO =
8484 F->getMachineMemOperand(
8485 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8486 MachineMemOperand::MOStore, Offset,
8487 /*Size=*/16, /*Align=*/16);
8488 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8489 .addFrameIndex(RegSaveFrameIndex)
8490 .addImm(/*Scale=*/1)
8491 .addReg(/*IndexReg=*/0)
8492 .addImm(/*Disp=*/Offset)
8493 .addReg(/*Segment=*/0)
8494 .addReg(MI->getOperand(i).getReg())
8495 .addMemOperand(MMO);
8498 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8504 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8505 MachineBasicBlock *BB,
8506 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8507 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8508 DebugLoc DL = MI->getDebugLoc();
8510 // To "insert" a SELECT_CC instruction, we actually have to insert the
8511 // diamond control-flow pattern. The incoming instruction knows the
8512 // destination vreg to set, the condition code register to branch on, the
8513 // true/false values to select between, and a branch opcode to use.
8514 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8515 MachineFunction::iterator It = BB;
8521 // cmpTY ccX, r1, r2
8523 // fallthrough --> copy0MBB
8524 MachineBasicBlock *thisMBB = BB;
8525 MachineFunction *F = BB->getParent();
8526 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8527 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8529 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8530 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8531 F->insert(It, copy0MBB);
8532 F->insert(It, sinkMBB);
8533 // Update machine-CFG edges by first adding all successors of the current
8534 // block to the new block which will contain the Phi node for the select.
8535 // Also inform sdisel of the edge changes.
8536 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8537 E = BB->succ_end(); I != E; ++I) {
8538 EM->insert(std::make_pair(*I, sinkMBB));
8539 sinkMBB->addSuccessor(*I);
8541 // Next, remove all successors of the current block, and add the true
8542 // and fallthrough blocks as its successors.
8543 while (!BB->succ_empty())
8544 BB->removeSuccessor(BB->succ_begin());
8545 // Add the true and fallthrough blocks as its successors.
8546 BB->addSuccessor(copy0MBB);
8547 BB->addSuccessor(sinkMBB);
8550 // %FalseValue = ...
8551 // # fallthrough to sinkMBB
8554 // Update machine-CFG edges
8555 BB->addSuccessor(sinkMBB);
8558 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8561 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8562 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8563 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8565 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8570 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8571 MachineBasicBlock *BB,
8572 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8573 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8574 DebugLoc DL = MI->getDebugLoc();
8575 MachineFunction *F = BB->getParent();
8577 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8578 // non-trivial part is impdef of ESP.
8579 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8582 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8583 .addExternalSymbol("_alloca")
8584 .addReg(X86::EAX, RegState::Implicit)
8585 .addReg(X86::ESP, RegState::Implicit)
8586 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8587 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8589 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8594 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8595 MachineBasicBlock *BB,
8596 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8597 switch (MI->getOpcode()) {
8598 default: assert(false && "Unexpected instr type to insert");
8599 case X86::MINGW_ALLOCA:
8600 return EmitLoweredMingwAlloca(MI, BB, EM);
8602 case X86::CMOV_V1I64:
8603 case X86::CMOV_FR32:
8604 case X86::CMOV_FR64:
8605 case X86::CMOV_V4F32:
8606 case X86::CMOV_V2F64:
8607 case X86::CMOV_V2I64:
8608 case X86::CMOV_GR16:
8609 case X86::CMOV_GR32:
8610 case X86::CMOV_RFP32:
8611 case X86::CMOV_RFP64:
8612 case X86::CMOV_RFP80:
8613 return EmitLoweredSelect(MI, BB, EM);
8615 case X86::FP32_TO_INT16_IN_MEM:
8616 case X86::FP32_TO_INT32_IN_MEM:
8617 case X86::FP32_TO_INT64_IN_MEM:
8618 case X86::FP64_TO_INT16_IN_MEM:
8619 case X86::FP64_TO_INT32_IN_MEM:
8620 case X86::FP64_TO_INT64_IN_MEM:
8621 case X86::FP80_TO_INT16_IN_MEM:
8622 case X86::FP80_TO_INT32_IN_MEM:
8623 case X86::FP80_TO_INT64_IN_MEM: {
8624 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8625 DebugLoc DL = MI->getDebugLoc();
8627 // Change the floating point control register to use "round towards zero"
8628 // mode when truncating to an integer value.
8629 MachineFunction *F = BB->getParent();
8630 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8631 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8633 // Load the old value of the high byte of the control word...
8635 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8636 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8639 // Set the high part to be round to zero...
8640 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8643 // Reload the modified control word now...
8644 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8646 // Restore the memory image of control word to original value
8647 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8650 // Get the X86 opcode to use.
8652 switch (MI->getOpcode()) {
8653 default: llvm_unreachable("illegal opcode!");
8654 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8655 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8656 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8657 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8658 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8659 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8660 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8661 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8662 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8666 MachineOperand &Op = MI->getOperand(0);
8668 AM.BaseType = X86AddressMode::RegBase;
8669 AM.Base.Reg = Op.getReg();
8671 AM.BaseType = X86AddressMode::FrameIndexBase;
8672 AM.Base.FrameIndex = Op.getIndex();
8674 Op = MI->getOperand(1);
8676 AM.Scale = Op.getImm();
8677 Op = MI->getOperand(2);
8679 AM.IndexReg = Op.getImm();
8680 Op = MI->getOperand(3);
8681 if (Op.isGlobal()) {
8682 AM.GV = Op.getGlobal();
8684 AM.Disp = Op.getImm();
8686 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8687 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8689 // Reload the original control word now.
8690 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8692 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8695 // DBG_VALUE. Only the frame index case is done here.
8696 case X86::DBG_VALUE: {
8697 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8698 DebugLoc DL = MI->getDebugLoc();
8700 MachineFunction *F = BB->getParent();
8701 AM.BaseType = X86AddressMode::FrameIndexBase;
8702 AM.Base.FrameIndex = MI->getOperand(0).getImm();
8703 addFullAddress(BuildMI(BB, DL, TII->get(X86::DBG_VALUE)), AM).
8704 addImm(MI->getOperand(1).getImm()).
8705 addMetadata(MI->getOperand(2).getMetadata());
8706 F->DeleteMachineInstr(MI); // Remove pseudo.
8710 // String/text processing lowering.
8711 case X86::PCMPISTRM128REG:
8712 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8713 case X86::PCMPISTRM128MEM:
8714 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8715 case X86::PCMPESTRM128REG:
8716 return EmitPCMP(MI, BB, 5, false /* in mem */);
8717 case X86::PCMPESTRM128MEM:
8718 return EmitPCMP(MI, BB, 5, true /* in mem */);
8721 case X86::ATOMAND32:
8722 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8723 X86::AND32ri, X86::MOV32rm,
8724 X86::LCMPXCHG32, X86::MOV32rr,
8725 X86::NOT32r, X86::EAX,
8726 X86::GR32RegisterClass);
8728 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8729 X86::OR32ri, X86::MOV32rm,
8730 X86::LCMPXCHG32, X86::MOV32rr,
8731 X86::NOT32r, X86::EAX,
8732 X86::GR32RegisterClass);
8733 case X86::ATOMXOR32:
8734 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8735 X86::XOR32ri, X86::MOV32rm,
8736 X86::LCMPXCHG32, X86::MOV32rr,
8737 X86::NOT32r, X86::EAX,
8738 X86::GR32RegisterClass);
8739 case X86::ATOMNAND32:
8740 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8741 X86::AND32ri, X86::MOV32rm,
8742 X86::LCMPXCHG32, X86::MOV32rr,
8743 X86::NOT32r, X86::EAX,
8744 X86::GR32RegisterClass, true);
8745 case X86::ATOMMIN32:
8746 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8747 case X86::ATOMMAX32:
8748 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8749 case X86::ATOMUMIN32:
8750 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8751 case X86::ATOMUMAX32:
8752 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8754 case X86::ATOMAND16:
8755 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8756 X86::AND16ri, X86::MOV16rm,
8757 X86::LCMPXCHG16, X86::MOV16rr,
8758 X86::NOT16r, X86::AX,
8759 X86::GR16RegisterClass);
8761 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8762 X86::OR16ri, X86::MOV16rm,
8763 X86::LCMPXCHG16, X86::MOV16rr,
8764 X86::NOT16r, X86::AX,
8765 X86::GR16RegisterClass);
8766 case X86::ATOMXOR16:
8767 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8768 X86::XOR16ri, X86::MOV16rm,
8769 X86::LCMPXCHG16, X86::MOV16rr,
8770 X86::NOT16r, X86::AX,
8771 X86::GR16RegisterClass);
8772 case X86::ATOMNAND16:
8773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8774 X86::AND16ri, X86::MOV16rm,
8775 X86::LCMPXCHG16, X86::MOV16rr,
8776 X86::NOT16r, X86::AX,
8777 X86::GR16RegisterClass, true);
8778 case X86::ATOMMIN16:
8779 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8780 case X86::ATOMMAX16:
8781 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8782 case X86::ATOMUMIN16:
8783 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8784 case X86::ATOMUMAX16:
8785 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8788 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8789 X86::AND8ri, X86::MOV8rm,
8790 X86::LCMPXCHG8, X86::MOV8rr,
8791 X86::NOT8r, X86::AL,
8792 X86::GR8RegisterClass);
8794 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8795 X86::OR8ri, X86::MOV8rm,
8796 X86::LCMPXCHG8, X86::MOV8rr,
8797 X86::NOT8r, X86::AL,
8798 X86::GR8RegisterClass);
8800 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8801 X86::XOR8ri, X86::MOV8rm,
8802 X86::LCMPXCHG8, X86::MOV8rr,
8803 X86::NOT8r, X86::AL,
8804 X86::GR8RegisterClass);
8805 case X86::ATOMNAND8:
8806 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8807 X86::AND8ri, X86::MOV8rm,
8808 X86::LCMPXCHG8, X86::MOV8rr,
8809 X86::NOT8r, X86::AL,
8810 X86::GR8RegisterClass, true);
8811 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8812 // This group is for 64-bit host.
8813 case X86::ATOMAND64:
8814 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8815 X86::AND64ri32, X86::MOV64rm,
8816 X86::LCMPXCHG64, X86::MOV64rr,
8817 X86::NOT64r, X86::RAX,
8818 X86::GR64RegisterClass);
8820 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8821 X86::OR64ri32, X86::MOV64rm,
8822 X86::LCMPXCHG64, X86::MOV64rr,
8823 X86::NOT64r, X86::RAX,
8824 X86::GR64RegisterClass);
8825 case X86::ATOMXOR64:
8826 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8827 X86::XOR64ri32, X86::MOV64rm,
8828 X86::LCMPXCHG64, X86::MOV64rr,
8829 X86::NOT64r, X86::RAX,
8830 X86::GR64RegisterClass);
8831 case X86::ATOMNAND64:
8832 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8833 X86::AND64ri32, X86::MOV64rm,
8834 X86::LCMPXCHG64, X86::MOV64rr,
8835 X86::NOT64r, X86::RAX,
8836 X86::GR64RegisterClass, true);
8837 case X86::ATOMMIN64:
8838 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8839 case X86::ATOMMAX64:
8840 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8841 case X86::ATOMUMIN64:
8842 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8843 case X86::ATOMUMAX64:
8844 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8846 // This group does 64-bit operations on a 32-bit host.
8847 case X86::ATOMAND6432:
8848 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8849 X86::AND32rr, X86::AND32rr,
8850 X86::AND32ri, X86::AND32ri,
8852 case X86::ATOMOR6432:
8853 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8854 X86::OR32rr, X86::OR32rr,
8855 X86::OR32ri, X86::OR32ri,
8857 case X86::ATOMXOR6432:
8858 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8859 X86::XOR32rr, X86::XOR32rr,
8860 X86::XOR32ri, X86::XOR32ri,
8862 case X86::ATOMNAND6432:
8863 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8864 X86::AND32rr, X86::AND32rr,
8865 X86::AND32ri, X86::AND32ri,
8867 case X86::ATOMADD6432:
8868 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8869 X86::ADD32rr, X86::ADC32rr,
8870 X86::ADD32ri, X86::ADC32ri,
8872 case X86::ATOMSUB6432:
8873 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8874 X86::SUB32rr, X86::SBB32rr,
8875 X86::SUB32ri, X86::SBB32ri,
8877 case X86::ATOMSWAP6432:
8878 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8879 X86::MOV32rr, X86::MOV32rr,
8880 X86::MOV32ri, X86::MOV32ri,
8882 case X86::VASTART_SAVE_XMM_REGS:
8883 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8887 //===----------------------------------------------------------------------===//
8888 // X86 Optimization Hooks
8889 //===----------------------------------------------------------------------===//
8891 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8895 const SelectionDAG &DAG,
8896 unsigned Depth) const {
8897 unsigned Opc = Op.getOpcode();
8898 assert((Opc >= ISD::BUILTIN_OP_END ||
8899 Opc == ISD::INTRINSIC_WO_CHAIN ||
8900 Opc == ISD::INTRINSIC_W_CHAIN ||
8901 Opc == ISD::INTRINSIC_VOID) &&
8902 "Should use MaskedValueIsZero if you don't know whether Op"
8903 " is a target node!");
8905 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8917 // These nodes' second result is a boolean.
8918 if (Op.getResNo() == 0)
8922 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8923 Mask.getBitWidth() - 1);
8928 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8929 /// node is a GlobalAddress + offset.
8930 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8931 const GlobalValue* &GA,
8932 int64_t &Offset) const {
8933 if (N->getOpcode() == X86ISD::Wrapper) {
8934 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8935 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8936 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8940 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8943 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8944 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8945 /// if the load addresses are consecutive, non-overlapping, and in the right
8947 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8948 const TargetLowering &TLI) {
8949 DebugLoc dl = N->getDebugLoc();
8950 EVT VT = N->getValueType(0);
8951 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8953 if (VT.getSizeInBits() != 128)
8956 SmallVector<SDValue, 16> Elts;
8957 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8958 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8960 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8963 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8964 /// and convert it from being a bunch of shuffles and extracts to a simple
8965 /// store and scalar loads to extract the elements.
8966 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8967 const TargetLowering &TLI) {
8968 SDValue InputVector = N->getOperand(0);
8970 // Only operate on vectors of 4 elements, where the alternative shuffling
8971 // gets to be more expensive.
8972 if (InputVector.getValueType() != MVT::v4i32)
8975 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8976 // single use which is a sign-extend or zero-extend, and all elements are
8978 SmallVector<SDNode *, 4> Uses;
8979 unsigned ExtractedElements = 0;
8980 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8981 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8982 if (UI.getUse().getResNo() != InputVector.getResNo())
8985 SDNode *Extract = *UI;
8986 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8989 if (Extract->getValueType(0) != MVT::i32)
8991 if (!Extract->hasOneUse())
8993 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8994 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8996 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8999 // Record which element was extracted.
9000 ExtractedElements |=
9001 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9003 Uses.push_back(Extract);
9006 // If not all the elements were used, this may not be worthwhile.
9007 if (ExtractedElements != 15)
9010 // Ok, we've now decided to do the transformation.
9011 DebugLoc dl = InputVector.getDebugLoc();
9013 // Store the value to a temporary stack slot.
9014 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9015 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9018 // Replace each use (extract) with a load of the appropriate element.
9019 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9020 UE = Uses.end(); UI != UE; ++UI) {
9021 SDNode *Extract = *UI;
9023 // Compute the element's address.
9024 SDValue Idx = Extract->getOperand(1);
9026 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9027 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9028 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9030 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9033 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9034 NULL, 0, false, false, 0);
9036 // Replace the exact with the load.
9037 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9040 // The replacement was made in place; don't return anything.
9044 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9045 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9046 const X86Subtarget *Subtarget) {
9047 DebugLoc DL = N->getDebugLoc();
9048 SDValue Cond = N->getOperand(0);
9049 // Get the LHS/RHS of the select.
9050 SDValue LHS = N->getOperand(1);
9051 SDValue RHS = N->getOperand(2);
9053 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9054 // instructions match the semantics of the common C idiom x<y?x:y but not
9055 // x<=y?x:y, because of how they handle negative zero (which can be
9056 // ignored in unsafe-math mode).
9057 if (Subtarget->hasSSE2() &&
9058 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9059 Cond.getOpcode() == ISD::SETCC) {
9060 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9062 unsigned Opcode = 0;
9063 // Check for x CC y ? x : y.
9064 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9065 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9069 // Converting this to a min would handle NaNs incorrectly, and swapping
9070 // the operands would cause it to handle comparisons between positive
9071 // and negative zero incorrectly.
9072 if (!FiniteOnlyFPMath() &&
9073 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9074 if (!UnsafeFPMath &&
9075 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9077 std::swap(LHS, RHS);
9079 Opcode = X86ISD::FMIN;
9082 // Converting this to a min would handle comparisons between positive
9083 // and negative zero incorrectly.
9084 if (!UnsafeFPMath &&
9085 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9087 Opcode = X86ISD::FMIN;
9090 // Converting this to a min would handle both negative zeros and NaNs
9091 // incorrectly, but we can swap the operands to fix both.
9092 std::swap(LHS, RHS);
9096 Opcode = X86ISD::FMIN;
9100 // Converting this to a max would handle comparisons between positive
9101 // and negative zero incorrectly.
9102 if (!UnsafeFPMath &&
9103 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9105 Opcode = X86ISD::FMAX;
9108 // Converting this to a max would handle NaNs incorrectly, and swapping
9109 // the operands would cause it to handle comparisons between positive
9110 // and negative zero incorrectly.
9111 if (!FiniteOnlyFPMath() &&
9112 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9113 if (!UnsafeFPMath &&
9114 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9116 std::swap(LHS, RHS);
9118 Opcode = X86ISD::FMAX;
9121 // Converting this to a max would handle both negative zeros and NaNs
9122 // incorrectly, but we can swap the operands to fix both.
9123 std::swap(LHS, RHS);
9127 Opcode = X86ISD::FMAX;
9130 // Check for x CC y ? y : x -- a min/max with reversed arms.
9131 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9132 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9136 // Converting this to a min would handle comparisons between positive
9137 // and negative zero incorrectly, and swapping the operands would
9138 // cause it to handle NaNs incorrectly.
9139 if (!UnsafeFPMath &&
9140 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9141 if (!FiniteOnlyFPMath() &&
9142 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9144 std::swap(LHS, RHS);
9146 Opcode = X86ISD::FMIN;
9149 // Converting this to a min would handle NaNs incorrectly.
9150 if (!UnsafeFPMath &&
9151 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9153 Opcode = X86ISD::FMIN;
9156 // Converting this to a min would handle both negative zeros and NaNs
9157 // incorrectly, but we can swap the operands to fix both.
9158 std::swap(LHS, RHS);
9162 Opcode = X86ISD::FMIN;
9166 // Converting this to a max would handle NaNs incorrectly.
9167 if (!FiniteOnlyFPMath() &&
9168 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9170 Opcode = X86ISD::FMAX;
9173 // Converting this to a max would handle comparisons between positive
9174 // and negative zero incorrectly, and swapping the operands would
9175 // cause it to handle NaNs incorrectly.
9176 if (!UnsafeFPMath &&
9177 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9178 if (!FiniteOnlyFPMath() &&
9179 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9181 std::swap(LHS, RHS);
9183 Opcode = X86ISD::FMAX;
9186 // Converting this to a max would handle both negative zeros and NaNs
9187 // incorrectly, but we can swap the operands to fix both.
9188 std::swap(LHS, RHS);
9192 Opcode = X86ISD::FMAX;
9198 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9201 // If this is a select between two integer constants, try to do some
9203 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9204 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9205 // Don't do this for crazy integer types.
9206 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9207 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9208 // so that TrueC (the true value) is larger than FalseC.
9209 bool NeedsCondInvert = false;
9211 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9212 // Efficiently invertible.
9213 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9214 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9215 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9216 NeedsCondInvert = true;
9217 std::swap(TrueC, FalseC);
9220 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9221 if (FalseC->getAPIntValue() == 0 &&
9222 TrueC->getAPIntValue().isPowerOf2()) {
9223 if (NeedsCondInvert) // Invert the condition if needed.
9224 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9225 DAG.getConstant(1, Cond.getValueType()));
9227 // Zero extend the condition if needed.
9228 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9230 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9231 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9232 DAG.getConstant(ShAmt, MVT::i8));
9235 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9236 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9237 if (NeedsCondInvert) // Invert the condition if needed.
9238 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9239 DAG.getConstant(1, Cond.getValueType()));
9241 // Zero extend the condition if needed.
9242 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9243 FalseC->getValueType(0), Cond);
9244 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9245 SDValue(FalseC, 0));
9248 // Optimize cases that will turn into an LEA instruction. This requires
9249 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9250 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9251 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9252 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9254 bool isFastMultiplier = false;
9256 switch ((unsigned char)Diff) {
9258 case 1: // result = add base, cond
9259 case 2: // result = lea base( , cond*2)
9260 case 3: // result = lea base(cond, cond*2)
9261 case 4: // result = lea base( , cond*4)
9262 case 5: // result = lea base(cond, cond*4)
9263 case 8: // result = lea base( , cond*8)
9264 case 9: // result = lea base(cond, cond*8)
9265 isFastMultiplier = true;
9270 if (isFastMultiplier) {
9271 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9272 if (NeedsCondInvert) // Invert the condition if needed.
9273 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9274 DAG.getConstant(1, Cond.getValueType()));
9276 // Zero extend the condition if needed.
9277 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9279 // Scale the condition by the difference.
9281 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9282 DAG.getConstant(Diff, Cond.getValueType()));
9284 // Add the base if non-zero.
9285 if (FalseC->getAPIntValue() != 0)
9286 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9287 SDValue(FalseC, 0));
9297 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9298 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9299 TargetLowering::DAGCombinerInfo &DCI) {
9300 DebugLoc DL = N->getDebugLoc();
9302 // If the flag operand isn't dead, don't touch this CMOV.
9303 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9306 // If this is a select between two integer constants, try to do some
9307 // optimizations. Note that the operands are ordered the opposite of SELECT
9309 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9310 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9311 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9312 // larger than FalseC (the false value).
9313 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9315 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9316 CC = X86::GetOppositeBranchCondition(CC);
9317 std::swap(TrueC, FalseC);
9320 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9321 // This is efficient for any integer data type (including i8/i16) and
9323 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9324 SDValue Cond = N->getOperand(3);
9325 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9326 DAG.getConstant(CC, MVT::i8), Cond);
9328 // Zero extend the condition if needed.
9329 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9331 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9332 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9333 DAG.getConstant(ShAmt, MVT::i8));
9334 if (N->getNumValues() == 2) // Dead flag value?
9335 return DCI.CombineTo(N, Cond, SDValue());
9339 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9340 // for any integer data type, including i8/i16.
9341 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9342 SDValue Cond = N->getOperand(3);
9343 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9344 DAG.getConstant(CC, MVT::i8), Cond);
9346 // Zero extend the condition if needed.
9347 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9348 FalseC->getValueType(0), Cond);
9349 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9350 SDValue(FalseC, 0));
9352 if (N->getNumValues() == 2) // Dead flag value?
9353 return DCI.CombineTo(N, Cond, SDValue());
9357 // Optimize cases that will turn into an LEA instruction. This requires
9358 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9359 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9360 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9361 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9363 bool isFastMultiplier = false;
9365 switch ((unsigned char)Diff) {
9367 case 1: // result = add base, cond
9368 case 2: // result = lea base( , cond*2)
9369 case 3: // result = lea base(cond, cond*2)
9370 case 4: // result = lea base( , cond*4)
9371 case 5: // result = lea base(cond, cond*4)
9372 case 8: // result = lea base( , cond*8)
9373 case 9: // result = lea base(cond, cond*8)
9374 isFastMultiplier = true;
9379 if (isFastMultiplier) {
9380 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9381 SDValue Cond = N->getOperand(3);
9382 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9383 DAG.getConstant(CC, MVT::i8), Cond);
9384 // Zero extend the condition if needed.
9385 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9387 // Scale the condition by the difference.
9389 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9390 DAG.getConstant(Diff, Cond.getValueType()));
9392 // Add the base if non-zero.
9393 if (FalseC->getAPIntValue() != 0)
9394 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9395 SDValue(FalseC, 0));
9396 if (N->getNumValues() == 2) // Dead flag value?
9397 return DCI.CombineTo(N, Cond, SDValue());
9407 /// PerformMulCombine - Optimize a single multiply with constant into two
9408 /// in order to implement it with two cheaper instructions, e.g.
9409 /// LEA + SHL, LEA + LEA.
9410 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9411 TargetLowering::DAGCombinerInfo &DCI) {
9412 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9415 EVT VT = N->getValueType(0);
9419 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9422 uint64_t MulAmt = C->getZExtValue();
9423 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9426 uint64_t MulAmt1 = 0;
9427 uint64_t MulAmt2 = 0;
9428 if ((MulAmt % 9) == 0) {
9430 MulAmt2 = MulAmt / 9;
9431 } else if ((MulAmt % 5) == 0) {
9433 MulAmt2 = MulAmt / 5;
9434 } else if ((MulAmt % 3) == 0) {
9436 MulAmt2 = MulAmt / 3;
9439 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9440 DebugLoc DL = N->getDebugLoc();
9442 if (isPowerOf2_64(MulAmt2) &&
9443 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9444 // If second multiplifer is pow2, issue it first. We want the multiply by
9445 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9447 std::swap(MulAmt1, MulAmt2);
9450 if (isPowerOf2_64(MulAmt1))
9451 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9452 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9454 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9455 DAG.getConstant(MulAmt1, VT));
9457 if (isPowerOf2_64(MulAmt2))
9458 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9459 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9461 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9462 DAG.getConstant(MulAmt2, VT));
9464 // Do not add new nodes to DAG combiner worklist.
9465 DCI.CombineTo(N, NewMul, false);
9470 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9471 SDValue N0 = N->getOperand(0);
9472 SDValue N1 = N->getOperand(1);
9473 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9474 EVT VT = N0.getValueType();
9476 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9477 // since the result of setcc_c is all zero's or all ones.
9478 if (N1C && N0.getOpcode() == ISD::AND &&
9479 N0.getOperand(1).getOpcode() == ISD::Constant) {
9480 SDValue N00 = N0.getOperand(0);
9481 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9482 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9483 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9484 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9485 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9486 APInt ShAmt = N1C->getAPIntValue();
9487 Mask = Mask.shl(ShAmt);
9489 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9490 N00, DAG.getConstant(Mask, VT));
9497 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9499 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9500 const X86Subtarget *Subtarget) {
9501 EVT VT = N->getValueType(0);
9502 if (!VT.isVector() && VT.isInteger() &&
9503 N->getOpcode() == ISD::SHL)
9504 return PerformSHLCombine(N, DAG);
9506 // On X86 with SSE2 support, we can transform this to a vector shift if
9507 // all elements are shifted by the same amount. We can't do this in legalize
9508 // because the a constant vector is typically transformed to a constant pool
9509 // so we have no knowledge of the shift amount.
9510 if (!Subtarget->hasSSE2())
9513 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9516 SDValue ShAmtOp = N->getOperand(1);
9517 EVT EltVT = VT.getVectorElementType();
9518 DebugLoc DL = N->getDebugLoc();
9519 SDValue BaseShAmt = SDValue();
9520 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9521 unsigned NumElts = VT.getVectorNumElements();
9523 for (; i != NumElts; ++i) {
9524 SDValue Arg = ShAmtOp.getOperand(i);
9525 if (Arg.getOpcode() == ISD::UNDEF) continue;
9529 for (; i != NumElts; ++i) {
9530 SDValue Arg = ShAmtOp.getOperand(i);
9531 if (Arg.getOpcode() == ISD::UNDEF) continue;
9532 if (Arg != BaseShAmt) {
9536 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9537 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9538 SDValue InVec = ShAmtOp.getOperand(0);
9539 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9540 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9542 for (; i != NumElts; ++i) {
9543 SDValue Arg = InVec.getOperand(i);
9544 if (Arg.getOpcode() == ISD::UNDEF) continue;
9548 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9549 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9550 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9551 if (C->getZExtValue() == SplatIdx)
9552 BaseShAmt = InVec.getOperand(1);
9555 if (BaseShAmt.getNode() == 0)
9556 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9557 DAG.getIntPtrConstant(0));
9561 // The shift amount is an i32.
9562 if (EltVT.bitsGT(MVT::i32))
9563 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9564 else if (EltVT.bitsLT(MVT::i32))
9565 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9567 // The shift amount is identical so we can do a vector shift.
9568 SDValue ValOp = N->getOperand(0);
9569 switch (N->getOpcode()) {
9571 llvm_unreachable("Unknown shift opcode!");
9574 if (VT == MVT::v2i64)
9575 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9576 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9578 if (VT == MVT::v4i32)
9579 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9580 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9582 if (VT == MVT::v8i16)
9583 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9584 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9588 if (VT == MVT::v4i32)
9589 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9590 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9592 if (VT == MVT::v8i16)
9593 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9594 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9598 if (VT == MVT::v2i64)
9599 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9600 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9602 if (VT == MVT::v4i32)
9603 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9604 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9606 if (VT == MVT::v8i16)
9607 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9608 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9615 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9616 const X86Subtarget *Subtarget) {
9617 EVT VT = N->getValueType(0);
9618 if (VT != MVT::i64 || !Subtarget->is64Bit())
9621 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9622 SDValue N0 = N->getOperand(0);
9623 SDValue N1 = N->getOperand(1);
9624 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9626 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9629 SDValue ShAmt0 = N0.getOperand(1);
9630 if (ShAmt0.getValueType() != MVT::i8)
9632 SDValue ShAmt1 = N1.getOperand(1);
9633 if (ShAmt1.getValueType() != MVT::i8)
9635 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9636 ShAmt0 = ShAmt0.getOperand(0);
9637 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9638 ShAmt1 = ShAmt1.getOperand(0);
9640 DebugLoc DL = N->getDebugLoc();
9641 unsigned Opc = X86ISD::SHLD;
9642 SDValue Op0 = N0.getOperand(0);
9643 SDValue Op1 = N1.getOperand(0);
9644 if (ShAmt0.getOpcode() == ISD::SUB) {
9646 std::swap(Op0, Op1);
9647 std::swap(ShAmt0, ShAmt1);
9650 if (ShAmt1.getOpcode() == ISD::SUB) {
9651 SDValue Sum = ShAmt1.getOperand(0);
9652 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9653 if (SumC->getSExtValue() == 64 &&
9654 ShAmt1.getOperand(1) == ShAmt0)
9655 return DAG.getNode(Opc, DL, VT,
9657 DAG.getNode(ISD::TRUNCATE, DL,
9660 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9661 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9663 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9664 return DAG.getNode(Opc, DL, VT,
9665 N0.getOperand(0), N1.getOperand(0),
9666 DAG.getNode(ISD::TRUNCATE, DL,
9673 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9674 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9675 const X86Subtarget *Subtarget) {
9676 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9677 // the FP state in cases where an emms may be missing.
9678 // A preferable solution to the general problem is to figure out the right
9679 // places to insert EMMS. This qualifies as a quick hack.
9681 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9682 StoreSDNode *St = cast<StoreSDNode>(N);
9683 EVT VT = St->getValue().getValueType();
9684 if (VT.getSizeInBits() != 64)
9687 const Function *F = DAG.getMachineFunction().getFunction();
9688 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9689 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9690 && Subtarget->hasSSE2();
9691 if ((VT.isVector() ||
9692 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9693 isa<LoadSDNode>(St->getValue()) &&
9694 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9695 St->getChain().hasOneUse() && !St->isVolatile()) {
9696 SDNode* LdVal = St->getValue().getNode();
9698 int TokenFactorIndex = -1;
9699 SmallVector<SDValue, 8> Ops;
9700 SDNode* ChainVal = St->getChain().getNode();
9701 // Must be a store of a load. We currently handle two cases: the load
9702 // is a direct child, and it's under an intervening TokenFactor. It is
9703 // possible to dig deeper under nested TokenFactors.
9704 if (ChainVal == LdVal)
9705 Ld = cast<LoadSDNode>(St->getChain());
9706 else if (St->getValue().hasOneUse() &&
9707 ChainVal->getOpcode() == ISD::TokenFactor) {
9708 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9709 if (ChainVal->getOperand(i).getNode() == LdVal) {
9710 TokenFactorIndex = i;
9711 Ld = cast<LoadSDNode>(St->getValue());
9713 Ops.push_back(ChainVal->getOperand(i));
9717 if (!Ld || !ISD::isNormalLoad(Ld))
9720 // If this is not the MMX case, i.e. we are just turning i64 load/store
9721 // into f64 load/store, avoid the transformation if there are multiple
9722 // uses of the loaded value.
9723 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9726 DebugLoc LdDL = Ld->getDebugLoc();
9727 DebugLoc StDL = N->getDebugLoc();
9728 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9729 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9731 if (Subtarget->is64Bit() || F64IsLegal) {
9732 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9733 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9734 Ld->getBasePtr(), Ld->getSrcValue(),
9735 Ld->getSrcValueOffset(), Ld->isVolatile(),
9736 Ld->isNonTemporal(), Ld->getAlignment());
9737 SDValue NewChain = NewLd.getValue(1);
9738 if (TokenFactorIndex != -1) {
9739 Ops.push_back(NewChain);
9740 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9743 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9744 St->getSrcValue(), St->getSrcValueOffset(),
9745 St->isVolatile(), St->isNonTemporal(),
9746 St->getAlignment());
9749 // Otherwise, lower to two pairs of 32-bit loads / stores.
9750 SDValue LoAddr = Ld->getBasePtr();
9751 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9752 DAG.getConstant(4, MVT::i32));
9754 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9755 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9756 Ld->isVolatile(), Ld->isNonTemporal(),
9757 Ld->getAlignment());
9758 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9759 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9760 Ld->isVolatile(), Ld->isNonTemporal(),
9761 MinAlign(Ld->getAlignment(), 4));
9763 SDValue NewChain = LoLd.getValue(1);
9764 if (TokenFactorIndex != -1) {
9765 Ops.push_back(LoLd);
9766 Ops.push_back(HiLd);
9767 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9771 LoAddr = St->getBasePtr();
9772 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9773 DAG.getConstant(4, MVT::i32));
9775 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9776 St->getSrcValue(), St->getSrcValueOffset(),
9777 St->isVolatile(), St->isNonTemporal(),
9778 St->getAlignment());
9779 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9781 St->getSrcValueOffset() + 4,
9783 St->isNonTemporal(),
9784 MinAlign(St->getAlignment(), 4));
9785 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9790 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9791 /// X86ISD::FXOR nodes.
9792 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9793 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9794 // F[X]OR(0.0, x) -> x
9795 // F[X]OR(x, 0.0) -> x
9796 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9797 if (C->getValueAPF().isPosZero())
9798 return N->getOperand(1);
9799 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9800 if (C->getValueAPF().isPosZero())
9801 return N->getOperand(0);
9805 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9806 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9807 // FAND(0.0, x) -> 0.0
9808 // FAND(x, 0.0) -> 0.0
9809 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9810 if (C->getValueAPF().isPosZero())
9811 return N->getOperand(0);
9812 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9813 if (C->getValueAPF().isPosZero())
9814 return N->getOperand(1);
9818 static SDValue PerformBTCombine(SDNode *N,
9820 TargetLowering::DAGCombinerInfo &DCI) {
9821 // BT ignores high bits in the bit index operand.
9822 SDValue Op1 = N->getOperand(1);
9823 if (Op1.hasOneUse()) {
9824 unsigned BitWidth = Op1.getValueSizeInBits();
9825 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9826 APInt KnownZero, KnownOne;
9827 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9828 !DCI.isBeforeLegalizeOps());
9829 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9830 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9831 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9832 DCI.CommitTargetLoweringOpt(TLO);
9837 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9838 SDValue Op = N->getOperand(0);
9839 if (Op.getOpcode() == ISD::BIT_CONVERT)
9840 Op = Op.getOperand(0);
9841 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9842 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9843 VT.getVectorElementType().getSizeInBits() ==
9844 OpVT.getVectorElementType().getSizeInBits()) {
9845 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9850 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9851 // Locked instructions, in turn, have implicit fence semantics (all memory
9852 // operations are flushed before issuing the locked instruction, and the
9853 // are not buffered), so we can fold away the common pattern of
9854 // fence-atomic-fence.
9855 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9856 SDValue atomic = N->getOperand(0);
9857 switch (atomic.getOpcode()) {
9858 case ISD::ATOMIC_CMP_SWAP:
9859 case ISD::ATOMIC_SWAP:
9860 case ISD::ATOMIC_LOAD_ADD:
9861 case ISD::ATOMIC_LOAD_SUB:
9862 case ISD::ATOMIC_LOAD_AND:
9863 case ISD::ATOMIC_LOAD_OR:
9864 case ISD::ATOMIC_LOAD_XOR:
9865 case ISD::ATOMIC_LOAD_NAND:
9866 case ISD::ATOMIC_LOAD_MIN:
9867 case ISD::ATOMIC_LOAD_MAX:
9868 case ISD::ATOMIC_LOAD_UMIN:
9869 case ISD::ATOMIC_LOAD_UMAX:
9875 SDValue fence = atomic.getOperand(0);
9876 if (fence.getOpcode() != ISD::MEMBARRIER)
9879 switch (atomic.getOpcode()) {
9880 case ISD::ATOMIC_CMP_SWAP:
9881 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9882 atomic.getOperand(1), atomic.getOperand(2),
9883 atomic.getOperand(3));
9884 case ISD::ATOMIC_SWAP:
9885 case ISD::ATOMIC_LOAD_ADD:
9886 case ISD::ATOMIC_LOAD_SUB:
9887 case ISD::ATOMIC_LOAD_AND:
9888 case ISD::ATOMIC_LOAD_OR:
9889 case ISD::ATOMIC_LOAD_XOR:
9890 case ISD::ATOMIC_LOAD_NAND:
9891 case ISD::ATOMIC_LOAD_MIN:
9892 case ISD::ATOMIC_LOAD_MAX:
9893 case ISD::ATOMIC_LOAD_UMIN:
9894 case ISD::ATOMIC_LOAD_UMAX:
9895 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9896 atomic.getOperand(1), atomic.getOperand(2));
9902 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9903 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9904 // (and (i32 x86isd::setcc_carry), 1)
9905 // This eliminates the zext. This transformation is necessary because
9906 // ISD::SETCC is always legalized to i8.
9907 DebugLoc dl = N->getDebugLoc();
9908 SDValue N0 = N->getOperand(0);
9909 EVT VT = N->getValueType(0);
9910 if (N0.getOpcode() == ISD::AND &&
9912 N0.getOperand(0).hasOneUse()) {
9913 SDValue N00 = N0.getOperand(0);
9914 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9916 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9917 if (!C || C->getZExtValue() != 1)
9919 return DAG.getNode(ISD::AND, dl, VT,
9920 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9921 N00.getOperand(0), N00.getOperand(1)),
9922 DAG.getConstant(1, VT));
9928 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9929 DAGCombinerInfo &DCI) const {
9930 SelectionDAG &DAG = DCI.DAG;
9931 switch (N->getOpcode()) {
9933 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9934 case ISD::EXTRACT_VECTOR_ELT:
9935 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9936 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9937 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9938 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9941 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9942 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9943 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9945 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9946 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9947 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9948 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9949 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9950 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9956 /// isTypeDesirableForOp - Return true if the target has native support for
9957 /// the specified value type and it is 'desirable' to use the type for the
9958 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9959 /// instruction encodings are longer and some i16 instructions are slow.
9960 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9961 if (!isTypeLegal(VT))
9963 if (!Promote16Bit || VT != MVT::i16)
9970 case ISD::SIGN_EXTEND:
9971 case ISD::ZERO_EXTEND:
9972 case ISD::ANY_EXTEND:
9988 /// IsDesirableToPromoteOp - This method query the target whether it is
9989 /// beneficial for dag combiner to promote the specified node. If true, it
9990 /// should return the desired promotion type by reference.
9991 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9995 EVT VT = Op.getValueType();
9999 bool Promote = false;
10000 bool Commute = false;
10001 switch (Op.getOpcode()) {
10004 LoadSDNode *LD = cast<LoadSDNode>(Op);
10005 // If the non-extending load has a single use and it's not live out, then it
10006 // might be folded.
10007 if (LD->getExtensionType() == ISD::NON_EXTLOAD &&
10009 Op.getNode()->use_begin()->getOpcode() != ISD::CopyToReg)
10014 case ISD::SIGN_EXTEND:
10015 case ISD::ZERO_EXTEND:
10016 case ISD::ANY_EXTEND:
10024 SDValue N0 = Op.getOperand(0);
10025 // Look out for (store (shl (load), x)).
10026 if (isa<LoadSDNode>(N0) && N0.hasOneUse() &&
10027 Op.hasOneUse() && Op.getNode()->use_begin()->getOpcode() == ISD::STORE)
10040 SDValue N0 = Op.getOperand(0);
10041 SDValue N1 = Op.getOperand(1);
10042 if (!Commute && isa<LoadSDNode>(N1))
10044 // Avoid disabling potential load folding opportunities.
10045 if ((isa<LoadSDNode>(N0) && N0.hasOneUse()) && !isa<ConstantSDNode>(N1))
10047 if ((isa<LoadSDNode>(N1) && N1.hasOneUse()) && !isa<ConstantSDNode>(N0))
10057 //===----------------------------------------------------------------------===//
10058 // X86 Inline Assembly Support
10059 //===----------------------------------------------------------------------===//
10061 static bool LowerToBSwap(CallInst *CI) {
10062 // FIXME: this should verify that we are targetting a 486 or better. If not,
10063 // we will turn this bswap into something that will be lowered to logical ops
10064 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10065 // so don't worry about this.
10067 // Verify this is a simple bswap.
10068 if (CI->getNumOperands() != 2 ||
10069 CI->getType() != CI->getOperand(1)->getType() ||
10070 !CI->getType()->isIntegerTy())
10073 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10074 if (!Ty || Ty->getBitWidth() % 16 != 0)
10077 // Okay, we can do this xform, do so now.
10078 const Type *Tys[] = { Ty };
10079 Module *M = CI->getParent()->getParent()->getParent();
10080 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10082 Value *Op = CI->getOperand(1);
10083 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10085 CI->replaceAllUsesWith(Op);
10086 CI->eraseFromParent();
10090 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10091 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10092 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10094 std::string AsmStr = IA->getAsmString();
10096 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10097 SmallVector<StringRef, 4> AsmPieces;
10098 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10100 switch (AsmPieces.size()) {
10101 default: return false;
10103 AsmStr = AsmPieces[0];
10105 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10108 if (AsmPieces.size() == 2 &&
10109 (AsmPieces[0] == "bswap" ||
10110 AsmPieces[0] == "bswapq" ||
10111 AsmPieces[0] == "bswapl") &&
10112 (AsmPieces[1] == "$0" ||
10113 AsmPieces[1] == "${0:q}")) {
10114 // No need to check constraints, nothing other than the equivalent of
10115 // "=r,0" would be valid here.
10116 return LowerToBSwap(CI);
10118 // rorw $$8, ${0:w} --> llvm.bswap.i16
10119 if (CI->getType()->isIntegerTy(16) &&
10120 AsmPieces.size() == 3 &&
10121 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10122 AsmPieces[1] == "$$8," &&
10123 AsmPieces[2] == "${0:w}" &&
10124 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10126 const std::string &Constraints = IA->getConstraintString();
10127 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10128 std::sort(AsmPieces.begin(), AsmPieces.end());
10129 if (AsmPieces.size() == 4 &&
10130 AsmPieces[0] == "~{cc}" &&
10131 AsmPieces[1] == "~{dirflag}" &&
10132 AsmPieces[2] == "~{flags}" &&
10133 AsmPieces[3] == "~{fpsr}") {
10134 return LowerToBSwap(CI);
10139 if (CI->getType()->isIntegerTy(64) &&
10140 Constraints.size() >= 2 &&
10141 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10142 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10143 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10144 SmallVector<StringRef, 4> Words;
10145 SplitString(AsmPieces[0], Words, " \t");
10146 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10148 SplitString(AsmPieces[1], Words, " \t");
10149 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10151 SplitString(AsmPieces[2], Words, " \t,");
10152 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10153 Words[2] == "%edx") {
10154 return LowerToBSwap(CI);
10166 /// getConstraintType - Given a constraint letter, return the type of
10167 /// constraint it is for this target.
10168 X86TargetLowering::ConstraintType
10169 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10170 if (Constraint.size() == 1) {
10171 switch (Constraint[0]) {
10183 return C_RegisterClass;
10191 return TargetLowering::getConstraintType(Constraint);
10194 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10195 /// with another that has more specific requirements based on the type of the
10196 /// corresponding operand.
10197 const char *X86TargetLowering::
10198 LowerXConstraint(EVT ConstraintVT) const {
10199 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10200 // 'f' like normal targets.
10201 if (ConstraintVT.isFloatingPoint()) {
10202 if (Subtarget->hasSSE2())
10204 if (Subtarget->hasSSE1())
10208 return TargetLowering::LowerXConstraint(ConstraintVT);
10211 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10212 /// vector. If it is invalid, don't add anything to Ops.
10213 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10216 std::vector<SDValue>&Ops,
10217 SelectionDAG &DAG) const {
10218 SDValue Result(0, 0);
10220 switch (Constraint) {
10223 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10224 if (C->getZExtValue() <= 31) {
10225 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10231 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10232 if (C->getZExtValue() <= 63) {
10233 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10239 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10240 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10241 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10248 if (C->getZExtValue() <= 255) {
10249 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10255 // 32-bit signed value
10256 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10257 const ConstantInt *CI = C->getConstantIntValue();
10258 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10259 C->getSExtValue())) {
10260 // Widen to 64 bits here to get it sign extended.
10261 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10264 // FIXME gcc accepts some relocatable values here too, but only in certain
10265 // memory models; it's complicated.
10270 // 32-bit unsigned value
10271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10272 const ConstantInt *CI = C->getConstantIntValue();
10273 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10274 C->getZExtValue())) {
10275 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10279 // FIXME gcc accepts some relocatable values here too, but only in certain
10280 // memory models; it's complicated.
10284 // Literal immediates are always ok.
10285 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10286 // Widen to 64 bits here to get it sign extended.
10287 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10291 // If we are in non-pic codegen mode, we allow the address of a global (with
10292 // an optional displacement) to be used with 'i'.
10293 GlobalAddressSDNode *GA = 0;
10294 int64_t Offset = 0;
10296 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10298 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10299 Offset += GA->getOffset();
10301 } else if (Op.getOpcode() == ISD::ADD) {
10302 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10303 Offset += C->getZExtValue();
10304 Op = Op.getOperand(0);
10307 } else if (Op.getOpcode() == ISD::SUB) {
10308 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10309 Offset += -C->getZExtValue();
10310 Op = Op.getOperand(0);
10315 // Otherwise, this isn't something we can handle, reject it.
10319 const GlobalValue *GV = GA->getGlobal();
10320 // If we require an extra load to get this address, as in PIC mode, we
10321 // can't accept it.
10322 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10323 getTargetMachine())))
10327 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10329 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10335 if (Result.getNode()) {
10336 Ops.push_back(Result);
10339 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10343 std::vector<unsigned> X86TargetLowering::
10344 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10346 if (Constraint.size() == 1) {
10347 // FIXME: not handling fp-stack yet!
10348 switch (Constraint[0]) { // GCC X86 Constraint Letters
10349 default: break; // Unknown constraint letter
10350 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10351 if (Subtarget->is64Bit()) {
10352 if (VT == MVT::i32)
10353 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10354 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10355 X86::R10D,X86::R11D,X86::R12D,
10356 X86::R13D,X86::R14D,X86::R15D,
10357 X86::EBP, X86::ESP, 0);
10358 else if (VT == MVT::i16)
10359 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10360 X86::SI, X86::DI, X86::R8W,X86::R9W,
10361 X86::R10W,X86::R11W,X86::R12W,
10362 X86::R13W,X86::R14W,X86::R15W,
10363 X86::BP, X86::SP, 0);
10364 else if (VT == MVT::i8)
10365 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10366 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10367 X86::R10B,X86::R11B,X86::R12B,
10368 X86::R13B,X86::R14B,X86::R15B,
10369 X86::BPL, X86::SPL, 0);
10371 else if (VT == MVT::i64)
10372 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10373 X86::RSI, X86::RDI, X86::R8, X86::R9,
10374 X86::R10, X86::R11, X86::R12,
10375 X86::R13, X86::R14, X86::R15,
10376 X86::RBP, X86::RSP, 0);
10380 // 32-bit fallthrough
10381 case 'Q': // Q_REGS
10382 if (VT == MVT::i32)
10383 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10384 else if (VT == MVT::i16)
10385 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10386 else if (VT == MVT::i8)
10387 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10388 else if (VT == MVT::i64)
10389 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10394 return std::vector<unsigned>();
10397 std::pair<unsigned, const TargetRegisterClass*>
10398 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10400 // First, see if this is a constraint that directly corresponds to an LLVM
10402 if (Constraint.size() == 1) {
10403 // GCC Constraint Letters
10404 switch (Constraint[0]) {
10406 case 'r': // GENERAL_REGS
10407 case 'l': // INDEX_REGS
10409 return std::make_pair(0U, X86::GR8RegisterClass);
10410 if (VT == MVT::i16)
10411 return std::make_pair(0U, X86::GR16RegisterClass);
10412 if (VT == MVT::i32 || !Subtarget->is64Bit())
10413 return std::make_pair(0U, X86::GR32RegisterClass);
10414 return std::make_pair(0U, X86::GR64RegisterClass);
10415 case 'R': // LEGACY_REGS
10417 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10418 if (VT == MVT::i16)
10419 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10420 if (VT == MVT::i32 || !Subtarget->is64Bit())
10421 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10422 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10423 case 'f': // FP Stack registers.
10424 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10425 // value to the correct fpstack register class.
10426 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10427 return std::make_pair(0U, X86::RFP32RegisterClass);
10428 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10429 return std::make_pair(0U, X86::RFP64RegisterClass);
10430 return std::make_pair(0U, X86::RFP80RegisterClass);
10431 case 'y': // MMX_REGS if MMX allowed.
10432 if (!Subtarget->hasMMX()) break;
10433 return std::make_pair(0U, X86::VR64RegisterClass);
10434 case 'Y': // SSE_REGS if SSE2 allowed
10435 if (!Subtarget->hasSSE2()) break;
10437 case 'x': // SSE_REGS if SSE1 allowed
10438 if (!Subtarget->hasSSE1()) break;
10440 switch (VT.getSimpleVT().SimpleTy) {
10442 // Scalar SSE types.
10445 return std::make_pair(0U, X86::FR32RegisterClass);
10448 return std::make_pair(0U, X86::FR64RegisterClass);
10456 return std::make_pair(0U, X86::VR128RegisterClass);
10462 // Use the default implementation in TargetLowering to convert the register
10463 // constraint into a member of a register class.
10464 std::pair<unsigned, const TargetRegisterClass*> Res;
10465 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10467 // Not found as a standard register?
10468 if (Res.second == 0) {
10469 // Map st(0) -> st(7) -> ST0
10470 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10471 tolower(Constraint[1]) == 's' &&
10472 tolower(Constraint[2]) == 't' &&
10473 Constraint[3] == '(' &&
10474 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10475 Constraint[5] == ')' &&
10476 Constraint[6] == '}') {
10478 Res.first = X86::ST0+Constraint[4]-'0';
10479 Res.second = X86::RFP80RegisterClass;
10483 // GCC allows "st(0)" to be called just plain "st".
10484 if (StringRef("{st}").equals_lower(Constraint)) {
10485 Res.first = X86::ST0;
10486 Res.second = X86::RFP80RegisterClass;
10491 if (StringRef("{flags}").equals_lower(Constraint)) {
10492 Res.first = X86::EFLAGS;
10493 Res.second = X86::CCRRegisterClass;
10497 // 'A' means EAX + EDX.
10498 if (Constraint == "A") {
10499 Res.first = X86::EAX;
10500 Res.second = X86::GR32_ADRegisterClass;
10506 // Otherwise, check to see if this is a register class of the wrong value
10507 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10508 // turn into {ax},{dx}.
10509 if (Res.second->hasType(VT))
10510 return Res; // Correct type already, nothing to do.
10512 // All of the single-register GCC register classes map their values onto
10513 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10514 // really want an 8-bit or 32-bit register, map to the appropriate register
10515 // class and return the appropriate register.
10516 if (Res.second == X86::GR16RegisterClass) {
10517 if (VT == MVT::i8) {
10518 unsigned DestReg = 0;
10519 switch (Res.first) {
10521 case X86::AX: DestReg = X86::AL; break;
10522 case X86::DX: DestReg = X86::DL; break;
10523 case X86::CX: DestReg = X86::CL; break;
10524 case X86::BX: DestReg = X86::BL; break;
10527 Res.first = DestReg;
10528 Res.second = X86::GR8RegisterClass;
10530 } else if (VT == MVT::i32) {
10531 unsigned DestReg = 0;
10532 switch (Res.first) {
10534 case X86::AX: DestReg = X86::EAX; break;
10535 case X86::DX: DestReg = X86::EDX; break;
10536 case X86::CX: DestReg = X86::ECX; break;
10537 case X86::BX: DestReg = X86::EBX; break;
10538 case X86::SI: DestReg = X86::ESI; break;
10539 case X86::DI: DestReg = X86::EDI; break;
10540 case X86::BP: DestReg = X86::EBP; break;
10541 case X86::SP: DestReg = X86::ESP; break;
10544 Res.first = DestReg;
10545 Res.second = X86::GR32RegisterClass;
10547 } else if (VT == MVT::i64) {
10548 unsigned DestReg = 0;
10549 switch (Res.first) {
10551 case X86::AX: DestReg = X86::RAX; break;
10552 case X86::DX: DestReg = X86::RDX; break;
10553 case X86::CX: DestReg = X86::RCX; break;
10554 case X86::BX: DestReg = X86::RBX; break;
10555 case X86::SI: DestReg = X86::RSI; break;
10556 case X86::DI: DestReg = X86::RDI; break;
10557 case X86::BP: DestReg = X86::RBP; break;
10558 case X86::SP: DestReg = X86::RSP; break;
10561 Res.first = DestReg;
10562 Res.second = X86::GR64RegisterClass;
10565 } else if (Res.second == X86::FR32RegisterClass ||
10566 Res.second == X86::FR64RegisterClass ||
10567 Res.second == X86::VR128RegisterClass) {
10568 // Handle references to XMM physical registers that got mapped into the
10569 // wrong class. This can happen with constraints like {xmm0} where the
10570 // target independent register mapper will just pick the first match it can
10571 // find, ignoring the required type.
10572 if (VT == MVT::f32)
10573 Res.second = X86::FR32RegisterClass;
10574 else if (VT == MVT::f64)
10575 Res.second = X86::FR64RegisterClass;
10576 else if (X86::VR128RegisterClass->hasType(VT))
10577 Res.second = X86::VR128RegisterClass;