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 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
65 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
66 default: llvm_unreachable("unknown subtarget type");
67 case X86Subtarget::isDarwin:
68 if (TM.getSubtarget<X86Subtarget>().is64Bit())
69 return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 case X86Subtarget::isELF:
72 if (TM.getSubtarget<X86Subtarget>().is64Bit())
73 return new X8664_ELFTargetObjectFile(TM);
74 return new X8632_ELFTargetObjectFile(TM);
75 case X86Subtarget::isMingw:
76 case X86Subtarget::isCygwin:
77 case X86Subtarget::isWindows:
78 return new TargetLoweringObjectFileCOFF();
82 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
83 : TargetLowering(TM, createTLOF(TM)) {
84 Subtarget = &TM.getSubtarget<X86Subtarget>();
85 X86ScalarSSEf64 = Subtarget->hasSSE2();
86 X86ScalarSSEf32 = Subtarget->hasSSE1();
87 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
89 RegInfo = TM.getRegisterInfo();
92 // Set up the TargetLowering object.
94 // X86 is weird, it always uses i8 for shift amounts and setcc results.
95 setShiftAmountType(MVT::i8);
96 setBooleanContents(ZeroOrOneBooleanContent);
97 setSchedulingPreference(Sched::RegPressure);
98 setStackPointerRegisterToSaveRestore(X86StackPtr);
100 if (Subtarget->isTargetDarwin()) {
101 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
102 setUseUnderscoreSetJmp(false);
103 setUseUnderscoreLongJmp(false);
104 } else if (Subtarget->isTargetMingw()) {
105 // MS runtime is weird: it exports _setjmp, but longjmp!
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(false);
109 setUseUnderscoreSetJmp(true);
110 setUseUnderscoreLongJmp(true);
113 // Set up the register classes.
114 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
115 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
116 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
117 if (Subtarget->is64Bit())
118 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
120 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
122 // We don't accept any truncstore of integer registers.
123 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 // We have an algorithm for SSE2->double, and we turn this into a
149 // 64-bit FILD followed by conditional FADD for other targets.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
151 // We have an algorithm for SSE2, and we turn this into a 64-bit
152 // FILD for other targets.
153 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
156 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
158 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
159 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
162 // SSE has no i16 to fp conversion, only i32
163 if (X86ScalarSSEf32) {
164 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
165 // f32 and f64 cases are Legal, f80 case is not
166 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
172 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
176 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
177 // are Legal, f80 is custom lowered.
178 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
179 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
181 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
183 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
184 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
186 if (X86ScalarSSEf32) {
187 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
188 // f32 and f64 cases are Legal, f80 case is not
189 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
191 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
195 // Handle FP_TO_UINT by promoting the destination to a larger signed
197 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
198 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
201 if (Subtarget->is64Bit()) {
202 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
203 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
204 } else if (!UseSoftFloat) {
205 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
206 // Expand FP_TO_UINT into a select.
207 // FIXME: We would like to use a Custom expander here eventually to do
208 // the optimal thing for SSE vs. the default expansion in the legalizer.
209 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
211 // With SSE3 we can use fisttpll to convert to a signed i64; without
212 // SSE, we're stuck with a fistpll.
213 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
216 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
217 if (!X86ScalarSSEf64) {
218 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
219 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
220 if (Subtarget->is64Bit()) {
221 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
222 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
223 if (Subtarget->hasMMX() && !DisableMMX)
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
226 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
230 // Scalar integer divide and remainder are lowered to use operations that
231 // produce two results, to match the available instructions. This exposes
232 // the two-result form to trivial CSE, which is able to combine x/y and x%y
233 // into a single instruction.
235 // Scalar integer multiply-high is also lowered to use two-result
236 // operations, to match the available instructions. However, plain multiply
237 // (low) operations are left as Legal, as there are single-result
238 // instructions for this in x86. Using the two-result multiply instructions
239 // when both high and low results are needed must be arranged by dagcombine.
240 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
241 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
242 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
243 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
244 setOperationAction(ISD::SREM , MVT::i8 , Expand);
245 setOperationAction(ISD::UREM , MVT::i8 , Expand);
246 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
247 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
248 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
249 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
250 setOperationAction(ISD::SREM , MVT::i16 , Expand);
251 setOperationAction(ISD::UREM , MVT::i16 , Expand);
252 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
253 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
254 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
255 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
256 setOperationAction(ISD::SREM , MVT::i32 , Expand);
257 setOperationAction(ISD::UREM , MVT::i32 , Expand);
258 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
259 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
260 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
261 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
262 setOperationAction(ISD::SREM , MVT::i64 , Expand);
263 setOperationAction(ISD::UREM , MVT::i64 , Expand);
265 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
266 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
267 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
268 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
269 if (Subtarget->is64Bit())
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
274 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
275 setOperationAction(ISD::FREM , MVT::f32 , Expand);
276 setOperationAction(ISD::FREM , MVT::f64 , Expand);
277 setOperationAction(ISD::FREM , MVT::f80 , Expand);
278 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
280 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
283 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
284 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
285 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
286 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
287 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
288 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
289 if (Subtarget->is64Bit()) {
290 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
291 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
292 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
295 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
296 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
298 // These should be promoted to a larger select which is supported.
299 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
300 // X86 wants to expand cmov itself.
301 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
302 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
303 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
309 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
311 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
312 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
313 if (Subtarget->is64Bit()) {
314 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
315 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
317 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
320 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
321 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
322 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
324 if (Subtarget->is64Bit())
325 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
326 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
327 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
328 if (Subtarget->is64Bit()) {
329 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
330 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
331 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
332 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
333 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
335 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
336 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
337 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
338 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
339 if (Subtarget->is64Bit()) {
340 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
345 if (Subtarget->hasSSE1())
346 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
348 if (!Subtarget->hasSSE2())
349 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
351 // Expand certain atomics
352 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
353 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
354 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
355 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
357 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
362 if (!Subtarget->is64Bit()) {
363 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
372 // FIXME - use subtarget debug flags
373 if (!Subtarget->isTargetDarwin() &&
374 !Subtarget->isTargetELF() &&
375 !Subtarget->isTargetCygMing()) {
376 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
379 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
380 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
381 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
382 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
383 if (Subtarget->is64Bit()) {
384 setExceptionPointerRegister(X86::RAX);
385 setExceptionSelectorRegister(X86::RDX);
387 setExceptionPointerRegister(X86::EAX);
388 setExceptionSelectorRegister(X86::EDX);
390 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
391 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
393 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
395 setOperationAction(ISD::TRAP, MVT::Other, Legal);
397 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
398 setOperationAction(ISD::VASTART , MVT::Other, Custom);
399 setOperationAction(ISD::VAEND , MVT::Other, Expand);
400 if (Subtarget->is64Bit()) {
401 setOperationAction(ISD::VAARG , MVT::Other, Custom);
402 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
404 setOperationAction(ISD::VAARG , MVT::Other, Expand);
405 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
408 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
409 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
410 if (Subtarget->is64Bit())
411 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
412 if (Subtarget->isTargetCygMing())
413 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
417 if (!UseSoftFloat && X86ScalarSSEf64) {
418 // f32 and f64 use SSE.
419 // Set up the FP register classes.
420 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
421 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
423 // Use ANDPD to simulate FABS.
424 setOperationAction(ISD::FABS , MVT::f64, Custom);
425 setOperationAction(ISD::FABS , MVT::f32, Custom);
427 // Use XORP to simulate FNEG.
428 setOperationAction(ISD::FNEG , MVT::f64, Custom);
429 setOperationAction(ISD::FNEG , MVT::f32, Custom);
431 // Use ANDPD and ORPD to simulate FCOPYSIGN.
432 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
433 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
435 // We don't support sin/cos/fmod
436 setOperationAction(ISD::FSIN , MVT::f64, Expand);
437 setOperationAction(ISD::FCOS , MVT::f64, Expand);
438 setOperationAction(ISD::FSIN , MVT::f32, Expand);
439 setOperationAction(ISD::FCOS , MVT::f32, Expand);
441 // Expand FP immediates into loads from the stack, except for the special
443 addLegalFPImmediate(APFloat(+0.0)); // xorpd
444 addLegalFPImmediate(APFloat(+0.0f)); // xorps
445 } else if (!UseSoftFloat && X86ScalarSSEf32) {
446 // Use SSE for f32, x87 for f64.
447 // Set up the FP register classes.
448 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
449 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
451 // Use ANDPS to simulate FABS.
452 setOperationAction(ISD::FABS , MVT::f32, Custom);
454 // Use XORP to simulate FNEG.
455 setOperationAction(ISD::FNEG , MVT::f32, Custom);
457 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
459 // Use ANDPS and ORPS to simulate FCOPYSIGN.
460 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
461 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
463 // We don't support sin/cos/fmod
464 setOperationAction(ISD::FSIN , MVT::f32, Expand);
465 setOperationAction(ISD::FCOS , MVT::f32, Expand);
467 // Special cases we handle for FP constants.
468 addLegalFPImmediate(APFloat(+0.0f)); // xorps
469 addLegalFPImmediate(APFloat(+0.0)); // FLD0
470 addLegalFPImmediate(APFloat(+1.0)); // FLD1
471 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
472 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
475 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
476 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
478 } else if (!UseSoftFloat) {
479 // f32 and f64 in x87.
480 // Set up the FP register classes.
481 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
482 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
484 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
485 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
486 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
487 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 addLegalFPImmediate(APFloat(+0.0)); // FLD0
494 addLegalFPImmediate(APFloat(+1.0)); // FLD1
495 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
496 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
497 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
503 // Long double always uses X87.
505 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
506 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
510 APFloat TmpFlt(+0.0);
511 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
513 addLegalFPImmediate(TmpFlt); // FLD0
515 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
516 APFloat TmpFlt2(+1.0);
517 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt2); // FLD1
520 TmpFlt2.changeSign();
521 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
525 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
526 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
530 // Always use a library call for pow.
531 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
533 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
535 setOperationAction(ISD::FLOG, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
537 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP, MVT::f80, Expand);
539 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
541 // First set operation action for all vector types to either promote
542 // (for widening) or expand (for scalarization). Then we will selectively
543 // turn on ones that can be effectively codegen'd.
544 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
545 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
546 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
561 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
562 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
595 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
599 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
600 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
601 setTruncStoreAction((MVT::SimpleValueType)VT,
602 (MVT::SimpleValueType)InnerVT, Expand);
603 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
604 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
605 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
609 // with -msoft-float, disable use of MMX as well.
610 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
611 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
612 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
613 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
614 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
615 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
617 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
618 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
619 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
620 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
622 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
623 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
624 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
625 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
627 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
628 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
630 setOperationAction(ISD::AND, MVT::v8i8, Promote);
631 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
632 setOperationAction(ISD::AND, MVT::v4i16, Promote);
633 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
634 setOperationAction(ISD::AND, MVT::v2i32, Promote);
635 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v1i64, Legal);
638 setOperationAction(ISD::OR, MVT::v8i8, Promote);
639 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
640 setOperationAction(ISD::OR, MVT::v4i16, Promote);
641 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
642 setOperationAction(ISD::OR, MVT::v2i32, Promote);
643 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v1i64, Legal);
646 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
647 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
648 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
649 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
650 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
654 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
655 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
656 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
657 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
658 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
664 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
665 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
670 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
671 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
675 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
676 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
680 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
682 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
683 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
684 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
685 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
690 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
691 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
692 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2f32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector()) {
799 setOperationAction(ISD::AND, SVT, Promote);
800 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
801 setOperationAction(ISD::OR, SVT, Promote);
802 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
803 setOperationAction(ISD::XOR, SVT, Promote);
804 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
805 setOperationAction(ISD::LOAD, SVT, Promote);
806 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
807 setOperationAction(ISD::SELECT, SVT, Promote);
808 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
811 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
813 // Custom lower v2i64 and v2f64 selects.
814 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
815 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
816 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
817 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
819 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
820 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
821 if (!DisableMMX && Subtarget->hasMMX()) {
822 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
827 if (Subtarget->hasSSE41()) {
828 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
829 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
830 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
831 setOperationAction(ISD::FRINT, MVT::f32, Legal);
832 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
833 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
834 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
835 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
836 setOperationAction(ISD::FRINT, MVT::f64, Legal);
837 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
839 // FIXME: Do we need to handle scalar-to-vector here?
840 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
842 // i8 and i16 vectors are custom , because the source register and source
843 // source memory operand types are not the same width. f32 vectors are
844 // custom since the immediate controlling the insert encodes additional
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
852 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
853 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
856 if (Subtarget->is64Bit()) {
857 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
858 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
862 if (Subtarget->hasSSE42()) {
863 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
866 if (!UseSoftFloat && Subtarget->hasAVX()) {
867 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
868 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
869 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
870 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
872 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
873 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
874 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
875 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
876 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
877 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
878 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
879 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
880 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
881 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
882 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
883 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
884 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
885 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
886 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
888 // Operations to consider commented out -v16i16 v32i8
889 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
890 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
891 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
892 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
893 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
894 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
895 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
896 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
897 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
898 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
899 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
900 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
901 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
902 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
904 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
905 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
906 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
907 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
909 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
910 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
911 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
913 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
916 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
917 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
918 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
919 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
920 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
923 // Not sure we want to do this since there are no 256-bit integer
926 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
927 // This includes 256-bit vectors
928 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
929 EVT VT = (MVT::SimpleValueType)i;
931 // Do not attempt to custom lower non-power-of-2 vectors
932 if (!isPowerOf2_32(VT.getVectorNumElements()))
935 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
936 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
937 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
940 if (Subtarget->is64Bit()) {
941 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
942 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
947 // Not sure we want to do this since there are no 256-bit integer
950 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
951 // Including 256-bit vectors
952 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
953 EVT VT = (MVT::SimpleValueType)i;
955 if (!VT.is256BitVector()) {
958 setOperationAction(ISD::AND, VT, Promote);
959 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
960 setOperationAction(ISD::OR, VT, Promote);
961 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
962 setOperationAction(ISD::XOR, VT, Promote);
963 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
964 setOperationAction(ISD::LOAD, VT, Promote);
965 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
966 setOperationAction(ISD::SELECT, VT, Promote);
967 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
970 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
974 // We want to custom lower some of our intrinsics.
975 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
977 // Add/Sub/Mul with overflow operations are custom lowered.
978 setOperationAction(ISD::SADDO, MVT::i32, Custom);
979 setOperationAction(ISD::UADDO, MVT::i32, Custom);
980 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
981 setOperationAction(ISD::USUBO, MVT::i32, Custom);
982 setOperationAction(ISD::SMULO, MVT::i32, Custom);
984 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
985 // handle type legalization for these operations here.
987 // FIXME: We really should do custom legalization for addition and
988 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
989 // than generic legalization for 64-bit multiplication-with-overflow, though.
990 if (Subtarget->is64Bit()) {
991 setOperationAction(ISD::SADDO, MVT::i64, Custom);
992 setOperationAction(ISD::UADDO, MVT::i64, Custom);
993 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
994 setOperationAction(ISD::USUBO, MVT::i64, Custom);
995 setOperationAction(ISD::SMULO, MVT::i64, Custom);
998 if (!Subtarget->is64Bit()) {
999 // These libcalls are not available in 32-bit.
1000 setLibcallName(RTLIB::SHL_I128, 0);
1001 setLibcallName(RTLIB::SRL_I128, 0);
1002 setLibcallName(RTLIB::SRA_I128, 0);
1005 // We have target-specific dag combine patterns for the following nodes:
1006 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1007 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1008 setTargetDAGCombine(ISD::BUILD_VECTOR);
1009 setTargetDAGCombine(ISD::SELECT);
1010 setTargetDAGCombine(ISD::SHL);
1011 setTargetDAGCombine(ISD::SRA);
1012 setTargetDAGCombine(ISD::SRL);
1013 setTargetDAGCombine(ISD::OR);
1014 setTargetDAGCombine(ISD::STORE);
1015 setTargetDAGCombine(ISD::MEMBARRIER);
1016 setTargetDAGCombine(ISD::ZERO_EXTEND);
1017 if (Subtarget->is64Bit())
1018 setTargetDAGCombine(ISD::MUL);
1020 computeRegisterProperties();
1022 // FIXME: These should be based on subtarget info. Plus, the values should
1023 // be smaller when we are in optimizing for size mode.
1024 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1025 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1026 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1027 setPrefLoopAlignment(16);
1028 benefitFromCodePlacementOpt = true;
1032 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1037 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1038 /// the desired ByVal argument alignment.
1039 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1042 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1043 if (VTy->getBitWidth() == 128)
1045 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1046 unsigned EltAlign = 0;
1047 getMaxByValAlign(ATy->getElementType(), EltAlign);
1048 if (EltAlign > MaxAlign)
1049 MaxAlign = EltAlign;
1050 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1051 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1052 unsigned EltAlign = 0;
1053 getMaxByValAlign(STy->getElementType(i), EltAlign);
1054 if (EltAlign > MaxAlign)
1055 MaxAlign = EltAlign;
1063 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1064 /// function arguments in the caller parameter area. For X86, aggregates
1065 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1066 /// are at 4-byte boundaries.
1067 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1068 if (Subtarget->is64Bit()) {
1069 // Max of 8 and alignment of type.
1070 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1077 if (Subtarget->hasSSE1())
1078 getMaxByValAlign(Ty, Align);
1082 /// getOptimalMemOpType - Returns the target specific optimal type for load
1083 /// and store operations as a result of memset, memcpy, and memmove
1084 /// lowering. If DstAlign is zero that means it's safe to destination
1085 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1086 /// means there isn't a need to check it against alignment requirement,
1087 /// probably because the source does not need to be loaded. If
1088 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1089 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1090 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1091 /// constant so it does not need to be loaded.
1092 /// It returns EVT::Other if the type should be determined using generic
1093 /// target-independent logic.
1095 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1096 unsigned DstAlign, unsigned SrcAlign,
1097 bool NonScalarIntSafe,
1099 MachineFunction &MF) const {
1100 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1101 // linux. This is because the stack realignment code can't handle certain
1102 // cases like PR2962. This should be removed when PR2962 is fixed.
1103 const Function *F = MF.getFunction();
1104 if (NonScalarIntSafe &&
1105 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1107 (Subtarget->isUnalignedMemAccessFast() ||
1108 ((DstAlign == 0 || DstAlign >= 16) &&
1109 (SrcAlign == 0 || SrcAlign >= 16))) &&
1110 Subtarget->getStackAlignment() >= 16) {
1111 if (Subtarget->hasSSE2())
1113 if (Subtarget->hasSSE1())
1115 } else if (!MemcpyStrSrc && Size >= 8 &&
1116 !Subtarget->is64Bit() &&
1117 Subtarget->getStackAlignment() >= 8 &&
1118 Subtarget->hasSSE2()) {
1119 // Do not use f64 to lower memcpy if source is string constant. It's
1120 // better to use i32 to avoid the loads.
1124 if (Subtarget->is64Bit() && Size >= 8)
1129 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1130 /// current function. The returned value is a member of the
1131 /// MachineJumpTableInfo::JTEntryKind enum.
1132 unsigned X86TargetLowering::getJumpTableEncoding() const {
1133 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1135 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1136 Subtarget->isPICStyleGOT())
1137 return MachineJumpTableInfo::EK_Custom32;
1139 // Otherwise, use the normal jump table encoding heuristics.
1140 return TargetLowering::getJumpTableEncoding();
1143 /// getPICBaseSymbol - Return the X86-32 PIC base.
1145 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1146 MCContext &Ctx) const {
1147 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1148 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1149 Twine(MF->getFunctionNumber())+"$pb");
1154 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1155 const MachineBasicBlock *MBB,
1156 unsigned uid,MCContext &Ctx) const{
1157 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1158 Subtarget->isPICStyleGOT());
1159 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1161 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1162 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1165 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1167 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1168 SelectionDAG &DAG) const {
1169 if (!Subtarget->is64Bit())
1170 // This doesn't have DebugLoc associated with it, but is not really the
1171 // same as a Register.
1172 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1176 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1177 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1179 const MCExpr *X86TargetLowering::
1180 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1181 MCContext &Ctx) const {
1182 // X86-64 uses RIP relative addressing based on the jump table label.
1183 if (Subtarget->isPICStyleRIPRel())
1184 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1186 // Otherwise, the reference is relative to the PIC base.
1187 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1190 /// getFunctionAlignment - Return the Log2 alignment of this function.
1191 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1192 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1195 //===----------------------------------------------------------------------===//
1196 // Return Value Calling Convention Implementation
1197 //===----------------------------------------------------------------------===//
1199 #include "X86GenCallingConv.inc"
1202 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1203 const SmallVectorImpl<EVT> &OutTys,
1204 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1205 SelectionDAG &DAG) const {
1206 SmallVector<CCValAssign, 16> RVLocs;
1207 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1208 RVLocs, *DAG.getContext());
1209 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1213 X86TargetLowering::LowerReturn(SDValue Chain,
1214 CallingConv::ID CallConv, bool isVarArg,
1215 const SmallVectorImpl<ISD::OutputArg> &Outs,
1216 DebugLoc dl, SelectionDAG &DAG) const {
1217 MachineFunction &MF = DAG.getMachineFunction();
1218 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1220 SmallVector<CCValAssign, 16> RVLocs;
1221 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1222 RVLocs, *DAG.getContext());
1223 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1225 // Add the regs to the liveout set for the function.
1226 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1227 for (unsigned i = 0; i != RVLocs.size(); ++i)
1228 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1229 MRI.addLiveOut(RVLocs[i].getLocReg());
1233 SmallVector<SDValue, 6> RetOps;
1234 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1235 // Operand #1 = Bytes To Pop
1236 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1239 // Copy the result values into the output registers.
1240 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1241 CCValAssign &VA = RVLocs[i];
1242 assert(VA.isRegLoc() && "Can only return in registers!");
1243 SDValue ValToCopy = Outs[i].Val;
1245 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1246 // the RET instruction and handled by the FP Stackifier.
1247 if (VA.getLocReg() == X86::ST0 ||
1248 VA.getLocReg() == X86::ST1) {
1249 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1250 // change the value to the FP stack register class.
1251 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1252 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1253 RetOps.push_back(ValToCopy);
1254 // Don't emit a copytoreg.
1258 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1259 // which is returned in RAX / RDX.
1260 if (Subtarget->is64Bit()) {
1261 EVT ValVT = ValToCopy.getValueType();
1262 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1263 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1264 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1265 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1269 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1270 Flag = Chain.getValue(1);
1273 // The x86-64 ABI for returning structs by value requires that we copy
1274 // the sret argument into %rax for the return. We saved the argument into
1275 // a virtual register in the entry block, so now we copy the value out
1277 if (Subtarget->is64Bit() &&
1278 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1279 MachineFunction &MF = DAG.getMachineFunction();
1280 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1281 unsigned Reg = FuncInfo->getSRetReturnReg();
1283 "SRetReturnReg should have been set in LowerFormalArguments().");
1284 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1286 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1287 Flag = Chain.getValue(1);
1289 // RAX now acts like a return value.
1290 MRI.addLiveOut(X86::RAX);
1293 RetOps[0] = Chain; // Update chain.
1295 // Add the flag if we have it.
1297 RetOps.push_back(Flag);
1299 return DAG.getNode(X86ISD::RET_FLAG, dl,
1300 MVT::Other, &RetOps[0], RetOps.size());
1303 /// LowerCallResult - Lower the result values of a call into the
1304 /// appropriate copies out of appropriate physical registers.
1307 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1308 CallingConv::ID CallConv, bool isVarArg,
1309 const SmallVectorImpl<ISD::InputArg> &Ins,
1310 DebugLoc dl, SelectionDAG &DAG,
1311 SmallVectorImpl<SDValue> &InVals) const {
1313 // Assign locations to each value returned by this call.
1314 SmallVector<CCValAssign, 16> RVLocs;
1315 bool Is64Bit = Subtarget->is64Bit();
1316 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1317 RVLocs, *DAG.getContext());
1318 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1320 // Copy all of the result registers out of their specified physreg.
1321 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1322 CCValAssign &VA = RVLocs[i];
1323 EVT CopyVT = VA.getValVT();
1325 // If this is x86-64, and we disabled SSE, we can't return FP values
1326 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1327 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1328 report_fatal_error("SSE register return with SSE disabled");
1331 // If this is a call to a function that returns an fp value on the floating
1332 // point stack, but where we prefer to use the value in xmm registers, copy
1333 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1334 if ((VA.getLocReg() == X86::ST0 ||
1335 VA.getLocReg() == X86::ST1) &&
1336 isScalarFPTypeInSSEReg(VA.getValVT())) {
1341 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1342 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1343 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1344 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1345 MVT::v2i64, InFlag).getValue(1);
1346 Val = Chain.getValue(0);
1347 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1348 Val, DAG.getConstant(0, MVT::i64));
1350 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1351 MVT::i64, InFlag).getValue(1);
1352 Val = Chain.getValue(0);
1354 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1356 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1357 CopyVT, InFlag).getValue(1);
1358 Val = Chain.getValue(0);
1360 InFlag = Chain.getValue(2);
1362 if (CopyVT != VA.getValVT()) {
1363 // Round the F80 the right size, which also moves to the appropriate xmm
1365 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1366 // This truncation won't change the value.
1367 DAG.getIntPtrConstant(1));
1370 InVals.push_back(Val);
1377 //===----------------------------------------------------------------------===//
1378 // C & StdCall & Fast Calling Convention implementation
1379 //===----------------------------------------------------------------------===//
1380 // StdCall calling convention seems to be standard for many Windows' API
1381 // routines and around. It differs from C calling convention just a little:
1382 // callee should clean up the stack, not caller. Symbols should be also
1383 // decorated in some fancy way :) It doesn't support any vector arguments.
1384 // For info on fast calling convention see Fast Calling Convention (tail call)
1385 // implementation LowerX86_32FastCCCallTo.
1387 /// CallIsStructReturn - Determines whether a call uses struct return
1389 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1393 return Outs[0].Flags.isSRet();
1396 /// ArgsAreStructReturn - Determines whether a function uses struct
1397 /// return semantics.
1399 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1403 return Ins[0].Flags.isSRet();
1406 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1407 /// given CallingConvention value.
1408 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1409 if (Subtarget->is64Bit()) {
1410 if (CC == CallingConv::GHC)
1411 return CC_X86_64_GHC;
1412 else if (Subtarget->isTargetWin64())
1413 return CC_X86_Win64_C;
1418 if (CC == CallingConv::X86_FastCall)
1419 return CC_X86_32_FastCall;
1420 else if (CC == CallingConv::X86_ThisCall)
1421 return CC_X86_32_ThisCall;
1422 else if (CC == CallingConv::Fast)
1423 return CC_X86_32_FastCC;
1424 else if (CC == CallingConv::GHC)
1425 return CC_X86_32_GHC;
1430 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1431 /// by "Src" to address "Dst" with size and alignment information specified by
1432 /// the specific parameter attribute. The copy will be passed as a byval
1433 /// function parameter.
1435 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1436 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1438 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1439 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1440 /*isVolatile*/false, /*AlwaysInline=*/true,
1444 /// IsTailCallConvention - Return true if the calling convention is one that
1445 /// supports tail call optimization.
1446 static bool IsTailCallConvention(CallingConv::ID CC) {
1447 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1450 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1451 /// a tailcall target by changing its ABI.
1452 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1453 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1457 X86TargetLowering::LowerMemArgument(SDValue Chain,
1458 CallingConv::ID CallConv,
1459 const SmallVectorImpl<ISD::InputArg> &Ins,
1460 DebugLoc dl, SelectionDAG &DAG,
1461 const CCValAssign &VA,
1462 MachineFrameInfo *MFI,
1464 // Create the nodes corresponding to a load from this parameter slot.
1465 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1466 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1467 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1470 // If value is passed by pointer we have address passed instead of the value
1472 if (VA.getLocInfo() == CCValAssign::Indirect)
1473 ValVT = VA.getLocVT();
1475 ValVT = VA.getValVT();
1477 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1478 // changed with more analysis.
1479 // In case of tail call optimization mark all arguments mutable. Since they
1480 // could be overwritten by lowering of arguments in case of a tail call.
1481 if (Flags.isByVal()) {
1482 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1483 VA.getLocMemOffset(), isImmutable, false);
1484 return DAG.getFrameIndex(FI, getPointerTy());
1486 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1487 VA.getLocMemOffset(), isImmutable, false);
1488 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1489 return DAG.getLoad(ValVT, dl, Chain, FIN,
1490 PseudoSourceValue::getFixedStack(FI), 0,
1496 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1497 CallingConv::ID CallConv,
1499 const SmallVectorImpl<ISD::InputArg> &Ins,
1502 SmallVectorImpl<SDValue> &InVals)
1504 MachineFunction &MF = DAG.getMachineFunction();
1505 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1507 const Function* Fn = MF.getFunction();
1508 if (Fn->hasExternalLinkage() &&
1509 Subtarget->isTargetCygMing() &&
1510 Fn->getName() == "main")
1511 FuncInfo->setForceFramePointer(true);
1513 MachineFrameInfo *MFI = MF.getFrameInfo();
1514 bool Is64Bit = Subtarget->is64Bit();
1515 bool IsWin64 = Subtarget->isTargetWin64();
1517 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1518 "Var args not supported with calling convention fastcc or ghc");
1520 // Assign locations to all of the incoming arguments.
1521 SmallVector<CCValAssign, 16> ArgLocs;
1522 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1523 ArgLocs, *DAG.getContext());
1524 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1526 unsigned LastVal = ~0U;
1528 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1529 CCValAssign &VA = ArgLocs[i];
1530 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1532 assert(VA.getValNo() != LastVal &&
1533 "Don't support value assigned to multiple locs yet");
1534 LastVal = VA.getValNo();
1536 if (VA.isRegLoc()) {
1537 EVT RegVT = VA.getLocVT();
1538 TargetRegisterClass *RC = NULL;
1539 if (RegVT == MVT::i32)
1540 RC = X86::GR32RegisterClass;
1541 else if (Is64Bit && RegVT == MVT::i64)
1542 RC = X86::GR64RegisterClass;
1543 else if (RegVT == MVT::f32)
1544 RC = X86::FR32RegisterClass;
1545 else if (RegVT == MVT::f64)
1546 RC = X86::FR64RegisterClass;
1547 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1548 RC = X86::VR128RegisterClass;
1549 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1550 RC = X86::VR64RegisterClass;
1552 llvm_unreachable("Unknown argument type!");
1554 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1555 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1557 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1558 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1560 if (VA.getLocInfo() == CCValAssign::SExt)
1561 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1562 DAG.getValueType(VA.getValVT()));
1563 else if (VA.getLocInfo() == CCValAssign::ZExt)
1564 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1565 DAG.getValueType(VA.getValVT()));
1566 else if (VA.getLocInfo() == CCValAssign::BCvt)
1567 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1569 if (VA.isExtInLoc()) {
1570 // Handle MMX values passed in XMM regs.
1571 if (RegVT.isVector()) {
1572 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1573 ArgValue, DAG.getConstant(0, MVT::i64));
1574 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1576 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1579 assert(VA.isMemLoc());
1580 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1583 // If value is passed via pointer - do a load.
1584 if (VA.getLocInfo() == CCValAssign::Indirect)
1585 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1588 InVals.push_back(ArgValue);
1591 // The x86-64 ABI for returning structs by value requires that we copy
1592 // the sret argument into %rax for the return. Save the argument into
1593 // a virtual register so that we can access it from the return points.
1594 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1595 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1596 unsigned Reg = FuncInfo->getSRetReturnReg();
1598 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1599 FuncInfo->setSRetReturnReg(Reg);
1601 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1602 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1605 unsigned StackSize = CCInfo.getNextStackOffset();
1606 // Align stack specially for tail calls.
1607 if (FuncIsMadeTailCallSafe(CallConv))
1608 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1610 // If the function takes variable number of arguments, make a frame index for
1611 // the start of the first vararg value... for expansion of llvm.va_start.
1613 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1614 CallConv != CallingConv::X86_ThisCall)) {
1615 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
1619 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1621 // FIXME: We should really autogenerate these arrays
1622 static const unsigned GPR64ArgRegsWin64[] = {
1623 X86::RCX, X86::RDX, X86::R8, X86::R9
1625 static const unsigned XMMArgRegsWin64[] = {
1626 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1628 static const unsigned GPR64ArgRegs64Bit[] = {
1629 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1631 static const unsigned XMMArgRegs64Bit[] = {
1632 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1633 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1635 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1638 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1639 GPR64ArgRegs = GPR64ArgRegsWin64;
1640 XMMArgRegs = XMMArgRegsWin64;
1642 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1643 GPR64ArgRegs = GPR64ArgRegs64Bit;
1644 XMMArgRegs = XMMArgRegs64Bit;
1646 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1648 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1651 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1652 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1653 "SSE register cannot be used when SSE is disabled!");
1654 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1655 "SSE register cannot be used when SSE is disabled!");
1656 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1657 // Kernel mode asks for SSE to be disabled, so don't push them
1659 TotalNumXMMRegs = 0;
1661 // For X86-64, if there are vararg parameters that are passed via
1662 // registers, then we must store them to their spots on the stack so they
1663 // may be loaded by deferencing the result of va_next.
1664 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1665 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1666 FuncInfo->setRegSaveFrameIndex(
1667 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1670 // Store the integer parameter registers.
1671 SmallVector<SDValue, 8> MemOps;
1672 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1674 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1675 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1676 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1677 DAG.getIntPtrConstant(Offset));
1678 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1679 X86::GR64RegisterClass);
1680 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1682 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1683 PseudoSourceValue::getFixedStack(
1684 FuncInfo->getRegSaveFrameIndex()),
1685 Offset, false, false, 0);
1686 MemOps.push_back(Store);
1690 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1691 // Now store the XMM (fp + vector) parameter registers.
1692 SmallVector<SDValue, 11> SaveXMMOps;
1693 SaveXMMOps.push_back(Chain);
1695 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1696 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1697 SaveXMMOps.push_back(ALVal);
1699 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1700 FuncInfo->getRegSaveFrameIndex()));
1701 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1702 FuncInfo->getVarArgsFPOffset()));
1704 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1705 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1706 X86::VR128RegisterClass);
1707 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1708 SaveXMMOps.push_back(Val);
1710 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1712 &SaveXMMOps[0], SaveXMMOps.size()));
1715 if (!MemOps.empty())
1716 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1717 &MemOps[0], MemOps.size());
1721 // Some CCs need callee pop.
1722 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1723 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1725 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1726 // If this is an sret function, the return should pop the hidden pointer.
1727 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1728 FuncInfo->setBytesToPopOnReturn(4);
1732 // RegSaveFrameIndex is X86-64 only.
1733 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1734 if (CallConv == CallingConv::X86_FastCall ||
1735 CallConv == CallingConv::X86_ThisCall)
1736 // fastcc functions can't have varargs.
1737 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1744 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1745 SDValue StackPtr, SDValue Arg,
1746 DebugLoc dl, SelectionDAG &DAG,
1747 const CCValAssign &VA,
1748 ISD::ArgFlagsTy Flags) const {
1749 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1750 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1751 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1752 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1753 if (Flags.isByVal()) {
1754 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1756 return DAG.getStore(Chain, dl, Arg, PtrOff,
1757 PseudoSourceValue::getStack(), LocMemOffset,
1761 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1762 /// optimization is performed and it is required.
1764 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1765 SDValue &OutRetAddr, SDValue Chain,
1766 bool IsTailCall, bool Is64Bit,
1767 int FPDiff, DebugLoc dl) const {
1768 // Adjust the Return address stack slot.
1769 EVT VT = getPointerTy();
1770 OutRetAddr = getReturnAddressFrameIndex(DAG);
1772 // Load the "old" Return address.
1773 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1774 return SDValue(OutRetAddr.getNode(), 1);
1777 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1778 /// optimization is performed and it is required (FPDiff!=0).
1780 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1781 SDValue Chain, SDValue RetAddrFrIdx,
1782 bool Is64Bit, int FPDiff, DebugLoc dl) {
1783 // Store the return address to the appropriate stack slot.
1784 if (!FPDiff) return Chain;
1785 // Calculate the new stack slot for the return address.
1786 int SlotSize = Is64Bit ? 8 : 4;
1787 int NewReturnAddrFI =
1788 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1789 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1790 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1791 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1792 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1798 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1799 CallingConv::ID CallConv, bool isVarArg,
1801 const SmallVectorImpl<ISD::OutputArg> &Outs,
1802 const SmallVectorImpl<ISD::InputArg> &Ins,
1803 DebugLoc dl, SelectionDAG &DAG,
1804 SmallVectorImpl<SDValue> &InVals) const {
1805 MachineFunction &MF = DAG.getMachineFunction();
1806 bool Is64Bit = Subtarget->is64Bit();
1807 bool IsStructRet = CallIsStructReturn(Outs);
1808 bool IsSibcall = false;
1811 // Check if it's really possible to do a tail call.
1812 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1813 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1816 // Sibcalls are automatically detected tailcalls which do not require
1818 if (!GuaranteedTailCallOpt && isTailCall)
1825 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1826 "Var args not supported with calling convention fastcc or ghc");
1828 // Analyze operands of the call, assigning locations to each operand.
1829 SmallVector<CCValAssign, 16> ArgLocs;
1830 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1831 ArgLocs, *DAG.getContext());
1832 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1834 // Get a count of how many bytes are to be pushed on the stack.
1835 unsigned NumBytes = CCInfo.getNextStackOffset();
1837 // This is a sibcall. The memory operands are available in caller's
1838 // own caller's stack.
1840 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1841 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1844 if (isTailCall && !IsSibcall) {
1845 // Lower arguments at fp - stackoffset + fpdiff.
1846 unsigned NumBytesCallerPushed =
1847 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1848 FPDiff = NumBytesCallerPushed - NumBytes;
1850 // Set the delta of movement of the returnaddr stackslot.
1851 // But only set if delta is greater than previous delta.
1852 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1853 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1857 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1859 SDValue RetAddrFrIdx;
1860 // Load return adress for tail calls.
1861 if (isTailCall && FPDiff)
1862 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1863 Is64Bit, FPDiff, dl);
1865 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1866 SmallVector<SDValue, 8> MemOpChains;
1869 // Walk the register/memloc assignments, inserting copies/loads. In the case
1870 // of tail call optimization arguments are handle later.
1871 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1872 CCValAssign &VA = ArgLocs[i];
1873 EVT RegVT = VA.getLocVT();
1874 SDValue Arg = Outs[i].Val;
1875 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1876 bool isByVal = Flags.isByVal();
1878 // Promote the value if needed.
1879 switch (VA.getLocInfo()) {
1880 default: llvm_unreachable("Unknown loc info!");
1881 case CCValAssign::Full: break;
1882 case CCValAssign::SExt:
1883 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1885 case CCValAssign::ZExt:
1886 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1888 case CCValAssign::AExt:
1889 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1890 // Special case: passing MMX values in XMM registers.
1891 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1892 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1893 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1895 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1897 case CCValAssign::BCvt:
1898 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1900 case CCValAssign::Indirect: {
1901 // Store the argument.
1902 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1903 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1904 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1905 PseudoSourceValue::getFixedStack(FI), 0,
1912 if (VA.isRegLoc()) {
1913 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1914 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1915 assert(VA.isMemLoc());
1916 if (StackPtr.getNode() == 0)
1917 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1918 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1919 dl, DAG, VA, Flags));
1923 if (!MemOpChains.empty())
1924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1925 &MemOpChains[0], MemOpChains.size());
1927 // Build a sequence of copy-to-reg nodes chained together with token chain
1928 // and flag operands which copy the outgoing args into registers.
1930 // Tail call byval lowering might overwrite argument registers so in case of
1931 // tail call optimization the copies to registers are lowered later.
1933 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1934 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1935 RegsToPass[i].second, InFlag);
1936 InFlag = Chain.getValue(1);
1939 if (Subtarget->isPICStyleGOT()) {
1940 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1943 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1944 DAG.getNode(X86ISD::GlobalBaseReg,
1945 DebugLoc(), getPointerTy()),
1947 InFlag = Chain.getValue(1);
1949 // If we are tail calling and generating PIC/GOT style code load the
1950 // address of the callee into ECX. The value in ecx is used as target of
1951 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1952 // for tail calls on PIC/GOT architectures. Normally we would just put the
1953 // address of GOT into ebx and then call target@PLT. But for tail calls
1954 // ebx would be restored (since ebx is callee saved) before jumping to the
1957 // Note: The actual moving to ECX is done further down.
1958 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1959 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1960 !G->getGlobal()->hasProtectedVisibility())
1961 Callee = LowerGlobalAddress(Callee, DAG);
1962 else if (isa<ExternalSymbolSDNode>(Callee))
1963 Callee = LowerExternalSymbol(Callee, DAG);
1967 if (Is64Bit && isVarArg) {
1968 // From AMD64 ABI document:
1969 // For calls that may call functions that use varargs or stdargs
1970 // (prototype-less calls or calls to functions containing ellipsis (...) in
1971 // the declaration) %al is used as hidden argument to specify the number
1972 // of SSE registers used. The contents of %al do not need to match exactly
1973 // the number of registers, but must be an ubound on the number of SSE
1974 // registers used and is in the range 0 - 8 inclusive.
1976 // FIXME: Verify this on Win64
1977 // Count the number of XMM registers allocated.
1978 static const unsigned XMMArgRegs[] = {
1979 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1980 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1982 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1983 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1984 && "SSE registers cannot be used when SSE is disabled");
1986 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1987 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1988 InFlag = Chain.getValue(1);
1992 // For tail calls lower the arguments to the 'real' stack slot.
1994 // Force all the incoming stack arguments to be loaded from the stack
1995 // before any new outgoing arguments are stored to the stack, because the
1996 // outgoing stack slots may alias the incoming argument stack slots, and
1997 // the alias isn't otherwise explicit. This is slightly more conservative
1998 // than necessary, because it means that each store effectively depends
1999 // on every argument instead of just those arguments it would clobber.
2000 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2002 SmallVector<SDValue, 8> MemOpChains2;
2005 // Do not flag preceeding copytoreg stuff together with the following stuff.
2007 if (GuaranteedTailCallOpt) {
2008 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2009 CCValAssign &VA = ArgLocs[i];
2012 assert(VA.isMemLoc());
2013 SDValue Arg = Outs[i].Val;
2014 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2015 // Create frame index.
2016 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2017 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2018 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2019 FIN = DAG.getFrameIndex(FI, getPointerTy());
2021 if (Flags.isByVal()) {
2022 // Copy relative to framepointer.
2023 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2024 if (StackPtr.getNode() == 0)
2025 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2027 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2029 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2033 // Store relative to framepointer.
2034 MemOpChains2.push_back(
2035 DAG.getStore(ArgChain, dl, Arg, FIN,
2036 PseudoSourceValue::getFixedStack(FI), 0,
2042 if (!MemOpChains2.empty())
2043 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2044 &MemOpChains2[0], MemOpChains2.size());
2046 // Copy arguments to their registers.
2047 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2048 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2049 RegsToPass[i].second, InFlag);
2050 InFlag = Chain.getValue(1);
2054 // Store the return address to the appropriate stack slot.
2055 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2059 bool WasGlobalOrExternal = false;
2060 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2061 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2062 // In the 64-bit large code model, we have to make all calls
2063 // through a register, since the call instruction's 32-bit
2064 // pc-relative offset may not be large enough to hold the whole
2066 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2067 WasGlobalOrExternal = true;
2068 // If the callee is a GlobalAddress node (quite common, every direct call
2069 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2072 // We should use extra load for direct calls to dllimported functions in
2074 const GlobalValue *GV = G->getGlobal();
2075 if (!GV->hasDLLImportLinkage()) {
2076 unsigned char OpFlags = 0;
2078 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2079 // external symbols most go through the PLT in PIC mode. If the symbol
2080 // has hidden or protected visibility, or if it is static or local, then
2081 // we don't need to use the PLT - we can directly call it.
2082 if (Subtarget->isTargetELF() &&
2083 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2084 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2085 OpFlags = X86II::MO_PLT;
2086 } else if (Subtarget->isPICStyleStubAny() &&
2087 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2088 Subtarget->getDarwinVers() < 9) {
2089 // PC-relative references to external symbols should go through $stub,
2090 // unless we're building with the leopard linker or later, which
2091 // automatically synthesizes these stubs.
2092 OpFlags = X86II::MO_DARWIN_STUB;
2095 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2096 G->getOffset(), OpFlags);
2098 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2099 WasGlobalOrExternal = true;
2100 unsigned char OpFlags = 0;
2102 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2103 // symbols should go through the PLT.
2104 if (Subtarget->isTargetELF() &&
2105 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2106 OpFlags = X86II::MO_PLT;
2107 } else if (Subtarget->isPICStyleStubAny() &&
2108 Subtarget->getDarwinVers() < 9) {
2109 // PC-relative references to external symbols should go through $stub,
2110 // unless we're building with the leopard linker or later, which
2111 // automatically synthesizes these stubs.
2112 OpFlags = X86II::MO_DARWIN_STUB;
2115 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2119 // Returns a chain & a flag for retval copy to use.
2120 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2121 SmallVector<SDValue, 8> Ops;
2123 if (!IsSibcall && isTailCall) {
2124 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2125 DAG.getIntPtrConstant(0, true), InFlag);
2126 InFlag = Chain.getValue(1);
2129 Ops.push_back(Chain);
2130 Ops.push_back(Callee);
2133 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2135 // Add argument registers to the end of the list so that they are known live
2137 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2138 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2139 RegsToPass[i].second.getValueType()));
2141 // Add an implicit use GOT pointer in EBX.
2142 if (!isTailCall && Subtarget->isPICStyleGOT())
2143 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2145 // Add an implicit use of AL for x86 vararg functions.
2146 if (Is64Bit && isVarArg)
2147 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2149 if (InFlag.getNode())
2150 Ops.push_back(InFlag);
2153 // If this is the first return lowered for this function, add the regs
2154 // to the liveout set for the function.
2155 if (MF.getRegInfo().liveout_empty()) {
2156 SmallVector<CCValAssign, 16> RVLocs;
2157 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2159 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2160 for (unsigned i = 0; i != RVLocs.size(); ++i)
2161 if (RVLocs[i].isRegLoc())
2162 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2164 return DAG.getNode(X86ISD::TC_RETURN, dl,
2165 NodeTys, &Ops[0], Ops.size());
2168 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2169 InFlag = Chain.getValue(1);
2171 // Create the CALLSEQ_END node.
2172 unsigned NumBytesForCalleeToPush;
2173 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2174 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2175 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2176 // If this is a call to a struct-return function, the callee
2177 // pops the hidden struct pointer, so we have to push it back.
2178 // This is common for Darwin/X86, Linux & Mingw32 targets.
2179 NumBytesForCalleeToPush = 4;
2181 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2183 // Returns a flag for retval copy to use.
2185 Chain = DAG.getCALLSEQ_END(Chain,
2186 DAG.getIntPtrConstant(NumBytes, true),
2187 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2190 InFlag = Chain.getValue(1);
2193 // Handle result values, copying them out of physregs into vregs that we
2195 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2196 Ins, dl, DAG, InVals);
2200 //===----------------------------------------------------------------------===//
2201 // Fast Calling Convention (tail call) implementation
2202 //===----------------------------------------------------------------------===//
2204 // Like std call, callee cleans arguments, convention except that ECX is
2205 // reserved for storing the tail called function address. Only 2 registers are
2206 // free for argument passing (inreg). Tail call optimization is performed
2208 // * tailcallopt is enabled
2209 // * caller/callee are fastcc
2210 // On X86_64 architecture with GOT-style position independent code only local
2211 // (within module) calls are supported at the moment.
2212 // To keep the stack aligned according to platform abi the function
2213 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2214 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2215 // If a tail called function callee has more arguments than the caller the
2216 // caller needs to make sure that there is room to move the RETADDR to. This is
2217 // achieved by reserving an area the size of the argument delta right after the
2218 // original REtADDR, but before the saved framepointer or the spilled registers
2219 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2231 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2232 /// for a 16 byte align requirement.
2234 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2235 SelectionDAG& DAG) const {
2236 MachineFunction &MF = DAG.getMachineFunction();
2237 const TargetMachine &TM = MF.getTarget();
2238 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2239 unsigned StackAlignment = TFI.getStackAlignment();
2240 uint64_t AlignMask = StackAlignment - 1;
2241 int64_t Offset = StackSize;
2242 uint64_t SlotSize = TD->getPointerSize();
2243 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2244 // Number smaller than 12 so just add the difference.
2245 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2247 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2248 Offset = ((~AlignMask) & Offset) + StackAlignment +
2249 (StackAlignment-SlotSize);
2254 /// MatchingStackOffset - Return true if the given stack call argument is
2255 /// already available in the same position (relatively) of the caller's
2256 /// incoming argument stack.
2258 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2259 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2260 const X86InstrInfo *TII) {
2261 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2263 if (Arg.getOpcode() == ISD::CopyFromReg) {
2264 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2265 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2267 MachineInstr *Def = MRI->getVRegDef(VR);
2270 if (!Flags.isByVal()) {
2271 if (!TII->isLoadFromStackSlot(Def, FI))
2274 unsigned Opcode = Def->getOpcode();
2275 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2276 Def->getOperand(1).isFI()) {
2277 FI = Def->getOperand(1).getIndex();
2278 Bytes = Flags.getByValSize();
2282 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2283 if (Flags.isByVal())
2284 // ByVal argument is passed in as a pointer but it's now being
2285 // dereferenced. e.g.
2286 // define @foo(%struct.X* %A) {
2287 // tail call @bar(%struct.X* byval %A)
2290 SDValue Ptr = Ld->getBasePtr();
2291 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2294 FI = FINode->getIndex();
2298 assert(FI != INT_MAX);
2299 if (!MFI->isFixedObjectIndex(FI))
2301 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2304 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2305 /// for tail call optimization. Targets which want to do tail call
2306 /// optimization should implement this function.
2308 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2309 CallingConv::ID CalleeCC,
2311 bool isCalleeStructRet,
2312 bool isCallerStructRet,
2313 const SmallVectorImpl<ISD::OutputArg> &Outs,
2314 const SmallVectorImpl<ISD::InputArg> &Ins,
2315 SelectionDAG& DAG) const {
2316 if (!IsTailCallConvention(CalleeCC) &&
2317 CalleeCC != CallingConv::C)
2320 // If -tailcallopt is specified, make fastcc functions tail-callable.
2321 const MachineFunction &MF = DAG.getMachineFunction();
2322 const Function *CallerF = DAG.getMachineFunction().getFunction();
2323 CallingConv::ID CallerCC = CallerF->getCallingConv();
2324 bool CCMatch = CallerCC == CalleeCC;
2326 if (GuaranteedTailCallOpt) {
2327 if (IsTailCallConvention(CalleeCC) && CCMatch)
2332 // Look for obvious safe cases to perform tail call optimization that do not
2333 // require ABI changes. This is what gcc calls sibcall.
2335 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2336 // emit a special epilogue.
2337 if (RegInfo->needsStackRealignment(MF))
2340 // Do not sibcall optimize vararg calls unless the call site is not passing any
2342 if (isVarArg && !Outs.empty())
2345 // Also avoid sibcall optimization if either caller or callee uses struct
2346 // return semantics.
2347 if (isCalleeStructRet || isCallerStructRet)
2350 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2351 // Therefore if it's not used by the call it is not safe to optimize this into
2353 bool Unused = false;
2354 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2361 SmallVector<CCValAssign, 16> RVLocs;
2362 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2363 RVLocs, *DAG.getContext());
2364 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2365 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2366 CCValAssign &VA = RVLocs[i];
2367 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2372 // If the calling conventions do not match, then we'd better make sure the
2373 // results are returned in the same way as what the caller expects.
2375 SmallVector<CCValAssign, 16> RVLocs1;
2376 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2377 RVLocs1, *DAG.getContext());
2378 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2380 SmallVector<CCValAssign, 16> RVLocs2;
2381 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2382 RVLocs2, *DAG.getContext());
2383 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2385 if (RVLocs1.size() != RVLocs2.size())
2387 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2388 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2390 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2392 if (RVLocs1[i].isRegLoc()) {
2393 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2396 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2402 // If the callee takes no arguments then go on to check the results of the
2404 if (!Outs.empty()) {
2405 // Check if stack adjustment is needed. For now, do not do this if any
2406 // argument is passed on the stack.
2407 SmallVector<CCValAssign, 16> ArgLocs;
2408 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2409 ArgLocs, *DAG.getContext());
2410 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2411 if (CCInfo.getNextStackOffset()) {
2412 MachineFunction &MF = DAG.getMachineFunction();
2413 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2415 if (Subtarget->isTargetWin64())
2416 // Win64 ABI has additional complications.
2419 // Check if the arguments are already laid out in the right way as
2420 // the caller's fixed stack objects.
2421 MachineFrameInfo *MFI = MF.getFrameInfo();
2422 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2423 const X86InstrInfo *TII =
2424 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2425 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2426 CCValAssign &VA = ArgLocs[i];
2427 EVT RegVT = VA.getLocVT();
2428 SDValue Arg = Outs[i].Val;
2429 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2430 if (VA.getLocInfo() == CCValAssign::Indirect)
2432 if (!VA.isRegLoc()) {
2433 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2440 // If the tailcall address may be in a register, then make sure it's
2441 // possible to register allocate for it. In 32-bit, the call address can
2442 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2443 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2444 // RDI, R8, R9, R11.
2445 if (!isa<GlobalAddressSDNode>(Callee) &&
2446 !isa<ExternalSymbolSDNode>(Callee)) {
2447 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2448 unsigned NumInRegs = 0;
2449 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2450 CCValAssign &VA = ArgLocs[i];
2451 if (VA.isRegLoc()) {
2452 if (++NumInRegs == Limit)
2463 X86TargetLowering::createFastISel(MachineFunction &mf,
2464 DenseMap<const Value *, unsigned> &vm,
2465 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2466 DenseMap<const AllocaInst *, int> &am,
2467 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2469 , SmallSet<const Instruction *, 8> &cil
2472 return X86::createFastISel(mf, vm, bm, am, pn
2480 //===----------------------------------------------------------------------===//
2481 // Other Lowering Hooks
2482 //===----------------------------------------------------------------------===//
2485 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2486 MachineFunction &MF = DAG.getMachineFunction();
2487 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2488 int ReturnAddrIndex = FuncInfo->getRAIndex();
2490 if (ReturnAddrIndex == 0) {
2491 // Set up a frame object for the return address.
2492 uint64_t SlotSize = TD->getPointerSize();
2493 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2495 FuncInfo->setRAIndex(ReturnAddrIndex);
2498 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2502 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2503 bool hasSymbolicDisplacement) {
2504 // Offset should fit into 32 bit immediate field.
2505 if (!isInt<32>(Offset))
2508 // If we don't have a symbolic displacement - we don't have any extra
2510 if (!hasSymbolicDisplacement)
2513 // FIXME: Some tweaks might be needed for medium code model.
2514 if (M != CodeModel::Small && M != CodeModel::Kernel)
2517 // For small code model we assume that latest object is 16MB before end of 31
2518 // bits boundary. We may also accept pretty large negative constants knowing
2519 // that all objects are in the positive half of address space.
2520 if (M == CodeModel::Small && Offset < 16*1024*1024)
2523 // For kernel code model we know that all object resist in the negative half
2524 // of 32bits address space. We may not accept negative offsets, since they may
2525 // be just off and we may accept pretty large positive ones.
2526 if (M == CodeModel::Kernel && Offset > 0)
2532 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2533 /// specific condition code, returning the condition code and the LHS/RHS of the
2534 /// comparison to make.
2535 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2536 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2538 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2539 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2540 // X > -1 -> X == 0, jump !sign.
2541 RHS = DAG.getConstant(0, RHS.getValueType());
2542 return X86::COND_NS;
2543 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2544 // X < 0 -> X == 0, jump on sign.
2546 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2548 RHS = DAG.getConstant(0, RHS.getValueType());
2549 return X86::COND_LE;
2553 switch (SetCCOpcode) {
2554 default: llvm_unreachable("Invalid integer condition!");
2555 case ISD::SETEQ: return X86::COND_E;
2556 case ISD::SETGT: return X86::COND_G;
2557 case ISD::SETGE: return X86::COND_GE;
2558 case ISD::SETLT: return X86::COND_L;
2559 case ISD::SETLE: return X86::COND_LE;
2560 case ISD::SETNE: return X86::COND_NE;
2561 case ISD::SETULT: return X86::COND_B;
2562 case ISD::SETUGT: return X86::COND_A;
2563 case ISD::SETULE: return X86::COND_BE;
2564 case ISD::SETUGE: return X86::COND_AE;
2568 // First determine if it is required or is profitable to flip the operands.
2570 // If LHS is a foldable load, but RHS is not, flip the condition.
2571 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2572 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2573 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2574 std::swap(LHS, RHS);
2577 switch (SetCCOpcode) {
2583 std::swap(LHS, RHS);
2587 // On a floating point condition, the flags are set as follows:
2589 // 0 | 0 | 0 | X > Y
2590 // 0 | 0 | 1 | X < Y
2591 // 1 | 0 | 0 | X == Y
2592 // 1 | 1 | 1 | unordered
2593 switch (SetCCOpcode) {
2594 default: llvm_unreachable("Condcode should be pre-legalized away");
2596 case ISD::SETEQ: return X86::COND_E;
2597 case ISD::SETOLT: // flipped
2599 case ISD::SETGT: return X86::COND_A;
2600 case ISD::SETOLE: // flipped
2602 case ISD::SETGE: return X86::COND_AE;
2603 case ISD::SETUGT: // flipped
2605 case ISD::SETLT: return X86::COND_B;
2606 case ISD::SETUGE: // flipped
2608 case ISD::SETLE: return X86::COND_BE;
2610 case ISD::SETNE: return X86::COND_NE;
2611 case ISD::SETUO: return X86::COND_P;
2612 case ISD::SETO: return X86::COND_NP;
2614 case ISD::SETUNE: return X86::COND_INVALID;
2618 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2619 /// code. Current x86 isa includes the following FP cmov instructions:
2620 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2621 static bool hasFPCMov(unsigned X86CC) {
2637 /// isFPImmLegal - Returns true if the target can instruction select the
2638 /// specified FP immediate natively. If false, the legalizer will
2639 /// materialize the FP immediate as a load from a constant pool.
2640 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2641 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2642 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2648 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2649 /// the specified range (L, H].
2650 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2651 return (Val < 0) || (Val >= Low && Val < Hi);
2654 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2655 /// specified value.
2656 static bool isUndefOrEqual(int Val, int CmpVal) {
2657 if (Val < 0 || Val == CmpVal)
2662 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2663 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2664 /// the second operand.
2665 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2666 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2667 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2668 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2669 return (Mask[0] < 2 && Mask[1] < 2);
2673 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2674 SmallVector<int, 8> M;
2676 return ::isPSHUFDMask(M, N->getValueType(0));
2679 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2680 /// is suitable for input to PSHUFHW.
2681 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2682 if (VT != MVT::v8i16)
2685 // Lower quadword copied in order or undef.
2686 for (int i = 0; i != 4; ++i)
2687 if (Mask[i] >= 0 && Mask[i] != i)
2690 // Upper quadword shuffled.
2691 for (int i = 4; i != 8; ++i)
2692 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2698 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2699 SmallVector<int, 8> M;
2701 return ::isPSHUFHWMask(M, N->getValueType(0));
2704 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2705 /// is suitable for input to PSHUFLW.
2706 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2707 if (VT != MVT::v8i16)
2710 // Upper quadword copied in order.
2711 for (int i = 4; i != 8; ++i)
2712 if (Mask[i] >= 0 && Mask[i] != i)
2715 // Lower quadword shuffled.
2716 for (int i = 0; i != 4; ++i)
2723 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2724 SmallVector<int, 8> M;
2726 return ::isPSHUFLWMask(M, N->getValueType(0));
2729 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2730 /// is suitable for input to PALIGNR.
2731 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2733 int i, e = VT.getVectorNumElements();
2735 // Do not handle v2i64 / v2f64 shuffles with palignr.
2736 if (e < 4 || !hasSSSE3)
2739 for (i = 0; i != e; ++i)
2743 // All undef, not a palignr.
2747 // Determine if it's ok to perform a palignr with only the LHS, since we
2748 // don't have access to the actual shuffle elements to see if RHS is undef.
2749 bool Unary = Mask[i] < (int)e;
2750 bool NeedsUnary = false;
2752 int s = Mask[i] - i;
2754 // Check the rest of the elements to see if they are consecutive.
2755 for (++i; i != e; ++i) {
2760 Unary = Unary && (m < (int)e);
2761 NeedsUnary = NeedsUnary || (m < s);
2763 if (NeedsUnary && !Unary)
2765 if (Unary && m != ((s+i) & (e-1)))
2767 if (!Unary && m != (s+i))
2773 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2774 SmallVector<int, 8> M;
2776 return ::isPALIGNRMask(M, N->getValueType(0), true);
2779 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2780 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2781 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2782 int NumElems = VT.getVectorNumElements();
2783 if (NumElems != 2 && NumElems != 4)
2786 int Half = NumElems / 2;
2787 for (int i = 0; i < Half; ++i)
2788 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2790 for (int i = Half; i < NumElems; ++i)
2791 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2797 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2798 SmallVector<int, 8> M;
2800 return ::isSHUFPMask(M, N->getValueType(0));
2803 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2804 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2805 /// half elements to come from vector 1 (which would equal the dest.) and
2806 /// the upper half to come from vector 2.
2807 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2808 int NumElems = VT.getVectorNumElements();
2810 if (NumElems != 2 && NumElems != 4)
2813 int Half = NumElems / 2;
2814 for (int i = 0; i < Half; ++i)
2815 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2817 for (int i = Half; i < NumElems; ++i)
2818 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2823 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2824 SmallVector<int, 8> M;
2826 return isCommutedSHUFPMask(M, N->getValueType(0));
2829 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2830 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2831 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2832 if (N->getValueType(0).getVectorNumElements() != 4)
2835 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2836 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2837 isUndefOrEqual(N->getMaskElt(1), 7) &&
2838 isUndefOrEqual(N->getMaskElt(2), 2) &&
2839 isUndefOrEqual(N->getMaskElt(3), 3);
2842 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2843 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2845 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2846 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2851 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2852 isUndefOrEqual(N->getMaskElt(1), 3) &&
2853 isUndefOrEqual(N->getMaskElt(2), 2) &&
2854 isUndefOrEqual(N->getMaskElt(3), 3);
2857 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2858 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2859 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2860 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2862 if (NumElems != 2 && NumElems != 4)
2865 for (unsigned i = 0; i < NumElems/2; ++i)
2866 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2869 for (unsigned i = NumElems/2; i < NumElems; ++i)
2870 if (!isUndefOrEqual(N->getMaskElt(i), i))
2876 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2877 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2878 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2879 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2881 if (NumElems != 2 && NumElems != 4)
2884 for (unsigned i = 0; i < NumElems/2; ++i)
2885 if (!isUndefOrEqual(N->getMaskElt(i), i))
2888 for (unsigned i = 0; i < NumElems/2; ++i)
2889 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2895 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2896 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2897 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2898 bool V2IsSplat = false) {
2899 int NumElts = VT.getVectorNumElements();
2900 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2903 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2905 int BitI1 = Mask[i+1];
2906 if (!isUndefOrEqual(BitI, j))
2909 if (!isUndefOrEqual(BitI1, NumElts))
2912 if (!isUndefOrEqual(BitI1, j + NumElts))
2919 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2920 SmallVector<int, 8> M;
2922 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2925 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2926 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2927 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2928 bool V2IsSplat = false) {
2929 int NumElts = VT.getVectorNumElements();
2930 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2933 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2935 int BitI1 = Mask[i+1];
2936 if (!isUndefOrEqual(BitI, j + NumElts/2))
2939 if (isUndefOrEqual(BitI1, NumElts))
2942 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2949 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2950 SmallVector<int, 8> M;
2952 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2955 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2956 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2958 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2959 int NumElems = VT.getVectorNumElements();
2960 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2963 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2965 int BitI1 = Mask[i+1];
2966 if (!isUndefOrEqual(BitI, j))
2968 if (!isUndefOrEqual(BitI1, j))
2974 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2975 SmallVector<int, 8> M;
2977 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2980 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2981 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2983 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2984 int NumElems = VT.getVectorNumElements();
2985 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2988 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2990 int BitI1 = Mask[i+1];
2991 if (!isUndefOrEqual(BitI, j))
2993 if (!isUndefOrEqual(BitI1, j))
2999 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3000 SmallVector<int, 8> M;
3002 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3005 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3006 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3007 /// MOVSD, and MOVD, i.e. setting the lowest element.
3008 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3009 if (VT.getVectorElementType().getSizeInBits() < 32)
3012 int NumElts = VT.getVectorNumElements();
3014 if (!isUndefOrEqual(Mask[0], NumElts))
3017 for (int i = 1; i < NumElts; ++i)
3018 if (!isUndefOrEqual(Mask[i], i))
3024 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3025 SmallVector<int, 8> M;
3027 return ::isMOVLMask(M, N->getValueType(0));
3030 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3031 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3032 /// element of vector 2 and the other elements to come from vector 1 in order.
3033 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3034 bool V2IsSplat = false, bool V2IsUndef = false) {
3035 int NumOps = VT.getVectorNumElements();
3036 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3039 if (!isUndefOrEqual(Mask[0], 0))
3042 for (int i = 1; i < NumOps; ++i)
3043 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3044 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3045 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3051 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3052 bool V2IsUndef = false) {
3053 SmallVector<int, 8> M;
3055 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3058 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3059 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3060 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3061 if (N->getValueType(0).getVectorNumElements() != 4)
3064 // Expect 1, 1, 3, 3
3065 for (unsigned i = 0; i < 2; ++i) {
3066 int Elt = N->getMaskElt(i);
3067 if (Elt >= 0 && Elt != 1)
3072 for (unsigned i = 2; i < 4; ++i) {
3073 int Elt = N->getMaskElt(i);
3074 if (Elt >= 0 && Elt != 3)
3079 // Don't use movshdup if it can be done with a shufps.
3080 // FIXME: verify that matching u, u, 3, 3 is what we want.
3084 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3085 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3086 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3087 if (N->getValueType(0).getVectorNumElements() != 4)
3090 // Expect 0, 0, 2, 2
3091 for (unsigned i = 0; i < 2; ++i)
3092 if (N->getMaskElt(i) > 0)
3096 for (unsigned i = 2; i < 4; ++i) {
3097 int Elt = N->getMaskElt(i);
3098 if (Elt >= 0 && Elt != 2)
3103 // Don't use movsldup if it can be done with a shufps.
3107 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3108 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3109 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3110 int e = N->getValueType(0).getVectorNumElements() / 2;
3112 for (int i = 0; i < e; ++i)
3113 if (!isUndefOrEqual(N->getMaskElt(i), i))
3115 for (int i = 0; i < e; ++i)
3116 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3121 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3122 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3123 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3124 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3125 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3127 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3129 for (int i = 0; i < NumOperands; ++i) {
3130 int Val = SVOp->getMaskElt(NumOperands-i-1);
3131 if (Val < 0) Val = 0;
3132 if (Val >= NumOperands) Val -= NumOperands;
3134 if (i != NumOperands - 1)
3140 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3141 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3142 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3143 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3145 // 8 nodes, but we only care about the last 4.
3146 for (unsigned i = 7; i >= 4; --i) {
3147 int Val = SVOp->getMaskElt(i);
3156 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3157 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3158 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3159 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3161 // 8 nodes, but we only care about the first 4.
3162 for (int i = 3; i >= 0; --i) {
3163 int Val = SVOp->getMaskElt(i);
3172 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3173 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3174 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3175 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3176 EVT VVT = N->getValueType(0);
3177 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3181 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3182 Val = SVOp->getMaskElt(i);
3186 return (Val - i) * EltSize;
3189 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3191 bool X86::isZeroNode(SDValue Elt) {
3192 return ((isa<ConstantSDNode>(Elt) &&
3193 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3194 (isa<ConstantFPSDNode>(Elt) &&
3195 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3198 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3199 /// their permute mask.
3200 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3201 SelectionDAG &DAG) {
3202 EVT VT = SVOp->getValueType(0);
3203 unsigned NumElems = VT.getVectorNumElements();
3204 SmallVector<int, 8> MaskVec;
3206 for (unsigned i = 0; i != NumElems; ++i) {
3207 int idx = SVOp->getMaskElt(i);
3209 MaskVec.push_back(idx);
3210 else if (idx < (int)NumElems)
3211 MaskVec.push_back(idx + NumElems);
3213 MaskVec.push_back(idx - NumElems);
3215 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3216 SVOp->getOperand(0), &MaskVec[0]);
3219 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3220 /// the two vector operands have swapped position.
3221 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3222 unsigned NumElems = VT.getVectorNumElements();
3223 for (unsigned i = 0; i != NumElems; ++i) {
3227 else if (idx < (int)NumElems)
3228 Mask[i] = idx + NumElems;
3230 Mask[i] = idx - NumElems;
3234 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3235 /// match movhlps. The lower half elements should come from upper half of
3236 /// V1 (and in order), and the upper half elements should come from the upper
3237 /// half of V2 (and in order).
3238 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3239 if (Op->getValueType(0).getVectorNumElements() != 4)
3241 for (unsigned i = 0, e = 2; i != e; ++i)
3242 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3244 for (unsigned i = 2; i != 4; ++i)
3245 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3250 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3251 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3253 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3254 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3256 N = N->getOperand(0).getNode();
3257 if (!ISD::isNON_EXTLoad(N))
3260 *LD = cast<LoadSDNode>(N);
3264 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3265 /// match movlp{s|d}. The lower half elements should come from lower half of
3266 /// V1 (and in order), and the upper half elements should come from the upper
3267 /// half of V2 (and in order). And since V1 will become the source of the
3268 /// MOVLP, it must be either a vector load or a scalar load to vector.
3269 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3270 ShuffleVectorSDNode *Op) {
3271 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3273 // Is V2 is a vector load, don't do this transformation. We will try to use
3274 // load folding shufps op.
3275 if (ISD::isNON_EXTLoad(V2))
3278 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3280 if (NumElems != 2 && NumElems != 4)
3282 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3283 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3285 for (unsigned i = NumElems/2; i != NumElems; ++i)
3286 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3291 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3293 static bool isSplatVector(SDNode *N) {
3294 if (N->getOpcode() != ISD::BUILD_VECTOR)
3297 SDValue SplatValue = N->getOperand(0);
3298 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3299 if (N->getOperand(i) != SplatValue)
3304 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3305 /// to an zero vector.
3306 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3307 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3308 SDValue V1 = N->getOperand(0);
3309 SDValue V2 = N->getOperand(1);
3310 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3311 for (unsigned i = 0; i != NumElems; ++i) {
3312 int Idx = N->getMaskElt(i);
3313 if (Idx >= (int)NumElems) {
3314 unsigned Opc = V2.getOpcode();
3315 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3317 if (Opc != ISD::BUILD_VECTOR ||
3318 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3320 } else if (Idx >= 0) {
3321 unsigned Opc = V1.getOpcode();
3322 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3324 if (Opc != ISD::BUILD_VECTOR ||
3325 !X86::isZeroNode(V1.getOperand(Idx)))
3332 /// getZeroVector - Returns a vector of specified type with all zero elements.
3334 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3336 assert(VT.isVector() && "Expected a vector type");
3338 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3339 // type. This ensures they get CSE'd.
3341 if (VT.getSizeInBits() == 64) { // MMX
3342 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3343 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3344 } else if (HasSSE2) { // SSE2
3345 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3346 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3348 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3349 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3351 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3354 /// getOnesVector - Returns a vector of specified type with all bits set.
3356 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3357 assert(VT.isVector() && "Expected a vector type");
3359 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3360 // type. This ensures they get CSE'd.
3361 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3363 if (VT.getSizeInBits() == 64) // MMX
3364 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3366 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3367 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3371 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3372 /// that point to V2 points to its first element.
3373 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3374 EVT VT = SVOp->getValueType(0);
3375 unsigned NumElems = VT.getVectorNumElements();
3377 bool Changed = false;
3378 SmallVector<int, 8> MaskVec;
3379 SVOp->getMask(MaskVec);
3381 for (unsigned i = 0; i != NumElems; ++i) {
3382 if (MaskVec[i] > (int)NumElems) {
3383 MaskVec[i] = NumElems;
3388 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3389 SVOp->getOperand(1), &MaskVec[0]);
3390 return SDValue(SVOp, 0);
3393 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3394 /// operation of specified width.
3395 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3397 unsigned NumElems = VT.getVectorNumElements();
3398 SmallVector<int, 8> Mask;
3399 Mask.push_back(NumElems);
3400 for (unsigned i = 1; i != NumElems; ++i)
3402 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3405 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3406 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3408 unsigned NumElems = VT.getVectorNumElements();
3409 SmallVector<int, 8> Mask;
3410 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3412 Mask.push_back(i + NumElems);
3414 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3417 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3418 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3420 unsigned NumElems = VT.getVectorNumElements();
3421 unsigned Half = NumElems/2;
3422 SmallVector<int, 8> Mask;
3423 for (unsigned i = 0; i != Half; ++i) {
3424 Mask.push_back(i + Half);
3425 Mask.push_back(i + NumElems + Half);
3427 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3430 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3431 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3433 if (SV->getValueType(0).getVectorNumElements() <= 4)
3434 return SDValue(SV, 0);
3436 EVT PVT = MVT::v4f32;
3437 EVT VT = SV->getValueType(0);
3438 DebugLoc dl = SV->getDebugLoc();
3439 SDValue V1 = SV->getOperand(0);
3440 int NumElems = VT.getVectorNumElements();
3441 int EltNo = SV->getSplatIndex();
3443 // unpack elements to the correct location
3444 while (NumElems > 4) {
3445 if (EltNo < NumElems/2) {
3446 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3448 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3449 EltNo -= NumElems/2;
3454 // Perform the splat.
3455 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3456 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3457 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3458 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3461 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3462 /// vector of zero or undef vector. This produces a shuffle where the low
3463 /// element of V2 is swizzled into the zero/undef vector, landing at element
3464 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3465 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3466 bool isZero, bool HasSSE2,
3467 SelectionDAG &DAG) {
3468 EVT VT = V2.getValueType();
3470 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3471 unsigned NumElems = VT.getVectorNumElements();
3472 SmallVector<int, 16> MaskVec;
3473 for (unsigned i = 0; i != NumElems; ++i)
3474 // If this is the insertion idx, put the low elt of V2 here.
3475 MaskVec.push_back(i == Idx ? NumElems : i);
3476 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3479 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3480 /// a shuffle that is zero.
3482 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3483 bool Low, SelectionDAG &DAG) {
3484 unsigned NumZeros = 0;
3485 for (int i = 0; i < NumElems; ++i) {
3486 unsigned Index = Low ? i : NumElems-i-1;
3487 int Idx = SVOp->getMaskElt(Index);
3492 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3493 if (Elt.getNode() && X86::isZeroNode(Elt))
3501 /// isVectorShift - Returns true if the shuffle can be implemented as a
3502 /// logical left or right shift of a vector.
3503 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3504 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3505 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3506 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3509 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3512 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3516 bool SeenV1 = false;
3517 bool SeenV2 = false;
3518 for (unsigned i = NumZeros; i < NumElems; ++i) {
3519 unsigned Val = isLeft ? (i - NumZeros) : i;
3520 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3523 unsigned Idx = (unsigned) Idx_;
3533 if (SeenV1 && SeenV2)
3536 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3542 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3544 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3545 unsigned NumNonZero, unsigned NumZero,
3547 const TargetLowering &TLI) {
3551 DebugLoc dl = Op.getDebugLoc();
3554 for (unsigned i = 0; i < 16; ++i) {
3555 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3556 if (ThisIsNonZero && First) {
3558 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3560 V = DAG.getUNDEF(MVT::v8i16);
3565 SDValue ThisElt(0, 0), LastElt(0, 0);
3566 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3567 if (LastIsNonZero) {
3568 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3569 MVT::i16, Op.getOperand(i-1));
3571 if (ThisIsNonZero) {
3572 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3573 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3574 ThisElt, DAG.getConstant(8, MVT::i8));
3576 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3580 if (ThisElt.getNode())
3581 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3582 DAG.getIntPtrConstant(i/2));
3586 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3589 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3591 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3592 unsigned NumNonZero, unsigned NumZero,
3594 const TargetLowering &TLI) {
3598 DebugLoc dl = Op.getDebugLoc();
3601 for (unsigned i = 0; i < 8; ++i) {
3602 bool isNonZero = (NonZeros & (1 << i)) != 0;
3606 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3608 V = DAG.getUNDEF(MVT::v8i16);
3611 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3612 MVT::v8i16, V, Op.getOperand(i),
3613 DAG.getIntPtrConstant(i));
3620 /// getVShift - Return a vector logical shift node.
3622 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3623 unsigned NumBits, SelectionDAG &DAG,
3624 const TargetLowering &TLI, DebugLoc dl) {
3625 bool isMMX = VT.getSizeInBits() == 64;
3626 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3627 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3628 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3629 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3630 DAG.getNode(Opc, dl, ShVT, SrcOp,
3631 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3635 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3636 SelectionDAG &DAG) const {
3638 // Check if the scalar load can be widened into a vector load. And if
3639 // the address is "base + cst" see if the cst can be "absorbed" into
3640 // the shuffle mask.
3641 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3642 SDValue Ptr = LD->getBasePtr();
3643 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3645 EVT PVT = LD->getValueType(0);
3646 if (PVT != MVT::i32 && PVT != MVT::f32)
3651 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3652 FI = FINode->getIndex();
3654 } else if (Ptr.getOpcode() == ISD::ADD &&
3655 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3656 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3657 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3658 Offset = Ptr.getConstantOperandVal(1);
3659 Ptr = Ptr.getOperand(0);
3664 SDValue Chain = LD->getChain();
3665 // Make sure the stack object alignment is at least 16.
3666 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3667 if (DAG.InferPtrAlignment(Ptr) < 16) {
3668 if (MFI->isFixedObjectIndex(FI)) {
3669 // Can't change the alignment. FIXME: It's possible to compute
3670 // the exact stack offset and reference FI + adjust offset instead.
3671 // If someone *really* cares about this. That's the way to implement it.
3674 MFI->setObjectAlignment(FI, 16);
3678 // (Offset % 16) must be multiple of 4. Then address is then
3679 // Ptr + (Offset & ~15).
3682 if ((Offset % 16) & 3)
3684 int64_t StartOffset = Offset & ~15;
3686 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3687 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3689 int EltNo = (Offset - StartOffset) >> 2;
3690 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3691 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3692 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3694 // Canonicalize it to a v4i32 shuffle.
3695 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3696 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3697 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3698 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3704 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3705 /// vector of type 'VT', see if the elements can be replaced by a single large
3706 /// load which has the same value as a build_vector whose operands are 'elts'.
3708 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3710 /// FIXME: we'd also like to handle the case where the last elements are zero
3711 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3712 /// There's even a handy isZeroNode for that purpose.
3713 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3714 DebugLoc &dl, SelectionDAG &DAG) {
3715 EVT EltVT = VT.getVectorElementType();
3716 unsigned NumElems = Elts.size();
3718 LoadSDNode *LDBase = NULL;
3719 unsigned LastLoadedElt = -1U;
3721 // For each element in the initializer, see if we've found a load or an undef.
3722 // If we don't find an initial load element, or later load elements are
3723 // non-consecutive, bail out.
3724 for (unsigned i = 0; i < NumElems; ++i) {
3725 SDValue Elt = Elts[i];
3727 if (!Elt.getNode() ||
3728 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3731 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3733 LDBase = cast<LoadSDNode>(Elt.getNode());
3737 if (Elt.getOpcode() == ISD::UNDEF)
3740 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3741 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3746 // If we have found an entire vector of loads and undefs, then return a large
3747 // load of the entire vector width starting at the base pointer. If we found
3748 // consecutive loads for the low half, generate a vzext_load node.
3749 if (LastLoadedElt == NumElems - 1) {
3750 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3751 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3752 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3753 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3754 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3755 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3756 LDBase->isVolatile(), LDBase->isNonTemporal(),
3757 LDBase->getAlignment());
3758 } else if (NumElems == 4 && LastLoadedElt == 1) {
3759 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3760 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3761 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3762 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3768 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3769 DebugLoc dl = Op.getDebugLoc();
3770 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3771 if (ISD::isBuildVectorAllZeros(Op.getNode())
3772 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3773 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3774 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3775 // eliminated on x86-32 hosts.
3776 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3779 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3780 return getOnesVector(Op.getValueType(), DAG, dl);
3781 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3784 EVT VT = Op.getValueType();
3785 EVT ExtVT = VT.getVectorElementType();
3786 unsigned EVTBits = ExtVT.getSizeInBits();
3788 unsigned NumElems = Op.getNumOperands();
3789 unsigned NumZero = 0;
3790 unsigned NumNonZero = 0;
3791 unsigned NonZeros = 0;
3792 bool IsAllConstants = true;
3793 SmallSet<SDValue, 8> Values;
3794 for (unsigned i = 0; i < NumElems; ++i) {
3795 SDValue Elt = Op.getOperand(i);
3796 if (Elt.getOpcode() == ISD::UNDEF)
3799 if (Elt.getOpcode() != ISD::Constant &&
3800 Elt.getOpcode() != ISD::ConstantFP)
3801 IsAllConstants = false;
3802 if (X86::isZeroNode(Elt))
3805 NonZeros |= (1 << i);
3810 if (NumNonZero == 0) {
3811 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3812 return DAG.getUNDEF(VT);
3815 // Special case for single non-zero, non-undef, element.
3816 if (NumNonZero == 1) {
3817 unsigned Idx = CountTrailingZeros_32(NonZeros);
3818 SDValue Item = Op.getOperand(Idx);
3820 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3821 // the value are obviously zero, truncate the value to i32 and do the
3822 // insertion that way. Only do this if the value is non-constant or if the
3823 // value is a constant being inserted into element 0. It is cheaper to do
3824 // a constant pool load than it is to do a movd + shuffle.
3825 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3826 (!IsAllConstants || Idx == 0)) {
3827 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3828 // Handle MMX and SSE both.
3829 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3830 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3832 // Truncate the value (which may itself be a constant) to i32, and
3833 // convert it to a vector with movd (S2V+shuffle to zero extend).
3834 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3835 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3836 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3837 Subtarget->hasSSE2(), DAG);
3839 // Now we have our 32-bit value zero extended in the low element of
3840 // a vector. If Idx != 0, swizzle it into place.
3842 SmallVector<int, 4> Mask;
3843 Mask.push_back(Idx);
3844 for (unsigned i = 1; i != VecElts; ++i)
3846 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3847 DAG.getUNDEF(Item.getValueType()),
3850 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3854 // If we have a constant or non-constant insertion into the low element of
3855 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3856 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3857 // depending on what the source datatype is.
3860 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3861 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3862 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3863 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3864 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3865 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3867 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3868 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3869 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3870 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3871 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3872 Subtarget->hasSSE2(), DAG);
3873 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3877 // Is it a vector logical left shift?
3878 if (NumElems == 2 && Idx == 1 &&
3879 X86::isZeroNode(Op.getOperand(0)) &&
3880 !X86::isZeroNode(Op.getOperand(1))) {
3881 unsigned NumBits = VT.getSizeInBits();
3882 return getVShift(true, VT,
3883 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3884 VT, Op.getOperand(1)),
3885 NumBits/2, DAG, *this, dl);
3888 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3891 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3892 // is a non-constant being inserted into an element other than the low one,
3893 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3894 // movd/movss) to move this into the low element, then shuffle it into
3896 if (EVTBits == 32) {
3897 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3899 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3900 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3901 Subtarget->hasSSE2(), DAG);
3902 SmallVector<int, 8> MaskVec;
3903 for (unsigned i = 0; i < NumElems; i++)
3904 MaskVec.push_back(i == Idx ? 0 : 1);
3905 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3909 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3910 if (Values.size() == 1) {
3911 if (EVTBits == 32) {
3912 // Instead of a shuffle like this:
3913 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3914 // Check if it's possible to issue this instead.
3915 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3916 unsigned Idx = CountTrailingZeros_32(NonZeros);
3917 SDValue Item = Op.getOperand(Idx);
3918 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3919 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3924 // A vector full of immediates; various special cases are already
3925 // handled, so this is best done with a single constant-pool load.
3929 // Let legalizer expand 2-wide build_vectors.
3930 if (EVTBits == 64) {
3931 if (NumNonZero == 1) {
3932 // One half is zero or undef.
3933 unsigned Idx = CountTrailingZeros_32(NonZeros);
3934 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3935 Op.getOperand(Idx));
3936 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3937 Subtarget->hasSSE2(), DAG);
3942 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3943 if (EVTBits == 8 && NumElems == 16) {
3944 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3946 if (V.getNode()) return V;
3949 if (EVTBits == 16 && NumElems == 8) {
3950 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3952 if (V.getNode()) return V;
3955 // If element VT is == 32 bits, turn it into a number of shuffles.
3956 SmallVector<SDValue, 8> V;
3958 if (NumElems == 4 && NumZero > 0) {
3959 for (unsigned i = 0; i < 4; ++i) {
3960 bool isZero = !(NonZeros & (1 << i));
3962 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3964 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3967 for (unsigned i = 0; i < 2; ++i) {
3968 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3971 V[i] = V[i*2]; // Must be a zero vector.
3974 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3977 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3980 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3985 SmallVector<int, 8> MaskVec;
3986 bool Reverse = (NonZeros & 0x3) == 2;
3987 for (unsigned i = 0; i < 2; ++i)
3988 MaskVec.push_back(Reverse ? 1-i : i);
3989 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3990 for (unsigned i = 0; i < 2; ++i)
3991 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3992 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3995 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3996 // Check for a build vector of consecutive loads.
3997 for (unsigned i = 0; i < NumElems; ++i)
3998 V[i] = Op.getOperand(i);
4000 // Check for elements which are consecutive loads.
4001 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4005 // For SSE 4.1, use inserts into undef.
4006 if (getSubtarget()->hasSSE41()) {
4007 V[0] = DAG.getUNDEF(VT);
4008 for (unsigned i = 0; i < NumElems; ++i)
4009 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4010 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4011 Op.getOperand(i), DAG.getIntPtrConstant(i));
4015 // Otherwise, expand into a number of unpckl*
4017 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4018 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4019 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4020 for (unsigned i = 0; i < NumElems; ++i)
4021 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4023 while (NumElems != 0) {
4024 for (unsigned i = 0; i < NumElems; ++i)
4025 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4034 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4035 // We support concatenate two MMX registers and place them in a MMX
4036 // register. This is better than doing a stack convert.
4037 DebugLoc dl = Op.getDebugLoc();
4038 EVT ResVT = Op.getValueType();
4039 assert(Op.getNumOperands() == 2);
4040 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4041 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4043 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4044 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4045 InVec = Op.getOperand(1);
4046 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4047 unsigned NumElts = ResVT.getVectorNumElements();
4048 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4049 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4050 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4052 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4053 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4054 Mask[0] = 0; Mask[1] = 2;
4055 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4057 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4060 // v8i16 shuffles - Prefer shuffles in the following order:
4061 // 1. [all] pshuflw, pshufhw, optional move
4062 // 2. [ssse3] 1 x pshufb
4063 // 3. [ssse3] 2 x pshufb + 1 x por
4064 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4066 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4068 const X86TargetLowering &TLI) {
4069 SDValue V1 = SVOp->getOperand(0);
4070 SDValue V2 = SVOp->getOperand(1);
4071 DebugLoc dl = SVOp->getDebugLoc();
4072 SmallVector<int, 8> MaskVals;
4074 // Determine if more than 1 of the words in each of the low and high quadwords
4075 // of the result come from the same quadword of one of the two inputs. Undef
4076 // mask values count as coming from any quadword, for better codegen.
4077 SmallVector<unsigned, 4> LoQuad(4);
4078 SmallVector<unsigned, 4> HiQuad(4);
4079 BitVector InputQuads(4);
4080 for (unsigned i = 0; i < 8; ++i) {
4081 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4082 int EltIdx = SVOp->getMaskElt(i);
4083 MaskVals.push_back(EltIdx);
4092 InputQuads.set(EltIdx / 4);
4095 int BestLoQuad = -1;
4096 unsigned MaxQuad = 1;
4097 for (unsigned i = 0; i < 4; ++i) {
4098 if (LoQuad[i] > MaxQuad) {
4100 MaxQuad = LoQuad[i];
4104 int BestHiQuad = -1;
4106 for (unsigned i = 0; i < 4; ++i) {
4107 if (HiQuad[i] > MaxQuad) {
4109 MaxQuad = HiQuad[i];
4113 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4114 // of the two input vectors, shuffle them into one input vector so only a
4115 // single pshufb instruction is necessary. If There are more than 2 input
4116 // quads, disable the next transformation since it does not help SSSE3.
4117 bool V1Used = InputQuads[0] || InputQuads[1];
4118 bool V2Used = InputQuads[2] || InputQuads[3];
4119 if (TLI.getSubtarget()->hasSSSE3()) {
4120 if (InputQuads.count() == 2 && V1Used && V2Used) {
4121 BestLoQuad = InputQuads.find_first();
4122 BestHiQuad = InputQuads.find_next(BestLoQuad);
4124 if (InputQuads.count() > 2) {
4130 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4131 // the shuffle mask. If a quad is scored as -1, that means that it contains
4132 // words from all 4 input quadwords.
4134 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4135 SmallVector<int, 8> MaskV;
4136 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4137 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4138 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4139 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4140 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4141 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4143 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4144 // source words for the shuffle, to aid later transformations.
4145 bool AllWordsInNewV = true;
4146 bool InOrder[2] = { true, true };
4147 for (unsigned i = 0; i != 8; ++i) {
4148 int idx = MaskVals[i];
4150 InOrder[i/4] = false;
4151 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4153 AllWordsInNewV = false;
4157 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4158 if (AllWordsInNewV) {
4159 for (int i = 0; i != 8; ++i) {
4160 int idx = MaskVals[i];
4163 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4164 if ((idx != i) && idx < 4)
4166 if ((idx != i) && idx > 3)
4175 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4176 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4177 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4178 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4179 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4183 // If we have SSSE3, and all words of the result are from 1 input vector,
4184 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4185 // is present, fall back to case 4.
4186 if (TLI.getSubtarget()->hasSSSE3()) {
4187 SmallVector<SDValue,16> pshufbMask;
4189 // If we have elements from both input vectors, set the high bit of the
4190 // shuffle mask element to zero out elements that come from V2 in the V1
4191 // mask, and elements that come from V1 in the V2 mask, so that the two
4192 // results can be OR'd together.
4193 bool TwoInputs = V1Used && V2Used;
4194 for (unsigned i = 0; i != 8; ++i) {
4195 int EltIdx = MaskVals[i] * 2;
4196 if (TwoInputs && (EltIdx >= 16)) {
4197 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4198 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4201 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4202 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4204 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4205 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4206 DAG.getNode(ISD::BUILD_VECTOR, dl,
4207 MVT::v16i8, &pshufbMask[0], 16));
4209 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4211 // Calculate the shuffle mask for the second input, shuffle it, and
4212 // OR it with the first shuffled input.
4214 for (unsigned i = 0; i != 8; ++i) {
4215 int EltIdx = MaskVals[i] * 2;
4217 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4218 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4221 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4222 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4224 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4225 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4226 DAG.getNode(ISD::BUILD_VECTOR, dl,
4227 MVT::v16i8, &pshufbMask[0], 16));
4228 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4229 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4232 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4233 // and update MaskVals with new element order.
4234 BitVector InOrder(8);
4235 if (BestLoQuad >= 0) {
4236 SmallVector<int, 8> MaskV;
4237 for (int i = 0; i != 4; ++i) {
4238 int idx = MaskVals[i];
4240 MaskV.push_back(-1);
4242 } else if ((idx / 4) == BestLoQuad) {
4243 MaskV.push_back(idx & 3);
4246 MaskV.push_back(-1);
4249 for (unsigned i = 4; i != 8; ++i)
4251 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4255 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4256 // and update MaskVals with the new element order.
4257 if (BestHiQuad >= 0) {
4258 SmallVector<int, 8> MaskV;
4259 for (unsigned i = 0; i != 4; ++i)
4261 for (unsigned i = 4; i != 8; ++i) {
4262 int idx = MaskVals[i];
4264 MaskV.push_back(-1);
4266 } else if ((idx / 4) == BestHiQuad) {
4267 MaskV.push_back((idx & 3) + 4);
4270 MaskV.push_back(-1);
4273 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4277 // In case BestHi & BestLo were both -1, which means each quadword has a word
4278 // from each of the four input quadwords, calculate the InOrder bitvector now
4279 // before falling through to the insert/extract cleanup.
4280 if (BestLoQuad == -1 && BestHiQuad == -1) {
4282 for (int i = 0; i != 8; ++i)
4283 if (MaskVals[i] < 0 || MaskVals[i] == i)
4287 // The other elements are put in the right place using pextrw and pinsrw.
4288 for (unsigned i = 0; i != 8; ++i) {
4291 int EltIdx = MaskVals[i];
4294 SDValue ExtOp = (EltIdx < 8)
4295 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4296 DAG.getIntPtrConstant(EltIdx))
4297 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4298 DAG.getIntPtrConstant(EltIdx - 8));
4299 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4300 DAG.getIntPtrConstant(i));
4305 // v16i8 shuffles - Prefer shuffles in the following order:
4306 // 1. [ssse3] 1 x pshufb
4307 // 2. [ssse3] 2 x pshufb + 1 x por
4308 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4310 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4312 const X86TargetLowering &TLI) {
4313 SDValue V1 = SVOp->getOperand(0);
4314 SDValue V2 = SVOp->getOperand(1);
4315 DebugLoc dl = SVOp->getDebugLoc();
4316 SmallVector<int, 16> MaskVals;
4317 SVOp->getMask(MaskVals);
4319 // If we have SSSE3, case 1 is generated when all result bytes come from
4320 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4321 // present, fall back to case 3.
4322 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4325 for (unsigned i = 0; i < 16; ++i) {
4326 int EltIdx = MaskVals[i];
4335 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4336 if (TLI.getSubtarget()->hasSSSE3()) {
4337 SmallVector<SDValue,16> pshufbMask;
4339 // If all result elements are from one input vector, then only translate
4340 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4342 // Otherwise, we have elements from both input vectors, and must zero out
4343 // elements that come from V2 in the first mask, and V1 in the second mask
4344 // so that we can OR them together.
4345 bool TwoInputs = !(V1Only || V2Only);
4346 for (unsigned i = 0; i != 16; ++i) {
4347 int EltIdx = MaskVals[i];
4348 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4349 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4352 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4354 // If all the elements are from V2, assign it to V1 and return after
4355 // building the first pshufb.
4358 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4359 DAG.getNode(ISD::BUILD_VECTOR, dl,
4360 MVT::v16i8, &pshufbMask[0], 16));
4364 // Calculate the shuffle mask for the second input, shuffle it, and
4365 // OR it with the first shuffled input.
4367 for (unsigned i = 0; i != 16; ++i) {
4368 int EltIdx = MaskVals[i];
4370 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4373 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4375 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4376 DAG.getNode(ISD::BUILD_VECTOR, dl,
4377 MVT::v16i8, &pshufbMask[0], 16));
4378 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4381 // No SSSE3 - Calculate in place words and then fix all out of place words
4382 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4383 // the 16 different words that comprise the two doublequadword input vectors.
4384 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4385 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4386 SDValue NewV = V2Only ? V2 : V1;
4387 for (int i = 0; i != 8; ++i) {
4388 int Elt0 = MaskVals[i*2];
4389 int Elt1 = MaskVals[i*2+1];
4391 // This word of the result is all undef, skip it.
4392 if (Elt0 < 0 && Elt1 < 0)
4395 // This word of the result is already in the correct place, skip it.
4396 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4398 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4401 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4402 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4405 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4406 // using a single extract together, load it and store it.
4407 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4408 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4409 DAG.getIntPtrConstant(Elt1 / 2));
4410 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4411 DAG.getIntPtrConstant(i));
4415 // If Elt1 is defined, extract it from the appropriate source. If the
4416 // source byte is not also odd, shift the extracted word left 8 bits
4417 // otherwise clear the bottom 8 bits if we need to do an or.
4419 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4420 DAG.getIntPtrConstant(Elt1 / 2));
4421 if ((Elt1 & 1) == 0)
4422 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4423 DAG.getConstant(8, TLI.getShiftAmountTy()));
4425 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4426 DAG.getConstant(0xFF00, MVT::i16));
4428 // If Elt0 is defined, extract it from the appropriate source. If the
4429 // source byte is not also even, shift the extracted word right 8 bits. If
4430 // Elt1 was also defined, OR the extracted values together before
4431 // inserting them in the result.
4433 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4434 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4435 if ((Elt0 & 1) != 0)
4436 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4437 DAG.getConstant(8, TLI.getShiftAmountTy()));
4439 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4440 DAG.getConstant(0x00FF, MVT::i16));
4441 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4444 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4445 DAG.getIntPtrConstant(i));
4447 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4450 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4451 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4452 /// done when every pair / quad of shuffle mask elements point to elements in
4453 /// the right sequence. e.g.
4454 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4456 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4458 const TargetLowering &TLI, DebugLoc dl) {
4459 EVT VT = SVOp->getValueType(0);
4460 SDValue V1 = SVOp->getOperand(0);
4461 SDValue V2 = SVOp->getOperand(1);
4462 unsigned NumElems = VT.getVectorNumElements();
4463 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4464 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4465 EVT MaskEltVT = MaskVT.getVectorElementType();
4467 switch (VT.getSimpleVT().SimpleTy) {
4468 default: assert(false && "Unexpected!");
4469 case MVT::v4f32: NewVT = MVT::v2f64; break;
4470 case MVT::v4i32: NewVT = MVT::v2i64; break;
4471 case MVT::v8i16: NewVT = MVT::v4i32; break;
4472 case MVT::v16i8: NewVT = MVT::v4i32; break;
4475 if (NewWidth == 2) {
4481 int Scale = NumElems / NewWidth;
4482 SmallVector<int, 8> MaskVec;
4483 for (unsigned i = 0; i < NumElems; i += Scale) {
4485 for (int j = 0; j < Scale; ++j) {
4486 int EltIdx = SVOp->getMaskElt(i+j);
4490 StartIdx = EltIdx - (EltIdx % Scale);
4491 if (EltIdx != StartIdx + j)
4495 MaskVec.push_back(-1);
4497 MaskVec.push_back(StartIdx / Scale);
4500 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4501 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4502 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4505 /// getVZextMovL - Return a zero-extending vector move low node.
4507 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4508 SDValue SrcOp, SelectionDAG &DAG,
4509 const X86Subtarget *Subtarget, DebugLoc dl) {
4510 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4511 LoadSDNode *LD = NULL;
4512 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4513 LD = dyn_cast<LoadSDNode>(SrcOp);
4515 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4517 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4518 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4519 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4520 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4521 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4523 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4524 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4525 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4526 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4534 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4535 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4536 DAG.getNode(ISD::BIT_CONVERT, dl,
4540 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4543 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4544 SDValue V1 = SVOp->getOperand(0);
4545 SDValue V2 = SVOp->getOperand(1);
4546 DebugLoc dl = SVOp->getDebugLoc();
4547 EVT VT = SVOp->getValueType(0);
4549 SmallVector<std::pair<int, int>, 8> Locs;
4551 SmallVector<int, 8> Mask1(4U, -1);
4552 SmallVector<int, 8> PermMask;
4553 SVOp->getMask(PermMask);
4557 for (unsigned i = 0; i != 4; ++i) {
4558 int Idx = PermMask[i];
4560 Locs[i] = std::make_pair(-1, -1);
4562 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4564 Locs[i] = std::make_pair(0, NumLo);
4568 Locs[i] = std::make_pair(1, NumHi);
4570 Mask1[2+NumHi] = Idx;
4576 if (NumLo <= 2 && NumHi <= 2) {
4577 // If no more than two elements come from either vector. This can be
4578 // implemented with two shuffles. First shuffle gather the elements.
4579 // The second shuffle, which takes the first shuffle as both of its
4580 // vector operands, put the elements into the right order.
4581 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4583 SmallVector<int, 8> Mask2(4U, -1);
4585 for (unsigned i = 0; i != 4; ++i) {
4586 if (Locs[i].first == -1)
4589 unsigned Idx = (i < 2) ? 0 : 4;
4590 Idx += Locs[i].first * 2 + Locs[i].second;
4595 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4596 } else if (NumLo == 3 || NumHi == 3) {
4597 // Otherwise, we must have three elements from one vector, call it X, and
4598 // one element from the other, call it Y. First, use a shufps to build an
4599 // intermediate vector with the one element from Y and the element from X
4600 // that will be in the same half in the final destination (the indexes don't
4601 // matter). Then, use a shufps to build the final vector, taking the half
4602 // containing the element from Y from the intermediate, and the other half
4605 // Normalize it so the 3 elements come from V1.
4606 CommuteVectorShuffleMask(PermMask, VT);
4610 // Find the element from V2.
4612 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4613 int Val = PermMask[HiIndex];
4620 Mask1[0] = PermMask[HiIndex];
4622 Mask1[2] = PermMask[HiIndex^1];
4624 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4627 Mask1[0] = PermMask[0];
4628 Mask1[1] = PermMask[1];
4629 Mask1[2] = HiIndex & 1 ? 6 : 4;
4630 Mask1[3] = HiIndex & 1 ? 4 : 6;
4631 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4633 Mask1[0] = HiIndex & 1 ? 2 : 0;
4634 Mask1[1] = HiIndex & 1 ? 0 : 2;
4635 Mask1[2] = PermMask[2];
4636 Mask1[3] = PermMask[3];
4641 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4645 // Break it into (shuffle shuffle_hi, shuffle_lo).
4647 SmallVector<int,8> LoMask(4U, -1);
4648 SmallVector<int,8> HiMask(4U, -1);
4650 SmallVector<int,8> *MaskPtr = &LoMask;
4651 unsigned MaskIdx = 0;
4654 for (unsigned i = 0; i != 4; ++i) {
4661 int Idx = PermMask[i];
4663 Locs[i] = std::make_pair(-1, -1);
4664 } else if (Idx < 4) {
4665 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4666 (*MaskPtr)[LoIdx] = Idx;
4669 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4670 (*MaskPtr)[HiIdx] = Idx;
4675 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4676 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4677 SmallVector<int, 8> MaskOps;
4678 for (unsigned i = 0; i != 4; ++i) {
4679 if (Locs[i].first == -1) {
4680 MaskOps.push_back(-1);
4682 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4683 MaskOps.push_back(Idx);
4686 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4690 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4691 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4692 SDValue V1 = Op.getOperand(0);
4693 SDValue V2 = Op.getOperand(1);
4694 EVT VT = Op.getValueType();
4695 DebugLoc dl = Op.getDebugLoc();
4696 unsigned NumElems = VT.getVectorNumElements();
4697 bool isMMX = VT.getSizeInBits() == 64;
4698 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4699 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4700 bool V1IsSplat = false;
4701 bool V2IsSplat = false;
4703 if (isZeroShuffle(SVOp))
4704 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4706 // Promote splats to v4f32.
4707 if (SVOp->isSplat()) {
4708 if (isMMX || NumElems < 4)
4710 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4713 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4715 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4716 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4717 if (NewOp.getNode())
4718 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4719 LowerVECTOR_SHUFFLE(NewOp, DAG));
4720 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4721 // FIXME: Figure out a cleaner way to do this.
4722 // Try to make use of movq to zero out the top part.
4723 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4724 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4725 if (NewOp.getNode()) {
4726 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4727 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4728 DAG, Subtarget, dl);
4730 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4731 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4732 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4733 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4734 DAG, Subtarget, dl);
4738 if (X86::isPSHUFDMask(SVOp))
4741 // Check if this can be converted into a logical shift.
4742 bool isLeft = false;
4745 bool isShift = getSubtarget()->hasSSE2() &&
4746 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4747 if (isShift && ShVal.hasOneUse()) {
4748 // If the shifted value has multiple uses, it may be cheaper to use
4749 // v_set0 + movlhps or movhlps, etc.
4750 EVT EltVT = VT.getVectorElementType();
4751 ShAmt *= EltVT.getSizeInBits();
4752 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4755 if (X86::isMOVLMask(SVOp)) {
4758 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4759 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4764 // FIXME: fold these into legal mask.
4765 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4766 X86::isMOVSLDUPMask(SVOp) ||
4767 X86::isMOVHLPSMask(SVOp) ||
4768 X86::isMOVLHPSMask(SVOp) ||
4769 X86::isMOVLPMask(SVOp)))
4772 if (ShouldXformToMOVHLPS(SVOp) ||
4773 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4774 return CommuteVectorShuffle(SVOp, DAG);
4777 // No better options. Use a vshl / vsrl.
4778 EVT EltVT = VT.getVectorElementType();
4779 ShAmt *= EltVT.getSizeInBits();
4780 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4783 bool Commuted = false;
4784 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4785 // 1,1,1,1 -> v8i16 though.
4786 V1IsSplat = isSplatVector(V1.getNode());
4787 V2IsSplat = isSplatVector(V2.getNode());
4789 // Canonicalize the splat or undef, if present, to be on the RHS.
4790 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4791 Op = CommuteVectorShuffle(SVOp, DAG);
4792 SVOp = cast<ShuffleVectorSDNode>(Op);
4793 V1 = SVOp->getOperand(0);
4794 V2 = SVOp->getOperand(1);
4795 std::swap(V1IsSplat, V2IsSplat);
4796 std::swap(V1IsUndef, V2IsUndef);
4800 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4801 // Shuffling low element of v1 into undef, just return v1.
4804 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4805 // the instruction selector will not match, so get a canonical MOVL with
4806 // swapped operands to undo the commute.
4807 return getMOVL(DAG, dl, VT, V2, V1);
4810 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4811 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4812 X86::isUNPCKLMask(SVOp) ||
4813 X86::isUNPCKHMask(SVOp))
4817 // Normalize mask so all entries that point to V2 points to its first
4818 // element then try to match unpck{h|l} again. If match, return a
4819 // new vector_shuffle with the corrected mask.
4820 SDValue NewMask = NormalizeMask(SVOp, DAG);
4821 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4822 if (NSVOp != SVOp) {
4823 if (X86::isUNPCKLMask(NSVOp, true)) {
4825 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4832 // Commute is back and try unpck* again.
4833 // FIXME: this seems wrong.
4834 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4835 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4836 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4837 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4838 X86::isUNPCKLMask(NewSVOp) ||
4839 X86::isUNPCKHMask(NewSVOp))
4843 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4845 // Normalize the node to match x86 shuffle ops if needed
4846 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4847 return CommuteVectorShuffle(SVOp, DAG);
4849 // Check for legal shuffle and return?
4850 SmallVector<int, 16> PermMask;
4851 SVOp->getMask(PermMask);
4852 if (isShuffleMaskLegal(PermMask, VT))
4855 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4856 if (VT == MVT::v8i16) {
4857 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4858 if (NewOp.getNode())
4862 if (VT == MVT::v16i8) {
4863 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4864 if (NewOp.getNode())
4868 // Handle all 4 wide cases with a number of shuffles except for MMX.
4869 if (NumElems == 4 && !isMMX)
4870 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4876 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4877 SelectionDAG &DAG) const {
4878 EVT VT = Op.getValueType();
4879 DebugLoc dl = Op.getDebugLoc();
4880 if (VT.getSizeInBits() == 8) {
4881 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4882 Op.getOperand(0), Op.getOperand(1));
4883 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4884 DAG.getValueType(VT));
4885 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4886 } else if (VT.getSizeInBits() == 16) {
4887 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4888 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4890 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4891 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4892 DAG.getNode(ISD::BIT_CONVERT, dl,
4896 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4897 Op.getOperand(0), Op.getOperand(1));
4898 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4899 DAG.getValueType(VT));
4900 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4901 } else if (VT == MVT::f32) {
4902 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4903 // the result back to FR32 register. It's only worth matching if the
4904 // result has a single use which is a store or a bitcast to i32. And in
4905 // the case of a store, it's not worth it if the index is a constant 0,
4906 // because a MOVSSmr can be used instead, which is smaller and faster.
4907 if (!Op.hasOneUse())
4909 SDNode *User = *Op.getNode()->use_begin();
4910 if ((User->getOpcode() != ISD::STORE ||
4911 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4912 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4913 (User->getOpcode() != ISD::BIT_CONVERT ||
4914 User->getValueType(0) != MVT::i32))
4916 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4917 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4920 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4921 } else if (VT == MVT::i32) {
4922 // ExtractPS works with constant index.
4923 if (isa<ConstantSDNode>(Op.getOperand(1)))
4931 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4932 SelectionDAG &DAG) const {
4933 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4936 if (Subtarget->hasSSE41()) {
4937 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4942 EVT VT = Op.getValueType();
4943 DebugLoc dl = Op.getDebugLoc();
4944 // TODO: handle v16i8.
4945 if (VT.getSizeInBits() == 16) {
4946 SDValue Vec = Op.getOperand(0);
4947 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4949 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4950 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4951 DAG.getNode(ISD::BIT_CONVERT, dl,
4954 // Transform it so it match pextrw which produces a 32-bit result.
4955 EVT EltVT = MVT::i32;
4956 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4957 Op.getOperand(0), Op.getOperand(1));
4958 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4959 DAG.getValueType(VT));
4960 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4961 } else if (VT.getSizeInBits() == 32) {
4962 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4966 // SHUFPS the element to the lowest double word, then movss.
4967 int Mask[4] = { Idx, -1, -1, -1 };
4968 EVT VVT = Op.getOperand(0).getValueType();
4969 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4970 DAG.getUNDEF(VVT), Mask);
4971 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4972 DAG.getIntPtrConstant(0));
4973 } else if (VT.getSizeInBits() == 64) {
4974 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4975 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4976 // to match extract_elt for f64.
4977 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4981 // UNPCKHPD the element to the lowest double word, then movsd.
4982 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4983 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4984 int Mask[2] = { 1, -1 };
4985 EVT VVT = Op.getOperand(0).getValueType();
4986 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4987 DAG.getUNDEF(VVT), Mask);
4988 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4989 DAG.getIntPtrConstant(0));
4996 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4997 SelectionDAG &DAG) const {
4998 EVT VT = Op.getValueType();
4999 EVT EltVT = VT.getVectorElementType();
5000 DebugLoc dl = Op.getDebugLoc();
5002 SDValue N0 = Op.getOperand(0);
5003 SDValue N1 = Op.getOperand(1);
5004 SDValue N2 = Op.getOperand(2);
5006 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5007 isa<ConstantSDNode>(N2)) {
5009 if (VT == MVT::v8i16)
5010 Opc = X86ISD::PINSRW;
5011 else if (VT == MVT::v4i16)
5012 Opc = X86ISD::MMX_PINSRW;
5013 else if (VT == MVT::v16i8)
5014 Opc = X86ISD::PINSRB;
5016 Opc = X86ISD::PINSRB;
5018 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5020 if (N1.getValueType() != MVT::i32)
5021 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5022 if (N2.getValueType() != MVT::i32)
5023 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5024 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5025 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5026 // Bits [7:6] of the constant are the source select. This will always be
5027 // zero here. The DAG Combiner may combine an extract_elt index into these
5028 // bits. For example (insert (extract, 3), 2) could be matched by putting
5029 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5030 // Bits [5:4] of the constant are the destination select. This is the
5031 // value of the incoming immediate.
5032 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5033 // combine either bitwise AND or insert of float 0.0 to set these bits.
5034 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5035 // Create this as a scalar to vector..
5036 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5037 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5038 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5039 // PINSR* works with constant index.
5046 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5047 EVT VT = Op.getValueType();
5048 EVT EltVT = VT.getVectorElementType();
5050 if (Subtarget->hasSSE41())
5051 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5053 if (EltVT == MVT::i8)
5056 DebugLoc dl = Op.getDebugLoc();
5057 SDValue N0 = Op.getOperand(0);
5058 SDValue N1 = Op.getOperand(1);
5059 SDValue N2 = Op.getOperand(2);
5061 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5062 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5063 // as its second argument.
5064 if (N1.getValueType() != MVT::i32)
5065 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5066 if (N2.getValueType() != MVT::i32)
5067 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5068 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5069 dl, VT, N0, N1, N2);
5075 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5076 DebugLoc dl = Op.getDebugLoc();
5077 if (Op.getValueType() == MVT::v2f32)
5078 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5079 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5080 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5081 Op.getOperand(0))));
5083 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5084 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5086 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5087 EVT VT = MVT::v2i32;
5088 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5095 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5096 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5099 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5100 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5101 // one of the above mentioned nodes. It has to be wrapped because otherwise
5102 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5103 // be used to form addressing mode. These wrapped nodes will be selected
5106 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5107 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5109 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5111 unsigned char OpFlag = 0;
5112 unsigned WrapperKind = X86ISD::Wrapper;
5113 CodeModel::Model M = getTargetMachine().getCodeModel();
5115 if (Subtarget->isPICStyleRIPRel() &&
5116 (M == CodeModel::Small || M == CodeModel::Kernel))
5117 WrapperKind = X86ISD::WrapperRIP;
5118 else if (Subtarget->isPICStyleGOT())
5119 OpFlag = X86II::MO_GOTOFF;
5120 else if (Subtarget->isPICStyleStubPIC())
5121 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5123 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5125 CP->getOffset(), OpFlag);
5126 DebugLoc DL = CP->getDebugLoc();
5127 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5128 // With PIC, the address is actually $g + Offset.
5130 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5131 DAG.getNode(X86ISD::GlobalBaseReg,
5132 DebugLoc(), getPointerTy()),
5139 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5140 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5142 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5144 unsigned char OpFlag = 0;
5145 unsigned WrapperKind = X86ISD::Wrapper;
5146 CodeModel::Model M = getTargetMachine().getCodeModel();
5148 if (Subtarget->isPICStyleRIPRel() &&
5149 (M == CodeModel::Small || M == CodeModel::Kernel))
5150 WrapperKind = X86ISD::WrapperRIP;
5151 else if (Subtarget->isPICStyleGOT())
5152 OpFlag = X86II::MO_GOTOFF;
5153 else if (Subtarget->isPICStyleStubPIC())
5154 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5156 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5158 DebugLoc DL = JT->getDebugLoc();
5159 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5161 // With PIC, the address is actually $g + Offset.
5163 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5164 DAG.getNode(X86ISD::GlobalBaseReg,
5165 DebugLoc(), getPointerTy()),
5173 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5174 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5176 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5178 unsigned char OpFlag = 0;
5179 unsigned WrapperKind = X86ISD::Wrapper;
5180 CodeModel::Model M = getTargetMachine().getCodeModel();
5182 if (Subtarget->isPICStyleRIPRel() &&
5183 (M == CodeModel::Small || M == CodeModel::Kernel))
5184 WrapperKind = X86ISD::WrapperRIP;
5185 else if (Subtarget->isPICStyleGOT())
5186 OpFlag = X86II::MO_GOTOFF;
5187 else if (Subtarget->isPICStyleStubPIC())
5188 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5190 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5192 DebugLoc DL = Op.getDebugLoc();
5193 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5196 // With PIC, the address is actually $g + Offset.
5197 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5198 !Subtarget->is64Bit()) {
5199 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5200 DAG.getNode(X86ISD::GlobalBaseReg,
5201 DebugLoc(), getPointerTy()),
5209 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5210 // Create the TargetBlockAddressAddress node.
5211 unsigned char OpFlags =
5212 Subtarget->ClassifyBlockAddressReference();
5213 CodeModel::Model M = getTargetMachine().getCodeModel();
5214 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5215 DebugLoc dl = Op.getDebugLoc();
5216 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5217 /*isTarget=*/true, OpFlags);
5219 if (Subtarget->isPICStyleRIPRel() &&
5220 (M == CodeModel::Small || M == CodeModel::Kernel))
5221 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5223 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5225 // With PIC, the address is actually $g + Offset.
5226 if (isGlobalRelativeToPICBase(OpFlags)) {
5227 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5228 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5236 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5238 SelectionDAG &DAG) const {
5239 // Create the TargetGlobalAddress node, folding in the constant
5240 // offset if it is legal.
5241 unsigned char OpFlags =
5242 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5243 CodeModel::Model M = getTargetMachine().getCodeModel();
5245 if (OpFlags == X86II::MO_NO_FLAG &&
5246 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5247 // A direct static reference to a global.
5248 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5251 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5254 if (Subtarget->isPICStyleRIPRel() &&
5255 (M == CodeModel::Small || M == CodeModel::Kernel))
5256 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5258 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5260 // With PIC, the address is actually $g + Offset.
5261 if (isGlobalRelativeToPICBase(OpFlags)) {
5262 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5263 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5267 // For globals that require a load from a stub to get the address, emit the
5269 if (isGlobalStubReference(OpFlags))
5270 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5271 PseudoSourceValue::getGOT(), 0, false, false, 0);
5273 // If there was a non-zero offset that we didn't fold, create an explicit
5276 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5277 DAG.getConstant(Offset, getPointerTy()));
5283 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5284 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5285 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5286 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5290 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5291 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5292 unsigned char OperandFlags) {
5293 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5294 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5295 DebugLoc dl = GA->getDebugLoc();
5296 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5297 GA->getValueType(0),
5301 SDValue Ops[] = { Chain, TGA, *InFlag };
5302 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5304 SDValue Ops[] = { Chain, TGA };
5305 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5308 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5309 MFI->setAdjustsStack(true);
5311 SDValue Flag = Chain.getValue(1);
5312 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5315 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5317 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5320 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5321 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5322 DAG.getNode(X86ISD::GlobalBaseReg,
5323 DebugLoc(), PtrVT), InFlag);
5324 InFlag = Chain.getValue(1);
5326 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5329 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5331 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5333 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5334 X86::RAX, X86II::MO_TLSGD);
5337 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5338 // "local exec" model.
5339 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5340 const EVT PtrVT, TLSModel::Model model,
5342 DebugLoc dl = GA->getDebugLoc();
5343 // Get the Thread Pointer
5344 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5346 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5349 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5350 NULL, 0, false, false, 0);
5352 unsigned char OperandFlags = 0;
5353 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5355 unsigned WrapperKind = X86ISD::Wrapper;
5356 if (model == TLSModel::LocalExec) {
5357 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5358 } else if (is64Bit) {
5359 assert(model == TLSModel::InitialExec);
5360 OperandFlags = X86II::MO_GOTTPOFF;
5361 WrapperKind = X86ISD::WrapperRIP;
5363 assert(model == TLSModel::InitialExec);
5364 OperandFlags = X86II::MO_INDNTPOFF;
5367 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5369 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5370 GA->getOffset(), OperandFlags);
5371 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5373 if (model == TLSModel::InitialExec)
5374 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5375 PseudoSourceValue::getGOT(), 0, false, false, 0);
5377 // The address of the thread local variable is the add of the thread
5378 // pointer with the offset of the variable.
5379 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5383 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5385 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5386 const GlobalValue *GV = GA->getGlobal();
5388 if (Subtarget->isTargetELF()) {
5389 // TODO: implement the "local dynamic" model
5390 // TODO: implement the "initial exec"model for pic executables
5392 // If GV is an alias then use the aliasee for determining
5393 // thread-localness.
5394 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5395 GV = GA->resolveAliasedGlobal(false);
5397 TLSModel::Model model
5398 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5401 case TLSModel::GeneralDynamic:
5402 case TLSModel::LocalDynamic: // not implemented
5403 if (Subtarget->is64Bit())
5404 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5405 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5407 case TLSModel::InitialExec:
5408 case TLSModel::LocalExec:
5409 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5410 Subtarget->is64Bit());
5412 } else if (Subtarget->isTargetDarwin()) {
5413 // Darwin only has one model of TLS. Lower to that.
5414 unsigned char OpFlag = 0;
5415 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5416 X86ISD::WrapperRIP : X86ISD::Wrapper;
5418 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5420 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5421 !Subtarget->is64Bit();
5423 OpFlag = X86II::MO_TLVP_PIC_BASE;
5425 OpFlag = X86II::MO_TLVP;
5427 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
5429 GA->getOffset(), OpFlag);
5431 DebugLoc DL = Op.getDebugLoc();
5432 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5434 // With PIC32, the address is actually $g + Offset.
5436 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5437 DAG.getNode(X86ISD::GlobalBaseReg,
5438 DebugLoc(), getPointerTy()),
5441 // Lowering the machine isd will make sure everything is in the right
5443 SDValue Args[] = { Offset };
5444 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5446 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5447 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5448 MFI->setAdjustsStack(true);
5450 // And our return value (tls address) is in the standard call return value
5452 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5453 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5457 "TLS not implemented for this target.");
5459 llvm_unreachable("Unreachable");
5464 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5465 /// take a 2 x i32 value to shift plus a shift amount.
5466 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5467 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5468 EVT VT = Op.getValueType();
5469 unsigned VTBits = VT.getSizeInBits();
5470 DebugLoc dl = Op.getDebugLoc();
5471 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5472 SDValue ShOpLo = Op.getOperand(0);
5473 SDValue ShOpHi = Op.getOperand(1);
5474 SDValue ShAmt = Op.getOperand(2);
5475 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5476 DAG.getConstant(VTBits - 1, MVT::i8))
5477 : DAG.getConstant(0, VT);
5480 if (Op.getOpcode() == ISD::SHL_PARTS) {
5481 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5482 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5484 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5485 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5488 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5489 DAG.getConstant(VTBits, MVT::i8));
5490 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5491 AndNode, DAG.getConstant(0, MVT::i8));
5494 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5495 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5496 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5498 if (Op.getOpcode() == ISD::SHL_PARTS) {
5499 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5500 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5502 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5503 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5506 SDValue Ops[2] = { Lo, Hi };
5507 return DAG.getMergeValues(Ops, 2, dl);
5510 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5511 SelectionDAG &DAG) const {
5512 EVT SrcVT = Op.getOperand(0).getValueType();
5514 if (SrcVT.isVector()) {
5515 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5521 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5522 "Unknown SINT_TO_FP to lower!");
5524 // These are really Legal; return the operand so the caller accepts it as
5526 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5528 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5529 Subtarget->is64Bit()) {
5533 DebugLoc dl = Op.getDebugLoc();
5534 unsigned Size = SrcVT.getSizeInBits()/8;
5535 MachineFunction &MF = DAG.getMachineFunction();
5536 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5537 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5538 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5540 PseudoSourceValue::getFixedStack(SSFI), 0,
5542 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5545 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5547 SelectionDAG &DAG) const {
5549 DebugLoc dl = Op.getDebugLoc();
5551 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5553 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5555 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5556 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5557 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5558 Tys, Ops, array_lengthof(Ops));
5561 Chain = Result.getValue(1);
5562 SDValue InFlag = Result.getValue(2);
5564 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5565 // shouldn't be necessary except that RFP cannot be live across
5566 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5567 MachineFunction &MF = DAG.getMachineFunction();
5568 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5569 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5570 Tys = DAG.getVTList(MVT::Other);
5572 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5574 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5575 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5576 PseudoSourceValue::getFixedStack(SSFI), 0,
5583 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5584 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5585 SelectionDAG &DAG) const {
5586 // This algorithm is not obvious. Here it is in C code, more or less:
5588 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5589 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5590 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5592 // Copy ints to xmm registers.
5593 __m128i xh = _mm_cvtsi32_si128( hi );
5594 __m128i xl = _mm_cvtsi32_si128( lo );
5596 // Combine into low half of a single xmm register.
5597 __m128i x = _mm_unpacklo_epi32( xh, xl );
5601 // Merge in appropriate exponents to give the integer bits the right
5603 x = _mm_unpacklo_epi32( x, exp );
5605 // Subtract away the biases to deal with the IEEE-754 double precision
5607 d = _mm_sub_pd( (__m128d) x, bias );
5609 // All conversions up to here are exact. The correctly rounded result is
5610 // calculated using the current rounding mode using the following
5612 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5613 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5614 // store doesn't really need to be here (except
5615 // maybe to zero the other double)
5620 DebugLoc dl = Op.getDebugLoc();
5621 LLVMContext *Context = DAG.getContext();
5623 // Build some magic constants.
5624 std::vector<Constant*> CV0;
5625 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5626 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5627 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5628 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5629 Constant *C0 = ConstantVector::get(CV0);
5630 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5632 std::vector<Constant*> CV1;
5634 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5636 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5637 Constant *C1 = ConstantVector::get(CV1);
5638 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5640 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5641 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5643 DAG.getIntPtrConstant(1)));
5644 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5645 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5647 DAG.getIntPtrConstant(0)));
5648 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5649 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5650 PseudoSourceValue::getConstantPool(), 0,
5652 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5653 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5654 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5655 PseudoSourceValue::getConstantPool(), 0,
5657 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5659 // Add the halves; easiest way is to swap them into another reg first.
5660 int ShufMask[2] = { 1, -1 };
5661 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5662 DAG.getUNDEF(MVT::v2f64), ShufMask);
5663 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5664 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5665 DAG.getIntPtrConstant(0));
5668 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5669 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5670 SelectionDAG &DAG) const {
5671 DebugLoc dl = Op.getDebugLoc();
5672 // FP constant to bias correct the final result.
5673 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5676 // Load the 32-bit value into an XMM register.
5677 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5678 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5680 DAG.getIntPtrConstant(0)));
5682 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5683 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5684 DAG.getIntPtrConstant(0));
5686 // Or the load with the bias.
5687 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5688 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5689 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5691 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5692 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5693 MVT::v2f64, Bias)));
5694 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5695 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5696 DAG.getIntPtrConstant(0));
5698 // Subtract the bias.
5699 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5701 // Handle final rounding.
5702 EVT DestVT = Op.getValueType();
5704 if (DestVT.bitsLT(MVT::f64)) {
5705 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5706 DAG.getIntPtrConstant(0));
5707 } else if (DestVT.bitsGT(MVT::f64)) {
5708 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5711 // Handle final rounding.
5715 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5716 SelectionDAG &DAG) const {
5717 SDValue N0 = Op.getOperand(0);
5718 DebugLoc dl = Op.getDebugLoc();
5720 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5721 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5722 // the optimization here.
5723 if (DAG.SignBitIsZero(N0))
5724 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5726 EVT SrcVT = N0.getValueType();
5727 EVT DstVT = Op.getValueType();
5728 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5729 return LowerUINT_TO_FP_i64(Op, DAG);
5730 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5731 return LowerUINT_TO_FP_i32(Op, DAG);
5733 // Make a 64-bit buffer, and use it to build an FILD.
5734 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5735 if (SrcVT == MVT::i32) {
5736 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5737 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5738 getPointerTy(), StackSlot, WordOff);
5739 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5740 StackSlot, NULL, 0, false, false, 0);
5741 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5742 OffsetSlot, NULL, 0, false, false, 0);
5743 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5747 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5748 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5749 StackSlot, NULL, 0, false, false, 0);
5750 // For i64 source, we need to add the appropriate power of 2 if the input
5751 // was negative. This is the same as the optimization in
5752 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5753 // we must be careful to do the computation in x87 extended precision, not
5754 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5755 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5756 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5757 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5759 APInt FF(32, 0x5F800000ULL);
5761 // Check whether the sign bit is set.
5762 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5763 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5766 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5767 SDValue FudgePtr = DAG.getConstantPool(
5768 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5771 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5772 SDValue Zero = DAG.getIntPtrConstant(0);
5773 SDValue Four = DAG.getIntPtrConstant(4);
5774 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5776 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5778 // Load the value out, extending it from f32 to f80.
5779 // FIXME: Avoid the extend by constructing the right constant pool?
5780 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5781 FudgePtr, PseudoSourceValue::getConstantPool(),
5782 0, MVT::f32, false, false, 4);
5783 // Extend everything to 80 bits to force it to be done on x87.
5784 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5785 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5788 std::pair<SDValue,SDValue> X86TargetLowering::
5789 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5790 DebugLoc dl = Op.getDebugLoc();
5792 EVT DstTy = Op.getValueType();
5795 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5799 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5800 DstTy.getSimpleVT() >= MVT::i16 &&
5801 "Unknown FP_TO_SINT to lower!");
5803 // These are really Legal.
5804 if (DstTy == MVT::i32 &&
5805 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5806 return std::make_pair(SDValue(), SDValue());
5807 if (Subtarget->is64Bit() &&
5808 DstTy == MVT::i64 &&
5809 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5810 return std::make_pair(SDValue(), SDValue());
5812 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5814 MachineFunction &MF = DAG.getMachineFunction();
5815 unsigned MemSize = DstTy.getSizeInBits()/8;
5816 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5817 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5820 switch (DstTy.getSimpleVT().SimpleTy) {
5821 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5822 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5823 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5824 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5827 SDValue Chain = DAG.getEntryNode();
5828 SDValue Value = Op.getOperand(0);
5829 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5830 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5831 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5832 PseudoSourceValue::getFixedStack(SSFI), 0,
5834 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5836 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5838 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5839 Chain = Value.getValue(1);
5840 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5841 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5844 // Build the FP_TO_INT*_IN_MEM
5845 SDValue Ops[] = { Chain, Value, StackSlot };
5846 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5848 return std::make_pair(FIST, StackSlot);
5851 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5852 SelectionDAG &DAG) const {
5853 if (Op.getValueType().isVector()) {
5854 if (Op.getValueType() == MVT::v2i32 &&
5855 Op.getOperand(0).getValueType() == MVT::v2f64) {
5861 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5862 SDValue FIST = Vals.first, StackSlot = Vals.second;
5863 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5864 if (FIST.getNode() == 0) return Op;
5867 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5868 FIST, StackSlot, NULL, 0, false, false, 0);
5871 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5872 SelectionDAG &DAG) const {
5873 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5874 SDValue FIST = Vals.first, StackSlot = Vals.second;
5875 assert(FIST.getNode() && "Unexpected failure");
5878 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5879 FIST, StackSlot, NULL, 0, false, false, 0);
5882 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5883 SelectionDAG &DAG) const {
5884 LLVMContext *Context = DAG.getContext();
5885 DebugLoc dl = Op.getDebugLoc();
5886 EVT VT = Op.getValueType();
5889 EltVT = VT.getVectorElementType();
5890 std::vector<Constant*> CV;
5891 if (EltVT == MVT::f64) {
5892 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5896 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5902 Constant *C = ConstantVector::get(CV);
5903 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5904 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5905 PseudoSourceValue::getConstantPool(), 0,
5907 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5910 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5911 LLVMContext *Context = DAG.getContext();
5912 DebugLoc dl = Op.getDebugLoc();
5913 EVT VT = Op.getValueType();
5916 EltVT = VT.getVectorElementType();
5917 std::vector<Constant*> CV;
5918 if (EltVT == MVT::f64) {
5919 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5923 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5929 Constant *C = ConstantVector::get(CV);
5930 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5931 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5932 PseudoSourceValue::getConstantPool(), 0,
5934 if (VT.isVector()) {
5935 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5936 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5937 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5939 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5941 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5945 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5946 LLVMContext *Context = DAG.getContext();
5947 SDValue Op0 = Op.getOperand(0);
5948 SDValue Op1 = Op.getOperand(1);
5949 DebugLoc dl = Op.getDebugLoc();
5950 EVT VT = Op.getValueType();
5951 EVT SrcVT = Op1.getValueType();
5953 // If second operand is smaller, extend it first.
5954 if (SrcVT.bitsLT(VT)) {
5955 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5958 // And if it is bigger, shrink it first.
5959 if (SrcVT.bitsGT(VT)) {
5960 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5964 // At this point the operands and the result should have the same
5965 // type, and that won't be f80 since that is not custom lowered.
5967 // First get the sign bit of second operand.
5968 std::vector<Constant*> CV;
5969 if (SrcVT == MVT::f64) {
5970 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5971 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5973 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5974 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5975 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5976 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5978 Constant *C = ConstantVector::get(CV);
5979 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5980 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5981 PseudoSourceValue::getConstantPool(), 0,
5983 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5985 // Shift sign bit right or left if the two operands have different types.
5986 if (SrcVT.bitsGT(VT)) {
5987 // Op0 is MVT::f32, Op1 is MVT::f64.
5988 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5989 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5990 DAG.getConstant(32, MVT::i32));
5991 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5992 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5993 DAG.getIntPtrConstant(0));
5996 // Clear first operand sign bit.
5998 if (VT == MVT::f64) {
5999 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6000 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6002 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6003 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6004 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6005 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6007 C = ConstantVector::get(CV);
6008 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6009 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6010 PseudoSourceValue::getConstantPool(), 0,
6012 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6014 // Or the value with the sign bit.
6015 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6018 /// Emit nodes that will be selected as "test Op0,Op0", or something
6020 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6021 SelectionDAG &DAG) const {
6022 DebugLoc dl = Op.getDebugLoc();
6024 // CF and OF aren't always set the way we want. Determine which
6025 // of these we need.
6026 bool NeedCF = false;
6027 bool NeedOF = false;
6029 case X86::COND_A: case X86::COND_AE:
6030 case X86::COND_B: case X86::COND_BE:
6033 case X86::COND_G: case X86::COND_GE:
6034 case X86::COND_L: case X86::COND_LE:
6035 case X86::COND_O: case X86::COND_NO:
6041 // See if we can use the EFLAGS value from the operand instead of
6042 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6043 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6044 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
6045 unsigned Opcode = 0;
6046 unsigned NumOperands = 0;
6047 switch (Op.getNode()->getOpcode()) {
6049 // Due to an isel shortcoming, be conservative if this add is
6050 // likely to be selected as part of a load-modify-store
6051 // instruction. When the root node in a match is a store, isel
6052 // doesn't know how to remap non-chain non-flag uses of other
6053 // nodes in the match, such as the ADD in this case. This leads
6054 // to the ADD being left around and reselected, with the result
6055 // being two adds in the output. Alas, even if none our users
6056 // are stores, that doesn't prove we're O.K. Ergo, if we have
6057 // any parents that aren't CopyToReg or SETCC, eschew INC/DEC.
6058 // A better fix seems to require climbing the DAG back to the
6059 // root, and it doesn't seem to be worth the effort.
6060 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6061 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6062 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6064 if (ConstantSDNode *C =
6065 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6066 // An add of one will be selected as an INC.
6067 if (C->getAPIntValue() == 1) {
6068 Opcode = X86ISD::INC;
6072 // An add of negative one (subtract of one) will be selected as a DEC.
6073 if (C->getAPIntValue().isAllOnesValue()) {
6074 Opcode = X86ISD::DEC;
6079 // Otherwise use a regular EFLAGS-setting add.
6080 Opcode = X86ISD::ADD;
6084 // If the primary and result isn't used, don't bother using X86ISD::AND,
6085 // because a TEST instruction will be better.
6086 bool NonFlagUse = false;
6087 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6088 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6090 unsigned UOpNo = UI.getOperandNo();
6091 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6092 // Look pass truncate.
6093 UOpNo = User->use_begin().getOperandNo();
6094 User = *User->use_begin();
6096 if (User->getOpcode() != ISD::BRCOND &&
6097 User->getOpcode() != ISD::SETCC &&
6098 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6110 // Due to the ISEL shortcoming noted above, be conservative if this op is
6111 // likely to be selected as part of a load-modify-store instruction.
6112 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6113 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6114 if (UI->getOpcode() == ISD::STORE)
6116 // Otherwise use a regular EFLAGS-setting instruction.
6117 switch (Op.getNode()->getOpcode()) {
6118 case ISD::SUB: Opcode = X86ISD::SUB; break;
6119 case ISD::OR: Opcode = X86ISD::OR; break;
6120 case ISD::XOR: Opcode = X86ISD::XOR; break;
6121 case ISD::AND: Opcode = X86ISD::AND; break;
6122 default: llvm_unreachable("unexpected operator!");
6133 return SDValue(Op.getNode(), 1);
6139 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6140 SmallVector<SDValue, 4> Ops;
6141 for (unsigned i = 0; i != NumOperands; ++i)
6142 Ops.push_back(Op.getOperand(i));
6143 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6144 DAG.ReplaceAllUsesWith(Op, New);
6145 return SDValue(New.getNode(), 1);
6149 // Otherwise just emit a CMP with 0, which is the TEST pattern.
6150 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6151 DAG.getConstant(0, Op.getValueType()));
6154 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6156 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6157 SelectionDAG &DAG) const {
6158 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6159 if (C->getAPIntValue() == 0)
6160 return EmitTest(Op0, X86CC, DAG);
6162 DebugLoc dl = Op0.getDebugLoc();
6163 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6166 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6167 /// if it's possible.
6168 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6169 DebugLoc dl, SelectionDAG &DAG) const {
6170 SDValue Op0 = And.getOperand(0);
6171 SDValue Op1 = And.getOperand(1);
6172 if (Op0.getOpcode() == ISD::TRUNCATE)
6173 Op0 = Op0.getOperand(0);
6174 if (Op1.getOpcode() == ISD::TRUNCATE)
6175 Op1 = Op1.getOperand(0);
6178 if (Op1.getOpcode() == ISD::SHL) {
6179 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
6180 if (And10C->getZExtValue() == 1) {
6182 RHS = Op1.getOperand(1);
6184 } else if (Op0.getOpcode() == ISD::SHL) {
6185 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6186 if (And00C->getZExtValue() == 1) {
6188 RHS = Op0.getOperand(1);
6190 } else if (Op1.getOpcode() == ISD::Constant) {
6191 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6192 SDValue AndLHS = Op0;
6193 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6194 LHS = AndLHS.getOperand(0);
6195 RHS = AndLHS.getOperand(1);
6199 if (LHS.getNode()) {
6200 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6201 // instruction. Since the shift amount is in-range-or-undefined, we know
6202 // that doing a bittest on the i32 value is ok. We extend to i32 because
6203 // the encoding for the i16 version is larger than the i32 version.
6204 // Also promote i16 to i32 for performance / code size reason.
6205 if (LHS.getValueType() == MVT::i8 ||
6206 LHS.getValueType() == MVT::i16)
6207 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6209 // If the operand types disagree, extend the shift amount to match. Since
6210 // BT ignores high bits (like shifts) we can use anyextend.
6211 if (LHS.getValueType() != RHS.getValueType())
6212 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6214 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6215 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6216 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6217 DAG.getConstant(Cond, MVT::i8), BT);
6223 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6224 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6225 SDValue Op0 = Op.getOperand(0);
6226 SDValue Op1 = Op.getOperand(1);
6227 DebugLoc dl = Op.getDebugLoc();
6228 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6230 // Optimize to BT if possible.
6231 // Lower (X & (1 << N)) == 0 to BT(X, N).
6232 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6233 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6234 if (Op0.getOpcode() == ISD::AND &&
6236 Op1.getOpcode() == ISD::Constant &&
6237 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
6238 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6239 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6240 if (NewSetCC.getNode())
6244 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6245 if (Op0.getOpcode() == X86ISD::SETCC &&
6246 Op1.getOpcode() == ISD::Constant &&
6247 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6248 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6249 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6250 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6251 bool Invert = (CC == ISD::SETNE) ^
6252 cast<ConstantSDNode>(Op1)->isNullValue();
6254 CCode = X86::GetOppositeBranchCondition(CCode);
6255 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6256 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6259 bool isFP = Op1.getValueType().isFloatingPoint();
6260 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6261 if (X86CC == X86::COND_INVALID)
6264 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6266 // Use sbb x, x to materialize carry bit into a GPR.
6267 if (X86CC == X86::COND_B)
6268 return DAG.getNode(ISD::AND, dl, MVT::i8,
6269 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6270 DAG.getConstant(X86CC, MVT::i8), Cond),
6271 DAG.getConstant(1, MVT::i8));
6273 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6274 DAG.getConstant(X86CC, MVT::i8), Cond);
6277 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6279 SDValue Op0 = Op.getOperand(0);
6280 SDValue Op1 = Op.getOperand(1);
6281 SDValue CC = Op.getOperand(2);
6282 EVT VT = Op.getValueType();
6283 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6284 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6285 DebugLoc dl = Op.getDebugLoc();
6289 EVT VT0 = Op0.getValueType();
6290 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6291 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6294 switch (SetCCOpcode) {
6297 case ISD::SETEQ: SSECC = 0; break;
6299 case ISD::SETGT: Swap = true; // Fallthrough
6301 case ISD::SETOLT: SSECC = 1; break;
6303 case ISD::SETGE: Swap = true; // Fallthrough
6305 case ISD::SETOLE: SSECC = 2; break;
6306 case ISD::SETUO: SSECC = 3; break;
6308 case ISD::SETNE: SSECC = 4; break;
6309 case ISD::SETULE: Swap = true;
6310 case ISD::SETUGE: SSECC = 5; break;
6311 case ISD::SETULT: Swap = true;
6312 case ISD::SETUGT: SSECC = 6; break;
6313 case ISD::SETO: SSECC = 7; break;
6316 std::swap(Op0, Op1);
6318 // In the two special cases we can't handle, emit two comparisons.
6320 if (SetCCOpcode == ISD::SETUEQ) {
6322 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6323 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6324 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6326 else if (SetCCOpcode == ISD::SETONE) {
6328 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6329 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6330 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6332 llvm_unreachable("Illegal FP comparison");
6334 // Handle all other FP comparisons here.
6335 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6338 // We are handling one of the integer comparisons here. Since SSE only has
6339 // GT and EQ comparisons for integer, swapping operands and multiple
6340 // operations may be required for some comparisons.
6341 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6342 bool Swap = false, Invert = false, FlipSigns = false;
6344 switch (VT.getSimpleVT().SimpleTy) {
6347 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6349 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6351 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6352 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6355 switch (SetCCOpcode) {
6357 case ISD::SETNE: Invert = true;
6358 case ISD::SETEQ: Opc = EQOpc; break;
6359 case ISD::SETLT: Swap = true;
6360 case ISD::SETGT: Opc = GTOpc; break;
6361 case ISD::SETGE: Swap = true;
6362 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6363 case ISD::SETULT: Swap = true;
6364 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6365 case ISD::SETUGE: Swap = true;
6366 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6369 std::swap(Op0, Op1);
6371 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6372 // bits of the inputs before performing those operations.
6374 EVT EltVT = VT.getVectorElementType();
6375 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6377 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6378 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6380 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6381 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6384 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6386 // If the logical-not of the result is required, perform that now.
6388 Result = DAG.getNOT(dl, Result, VT);
6393 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6394 static bool isX86LogicalCmp(SDValue Op) {
6395 unsigned Opc = Op.getNode()->getOpcode();
6396 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6398 if (Op.getResNo() == 1 &&
6399 (Opc == X86ISD::ADD ||
6400 Opc == X86ISD::SUB ||
6401 Opc == X86ISD::SMUL ||
6402 Opc == X86ISD::UMUL ||
6403 Opc == X86ISD::INC ||
6404 Opc == X86ISD::DEC ||
6405 Opc == X86ISD::OR ||
6406 Opc == X86ISD::XOR ||
6407 Opc == X86ISD::AND))
6413 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6414 bool addTest = true;
6415 SDValue Cond = Op.getOperand(0);
6416 DebugLoc dl = Op.getDebugLoc();
6419 if (Cond.getOpcode() == ISD::SETCC) {
6420 SDValue NewCond = LowerSETCC(Cond, DAG);
6421 if (NewCond.getNode())
6425 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6426 SDValue Op1 = Op.getOperand(1);
6427 SDValue Op2 = Op.getOperand(2);
6428 if (Cond.getOpcode() == X86ISD::SETCC &&
6429 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6430 SDValue Cmp = Cond.getOperand(1);
6431 if (Cmp.getOpcode() == X86ISD::CMP) {
6432 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6433 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6434 ConstantSDNode *RHSC =
6435 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6436 if (N1C && N1C->isAllOnesValue() &&
6437 N2C && N2C->isNullValue() &&
6438 RHSC && RHSC->isNullValue()) {
6439 SDValue CmpOp0 = Cmp.getOperand(0);
6440 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6441 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6442 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6443 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6448 // Look pass (and (setcc_carry (cmp ...)), 1).
6449 if (Cond.getOpcode() == ISD::AND &&
6450 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6451 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6452 if (C && C->getAPIntValue() == 1)
6453 Cond = Cond.getOperand(0);
6456 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6457 // setting operand in place of the X86ISD::SETCC.
6458 if (Cond.getOpcode() == X86ISD::SETCC ||
6459 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6460 CC = Cond.getOperand(0);
6462 SDValue Cmp = Cond.getOperand(1);
6463 unsigned Opc = Cmp.getOpcode();
6464 EVT VT = Op.getValueType();
6466 bool IllegalFPCMov = false;
6467 if (VT.isFloatingPoint() && !VT.isVector() &&
6468 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6469 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6471 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6472 Opc == X86ISD::BT) { // FIXME
6479 // Look pass the truncate.
6480 if (Cond.getOpcode() == ISD::TRUNCATE)
6481 Cond = Cond.getOperand(0);
6483 // We know the result of AND is compared against zero. Try to match
6485 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6486 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6487 if (NewSetCC.getNode()) {
6488 CC = NewSetCC.getOperand(0);
6489 Cond = NewSetCC.getOperand(1);
6496 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6497 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6500 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6501 // condition is true.
6502 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6503 SDValue Ops[] = { Op2, Op1, CC, Cond };
6504 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6507 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6508 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6509 // from the AND / OR.
6510 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6511 Opc = Op.getOpcode();
6512 if (Opc != ISD::OR && Opc != ISD::AND)
6514 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6515 Op.getOperand(0).hasOneUse() &&
6516 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6517 Op.getOperand(1).hasOneUse());
6520 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6521 // 1 and that the SETCC node has a single use.
6522 static bool isXor1OfSetCC(SDValue Op) {
6523 if (Op.getOpcode() != ISD::XOR)
6525 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6526 if (N1C && N1C->getAPIntValue() == 1) {
6527 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6528 Op.getOperand(0).hasOneUse();
6533 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6534 bool addTest = true;
6535 SDValue Chain = Op.getOperand(0);
6536 SDValue Cond = Op.getOperand(1);
6537 SDValue Dest = Op.getOperand(2);
6538 DebugLoc dl = Op.getDebugLoc();
6541 if (Cond.getOpcode() == ISD::SETCC) {
6542 SDValue NewCond = LowerSETCC(Cond, DAG);
6543 if (NewCond.getNode())
6547 // FIXME: LowerXALUO doesn't handle these!!
6548 else if (Cond.getOpcode() == X86ISD::ADD ||
6549 Cond.getOpcode() == X86ISD::SUB ||
6550 Cond.getOpcode() == X86ISD::SMUL ||
6551 Cond.getOpcode() == X86ISD::UMUL)
6552 Cond = LowerXALUO(Cond, DAG);
6555 // Look pass (and (setcc_carry (cmp ...)), 1).
6556 if (Cond.getOpcode() == ISD::AND &&
6557 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6558 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6559 if (C && C->getAPIntValue() == 1)
6560 Cond = Cond.getOperand(0);
6563 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6564 // setting operand in place of the X86ISD::SETCC.
6565 if (Cond.getOpcode() == X86ISD::SETCC ||
6566 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6567 CC = Cond.getOperand(0);
6569 SDValue Cmp = Cond.getOperand(1);
6570 unsigned Opc = Cmp.getOpcode();
6571 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6572 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6576 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6580 // These can only come from an arithmetic instruction with overflow,
6581 // e.g. SADDO, UADDO.
6582 Cond = Cond.getNode()->getOperand(1);
6589 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6590 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6591 if (CondOpc == ISD::OR) {
6592 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6593 // two branches instead of an explicit OR instruction with a
6595 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6596 isX86LogicalCmp(Cmp)) {
6597 CC = Cond.getOperand(0).getOperand(0);
6598 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6599 Chain, Dest, CC, Cmp);
6600 CC = Cond.getOperand(1).getOperand(0);
6604 } else { // ISD::AND
6605 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6606 // two branches instead of an explicit AND instruction with a
6607 // separate test. However, we only do this if this block doesn't
6608 // have a fall-through edge, because this requires an explicit
6609 // jmp when the condition is false.
6610 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6611 isX86LogicalCmp(Cmp) &&
6612 Op.getNode()->hasOneUse()) {
6613 X86::CondCode CCode =
6614 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6615 CCode = X86::GetOppositeBranchCondition(CCode);
6616 CC = DAG.getConstant(CCode, MVT::i8);
6617 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6618 // Look for an unconditional branch following this conditional branch.
6619 // We need this because we need to reverse the successors in order
6620 // to implement FCMP_OEQ.
6621 if (User.getOpcode() == ISD::BR) {
6622 SDValue FalseBB = User.getOperand(1);
6624 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6625 assert(NewBR == User);
6628 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6629 Chain, Dest, CC, Cmp);
6630 X86::CondCode CCode =
6631 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6632 CCode = X86::GetOppositeBranchCondition(CCode);
6633 CC = DAG.getConstant(CCode, MVT::i8);
6639 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6640 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6641 // It should be transformed during dag combiner except when the condition
6642 // is set by a arithmetics with overflow node.
6643 X86::CondCode CCode =
6644 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6645 CCode = X86::GetOppositeBranchCondition(CCode);
6646 CC = DAG.getConstant(CCode, MVT::i8);
6647 Cond = Cond.getOperand(0).getOperand(1);
6653 // Look pass the truncate.
6654 if (Cond.getOpcode() == ISD::TRUNCATE)
6655 Cond = Cond.getOperand(0);
6657 // We know the result of AND is compared against zero. Try to match
6659 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6660 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6661 if (NewSetCC.getNode()) {
6662 CC = NewSetCC.getOperand(0);
6663 Cond = NewSetCC.getOperand(1);
6670 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6671 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6673 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6674 Chain, Dest, CC, Cond);
6678 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6679 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6680 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6681 // that the guard pages used by the OS virtual memory manager are allocated in
6682 // correct sequence.
6684 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6685 SelectionDAG &DAG) const {
6686 assert(Subtarget->isTargetCygMing() &&
6687 "This should be used only on Cygwin/Mingw targets");
6688 DebugLoc dl = Op.getDebugLoc();
6691 SDValue Chain = Op.getOperand(0);
6692 SDValue Size = Op.getOperand(1);
6693 // FIXME: Ensure alignment here
6697 EVT IntPtr = getPointerTy();
6698 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6700 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6701 Flag = Chain.getValue(1);
6703 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6705 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6706 Flag = Chain.getValue(1);
6708 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6710 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6711 return DAG.getMergeValues(Ops1, 2, dl);
6714 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6715 MachineFunction &MF = DAG.getMachineFunction();
6716 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6718 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6719 DebugLoc dl = Op.getDebugLoc();
6721 if (!Subtarget->is64Bit()) {
6722 // vastart just stores the address of the VarArgsFrameIndex slot into the
6723 // memory location argument.
6724 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6726 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6731 // gp_offset (0 - 6 * 8)
6732 // fp_offset (48 - 48 + 8 * 16)
6733 // overflow_arg_area (point to parameters coming in memory).
6735 SmallVector<SDValue, 8> MemOps;
6736 SDValue FIN = Op.getOperand(1);
6738 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6739 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6741 FIN, SV, 0, false, false, 0);
6742 MemOps.push_back(Store);
6745 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6746 FIN, DAG.getIntPtrConstant(4));
6747 Store = DAG.getStore(Op.getOperand(0), dl,
6748 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6750 FIN, SV, 0, false, false, 0);
6751 MemOps.push_back(Store);
6753 // Store ptr to overflow_arg_area
6754 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6755 FIN, DAG.getIntPtrConstant(4));
6756 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6758 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6760 MemOps.push_back(Store);
6762 // Store ptr to reg_save_area.
6763 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6764 FIN, DAG.getIntPtrConstant(8));
6765 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6767 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6769 MemOps.push_back(Store);
6770 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6771 &MemOps[0], MemOps.size());
6774 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6775 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6776 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6777 SDValue Chain = Op.getOperand(0);
6778 SDValue SrcPtr = Op.getOperand(1);
6779 SDValue SrcSV = Op.getOperand(2);
6781 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6785 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6786 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6787 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6788 SDValue Chain = Op.getOperand(0);
6789 SDValue DstPtr = Op.getOperand(1);
6790 SDValue SrcPtr = Op.getOperand(2);
6791 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6792 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6793 DebugLoc dl = Op.getDebugLoc();
6795 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6796 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6797 false, DstSV, 0, SrcSV, 0);
6801 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6802 DebugLoc dl = Op.getDebugLoc();
6803 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6805 default: return SDValue(); // Don't custom lower most intrinsics.
6806 // Comparison intrinsics.
6807 case Intrinsic::x86_sse_comieq_ss:
6808 case Intrinsic::x86_sse_comilt_ss:
6809 case Intrinsic::x86_sse_comile_ss:
6810 case Intrinsic::x86_sse_comigt_ss:
6811 case Intrinsic::x86_sse_comige_ss:
6812 case Intrinsic::x86_sse_comineq_ss:
6813 case Intrinsic::x86_sse_ucomieq_ss:
6814 case Intrinsic::x86_sse_ucomilt_ss:
6815 case Intrinsic::x86_sse_ucomile_ss:
6816 case Intrinsic::x86_sse_ucomigt_ss:
6817 case Intrinsic::x86_sse_ucomige_ss:
6818 case Intrinsic::x86_sse_ucomineq_ss:
6819 case Intrinsic::x86_sse2_comieq_sd:
6820 case Intrinsic::x86_sse2_comilt_sd:
6821 case Intrinsic::x86_sse2_comile_sd:
6822 case Intrinsic::x86_sse2_comigt_sd:
6823 case Intrinsic::x86_sse2_comige_sd:
6824 case Intrinsic::x86_sse2_comineq_sd:
6825 case Intrinsic::x86_sse2_ucomieq_sd:
6826 case Intrinsic::x86_sse2_ucomilt_sd:
6827 case Intrinsic::x86_sse2_ucomile_sd:
6828 case Intrinsic::x86_sse2_ucomigt_sd:
6829 case Intrinsic::x86_sse2_ucomige_sd:
6830 case Intrinsic::x86_sse2_ucomineq_sd: {
6832 ISD::CondCode CC = ISD::SETCC_INVALID;
6835 case Intrinsic::x86_sse_comieq_ss:
6836 case Intrinsic::x86_sse2_comieq_sd:
6840 case Intrinsic::x86_sse_comilt_ss:
6841 case Intrinsic::x86_sse2_comilt_sd:
6845 case Intrinsic::x86_sse_comile_ss:
6846 case Intrinsic::x86_sse2_comile_sd:
6850 case Intrinsic::x86_sse_comigt_ss:
6851 case Intrinsic::x86_sse2_comigt_sd:
6855 case Intrinsic::x86_sse_comige_ss:
6856 case Intrinsic::x86_sse2_comige_sd:
6860 case Intrinsic::x86_sse_comineq_ss:
6861 case Intrinsic::x86_sse2_comineq_sd:
6865 case Intrinsic::x86_sse_ucomieq_ss:
6866 case Intrinsic::x86_sse2_ucomieq_sd:
6867 Opc = X86ISD::UCOMI;
6870 case Intrinsic::x86_sse_ucomilt_ss:
6871 case Intrinsic::x86_sse2_ucomilt_sd:
6872 Opc = X86ISD::UCOMI;
6875 case Intrinsic::x86_sse_ucomile_ss:
6876 case Intrinsic::x86_sse2_ucomile_sd:
6877 Opc = X86ISD::UCOMI;
6880 case Intrinsic::x86_sse_ucomigt_ss:
6881 case Intrinsic::x86_sse2_ucomigt_sd:
6882 Opc = X86ISD::UCOMI;
6885 case Intrinsic::x86_sse_ucomige_ss:
6886 case Intrinsic::x86_sse2_ucomige_sd:
6887 Opc = X86ISD::UCOMI;
6890 case Intrinsic::x86_sse_ucomineq_ss:
6891 case Intrinsic::x86_sse2_ucomineq_sd:
6892 Opc = X86ISD::UCOMI;
6897 SDValue LHS = Op.getOperand(1);
6898 SDValue RHS = Op.getOperand(2);
6899 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6900 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6901 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6902 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6903 DAG.getConstant(X86CC, MVT::i8), Cond);
6904 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6906 // ptest intrinsics. The intrinsic these come from are designed to return
6907 // an integer value, not just an instruction so lower it to the ptest
6908 // pattern and a setcc for the result.
6909 case Intrinsic::x86_sse41_ptestz:
6910 case Intrinsic::x86_sse41_ptestc:
6911 case Intrinsic::x86_sse41_ptestnzc:{
6914 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6915 case Intrinsic::x86_sse41_ptestz:
6917 X86CC = X86::COND_E;
6919 case Intrinsic::x86_sse41_ptestc:
6921 X86CC = X86::COND_B;
6923 case Intrinsic::x86_sse41_ptestnzc:
6925 X86CC = X86::COND_A;
6929 SDValue LHS = Op.getOperand(1);
6930 SDValue RHS = Op.getOperand(2);
6931 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6932 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6933 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6934 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6937 // Fix vector shift instructions where the last operand is a non-immediate
6939 case Intrinsic::x86_sse2_pslli_w:
6940 case Intrinsic::x86_sse2_pslli_d:
6941 case Intrinsic::x86_sse2_pslli_q:
6942 case Intrinsic::x86_sse2_psrli_w:
6943 case Intrinsic::x86_sse2_psrli_d:
6944 case Intrinsic::x86_sse2_psrli_q:
6945 case Intrinsic::x86_sse2_psrai_w:
6946 case Intrinsic::x86_sse2_psrai_d:
6947 case Intrinsic::x86_mmx_pslli_w:
6948 case Intrinsic::x86_mmx_pslli_d:
6949 case Intrinsic::x86_mmx_pslli_q:
6950 case Intrinsic::x86_mmx_psrli_w:
6951 case Intrinsic::x86_mmx_psrli_d:
6952 case Intrinsic::x86_mmx_psrli_q:
6953 case Intrinsic::x86_mmx_psrai_w:
6954 case Intrinsic::x86_mmx_psrai_d: {
6955 SDValue ShAmt = Op.getOperand(2);
6956 if (isa<ConstantSDNode>(ShAmt))
6959 unsigned NewIntNo = 0;
6960 EVT ShAmtVT = MVT::v4i32;
6962 case Intrinsic::x86_sse2_pslli_w:
6963 NewIntNo = Intrinsic::x86_sse2_psll_w;
6965 case Intrinsic::x86_sse2_pslli_d:
6966 NewIntNo = Intrinsic::x86_sse2_psll_d;
6968 case Intrinsic::x86_sse2_pslli_q:
6969 NewIntNo = Intrinsic::x86_sse2_psll_q;
6971 case Intrinsic::x86_sse2_psrli_w:
6972 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6974 case Intrinsic::x86_sse2_psrli_d:
6975 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6977 case Intrinsic::x86_sse2_psrli_q:
6978 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6980 case Intrinsic::x86_sse2_psrai_w:
6981 NewIntNo = Intrinsic::x86_sse2_psra_w;
6983 case Intrinsic::x86_sse2_psrai_d:
6984 NewIntNo = Intrinsic::x86_sse2_psra_d;
6987 ShAmtVT = MVT::v2i32;
6989 case Intrinsic::x86_mmx_pslli_w:
6990 NewIntNo = Intrinsic::x86_mmx_psll_w;
6992 case Intrinsic::x86_mmx_pslli_d:
6993 NewIntNo = Intrinsic::x86_mmx_psll_d;
6995 case Intrinsic::x86_mmx_pslli_q:
6996 NewIntNo = Intrinsic::x86_mmx_psll_q;
6998 case Intrinsic::x86_mmx_psrli_w:
6999 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7001 case Intrinsic::x86_mmx_psrli_d:
7002 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7004 case Intrinsic::x86_mmx_psrli_q:
7005 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7007 case Intrinsic::x86_mmx_psrai_w:
7008 NewIntNo = Intrinsic::x86_mmx_psra_w;
7010 case Intrinsic::x86_mmx_psrai_d:
7011 NewIntNo = Intrinsic::x86_mmx_psra_d;
7013 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7019 // The vector shift intrinsics with scalars uses 32b shift amounts but
7020 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7024 ShOps[1] = DAG.getConstant(0, MVT::i32);
7025 if (ShAmtVT == MVT::v4i32) {
7026 ShOps[2] = DAG.getUNDEF(MVT::i32);
7027 ShOps[3] = DAG.getUNDEF(MVT::i32);
7028 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7030 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7033 EVT VT = Op.getValueType();
7034 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7035 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7036 DAG.getConstant(NewIntNo, MVT::i32),
7037 Op.getOperand(1), ShAmt);
7042 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7043 SelectionDAG &DAG) const {
7044 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7045 MFI->setReturnAddressIsTaken(true);
7047 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7048 DebugLoc dl = Op.getDebugLoc();
7051 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7053 DAG.getConstant(TD->getPointerSize(),
7054 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7055 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7056 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7058 NULL, 0, false, false, 0);
7061 // Just load the return address.
7062 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7063 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7064 RetAddrFI, NULL, 0, false, false, 0);
7067 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7068 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7069 MFI->setFrameAddressIsTaken(true);
7071 EVT VT = Op.getValueType();
7072 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7073 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7074 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7075 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7077 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7082 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7083 SelectionDAG &DAG) const {
7084 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7087 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7088 MachineFunction &MF = DAG.getMachineFunction();
7089 SDValue Chain = Op.getOperand(0);
7090 SDValue Offset = Op.getOperand(1);
7091 SDValue Handler = Op.getOperand(2);
7092 DebugLoc dl = Op.getDebugLoc();
7094 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7096 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7098 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7099 DAG.getIntPtrConstant(-TD->getPointerSize()));
7100 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7101 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7102 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7103 MF.getRegInfo().addLiveOut(StoreAddrReg);
7105 return DAG.getNode(X86ISD::EH_RETURN, dl,
7107 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7110 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7111 SelectionDAG &DAG) const {
7112 SDValue Root = Op.getOperand(0);
7113 SDValue Trmp = Op.getOperand(1); // trampoline
7114 SDValue FPtr = Op.getOperand(2); // nested function
7115 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7116 DebugLoc dl = Op.getDebugLoc();
7118 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7120 if (Subtarget->is64Bit()) {
7121 SDValue OutChains[6];
7123 // Large code-model.
7124 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7125 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7127 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7128 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7130 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7132 // Load the pointer to the nested function into R11.
7133 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7134 SDValue Addr = Trmp;
7135 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7136 Addr, TrmpAddr, 0, false, false, 0);
7138 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7139 DAG.getConstant(2, MVT::i64));
7140 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7143 // Load the 'nest' parameter value into R10.
7144 // R10 is specified in X86CallingConv.td
7145 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7146 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7147 DAG.getConstant(10, MVT::i64));
7148 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7149 Addr, TrmpAddr, 10, false, false, 0);
7151 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7152 DAG.getConstant(12, MVT::i64));
7153 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7156 // Jump to the nested function.
7157 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7158 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7159 DAG.getConstant(20, MVT::i64));
7160 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7161 Addr, TrmpAddr, 20, false, false, 0);
7163 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7164 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7165 DAG.getConstant(22, MVT::i64));
7166 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7167 TrmpAddr, 22, false, false, 0);
7170 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7171 return DAG.getMergeValues(Ops, 2, dl);
7173 const Function *Func =
7174 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7175 CallingConv::ID CC = Func->getCallingConv();
7180 llvm_unreachable("Unsupported calling convention");
7181 case CallingConv::C:
7182 case CallingConv::X86_StdCall: {
7183 // Pass 'nest' parameter in ECX.
7184 // Must be kept in sync with X86CallingConv.td
7187 // Check that ECX wasn't needed by an 'inreg' parameter.
7188 const FunctionType *FTy = Func->getFunctionType();
7189 const AttrListPtr &Attrs = Func->getAttributes();
7191 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7192 unsigned InRegCount = 0;
7195 for (FunctionType::param_iterator I = FTy->param_begin(),
7196 E = FTy->param_end(); I != E; ++I, ++Idx)
7197 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7198 // FIXME: should only count parameters that are lowered to integers.
7199 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7201 if (InRegCount > 2) {
7202 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7207 case CallingConv::X86_FastCall:
7208 case CallingConv::X86_ThisCall:
7209 case CallingConv::Fast:
7210 // Pass 'nest' parameter in EAX.
7211 // Must be kept in sync with X86CallingConv.td
7216 SDValue OutChains[4];
7219 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7220 DAG.getConstant(10, MVT::i32));
7221 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7223 // This is storing the opcode for MOV32ri.
7224 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7225 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7226 OutChains[0] = DAG.getStore(Root, dl,
7227 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7228 Trmp, TrmpAddr, 0, false, false, 0);
7230 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7231 DAG.getConstant(1, MVT::i32));
7232 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7235 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7236 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7237 DAG.getConstant(5, MVT::i32));
7238 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7239 TrmpAddr, 5, false, false, 1);
7241 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7242 DAG.getConstant(6, MVT::i32));
7243 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7247 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7248 return DAG.getMergeValues(Ops, 2, dl);
7252 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7253 SelectionDAG &DAG) const {
7255 The rounding mode is in bits 11:10 of FPSR, and has the following
7262 FLT_ROUNDS, on the other hand, expects the following:
7269 To perform the conversion, we do:
7270 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7273 MachineFunction &MF = DAG.getMachineFunction();
7274 const TargetMachine &TM = MF.getTarget();
7275 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7276 unsigned StackAlignment = TFI.getStackAlignment();
7277 EVT VT = Op.getValueType();
7278 DebugLoc dl = Op.getDebugLoc();
7280 // Save FP Control Word to stack slot
7281 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7282 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7284 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7285 DAG.getEntryNode(), StackSlot);
7287 // Load FP Control Word from stack slot
7288 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7291 // Transform as necessary
7293 DAG.getNode(ISD::SRL, dl, MVT::i16,
7294 DAG.getNode(ISD::AND, dl, MVT::i16,
7295 CWD, DAG.getConstant(0x800, MVT::i16)),
7296 DAG.getConstant(11, MVT::i8));
7298 DAG.getNode(ISD::SRL, dl, MVT::i16,
7299 DAG.getNode(ISD::AND, dl, MVT::i16,
7300 CWD, DAG.getConstant(0x400, MVT::i16)),
7301 DAG.getConstant(9, MVT::i8));
7304 DAG.getNode(ISD::AND, dl, MVT::i16,
7305 DAG.getNode(ISD::ADD, dl, MVT::i16,
7306 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7307 DAG.getConstant(1, MVT::i16)),
7308 DAG.getConstant(3, MVT::i16));
7311 return DAG.getNode((VT.getSizeInBits() < 16 ?
7312 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7315 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7316 EVT VT = Op.getValueType();
7318 unsigned NumBits = VT.getSizeInBits();
7319 DebugLoc dl = Op.getDebugLoc();
7321 Op = Op.getOperand(0);
7322 if (VT == MVT::i8) {
7323 // Zero extend to i32 since there is not an i8 bsr.
7325 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7328 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7329 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7330 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7332 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7335 DAG.getConstant(NumBits+NumBits-1, OpVT),
7336 DAG.getConstant(X86::COND_E, MVT::i8),
7339 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7341 // Finally xor with NumBits-1.
7342 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7345 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7349 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7350 EVT VT = Op.getValueType();
7352 unsigned NumBits = VT.getSizeInBits();
7353 DebugLoc dl = Op.getDebugLoc();
7355 Op = Op.getOperand(0);
7356 if (VT == MVT::i8) {
7358 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7361 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7362 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7363 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7365 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7368 DAG.getConstant(NumBits, OpVT),
7369 DAG.getConstant(X86::COND_E, MVT::i8),
7372 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7375 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7379 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7380 EVT VT = Op.getValueType();
7381 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7382 DebugLoc dl = Op.getDebugLoc();
7384 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7385 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7386 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7387 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7388 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7390 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7391 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7392 // return AloBlo + AloBhi + AhiBlo;
7394 SDValue A = Op.getOperand(0);
7395 SDValue B = Op.getOperand(1);
7397 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7398 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7399 A, DAG.getConstant(32, MVT::i32));
7400 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7401 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7402 B, DAG.getConstant(32, MVT::i32));
7403 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7404 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7406 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7407 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7409 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7410 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7412 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7413 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7414 AloBhi, DAG.getConstant(32, MVT::i32));
7415 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7416 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7417 AhiBlo, DAG.getConstant(32, MVT::i32));
7418 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7419 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7424 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7425 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7426 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7427 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7428 // has only one use.
7429 SDNode *N = Op.getNode();
7430 SDValue LHS = N->getOperand(0);
7431 SDValue RHS = N->getOperand(1);
7432 unsigned BaseOp = 0;
7434 DebugLoc dl = Op.getDebugLoc();
7436 switch (Op.getOpcode()) {
7437 default: llvm_unreachable("Unknown ovf instruction!");
7439 // A subtract of one will be selected as a INC. Note that INC doesn't
7440 // set CF, so we can't do this for UADDO.
7441 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7442 if (C->getAPIntValue() == 1) {
7443 BaseOp = X86ISD::INC;
7447 BaseOp = X86ISD::ADD;
7451 BaseOp = X86ISD::ADD;
7455 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7456 // set CF, so we can't do this for USUBO.
7457 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7458 if (C->getAPIntValue() == 1) {
7459 BaseOp = X86ISD::DEC;
7463 BaseOp = X86ISD::SUB;
7467 BaseOp = X86ISD::SUB;
7471 BaseOp = X86ISD::SMUL;
7475 BaseOp = X86ISD::UMUL;
7480 // Also sets EFLAGS.
7481 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7482 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7485 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7486 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7488 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7492 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7493 EVT T = Op.getValueType();
7494 DebugLoc dl = Op.getDebugLoc();
7497 switch(T.getSimpleVT().SimpleTy) {
7499 assert(false && "Invalid value type!");
7500 case MVT::i8: Reg = X86::AL; size = 1; break;
7501 case MVT::i16: Reg = X86::AX; size = 2; break;
7502 case MVT::i32: Reg = X86::EAX; size = 4; break;
7504 assert(Subtarget->is64Bit() && "Node not type legal!");
7505 Reg = X86::RAX; size = 8;
7508 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7509 Op.getOperand(2), SDValue());
7510 SDValue Ops[] = { cpIn.getValue(0),
7513 DAG.getTargetConstant(size, MVT::i8),
7515 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7516 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7518 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7522 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7523 SelectionDAG &DAG) const {
7524 assert(Subtarget->is64Bit() && "Result not type legalized?");
7525 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7526 SDValue TheChain = Op.getOperand(0);
7527 DebugLoc dl = Op.getDebugLoc();
7528 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7529 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7530 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7532 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7533 DAG.getConstant(32, MVT::i8));
7535 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7538 return DAG.getMergeValues(Ops, 2, dl);
7541 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7542 SelectionDAG &DAG) const {
7543 EVT SrcVT = Op.getOperand(0).getValueType();
7544 EVT DstVT = Op.getValueType();
7545 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7546 Subtarget->hasMMX() && !DisableMMX) &&
7547 "Unexpected custom BIT_CONVERT");
7548 assert((DstVT == MVT::i64 ||
7549 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7550 "Unexpected custom BIT_CONVERT");
7551 // i64 <=> MMX conversions are Legal.
7552 if (SrcVT==MVT::i64 && DstVT.isVector())
7554 if (DstVT==MVT::i64 && SrcVT.isVector())
7556 // MMX <=> MMX conversions are Legal.
7557 if (SrcVT.isVector() && DstVT.isVector())
7559 // All other conversions need to be expanded.
7562 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7563 SDNode *Node = Op.getNode();
7564 DebugLoc dl = Node->getDebugLoc();
7565 EVT T = Node->getValueType(0);
7566 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7567 DAG.getConstant(0, T), Node->getOperand(2));
7568 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7569 cast<AtomicSDNode>(Node)->getMemoryVT(),
7570 Node->getOperand(0),
7571 Node->getOperand(1), negOp,
7572 cast<AtomicSDNode>(Node)->getSrcValue(),
7573 cast<AtomicSDNode>(Node)->getAlignment());
7576 /// LowerOperation - Provide custom lowering hooks for some operations.
7578 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7579 switch (Op.getOpcode()) {
7580 default: llvm_unreachable("Should not custom lower this!");
7581 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7582 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7583 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7584 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7585 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7586 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7587 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7588 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7589 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7590 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7591 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7592 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7593 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7594 case ISD::SHL_PARTS:
7595 case ISD::SRA_PARTS:
7596 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7597 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7598 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7599 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7600 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7601 case ISD::FABS: return LowerFABS(Op, DAG);
7602 case ISD::FNEG: return LowerFNEG(Op, DAG);
7603 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7604 case ISD::SETCC: return LowerSETCC(Op, DAG);
7605 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7606 case ISD::SELECT: return LowerSELECT(Op, DAG);
7607 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7608 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7609 case ISD::VASTART: return LowerVASTART(Op, DAG);
7610 case ISD::VAARG: return LowerVAARG(Op, DAG);
7611 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7612 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7613 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7614 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7615 case ISD::FRAME_TO_ARGS_OFFSET:
7616 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7617 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7618 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7619 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7620 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7621 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7622 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7623 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7629 case ISD::UMULO: return LowerXALUO(Op, DAG);
7630 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7631 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7635 void X86TargetLowering::
7636 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7637 SelectionDAG &DAG, unsigned NewOp) const {
7638 EVT T = Node->getValueType(0);
7639 DebugLoc dl = Node->getDebugLoc();
7640 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7642 SDValue Chain = Node->getOperand(0);
7643 SDValue In1 = Node->getOperand(1);
7644 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7645 Node->getOperand(2), DAG.getIntPtrConstant(0));
7646 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7647 Node->getOperand(2), DAG.getIntPtrConstant(1));
7648 SDValue Ops[] = { Chain, In1, In2L, In2H };
7649 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7651 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7652 cast<MemSDNode>(Node)->getMemOperand());
7653 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7654 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7655 Results.push_back(Result.getValue(2));
7658 /// ReplaceNodeResults - Replace a node with an illegal result type
7659 /// with a new node built out of custom code.
7660 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7661 SmallVectorImpl<SDValue>&Results,
7662 SelectionDAG &DAG) const {
7663 DebugLoc dl = N->getDebugLoc();
7664 switch (N->getOpcode()) {
7666 assert(false && "Do not know how to custom type legalize this operation!");
7668 case ISD::FP_TO_SINT: {
7669 std::pair<SDValue,SDValue> Vals =
7670 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7671 SDValue FIST = Vals.first, StackSlot = Vals.second;
7672 if (FIST.getNode() != 0) {
7673 EVT VT = N->getValueType(0);
7674 // Return a load from the stack slot.
7675 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7680 case ISD::READCYCLECOUNTER: {
7681 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7682 SDValue TheChain = N->getOperand(0);
7683 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7684 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7686 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7688 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7689 SDValue Ops[] = { eax, edx };
7690 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7691 Results.push_back(edx.getValue(1));
7694 case ISD::ATOMIC_CMP_SWAP: {
7695 EVT T = N->getValueType(0);
7696 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7697 SDValue cpInL, cpInH;
7698 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7699 DAG.getConstant(0, MVT::i32));
7700 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7701 DAG.getConstant(1, MVT::i32));
7702 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7703 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7705 SDValue swapInL, swapInH;
7706 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7707 DAG.getConstant(0, MVT::i32));
7708 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7709 DAG.getConstant(1, MVT::i32));
7710 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7712 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7713 swapInL.getValue(1));
7714 SDValue Ops[] = { swapInH.getValue(0),
7716 swapInH.getValue(1) };
7717 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7718 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7719 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7720 MVT::i32, Result.getValue(1));
7721 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7722 MVT::i32, cpOutL.getValue(2));
7723 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7724 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7725 Results.push_back(cpOutH.getValue(1));
7728 case ISD::ATOMIC_LOAD_ADD:
7729 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7731 case ISD::ATOMIC_LOAD_AND:
7732 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7734 case ISD::ATOMIC_LOAD_NAND:
7735 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7737 case ISD::ATOMIC_LOAD_OR:
7738 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7740 case ISD::ATOMIC_LOAD_SUB:
7741 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7743 case ISD::ATOMIC_LOAD_XOR:
7744 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7746 case ISD::ATOMIC_SWAP:
7747 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7752 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7754 default: return NULL;
7755 case X86ISD::BSF: return "X86ISD::BSF";
7756 case X86ISD::BSR: return "X86ISD::BSR";
7757 case X86ISD::SHLD: return "X86ISD::SHLD";
7758 case X86ISD::SHRD: return "X86ISD::SHRD";
7759 case X86ISD::FAND: return "X86ISD::FAND";
7760 case X86ISD::FOR: return "X86ISD::FOR";
7761 case X86ISD::FXOR: return "X86ISD::FXOR";
7762 case X86ISD::FSRL: return "X86ISD::FSRL";
7763 case X86ISD::FILD: return "X86ISD::FILD";
7764 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7765 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7766 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7767 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7768 case X86ISD::FLD: return "X86ISD::FLD";
7769 case X86ISD::FST: return "X86ISD::FST";
7770 case X86ISD::CALL: return "X86ISD::CALL";
7771 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7772 case X86ISD::BT: return "X86ISD::BT";
7773 case X86ISD::CMP: return "X86ISD::CMP";
7774 case X86ISD::COMI: return "X86ISD::COMI";
7775 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7776 case X86ISD::SETCC: return "X86ISD::SETCC";
7777 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7778 case X86ISD::CMOV: return "X86ISD::CMOV";
7779 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7780 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7781 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7782 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7783 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7784 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7785 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7786 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7787 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7788 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7789 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7790 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7791 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7792 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7793 case X86ISD::FMAX: return "X86ISD::FMAX";
7794 case X86ISD::FMIN: return "X86ISD::FMIN";
7795 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7796 case X86ISD::FRCP: return "X86ISD::FRCP";
7797 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7798 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7799 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7800 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7801 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7802 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7803 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7804 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7805 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7806 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7807 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7808 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7809 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7810 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7811 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7812 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7813 case X86ISD::VSHL: return "X86ISD::VSHL";
7814 case X86ISD::VSRL: return "X86ISD::VSRL";
7815 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7816 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7817 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7818 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7819 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7820 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7821 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7822 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7823 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7824 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7825 case X86ISD::ADD: return "X86ISD::ADD";
7826 case X86ISD::SUB: return "X86ISD::SUB";
7827 case X86ISD::SMUL: return "X86ISD::SMUL";
7828 case X86ISD::UMUL: return "X86ISD::UMUL";
7829 case X86ISD::INC: return "X86ISD::INC";
7830 case X86ISD::DEC: return "X86ISD::DEC";
7831 case X86ISD::OR: return "X86ISD::OR";
7832 case X86ISD::XOR: return "X86ISD::XOR";
7833 case X86ISD::AND: return "X86ISD::AND";
7834 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7835 case X86ISD::PTEST: return "X86ISD::PTEST";
7836 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7837 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7841 // isLegalAddressingMode - Return true if the addressing mode represented
7842 // by AM is legal for this target, for a load/store of the specified type.
7843 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7844 const Type *Ty) const {
7845 // X86 supports extremely general addressing modes.
7846 CodeModel::Model M = getTargetMachine().getCodeModel();
7848 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7849 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7854 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7856 // If a reference to this global requires an extra load, we can't fold it.
7857 if (isGlobalStubReference(GVFlags))
7860 // If BaseGV requires a register for the PIC base, we cannot also have a
7861 // BaseReg specified.
7862 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7865 // If lower 4G is not available, then we must use rip-relative addressing.
7866 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7876 // These scales always work.
7881 // These scales are formed with basereg+scalereg. Only accept if there is
7886 default: // Other stuff never works.
7894 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7895 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7897 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7898 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7899 if (NumBits1 <= NumBits2)
7904 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7905 if (!VT1.isInteger() || !VT2.isInteger())
7907 unsigned NumBits1 = VT1.getSizeInBits();
7908 unsigned NumBits2 = VT2.getSizeInBits();
7909 if (NumBits1 <= NumBits2)
7914 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7915 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7916 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7919 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7920 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7921 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7924 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7925 // i16 instructions are longer (0x66 prefix) and potentially slower.
7926 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7929 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7930 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7931 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7932 /// are assumed to be legal.
7934 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7936 // Very little shuffling can be done for 64-bit vectors right now.
7937 if (VT.getSizeInBits() == 64)
7938 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7940 // FIXME: pshufb, blends, shifts.
7941 return (VT.getVectorNumElements() == 2 ||
7942 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7943 isMOVLMask(M, VT) ||
7944 isSHUFPMask(M, VT) ||
7945 isPSHUFDMask(M, VT) ||
7946 isPSHUFHWMask(M, VT) ||
7947 isPSHUFLWMask(M, VT) ||
7948 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7949 isUNPCKLMask(M, VT) ||
7950 isUNPCKHMask(M, VT) ||
7951 isUNPCKL_v_undef_Mask(M, VT) ||
7952 isUNPCKH_v_undef_Mask(M, VT));
7956 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7958 unsigned NumElts = VT.getVectorNumElements();
7959 // FIXME: This collection of masks seems suspect.
7962 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7963 return (isMOVLMask(Mask, VT) ||
7964 isCommutedMOVLMask(Mask, VT, true) ||
7965 isSHUFPMask(Mask, VT) ||
7966 isCommutedSHUFPMask(Mask, VT));
7971 //===----------------------------------------------------------------------===//
7972 // X86 Scheduler Hooks
7973 //===----------------------------------------------------------------------===//
7975 // private utility function
7977 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7978 MachineBasicBlock *MBB,
7986 TargetRegisterClass *RC,
7987 bool invSrc) const {
7988 // For the atomic bitwise operator, we generate
7991 // ld t1 = [bitinstr.addr]
7992 // op t2 = t1, [bitinstr.val]
7994 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7996 // fallthrough -->nextMBB
7997 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7998 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7999 MachineFunction::iterator MBBIter = MBB;
8002 /// First build the CFG
8003 MachineFunction *F = MBB->getParent();
8004 MachineBasicBlock *thisMBB = MBB;
8005 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8006 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8007 F->insert(MBBIter, newMBB);
8008 F->insert(MBBIter, nextMBB);
8010 // Move all successors to thisMBB to nextMBB
8011 nextMBB->transferSuccessors(thisMBB);
8013 // Update thisMBB to fall through to newMBB
8014 thisMBB->addSuccessor(newMBB);
8016 // newMBB jumps to itself and fall through to nextMBB
8017 newMBB->addSuccessor(nextMBB);
8018 newMBB->addSuccessor(newMBB);
8020 // Insert instructions into newMBB based on incoming instruction
8021 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8022 "unexpected number of operands");
8023 DebugLoc dl = bInstr->getDebugLoc();
8024 MachineOperand& destOper = bInstr->getOperand(0);
8025 MachineOperand* argOpers[2 + X86AddrNumOperands];
8026 int numArgs = bInstr->getNumOperands() - 1;
8027 for (int i=0; i < numArgs; ++i)
8028 argOpers[i] = &bInstr->getOperand(i+1);
8030 // x86 address has 4 operands: base, index, scale, and displacement
8031 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8032 int valArgIndx = lastAddrIndx + 1;
8034 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8035 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8036 for (int i=0; i <= lastAddrIndx; ++i)
8037 (*MIB).addOperand(*argOpers[i]);
8039 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8041 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8046 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8047 assert((argOpers[valArgIndx]->isReg() ||
8048 argOpers[valArgIndx]->isImm()) &&
8050 if (argOpers[valArgIndx]->isReg())
8051 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8053 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8055 (*MIB).addOperand(*argOpers[valArgIndx]);
8057 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8060 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8061 for (int i=0; i <= lastAddrIndx; ++i)
8062 (*MIB).addOperand(*argOpers[i]);
8064 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8065 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8066 bInstr->memoperands_end());
8068 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8072 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8074 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8078 // private utility function: 64 bit atomics on 32 bit host.
8080 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8081 MachineBasicBlock *MBB,
8086 bool invSrc) const {
8087 // For the atomic bitwise operator, we generate
8088 // thisMBB (instructions are in pairs, except cmpxchg8b)
8089 // ld t1,t2 = [bitinstr.addr]
8091 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8092 // op t5, t6 <- out1, out2, [bitinstr.val]
8093 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8094 // mov ECX, EBX <- t5, t6
8095 // mov EAX, EDX <- t1, t2
8096 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8097 // mov t3, t4 <- EAX, EDX
8099 // result in out1, out2
8100 // fallthrough -->nextMBB
8102 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8103 const unsigned LoadOpc = X86::MOV32rm;
8104 const unsigned copyOpc = X86::MOV32rr;
8105 const unsigned NotOpc = X86::NOT32r;
8106 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8107 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8108 MachineFunction::iterator MBBIter = MBB;
8111 /// First build the CFG
8112 MachineFunction *F = MBB->getParent();
8113 MachineBasicBlock *thisMBB = MBB;
8114 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8115 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8116 F->insert(MBBIter, newMBB);
8117 F->insert(MBBIter, nextMBB);
8119 // Move all successors to thisMBB to nextMBB
8120 nextMBB->transferSuccessors(thisMBB);
8122 // Update thisMBB to fall through to newMBB
8123 thisMBB->addSuccessor(newMBB);
8125 // newMBB jumps to itself and fall through to nextMBB
8126 newMBB->addSuccessor(nextMBB);
8127 newMBB->addSuccessor(newMBB);
8129 DebugLoc dl = bInstr->getDebugLoc();
8130 // Insert instructions into newMBB based on incoming instruction
8131 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8132 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8133 "unexpected number of operands");
8134 MachineOperand& dest1Oper = bInstr->getOperand(0);
8135 MachineOperand& dest2Oper = bInstr->getOperand(1);
8136 MachineOperand* argOpers[2 + X86AddrNumOperands];
8137 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8138 argOpers[i] = &bInstr->getOperand(i+2);
8140 // We use some of the operands multiple times, so conservatively just
8141 // clear any kill flags that might be present.
8142 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8143 argOpers[i]->setIsKill(false);
8146 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8147 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8149 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8150 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8151 for (int i=0; i <= lastAddrIndx; ++i)
8152 (*MIB).addOperand(*argOpers[i]);
8153 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8154 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8155 // add 4 to displacement.
8156 for (int i=0; i <= lastAddrIndx-2; ++i)
8157 (*MIB).addOperand(*argOpers[i]);
8158 MachineOperand newOp3 = *(argOpers[3]);
8160 newOp3.setImm(newOp3.getImm()+4);
8162 newOp3.setOffset(newOp3.getOffset()+4);
8163 (*MIB).addOperand(newOp3);
8164 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8166 // t3/4 are defined later, at the bottom of the loop
8167 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8168 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8169 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8170 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8171 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8172 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8174 // The subsequent operations should be using the destination registers of
8175 //the PHI instructions.
8177 t1 = F->getRegInfo().createVirtualRegister(RC);
8178 t2 = F->getRegInfo().createVirtualRegister(RC);
8179 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8180 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8182 t1 = dest1Oper.getReg();
8183 t2 = dest2Oper.getReg();
8186 int valArgIndx = lastAddrIndx + 1;
8187 assert((argOpers[valArgIndx]->isReg() ||
8188 argOpers[valArgIndx]->isImm()) &&
8190 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8191 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8192 if (argOpers[valArgIndx]->isReg())
8193 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8195 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8196 if (regOpcL != X86::MOV32rr)
8198 (*MIB).addOperand(*argOpers[valArgIndx]);
8199 assert(argOpers[valArgIndx + 1]->isReg() ==
8200 argOpers[valArgIndx]->isReg());
8201 assert(argOpers[valArgIndx + 1]->isImm() ==
8202 argOpers[valArgIndx]->isImm());
8203 if (argOpers[valArgIndx + 1]->isReg())
8204 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8206 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8207 if (regOpcH != X86::MOV32rr)
8209 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8211 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8213 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8216 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8218 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8221 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8222 for (int i=0; i <= lastAddrIndx; ++i)
8223 (*MIB).addOperand(*argOpers[i]);
8225 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8226 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8227 bInstr->memoperands_end());
8229 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8230 MIB.addReg(X86::EAX);
8231 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8232 MIB.addReg(X86::EDX);
8235 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8237 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8241 // private utility function
8243 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8244 MachineBasicBlock *MBB,
8245 unsigned cmovOpc) const {
8246 // For the atomic min/max operator, we generate
8249 // ld t1 = [min/max.addr]
8250 // mov t2 = [min/max.val]
8252 // cmov[cond] t2 = t1
8254 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8256 // fallthrough -->nextMBB
8258 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8259 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8260 MachineFunction::iterator MBBIter = MBB;
8263 /// First build the CFG
8264 MachineFunction *F = MBB->getParent();
8265 MachineBasicBlock *thisMBB = MBB;
8266 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8267 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8268 F->insert(MBBIter, newMBB);
8269 F->insert(MBBIter, nextMBB);
8271 // Move all successors of thisMBB to nextMBB
8272 nextMBB->transferSuccessors(thisMBB);
8274 // Update thisMBB to fall through to newMBB
8275 thisMBB->addSuccessor(newMBB);
8277 // newMBB jumps to newMBB and fall through to nextMBB
8278 newMBB->addSuccessor(nextMBB);
8279 newMBB->addSuccessor(newMBB);
8281 DebugLoc dl = mInstr->getDebugLoc();
8282 // Insert instructions into newMBB based on incoming instruction
8283 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8284 "unexpected number of operands");
8285 MachineOperand& destOper = mInstr->getOperand(0);
8286 MachineOperand* argOpers[2 + X86AddrNumOperands];
8287 int numArgs = mInstr->getNumOperands() - 1;
8288 for (int i=0; i < numArgs; ++i)
8289 argOpers[i] = &mInstr->getOperand(i+1);
8291 // x86 address has 4 operands: base, index, scale, and displacement
8292 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8293 int valArgIndx = lastAddrIndx + 1;
8295 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8296 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8297 for (int i=0; i <= lastAddrIndx; ++i)
8298 (*MIB).addOperand(*argOpers[i]);
8300 // We only support register and immediate values
8301 assert((argOpers[valArgIndx]->isReg() ||
8302 argOpers[valArgIndx]->isImm()) &&
8305 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8306 if (argOpers[valArgIndx]->isReg())
8307 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8309 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8310 (*MIB).addOperand(*argOpers[valArgIndx]);
8312 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8315 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8320 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8321 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8325 // Cmp and exchange if none has modified the memory location
8326 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8327 for (int i=0; i <= lastAddrIndx; ++i)
8328 (*MIB).addOperand(*argOpers[i]);
8330 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8331 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8332 mInstr->memoperands_end());
8334 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8335 MIB.addReg(X86::EAX);
8338 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8340 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8344 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8345 // all of this code can be replaced with that in the .td file.
8347 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8348 unsigned numArgs, bool memArg) const {
8350 MachineFunction *F = BB->getParent();
8351 DebugLoc dl = MI->getDebugLoc();
8352 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8356 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8358 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8360 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8362 for (unsigned i = 0; i < numArgs; ++i) {
8363 MachineOperand &Op = MI->getOperand(i+1);
8365 if (!(Op.isReg() && Op.isImplicit()))
8369 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8372 F->DeleteMachineInstr(MI);
8378 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8380 MachineBasicBlock *MBB) const {
8381 // Emit code to save XMM registers to the stack. The ABI says that the
8382 // number of registers to save is given in %al, so it's theoretically
8383 // possible to do an indirect jump trick to avoid saving all of them,
8384 // however this code takes a simpler approach and just executes all
8385 // of the stores if %al is non-zero. It's less code, and it's probably
8386 // easier on the hardware branch predictor, and stores aren't all that
8387 // expensive anyway.
8389 // Create the new basic blocks. One block contains all the XMM stores,
8390 // and one block is the final destination regardless of whether any
8391 // stores were performed.
8392 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8393 MachineFunction *F = MBB->getParent();
8394 MachineFunction::iterator MBBIter = MBB;
8396 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8397 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8398 F->insert(MBBIter, XMMSaveMBB);
8399 F->insert(MBBIter, EndMBB);
8402 // Move any original successors of MBB to the end block.
8403 EndMBB->transferSuccessors(MBB);
8404 // The original block will now fall through to the XMM save block.
8405 MBB->addSuccessor(XMMSaveMBB);
8406 // The XMMSaveMBB will fall through to the end block.
8407 XMMSaveMBB->addSuccessor(EndMBB);
8409 // Now add the instructions.
8410 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8411 DebugLoc DL = MI->getDebugLoc();
8413 unsigned CountReg = MI->getOperand(0).getReg();
8414 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8415 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8417 if (!Subtarget->isTargetWin64()) {
8418 // If %al is 0, branch around the XMM save block.
8419 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8420 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8421 MBB->addSuccessor(EndMBB);
8424 // In the XMM save block, save all the XMM argument registers.
8425 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8426 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8427 MachineMemOperand *MMO =
8428 F->getMachineMemOperand(
8429 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8430 MachineMemOperand::MOStore, Offset,
8431 /*Size=*/16, /*Align=*/16);
8432 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8433 .addFrameIndex(RegSaveFrameIndex)
8434 .addImm(/*Scale=*/1)
8435 .addReg(/*IndexReg=*/0)
8436 .addImm(/*Disp=*/Offset)
8437 .addReg(/*Segment=*/0)
8438 .addReg(MI->getOperand(i).getReg())
8439 .addMemOperand(MMO);
8442 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8448 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8449 MachineBasicBlock *BB) const {
8450 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8451 DebugLoc DL = MI->getDebugLoc();
8453 // To "insert" a SELECT_CC instruction, we actually have to insert the
8454 // diamond control-flow pattern. The incoming instruction knows the
8455 // destination vreg to set, the condition code register to branch on, the
8456 // true/false values to select between, and a branch opcode to use.
8457 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8458 MachineFunction::iterator It = BB;
8464 // cmpTY ccX, r1, r2
8466 // fallthrough --> copy0MBB
8467 MachineBasicBlock *thisMBB = BB;
8468 MachineFunction *F = BB->getParent();
8469 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8470 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8472 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8473 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8474 F->insert(It, copy0MBB);
8475 F->insert(It, sinkMBB);
8476 // Update machine-CFG edges by first adding all successors of the current
8477 // block to the new block which will contain the Phi node for the select.
8478 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8479 E = BB->succ_end(); I != E; ++I)
8480 sinkMBB->addSuccessor(*I);
8481 // Next, remove all successors of the current block, and add the true
8482 // and fallthrough blocks as its successors.
8483 while (!BB->succ_empty())
8484 BB->removeSuccessor(BB->succ_begin());
8485 // Add the true and fallthrough blocks as its successors.
8486 BB->addSuccessor(copy0MBB);
8487 BB->addSuccessor(sinkMBB);
8490 // %FalseValue = ...
8491 // # fallthrough to sinkMBB
8492 copy0MBB->addSuccessor(sinkMBB);
8495 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8497 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8498 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8499 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8501 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8506 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8507 MachineBasicBlock *BB) const {
8508 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8509 DebugLoc DL = MI->getDebugLoc();
8510 MachineFunction *F = BB->getParent();
8512 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8513 // non-trivial part is impdef of ESP.
8514 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8517 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8518 .addExternalSymbol("_alloca")
8519 .addReg(X86::EAX, RegState::Implicit)
8520 .addReg(X86::ESP, RegState::Implicit)
8521 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8522 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8524 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8529 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8530 MachineBasicBlock *BB) const {
8531 // This is pretty easy. We're taking the value that we received from
8532 // our load from the relocation, sticking it in either RDI (x86-64)
8533 // or EAX and doing an indirect call. The return value will then
8534 // be in the normal return register.
8535 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8536 DebugLoc DL = MI->getDebugLoc();
8537 MachineFunction *F = BB->getParent();
8539 if (Subtarget->is64Bit()) {
8540 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rr), X86::RDI)
8541 .addReg(MI->getOperand(0).getReg());
8542 MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
8543 addDirectMem(MIB, X86::RDI).addReg(0);
8545 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rr), X86::EAX)
8546 .addReg(MI->getOperand(0).getReg());
8547 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8548 addDirectMem(MIB, X86::EAX).addReg(0);
8551 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8556 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8557 MachineBasicBlock *BB) const {
8558 switch (MI->getOpcode()) {
8559 default: assert(false && "Unexpected instr type to insert");
8560 case X86::MINGW_ALLOCA:
8561 return EmitLoweredMingwAlloca(MI, BB);
8562 case X86::TLSCall_32:
8563 case X86::TLSCall_64:
8564 return EmitLoweredTLSCall(MI, BB);
8566 case X86::CMOV_V1I64:
8567 case X86::CMOV_FR32:
8568 case X86::CMOV_FR64:
8569 case X86::CMOV_V4F32:
8570 case X86::CMOV_V2F64:
8571 case X86::CMOV_V2I64:
8572 case X86::CMOV_GR16:
8573 case X86::CMOV_GR32:
8574 case X86::CMOV_RFP32:
8575 case X86::CMOV_RFP64:
8576 case X86::CMOV_RFP80:
8577 return EmitLoweredSelect(MI, BB);
8579 case X86::FP32_TO_INT16_IN_MEM:
8580 case X86::FP32_TO_INT32_IN_MEM:
8581 case X86::FP32_TO_INT64_IN_MEM:
8582 case X86::FP64_TO_INT16_IN_MEM:
8583 case X86::FP64_TO_INT32_IN_MEM:
8584 case X86::FP64_TO_INT64_IN_MEM:
8585 case X86::FP80_TO_INT16_IN_MEM:
8586 case X86::FP80_TO_INT32_IN_MEM:
8587 case X86::FP80_TO_INT64_IN_MEM: {
8588 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8589 DebugLoc DL = MI->getDebugLoc();
8591 // Change the floating point control register to use "round towards zero"
8592 // mode when truncating to an integer value.
8593 MachineFunction *F = BB->getParent();
8594 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8595 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8597 // Load the old value of the high byte of the control word...
8599 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8600 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8603 // Set the high part to be round to zero...
8604 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8607 // Reload the modified control word now...
8608 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8610 // Restore the memory image of control word to original value
8611 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8614 // Get the X86 opcode to use.
8616 switch (MI->getOpcode()) {
8617 default: llvm_unreachable("illegal opcode!");
8618 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8619 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8620 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8621 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8622 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8623 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8624 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8625 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8626 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8630 MachineOperand &Op = MI->getOperand(0);
8632 AM.BaseType = X86AddressMode::RegBase;
8633 AM.Base.Reg = Op.getReg();
8635 AM.BaseType = X86AddressMode::FrameIndexBase;
8636 AM.Base.FrameIndex = Op.getIndex();
8638 Op = MI->getOperand(1);
8640 AM.Scale = Op.getImm();
8641 Op = MI->getOperand(2);
8643 AM.IndexReg = Op.getImm();
8644 Op = MI->getOperand(3);
8645 if (Op.isGlobal()) {
8646 AM.GV = Op.getGlobal();
8648 AM.Disp = Op.getImm();
8650 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8651 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8653 // Reload the original control word now.
8654 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8656 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8659 // String/text processing lowering.
8660 case X86::PCMPISTRM128REG:
8661 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8662 case X86::PCMPISTRM128MEM:
8663 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8664 case X86::PCMPESTRM128REG:
8665 return EmitPCMP(MI, BB, 5, false /* in mem */);
8666 case X86::PCMPESTRM128MEM:
8667 return EmitPCMP(MI, BB, 5, true /* in mem */);
8670 case X86::ATOMAND32:
8671 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8672 X86::AND32ri, X86::MOV32rm,
8673 X86::LCMPXCHG32, X86::MOV32rr,
8674 X86::NOT32r, X86::EAX,
8675 X86::GR32RegisterClass);
8677 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8678 X86::OR32ri, X86::MOV32rm,
8679 X86::LCMPXCHG32, X86::MOV32rr,
8680 X86::NOT32r, X86::EAX,
8681 X86::GR32RegisterClass);
8682 case X86::ATOMXOR32:
8683 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8684 X86::XOR32ri, X86::MOV32rm,
8685 X86::LCMPXCHG32, X86::MOV32rr,
8686 X86::NOT32r, X86::EAX,
8687 X86::GR32RegisterClass);
8688 case X86::ATOMNAND32:
8689 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8690 X86::AND32ri, X86::MOV32rm,
8691 X86::LCMPXCHG32, X86::MOV32rr,
8692 X86::NOT32r, X86::EAX,
8693 X86::GR32RegisterClass, true);
8694 case X86::ATOMMIN32:
8695 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8696 case X86::ATOMMAX32:
8697 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8698 case X86::ATOMUMIN32:
8699 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8700 case X86::ATOMUMAX32:
8701 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8703 case X86::ATOMAND16:
8704 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8705 X86::AND16ri, X86::MOV16rm,
8706 X86::LCMPXCHG16, X86::MOV16rr,
8707 X86::NOT16r, X86::AX,
8708 X86::GR16RegisterClass);
8710 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8711 X86::OR16ri, X86::MOV16rm,
8712 X86::LCMPXCHG16, X86::MOV16rr,
8713 X86::NOT16r, X86::AX,
8714 X86::GR16RegisterClass);
8715 case X86::ATOMXOR16:
8716 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8717 X86::XOR16ri, X86::MOV16rm,
8718 X86::LCMPXCHG16, X86::MOV16rr,
8719 X86::NOT16r, X86::AX,
8720 X86::GR16RegisterClass);
8721 case X86::ATOMNAND16:
8722 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8723 X86::AND16ri, X86::MOV16rm,
8724 X86::LCMPXCHG16, X86::MOV16rr,
8725 X86::NOT16r, X86::AX,
8726 X86::GR16RegisterClass, true);
8727 case X86::ATOMMIN16:
8728 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8729 case X86::ATOMMAX16:
8730 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8731 case X86::ATOMUMIN16:
8732 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8733 case X86::ATOMUMAX16:
8734 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8737 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8738 X86::AND8ri, X86::MOV8rm,
8739 X86::LCMPXCHG8, X86::MOV8rr,
8740 X86::NOT8r, X86::AL,
8741 X86::GR8RegisterClass);
8743 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8744 X86::OR8ri, X86::MOV8rm,
8745 X86::LCMPXCHG8, X86::MOV8rr,
8746 X86::NOT8r, X86::AL,
8747 X86::GR8RegisterClass);
8749 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8750 X86::XOR8ri, X86::MOV8rm,
8751 X86::LCMPXCHG8, X86::MOV8rr,
8752 X86::NOT8r, X86::AL,
8753 X86::GR8RegisterClass);
8754 case X86::ATOMNAND8:
8755 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8756 X86::AND8ri, X86::MOV8rm,
8757 X86::LCMPXCHG8, X86::MOV8rr,
8758 X86::NOT8r, X86::AL,
8759 X86::GR8RegisterClass, true);
8760 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8761 // This group is for 64-bit host.
8762 case X86::ATOMAND64:
8763 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8764 X86::AND64ri32, X86::MOV64rm,
8765 X86::LCMPXCHG64, X86::MOV64rr,
8766 X86::NOT64r, X86::RAX,
8767 X86::GR64RegisterClass);
8769 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8770 X86::OR64ri32, X86::MOV64rm,
8771 X86::LCMPXCHG64, X86::MOV64rr,
8772 X86::NOT64r, X86::RAX,
8773 X86::GR64RegisterClass);
8774 case X86::ATOMXOR64:
8775 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8776 X86::XOR64ri32, X86::MOV64rm,
8777 X86::LCMPXCHG64, X86::MOV64rr,
8778 X86::NOT64r, X86::RAX,
8779 X86::GR64RegisterClass);
8780 case X86::ATOMNAND64:
8781 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8782 X86::AND64ri32, X86::MOV64rm,
8783 X86::LCMPXCHG64, X86::MOV64rr,
8784 X86::NOT64r, X86::RAX,
8785 X86::GR64RegisterClass, true);
8786 case X86::ATOMMIN64:
8787 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8788 case X86::ATOMMAX64:
8789 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8790 case X86::ATOMUMIN64:
8791 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8792 case X86::ATOMUMAX64:
8793 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8795 // This group does 64-bit operations on a 32-bit host.
8796 case X86::ATOMAND6432:
8797 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8798 X86::AND32rr, X86::AND32rr,
8799 X86::AND32ri, X86::AND32ri,
8801 case X86::ATOMOR6432:
8802 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8803 X86::OR32rr, X86::OR32rr,
8804 X86::OR32ri, X86::OR32ri,
8806 case X86::ATOMXOR6432:
8807 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8808 X86::XOR32rr, X86::XOR32rr,
8809 X86::XOR32ri, X86::XOR32ri,
8811 case X86::ATOMNAND6432:
8812 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8813 X86::AND32rr, X86::AND32rr,
8814 X86::AND32ri, X86::AND32ri,
8816 case X86::ATOMADD6432:
8817 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8818 X86::ADD32rr, X86::ADC32rr,
8819 X86::ADD32ri, X86::ADC32ri,
8821 case X86::ATOMSUB6432:
8822 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8823 X86::SUB32rr, X86::SBB32rr,
8824 X86::SUB32ri, X86::SBB32ri,
8826 case X86::ATOMSWAP6432:
8827 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8828 X86::MOV32rr, X86::MOV32rr,
8829 X86::MOV32ri, X86::MOV32ri,
8831 case X86::VASTART_SAVE_XMM_REGS:
8832 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8836 //===----------------------------------------------------------------------===//
8837 // X86 Optimization Hooks
8838 //===----------------------------------------------------------------------===//
8840 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8844 const SelectionDAG &DAG,
8845 unsigned Depth) const {
8846 unsigned Opc = Op.getOpcode();
8847 assert((Opc >= ISD::BUILTIN_OP_END ||
8848 Opc == ISD::INTRINSIC_WO_CHAIN ||
8849 Opc == ISD::INTRINSIC_W_CHAIN ||
8850 Opc == ISD::INTRINSIC_VOID) &&
8851 "Should use MaskedValueIsZero if you don't know whether Op"
8852 " is a target node!");
8854 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8866 // These nodes' second result is a boolean.
8867 if (Op.getResNo() == 0)
8871 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8872 Mask.getBitWidth() - 1);
8877 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8878 /// node is a GlobalAddress + offset.
8879 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8880 const GlobalValue* &GA,
8881 int64_t &Offset) const {
8882 if (N->getOpcode() == X86ISD::Wrapper) {
8883 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8884 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8885 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8889 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8892 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8893 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8894 /// if the load addresses are consecutive, non-overlapping, and in the right
8896 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8897 const TargetLowering &TLI) {
8898 DebugLoc dl = N->getDebugLoc();
8899 EVT VT = N->getValueType(0);
8900 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8902 if (VT.getSizeInBits() != 128)
8905 SmallVector<SDValue, 16> Elts;
8906 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8907 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8909 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8912 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8913 /// and convert it from being a bunch of shuffles and extracts to a simple
8914 /// store and scalar loads to extract the elements.
8915 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8916 const TargetLowering &TLI) {
8917 SDValue InputVector = N->getOperand(0);
8919 // Only operate on vectors of 4 elements, where the alternative shuffling
8920 // gets to be more expensive.
8921 if (InputVector.getValueType() != MVT::v4i32)
8924 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8925 // single use which is a sign-extend or zero-extend, and all elements are
8927 SmallVector<SDNode *, 4> Uses;
8928 unsigned ExtractedElements = 0;
8929 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8930 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8931 if (UI.getUse().getResNo() != InputVector.getResNo())
8934 SDNode *Extract = *UI;
8935 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8938 if (Extract->getValueType(0) != MVT::i32)
8940 if (!Extract->hasOneUse())
8942 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8943 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8945 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8948 // Record which element was extracted.
8949 ExtractedElements |=
8950 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
8952 Uses.push_back(Extract);
8955 // If not all the elements were used, this may not be worthwhile.
8956 if (ExtractedElements != 15)
8959 // Ok, we've now decided to do the transformation.
8960 DebugLoc dl = InputVector.getDebugLoc();
8962 // Store the value to a temporary stack slot.
8963 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
8964 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
8967 // Replace each use (extract) with a load of the appropriate element.
8968 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
8969 UE = Uses.end(); UI != UE; ++UI) {
8970 SDNode *Extract = *UI;
8972 // Compute the element's address.
8973 SDValue Idx = Extract->getOperand(1);
8975 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
8976 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
8977 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
8979 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
8982 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
8983 NULL, 0, false, false, 0);
8985 // Replace the exact with the load.
8986 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
8989 // The replacement was made in place; don't return anything.
8993 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8994 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8995 const X86Subtarget *Subtarget) {
8996 DebugLoc DL = N->getDebugLoc();
8997 SDValue Cond = N->getOperand(0);
8998 // Get the LHS/RHS of the select.
8999 SDValue LHS = N->getOperand(1);
9000 SDValue RHS = N->getOperand(2);
9002 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9003 // instructions match the semantics of the common C idiom x<y?x:y but not
9004 // x<=y?x:y, because of how they handle negative zero (which can be
9005 // ignored in unsafe-math mode).
9006 if (Subtarget->hasSSE2() &&
9007 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9008 Cond.getOpcode() == ISD::SETCC) {
9009 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9011 unsigned Opcode = 0;
9012 // Check for x CC y ? x : y.
9013 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9014 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9018 // Converting this to a min would handle NaNs incorrectly, and swapping
9019 // the operands would cause it to handle comparisons between positive
9020 // and negative zero incorrectly.
9021 if (!FiniteOnlyFPMath() &&
9022 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9023 if (!UnsafeFPMath &&
9024 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9026 std::swap(LHS, RHS);
9028 Opcode = X86ISD::FMIN;
9031 // Converting this to a min would handle comparisons between positive
9032 // and negative zero incorrectly.
9033 if (!UnsafeFPMath &&
9034 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9036 Opcode = X86ISD::FMIN;
9039 // Converting this to a min would handle both negative zeros and NaNs
9040 // incorrectly, but we can swap the operands to fix both.
9041 std::swap(LHS, RHS);
9045 Opcode = X86ISD::FMIN;
9049 // Converting this to a max would handle comparisons between positive
9050 // and negative zero incorrectly.
9051 if (!UnsafeFPMath &&
9052 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9054 Opcode = X86ISD::FMAX;
9057 // Converting this to a max would handle NaNs incorrectly, and swapping
9058 // the operands would cause it to handle comparisons between positive
9059 // and negative zero incorrectly.
9060 if (!FiniteOnlyFPMath() &&
9061 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9062 if (!UnsafeFPMath &&
9063 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9065 std::swap(LHS, RHS);
9067 Opcode = X86ISD::FMAX;
9070 // Converting this to a max would handle both negative zeros and NaNs
9071 // incorrectly, but we can swap the operands to fix both.
9072 std::swap(LHS, RHS);
9076 Opcode = X86ISD::FMAX;
9079 // Check for x CC y ? y : x -- a min/max with reversed arms.
9080 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9081 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9085 // Converting this to a min would handle comparisons between positive
9086 // and negative zero incorrectly, and swapping the operands would
9087 // cause it to handle NaNs incorrectly.
9088 if (!UnsafeFPMath &&
9089 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9090 if (!FiniteOnlyFPMath() &&
9091 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9093 std::swap(LHS, RHS);
9095 Opcode = X86ISD::FMIN;
9098 // Converting this to a min would handle NaNs incorrectly.
9099 if (!UnsafeFPMath &&
9100 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9102 Opcode = X86ISD::FMIN;
9105 // Converting this to a min would handle both negative zeros and NaNs
9106 // incorrectly, but we can swap the operands to fix both.
9107 std::swap(LHS, RHS);
9111 Opcode = X86ISD::FMIN;
9115 // Converting this to a max would handle NaNs incorrectly.
9116 if (!FiniteOnlyFPMath() &&
9117 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9119 Opcode = X86ISD::FMAX;
9122 // Converting this to a max would handle comparisons between positive
9123 // and negative zero incorrectly, and swapping the operands would
9124 // cause it to handle NaNs incorrectly.
9125 if (!UnsafeFPMath &&
9126 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9127 if (!FiniteOnlyFPMath() &&
9128 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9130 std::swap(LHS, RHS);
9132 Opcode = X86ISD::FMAX;
9135 // Converting this to a max would handle both negative zeros and NaNs
9136 // incorrectly, but we can swap the operands to fix both.
9137 std::swap(LHS, RHS);
9141 Opcode = X86ISD::FMAX;
9147 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9150 // If this is a select between two integer constants, try to do some
9152 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9153 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9154 // Don't do this for crazy integer types.
9155 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9156 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9157 // so that TrueC (the true value) is larger than FalseC.
9158 bool NeedsCondInvert = false;
9160 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9161 // Efficiently invertible.
9162 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9163 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9164 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9165 NeedsCondInvert = true;
9166 std::swap(TrueC, FalseC);
9169 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9170 if (FalseC->getAPIntValue() == 0 &&
9171 TrueC->getAPIntValue().isPowerOf2()) {
9172 if (NeedsCondInvert) // Invert the condition if needed.
9173 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9174 DAG.getConstant(1, Cond.getValueType()));
9176 // Zero extend the condition if needed.
9177 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9179 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9180 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9181 DAG.getConstant(ShAmt, MVT::i8));
9184 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9185 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9186 if (NeedsCondInvert) // Invert the condition if needed.
9187 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9188 DAG.getConstant(1, Cond.getValueType()));
9190 // Zero extend the condition if needed.
9191 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9192 FalseC->getValueType(0), Cond);
9193 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9194 SDValue(FalseC, 0));
9197 // Optimize cases that will turn into an LEA instruction. This requires
9198 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9199 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9200 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9201 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9203 bool isFastMultiplier = false;
9205 switch ((unsigned char)Diff) {
9207 case 1: // result = add base, cond
9208 case 2: // result = lea base( , cond*2)
9209 case 3: // result = lea base(cond, cond*2)
9210 case 4: // result = lea base( , cond*4)
9211 case 5: // result = lea base(cond, cond*4)
9212 case 8: // result = lea base( , cond*8)
9213 case 9: // result = lea base(cond, cond*8)
9214 isFastMultiplier = true;
9219 if (isFastMultiplier) {
9220 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9221 if (NeedsCondInvert) // Invert the condition if needed.
9222 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9223 DAG.getConstant(1, Cond.getValueType()));
9225 // Zero extend the condition if needed.
9226 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9228 // Scale the condition by the difference.
9230 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9231 DAG.getConstant(Diff, Cond.getValueType()));
9233 // Add the base if non-zero.
9234 if (FalseC->getAPIntValue() != 0)
9235 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9236 SDValue(FalseC, 0));
9246 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9247 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9248 TargetLowering::DAGCombinerInfo &DCI) {
9249 DebugLoc DL = N->getDebugLoc();
9251 // If the flag operand isn't dead, don't touch this CMOV.
9252 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9255 // If this is a select between two integer constants, try to do some
9256 // optimizations. Note that the operands are ordered the opposite of SELECT
9258 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9259 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9260 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9261 // larger than FalseC (the false value).
9262 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9264 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9265 CC = X86::GetOppositeBranchCondition(CC);
9266 std::swap(TrueC, FalseC);
9269 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9270 // This is efficient for any integer data type (including i8/i16) and
9272 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9273 SDValue Cond = N->getOperand(3);
9274 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9275 DAG.getConstant(CC, MVT::i8), Cond);
9277 // Zero extend the condition if needed.
9278 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9280 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9281 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9282 DAG.getConstant(ShAmt, MVT::i8));
9283 if (N->getNumValues() == 2) // Dead flag value?
9284 return DCI.CombineTo(N, Cond, SDValue());
9288 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9289 // for any integer data type, including i8/i16.
9290 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9291 SDValue Cond = N->getOperand(3);
9292 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9293 DAG.getConstant(CC, MVT::i8), Cond);
9295 // Zero extend the condition if needed.
9296 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9297 FalseC->getValueType(0), Cond);
9298 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9299 SDValue(FalseC, 0));
9301 if (N->getNumValues() == 2) // Dead flag value?
9302 return DCI.CombineTo(N, Cond, SDValue());
9306 // Optimize cases that will turn into an LEA instruction. This requires
9307 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9308 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9309 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9310 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9312 bool isFastMultiplier = false;
9314 switch ((unsigned char)Diff) {
9316 case 1: // result = add base, cond
9317 case 2: // result = lea base( , cond*2)
9318 case 3: // result = lea base(cond, cond*2)
9319 case 4: // result = lea base( , cond*4)
9320 case 5: // result = lea base(cond, cond*4)
9321 case 8: // result = lea base( , cond*8)
9322 case 9: // result = lea base(cond, cond*8)
9323 isFastMultiplier = true;
9328 if (isFastMultiplier) {
9329 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9330 SDValue Cond = N->getOperand(3);
9331 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9332 DAG.getConstant(CC, MVT::i8), Cond);
9333 // Zero extend the condition if needed.
9334 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9336 // Scale the condition by the difference.
9338 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9339 DAG.getConstant(Diff, Cond.getValueType()));
9341 // Add the base if non-zero.
9342 if (FalseC->getAPIntValue() != 0)
9343 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9344 SDValue(FalseC, 0));
9345 if (N->getNumValues() == 2) // Dead flag value?
9346 return DCI.CombineTo(N, Cond, SDValue());
9356 /// PerformMulCombine - Optimize a single multiply with constant into two
9357 /// in order to implement it with two cheaper instructions, e.g.
9358 /// LEA + SHL, LEA + LEA.
9359 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9360 TargetLowering::DAGCombinerInfo &DCI) {
9361 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9364 EVT VT = N->getValueType(0);
9368 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9371 uint64_t MulAmt = C->getZExtValue();
9372 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9375 uint64_t MulAmt1 = 0;
9376 uint64_t MulAmt2 = 0;
9377 if ((MulAmt % 9) == 0) {
9379 MulAmt2 = MulAmt / 9;
9380 } else if ((MulAmt % 5) == 0) {
9382 MulAmt2 = MulAmt / 5;
9383 } else if ((MulAmt % 3) == 0) {
9385 MulAmt2 = MulAmt / 3;
9388 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9389 DebugLoc DL = N->getDebugLoc();
9391 if (isPowerOf2_64(MulAmt2) &&
9392 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9393 // If second multiplifer is pow2, issue it first. We want the multiply by
9394 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9396 std::swap(MulAmt1, MulAmt2);
9399 if (isPowerOf2_64(MulAmt1))
9400 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9401 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9403 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9404 DAG.getConstant(MulAmt1, VT));
9406 if (isPowerOf2_64(MulAmt2))
9407 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9408 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9410 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9411 DAG.getConstant(MulAmt2, VT));
9413 // Do not add new nodes to DAG combiner worklist.
9414 DCI.CombineTo(N, NewMul, false);
9419 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9420 SDValue N0 = N->getOperand(0);
9421 SDValue N1 = N->getOperand(1);
9422 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9423 EVT VT = N0.getValueType();
9425 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9426 // since the result of setcc_c is all zero's or all ones.
9427 if (N1C && N0.getOpcode() == ISD::AND &&
9428 N0.getOperand(1).getOpcode() == ISD::Constant) {
9429 SDValue N00 = N0.getOperand(0);
9430 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9431 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9432 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9433 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9434 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9435 APInt ShAmt = N1C->getAPIntValue();
9436 Mask = Mask.shl(ShAmt);
9438 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9439 N00, DAG.getConstant(Mask, VT));
9446 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9448 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9449 const X86Subtarget *Subtarget) {
9450 EVT VT = N->getValueType(0);
9451 if (!VT.isVector() && VT.isInteger() &&
9452 N->getOpcode() == ISD::SHL)
9453 return PerformSHLCombine(N, DAG);
9455 // On X86 with SSE2 support, we can transform this to a vector shift if
9456 // all elements are shifted by the same amount. We can't do this in legalize
9457 // because the a constant vector is typically transformed to a constant pool
9458 // so we have no knowledge of the shift amount.
9459 if (!Subtarget->hasSSE2())
9462 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9465 SDValue ShAmtOp = N->getOperand(1);
9466 EVT EltVT = VT.getVectorElementType();
9467 DebugLoc DL = N->getDebugLoc();
9468 SDValue BaseShAmt = SDValue();
9469 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9470 unsigned NumElts = VT.getVectorNumElements();
9472 for (; i != NumElts; ++i) {
9473 SDValue Arg = ShAmtOp.getOperand(i);
9474 if (Arg.getOpcode() == ISD::UNDEF) continue;
9478 for (; i != NumElts; ++i) {
9479 SDValue Arg = ShAmtOp.getOperand(i);
9480 if (Arg.getOpcode() == ISD::UNDEF) continue;
9481 if (Arg != BaseShAmt) {
9485 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9486 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9487 SDValue InVec = ShAmtOp.getOperand(0);
9488 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9489 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9491 for (; i != NumElts; ++i) {
9492 SDValue Arg = InVec.getOperand(i);
9493 if (Arg.getOpcode() == ISD::UNDEF) continue;
9497 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9498 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9499 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9500 if (C->getZExtValue() == SplatIdx)
9501 BaseShAmt = InVec.getOperand(1);
9504 if (BaseShAmt.getNode() == 0)
9505 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9506 DAG.getIntPtrConstant(0));
9510 // The shift amount is an i32.
9511 if (EltVT.bitsGT(MVT::i32))
9512 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9513 else if (EltVT.bitsLT(MVT::i32))
9514 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9516 // The shift amount is identical so we can do a vector shift.
9517 SDValue ValOp = N->getOperand(0);
9518 switch (N->getOpcode()) {
9520 llvm_unreachable("Unknown shift opcode!");
9523 if (VT == MVT::v2i64)
9524 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9525 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9527 if (VT == MVT::v4i32)
9528 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9529 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9531 if (VT == MVT::v8i16)
9532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9533 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9537 if (VT == MVT::v4i32)
9538 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9539 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9541 if (VT == MVT::v8i16)
9542 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9543 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9547 if (VT == MVT::v2i64)
9548 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9549 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9551 if (VT == MVT::v4i32)
9552 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9553 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9555 if (VT == MVT::v8i16)
9556 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9557 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9564 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9565 TargetLowering::DAGCombinerInfo &DCI,
9566 const X86Subtarget *Subtarget) {
9567 if (DCI.isBeforeLegalizeOps())
9570 EVT VT = N->getValueType(0);
9571 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9574 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9575 SDValue N0 = N->getOperand(0);
9576 SDValue N1 = N->getOperand(1);
9577 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9579 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9581 if (!N0.hasOneUse() || !N1.hasOneUse())
9584 SDValue ShAmt0 = N0.getOperand(1);
9585 if (ShAmt0.getValueType() != MVT::i8)
9587 SDValue ShAmt1 = N1.getOperand(1);
9588 if (ShAmt1.getValueType() != MVT::i8)
9590 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9591 ShAmt0 = ShAmt0.getOperand(0);
9592 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9593 ShAmt1 = ShAmt1.getOperand(0);
9595 DebugLoc DL = N->getDebugLoc();
9596 unsigned Opc = X86ISD::SHLD;
9597 SDValue Op0 = N0.getOperand(0);
9598 SDValue Op1 = N1.getOperand(0);
9599 if (ShAmt0.getOpcode() == ISD::SUB) {
9601 std::swap(Op0, Op1);
9602 std::swap(ShAmt0, ShAmt1);
9605 unsigned Bits = VT.getSizeInBits();
9606 if (ShAmt1.getOpcode() == ISD::SUB) {
9607 SDValue Sum = ShAmt1.getOperand(0);
9608 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9609 if (SumC->getSExtValue() == Bits &&
9610 ShAmt1.getOperand(1) == ShAmt0)
9611 return DAG.getNode(Opc, DL, VT,
9613 DAG.getNode(ISD::TRUNCATE, DL,
9616 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9617 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9619 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9620 return DAG.getNode(Opc, DL, VT,
9621 N0.getOperand(0), N1.getOperand(0),
9622 DAG.getNode(ISD::TRUNCATE, DL,
9629 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9630 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9631 const X86Subtarget *Subtarget) {
9632 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9633 // the FP state in cases where an emms may be missing.
9634 // A preferable solution to the general problem is to figure out the right
9635 // places to insert EMMS. This qualifies as a quick hack.
9637 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9638 StoreSDNode *St = cast<StoreSDNode>(N);
9639 EVT VT = St->getValue().getValueType();
9640 if (VT.getSizeInBits() != 64)
9643 const Function *F = DAG.getMachineFunction().getFunction();
9644 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9645 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9646 && Subtarget->hasSSE2();
9647 if ((VT.isVector() ||
9648 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9649 isa<LoadSDNode>(St->getValue()) &&
9650 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9651 St->getChain().hasOneUse() && !St->isVolatile()) {
9652 SDNode* LdVal = St->getValue().getNode();
9654 int TokenFactorIndex = -1;
9655 SmallVector<SDValue, 8> Ops;
9656 SDNode* ChainVal = St->getChain().getNode();
9657 // Must be a store of a load. We currently handle two cases: the load
9658 // is a direct child, and it's under an intervening TokenFactor. It is
9659 // possible to dig deeper under nested TokenFactors.
9660 if (ChainVal == LdVal)
9661 Ld = cast<LoadSDNode>(St->getChain());
9662 else if (St->getValue().hasOneUse() &&
9663 ChainVal->getOpcode() == ISD::TokenFactor) {
9664 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9665 if (ChainVal->getOperand(i).getNode() == LdVal) {
9666 TokenFactorIndex = i;
9667 Ld = cast<LoadSDNode>(St->getValue());
9669 Ops.push_back(ChainVal->getOperand(i));
9673 if (!Ld || !ISD::isNormalLoad(Ld))
9676 // If this is not the MMX case, i.e. we are just turning i64 load/store
9677 // into f64 load/store, avoid the transformation if there are multiple
9678 // uses of the loaded value.
9679 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9682 DebugLoc LdDL = Ld->getDebugLoc();
9683 DebugLoc StDL = N->getDebugLoc();
9684 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9685 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9687 if (Subtarget->is64Bit() || F64IsLegal) {
9688 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9689 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9690 Ld->getBasePtr(), Ld->getSrcValue(),
9691 Ld->getSrcValueOffset(), Ld->isVolatile(),
9692 Ld->isNonTemporal(), Ld->getAlignment());
9693 SDValue NewChain = NewLd.getValue(1);
9694 if (TokenFactorIndex != -1) {
9695 Ops.push_back(NewChain);
9696 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9699 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9700 St->getSrcValue(), St->getSrcValueOffset(),
9701 St->isVolatile(), St->isNonTemporal(),
9702 St->getAlignment());
9705 // Otherwise, lower to two pairs of 32-bit loads / stores.
9706 SDValue LoAddr = Ld->getBasePtr();
9707 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9708 DAG.getConstant(4, MVT::i32));
9710 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9711 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9712 Ld->isVolatile(), Ld->isNonTemporal(),
9713 Ld->getAlignment());
9714 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9715 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9716 Ld->isVolatile(), Ld->isNonTemporal(),
9717 MinAlign(Ld->getAlignment(), 4));
9719 SDValue NewChain = LoLd.getValue(1);
9720 if (TokenFactorIndex != -1) {
9721 Ops.push_back(LoLd);
9722 Ops.push_back(HiLd);
9723 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9727 LoAddr = St->getBasePtr();
9728 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9729 DAG.getConstant(4, MVT::i32));
9731 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9732 St->getSrcValue(), St->getSrcValueOffset(),
9733 St->isVolatile(), St->isNonTemporal(),
9734 St->getAlignment());
9735 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9737 St->getSrcValueOffset() + 4,
9739 St->isNonTemporal(),
9740 MinAlign(St->getAlignment(), 4));
9741 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9746 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9747 /// X86ISD::FXOR nodes.
9748 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9749 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9750 // F[X]OR(0.0, x) -> x
9751 // F[X]OR(x, 0.0) -> x
9752 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9753 if (C->getValueAPF().isPosZero())
9754 return N->getOperand(1);
9755 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9756 if (C->getValueAPF().isPosZero())
9757 return N->getOperand(0);
9761 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9762 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9763 // FAND(0.0, x) -> 0.0
9764 // FAND(x, 0.0) -> 0.0
9765 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9766 if (C->getValueAPF().isPosZero())
9767 return N->getOperand(0);
9768 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9769 if (C->getValueAPF().isPosZero())
9770 return N->getOperand(1);
9774 static SDValue PerformBTCombine(SDNode *N,
9776 TargetLowering::DAGCombinerInfo &DCI) {
9777 // BT ignores high bits in the bit index operand.
9778 SDValue Op1 = N->getOperand(1);
9779 if (Op1.hasOneUse()) {
9780 unsigned BitWidth = Op1.getValueSizeInBits();
9781 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9782 APInt KnownZero, KnownOne;
9783 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9784 !DCI.isBeforeLegalizeOps());
9785 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9786 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9787 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9788 DCI.CommitTargetLoweringOpt(TLO);
9793 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9794 SDValue Op = N->getOperand(0);
9795 if (Op.getOpcode() == ISD::BIT_CONVERT)
9796 Op = Op.getOperand(0);
9797 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9798 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9799 VT.getVectorElementType().getSizeInBits() ==
9800 OpVT.getVectorElementType().getSizeInBits()) {
9801 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9806 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9807 // Locked instructions, in turn, have implicit fence semantics (all memory
9808 // operations are flushed before issuing the locked instruction, and the
9809 // are not buffered), so we can fold away the common pattern of
9810 // fence-atomic-fence.
9811 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9812 SDValue atomic = N->getOperand(0);
9813 switch (atomic.getOpcode()) {
9814 case ISD::ATOMIC_CMP_SWAP:
9815 case ISD::ATOMIC_SWAP:
9816 case ISD::ATOMIC_LOAD_ADD:
9817 case ISD::ATOMIC_LOAD_SUB:
9818 case ISD::ATOMIC_LOAD_AND:
9819 case ISD::ATOMIC_LOAD_OR:
9820 case ISD::ATOMIC_LOAD_XOR:
9821 case ISD::ATOMIC_LOAD_NAND:
9822 case ISD::ATOMIC_LOAD_MIN:
9823 case ISD::ATOMIC_LOAD_MAX:
9824 case ISD::ATOMIC_LOAD_UMIN:
9825 case ISD::ATOMIC_LOAD_UMAX:
9831 SDValue fence = atomic.getOperand(0);
9832 if (fence.getOpcode() != ISD::MEMBARRIER)
9835 switch (atomic.getOpcode()) {
9836 case ISD::ATOMIC_CMP_SWAP:
9837 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9838 atomic.getOperand(1), atomic.getOperand(2),
9839 atomic.getOperand(3));
9840 case ISD::ATOMIC_SWAP:
9841 case ISD::ATOMIC_LOAD_ADD:
9842 case ISD::ATOMIC_LOAD_SUB:
9843 case ISD::ATOMIC_LOAD_AND:
9844 case ISD::ATOMIC_LOAD_OR:
9845 case ISD::ATOMIC_LOAD_XOR:
9846 case ISD::ATOMIC_LOAD_NAND:
9847 case ISD::ATOMIC_LOAD_MIN:
9848 case ISD::ATOMIC_LOAD_MAX:
9849 case ISD::ATOMIC_LOAD_UMIN:
9850 case ISD::ATOMIC_LOAD_UMAX:
9851 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9852 atomic.getOperand(1), atomic.getOperand(2));
9858 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9859 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9860 // (and (i32 x86isd::setcc_carry), 1)
9861 // This eliminates the zext. This transformation is necessary because
9862 // ISD::SETCC is always legalized to i8.
9863 DebugLoc dl = N->getDebugLoc();
9864 SDValue N0 = N->getOperand(0);
9865 EVT VT = N->getValueType(0);
9866 if (N0.getOpcode() == ISD::AND &&
9868 N0.getOperand(0).hasOneUse()) {
9869 SDValue N00 = N0.getOperand(0);
9870 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9872 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9873 if (!C || C->getZExtValue() != 1)
9875 return DAG.getNode(ISD::AND, dl, VT,
9876 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9877 N00.getOperand(0), N00.getOperand(1)),
9878 DAG.getConstant(1, VT));
9884 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9885 DAGCombinerInfo &DCI) const {
9886 SelectionDAG &DAG = DCI.DAG;
9887 switch (N->getOpcode()) {
9889 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9890 case ISD::EXTRACT_VECTOR_ELT:
9891 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9892 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9893 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9894 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9897 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9898 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9899 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9901 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9902 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9903 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9904 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9905 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9906 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9912 /// isTypeDesirableForOp - Return true if the target has native support for
9913 /// the specified value type and it is 'desirable' to use the type for the
9914 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9915 /// instruction encodings are longer and some i16 instructions are slow.
9916 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9917 if (!isTypeLegal(VT))
9926 case ISD::SIGN_EXTEND:
9927 case ISD::ZERO_EXTEND:
9928 case ISD::ANY_EXTEND:
9941 static bool MayFoldLoad(SDValue Op) {
9942 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9945 static bool MayFoldIntoStore(SDValue Op) {
9946 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9949 /// IsDesirableToPromoteOp - This method query the target whether it is
9950 /// beneficial for dag combiner to promote the specified node. If true, it
9951 /// should return the desired promotion type by reference.
9952 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9953 EVT VT = Op.getValueType();
9957 bool Promote = false;
9958 bool Commute = false;
9959 switch (Op.getOpcode()) {
9962 LoadSDNode *LD = cast<LoadSDNode>(Op);
9963 // If the non-extending load has a single use and it's not live out, then it
9965 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9967 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9968 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9969 // The only case where we'd want to promote LOAD (rather then it being
9970 // promoted as an operand is when it's only use is liveout.
9971 if (UI->getOpcode() != ISD::CopyToReg)
9978 case ISD::SIGN_EXTEND:
9979 case ISD::ZERO_EXTEND:
9980 case ISD::ANY_EXTEND:
9985 SDValue N0 = Op.getOperand(0);
9986 // Look out for (store (shl (load), x)).
9987 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10000 SDValue N0 = Op.getOperand(0);
10001 SDValue N1 = Op.getOperand(1);
10002 if (!Commute && MayFoldLoad(N1))
10004 // Avoid disabling potential load folding opportunities.
10005 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10007 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10017 //===----------------------------------------------------------------------===//
10018 // X86 Inline Assembly Support
10019 //===----------------------------------------------------------------------===//
10021 static bool LowerToBSwap(CallInst *CI) {
10022 // FIXME: this should verify that we are targetting a 486 or better. If not,
10023 // we will turn this bswap into something that will be lowered to logical ops
10024 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10025 // so don't worry about this.
10027 // Verify this is a simple bswap.
10028 if (CI->getNumOperands() != 2 ||
10029 CI->getType() != CI->getOperand(1)->getType() ||
10030 !CI->getType()->isIntegerTy())
10033 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10034 if (!Ty || Ty->getBitWidth() % 16 != 0)
10037 // Okay, we can do this xform, do so now.
10038 const Type *Tys[] = { Ty };
10039 Module *M = CI->getParent()->getParent()->getParent();
10040 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10042 Value *Op = CI->getOperand(1);
10043 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10045 CI->replaceAllUsesWith(Op);
10046 CI->eraseFromParent();
10050 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10051 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10052 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10054 std::string AsmStr = IA->getAsmString();
10056 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10057 SmallVector<StringRef, 4> AsmPieces;
10058 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10060 switch (AsmPieces.size()) {
10061 default: return false;
10063 AsmStr = AsmPieces[0];
10065 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10068 if (AsmPieces.size() == 2 &&
10069 (AsmPieces[0] == "bswap" ||
10070 AsmPieces[0] == "bswapq" ||
10071 AsmPieces[0] == "bswapl") &&
10072 (AsmPieces[1] == "$0" ||
10073 AsmPieces[1] == "${0:q}")) {
10074 // No need to check constraints, nothing other than the equivalent of
10075 // "=r,0" would be valid here.
10076 return LowerToBSwap(CI);
10078 // rorw $$8, ${0:w} --> llvm.bswap.i16
10079 if (CI->getType()->isIntegerTy(16) &&
10080 AsmPieces.size() == 3 &&
10081 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10082 AsmPieces[1] == "$$8," &&
10083 AsmPieces[2] == "${0:w}" &&
10084 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10086 const std::string &Constraints = IA->getConstraintString();
10087 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10088 std::sort(AsmPieces.begin(), AsmPieces.end());
10089 if (AsmPieces.size() == 4 &&
10090 AsmPieces[0] == "~{cc}" &&
10091 AsmPieces[1] == "~{dirflag}" &&
10092 AsmPieces[2] == "~{flags}" &&
10093 AsmPieces[3] == "~{fpsr}") {
10094 return LowerToBSwap(CI);
10099 if (CI->getType()->isIntegerTy(64) &&
10100 Constraints.size() >= 2 &&
10101 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10102 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10103 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10104 SmallVector<StringRef, 4> Words;
10105 SplitString(AsmPieces[0], Words, " \t");
10106 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10108 SplitString(AsmPieces[1], Words, " \t");
10109 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10111 SplitString(AsmPieces[2], Words, " \t,");
10112 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10113 Words[2] == "%edx") {
10114 return LowerToBSwap(CI);
10126 /// getConstraintType - Given a constraint letter, return the type of
10127 /// constraint it is for this target.
10128 X86TargetLowering::ConstraintType
10129 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10130 if (Constraint.size() == 1) {
10131 switch (Constraint[0]) {
10143 return C_RegisterClass;
10151 return TargetLowering::getConstraintType(Constraint);
10154 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10155 /// with another that has more specific requirements based on the type of the
10156 /// corresponding operand.
10157 const char *X86TargetLowering::
10158 LowerXConstraint(EVT ConstraintVT) const {
10159 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10160 // 'f' like normal targets.
10161 if (ConstraintVT.isFloatingPoint()) {
10162 if (Subtarget->hasSSE2())
10164 if (Subtarget->hasSSE1())
10168 return TargetLowering::LowerXConstraint(ConstraintVT);
10171 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10172 /// vector. If it is invalid, don't add anything to Ops.
10173 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10176 std::vector<SDValue>&Ops,
10177 SelectionDAG &DAG) const {
10178 SDValue Result(0, 0);
10180 switch (Constraint) {
10183 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10184 if (C->getZExtValue() <= 31) {
10185 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10191 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10192 if (C->getZExtValue() <= 63) {
10193 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10199 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10200 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10201 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10207 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10208 if (C->getZExtValue() <= 255) {
10209 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10215 // 32-bit signed value
10216 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10217 const ConstantInt *CI = C->getConstantIntValue();
10218 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10219 C->getSExtValue())) {
10220 // Widen to 64 bits here to get it sign extended.
10221 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10224 // FIXME gcc accepts some relocatable values here too, but only in certain
10225 // memory models; it's complicated.
10230 // 32-bit unsigned value
10231 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10232 const ConstantInt *CI = C->getConstantIntValue();
10233 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10234 C->getZExtValue())) {
10235 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10239 // FIXME gcc accepts some relocatable values here too, but only in certain
10240 // memory models; it's complicated.
10244 // Literal immediates are always ok.
10245 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10246 // Widen to 64 bits here to get it sign extended.
10247 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10251 // If we are in non-pic codegen mode, we allow the address of a global (with
10252 // an optional displacement) to be used with 'i'.
10253 GlobalAddressSDNode *GA = 0;
10254 int64_t Offset = 0;
10256 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10258 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10259 Offset += GA->getOffset();
10261 } else if (Op.getOpcode() == ISD::ADD) {
10262 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10263 Offset += C->getZExtValue();
10264 Op = Op.getOperand(0);
10267 } else if (Op.getOpcode() == ISD::SUB) {
10268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10269 Offset += -C->getZExtValue();
10270 Op = Op.getOperand(0);
10275 // Otherwise, this isn't something we can handle, reject it.
10279 const GlobalValue *GV = GA->getGlobal();
10280 // If we require an extra load to get this address, as in PIC mode, we
10281 // can't accept it.
10282 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10283 getTargetMachine())))
10287 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10289 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10295 if (Result.getNode()) {
10296 Ops.push_back(Result);
10299 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10303 std::vector<unsigned> X86TargetLowering::
10304 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10306 if (Constraint.size() == 1) {
10307 // FIXME: not handling fp-stack yet!
10308 switch (Constraint[0]) { // GCC X86 Constraint Letters
10309 default: break; // Unknown constraint letter
10310 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10311 if (Subtarget->is64Bit()) {
10312 if (VT == MVT::i32)
10313 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10314 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10315 X86::R10D,X86::R11D,X86::R12D,
10316 X86::R13D,X86::R14D,X86::R15D,
10317 X86::EBP, X86::ESP, 0);
10318 else if (VT == MVT::i16)
10319 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10320 X86::SI, X86::DI, X86::R8W,X86::R9W,
10321 X86::R10W,X86::R11W,X86::R12W,
10322 X86::R13W,X86::R14W,X86::R15W,
10323 X86::BP, X86::SP, 0);
10324 else if (VT == MVT::i8)
10325 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10326 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10327 X86::R10B,X86::R11B,X86::R12B,
10328 X86::R13B,X86::R14B,X86::R15B,
10329 X86::BPL, X86::SPL, 0);
10331 else if (VT == MVT::i64)
10332 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10333 X86::RSI, X86::RDI, X86::R8, X86::R9,
10334 X86::R10, X86::R11, X86::R12,
10335 X86::R13, X86::R14, X86::R15,
10336 X86::RBP, X86::RSP, 0);
10340 // 32-bit fallthrough
10341 case 'Q': // Q_REGS
10342 if (VT == MVT::i32)
10343 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10344 else if (VT == MVT::i16)
10345 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10346 else if (VT == MVT::i8)
10347 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10348 else if (VT == MVT::i64)
10349 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10354 return std::vector<unsigned>();
10357 std::pair<unsigned, const TargetRegisterClass*>
10358 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10360 // First, see if this is a constraint that directly corresponds to an LLVM
10362 if (Constraint.size() == 1) {
10363 // GCC Constraint Letters
10364 switch (Constraint[0]) {
10366 case 'r': // GENERAL_REGS
10367 case 'l': // INDEX_REGS
10369 return std::make_pair(0U, X86::GR8RegisterClass);
10370 if (VT == MVT::i16)
10371 return std::make_pair(0U, X86::GR16RegisterClass);
10372 if (VT == MVT::i32 || !Subtarget->is64Bit())
10373 return std::make_pair(0U, X86::GR32RegisterClass);
10374 return std::make_pair(0U, X86::GR64RegisterClass);
10375 case 'R': // LEGACY_REGS
10377 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10378 if (VT == MVT::i16)
10379 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10380 if (VT == MVT::i32 || !Subtarget->is64Bit())
10381 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10382 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10383 case 'f': // FP Stack registers.
10384 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10385 // value to the correct fpstack register class.
10386 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10387 return std::make_pair(0U, X86::RFP32RegisterClass);
10388 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10389 return std::make_pair(0U, X86::RFP64RegisterClass);
10390 return std::make_pair(0U, X86::RFP80RegisterClass);
10391 case 'y': // MMX_REGS if MMX allowed.
10392 if (!Subtarget->hasMMX()) break;
10393 return std::make_pair(0U, X86::VR64RegisterClass);
10394 case 'Y': // SSE_REGS if SSE2 allowed
10395 if (!Subtarget->hasSSE2()) break;
10397 case 'x': // SSE_REGS if SSE1 allowed
10398 if (!Subtarget->hasSSE1()) break;
10400 switch (VT.getSimpleVT().SimpleTy) {
10402 // Scalar SSE types.
10405 return std::make_pair(0U, X86::FR32RegisterClass);
10408 return std::make_pair(0U, X86::FR64RegisterClass);
10416 return std::make_pair(0U, X86::VR128RegisterClass);
10422 // Use the default implementation in TargetLowering to convert the register
10423 // constraint into a member of a register class.
10424 std::pair<unsigned, const TargetRegisterClass*> Res;
10425 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10427 // Not found as a standard register?
10428 if (Res.second == 0) {
10429 // Map st(0) -> st(7) -> ST0
10430 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10431 tolower(Constraint[1]) == 's' &&
10432 tolower(Constraint[2]) == 't' &&
10433 Constraint[3] == '(' &&
10434 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10435 Constraint[5] == ')' &&
10436 Constraint[6] == '}') {
10438 Res.first = X86::ST0+Constraint[4]-'0';
10439 Res.second = X86::RFP80RegisterClass;
10443 // GCC allows "st(0)" to be called just plain "st".
10444 if (StringRef("{st}").equals_lower(Constraint)) {
10445 Res.first = X86::ST0;
10446 Res.second = X86::RFP80RegisterClass;
10451 if (StringRef("{flags}").equals_lower(Constraint)) {
10452 Res.first = X86::EFLAGS;
10453 Res.second = X86::CCRRegisterClass;
10457 // 'A' means EAX + EDX.
10458 if (Constraint == "A") {
10459 Res.first = X86::EAX;
10460 Res.second = X86::GR32_ADRegisterClass;
10466 // Otherwise, check to see if this is a register class of the wrong value
10467 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10468 // turn into {ax},{dx}.
10469 if (Res.second->hasType(VT))
10470 return Res; // Correct type already, nothing to do.
10472 // All of the single-register GCC register classes map their values onto
10473 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10474 // really want an 8-bit or 32-bit register, map to the appropriate register
10475 // class and return the appropriate register.
10476 if (Res.second == X86::GR16RegisterClass) {
10477 if (VT == MVT::i8) {
10478 unsigned DestReg = 0;
10479 switch (Res.first) {
10481 case X86::AX: DestReg = X86::AL; break;
10482 case X86::DX: DestReg = X86::DL; break;
10483 case X86::CX: DestReg = X86::CL; break;
10484 case X86::BX: DestReg = X86::BL; break;
10487 Res.first = DestReg;
10488 Res.second = X86::GR8RegisterClass;
10490 } else if (VT == MVT::i32) {
10491 unsigned DestReg = 0;
10492 switch (Res.first) {
10494 case X86::AX: DestReg = X86::EAX; break;
10495 case X86::DX: DestReg = X86::EDX; break;
10496 case X86::CX: DestReg = X86::ECX; break;
10497 case X86::BX: DestReg = X86::EBX; break;
10498 case X86::SI: DestReg = X86::ESI; break;
10499 case X86::DI: DestReg = X86::EDI; break;
10500 case X86::BP: DestReg = X86::EBP; break;
10501 case X86::SP: DestReg = X86::ESP; break;
10504 Res.first = DestReg;
10505 Res.second = X86::GR32RegisterClass;
10507 } else if (VT == MVT::i64) {
10508 unsigned DestReg = 0;
10509 switch (Res.first) {
10511 case X86::AX: DestReg = X86::RAX; break;
10512 case X86::DX: DestReg = X86::RDX; break;
10513 case X86::CX: DestReg = X86::RCX; break;
10514 case X86::BX: DestReg = X86::RBX; break;
10515 case X86::SI: DestReg = X86::RSI; break;
10516 case X86::DI: DestReg = X86::RDI; break;
10517 case X86::BP: DestReg = X86::RBP; break;
10518 case X86::SP: DestReg = X86::RSP; break;
10521 Res.first = DestReg;
10522 Res.second = X86::GR64RegisterClass;
10525 } else if (Res.second == X86::FR32RegisterClass ||
10526 Res.second == X86::FR64RegisterClass ||
10527 Res.second == X86::VR128RegisterClass) {
10528 // Handle references to XMM physical registers that got mapped into the
10529 // wrong class. This can happen with constraints like {xmm0} where the
10530 // target independent register mapper will just pick the first match it can
10531 // find, ignoring the required type.
10532 if (VT == MVT::f32)
10533 Res.second = X86::FR32RegisterClass;
10534 else if (VT == MVT::f64)
10535 Res.second = X86::FR64RegisterClass;
10536 else if (X86::VR128RegisterClass->hasType(VT))
10537 Res.second = X86::VR128RegisterClass;