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);
350 // On X86 and X86-64, atomic operations are lowered to locked instructions.
351 // Locked instructions, in turn, have implicit fence semantics (all memory
352 // operations are flushed before issuing the locked instruction, and they
353 // are not buffered), so we can fold away the common pattern of
354 // fence-atomic-fence.
355 setShouldFoldAtomicFences(true);
357 // Expand certain atomics
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
361 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
368 if (!Subtarget->is64Bit()) {
369 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
378 // FIXME - use subtarget debug flags
379 if (!Subtarget->isTargetDarwin() &&
380 !Subtarget->isTargetELF() &&
381 !Subtarget->isTargetCygMing()) {
382 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
387 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
388 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
389 if (Subtarget->is64Bit()) {
390 setExceptionPointerRegister(X86::RAX);
391 setExceptionSelectorRegister(X86::RDX);
393 setExceptionPointerRegister(X86::EAX);
394 setExceptionSelectorRegister(X86::EDX);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
397 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
399 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
401 setOperationAction(ISD::TRAP, MVT::Other, Legal);
403 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
404 setOperationAction(ISD::VASTART , MVT::Other, Custom);
405 setOperationAction(ISD::VAEND , MVT::Other, Expand);
406 if (Subtarget->is64Bit()) {
407 setOperationAction(ISD::VAARG , MVT::Other, Custom);
408 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
410 setOperationAction(ISD::VAARG , MVT::Other, Expand);
411 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
414 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
415 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
416 if (Subtarget->is64Bit())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
418 if (Subtarget->isTargetCygMing())
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
423 if (!UseSoftFloat && X86ScalarSSEf64) {
424 // f32 and f64 use SSE.
425 // Set up the FP register classes.
426 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
427 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
429 // Use ANDPD to simulate FABS.
430 setOperationAction(ISD::FABS , MVT::f64, Custom);
431 setOperationAction(ISD::FABS , MVT::f32, Custom);
433 // Use XORP to simulate FNEG.
434 setOperationAction(ISD::FNEG , MVT::f64, Custom);
435 setOperationAction(ISD::FNEG , MVT::f32, Custom);
437 // Use ANDPD and ORPD to simulate FCOPYSIGN.
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 // We don't support sin/cos/fmod
442 setOperationAction(ISD::FSIN , MVT::f64, Expand);
443 setOperationAction(ISD::FCOS , MVT::f64, Expand);
444 setOperationAction(ISD::FSIN , MVT::f32, Expand);
445 setOperationAction(ISD::FCOS , MVT::f32, Expand);
447 // Expand FP immediates into loads from the stack, except for the special
449 addLegalFPImmediate(APFloat(+0.0)); // xorpd
450 addLegalFPImmediate(APFloat(+0.0f)); // xorps
451 } else if (!UseSoftFloat && X86ScalarSSEf32) {
452 // Use SSE for f32, x87 for f64.
453 // Set up the FP register classes.
454 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
455 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
457 // Use ANDPS to simulate FABS.
458 setOperationAction(ISD::FABS , MVT::f32, Custom);
460 // Use XORP to simulate FNEG.
461 setOperationAction(ISD::FNEG , MVT::f32, Custom);
463 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
465 // Use ANDPS and ORPS to simulate FCOPYSIGN.
466 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
467 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
469 // We don't support sin/cos/fmod
470 setOperationAction(ISD::FSIN , MVT::f32, Expand);
471 setOperationAction(ISD::FCOS , MVT::f32, Expand);
473 // Special cases we handle for FP constants.
474 addLegalFPImmediate(APFloat(+0.0f)); // xorps
475 addLegalFPImmediate(APFloat(+0.0)); // FLD0
476 addLegalFPImmediate(APFloat(+1.0)); // FLD1
477 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
478 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
481 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
482 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
484 } else if (!UseSoftFloat) {
485 // f32 and f64 in x87.
486 // Set up the FP register classes.
487 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
488 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
490 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
491 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
493 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
496 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
497 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
499 addLegalFPImmediate(APFloat(+0.0)); // FLD0
500 addLegalFPImmediate(APFloat(+1.0)); // FLD1
501 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
502 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
503 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
504 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
505 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
506 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
509 // Long double always uses X87.
511 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
512 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
513 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
516 APFloat TmpFlt(+0.0);
517 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt); // FLD0
521 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
522 APFloat TmpFlt2(+1.0);
523 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt2); // FLD1
526 TmpFlt2.changeSign();
527 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
531 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
532 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
536 // Always use a library call for pow.
537 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
539 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
541 setOperationAction(ISD::FLOG, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
543 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP, MVT::f80, Expand);
545 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
547 // First set operation action for all vector types to either promote
548 // (for widening) or expand (for scalarization). Then we will selectively
549 // turn on ones that can be effectively codegen'd.
550 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
551 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
552 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
568 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
601 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
605 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
606 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
607 setTruncStoreAction((MVT::SimpleValueType)VT,
608 (MVT::SimpleValueType)InnerVT, Expand);
609 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
611 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
614 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
615 // with -msoft-float, disable use of MMX as well.
616 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
617 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
619 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
620 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
621 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
623 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
624 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
625 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
626 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
628 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
629 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
630 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
631 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
633 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
634 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
636 setOperationAction(ISD::AND, MVT::v8i8, Promote);
637 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v4i16, Promote);
639 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v2i32, Promote);
641 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
642 setOperationAction(ISD::AND, MVT::v1i64, Legal);
644 setOperationAction(ISD::OR, MVT::v8i8, Promote);
645 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v4i16, Promote);
647 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v2i32, Promote);
649 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
650 setOperationAction(ISD::OR, MVT::v1i64, Legal);
652 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
657 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
658 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
660 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
667 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
668 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
673 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
674 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
677 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
679 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
681 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
683 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
684 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
686 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
688 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
689 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
690 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
691 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
692 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
693 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
694 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
696 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
697 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
698 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
699 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
700 setOperationAction(ISD::BIT_CONVERT, MVT::v2f32, Custom);
701 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
705 if (!UseSoftFloat && Subtarget->hasSSE1()) {
706 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
708 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
709 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
710 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
711 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
712 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
713 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
714 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
715 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
716 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
717 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
718 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
722 if (!UseSoftFloat && Subtarget->hasSSE2()) {
723 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
725 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
726 // registers cannot be used even for integer operations.
727 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
729 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
730 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
732 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
733 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
734 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
735 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
736 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
737 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
738 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
739 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
740 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
741 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
742 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
743 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
744 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
745 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
746 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
747 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
751 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
752 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
754 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
755 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
758 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
763 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
764 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
766 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
767 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
768 EVT VT = (MVT::SimpleValueType)i;
769 // Do not attempt to custom lower non-power-of-2 vectors
770 if (!isPowerOf2_32(VT.getVectorNumElements()))
772 // Do not attempt to custom lower non-128-bit vectors
773 if (!VT.is128BitVector())
775 setOperationAction(ISD::BUILD_VECTOR,
776 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::VECTOR_SHUFFLE,
778 VT.getSimpleVT().SimpleTy, Custom);
779 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
780 VT.getSimpleVT().SimpleTy, Custom);
783 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
784 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
785 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
786 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
787 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
790 if (Subtarget->is64Bit()) {
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
795 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
796 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
797 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
800 // Do not attempt to promote non-128-bit vectors
801 if (!VT.is128BitVector()) {
805 setOperationAction(ISD::AND, SVT, Promote);
806 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
807 setOperationAction(ISD::OR, SVT, Promote);
808 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
809 setOperationAction(ISD::XOR, SVT, Promote);
810 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
811 setOperationAction(ISD::LOAD, SVT, Promote);
812 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
813 setOperationAction(ISD::SELECT, SVT, Promote);
814 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
817 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
819 // Custom lower v2i64 and v2f64 selects.
820 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
821 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
822 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
823 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
825 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
827 if (!DisableMMX && Subtarget->hasMMX()) {
828 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
829 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
833 if (Subtarget->hasSSE41()) {
834 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
835 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
836 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
837 setOperationAction(ISD::FRINT, MVT::f32, Legal);
838 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
839 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
840 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
841 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
842 setOperationAction(ISD::FRINT, MVT::f64, Legal);
843 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
845 // FIXME: Do we need to handle scalar-to-vector here?
846 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
848 // i8 and i16 vectors are custom , because the source register and source
849 // source memory operand types are not the same width. f32 vectors are
850 // custom since the immediate controlling the insert encodes additional
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
853 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
854 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
858 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
859 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
860 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
862 if (Subtarget->is64Bit()) {
863 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
864 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
868 if (Subtarget->hasSSE42()) {
869 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
872 if (!UseSoftFloat && Subtarget->hasAVX()) {
873 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
875 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
876 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
878 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
879 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
880 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
881 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
882 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
883 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
884 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
885 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
886 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
887 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
888 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
889 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
890 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
891 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
892 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
894 // Operations to consider commented out -v16i16 v32i8
895 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
896 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
897 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
898 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
899 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
900 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
901 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
902 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
903 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
904 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
905 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
906 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
907 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
908 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
910 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
911 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
912 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
913 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
915 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
916 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
917 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
919 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
921 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
922 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
923 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
924 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
925 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
926 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
929 // Not sure we want to do this since there are no 256-bit integer
932 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
933 // This includes 256-bit vectors
934 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
935 EVT VT = (MVT::SimpleValueType)i;
937 // Do not attempt to custom lower non-power-of-2 vectors
938 if (!isPowerOf2_32(VT.getVectorNumElements()))
941 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
942 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
943 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
946 if (Subtarget->is64Bit()) {
947 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
948 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
953 // Not sure we want to do this since there are no 256-bit integer
956 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
957 // Including 256-bit vectors
958 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
959 EVT VT = (MVT::SimpleValueType)i;
961 if (!VT.is256BitVector()) {
964 setOperationAction(ISD::AND, VT, Promote);
965 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
966 setOperationAction(ISD::OR, VT, Promote);
967 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
968 setOperationAction(ISD::XOR, VT, Promote);
969 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
970 setOperationAction(ISD::LOAD, VT, Promote);
971 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
972 setOperationAction(ISD::SELECT, VT, Promote);
973 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
976 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
980 // We want to custom lower some of our intrinsics.
981 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
983 // Add/Sub/Mul with overflow operations are custom lowered.
984 setOperationAction(ISD::SADDO, MVT::i32, Custom);
985 setOperationAction(ISD::UADDO, MVT::i32, Custom);
986 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
987 setOperationAction(ISD::USUBO, MVT::i32, Custom);
988 setOperationAction(ISD::SMULO, MVT::i32, Custom);
990 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
991 // handle type legalization for these operations here.
993 // FIXME: We really should do custom legalization for addition and
994 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
995 // than generic legalization for 64-bit multiplication-with-overflow, though.
996 if (Subtarget->is64Bit()) {
997 setOperationAction(ISD::SADDO, MVT::i64, Custom);
998 setOperationAction(ISD::UADDO, MVT::i64, Custom);
999 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
1000 setOperationAction(ISD::USUBO, MVT::i64, Custom);
1001 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1004 if (!Subtarget->is64Bit()) {
1005 // These libcalls are not available in 32-bit.
1006 setLibcallName(RTLIB::SHL_I128, 0);
1007 setLibcallName(RTLIB::SRL_I128, 0);
1008 setLibcallName(RTLIB::SRA_I128, 0);
1011 // We have target-specific dag combine patterns for the following nodes:
1012 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1013 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1014 setTargetDAGCombine(ISD::BUILD_VECTOR);
1015 setTargetDAGCombine(ISD::SELECT);
1016 setTargetDAGCombine(ISD::SHL);
1017 setTargetDAGCombine(ISD::SRA);
1018 setTargetDAGCombine(ISD::SRL);
1019 setTargetDAGCombine(ISD::OR);
1020 setTargetDAGCombine(ISD::STORE);
1021 setTargetDAGCombine(ISD::ZERO_EXTEND);
1022 if (Subtarget->is64Bit())
1023 setTargetDAGCombine(ISD::MUL);
1025 computeRegisterProperties();
1027 // FIXME: These should be based on subtarget info. Plus, the values should
1028 // be smaller when we are in optimizing for size mode.
1029 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1030 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1031 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1032 setPrefLoopAlignment(16);
1033 benefitFromCodePlacementOpt = true;
1037 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1042 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1043 /// the desired ByVal argument alignment.
1044 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1047 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1048 if (VTy->getBitWidth() == 128)
1050 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1051 unsigned EltAlign = 0;
1052 getMaxByValAlign(ATy->getElementType(), EltAlign);
1053 if (EltAlign > MaxAlign)
1054 MaxAlign = EltAlign;
1055 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1056 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1057 unsigned EltAlign = 0;
1058 getMaxByValAlign(STy->getElementType(i), EltAlign);
1059 if (EltAlign > MaxAlign)
1060 MaxAlign = EltAlign;
1068 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1069 /// function arguments in the caller parameter area. For X86, aggregates
1070 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1071 /// are at 4-byte boundaries.
1072 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1073 if (Subtarget->is64Bit()) {
1074 // Max of 8 and alignment of type.
1075 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1082 if (Subtarget->hasSSE1())
1083 getMaxByValAlign(Ty, Align);
1087 /// getOptimalMemOpType - Returns the target specific optimal type for load
1088 /// and store operations as a result of memset, memcpy, and memmove
1089 /// lowering. If DstAlign is zero that means it's safe to destination
1090 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1091 /// means there isn't a need to check it against alignment requirement,
1092 /// probably because the source does not need to be loaded. If
1093 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1094 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1095 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1096 /// constant so it does not need to be loaded.
1097 /// It returns EVT::Other if the type should be determined using generic
1098 /// target-independent logic.
1100 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1101 unsigned DstAlign, unsigned SrcAlign,
1102 bool NonScalarIntSafe,
1104 MachineFunction &MF) const {
1105 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1106 // linux. This is because the stack realignment code can't handle certain
1107 // cases like PR2962. This should be removed when PR2962 is fixed.
1108 const Function *F = MF.getFunction();
1109 if (NonScalarIntSafe &&
1110 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1112 (Subtarget->isUnalignedMemAccessFast() ||
1113 ((DstAlign == 0 || DstAlign >= 16) &&
1114 (SrcAlign == 0 || SrcAlign >= 16))) &&
1115 Subtarget->getStackAlignment() >= 16) {
1116 if (Subtarget->hasSSE2())
1118 if (Subtarget->hasSSE1())
1120 } else if (!MemcpyStrSrc && Size >= 8 &&
1121 !Subtarget->is64Bit() &&
1122 Subtarget->getStackAlignment() >= 8 &&
1123 Subtarget->hasSSE2()) {
1124 // Do not use f64 to lower memcpy if source is string constant. It's
1125 // better to use i32 to avoid the loads.
1129 if (Subtarget->is64Bit() && Size >= 8)
1134 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1135 /// current function. The returned value is a member of the
1136 /// MachineJumpTableInfo::JTEntryKind enum.
1137 unsigned X86TargetLowering::getJumpTableEncoding() const {
1138 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1140 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1141 Subtarget->isPICStyleGOT())
1142 return MachineJumpTableInfo::EK_Custom32;
1144 // Otherwise, use the normal jump table encoding heuristics.
1145 return TargetLowering::getJumpTableEncoding();
1148 /// getPICBaseSymbol - Return the X86-32 PIC base.
1150 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1151 MCContext &Ctx) const {
1152 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1153 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1154 Twine(MF->getFunctionNumber())+"$pb");
1159 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1160 const MachineBasicBlock *MBB,
1161 unsigned uid,MCContext &Ctx) const{
1162 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1163 Subtarget->isPICStyleGOT());
1164 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1166 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1167 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1170 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1172 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1173 SelectionDAG &DAG) const {
1174 if (!Subtarget->is64Bit())
1175 // This doesn't have DebugLoc associated with it, but is not really the
1176 // same as a Register.
1177 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1181 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1182 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1184 const MCExpr *X86TargetLowering::
1185 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1186 MCContext &Ctx) const {
1187 // X86-64 uses RIP relative addressing based on the jump table label.
1188 if (Subtarget->isPICStyleRIPRel())
1189 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1191 // Otherwise, the reference is relative to the PIC base.
1192 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1195 /// getFunctionAlignment - Return the Log2 alignment of this function.
1196 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1197 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1200 //===----------------------------------------------------------------------===//
1201 // Return Value Calling Convention Implementation
1202 //===----------------------------------------------------------------------===//
1204 #include "X86GenCallingConv.inc"
1207 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1208 const SmallVectorImpl<EVT> &OutTys,
1209 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1210 SelectionDAG &DAG) const {
1211 SmallVector<CCValAssign, 16> RVLocs;
1212 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1213 RVLocs, *DAG.getContext());
1214 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1218 X86TargetLowering::LowerReturn(SDValue Chain,
1219 CallingConv::ID CallConv, bool isVarArg,
1220 const SmallVectorImpl<ISD::OutputArg> &Outs,
1221 DebugLoc dl, SelectionDAG &DAG) const {
1222 MachineFunction &MF = DAG.getMachineFunction();
1223 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1225 SmallVector<CCValAssign, 16> RVLocs;
1226 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1227 RVLocs, *DAG.getContext());
1228 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1230 // Add the regs to the liveout set for the function.
1231 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1232 for (unsigned i = 0; i != RVLocs.size(); ++i)
1233 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1234 MRI.addLiveOut(RVLocs[i].getLocReg());
1238 SmallVector<SDValue, 6> RetOps;
1239 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1240 // Operand #1 = Bytes To Pop
1241 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1244 // Copy the result values into the output registers.
1245 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1246 CCValAssign &VA = RVLocs[i];
1247 assert(VA.isRegLoc() && "Can only return in registers!");
1248 SDValue ValToCopy = Outs[i].Val;
1250 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1251 // the RET instruction and handled by the FP Stackifier.
1252 if (VA.getLocReg() == X86::ST0 ||
1253 VA.getLocReg() == X86::ST1) {
1254 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1255 // change the value to the FP stack register class.
1256 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1257 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1258 RetOps.push_back(ValToCopy);
1259 // Don't emit a copytoreg.
1263 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1264 // which is returned in RAX / RDX.
1265 if (Subtarget->is64Bit()) {
1266 EVT ValVT = ValToCopy.getValueType();
1267 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1268 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1269 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1270 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1274 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1275 Flag = Chain.getValue(1);
1278 // The x86-64 ABI for returning structs by value requires that we copy
1279 // the sret argument into %rax for the return. We saved the argument into
1280 // a virtual register in the entry block, so now we copy the value out
1282 if (Subtarget->is64Bit() &&
1283 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1284 MachineFunction &MF = DAG.getMachineFunction();
1285 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1286 unsigned Reg = FuncInfo->getSRetReturnReg();
1288 "SRetReturnReg should have been set in LowerFormalArguments().");
1289 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1291 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1292 Flag = Chain.getValue(1);
1294 // RAX now acts like a return value.
1295 MRI.addLiveOut(X86::RAX);
1298 RetOps[0] = Chain; // Update chain.
1300 // Add the flag if we have it.
1302 RetOps.push_back(Flag);
1304 return DAG.getNode(X86ISD::RET_FLAG, dl,
1305 MVT::Other, &RetOps[0], RetOps.size());
1308 /// LowerCallResult - Lower the result values of a call into the
1309 /// appropriate copies out of appropriate physical registers.
1312 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1313 CallingConv::ID CallConv, bool isVarArg,
1314 const SmallVectorImpl<ISD::InputArg> &Ins,
1315 DebugLoc dl, SelectionDAG &DAG,
1316 SmallVectorImpl<SDValue> &InVals) const {
1318 // Assign locations to each value returned by this call.
1319 SmallVector<CCValAssign, 16> RVLocs;
1320 bool Is64Bit = Subtarget->is64Bit();
1321 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1322 RVLocs, *DAG.getContext());
1323 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1325 // Copy all of the result registers out of their specified physreg.
1326 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1327 CCValAssign &VA = RVLocs[i];
1328 EVT CopyVT = VA.getValVT();
1330 // If this is x86-64, and we disabled SSE, we can't return FP values
1331 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1332 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1333 report_fatal_error("SSE register return with SSE disabled");
1336 // If this is a call to a function that returns an fp value on the floating
1337 // point stack, but where we prefer to use the value in xmm registers, copy
1338 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1339 if ((VA.getLocReg() == X86::ST0 ||
1340 VA.getLocReg() == X86::ST1) &&
1341 isScalarFPTypeInSSEReg(VA.getValVT())) {
1346 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1347 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1348 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1349 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1350 MVT::v2i64, InFlag).getValue(1);
1351 Val = Chain.getValue(0);
1352 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1353 Val, DAG.getConstant(0, MVT::i64));
1355 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1356 MVT::i64, InFlag).getValue(1);
1357 Val = Chain.getValue(0);
1359 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1361 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1362 CopyVT, InFlag).getValue(1);
1363 Val = Chain.getValue(0);
1365 InFlag = Chain.getValue(2);
1367 if (CopyVT != VA.getValVT()) {
1368 // Round the F80 the right size, which also moves to the appropriate xmm
1370 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1371 // This truncation won't change the value.
1372 DAG.getIntPtrConstant(1));
1375 InVals.push_back(Val);
1382 //===----------------------------------------------------------------------===//
1383 // C & StdCall & Fast Calling Convention implementation
1384 //===----------------------------------------------------------------------===//
1385 // StdCall calling convention seems to be standard for many Windows' API
1386 // routines and around. It differs from C calling convention just a little:
1387 // callee should clean up the stack, not caller. Symbols should be also
1388 // decorated in some fancy way :) It doesn't support any vector arguments.
1389 // For info on fast calling convention see Fast Calling Convention (tail call)
1390 // implementation LowerX86_32FastCCCallTo.
1392 /// CallIsStructReturn - Determines whether a call uses struct return
1394 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1398 return Outs[0].Flags.isSRet();
1401 /// ArgsAreStructReturn - Determines whether a function uses struct
1402 /// return semantics.
1404 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1408 return Ins[0].Flags.isSRet();
1411 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1412 /// given CallingConvention value.
1413 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1414 if (Subtarget->is64Bit()) {
1415 if (CC == CallingConv::GHC)
1416 return CC_X86_64_GHC;
1417 else if (Subtarget->isTargetWin64())
1418 return CC_X86_Win64_C;
1423 if (CC == CallingConv::X86_FastCall)
1424 return CC_X86_32_FastCall;
1425 else if (CC == CallingConv::X86_ThisCall)
1426 return CC_X86_32_ThisCall;
1427 else if (CC == CallingConv::Fast)
1428 return CC_X86_32_FastCC;
1429 else if (CC == CallingConv::GHC)
1430 return CC_X86_32_GHC;
1435 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1436 /// by "Src" to address "Dst" with size and alignment information specified by
1437 /// the specific parameter attribute. The copy will be passed as a byval
1438 /// function parameter.
1440 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1441 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1443 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1444 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1445 /*isVolatile*/false, /*AlwaysInline=*/true,
1449 /// IsTailCallConvention - Return true if the calling convention is one that
1450 /// supports tail call optimization.
1451 static bool IsTailCallConvention(CallingConv::ID CC) {
1452 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1455 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1456 /// a tailcall target by changing its ABI.
1457 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1458 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1462 X86TargetLowering::LowerMemArgument(SDValue Chain,
1463 CallingConv::ID CallConv,
1464 const SmallVectorImpl<ISD::InputArg> &Ins,
1465 DebugLoc dl, SelectionDAG &DAG,
1466 const CCValAssign &VA,
1467 MachineFrameInfo *MFI,
1469 // Create the nodes corresponding to a load from this parameter slot.
1470 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1471 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1472 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1475 // If value is passed by pointer we have address passed instead of the value
1477 if (VA.getLocInfo() == CCValAssign::Indirect)
1478 ValVT = VA.getLocVT();
1480 ValVT = VA.getValVT();
1482 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1483 // changed with more analysis.
1484 // In case of tail call optimization mark all arguments mutable. Since they
1485 // could be overwritten by lowering of arguments in case of a tail call.
1486 if (Flags.isByVal()) {
1487 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1488 VA.getLocMemOffset(), isImmutable, false);
1489 return DAG.getFrameIndex(FI, getPointerTy());
1491 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1492 VA.getLocMemOffset(), isImmutable, false);
1493 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1494 return DAG.getLoad(ValVT, dl, Chain, FIN,
1495 PseudoSourceValue::getFixedStack(FI), 0,
1501 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1502 CallingConv::ID CallConv,
1504 const SmallVectorImpl<ISD::InputArg> &Ins,
1507 SmallVectorImpl<SDValue> &InVals)
1509 MachineFunction &MF = DAG.getMachineFunction();
1510 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1512 const Function* Fn = MF.getFunction();
1513 if (Fn->hasExternalLinkage() &&
1514 Subtarget->isTargetCygMing() &&
1515 Fn->getName() == "main")
1516 FuncInfo->setForceFramePointer(true);
1518 MachineFrameInfo *MFI = MF.getFrameInfo();
1519 bool Is64Bit = Subtarget->is64Bit();
1520 bool IsWin64 = Subtarget->isTargetWin64();
1522 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1523 "Var args not supported with calling convention fastcc or ghc");
1525 // Assign locations to all of the incoming arguments.
1526 SmallVector<CCValAssign, 16> ArgLocs;
1527 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1528 ArgLocs, *DAG.getContext());
1529 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1531 unsigned LastVal = ~0U;
1533 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1534 CCValAssign &VA = ArgLocs[i];
1535 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1537 assert(VA.getValNo() != LastVal &&
1538 "Don't support value assigned to multiple locs yet");
1539 LastVal = VA.getValNo();
1541 if (VA.isRegLoc()) {
1542 EVT RegVT = VA.getLocVT();
1543 TargetRegisterClass *RC = NULL;
1544 if (RegVT == MVT::i32)
1545 RC = X86::GR32RegisterClass;
1546 else if (Is64Bit && RegVT == MVT::i64)
1547 RC = X86::GR64RegisterClass;
1548 else if (RegVT == MVT::f32)
1549 RC = X86::FR32RegisterClass;
1550 else if (RegVT == MVT::f64)
1551 RC = X86::FR64RegisterClass;
1552 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1553 RC = X86::VR128RegisterClass;
1554 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1555 RC = X86::VR64RegisterClass;
1557 llvm_unreachable("Unknown argument type!");
1559 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1560 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1562 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1563 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1565 if (VA.getLocInfo() == CCValAssign::SExt)
1566 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1567 DAG.getValueType(VA.getValVT()));
1568 else if (VA.getLocInfo() == CCValAssign::ZExt)
1569 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1570 DAG.getValueType(VA.getValVT()));
1571 else if (VA.getLocInfo() == CCValAssign::BCvt)
1572 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1574 if (VA.isExtInLoc()) {
1575 // Handle MMX values passed in XMM regs.
1576 if (RegVT.isVector()) {
1577 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1578 ArgValue, DAG.getConstant(0, MVT::i64));
1579 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1581 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1584 assert(VA.isMemLoc());
1585 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1588 // If value is passed via pointer - do a load.
1589 if (VA.getLocInfo() == CCValAssign::Indirect)
1590 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1593 InVals.push_back(ArgValue);
1596 // The x86-64 ABI for returning structs by value requires that we copy
1597 // the sret argument into %rax for the return. Save the argument into
1598 // a virtual register so that we can access it from the return points.
1599 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1600 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1601 unsigned Reg = FuncInfo->getSRetReturnReg();
1603 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1604 FuncInfo->setSRetReturnReg(Reg);
1606 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1607 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1610 unsigned StackSize = CCInfo.getNextStackOffset();
1611 // Align stack specially for tail calls.
1612 if (FuncIsMadeTailCallSafe(CallConv))
1613 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1615 // If the function takes variable number of arguments, make a frame index for
1616 // the start of the first vararg value... for expansion of llvm.va_start.
1618 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1619 CallConv != CallingConv::X86_ThisCall)) {
1620 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
1624 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1626 // FIXME: We should really autogenerate these arrays
1627 static const unsigned GPR64ArgRegsWin64[] = {
1628 X86::RCX, X86::RDX, X86::R8, X86::R9
1630 static const unsigned XMMArgRegsWin64[] = {
1631 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1633 static const unsigned GPR64ArgRegs64Bit[] = {
1634 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1636 static const unsigned XMMArgRegs64Bit[] = {
1637 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1638 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1640 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1643 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1644 GPR64ArgRegs = GPR64ArgRegsWin64;
1645 XMMArgRegs = XMMArgRegsWin64;
1647 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1648 GPR64ArgRegs = GPR64ArgRegs64Bit;
1649 XMMArgRegs = XMMArgRegs64Bit;
1651 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1653 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1656 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1657 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1658 "SSE register cannot be used when SSE is disabled!");
1659 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1660 "SSE register cannot be used when SSE is disabled!");
1661 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1662 // Kernel mode asks for SSE to be disabled, so don't push them
1664 TotalNumXMMRegs = 0;
1666 // For X86-64, if there are vararg parameters that are passed via
1667 // registers, then we must store them to their spots on the stack so they
1668 // may be loaded by deferencing the result of va_next.
1669 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1670 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1671 FuncInfo->setRegSaveFrameIndex(
1672 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1675 // Store the integer parameter registers.
1676 SmallVector<SDValue, 8> MemOps;
1677 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1679 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1680 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1681 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1682 DAG.getIntPtrConstant(Offset));
1683 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1684 X86::GR64RegisterClass);
1685 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1687 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1688 PseudoSourceValue::getFixedStack(
1689 FuncInfo->getRegSaveFrameIndex()),
1690 Offset, false, false, 0);
1691 MemOps.push_back(Store);
1695 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1696 // Now store the XMM (fp + vector) parameter registers.
1697 SmallVector<SDValue, 11> SaveXMMOps;
1698 SaveXMMOps.push_back(Chain);
1700 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1701 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1702 SaveXMMOps.push_back(ALVal);
1704 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1705 FuncInfo->getRegSaveFrameIndex()));
1706 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1707 FuncInfo->getVarArgsFPOffset()));
1709 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1710 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1711 X86::VR128RegisterClass);
1712 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1713 SaveXMMOps.push_back(Val);
1715 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1717 &SaveXMMOps[0], SaveXMMOps.size()));
1720 if (!MemOps.empty())
1721 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1722 &MemOps[0], MemOps.size());
1726 // Some CCs need callee pop.
1727 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1728 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1730 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1731 // If this is an sret function, the return should pop the hidden pointer.
1732 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1733 FuncInfo->setBytesToPopOnReturn(4);
1737 // RegSaveFrameIndex is X86-64 only.
1738 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1739 if (CallConv == CallingConv::X86_FastCall ||
1740 CallConv == CallingConv::X86_ThisCall)
1741 // fastcc functions can't have varargs.
1742 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1749 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1750 SDValue StackPtr, SDValue Arg,
1751 DebugLoc dl, SelectionDAG &DAG,
1752 const CCValAssign &VA,
1753 ISD::ArgFlagsTy Flags) const {
1754 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1755 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1756 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1757 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1758 if (Flags.isByVal()) {
1759 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1761 return DAG.getStore(Chain, dl, Arg, PtrOff,
1762 PseudoSourceValue::getStack(), LocMemOffset,
1766 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1767 /// optimization is performed and it is required.
1769 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1770 SDValue &OutRetAddr, SDValue Chain,
1771 bool IsTailCall, bool Is64Bit,
1772 int FPDiff, DebugLoc dl) const {
1773 // Adjust the Return address stack slot.
1774 EVT VT = getPointerTy();
1775 OutRetAddr = getReturnAddressFrameIndex(DAG);
1777 // Load the "old" Return address.
1778 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1779 return SDValue(OutRetAddr.getNode(), 1);
1782 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1783 /// optimization is performed and it is required (FPDiff!=0).
1785 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1786 SDValue Chain, SDValue RetAddrFrIdx,
1787 bool Is64Bit, int FPDiff, DebugLoc dl) {
1788 // Store the return address to the appropriate stack slot.
1789 if (!FPDiff) return Chain;
1790 // Calculate the new stack slot for the return address.
1791 int SlotSize = Is64Bit ? 8 : 4;
1792 int NewReturnAddrFI =
1793 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1794 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1795 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1796 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1797 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1803 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1804 CallingConv::ID CallConv, bool isVarArg,
1806 const SmallVectorImpl<ISD::OutputArg> &Outs,
1807 const SmallVectorImpl<ISD::InputArg> &Ins,
1808 DebugLoc dl, SelectionDAG &DAG,
1809 SmallVectorImpl<SDValue> &InVals) const {
1810 MachineFunction &MF = DAG.getMachineFunction();
1811 bool Is64Bit = Subtarget->is64Bit();
1812 bool IsStructRet = CallIsStructReturn(Outs);
1813 bool IsSibcall = false;
1816 // Check if it's really possible to do a tail call.
1817 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1818 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1821 // Sibcalls are automatically detected tailcalls which do not require
1823 if (!GuaranteedTailCallOpt && isTailCall)
1830 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1831 "Var args not supported with calling convention fastcc or ghc");
1833 // Analyze operands of the call, assigning locations to each operand.
1834 SmallVector<CCValAssign, 16> ArgLocs;
1835 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1836 ArgLocs, *DAG.getContext());
1837 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1839 // Get a count of how many bytes are to be pushed on the stack.
1840 unsigned NumBytes = CCInfo.getNextStackOffset();
1842 // This is a sibcall. The memory operands are available in caller's
1843 // own caller's stack.
1845 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1846 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1849 if (isTailCall && !IsSibcall) {
1850 // Lower arguments at fp - stackoffset + fpdiff.
1851 unsigned NumBytesCallerPushed =
1852 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1853 FPDiff = NumBytesCallerPushed - NumBytes;
1855 // Set the delta of movement of the returnaddr stackslot.
1856 // But only set if delta is greater than previous delta.
1857 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1858 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1862 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1864 SDValue RetAddrFrIdx;
1865 // Load return adress for tail calls.
1866 if (isTailCall && FPDiff)
1867 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1868 Is64Bit, FPDiff, dl);
1870 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1871 SmallVector<SDValue, 8> MemOpChains;
1874 // Walk the register/memloc assignments, inserting copies/loads. In the case
1875 // of tail call optimization arguments are handle later.
1876 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1877 CCValAssign &VA = ArgLocs[i];
1878 EVT RegVT = VA.getLocVT();
1879 SDValue Arg = Outs[i].Val;
1880 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1881 bool isByVal = Flags.isByVal();
1883 // Promote the value if needed.
1884 switch (VA.getLocInfo()) {
1885 default: llvm_unreachable("Unknown loc info!");
1886 case CCValAssign::Full: break;
1887 case CCValAssign::SExt:
1888 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1890 case CCValAssign::ZExt:
1891 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1893 case CCValAssign::AExt:
1894 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1895 // Special case: passing MMX values in XMM registers.
1896 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1897 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1898 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1900 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1902 case CCValAssign::BCvt:
1903 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1905 case CCValAssign::Indirect: {
1906 // Store the argument.
1907 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1908 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1909 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1910 PseudoSourceValue::getFixedStack(FI), 0,
1917 if (VA.isRegLoc()) {
1918 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1919 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1920 assert(VA.isMemLoc());
1921 if (StackPtr.getNode() == 0)
1922 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1923 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1924 dl, DAG, VA, Flags));
1928 if (!MemOpChains.empty())
1929 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1930 &MemOpChains[0], MemOpChains.size());
1932 // Build a sequence of copy-to-reg nodes chained together with token chain
1933 // and flag operands which copy the outgoing args into registers.
1935 // Tail call byval lowering might overwrite argument registers so in case of
1936 // tail call optimization the copies to registers are lowered later.
1938 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1939 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1940 RegsToPass[i].second, InFlag);
1941 InFlag = Chain.getValue(1);
1944 if (Subtarget->isPICStyleGOT()) {
1945 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1948 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1949 DAG.getNode(X86ISD::GlobalBaseReg,
1950 DebugLoc(), getPointerTy()),
1952 InFlag = Chain.getValue(1);
1954 // If we are tail calling and generating PIC/GOT style code load the
1955 // address of the callee into ECX. The value in ecx is used as target of
1956 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1957 // for tail calls on PIC/GOT architectures. Normally we would just put the
1958 // address of GOT into ebx and then call target@PLT. But for tail calls
1959 // ebx would be restored (since ebx is callee saved) before jumping to the
1962 // Note: The actual moving to ECX is done further down.
1963 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1964 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1965 !G->getGlobal()->hasProtectedVisibility())
1966 Callee = LowerGlobalAddress(Callee, DAG);
1967 else if (isa<ExternalSymbolSDNode>(Callee))
1968 Callee = LowerExternalSymbol(Callee, DAG);
1972 if (Is64Bit && isVarArg) {
1973 // From AMD64 ABI document:
1974 // For calls that may call functions that use varargs or stdargs
1975 // (prototype-less calls or calls to functions containing ellipsis (...) in
1976 // the declaration) %al is used as hidden argument to specify the number
1977 // of SSE registers used. The contents of %al do not need to match exactly
1978 // the number of registers, but must be an ubound on the number of SSE
1979 // registers used and is in the range 0 - 8 inclusive.
1981 // FIXME: Verify this on Win64
1982 // Count the number of XMM registers allocated.
1983 static const unsigned XMMArgRegs[] = {
1984 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1985 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1987 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1988 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1989 && "SSE registers cannot be used when SSE is disabled");
1991 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1992 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1993 InFlag = Chain.getValue(1);
1997 // For tail calls lower the arguments to the 'real' stack slot.
1999 // Force all the incoming stack arguments to be loaded from the stack
2000 // before any new outgoing arguments are stored to the stack, because the
2001 // outgoing stack slots may alias the incoming argument stack slots, and
2002 // the alias isn't otherwise explicit. This is slightly more conservative
2003 // than necessary, because it means that each store effectively depends
2004 // on every argument instead of just those arguments it would clobber.
2005 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2007 SmallVector<SDValue, 8> MemOpChains2;
2010 // Do not flag preceeding copytoreg stuff together with the following stuff.
2012 if (GuaranteedTailCallOpt) {
2013 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2014 CCValAssign &VA = ArgLocs[i];
2017 assert(VA.isMemLoc());
2018 SDValue Arg = Outs[i].Val;
2019 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2020 // Create frame index.
2021 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2022 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2023 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2024 FIN = DAG.getFrameIndex(FI, getPointerTy());
2026 if (Flags.isByVal()) {
2027 // Copy relative to framepointer.
2028 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2029 if (StackPtr.getNode() == 0)
2030 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2032 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2034 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2038 // Store relative to framepointer.
2039 MemOpChains2.push_back(
2040 DAG.getStore(ArgChain, dl, Arg, FIN,
2041 PseudoSourceValue::getFixedStack(FI), 0,
2047 if (!MemOpChains2.empty())
2048 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2049 &MemOpChains2[0], MemOpChains2.size());
2051 // Copy arguments to their registers.
2052 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2053 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2054 RegsToPass[i].second, InFlag);
2055 InFlag = Chain.getValue(1);
2059 // Store the return address to the appropriate stack slot.
2060 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2064 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2065 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2066 // In the 64-bit large code model, we have to make all calls
2067 // through a register, since the call instruction's 32-bit
2068 // pc-relative offset may not be large enough to hold the whole
2070 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2071 // If the callee is a GlobalAddress node (quite common, every direct call
2072 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2075 // We should use extra load for direct calls to dllimported functions in
2077 const GlobalValue *GV = G->getGlobal();
2078 if (!GV->hasDLLImportLinkage()) {
2079 unsigned char OpFlags = 0;
2081 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2082 // external symbols most go through the PLT in PIC mode. If the symbol
2083 // has hidden or protected visibility, or if it is static or local, then
2084 // we don't need to use the PLT - we can directly call it.
2085 if (Subtarget->isTargetELF() &&
2086 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2087 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2088 OpFlags = X86II::MO_PLT;
2089 } else if (Subtarget->isPICStyleStubAny() &&
2090 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2091 Subtarget->getDarwinVers() < 9) {
2092 // PC-relative references to external symbols should go through $stub,
2093 // unless we're building with the leopard linker or later, which
2094 // automatically synthesizes these stubs.
2095 OpFlags = X86II::MO_DARWIN_STUB;
2098 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2099 G->getOffset(), OpFlags);
2101 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2102 unsigned char OpFlags = 0;
2104 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2105 // symbols should go through the PLT.
2106 if (Subtarget->isTargetELF() &&
2107 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2108 OpFlags = X86II::MO_PLT;
2109 } else if (Subtarget->isPICStyleStubAny() &&
2110 Subtarget->getDarwinVers() < 9) {
2111 // PC-relative references to external symbols should go through $stub,
2112 // unless we're building with the leopard linker or later, which
2113 // automatically synthesizes these stubs.
2114 OpFlags = X86II::MO_DARWIN_STUB;
2117 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2121 // Returns a chain & a flag for retval copy to use.
2122 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2123 SmallVector<SDValue, 8> Ops;
2125 if (!IsSibcall && isTailCall) {
2126 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2127 DAG.getIntPtrConstant(0, true), InFlag);
2128 InFlag = Chain.getValue(1);
2131 Ops.push_back(Chain);
2132 Ops.push_back(Callee);
2135 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2137 // Add argument registers to the end of the list so that they are known live
2139 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2140 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2141 RegsToPass[i].second.getValueType()));
2143 // Add an implicit use GOT pointer in EBX.
2144 if (!isTailCall && Subtarget->isPICStyleGOT())
2145 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2147 // Add an implicit use of AL for x86 vararg functions.
2148 if (Is64Bit && isVarArg)
2149 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2151 if (InFlag.getNode())
2152 Ops.push_back(InFlag);
2156 //// If this is the first return lowered for this function, add the regs
2157 //// to the liveout set for the function.
2158 // This isn't right, although it's probably harmless on x86; liveouts
2159 // should be computed from returns not tail calls. Consider a void
2160 // function making a tail call to a function returning int.
2161 return DAG.getNode(X86ISD::TC_RETURN, dl,
2162 NodeTys, &Ops[0], Ops.size());
2165 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2166 InFlag = Chain.getValue(1);
2168 // Create the CALLSEQ_END node.
2169 unsigned NumBytesForCalleeToPush;
2170 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2171 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2172 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2173 // If this is a call to a struct-return function, the callee
2174 // pops the hidden struct pointer, so we have to push it back.
2175 // This is common for Darwin/X86, Linux & Mingw32 targets.
2176 NumBytesForCalleeToPush = 4;
2178 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2180 // Returns a flag for retval copy to use.
2182 Chain = DAG.getCALLSEQ_END(Chain,
2183 DAG.getIntPtrConstant(NumBytes, true),
2184 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2187 InFlag = Chain.getValue(1);
2190 // Handle result values, copying them out of physregs into vregs that we
2192 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2193 Ins, dl, DAG, InVals);
2197 //===----------------------------------------------------------------------===//
2198 // Fast Calling Convention (tail call) implementation
2199 //===----------------------------------------------------------------------===//
2201 // Like std call, callee cleans arguments, convention except that ECX is
2202 // reserved for storing the tail called function address. Only 2 registers are
2203 // free for argument passing (inreg). Tail call optimization is performed
2205 // * tailcallopt is enabled
2206 // * caller/callee are fastcc
2207 // On X86_64 architecture with GOT-style position independent code only local
2208 // (within module) calls are supported at the moment.
2209 // To keep the stack aligned according to platform abi the function
2210 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2211 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2212 // If a tail called function callee has more arguments than the caller the
2213 // caller needs to make sure that there is room to move the RETADDR to. This is
2214 // achieved by reserving an area the size of the argument delta right after the
2215 // original REtADDR, but before the saved framepointer or the spilled registers
2216 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2228 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2229 /// for a 16 byte align requirement.
2231 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2232 SelectionDAG& DAG) const {
2233 MachineFunction &MF = DAG.getMachineFunction();
2234 const TargetMachine &TM = MF.getTarget();
2235 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2236 unsigned StackAlignment = TFI.getStackAlignment();
2237 uint64_t AlignMask = StackAlignment - 1;
2238 int64_t Offset = StackSize;
2239 uint64_t SlotSize = TD->getPointerSize();
2240 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2241 // Number smaller than 12 so just add the difference.
2242 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2244 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2245 Offset = ((~AlignMask) & Offset) + StackAlignment +
2246 (StackAlignment-SlotSize);
2251 /// MatchingStackOffset - Return true if the given stack call argument is
2252 /// already available in the same position (relatively) of the caller's
2253 /// incoming argument stack.
2255 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2256 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2257 const X86InstrInfo *TII) {
2258 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2260 if (Arg.getOpcode() == ISD::CopyFromReg) {
2261 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2262 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2264 MachineInstr *Def = MRI->getVRegDef(VR);
2267 if (!Flags.isByVal()) {
2268 if (!TII->isLoadFromStackSlot(Def, FI))
2271 unsigned Opcode = Def->getOpcode();
2272 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2273 Def->getOperand(1).isFI()) {
2274 FI = Def->getOperand(1).getIndex();
2275 Bytes = Flags.getByValSize();
2279 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2280 if (Flags.isByVal())
2281 // ByVal argument is passed in as a pointer but it's now being
2282 // dereferenced. e.g.
2283 // define @foo(%struct.X* %A) {
2284 // tail call @bar(%struct.X* byval %A)
2287 SDValue Ptr = Ld->getBasePtr();
2288 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2291 FI = FINode->getIndex();
2295 assert(FI != INT_MAX);
2296 if (!MFI->isFixedObjectIndex(FI))
2298 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2301 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2302 /// for tail call optimization. Targets which want to do tail call
2303 /// optimization should implement this function.
2305 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2306 CallingConv::ID CalleeCC,
2308 bool isCalleeStructRet,
2309 bool isCallerStructRet,
2310 const SmallVectorImpl<ISD::OutputArg> &Outs,
2311 const SmallVectorImpl<ISD::InputArg> &Ins,
2312 SelectionDAG& DAG) const {
2313 if (!IsTailCallConvention(CalleeCC) &&
2314 CalleeCC != CallingConv::C)
2317 // If -tailcallopt is specified, make fastcc functions tail-callable.
2318 const MachineFunction &MF = DAG.getMachineFunction();
2319 const Function *CallerF = DAG.getMachineFunction().getFunction();
2320 CallingConv::ID CallerCC = CallerF->getCallingConv();
2321 bool CCMatch = CallerCC == CalleeCC;
2323 if (GuaranteedTailCallOpt) {
2324 if (IsTailCallConvention(CalleeCC) && CCMatch)
2329 // Look for obvious safe cases to perform tail call optimization that do not
2330 // require ABI changes. This is what gcc calls sibcall.
2332 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2333 // emit a special epilogue.
2334 if (RegInfo->needsStackRealignment(MF))
2337 // Do not sibcall optimize vararg calls unless the call site is not passing any
2339 if (isVarArg && !Outs.empty())
2342 // Also avoid sibcall optimization if either caller or callee uses struct
2343 // return semantics.
2344 if (isCalleeStructRet || isCallerStructRet)
2347 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2348 // Therefore if it's not used by the call it is not safe to optimize this into
2350 bool Unused = false;
2351 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2358 SmallVector<CCValAssign, 16> RVLocs;
2359 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2360 RVLocs, *DAG.getContext());
2361 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2362 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2363 CCValAssign &VA = RVLocs[i];
2364 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2369 // If the calling conventions do not match, then we'd better make sure the
2370 // results are returned in the same way as what the caller expects.
2372 SmallVector<CCValAssign, 16> RVLocs1;
2373 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2374 RVLocs1, *DAG.getContext());
2375 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2377 SmallVector<CCValAssign, 16> RVLocs2;
2378 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2379 RVLocs2, *DAG.getContext());
2380 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2382 if (RVLocs1.size() != RVLocs2.size())
2384 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2385 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2387 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2389 if (RVLocs1[i].isRegLoc()) {
2390 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2393 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2399 // If the callee takes no arguments then go on to check the results of the
2401 if (!Outs.empty()) {
2402 // Check if stack adjustment is needed. For now, do not do this if any
2403 // argument is passed on the stack.
2404 SmallVector<CCValAssign, 16> ArgLocs;
2405 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2406 ArgLocs, *DAG.getContext());
2407 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2408 if (CCInfo.getNextStackOffset()) {
2409 MachineFunction &MF = DAG.getMachineFunction();
2410 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2412 if (Subtarget->isTargetWin64())
2413 // Win64 ABI has additional complications.
2416 // Check if the arguments are already laid out in the right way as
2417 // the caller's fixed stack objects.
2418 MachineFrameInfo *MFI = MF.getFrameInfo();
2419 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2420 const X86InstrInfo *TII =
2421 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2422 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2423 CCValAssign &VA = ArgLocs[i];
2424 SDValue Arg = Outs[i].Val;
2425 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2426 if (VA.getLocInfo() == CCValAssign::Indirect)
2428 if (!VA.isRegLoc()) {
2429 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2436 // If the tailcall address may be in a register, then make sure it's
2437 // possible to register allocate for it. In 32-bit, the call address can
2438 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2439 // callee-saved registers are restored. In 64-bit, it's RAX, RCX, RDX, RSI,
2440 // RDI, R8, R9, R11.
2441 if (!isa<GlobalAddressSDNode>(Callee) &&
2442 !isa<ExternalSymbolSDNode>(Callee)) {
2443 unsigned Limit = Subtarget->is64Bit() ? 8 : 3;
2444 unsigned NumInRegs = 0;
2445 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2446 CCValAssign &VA = ArgLocs[i];
2447 if (VA.isRegLoc()) {
2448 if (++NumInRegs == Limit)
2459 X86TargetLowering::createFastISel(MachineFunction &mf,
2460 DenseMap<const Value *, unsigned> &vm,
2461 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2462 DenseMap<const AllocaInst *, int> &am,
2463 std::vector<std::pair<MachineInstr*, unsigned> > &pn
2465 , SmallSet<const Instruction *, 8> &cil
2468 return X86::createFastISel(mf, vm, bm, am, pn
2476 //===----------------------------------------------------------------------===//
2477 // Other Lowering Hooks
2478 //===----------------------------------------------------------------------===//
2481 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2482 MachineFunction &MF = DAG.getMachineFunction();
2483 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2484 int ReturnAddrIndex = FuncInfo->getRAIndex();
2486 if (ReturnAddrIndex == 0) {
2487 // Set up a frame object for the return address.
2488 uint64_t SlotSize = TD->getPointerSize();
2489 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2491 FuncInfo->setRAIndex(ReturnAddrIndex);
2494 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2498 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2499 bool hasSymbolicDisplacement) {
2500 // Offset should fit into 32 bit immediate field.
2501 if (!isInt<32>(Offset))
2504 // If we don't have a symbolic displacement - we don't have any extra
2506 if (!hasSymbolicDisplacement)
2509 // FIXME: Some tweaks might be needed for medium code model.
2510 if (M != CodeModel::Small && M != CodeModel::Kernel)
2513 // For small code model we assume that latest object is 16MB before end of 31
2514 // bits boundary. We may also accept pretty large negative constants knowing
2515 // that all objects are in the positive half of address space.
2516 if (M == CodeModel::Small && Offset < 16*1024*1024)
2519 // For kernel code model we know that all object resist in the negative half
2520 // of 32bits address space. We may not accept negative offsets, since they may
2521 // be just off and we may accept pretty large positive ones.
2522 if (M == CodeModel::Kernel && Offset > 0)
2528 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2529 /// specific condition code, returning the condition code and the LHS/RHS of the
2530 /// comparison to make.
2531 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2532 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2534 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2535 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2536 // X > -1 -> X == 0, jump !sign.
2537 RHS = DAG.getConstant(0, RHS.getValueType());
2538 return X86::COND_NS;
2539 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2540 // X < 0 -> X == 0, jump on sign.
2542 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2544 RHS = DAG.getConstant(0, RHS.getValueType());
2545 return X86::COND_LE;
2549 switch (SetCCOpcode) {
2550 default: llvm_unreachable("Invalid integer condition!");
2551 case ISD::SETEQ: return X86::COND_E;
2552 case ISD::SETGT: return X86::COND_G;
2553 case ISD::SETGE: return X86::COND_GE;
2554 case ISD::SETLT: return X86::COND_L;
2555 case ISD::SETLE: return X86::COND_LE;
2556 case ISD::SETNE: return X86::COND_NE;
2557 case ISD::SETULT: return X86::COND_B;
2558 case ISD::SETUGT: return X86::COND_A;
2559 case ISD::SETULE: return X86::COND_BE;
2560 case ISD::SETUGE: return X86::COND_AE;
2564 // First determine if it is required or is profitable to flip the operands.
2566 // If LHS is a foldable load, but RHS is not, flip the condition.
2567 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2568 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2569 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2570 std::swap(LHS, RHS);
2573 switch (SetCCOpcode) {
2579 std::swap(LHS, RHS);
2583 // On a floating point condition, the flags are set as follows:
2585 // 0 | 0 | 0 | X > Y
2586 // 0 | 0 | 1 | X < Y
2587 // 1 | 0 | 0 | X == Y
2588 // 1 | 1 | 1 | unordered
2589 switch (SetCCOpcode) {
2590 default: llvm_unreachable("Condcode should be pre-legalized away");
2592 case ISD::SETEQ: return X86::COND_E;
2593 case ISD::SETOLT: // flipped
2595 case ISD::SETGT: return X86::COND_A;
2596 case ISD::SETOLE: // flipped
2598 case ISD::SETGE: return X86::COND_AE;
2599 case ISD::SETUGT: // flipped
2601 case ISD::SETLT: return X86::COND_B;
2602 case ISD::SETUGE: // flipped
2604 case ISD::SETLE: return X86::COND_BE;
2606 case ISD::SETNE: return X86::COND_NE;
2607 case ISD::SETUO: return X86::COND_P;
2608 case ISD::SETO: return X86::COND_NP;
2610 case ISD::SETUNE: return X86::COND_INVALID;
2614 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2615 /// code. Current x86 isa includes the following FP cmov instructions:
2616 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2617 static bool hasFPCMov(unsigned X86CC) {
2633 /// isFPImmLegal - Returns true if the target can instruction select the
2634 /// specified FP immediate natively. If false, the legalizer will
2635 /// materialize the FP immediate as a load from a constant pool.
2636 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2637 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2638 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2644 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2645 /// the specified range (L, H].
2646 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2647 return (Val < 0) || (Val >= Low && Val < Hi);
2650 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2651 /// specified value.
2652 static bool isUndefOrEqual(int Val, int CmpVal) {
2653 if (Val < 0 || Val == CmpVal)
2658 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2659 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2660 /// the second operand.
2661 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2662 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2663 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2664 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2665 return (Mask[0] < 2 && Mask[1] < 2);
2669 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2670 SmallVector<int, 8> M;
2672 return ::isPSHUFDMask(M, N->getValueType(0));
2675 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2676 /// is suitable for input to PSHUFHW.
2677 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2678 if (VT != MVT::v8i16)
2681 // Lower quadword copied in order or undef.
2682 for (int i = 0; i != 4; ++i)
2683 if (Mask[i] >= 0 && Mask[i] != i)
2686 // Upper quadword shuffled.
2687 for (int i = 4; i != 8; ++i)
2688 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2694 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2695 SmallVector<int, 8> M;
2697 return ::isPSHUFHWMask(M, N->getValueType(0));
2700 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2701 /// is suitable for input to PSHUFLW.
2702 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2703 if (VT != MVT::v8i16)
2706 // Upper quadword copied in order.
2707 for (int i = 4; i != 8; ++i)
2708 if (Mask[i] >= 0 && Mask[i] != i)
2711 // Lower quadword shuffled.
2712 for (int i = 0; i != 4; ++i)
2719 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2720 SmallVector<int, 8> M;
2722 return ::isPSHUFLWMask(M, N->getValueType(0));
2725 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2726 /// is suitable for input to PALIGNR.
2727 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2729 int i, e = VT.getVectorNumElements();
2731 // Do not handle v2i64 / v2f64 shuffles with palignr.
2732 if (e < 4 || !hasSSSE3)
2735 for (i = 0; i != e; ++i)
2739 // All undef, not a palignr.
2743 // Determine if it's ok to perform a palignr with only the LHS, since we
2744 // don't have access to the actual shuffle elements to see if RHS is undef.
2745 bool Unary = Mask[i] < (int)e;
2746 bool NeedsUnary = false;
2748 int s = Mask[i] - i;
2750 // Check the rest of the elements to see if they are consecutive.
2751 for (++i; i != e; ++i) {
2756 Unary = Unary && (m < (int)e);
2757 NeedsUnary = NeedsUnary || (m < s);
2759 if (NeedsUnary && !Unary)
2761 if (Unary && m != ((s+i) & (e-1)))
2763 if (!Unary && m != (s+i))
2769 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2770 SmallVector<int, 8> M;
2772 return ::isPALIGNRMask(M, N->getValueType(0), true);
2775 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2776 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2777 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2778 int NumElems = VT.getVectorNumElements();
2779 if (NumElems != 2 && NumElems != 4)
2782 int Half = NumElems / 2;
2783 for (int i = 0; i < Half; ++i)
2784 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2786 for (int i = Half; i < NumElems; ++i)
2787 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2793 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2794 SmallVector<int, 8> M;
2796 return ::isSHUFPMask(M, N->getValueType(0));
2799 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2800 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2801 /// half elements to come from vector 1 (which would equal the dest.) and
2802 /// the upper half to come from vector 2.
2803 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2804 int NumElems = VT.getVectorNumElements();
2806 if (NumElems != 2 && NumElems != 4)
2809 int Half = NumElems / 2;
2810 for (int i = 0; i < Half; ++i)
2811 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2813 for (int i = Half; i < NumElems; ++i)
2814 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2819 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2820 SmallVector<int, 8> M;
2822 return isCommutedSHUFPMask(M, N->getValueType(0));
2825 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2826 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2827 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2828 if (N->getValueType(0).getVectorNumElements() != 4)
2831 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2832 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2833 isUndefOrEqual(N->getMaskElt(1), 7) &&
2834 isUndefOrEqual(N->getMaskElt(2), 2) &&
2835 isUndefOrEqual(N->getMaskElt(3), 3);
2838 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2839 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2841 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2842 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2847 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2848 isUndefOrEqual(N->getMaskElt(1), 3) &&
2849 isUndefOrEqual(N->getMaskElt(2), 2) &&
2850 isUndefOrEqual(N->getMaskElt(3), 3);
2853 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2854 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2855 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2856 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2858 if (NumElems != 2 && NumElems != 4)
2861 for (unsigned i = 0; i < NumElems/2; ++i)
2862 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2865 for (unsigned i = NumElems/2; i < NumElems; ++i)
2866 if (!isUndefOrEqual(N->getMaskElt(i), i))
2872 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2873 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2874 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2875 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2877 if (NumElems != 2 && NumElems != 4)
2880 for (unsigned i = 0; i < NumElems/2; ++i)
2881 if (!isUndefOrEqual(N->getMaskElt(i), i))
2884 for (unsigned i = 0; i < NumElems/2; ++i)
2885 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2891 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2892 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2893 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2894 bool V2IsSplat = false) {
2895 int NumElts = VT.getVectorNumElements();
2896 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2899 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2901 int BitI1 = Mask[i+1];
2902 if (!isUndefOrEqual(BitI, j))
2905 if (!isUndefOrEqual(BitI1, NumElts))
2908 if (!isUndefOrEqual(BitI1, j + NumElts))
2915 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2916 SmallVector<int, 8> M;
2918 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2921 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2922 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2923 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2924 bool V2IsSplat = false) {
2925 int NumElts = VT.getVectorNumElements();
2926 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2929 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2931 int BitI1 = Mask[i+1];
2932 if (!isUndefOrEqual(BitI, j + NumElts/2))
2935 if (isUndefOrEqual(BitI1, NumElts))
2938 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2945 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2946 SmallVector<int, 8> M;
2948 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2951 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2952 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2954 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2955 int NumElems = VT.getVectorNumElements();
2956 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2959 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2961 int BitI1 = Mask[i+1];
2962 if (!isUndefOrEqual(BitI, j))
2964 if (!isUndefOrEqual(BitI1, j))
2970 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2971 SmallVector<int, 8> M;
2973 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2976 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2977 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2979 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2980 int NumElems = VT.getVectorNumElements();
2981 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2984 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2986 int BitI1 = Mask[i+1];
2987 if (!isUndefOrEqual(BitI, j))
2989 if (!isUndefOrEqual(BitI1, j))
2995 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2996 SmallVector<int, 8> M;
2998 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3001 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3002 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3003 /// MOVSD, and MOVD, i.e. setting the lowest element.
3004 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3005 if (VT.getVectorElementType().getSizeInBits() < 32)
3008 int NumElts = VT.getVectorNumElements();
3010 if (!isUndefOrEqual(Mask[0], NumElts))
3013 for (int i = 1; i < NumElts; ++i)
3014 if (!isUndefOrEqual(Mask[i], i))
3020 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3021 SmallVector<int, 8> M;
3023 return ::isMOVLMask(M, N->getValueType(0));
3026 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3027 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3028 /// element of vector 2 and the other elements to come from vector 1 in order.
3029 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3030 bool V2IsSplat = false, bool V2IsUndef = false) {
3031 int NumOps = VT.getVectorNumElements();
3032 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3035 if (!isUndefOrEqual(Mask[0], 0))
3038 for (int i = 1; i < NumOps; ++i)
3039 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3040 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3041 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3047 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3048 bool V2IsUndef = false) {
3049 SmallVector<int, 8> M;
3051 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3054 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3055 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3056 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3057 if (N->getValueType(0).getVectorNumElements() != 4)
3060 // Expect 1, 1, 3, 3
3061 for (unsigned i = 0; i < 2; ++i) {
3062 int Elt = N->getMaskElt(i);
3063 if (Elt >= 0 && Elt != 1)
3068 for (unsigned i = 2; i < 4; ++i) {
3069 int Elt = N->getMaskElt(i);
3070 if (Elt >= 0 && Elt != 3)
3075 // Don't use movshdup if it can be done with a shufps.
3076 // FIXME: verify that matching u, u, 3, 3 is what we want.
3080 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3081 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3082 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3083 if (N->getValueType(0).getVectorNumElements() != 4)
3086 // Expect 0, 0, 2, 2
3087 for (unsigned i = 0; i < 2; ++i)
3088 if (N->getMaskElt(i) > 0)
3092 for (unsigned i = 2; i < 4; ++i) {
3093 int Elt = N->getMaskElt(i);
3094 if (Elt >= 0 && Elt != 2)
3099 // Don't use movsldup if it can be done with a shufps.
3103 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3104 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3105 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3106 int e = N->getValueType(0).getVectorNumElements() / 2;
3108 for (int i = 0; i < e; ++i)
3109 if (!isUndefOrEqual(N->getMaskElt(i), i))
3111 for (int i = 0; i < e; ++i)
3112 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3117 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3118 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3119 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3120 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3121 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3123 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3125 for (int i = 0; i < NumOperands; ++i) {
3126 int Val = SVOp->getMaskElt(NumOperands-i-1);
3127 if (Val < 0) Val = 0;
3128 if (Val >= NumOperands) Val -= NumOperands;
3130 if (i != NumOperands - 1)
3136 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3137 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3138 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3139 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3141 // 8 nodes, but we only care about the last 4.
3142 for (unsigned i = 7; i >= 4; --i) {
3143 int Val = SVOp->getMaskElt(i);
3152 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3153 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3154 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3155 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3157 // 8 nodes, but we only care about the first 4.
3158 for (int i = 3; i >= 0; --i) {
3159 int Val = SVOp->getMaskElt(i);
3168 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3169 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3170 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3171 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3172 EVT VVT = N->getValueType(0);
3173 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3177 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3178 Val = SVOp->getMaskElt(i);
3182 return (Val - i) * EltSize;
3185 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3187 bool X86::isZeroNode(SDValue Elt) {
3188 return ((isa<ConstantSDNode>(Elt) &&
3189 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3190 (isa<ConstantFPSDNode>(Elt) &&
3191 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3194 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3195 /// their permute mask.
3196 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3197 SelectionDAG &DAG) {
3198 EVT VT = SVOp->getValueType(0);
3199 unsigned NumElems = VT.getVectorNumElements();
3200 SmallVector<int, 8> MaskVec;
3202 for (unsigned i = 0; i != NumElems; ++i) {
3203 int idx = SVOp->getMaskElt(i);
3205 MaskVec.push_back(idx);
3206 else if (idx < (int)NumElems)
3207 MaskVec.push_back(idx + NumElems);
3209 MaskVec.push_back(idx - NumElems);
3211 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3212 SVOp->getOperand(0), &MaskVec[0]);
3215 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3216 /// the two vector operands have swapped position.
3217 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3218 unsigned NumElems = VT.getVectorNumElements();
3219 for (unsigned i = 0; i != NumElems; ++i) {
3223 else if (idx < (int)NumElems)
3224 Mask[i] = idx + NumElems;
3226 Mask[i] = idx - NumElems;
3230 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3231 /// match movhlps. The lower half elements should come from upper half of
3232 /// V1 (and in order), and the upper half elements should come from the upper
3233 /// half of V2 (and in order).
3234 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3235 if (Op->getValueType(0).getVectorNumElements() != 4)
3237 for (unsigned i = 0, e = 2; i != e; ++i)
3238 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3240 for (unsigned i = 2; i != 4; ++i)
3241 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3246 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3247 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3249 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3250 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3252 N = N->getOperand(0).getNode();
3253 if (!ISD::isNON_EXTLoad(N))
3256 *LD = cast<LoadSDNode>(N);
3260 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3261 /// match movlp{s|d}. The lower half elements should come from lower half of
3262 /// V1 (and in order), and the upper half elements should come from the upper
3263 /// half of V2 (and in order). And since V1 will become the source of the
3264 /// MOVLP, it must be either a vector load or a scalar load to vector.
3265 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3266 ShuffleVectorSDNode *Op) {
3267 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3269 // Is V2 is a vector load, don't do this transformation. We will try to use
3270 // load folding shufps op.
3271 if (ISD::isNON_EXTLoad(V2))
3274 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3276 if (NumElems != 2 && NumElems != 4)
3278 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3279 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3281 for (unsigned i = NumElems/2; i != NumElems; ++i)
3282 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3287 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3289 static bool isSplatVector(SDNode *N) {
3290 if (N->getOpcode() != ISD::BUILD_VECTOR)
3293 SDValue SplatValue = N->getOperand(0);
3294 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3295 if (N->getOperand(i) != SplatValue)
3300 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3301 /// to an zero vector.
3302 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3303 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3304 SDValue V1 = N->getOperand(0);
3305 SDValue V2 = N->getOperand(1);
3306 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3307 for (unsigned i = 0; i != NumElems; ++i) {
3308 int Idx = N->getMaskElt(i);
3309 if (Idx >= (int)NumElems) {
3310 unsigned Opc = V2.getOpcode();
3311 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3313 if (Opc != ISD::BUILD_VECTOR ||
3314 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3316 } else if (Idx >= 0) {
3317 unsigned Opc = V1.getOpcode();
3318 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3320 if (Opc != ISD::BUILD_VECTOR ||
3321 !X86::isZeroNode(V1.getOperand(Idx)))
3328 /// getZeroVector - Returns a vector of specified type with all zero elements.
3330 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3332 assert(VT.isVector() && "Expected a vector type");
3334 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3335 // type. This ensures they get CSE'd.
3337 if (VT.getSizeInBits() == 64) { // MMX
3338 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3339 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3340 } else if (HasSSE2) { // SSE2
3341 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3342 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3344 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3345 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3347 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3350 /// getOnesVector - Returns a vector of specified type with all bits set.
3352 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3353 assert(VT.isVector() && "Expected a vector type");
3355 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3356 // type. This ensures they get CSE'd.
3357 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3359 if (VT.getSizeInBits() == 64) // MMX
3360 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3362 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3363 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3367 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3368 /// that point to V2 points to its first element.
3369 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3370 EVT VT = SVOp->getValueType(0);
3371 unsigned NumElems = VT.getVectorNumElements();
3373 bool Changed = false;
3374 SmallVector<int, 8> MaskVec;
3375 SVOp->getMask(MaskVec);
3377 for (unsigned i = 0; i != NumElems; ++i) {
3378 if (MaskVec[i] > (int)NumElems) {
3379 MaskVec[i] = NumElems;
3384 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3385 SVOp->getOperand(1), &MaskVec[0]);
3386 return SDValue(SVOp, 0);
3389 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3390 /// operation of specified width.
3391 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3393 unsigned NumElems = VT.getVectorNumElements();
3394 SmallVector<int, 8> Mask;
3395 Mask.push_back(NumElems);
3396 for (unsigned i = 1; i != NumElems; ++i)
3398 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3401 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3402 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3404 unsigned NumElems = VT.getVectorNumElements();
3405 SmallVector<int, 8> Mask;
3406 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3408 Mask.push_back(i + NumElems);
3410 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3413 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3414 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3416 unsigned NumElems = VT.getVectorNumElements();
3417 unsigned Half = NumElems/2;
3418 SmallVector<int, 8> Mask;
3419 for (unsigned i = 0; i != Half; ++i) {
3420 Mask.push_back(i + Half);
3421 Mask.push_back(i + NumElems + Half);
3423 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3426 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3427 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3429 if (SV->getValueType(0).getVectorNumElements() <= 4)
3430 return SDValue(SV, 0);
3432 EVT PVT = MVT::v4f32;
3433 EVT VT = SV->getValueType(0);
3434 DebugLoc dl = SV->getDebugLoc();
3435 SDValue V1 = SV->getOperand(0);
3436 int NumElems = VT.getVectorNumElements();
3437 int EltNo = SV->getSplatIndex();
3439 // unpack elements to the correct location
3440 while (NumElems > 4) {
3441 if (EltNo < NumElems/2) {
3442 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3444 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3445 EltNo -= NumElems/2;
3450 // Perform the splat.
3451 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3452 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3453 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3454 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3457 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3458 /// vector of zero or undef vector. This produces a shuffle where the low
3459 /// element of V2 is swizzled into the zero/undef vector, landing at element
3460 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3461 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3462 bool isZero, bool HasSSE2,
3463 SelectionDAG &DAG) {
3464 EVT VT = V2.getValueType();
3466 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3467 unsigned NumElems = VT.getVectorNumElements();
3468 SmallVector<int, 16> MaskVec;
3469 for (unsigned i = 0; i != NumElems; ++i)
3470 // If this is the insertion idx, put the low elt of V2 here.
3471 MaskVec.push_back(i == Idx ? NumElems : i);
3472 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3475 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3476 /// a shuffle that is zero.
3478 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3479 bool Low, SelectionDAG &DAG) {
3480 unsigned NumZeros = 0;
3481 for (int i = 0; i < NumElems; ++i) {
3482 unsigned Index = Low ? i : NumElems-i-1;
3483 int Idx = SVOp->getMaskElt(Index);
3488 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3489 if (Elt.getNode() && X86::isZeroNode(Elt))
3497 /// isVectorShift - Returns true if the shuffle can be implemented as a
3498 /// logical left or right shift of a vector.
3499 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3500 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3501 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3502 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3505 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3508 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3512 bool SeenV1 = false;
3513 bool SeenV2 = false;
3514 for (unsigned i = NumZeros; i < NumElems; ++i) {
3515 unsigned Val = isLeft ? (i - NumZeros) : i;
3516 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3519 unsigned Idx = (unsigned) Idx_;
3529 if (SeenV1 && SeenV2)
3532 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3538 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3540 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3541 unsigned NumNonZero, unsigned NumZero,
3543 const TargetLowering &TLI) {
3547 DebugLoc dl = Op.getDebugLoc();
3550 for (unsigned i = 0; i < 16; ++i) {
3551 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3552 if (ThisIsNonZero && First) {
3554 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3556 V = DAG.getUNDEF(MVT::v8i16);
3561 SDValue ThisElt(0, 0), LastElt(0, 0);
3562 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3563 if (LastIsNonZero) {
3564 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3565 MVT::i16, Op.getOperand(i-1));
3567 if (ThisIsNonZero) {
3568 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3569 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3570 ThisElt, DAG.getConstant(8, MVT::i8));
3572 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3576 if (ThisElt.getNode())
3577 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3578 DAG.getIntPtrConstant(i/2));
3582 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3585 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3587 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3588 unsigned NumNonZero, unsigned NumZero,
3590 const TargetLowering &TLI) {
3594 DebugLoc dl = Op.getDebugLoc();
3597 for (unsigned i = 0; i < 8; ++i) {
3598 bool isNonZero = (NonZeros & (1 << i)) != 0;
3602 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3604 V = DAG.getUNDEF(MVT::v8i16);
3607 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3608 MVT::v8i16, V, Op.getOperand(i),
3609 DAG.getIntPtrConstant(i));
3616 /// getVShift - Return a vector logical shift node.
3618 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3619 unsigned NumBits, SelectionDAG &DAG,
3620 const TargetLowering &TLI, DebugLoc dl) {
3621 bool isMMX = VT.getSizeInBits() == 64;
3622 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3623 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3624 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3625 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3626 DAG.getNode(Opc, dl, ShVT, SrcOp,
3627 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3631 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3632 SelectionDAG &DAG) const {
3634 // Check if the scalar load can be widened into a vector load. And if
3635 // the address is "base + cst" see if the cst can be "absorbed" into
3636 // the shuffle mask.
3637 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3638 SDValue Ptr = LD->getBasePtr();
3639 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3641 EVT PVT = LD->getValueType(0);
3642 if (PVT != MVT::i32 && PVT != MVT::f32)
3647 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3648 FI = FINode->getIndex();
3650 } else if (Ptr.getOpcode() == ISD::ADD &&
3651 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3652 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3653 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3654 Offset = Ptr.getConstantOperandVal(1);
3655 Ptr = Ptr.getOperand(0);
3660 SDValue Chain = LD->getChain();
3661 // Make sure the stack object alignment is at least 16.
3662 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3663 if (DAG.InferPtrAlignment(Ptr) < 16) {
3664 if (MFI->isFixedObjectIndex(FI)) {
3665 // Can't change the alignment. FIXME: It's possible to compute
3666 // the exact stack offset and reference FI + adjust offset instead.
3667 // If someone *really* cares about this. That's the way to implement it.
3670 MFI->setObjectAlignment(FI, 16);
3674 // (Offset % 16) must be multiple of 4. Then address is then
3675 // Ptr + (Offset & ~15).
3678 if ((Offset % 16) & 3)
3680 int64_t StartOffset = Offset & ~15;
3682 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3683 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3685 int EltNo = (Offset - StartOffset) >> 2;
3686 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3687 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3688 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3690 // Canonicalize it to a v4i32 shuffle.
3691 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3692 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3693 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3694 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3700 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3701 /// vector of type 'VT', see if the elements can be replaced by a single large
3702 /// load which has the same value as a build_vector whose operands are 'elts'.
3704 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3706 /// FIXME: we'd also like to handle the case where the last elements are zero
3707 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3708 /// There's even a handy isZeroNode for that purpose.
3709 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3710 DebugLoc &dl, SelectionDAG &DAG) {
3711 EVT EltVT = VT.getVectorElementType();
3712 unsigned NumElems = Elts.size();
3714 LoadSDNode *LDBase = NULL;
3715 unsigned LastLoadedElt = -1U;
3717 // For each element in the initializer, see if we've found a load or an undef.
3718 // If we don't find an initial load element, or later load elements are
3719 // non-consecutive, bail out.
3720 for (unsigned i = 0; i < NumElems; ++i) {
3721 SDValue Elt = Elts[i];
3723 if (!Elt.getNode() ||
3724 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3727 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3729 LDBase = cast<LoadSDNode>(Elt.getNode());
3733 if (Elt.getOpcode() == ISD::UNDEF)
3736 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3737 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3742 // If we have found an entire vector of loads and undefs, then return a large
3743 // load of the entire vector width starting at the base pointer. If we found
3744 // consecutive loads for the low half, generate a vzext_load node.
3745 if (LastLoadedElt == NumElems - 1) {
3746 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3747 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3748 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3749 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3750 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3751 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3752 LDBase->isVolatile(), LDBase->isNonTemporal(),
3753 LDBase->getAlignment());
3754 } else if (NumElems == 4 && LastLoadedElt == 1) {
3755 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3756 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3757 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3758 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3764 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3765 DebugLoc dl = Op.getDebugLoc();
3766 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3767 if (ISD::isBuildVectorAllZeros(Op.getNode())
3768 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3769 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3770 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3771 // eliminated on x86-32 hosts.
3772 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3775 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3776 return getOnesVector(Op.getValueType(), DAG, dl);
3777 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3780 EVT VT = Op.getValueType();
3781 EVT ExtVT = VT.getVectorElementType();
3782 unsigned EVTBits = ExtVT.getSizeInBits();
3784 unsigned NumElems = Op.getNumOperands();
3785 unsigned NumZero = 0;
3786 unsigned NumNonZero = 0;
3787 unsigned NonZeros = 0;
3788 bool IsAllConstants = true;
3789 SmallSet<SDValue, 8> Values;
3790 for (unsigned i = 0; i < NumElems; ++i) {
3791 SDValue Elt = Op.getOperand(i);
3792 if (Elt.getOpcode() == ISD::UNDEF)
3795 if (Elt.getOpcode() != ISD::Constant &&
3796 Elt.getOpcode() != ISD::ConstantFP)
3797 IsAllConstants = false;
3798 if (X86::isZeroNode(Elt))
3801 NonZeros |= (1 << i);
3806 if (NumNonZero == 0) {
3807 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3808 return DAG.getUNDEF(VT);
3811 // Special case for single non-zero, non-undef, element.
3812 if (NumNonZero == 1) {
3813 unsigned Idx = CountTrailingZeros_32(NonZeros);
3814 SDValue Item = Op.getOperand(Idx);
3816 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3817 // the value are obviously zero, truncate the value to i32 and do the
3818 // insertion that way. Only do this if the value is non-constant or if the
3819 // value is a constant being inserted into element 0. It is cheaper to do
3820 // a constant pool load than it is to do a movd + shuffle.
3821 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3822 (!IsAllConstants || Idx == 0)) {
3823 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3824 // Handle MMX and SSE both.
3825 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3826 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3828 // Truncate the value (which may itself be a constant) to i32, and
3829 // convert it to a vector with movd (S2V+shuffle to zero extend).
3830 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3831 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3832 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3833 Subtarget->hasSSE2(), DAG);
3835 // Now we have our 32-bit value zero extended in the low element of
3836 // a vector. If Idx != 0, swizzle it into place.
3838 SmallVector<int, 4> Mask;
3839 Mask.push_back(Idx);
3840 for (unsigned i = 1; i != VecElts; ++i)
3842 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3843 DAG.getUNDEF(Item.getValueType()),
3846 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3850 // If we have a constant or non-constant insertion into the low element of
3851 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3852 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3853 // depending on what the source datatype is.
3856 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3857 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3858 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3859 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3860 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3861 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3863 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3864 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3865 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3866 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3867 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3868 Subtarget->hasSSE2(), DAG);
3869 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3873 // Is it a vector logical left shift?
3874 if (NumElems == 2 && Idx == 1 &&
3875 X86::isZeroNode(Op.getOperand(0)) &&
3876 !X86::isZeroNode(Op.getOperand(1))) {
3877 unsigned NumBits = VT.getSizeInBits();
3878 return getVShift(true, VT,
3879 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3880 VT, Op.getOperand(1)),
3881 NumBits/2, DAG, *this, dl);
3884 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3887 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3888 // is a non-constant being inserted into an element other than the low one,
3889 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3890 // movd/movss) to move this into the low element, then shuffle it into
3892 if (EVTBits == 32) {
3893 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3895 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3896 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3897 Subtarget->hasSSE2(), DAG);
3898 SmallVector<int, 8> MaskVec;
3899 for (unsigned i = 0; i < NumElems; i++)
3900 MaskVec.push_back(i == Idx ? 0 : 1);
3901 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3905 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3906 if (Values.size() == 1) {
3907 if (EVTBits == 32) {
3908 // Instead of a shuffle like this:
3909 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3910 // Check if it's possible to issue this instead.
3911 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3912 unsigned Idx = CountTrailingZeros_32(NonZeros);
3913 SDValue Item = Op.getOperand(Idx);
3914 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3915 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3920 // A vector full of immediates; various special cases are already
3921 // handled, so this is best done with a single constant-pool load.
3925 // Let legalizer expand 2-wide build_vectors.
3926 if (EVTBits == 64) {
3927 if (NumNonZero == 1) {
3928 // One half is zero or undef.
3929 unsigned Idx = CountTrailingZeros_32(NonZeros);
3930 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3931 Op.getOperand(Idx));
3932 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3933 Subtarget->hasSSE2(), DAG);
3938 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3939 if (EVTBits == 8 && NumElems == 16) {
3940 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3942 if (V.getNode()) return V;
3945 if (EVTBits == 16 && NumElems == 8) {
3946 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3948 if (V.getNode()) return V;
3951 // If element VT is == 32 bits, turn it into a number of shuffles.
3952 SmallVector<SDValue, 8> V;
3954 if (NumElems == 4 && NumZero > 0) {
3955 for (unsigned i = 0; i < 4; ++i) {
3956 bool isZero = !(NonZeros & (1 << i));
3958 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3960 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3963 for (unsigned i = 0; i < 2; ++i) {
3964 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3967 V[i] = V[i*2]; // Must be a zero vector.
3970 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3973 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3976 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3981 SmallVector<int, 8> MaskVec;
3982 bool Reverse = (NonZeros & 0x3) == 2;
3983 for (unsigned i = 0; i < 2; ++i)
3984 MaskVec.push_back(Reverse ? 1-i : i);
3985 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3986 for (unsigned i = 0; i < 2; ++i)
3987 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3988 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3991 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
3992 // Check for a build vector of consecutive loads.
3993 for (unsigned i = 0; i < NumElems; ++i)
3994 V[i] = Op.getOperand(i);
3996 // Check for elements which are consecutive loads.
3997 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4001 // For SSE 4.1, use inserts into undef.
4002 if (getSubtarget()->hasSSE41()) {
4003 V[0] = DAG.getUNDEF(VT);
4004 for (unsigned i = 0; i < NumElems; ++i)
4005 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4006 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4007 Op.getOperand(i), DAG.getIntPtrConstant(i));
4011 // Otherwise, expand into a number of unpckl*
4013 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4014 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4015 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4016 for (unsigned i = 0; i < NumElems; ++i)
4017 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4019 while (NumElems != 0) {
4020 for (unsigned i = 0; i < NumElems; ++i)
4021 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4030 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4031 // We support concatenate two MMX registers and place them in a MMX
4032 // register. This is better than doing a stack convert.
4033 DebugLoc dl = Op.getDebugLoc();
4034 EVT ResVT = Op.getValueType();
4035 assert(Op.getNumOperands() == 2);
4036 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4037 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4039 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4040 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4041 InVec = Op.getOperand(1);
4042 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4043 unsigned NumElts = ResVT.getVectorNumElements();
4044 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4045 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4046 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4048 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4049 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4050 Mask[0] = 0; Mask[1] = 2;
4051 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4053 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4056 // v8i16 shuffles - Prefer shuffles in the following order:
4057 // 1. [all] pshuflw, pshufhw, optional move
4058 // 2. [ssse3] 1 x pshufb
4059 // 3. [ssse3] 2 x pshufb + 1 x por
4060 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4062 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4064 const X86TargetLowering &TLI) {
4065 SDValue V1 = SVOp->getOperand(0);
4066 SDValue V2 = SVOp->getOperand(1);
4067 DebugLoc dl = SVOp->getDebugLoc();
4068 SmallVector<int, 8> MaskVals;
4070 // Determine if more than 1 of the words in each of the low and high quadwords
4071 // of the result come from the same quadword of one of the two inputs. Undef
4072 // mask values count as coming from any quadword, for better codegen.
4073 SmallVector<unsigned, 4> LoQuad(4);
4074 SmallVector<unsigned, 4> HiQuad(4);
4075 BitVector InputQuads(4);
4076 for (unsigned i = 0; i < 8; ++i) {
4077 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4078 int EltIdx = SVOp->getMaskElt(i);
4079 MaskVals.push_back(EltIdx);
4088 InputQuads.set(EltIdx / 4);
4091 int BestLoQuad = -1;
4092 unsigned MaxQuad = 1;
4093 for (unsigned i = 0; i < 4; ++i) {
4094 if (LoQuad[i] > MaxQuad) {
4096 MaxQuad = LoQuad[i];
4100 int BestHiQuad = -1;
4102 for (unsigned i = 0; i < 4; ++i) {
4103 if (HiQuad[i] > MaxQuad) {
4105 MaxQuad = HiQuad[i];
4109 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4110 // of the two input vectors, shuffle them into one input vector so only a
4111 // single pshufb instruction is necessary. If There are more than 2 input
4112 // quads, disable the next transformation since it does not help SSSE3.
4113 bool V1Used = InputQuads[0] || InputQuads[1];
4114 bool V2Used = InputQuads[2] || InputQuads[3];
4115 if (TLI.getSubtarget()->hasSSSE3()) {
4116 if (InputQuads.count() == 2 && V1Used && V2Used) {
4117 BestLoQuad = InputQuads.find_first();
4118 BestHiQuad = InputQuads.find_next(BestLoQuad);
4120 if (InputQuads.count() > 2) {
4126 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4127 // the shuffle mask. If a quad is scored as -1, that means that it contains
4128 // words from all 4 input quadwords.
4130 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4131 SmallVector<int, 8> MaskV;
4132 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4133 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4134 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4135 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4136 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4137 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4139 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4140 // source words for the shuffle, to aid later transformations.
4141 bool AllWordsInNewV = true;
4142 bool InOrder[2] = { true, true };
4143 for (unsigned i = 0; i != 8; ++i) {
4144 int idx = MaskVals[i];
4146 InOrder[i/4] = false;
4147 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4149 AllWordsInNewV = false;
4153 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4154 if (AllWordsInNewV) {
4155 for (int i = 0; i != 8; ++i) {
4156 int idx = MaskVals[i];
4159 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4160 if ((idx != i) && idx < 4)
4162 if ((idx != i) && idx > 3)
4171 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4172 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4173 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4174 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4175 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4179 // If we have SSSE3, and all words of the result are from 1 input vector,
4180 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4181 // is present, fall back to case 4.
4182 if (TLI.getSubtarget()->hasSSSE3()) {
4183 SmallVector<SDValue,16> pshufbMask;
4185 // If we have elements from both input vectors, set the high bit of the
4186 // shuffle mask element to zero out elements that come from V2 in the V1
4187 // mask, and elements that come from V1 in the V2 mask, so that the two
4188 // results can be OR'd together.
4189 bool TwoInputs = V1Used && V2Used;
4190 for (unsigned i = 0; i != 8; ++i) {
4191 int EltIdx = MaskVals[i] * 2;
4192 if (TwoInputs && (EltIdx >= 16)) {
4193 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4194 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4197 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4198 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4200 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4201 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4202 DAG.getNode(ISD::BUILD_VECTOR, dl,
4203 MVT::v16i8, &pshufbMask[0], 16));
4205 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4207 // Calculate the shuffle mask for the second input, shuffle it, and
4208 // OR it with the first shuffled input.
4210 for (unsigned i = 0; i != 8; ++i) {
4211 int EltIdx = MaskVals[i] * 2;
4213 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4214 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4217 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4218 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4220 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4221 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4222 DAG.getNode(ISD::BUILD_VECTOR, dl,
4223 MVT::v16i8, &pshufbMask[0], 16));
4224 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4225 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4228 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4229 // and update MaskVals with new element order.
4230 BitVector InOrder(8);
4231 if (BestLoQuad >= 0) {
4232 SmallVector<int, 8> MaskV;
4233 for (int i = 0; i != 4; ++i) {
4234 int idx = MaskVals[i];
4236 MaskV.push_back(-1);
4238 } else if ((idx / 4) == BestLoQuad) {
4239 MaskV.push_back(idx & 3);
4242 MaskV.push_back(-1);
4245 for (unsigned i = 4; i != 8; ++i)
4247 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4251 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4252 // and update MaskVals with the new element order.
4253 if (BestHiQuad >= 0) {
4254 SmallVector<int, 8> MaskV;
4255 for (unsigned i = 0; i != 4; ++i)
4257 for (unsigned i = 4; i != 8; ++i) {
4258 int idx = MaskVals[i];
4260 MaskV.push_back(-1);
4262 } else if ((idx / 4) == BestHiQuad) {
4263 MaskV.push_back((idx & 3) + 4);
4266 MaskV.push_back(-1);
4269 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4273 // In case BestHi & BestLo were both -1, which means each quadword has a word
4274 // from each of the four input quadwords, calculate the InOrder bitvector now
4275 // before falling through to the insert/extract cleanup.
4276 if (BestLoQuad == -1 && BestHiQuad == -1) {
4278 for (int i = 0; i != 8; ++i)
4279 if (MaskVals[i] < 0 || MaskVals[i] == i)
4283 // The other elements are put in the right place using pextrw and pinsrw.
4284 for (unsigned i = 0; i != 8; ++i) {
4287 int EltIdx = MaskVals[i];
4290 SDValue ExtOp = (EltIdx < 8)
4291 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4292 DAG.getIntPtrConstant(EltIdx))
4293 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4294 DAG.getIntPtrConstant(EltIdx - 8));
4295 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4296 DAG.getIntPtrConstant(i));
4301 // v16i8 shuffles - Prefer shuffles in the following order:
4302 // 1. [ssse3] 1 x pshufb
4303 // 2. [ssse3] 2 x pshufb + 1 x por
4304 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4306 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4308 const X86TargetLowering &TLI) {
4309 SDValue V1 = SVOp->getOperand(0);
4310 SDValue V2 = SVOp->getOperand(1);
4311 DebugLoc dl = SVOp->getDebugLoc();
4312 SmallVector<int, 16> MaskVals;
4313 SVOp->getMask(MaskVals);
4315 // If we have SSSE3, case 1 is generated when all result bytes come from
4316 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4317 // present, fall back to case 3.
4318 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4321 for (unsigned i = 0; i < 16; ++i) {
4322 int EltIdx = MaskVals[i];
4331 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4332 if (TLI.getSubtarget()->hasSSSE3()) {
4333 SmallVector<SDValue,16> pshufbMask;
4335 // If all result elements are from one input vector, then only translate
4336 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4338 // Otherwise, we have elements from both input vectors, and must zero out
4339 // elements that come from V2 in the first mask, and V1 in the second mask
4340 // so that we can OR them together.
4341 bool TwoInputs = !(V1Only || V2Only);
4342 for (unsigned i = 0; i != 16; ++i) {
4343 int EltIdx = MaskVals[i];
4344 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4345 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4348 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4350 // If all the elements are from V2, assign it to V1 and return after
4351 // building the first pshufb.
4354 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4355 DAG.getNode(ISD::BUILD_VECTOR, dl,
4356 MVT::v16i8, &pshufbMask[0], 16));
4360 // Calculate the shuffle mask for the second input, shuffle it, and
4361 // OR it with the first shuffled input.
4363 for (unsigned i = 0; i != 16; ++i) {
4364 int EltIdx = MaskVals[i];
4366 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4369 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4371 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4372 DAG.getNode(ISD::BUILD_VECTOR, dl,
4373 MVT::v16i8, &pshufbMask[0], 16));
4374 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4377 // No SSSE3 - Calculate in place words and then fix all out of place words
4378 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4379 // the 16 different words that comprise the two doublequadword input vectors.
4380 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4381 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4382 SDValue NewV = V2Only ? V2 : V1;
4383 for (int i = 0; i != 8; ++i) {
4384 int Elt0 = MaskVals[i*2];
4385 int Elt1 = MaskVals[i*2+1];
4387 // This word of the result is all undef, skip it.
4388 if (Elt0 < 0 && Elt1 < 0)
4391 // This word of the result is already in the correct place, skip it.
4392 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4394 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4397 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4398 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4401 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4402 // using a single extract together, load it and store it.
4403 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4404 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4405 DAG.getIntPtrConstant(Elt1 / 2));
4406 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4407 DAG.getIntPtrConstant(i));
4411 // If Elt1 is defined, extract it from the appropriate source. If the
4412 // source byte is not also odd, shift the extracted word left 8 bits
4413 // otherwise clear the bottom 8 bits if we need to do an or.
4415 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4416 DAG.getIntPtrConstant(Elt1 / 2));
4417 if ((Elt1 & 1) == 0)
4418 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4419 DAG.getConstant(8, TLI.getShiftAmountTy()));
4421 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4422 DAG.getConstant(0xFF00, MVT::i16));
4424 // If Elt0 is defined, extract it from the appropriate source. If the
4425 // source byte is not also even, shift the extracted word right 8 bits. If
4426 // Elt1 was also defined, OR the extracted values together before
4427 // inserting them in the result.
4429 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4430 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4431 if ((Elt0 & 1) != 0)
4432 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4433 DAG.getConstant(8, TLI.getShiftAmountTy()));
4435 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4436 DAG.getConstant(0x00FF, MVT::i16));
4437 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4440 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4441 DAG.getIntPtrConstant(i));
4443 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4446 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4447 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4448 /// done when every pair / quad of shuffle mask elements point to elements in
4449 /// the right sequence. e.g.
4450 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4452 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4454 const TargetLowering &TLI, DebugLoc dl) {
4455 EVT VT = SVOp->getValueType(0);
4456 SDValue V1 = SVOp->getOperand(0);
4457 SDValue V2 = SVOp->getOperand(1);
4458 unsigned NumElems = VT.getVectorNumElements();
4459 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4460 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4462 switch (VT.getSimpleVT().SimpleTy) {
4463 default: assert(false && "Unexpected!");
4464 case MVT::v4f32: NewVT = MVT::v2f64; break;
4465 case MVT::v4i32: NewVT = MVT::v2i64; break;
4466 case MVT::v8i16: NewVT = MVT::v4i32; break;
4467 case MVT::v16i8: NewVT = MVT::v4i32; break;
4470 if (NewWidth == 2) {
4476 int Scale = NumElems / NewWidth;
4477 SmallVector<int, 8> MaskVec;
4478 for (unsigned i = 0; i < NumElems; i += Scale) {
4480 for (int j = 0; j < Scale; ++j) {
4481 int EltIdx = SVOp->getMaskElt(i+j);
4485 StartIdx = EltIdx - (EltIdx % Scale);
4486 if (EltIdx != StartIdx + j)
4490 MaskVec.push_back(-1);
4492 MaskVec.push_back(StartIdx / Scale);
4495 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4496 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4497 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4500 /// getVZextMovL - Return a zero-extending vector move low node.
4502 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4503 SDValue SrcOp, SelectionDAG &DAG,
4504 const X86Subtarget *Subtarget, DebugLoc dl) {
4505 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4506 LoadSDNode *LD = NULL;
4507 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4508 LD = dyn_cast<LoadSDNode>(SrcOp);
4510 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4512 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4513 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4514 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4515 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4516 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4518 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4519 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4520 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4521 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4529 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4530 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4531 DAG.getNode(ISD::BIT_CONVERT, dl,
4535 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4538 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4539 SDValue V1 = SVOp->getOperand(0);
4540 SDValue V2 = SVOp->getOperand(1);
4541 DebugLoc dl = SVOp->getDebugLoc();
4542 EVT VT = SVOp->getValueType(0);
4544 SmallVector<std::pair<int, int>, 8> Locs;
4546 SmallVector<int, 8> Mask1(4U, -1);
4547 SmallVector<int, 8> PermMask;
4548 SVOp->getMask(PermMask);
4552 for (unsigned i = 0; i != 4; ++i) {
4553 int Idx = PermMask[i];
4555 Locs[i] = std::make_pair(-1, -1);
4557 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4559 Locs[i] = std::make_pair(0, NumLo);
4563 Locs[i] = std::make_pair(1, NumHi);
4565 Mask1[2+NumHi] = Idx;
4571 if (NumLo <= 2 && NumHi <= 2) {
4572 // If no more than two elements come from either vector. This can be
4573 // implemented with two shuffles. First shuffle gather the elements.
4574 // The second shuffle, which takes the first shuffle as both of its
4575 // vector operands, put the elements into the right order.
4576 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4578 SmallVector<int, 8> Mask2(4U, -1);
4580 for (unsigned i = 0; i != 4; ++i) {
4581 if (Locs[i].first == -1)
4584 unsigned Idx = (i < 2) ? 0 : 4;
4585 Idx += Locs[i].first * 2 + Locs[i].second;
4590 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4591 } else if (NumLo == 3 || NumHi == 3) {
4592 // Otherwise, we must have three elements from one vector, call it X, and
4593 // one element from the other, call it Y. First, use a shufps to build an
4594 // intermediate vector with the one element from Y and the element from X
4595 // that will be in the same half in the final destination (the indexes don't
4596 // matter). Then, use a shufps to build the final vector, taking the half
4597 // containing the element from Y from the intermediate, and the other half
4600 // Normalize it so the 3 elements come from V1.
4601 CommuteVectorShuffleMask(PermMask, VT);
4605 // Find the element from V2.
4607 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4608 int Val = PermMask[HiIndex];
4615 Mask1[0] = PermMask[HiIndex];
4617 Mask1[2] = PermMask[HiIndex^1];
4619 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4622 Mask1[0] = PermMask[0];
4623 Mask1[1] = PermMask[1];
4624 Mask1[2] = HiIndex & 1 ? 6 : 4;
4625 Mask1[3] = HiIndex & 1 ? 4 : 6;
4626 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4628 Mask1[0] = HiIndex & 1 ? 2 : 0;
4629 Mask1[1] = HiIndex & 1 ? 0 : 2;
4630 Mask1[2] = PermMask[2];
4631 Mask1[3] = PermMask[3];
4636 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4640 // Break it into (shuffle shuffle_hi, shuffle_lo).
4642 SmallVector<int,8> LoMask(4U, -1);
4643 SmallVector<int,8> HiMask(4U, -1);
4645 SmallVector<int,8> *MaskPtr = &LoMask;
4646 unsigned MaskIdx = 0;
4649 for (unsigned i = 0; i != 4; ++i) {
4656 int Idx = PermMask[i];
4658 Locs[i] = std::make_pair(-1, -1);
4659 } else if (Idx < 4) {
4660 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4661 (*MaskPtr)[LoIdx] = Idx;
4664 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4665 (*MaskPtr)[HiIdx] = Idx;
4670 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4671 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4672 SmallVector<int, 8> MaskOps;
4673 for (unsigned i = 0; i != 4; ++i) {
4674 if (Locs[i].first == -1) {
4675 MaskOps.push_back(-1);
4677 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4678 MaskOps.push_back(Idx);
4681 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4685 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4686 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4687 SDValue V1 = Op.getOperand(0);
4688 SDValue V2 = Op.getOperand(1);
4689 EVT VT = Op.getValueType();
4690 DebugLoc dl = Op.getDebugLoc();
4691 unsigned NumElems = VT.getVectorNumElements();
4692 bool isMMX = VT.getSizeInBits() == 64;
4693 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4694 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4695 bool V1IsSplat = false;
4696 bool V2IsSplat = false;
4698 if (isZeroShuffle(SVOp))
4699 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4701 // Promote splats to v4f32.
4702 if (SVOp->isSplat()) {
4703 if (isMMX || NumElems < 4)
4705 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4708 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4710 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4711 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4712 if (NewOp.getNode())
4713 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4714 LowerVECTOR_SHUFFLE(NewOp, DAG));
4715 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4716 // FIXME: Figure out a cleaner way to do this.
4717 // Try to make use of movq to zero out the top part.
4718 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4719 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4720 if (NewOp.getNode()) {
4721 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4722 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4723 DAG, Subtarget, dl);
4725 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4726 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4727 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4728 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4729 DAG, Subtarget, dl);
4733 if (X86::isPSHUFDMask(SVOp))
4736 // Check if this can be converted into a logical shift.
4737 bool isLeft = false;
4740 bool isShift = getSubtarget()->hasSSE2() &&
4741 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4742 if (isShift && ShVal.hasOneUse()) {
4743 // If the shifted value has multiple uses, it may be cheaper to use
4744 // v_set0 + movlhps or movhlps, etc.
4745 EVT EltVT = VT.getVectorElementType();
4746 ShAmt *= EltVT.getSizeInBits();
4747 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4750 if (X86::isMOVLMask(SVOp)) {
4753 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4754 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4759 // FIXME: fold these into legal mask.
4760 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4761 X86::isMOVSLDUPMask(SVOp) ||
4762 X86::isMOVHLPSMask(SVOp) ||
4763 X86::isMOVLHPSMask(SVOp) ||
4764 X86::isMOVLPMask(SVOp)))
4767 if (ShouldXformToMOVHLPS(SVOp) ||
4768 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4769 return CommuteVectorShuffle(SVOp, DAG);
4772 // No better options. Use a vshl / vsrl.
4773 EVT EltVT = VT.getVectorElementType();
4774 ShAmt *= EltVT.getSizeInBits();
4775 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4778 bool Commuted = false;
4779 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4780 // 1,1,1,1 -> v8i16 though.
4781 V1IsSplat = isSplatVector(V1.getNode());
4782 V2IsSplat = isSplatVector(V2.getNode());
4784 // Canonicalize the splat or undef, if present, to be on the RHS.
4785 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4786 Op = CommuteVectorShuffle(SVOp, DAG);
4787 SVOp = cast<ShuffleVectorSDNode>(Op);
4788 V1 = SVOp->getOperand(0);
4789 V2 = SVOp->getOperand(1);
4790 std::swap(V1IsSplat, V2IsSplat);
4791 std::swap(V1IsUndef, V2IsUndef);
4795 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4796 // Shuffling low element of v1 into undef, just return v1.
4799 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4800 // the instruction selector will not match, so get a canonical MOVL with
4801 // swapped operands to undo the commute.
4802 return getMOVL(DAG, dl, VT, V2, V1);
4805 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4806 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4807 X86::isUNPCKLMask(SVOp) ||
4808 X86::isUNPCKHMask(SVOp))
4812 // Normalize mask so all entries that point to V2 points to its first
4813 // element then try to match unpck{h|l} again. If match, return a
4814 // new vector_shuffle with the corrected mask.
4815 SDValue NewMask = NormalizeMask(SVOp, DAG);
4816 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4817 if (NSVOp != SVOp) {
4818 if (X86::isUNPCKLMask(NSVOp, true)) {
4820 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4827 // Commute is back and try unpck* again.
4828 // FIXME: this seems wrong.
4829 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4830 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4831 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4832 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4833 X86::isUNPCKLMask(NewSVOp) ||
4834 X86::isUNPCKHMask(NewSVOp))
4838 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4840 // Normalize the node to match x86 shuffle ops if needed
4841 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4842 return CommuteVectorShuffle(SVOp, DAG);
4844 // Check for legal shuffle and return?
4845 SmallVector<int, 16> PermMask;
4846 SVOp->getMask(PermMask);
4847 if (isShuffleMaskLegal(PermMask, VT))
4850 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4851 if (VT == MVT::v8i16) {
4852 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4853 if (NewOp.getNode())
4857 if (VT == MVT::v16i8) {
4858 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4859 if (NewOp.getNode())
4863 // Handle all 4 wide cases with a number of shuffles except for MMX.
4864 if (NumElems == 4 && !isMMX)
4865 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4871 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4872 SelectionDAG &DAG) const {
4873 EVT VT = Op.getValueType();
4874 DebugLoc dl = Op.getDebugLoc();
4875 if (VT.getSizeInBits() == 8) {
4876 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4877 Op.getOperand(0), Op.getOperand(1));
4878 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4879 DAG.getValueType(VT));
4880 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4881 } else if (VT.getSizeInBits() == 16) {
4882 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4883 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4885 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4886 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4887 DAG.getNode(ISD::BIT_CONVERT, dl,
4891 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4892 Op.getOperand(0), Op.getOperand(1));
4893 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4894 DAG.getValueType(VT));
4895 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4896 } else if (VT == MVT::f32) {
4897 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4898 // the result back to FR32 register. It's only worth matching if the
4899 // result has a single use which is a store or a bitcast to i32. And in
4900 // the case of a store, it's not worth it if the index is a constant 0,
4901 // because a MOVSSmr can be used instead, which is smaller and faster.
4902 if (!Op.hasOneUse())
4904 SDNode *User = *Op.getNode()->use_begin();
4905 if ((User->getOpcode() != ISD::STORE ||
4906 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4907 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4908 (User->getOpcode() != ISD::BIT_CONVERT ||
4909 User->getValueType(0) != MVT::i32))
4911 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4912 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4915 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4916 } else if (VT == MVT::i32) {
4917 // ExtractPS works with constant index.
4918 if (isa<ConstantSDNode>(Op.getOperand(1)))
4926 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
4927 SelectionDAG &DAG) const {
4928 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4931 if (Subtarget->hasSSE41()) {
4932 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4937 EVT VT = Op.getValueType();
4938 DebugLoc dl = Op.getDebugLoc();
4939 // TODO: handle v16i8.
4940 if (VT.getSizeInBits() == 16) {
4941 SDValue Vec = Op.getOperand(0);
4942 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4944 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4945 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4946 DAG.getNode(ISD::BIT_CONVERT, dl,
4949 // Transform it so it match pextrw which produces a 32-bit result.
4950 EVT EltVT = MVT::i32;
4951 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4952 Op.getOperand(0), Op.getOperand(1));
4953 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4954 DAG.getValueType(VT));
4955 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4956 } else if (VT.getSizeInBits() == 32) {
4957 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4961 // SHUFPS the element to the lowest double word, then movss.
4962 int Mask[4] = { Idx, -1, -1, -1 };
4963 EVT VVT = Op.getOperand(0).getValueType();
4964 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4965 DAG.getUNDEF(VVT), Mask);
4966 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4967 DAG.getIntPtrConstant(0));
4968 } else if (VT.getSizeInBits() == 64) {
4969 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4970 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4971 // to match extract_elt for f64.
4972 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4976 // UNPCKHPD the element to the lowest double word, then movsd.
4977 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4978 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4979 int Mask[2] = { 1, -1 };
4980 EVT VVT = Op.getOperand(0).getValueType();
4981 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4982 DAG.getUNDEF(VVT), Mask);
4983 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4984 DAG.getIntPtrConstant(0));
4991 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
4992 SelectionDAG &DAG) const {
4993 EVT VT = Op.getValueType();
4994 EVT EltVT = VT.getVectorElementType();
4995 DebugLoc dl = Op.getDebugLoc();
4997 SDValue N0 = Op.getOperand(0);
4998 SDValue N1 = Op.getOperand(1);
4999 SDValue N2 = Op.getOperand(2);
5001 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5002 isa<ConstantSDNode>(N2)) {
5004 if (VT == MVT::v8i16)
5005 Opc = X86ISD::PINSRW;
5006 else if (VT == MVT::v4i16)
5007 Opc = X86ISD::MMX_PINSRW;
5008 else if (VT == MVT::v16i8)
5009 Opc = X86ISD::PINSRB;
5011 Opc = X86ISD::PINSRB;
5013 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5015 if (N1.getValueType() != MVT::i32)
5016 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5017 if (N2.getValueType() != MVT::i32)
5018 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5019 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5020 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5021 // Bits [7:6] of the constant are the source select. This will always be
5022 // zero here. The DAG Combiner may combine an extract_elt index into these
5023 // bits. For example (insert (extract, 3), 2) could be matched by putting
5024 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5025 // Bits [5:4] of the constant are the destination select. This is the
5026 // value of the incoming immediate.
5027 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5028 // combine either bitwise AND or insert of float 0.0 to set these bits.
5029 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5030 // Create this as a scalar to vector..
5031 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5032 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5033 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5034 // PINSR* works with constant index.
5041 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5042 EVT VT = Op.getValueType();
5043 EVT EltVT = VT.getVectorElementType();
5045 if (Subtarget->hasSSE41())
5046 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5048 if (EltVT == MVT::i8)
5051 DebugLoc dl = Op.getDebugLoc();
5052 SDValue N0 = Op.getOperand(0);
5053 SDValue N1 = Op.getOperand(1);
5054 SDValue N2 = Op.getOperand(2);
5056 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5057 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5058 // as its second argument.
5059 if (N1.getValueType() != MVT::i32)
5060 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5061 if (N2.getValueType() != MVT::i32)
5062 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5063 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5064 dl, VT, N0, N1, N2);
5070 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5071 DebugLoc dl = Op.getDebugLoc();
5072 if (Op.getValueType() == MVT::v2f32)
5073 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
5074 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
5075 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
5076 Op.getOperand(0))));
5078 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
5079 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5081 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5082 EVT VT = MVT::v2i32;
5083 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5090 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5091 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5094 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5095 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5096 // one of the above mentioned nodes. It has to be wrapped because otherwise
5097 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5098 // be used to form addressing mode. These wrapped nodes will be selected
5101 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5102 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5104 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5106 unsigned char OpFlag = 0;
5107 unsigned WrapperKind = X86ISD::Wrapper;
5108 CodeModel::Model M = getTargetMachine().getCodeModel();
5110 if (Subtarget->isPICStyleRIPRel() &&
5111 (M == CodeModel::Small || M == CodeModel::Kernel))
5112 WrapperKind = X86ISD::WrapperRIP;
5113 else if (Subtarget->isPICStyleGOT())
5114 OpFlag = X86II::MO_GOTOFF;
5115 else if (Subtarget->isPICStyleStubPIC())
5116 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5118 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5120 CP->getOffset(), OpFlag);
5121 DebugLoc DL = CP->getDebugLoc();
5122 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5123 // With PIC, the address is actually $g + Offset.
5125 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5126 DAG.getNode(X86ISD::GlobalBaseReg,
5127 DebugLoc(), getPointerTy()),
5134 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5135 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5137 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5139 unsigned char OpFlag = 0;
5140 unsigned WrapperKind = X86ISD::Wrapper;
5141 CodeModel::Model M = getTargetMachine().getCodeModel();
5143 if (Subtarget->isPICStyleRIPRel() &&
5144 (M == CodeModel::Small || M == CodeModel::Kernel))
5145 WrapperKind = X86ISD::WrapperRIP;
5146 else if (Subtarget->isPICStyleGOT())
5147 OpFlag = X86II::MO_GOTOFF;
5148 else if (Subtarget->isPICStyleStubPIC())
5149 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5151 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5153 DebugLoc DL = JT->getDebugLoc();
5154 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5156 // With PIC, the address is actually $g + Offset.
5158 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5159 DAG.getNode(X86ISD::GlobalBaseReg,
5160 DebugLoc(), getPointerTy()),
5168 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5169 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5171 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5173 unsigned char OpFlag = 0;
5174 unsigned WrapperKind = X86ISD::Wrapper;
5175 CodeModel::Model M = getTargetMachine().getCodeModel();
5177 if (Subtarget->isPICStyleRIPRel() &&
5178 (M == CodeModel::Small || M == CodeModel::Kernel))
5179 WrapperKind = X86ISD::WrapperRIP;
5180 else if (Subtarget->isPICStyleGOT())
5181 OpFlag = X86II::MO_GOTOFF;
5182 else if (Subtarget->isPICStyleStubPIC())
5183 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5185 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5187 DebugLoc DL = Op.getDebugLoc();
5188 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5191 // With PIC, the address is actually $g + Offset.
5192 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5193 !Subtarget->is64Bit()) {
5194 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5195 DAG.getNode(X86ISD::GlobalBaseReg,
5196 DebugLoc(), getPointerTy()),
5204 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5205 // Create the TargetBlockAddressAddress node.
5206 unsigned char OpFlags =
5207 Subtarget->ClassifyBlockAddressReference();
5208 CodeModel::Model M = getTargetMachine().getCodeModel();
5209 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5210 DebugLoc dl = Op.getDebugLoc();
5211 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5212 /*isTarget=*/true, OpFlags);
5214 if (Subtarget->isPICStyleRIPRel() &&
5215 (M == CodeModel::Small || M == CodeModel::Kernel))
5216 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5218 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5220 // With PIC, the address is actually $g + Offset.
5221 if (isGlobalRelativeToPICBase(OpFlags)) {
5222 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5223 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5231 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5233 SelectionDAG &DAG) const {
5234 // Create the TargetGlobalAddress node, folding in the constant
5235 // offset if it is legal.
5236 unsigned char OpFlags =
5237 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5238 CodeModel::Model M = getTargetMachine().getCodeModel();
5240 if (OpFlags == X86II::MO_NO_FLAG &&
5241 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5242 // A direct static reference to a global.
5243 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5246 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5249 if (Subtarget->isPICStyleRIPRel() &&
5250 (M == CodeModel::Small || M == CodeModel::Kernel))
5251 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5253 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5255 // With PIC, the address is actually $g + Offset.
5256 if (isGlobalRelativeToPICBase(OpFlags)) {
5257 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5258 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5262 // For globals that require a load from a stub to get the address, emit the
5264 if (isGlobalStubReference(OpFlags))
5265 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5266 PseudoSourceValue::getGOT(), 0, false, false, 0);
5268 // If there was a non-zero offset that we didn't fold, create an explicit
5271 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5272 DAG.getConstant(Offset, getPointerTy()));
5278 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5279 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5280 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5281 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5285 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5286 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5287 unsigned char OperandFlags) {
5288 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5289 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5290 DebugLoc dl = GA->getDebugLoc();
5291 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5292 GA->getValueType(0),
5296 SDValue Ops[] = { Chain, TGA, *InFlag };
5297 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5299 SDValue Ops[] = { Chain, TGA };
5300 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5303 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5304 MFI->setAdjustsStack(true);
5306 SDValue Flag = Chain.getValue(1);
5307 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5310 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5312 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5315 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5316 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5317 DAG.getNode(X86ISD::GlobalBaseReg,
5318 DebugLoc(), PtrVT), InFlag);
5319 InFlag = Chain.getValue(1);
5321 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5324 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5326 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5328 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5329 X86::RAX, X86II::MO_TLSGD);
5332 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5333 // "local exec" model.
5334 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5335 const EVT PtrVT, TLSModel::Model model,
5337 DebugLoc dl = GA->getDebugLoc();
5338 // Get the Thread Pointer
5339 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5341 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5344 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5345 NULL, 0, false, false, 0);
5347 unsigned char OperandFlags = 0;
5348 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5350 unsigned WrapperKind = X86ISD::Wrapper;
5351 if (model == TLSModel::LocalExec) {
5352 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5353 } else if (is64Bit) {
5354 assert(model == TLSModel::InitialExec);
5355 OperandFlags = X86II::MO_GOTTPOFF;
5356 WrapperKind = X86ISD::WrapperRIP;
5358 assert(model == TLSModel::InitialExec);
5359 OperandFlags = X86II::MO_INDNTPOFF;
5362 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5364 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5365 GA->getOffset(), OperandFlags);
5366 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5368 if (model == TLSModel::InitialExec)
5369 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5370 PseudoSourceValue::getGOT(), 0, false, false, 0);
5372 // The address of the thread local variable is the add of the thread
5373 // pointer with the offset of the variable.
5374 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5378 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5380 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5381 const GlobalValue *GV = GA->getGlobal();
5383 if (Subtarget->isTargetELF()) {
5384 // TODO: implement the "local dynamic" model
5385 // TODO: implement the "initial exec"model for pic executables
5387 // If GV is an alias then use the aliasee for determining
5388 // thread-localness.
5389 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5390 GV = GA->resolveAliasedGlobal(false);
5392 TLSModel::Model model
5393 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5396 case TLSModel::GeneralDynamic:
5397 case TLSModel::LocalDynamic: // not implemented
5398 if (Subtarget->is64Bit())
5399 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5400 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5402 case TLSModel::InitialExec:
5403 case TLSModel::LocalExec:
5404 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5405 Subtarget->is64Bit());
5407 } else if (Subtarget->isTargetDarwin()) {
5408 // Darwin only has one model of TLS. Lower to that.
5409 unsigned char OpFlag = 0;
5410 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5411 X86ISD::WrapperRIP : X86ISD::Wrapper;
5413 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5415 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5416 !Subtarget->is64Bit();
5418 OpFlag = X86II::MO_TLVP_PIC_BASE;
5420 OpFlag = X86II::MO_TLVP;
5422 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
5424 GA->getOffset(), OpFlag);
5426 DebugLoc DL = Op.getDebugLoc();
5427 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5429 // With PIC32, the address is actually $g + Offset.
5431 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5432 DAG.getNode(X86ISD::GlobalBaseReg,
5433 DebugLoc(), getPointerTy()),
5436 // Lowering the machine isd will make sure everything is in the right
5438 SDValue Args[] = { Offset };
5439 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5441 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5442 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5443 MFI->setAdjustsStack(true);
5445 // And our return value (tls address) is in the standard call return value
5447 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5448 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5452 "TLS not implemented for this target.");
5454 llvm_unreachable("Unreachable");
5459 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5460 /// take a 2 x i32 value to shift plus a shift amount.
5461 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5462 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5463 EVT VT = Op.getValueType();
5464 unsigned VTBits = VT.getSizeInBits();
5465 DebugLoc dl = Op.getDebugLoc();
5466 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5467 SDValue ShOpLo = Op.getOperand(0);
5468 SDValue ShOpHi = Op.getOperand(1);
5469 SDValue ShAmt = Op.getOperand(2);
5470 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5471 DAG.getConstant(VTBits - 1, MVT::i8))
5472 : DAG.getConstant(0, VT);
5475 if (Op.getOpcode() == ISD::SHL_PARTS) {
5476 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5477 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5479 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5480 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5483 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5484 DAG.getConstant(VTBits, MVT::i8));
5485 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5486 AndNode, DAG.getConstant(0, MVT::i8));
5489 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5490 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5491 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5493 if (Op.getOpcode() == ISD::SHL_PARTS) {
5494 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5495 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5497 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5498 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5501 SDValue Ops[2] = { Lo, Hi };
5502 return DAG.getMergeValues(Ops, 2, dl);
5505 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5506 SelectionDAG &DAG) const {
5507 EVT SrcVT = Op.getOperand(0).getValueType();
5509 if (SrcVT.isVector()) {
5510 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5516 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5517 "Unknown SINT_TO_FP to lower!");
5519 // These are really Legal; return the operand so the caller accepts it as
5521 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5523 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5524 Subtarget->is64Bit()) {
5528 DebugLoc dl = Op.getDebugLoc();
5529 unsigned Size = SrcVT.getSizeInBits()/8;
5530 MachineFunction &MF = DAG.getMachineFunction();
5531 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5532 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5533 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5535 PseudoSourceValue::getFixedStack(SSFI), 0,
5537 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5540 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5542 SelectionDAG &DAG) const {
5544 DebugLoc dl = Op.getDebugLoc();
5546 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5548 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5550 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5551 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5552 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5553 Tys, Ops, array_lengthof(Ops));
5556 Chain = Result.getValue(1);
5557 SDValue InFlag = Result.getValue(2);
5559 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5560 // shouldn't be necessary except that RFP cannot be live across
5561 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5562 MachineFunction &MF = DAG.getMachineFunction();
5563 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5564 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5565 Tys = DAG.getVTList(MVT::Other);
5567 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5569 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5570 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5571 PseudoSourceValue::getFixedStack(SSFI), 0,
5578 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5579 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5580 SelectionDAG &DAG) const {
5581 // This algorithm is not obvious. Here it is in C code, more or less:
5583 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5584 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5585 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5587 // Copy ints to xmm registers.
5588 __m128i xh = _mm_cvtsi32_si128( hi );
5589 __m128i xl = _mm_cvtsi32_si128( lo );
5591 // Combine into low half of a single xmm register.
5592 __m128i x = _mm_unpacklo_epi32( xh, xl );
5596 // Merge in appropriate exponents to give the integer bits the right
5598 x = _mm_unpacklo_epi32( x, exp );
5600 // Subtract away the biases to deal with the IEEE-754 double precision
5602 d = _mm_sub_pd( (__m128d) x, bias );
5604 // All conversions up to here are exact. The correctly rounded result is
5605 // calculated using the current rounding mode using the following
5607 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5608 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5609 // store doesn't really need to be here (except
5610 // maybe to zero the other double)
5615 DebugLoc dl = Op.getDebugLoc();
5616 LLVMContext *Context = DAG.getContext();
5618 // Build some magic constants.
5619 std::vector<Constant*> CV0;
5620 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5621 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5622 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5623 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5624 Constant *C0 = ConstantVector::get(CV0);
5625 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5627 std::vector<Constant*> CV1;
5629 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5631 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5632 Constant *C1 = ConstantVector::get(CV1);
5633 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5635 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5636 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5638 DAG.getIntPtrConstant(1)));
5639 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5640 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5642 DAG.getIntPtrConstant(0)));
5643 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5644 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5645 PseudoSourceValue::getConstantPool(), 0,
5647 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5648 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5649 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5650 PseudoSourceValue::getConstantPool(), 0,
5652 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5654 // Add the halves; easiest way is to swap them into another reg first.
5655 int ShufMask[2] = { 1, -1 };
5656 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5657 DAG.getUNDEF(MVT::v2f64), ShufMask);
5658 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5659 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5660 DAG.getIntPtrConstant(0));
5663 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5664 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5665 SelectionDAG &DAG) const {
5666 DebugLoc dl = Op.getDebugLoc();
5667 // FP constant to bias correct the final result.
5668 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5671 // Load the 32-bit value into an XMM register.
5672 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5673 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5675 DAG.getIntPtrConstant(0)));
5677 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5678 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5679 DAG.getIntPtrConstant(0));
5681 // Or the load with the bias.
5682 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5683 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5684 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5686 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5687 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5688 MVT::v2f64, Bias)));
5689 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5690 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5691 DAG.getIntPtrConstant(0));
5693 // Subtract the bias.
5694 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5696 // Handle final rounding.
5697 EVT DestVT = Op.getValueType();
5699 if (DestVT.bitsLT(MVT::f64)) {
5700 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5701 DAG.getIntPtrConstant(0));
5702 } else if (DestVT.bitsGT(MVT::f64)) {
5703 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5706 // Handle final rounding.
5710 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5711 SelectionDAG &DAG) const {
5712 SDValue N0 = Op.getOperand(0);
5713 DebugLoc dl = Op.getDebugLoc();
5715 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5716 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5717 // the optimization here.
5718 if (DAG.SignBitIsZero(N0))
5719 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5721 EVT SrcVT = N0.getValueType();
5722 EVT DstVT = Op.getValueType();
5723 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5724 return LowerUINT_TO_FP_i64(Op, DAG);
5725 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5726 return LowerUINT_TO_FP_i32(Op, DAG);
5728 // Make a 64-bit buffer, and use it to build an FILD.
5729 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5730 if (SrcVT == MVT::i32) {
5731 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5732 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5733 getPointerTy(), StackSlot, WordOff);
5734 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5735 StackSlot, NULL, 0, false, false, 0);
5736 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5737 OffsetSlot, NULL, 0, false, false, 0);
5738 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5742 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5743 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5744 StackSlot, NULL, 0, false, false, 0);
5745 // For i64 source, we need to add the appropriate power of 2 if the input
5746 // was negative. This is the same as the optimization in
5747 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5748 // we must be careful to do the computation in x87 extended precision, not
5749 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5750 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5751 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5752 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5754 APInt FF(32, 0x5F800000ULL);
5756 // Check whether the sign bit is set.
5757 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5758 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5761 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5762 SDValue FudgePtr = DAG.getConstantPool(
5763 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5766 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5767 SDValue Zero = DAG.getIntPtrConstant(0);
5768 SDValue Four = DAG.getIntPtrConstant(4);
5769 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5771 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5773 // Load the value out, extending it from f32 to f80.
5774 // FIXME: Avoid the extend by constructing the right constant pool?
5775 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
5776 FudgePtr, PseudoSourceValue::getConstantPool(),
5777 0, MVT::f32, false, false, 4);
5778 // Extend everything to 80 bits to force it to be done on x87.
5779 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5780 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5783 std::pair<SDValue,SDValue> X86TargetLowering::
5784 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5785 DebugLoc dl = Op.getDebugLoc();
5787 EVT DstTy = Op.getValueType();
5790 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5794 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5795 DstTy.getSimpleVT() >= MVT::i16 &&
5796 "Unknown FP_TO_SINT to lower!");
5798 // These are really Legal.
5799 if (DstTy == MVT::i32 &&
5800 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5801 return std::make_pair(SDValue(), SDValue());
5802 if (Subtarget->is64Bit() &&
5803 DstTy == MVT::i64 &&
5804 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5805 return std::make_pair(SDValue(), SDValue());
5807 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5809 MachineFunction &MF = DAG.getMachineFunction();
5810 unsigned MemSize = DstTy.getSizeInBits()/8;
5811 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5812 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5815 switch (DstTy.getSimpleVT().SimpleTy) {
5816 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5817 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5818 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5819 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5822 SDValue Chain = DAG.getEntryNode();
5823 SDValue Value = Op.getOperand(0);
5824 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5825 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5826 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5827 PseudoSourceValue::getFixedStack(SSFI), 0,
5829 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5831 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5833 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5834 Chain = Value.getValue(1);
5835 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5836 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5839 // Build the FP_TO_INT*_IN_MEM
5840 SDValue Ops[] = { Chain, Value, StackSlot };
5841 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5843 return std::make_pair(FIST, StackSlot);
5846 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5847 SelectionDAG &DAG) const {
5848 if (Op.getValueType().isVector()) {
5849 if (Op.getValueType() == MVT::v2i32 &&
5850 Op.getOperand(0).getValueType() == MVT::v2f64) {
5856 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5857 SDValue FIST = Vals.first, StackSlot = Vals.second;
5858 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5859 if (FIST.getNode() == 0) return Op;
5862 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5863 FIST, StackSlot, NULL, 0, false, false, 0);
5866 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5867 SelectionDAG &DAG) const {
5868 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5869 SDValue FIST = Vals.first, StackSlot = Vals.second;
5870 assert(FIST.getNode() && "Unexpected failure");
5873 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5874 FIST, StackSlot, NULL, 0, false, false, 0);
5877 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5878 SelectionDAG &DAG) const {
5879 LLVMContext *Context = DAG.getContext();
5880 DebugLoc dl = Op.getDebugLoc();
5881 EVT VT = Op.getValueType();
5884 EltVT = VT.getVectorElementType();
5885 std::vector<Constant*> CV;
5886 if (EltVT == MVT::f64) {
5887 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5891 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5897 Constant *C = ConstantVector::get(CV);
5898 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5899 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5900 PseudoSourceValue::getConstantPool(), 0,
5902 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5905 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5906 LLVMContext *Context = DAG.getContext();
5907 DebugLoc dl = Op.getDebugLoc();
5908 EVT VT = Op.getValueType();
5911 EltVT = VT.getVectorElementType();
5912 std::vector<Constant*> CV;
5913 if (EltVT == MVT::f64) {
5914 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5918 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5924 Constant *C = ConstantVector::get(CV);
5925 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5926 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5927 PseudoSourceValue::getConstantPool(), 0,
5929 if (VT.isVector()) {
5930 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5931 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5932 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5934 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5936 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5940 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5941 LLVMContext *Context = DAG.getContext();
5942 SDValue Op0 = Op.getOperand(0);
5943 SDValue Op1 = Op.getOperand(1);
5944 DebugLoc dl = Op.getDebugLoc();
5945 EVT VT = Op.getValueType();
5946 EVT SrcVT = Op1.getValueType();
5948 // If second operand is smaller, extend it first.
5949 if (SrcVT.bitsLT(VT)) {
5950 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5953 // And if it is bigger, shrink it first.
5954 if (SrcVT.bitsGT(VT)) {
5955 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5959 // At this point the operands and the result should have the same
5960 // type, and that won't be f80 since that is not custom lowered.
5962 // First get the sign bit of second operand.
5963 std::vector<Constant*> CV;
5964 if (SrcVT == MVT::f64) {
5965 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5966 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5968 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5969 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5970 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5971 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5973 Constant *C = ConstantVector::get(CV);
5974 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5975 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5976 PseudoSourceValue::getConstantPool(), 0,
5978 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5980 // Shift sign bit right or left if the two operands have different types.
5981 if (SrcVT.bitsGT(VT)) {
5982 // Op0 is MVT::f32, Op1 is MVT::f64.
5983 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5984 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5985 DAG.getConstant(32, MVT::i32));
5986 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5987 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5988 DAG.getIntPtrConstant(0));
5991 // Clear first operand sign bit.
5993 if (VT == MVT::f64) {
5994 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5995 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5997 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5998 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5999 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6000 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6002 C = ConstantVector::get(CV);
6003 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6004 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6005 PseudoSourceValue::getConstantPool(), 0,
6007 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6009 // Or the value with the sign bit.
6010 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6013 /// Emit nodes that will be selected as "test Op0,Op0", or something
6015 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6016 SelectionDAG &DAG) const {
6017 DebugLoc dl = Op.getDebugLoc();
6019 // CF and OF aren't always set the way we want. Determine which
6020 // of these we need.
6021 bool NeedCF = false;
6022 bool NeedOF = false;
6025 case X86::COND_A: case X86::COND_AE:
6026 case X86::COND_B: case X86::COND_BE:
6029 case X86::COND_G: case X86::COND_GE:
6030 case X86::COND_L: case X86::COND_LE:
6031 case X86::COND_O: case X86::COND_NO:
6036 // See if we can use the EFLAGS value from the operand instead of
6037 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6038 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6039 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6040 // Emit a CMP with 0, which is the TEST pattern.
6041 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6042 DAG.getConstant(0, Op.getValueType()));
6044 unsigned Opcode = 0;
6045 unsigned NumOperands = 0;
6046 switch (Op.getNode()->getOpcode()) {
6048 // Due to an isel shortcoming, be conservative if this add is likely to be
6049 // selected as part of a load-modify-store instruction. When the root node
6050 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6051 // uses of other nodes in the match, such as the ADD in this case. This
6052 // leads to the ADD being left around and reselected, with the result being
6053 // two adds in the output. Alas, even if none our users are stores, that
6054 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6055 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6056 // climbing the DAG back to the root, and it doesn't seem to be worth the
6058 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6059 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6060 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6063 if (ConstantSDNode *C =
6064 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6065 // An add of one will be selected as an INC.
6066 if (C->getAPIntValue() == 1) {
6067 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;
6080 // Otherwise use a regular EFLAGS-setting add.
6081 Opcode = X86ISD::ADD;
6085 // If the primary and result isn't used, don't bother using X86ISD::AND,
6086 // because a TEST instruction will be better.
6087 bool NonFlagUse = false;
6088 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6089 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6091 unsigned UOpNo = UI.getOperandNo();
6092 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6093 // Look pass truncate.
6094 UOpNo = User->use_begin().getOperandNo();
6095 User = *User->use_begin();
6098 if (User->getOpcode() != ISD::BRCOND &&
6099 User->getOpcode() != ISD::SETCC &&
6100 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6113 // Due to the ISEL shortcoming noted above, be conservative if this op is
6114 // likely to be selected as part of a load-modify-store instruction.
6115 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6116 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6117 if (UI->getOpcode() == ISD::STORE)
6120 // Otherwise use a regular EFLAGS-setting instruction.
6121 switch (Op.getNode()->getOpcode()) {
6122 default: llvm_unreachable("unexpected operator!");
6123 case ISD::SUB: Opcode = X86ISD::SUB; break;
6124 case ISD::OR: Opcode = X86ISD::OR; break;
6125 case ISD::XOR: Opcode = X86ISD::XOR; break;
6126 case ISD::AND: Opcode = X86ISD::AND; break;
6138 return SDValue(Op.getNode(), 1);
6145 // Emit a CMP with 0, which is the TEST pattern.
6146 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6147 DAG.getConstant(0, Op.getValueType()));
6149 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6150 SmallVector<SDValue, 4> Ops;
6151 for (unsigned i = 0; i != NumOperands; ++i)
6152 Ops.push_back(Op.getOperand(i));
6154 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6155 DAG.ReplaceAllUsesWith(Op, New);
6156 return SDValue(New.getNode(), 1);
6159 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6161 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6162 SelectionDAG &DAG) const {
6163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6164 if (C->getAPIntValue() == 0)
6165 return EmitTest(Op0, X86CC, DAG);
6167 DebugLoc dl = Op0.getDebugLoc();
6168 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6171 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6172 /// if it's possible.
6173 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6174 DebugLoc dl, SelectionDAG &DAG) const {
6175 SDValue Op0 = And.getOperand(0);
6176 SDValue Op1 = And.getOperand(1);
6177 if (Op0.getOpcode() == ISD::TRUNCATE)
6178 Op0 = Op0.getOperand(0);
6179 if (Op1.getOpcode() == ISD::TRUNCATE)
6180 Op1 = Op1.getOperand(0);
6183 if (Op1.getOpcode() == ISD::SHL)
6184 std::swap(Op0, Op1);
6185 if (Op0.getOpcode() == ISD::SHL) {
6186 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6187 if (And00C->getZExtValue() == 1) {
6188 // If we looked past a truncate, check that it's only truncating away
6190 unsigned BitWidth = Op0.getValueSizeInBits();
6191 unsigned AndBitWidth = And.getValueSizeInBits();
6192 if (BitWidth > AndBitWidth) {
6193 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6194 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6195 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6199 RHS = Op0.getOperand(1);
6201 } else if (Op1.getOpcode() == ISD::Constant) {
6202 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6203 SDValue AndLHS = Op0;
6204 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6205 LHS = AndLHS.getOperand(0);
6206 RHS = AndLHS.getOperand(1);
6210 if (LHS.getNode()) {
6211 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6212 // instruction. Since the shift amount is in-range-or-undefined, we know
6213 // that doing a bittest on the i32 value is ok. We extend to i32 because
6214 // the encoding for the i16 version is larger than the i32 version.
6215 // Also promote i16 to i32 for performance / code size reason.
6216 if (LHS.getValueType() == MVT::i8 ||
6217 LHS.getValueType() == MVT::i16)
6218 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6220 // If the operand types disagree, extend the shift amount to match. Since
6221 // BT ignores high bits (like shifts) we can use anyextend.
6222 if (LHS.getValueType() != RHS.getValueType())
6223 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6225 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6226 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6227 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6228 DAG.getConstant(Cond, MVT::i8), BT);
6234 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6235 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6236 SDValue Op0 = Op.getOperand(0);
6237 SDValue Op1 = Op.getOperand(1);
6238 DebugLoc dl = Op.getDebugLoc();
6239 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6241 // Optimize to BT if possible.
6242 // Lower (X & (1 << N)) == 0 to BT(X, N).
6243 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6244 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6245 if (Op0.getOpcode() == ISD::AND &&
6247 Op1.getOpcode() == ISD::Constant &&
6248 cast<ConstantSDNode>(Op1)->isNullValue() &&
6249 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6250 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6251 if (NewSetCC.getNode())
6255 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6256 if (Op0.getOpcode() == X86ISD::SETCC &&
6257 Op1.getOpcode() == ISD::Constant &&
6258 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6259 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6260 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6261 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6262 bool Invert = (CC == ISD::SETNE) ^
6263 cast<ConstantSDNode>(Op1)->isNullValue();
6265 CCode = X86::GetOppositeBranchCondition(CCode);
6266 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6267 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6270 bool isFP = Op1.getValueType().isFloatingPoint();
6271 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6272 if (X86CC == X86::COND_INVALID)
6275 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6277 // Use sbb x, x to materialize carry bit into a GPR.
6278 if (X86CC == X86::COND_B)
6279 return DAG.getNode(ISD::AND, dl, MVT::i8,
6280 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6281 DAG.getConstant(X86CC, MVT::i8), Cond),
6282 DAG.getConstant(1, MVT::i8));
6284 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6285 DAG.getConstant(X86CC, MVT::i8), Cond);
6288 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6290 SDValue Op0 = Op.getOperand(0);
6291 SDValue Op1 = Op.getOperand(1);
6292 SDValue CC = Op.getOperand(2);
6293 EVT VT = Op.getValueType();
6294 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6295 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6296 DebugLoc dl = Op.getDebugLoc();
6300 EVT VT0 = Op0.getValueType();
6301 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6302 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6305 switch (SetCCOpcode) {
6308 case ISD::SETEQ: SSECC = 0; break;
6310 case ISD::SETGT: Swap = true; // Fallthrough
6312 case ISD::SETOLT: SSECC = 1; break;
6314 case ISD::SETGE: Swap = true; // Fallthrough
6316 case ISD::SETOLE: SSECC = 2; break;
6317 case ISD::SETUO: SSECC = 3; break;
6319 case ISD::SETNE: SSECC = 4; break;
6320 case ISD::SETULE: Swap = true;
6321 case ISD::SETUGE: SSECC = 5; break;
6322 case ISD::SETULT: Swap = true;
6323 case ISD::SETUGT: SSECC = 6; break;
6324 case ISD::SETO: SSECC = 7; break;
6327 std::swap(Op0, Op1);
6329 // In the two special cases we can't handle, emit two comparisons.
6331 if (SetCCOpcode == ISD::SETUEQ) {
6333 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6334 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6335 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6337 else if (SetCCOpcode == ISD::SETONE) {
6339 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6340 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6341 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6343 llvm_unreachable("Illegal FP comparison");
6345 // Handle all other FP comparisons here.
6346 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6349 // We are handling one of the integer comparisons here. Since SSE only has
6350 // GT and EQ comparisons for integer, swapping operands and multiple
6351 // operations may be required for some comparisons.
6352 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6353 bool Swap = false, Invert = false, FlipSigns = false;
6355 switch (VT.getSimpleVT().SimpleTy) {
6358 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6360 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6362 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6363 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6366 switch (SetCCOpcode) {
6368 case ISD::SETNE: Invert = true;
6369 case ISD::SETEQ: Opc = EQOpc; break;
6370 case ISD::SETLT: Swap = true;
6371 case ISD::SETGT: Opc = GTOpc; break;
6372 case ISD::SETGE: Swap = true;
6373 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6374 case ISD::SETULT: Swap = true;
6375 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6376 case ISD::SETUGE: Swap = true;
6377 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6380 std::swap(Op0, Op1);
6382 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6383 // bits of the inputs before performing those operations.
6385 EVT EltVT = VT.getVectorElementType();
6386 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6388 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6389 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6391 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6392 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6395 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6397 // If the logical-not of the result is required, perform that now.
6399 Result = DAG.getNOT(dl, Result, VT);
6404 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6405 static bool isX86LogicalCmp(SDValue Op) {
6406 unsigned Opc = Op.getNode()->getOpcode();
6407 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6409 if (Op.getResNo() == 1 &&
6410 (Opc == X86ISD::ADD ||
6411 Opc == X86ISD::SUB ||
6412 Opc == X86ISD::SMUL ||
6413 Opc == X86ISD::UMUL ||
6414 Opc == X86ISD::INC ||
6415 Opc == X86ISD::DEC ||
6416 Opc == X86ISD::OR ||
6417 Opc == X86ISD::XOR ||
6418 Opc == X86ISD::AND))
6424 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6425 bool addTest = true;
6426 SDValue Cond = Op.getOperand(0);
6427 DebugLoc dl = Op.getDebugLoc();
6430 if (Cond.getOpcode() == ISD::SETCC) {
6431 SDValue NewCond = LowerSETCC(Cond, DAG);
6432 if (NewCond.getNode())
6436 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6437 SDValue Op1 = Op.getOperand(1);
6438 SDValue Op2 = Op.getOperand(2);
6439 if (Cond.getOpcode() == X86ISD::SETCC &&
6440 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6441 SDValue Cmp = Cond.getOperand(1);
6442 if (Cmp.getOpcode() == X86ISD::CMP) {
6443 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6444 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6445 ConstantSDNode *RHSC =
6446 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6447 if (N1C && N1C->isAllOnesValue() &&
6448 N2C && N2C->isNullValue() &&
6449 RHSC && RHSC->isNullValue()) {
6450 SDValue CmpOp0 = Cmp.getOperand(0);
6451 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6452 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6453 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6454 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6459 // Look pass (and (setcc_carry (cmp ...)), 1).
6460 if (Cond.getOpcode() == ISD::AND &&
6461 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6462 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6463 if (C && C->getAPIntValue() == 1)
6464 Cond = Cond.getOperand(0);
6467 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6468 // setting operand in place of the X86ISD::SETCC.
6469 if (Cond.getOpcode() == X86ISD::SETCC ||
6470 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6471 CC = Cond.getOperand(0);
6473 SDValue Cmp = Cond.getOperand(1);
6474 unsigned Opc = Cmp.getOpcode();
6475 EVT VT = Op.getValueType();
6477 bool IllegalFPCMov = false;
6478 if (VT.isFloatingPoint() && !VT.isVector() &&
6479 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6480 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6482 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6483 Opc == X86ISD::BT) { // FIXME
6490 // Look pass the truncate.
6491 if (Cond.getOpcode() == ISD::TRUNCATE)
6492 Cond = Cond.getOperand(0);
6494 // We know the result of AND is compared against zero. Try to match
6496 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6497 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6498 if (NewSetCC.getNode()) {
6499 CC = NewSetCC.getOperand(0);
6500 Cond = NewSetCC.getOperand(1);
6507 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6508 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6511 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6512 // condition is true.
6513 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6514 SDValue Ops[] = { Op2, Op1, CC, Cond };
6515 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6518 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6519 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6520 // from the AND / OR.
6521 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6522 Opc = Op.getOpcode();
6523 if (Opc != ISD::OR && Opc != ISD::AND)
6525 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6526 Op.getOperand(0).hasOneUse() &&
6527 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6528 Op.getOperand(1).hasOneUse());
6531 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6532 // 1 and that the SETCC node has a single use.
6533 static bool isXor1OfSetCC(SDValue Op) {
6534 if (Op.getOpcode() != ISD::XOR)
6536 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6537 if (N1C && N1C->getAPIntValue() == 1) {
6538 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6539 Op.getOperand(0).hasOneUse();
6544 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6545 bool addTest = true;
6546 SDValue Chain = Op.getOperand(0);
6547 SDValue Cond = Op.getOperand(1);
6548 SDValue Dest = Op.getOperand(2);
6549 DebugLoc dl = Op.getDebugLoc();
6552 if (Cond.getOpcode() == ISD::SETCC) {
6553 SDValue NewCond = LowerSETCC(Cond, DAG);
6554 if (NewCond.getNode())
6558 // FIXME: LowerXALUO doesn't handle these!!
6559 else if (Cond.getOpcode() == X86ISD::ADD ||
6560 Cond.getOpcode() == X86ISD::SUB ||
6561 Cond.getOpcode() == X86ISD::SMUL ||
6562 Cond.getOpcode() == X86ISD::UMUL)
6563 Cond = LowerXALUO(Cond, DAG);
6566 // Look pass (and (setcc_carry (cmp ...)), 1).
6567 if (Cond.getOpcode() == ISD::AND &&
6568 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6569 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6570 if (C && C->getAPIntValue() == 1)
6571 Cond = Cond.getOperand(0);
6574 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6575 // setting operand in place of the X86ISD::SETCC.
6576 if (Cond.getOpcode() == X86ISD::SETCC ||
6577 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6578 CC = Cond.getOperand(0);
6580 SDValue Cmp = Cond.getOperand(1);
6581 unsigned Opc = Cmp.getOpcode();
6582 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6583 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6587 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6591 // These can only come from an arithmetic instruction with overflow,
6592 // e.g. SADDO, UADDO.
6593 Cond = Cond.getNode()->getOperand(1);
6600 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6601 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6602 if (CondOpc == ISD::OR) {
6603 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6604 // two branches instead of an explicit OR instruction with a
6606 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6607 isX86LogicalCmp(Cmp)) {
6608 CC = Cond.getOperand(0).getOperand(0);
6609 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6610 Chain, Dest, CC, Cmp);
6611 CC = Cond.getOperand(1).getOperand(0);
6615 } else { // ISD::AND
6616 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6617 // two branches instead of an explicit AND instruction with a
6618 // separate test. However, we only do this if this block doesn't
6619 // have a fall-through edge, because this requires an explicit
6620 // jmp when the condition is false.
6621 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6622 isX86LogicalCmp(Cmp) &&
6623 Op.getNode()->hasOneUse()) {
6624 X86::CondCode CCode =
6625 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6626 CCode = X86::GetOppositeBranchCondition(CCode);
6627 CC = DAG.getConstant(CCode, MVT::i8);
6628 SDNode *User = *Op.getNode()->use_begin();
6629 // Look for an unconditional branch following this conditional branch.
6630 // We need this because we need to reverse the successors in order
6631 // to implement FCMP_OEQ.
6632 if (User->getOpcode() == ISD::BR) {
6633 SDValue FalseBB = User->getOperand(1);
6635 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6636 assert(NewBR == User);
6640 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6641 Chain, Dest, CC, Cmp);
6642 X86::CondCode CCode =
6643 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6644 CCode = X86::GetOppositeBranchCondition(CCode);
6645 CC = DAG.getConstant(CCode, MVT::i8);
6651 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6652 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6653 // It should be transformed during dag combiner except when the condition
6654 // is set by a arithmetics with overflow node.
6655 X86::CondCode CCode =
6656 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6657 CCode = X86::GetOppositeBranchCondition(CCode);
6658 CC = DAG.getConstant(CCode, MVT::i8);
6659 Cond = Cond.getOperand(0).getOperand(1);
6665 // Look pass the truncate.
6666 if (Cond.getOpcode() == ISD::TRUNCATE)
6667 Cond = Cond.getOperand(0);
6669 // We know the result of AND is compared against zero. Try to match
6671 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6672 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6673 if (NewSetCC.getNode()) {
6674 CC = NewSetCC.getOperand(0);
6675 Cond = NewSetCC.getOperand(1);
6682 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6683 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6685 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6686 Chain, Dest, CC, Cond);
6690 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6691 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6692 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6693 // that the guard pages used by the OS virtual memory manager are allocated in
6694 // correct sequence.
6696 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6697 SelectionDAG &DAG) const {
6698 assert(Subtarget->isTargetCygMing() &&
6699 "This should be used only on Cygwin/Mingw targets");
6700 DebugLoc dl = Op.getDebugLoc();
6703 SDValue Chain = Op.getOperand(0);
6704 SDValue Size = Op.getOperand(1);
6705 // FIXME: Ensure alignment here
6709 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6711 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6712 Flag = Chain.getValue(1);
6714 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6716 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6717 Flag = Chain.getValue(1);
6719 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6721 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6722 return DAG.getMergeValues(Ops1, 2, dl);
6725 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6726 MachineFunction &MF = DAG.getMachineFunction();
6727 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6729 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6730 DebugLoc dl = Op.getDebugLoc();
6732 if (!Subtarget->is64Bit()) {
6733 // vastart just stores the address of the VarArgsFrameIndex slot into the
6734 // memory location argument.
6735 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6737 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6742 // gp_offset (0 - 6 * 8)
6743 // fp_offset (48 - 48 + 8 * 16)
6744 // overflow_arg_area (point to parameters coming in memory).
6746 SmallVector<SDValue, 8> MemOps;
6747 SDValue FIN = Op.getOperand(1);
6749 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6750 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6752 FIN, SV, 0, false, false, 0);
6753 MemOps.push_back(Store);
6756 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6757 FIN, DAG.getIntPtrConstant(4));
6758 Store = DAG.getStore(Op.getOperand(0), dl,
6759 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6761 FIN, SV, 0, false, false, 0);
6762 MemOps.push_back(Store);
6764 // Store ptr to overflow_arg_area
6765 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6766 FIN, DAG.getIntPtrConstant(4));
6767 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6769 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6771 MemOps.push_back(Store);
6773 // Store ptr to reg_save_area.
6774 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6775 FIN, DAG.getIntPtrConstant(8));
6776 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6778 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6780 MemOps.push_back(Store);
6781 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6782 &MemOps[0], MemOps.size());
6785 SDValue X86TargetLowering::LowerVAARG(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_arg!");
6789 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6793 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6794 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6795 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6796 SDValue Chain = Op.getOperand(0);
6797 SDValue DstPtr = Op.getOperand(1);
6798 SDValue SrcPtr = Op.getOperand(2);
6799 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6800 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6801 DebugLoc dl = Op.getDebugLoc();
6803 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6804 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6805 false, DstSV, 0, SrcSV, 0);
6809 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6810 DebugLoc dl = Op.getDebugLoc();
6811 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6813 default: return SDValue(); // Don't custom lower most intrinsics.
6814 // Comparison intrinsics.
6815 case Intrinsic::x86_sse_comieq_ss:
6816 case Intrinsic::x86_sse_comilt_ss:
6817 case Intrinsic::x86_sse_comile_ss:
6818 case Intrinsic::x86_sse_comigt_ss:
6819 case Intrinsic::x86_sse_comige_ss:
6820 case Intrinsic::x86_sse_comineq_ss:
6821 case Intrinsic::x86_sse_ucomieq_ss:
6822 case Intrinsic::x86_sse_ucomilt_ss:
6823 case Intrinsic::x86_sse_ucomile_ss:
6824 case Intrinsic::x86_sse_ucomigt_ss:
6825 case Intrinsic::x86_sse_ucomige_ss:
6826 case Intrinsic::x86_sse_ucomineq_ss:
6827 case Intrinsic::x86_sse2_comieq_sd:
6828 case Intrinsic::x86_sse2_comilt_sd:
6829 case Intrinsic::x86_sse2_comile_sd:
6830 case Intrinsic::x86_sse2_comigt_sd:
6831 case Intrinsic::x86_sse2_comige_sd:
6832 case Intrinsic::x86_sse2_comineq_sd:
6833 case Intrinsic::x86_sse2_ucomieq_sd:
6834 case Intrinsic::x86_sse2_ucomilt_sd:
6835 case Intrinsic::x86_sse2_ucomile_sd:
6836 case Intrinsic::x86_sse2_ucomigt_sd:
6837 case Intrinsic::x86_sse2_ucomige_sd:
6838 case Intrinsic::x86_sse2_ucomineq_sd: {
6840 ISD::CondCode CC = ISD::SETCC_INVALID;
6843 case Intrinsic::x86_sse_comieq_ss:
6844 case Intrinsic::x86_sse2_comieq_sd:
6848 case Intrinsic::x86_sse_comilt_ss:
6849 case Intrinsic::x86_sse2_comilt_sd:
6853 case Intrinsic::x86_sse_comile_ss:
6854 case Intrinsic::x86_sse2_comile_sd:
6858 case Intrinsic::x86_sse_comigt_ss:
6859 case Intrinsic::x86_sse2_comigt_sd:
6863 case Intrinsic::x86_sse_comige_ss:
6864 case Intrinsic::x86_sse2_comige_sd:
6868 case Intrinsic::x86_sse_comineq_ss:
6869 case Intrinsic::x86_sse2_comineq_sd:
6873 case Intrinsic::x86_sse_ucomieq_ss:
6874 case Intrinsic::x86_sse2_ucomieq_sd:
6875 Opc = X86ISD::UCOMI;
6878 case Intrinsic::x86_sse_ucomilt_ss:
6879 case Intrinsic::x86_sse2_ucomilt_sd:
6880 Opc = X86ISD::UCOMI;
6883 case Intrinsic::x86_sse_ucomile_ss:
6884 case Intrinsic::x86_sse2_ucomile_sd:
6885 Opc = X86ISD::UCOMI;
6888 case Intrinsic::x86_sse_ucomigt_ss:
6889 case Intrinsic::x86_sse2_ucomigt_sd:
6890 Opc = X86ISD::UCOMI;
6893 case Intrinsic::x86_sse_ucomige_ss:
6894 case Intrinsic::x86_sse2_ucomige_sd:
6895 Opc = X86ISD::UCOMI;
6898 case Intrinsic::x86_sse_ucomineq_ss:
6899 case Intrinsic::x86_sse2_ucomineq_sd:
6900 Opc = X86ISD::UCOMI;
6905 SDValue LHS = Op.getOperand(1);
6906 SDValue RHS = Op.getOperand(2);
6907 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6908 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6909 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6910 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6911 DAG.getConstant(X86CC, MVT::i8), Cond);
6912 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6914 // ptest intrinsics. The intrinsic these come from are designed to return
6915 // an integer value, not just an instruction so lower it to the ptest
6916 // pattern and a setcc for the result.
6917 case Intrinsic::x86_sse41_ptestz:
6918 case Intrinsic::x86_sse41_ptestc:
6919 case Intrinsic::x86_sse41_ptestnzc:{
6922 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6923 case Intrinsic::x86_sse41_ptestz:
6925 X86CC = X86::COND_E;
6927 case Intrinsic::x86_sse41_ptestc:
6929 X86CC = X86::COND_B;
6931 case Intrinsic::x86_sse41_ptestnzc:
6933 X86CC = X86::COND_A;
6937 SDValue LHS = Op.getOperand(1);
6938 SDValue RHS = Op.getOperand(2);
6939 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6940 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6941 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6942 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6945 // Fix vector shift instructions where the last operand is a non-immediate
6947 case Intrinsic::x86_sse2_pslli_w:
6948 case Intrinsic::x86_sse2_pslli_d:
6949 case Intrinsic::x86_sse2_pslli_q:
6950 case Intrinsic::x86_sse2_psrli_w:
6951 case Intrinsic::x86_sse2_psrli_d:
6952 case Intrinsic::x86_sse2_psrli_q:
6953 case Intrinsic::x86_sse2_psrai_w:
6954 case Intrinsic::x86_sse2_psrai_d:
6955 case Intrinsic::x86_mmx_pslli_w:
6956 case Intrinsic::x86_mmx_pslli_d:
6957 case Intrinsic::x86_mmx_pslli_q:
6958 case Intrinsic::x86_mmx_psrli_w:
6959 case Intrinsic::x86_mmx_psrli_d:
6960 case Intrinsic::x86_mmx_psrli_q:
6961 case Intrinsic::x86_mmx_psrai_w:
6962 case Intrinsic::x86_mmx_psrai_d: {
6963 SDValue ShAmt = Op.getOperand(2);
6964 if (isa<ConstantSDNode>(ShAmt))
6967 unsigned NewIntNo = 0;
6968 EVT ShAmtVT = MVT::v4i32;
6970 case Intrinsic::x86_sse2_pslli_w:
6971 NewIntNo = Intrinsic::x86_sse2_psll_w;
6973 case Intrinsic::x86_sse2_pslli_d:
6974 NewIntNo = Intrinsic::x86_sse2_psll_d;
6976 case Intrinsic::x86_sse2_pslli_q:
6977 NewIntNo = Intrinsic::x86_sse2_psll_q;
6979 case Intrinsic::x86_sse2_psrli_w:
6980 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6982 case Intrinsic::x86_sse2_psrli_d:
6983 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6985 case Intrinsic::x86_sse2_psrli_q:
6986 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6988 case Intrinsic::x86_sse2_psrai_w:
6989 NewIntNo = Intrinsic::x86_sse2_psra_w;
6991 case Intrinsic::x86_sse2_psrai_d:
6992 NewIntNo = Intrinsic::x86_sse2_psra_d;
6995 ShAmtVT = MVT::v2i32;
6997 case Intrinsic::x86_mmx_pslli_w:
6998 NewIntNo = Intrinsic::x86_mmx_psll_w;
7000 case Intrinsic::x86_mmx_pslli_d:
7001 NewIntNo = Intrinsic::x86_mmx_psll_d;
7003 case Intrinsic::x86_mmx_pslli_q:
7004 NewIntNo = Intrinsic::x86_mmx_psll_q;
7006 case Intrinsic::x86_mmx_psrli_w:
7007 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7009 case Intrinsic::x86_mmx_psrli_d:
7010 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7012 case Intrinsic::x86_mmx_psrli_q:
7013 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7015 case Intrinsic::x86_mmx_psrai_w:
7016 NewIntNo = Intrinsic::x86_mmx_psra_w;
7018 case Intrinsic::x86_mmx_psrai_d:
7019 NewIntNo = Intrinsic::x86_mmx_psra_d;
7021 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7027 // The vector shift intrinsics with scalars uses 32b shift amounts but
7028 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7032 ShOps[1] = DAG.getConstant(0, MVT::i32);
7033 if (ShAmtVT == MVT::v4i32) {
7034 ShOps[2] = DAG.getUNDEF(MVT::i32);
7035 ShOps[3] = DAG.getUNDEF(MVT::i32);
7036 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7038 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7041 EVT VT = Op.getValueType();
7042 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7043 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7044 DAG.getConstant(NewIntNo, MVT::i32),
7045 Op.getOperand(1), ShAmt);
7050 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7051 SelectionDAG &DAG) const {
7052 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7053 MFI->setReturnAddressIsTaken(true);
7055 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7056 DebugLoc dl = Op.getDebugLoc();
7059 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7061 DAG.getConstant(TD->getPointerSize(),
7062 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7063 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7064 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7066 NULL, 0, false, false, 0);
7069 // Just load the return address.
7070 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7071 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7072 RetAddrFI, NULL, 0, false, false, 0);
7075 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7076 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7077 MFI->setFrameAddressIsTaken(true);
7079 EVT VT = Op.getValueType();
7080 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7081 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7082 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7083 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7085 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7090 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7091 SelectionDAG &DAG) const {
7092 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7095 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7096 MachineFunction &MF = DAG.getMachineFunction();
7097 SDValue Chain = Op.getOperand(0);
7098 SDValue Offset = Op.getOperand(1);
7099 SDValue Handler = Op.getOperand(2);
7100 DebugLoc dl = Op.getDebugLoc();
7102 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7104 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7106 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7107 DAG.getIntPtrConstant(-TD->getPointerSize()));
7108 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7109 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7110 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7111 MF.getRegInfo().addLiveOut(StoreAddrReg);
7113 return DAG.getNode(X86ISD::EH_RETURN, dl,
7115 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7118 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7119 SelectionDAG &DAG) const {
7120 SDValue Root = Op.getOperand(0);
7121 SDValue Trmp = Op.getOperand(1); // trampoline
7122 SDValue FPtr = Op.getOperand(2); // nested function
7123 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7124 DebugLoc dl = Op.getDebugLoc();
7126 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7128 if (Subtarget->is64Bit()) {
7129 SDValue OutChains[6];
7131 // Large code-model.
7132 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7133 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7135 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7136 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7138 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7140 // Load the pointer to the nested function into R11.
7141 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7142 SDValue Addr = Trmp;
7143 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7144 Addr, TrmpAddr, 0, false, false, 0);
7146 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7147 DAG.getConstant(2, MVT::i64));
7148 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7151 // Load the 'nest' parameter value into R10.
7152 // R10 is specified in X86CallingConv.td
7153 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7154 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7155 DAG.getConstant(10, MVT::i64));
7156 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7157 Addr, TrmpAddr, 10, false, false, 0);
7159 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7160 DAG.getConstant(12, MVT::i64));
7161 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7164 // Jump to the nested function.
7165 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7166 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7167 DAG.getConstant(20, MVT::i64));
7168 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7169 Addr, TrmpAddr, 20, false, false, 0);
7171 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7172 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7173 DAG.getConstant(22, MVT::i64));
7174 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7175 TrmpAddr, 22, false, false, 0);
7178 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7179 return DAG.getMergeValues(Ops, 2, dl);
7181 const Function *Func =
7182 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7183 CallingConv::ID CC = Func->getCallingConv();
7188 llvm_unreachable("Unsupported calling convention");
7189 case CallingConv::C:
7190 case CallingConv::X86_StdCall: {
7191 // Pass 'nest' parameter in ECX.
7192 // Must be kept in sync with X86CallingConv.td
7195 // Check that ECX wasn't needed by an 'inreg' parameter.
7196 const FunctionType *FTy = Func->getFunctionType();
7197 const AttrListPtr &Attrs = Func->getAttributes();
7199 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7200 unsigned InRegCount = 0;
7203 for (FunctionType::param_iterator I = FTy->param_begin(),
7204 E = FTy->param_end(); I != E; ++I, ++Idx)
7205 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7206 // FIXME: should only count parameters that are lowered to integers.
7207 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7209 if (InRegCount > 2) {
7210 report_fatal_error("Nest register in use - reduce number of inreg parameters!");
7215 case CallingConv::X86_FastCall:
7216 case CallingConv::X86_ThisCall:
7217 case CallingConv::Fast:
7218 // Pass 'nest' parameter in EAX.
7219 // Must be kept in sync with X86CallingConv.td
7224 SDValue OutChains[4];
7227 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7228 DAG.getConstant(10, MVT::i32));
7229 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7231 // This is storing the opcode for MOV32ri.
7232 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7233 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7234 OutChains[0] = DAG.getStore(Root, dl,
7235 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7236 Trmp, TrmpAddr, 0, false, false, 0);
7238 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7239 DAG.getConstant(1, MVT::i32));
7240 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7243 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7244 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7245 DAG.getConstant(5, MVT::i32));
7246 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7247 TrmpAddr, 5, false, false, 1);
7249 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7250 DAG.getConstant(6, MVT::i32));
7251 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7255 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7256 return DAG.getMergeValues(Ops, 2, dl);
7260 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7261 SelectionDAG &DAG) const {
7263 The rounding mode is in bits 11:10 of FPSR, and has the following
7270 FLT_ROUNDS, on the other hand, expects the following:
7277 To perform the conversion, we do:
7278 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7281 MachineFunction &MF = DAG.getMachineFunction();
7282 const TargetMachine &TM = MF.getTarget();
7283 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7284 unsigned StackAlignment = TFI.getStackAlignment();
7285 EVT VT = Op.getValueType();
7286 DebugLoc dl = Op.getDebugLoc();
7288 // Save FP Control Word to stack slot
7289 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7290 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7292 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7293 DAG.getEntryNode(), StackSlot);
7295 // Load FP Control Word from stack slot
7296 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7299 // Transform as necessary
7301 DAG.getNode(ISD::SRL, dl, MVT::i16,
7302 DAG.getNode(ISD::AND, dl, MVT::i16,
7303 CWD, DAG.getConstant(0x800, MVT::i16)),
7304 DAG.getConstant(11, MVT::i8));
7306 DAG.getNode(ISD::SRL, dl, MVT::i16,
7307 DAG.getNode(ISD::AND, dl, MVT::i16,
7308 CWD, DAG.getConstant(0x400, MVT::i16)),
7309 DAG.getConstant(9, MVT::i8));
7312 DAG.getNode(ISD::AND, dl, MVT::i16,
7313 DAG.getNode(ISD::ADD, dl, MVT::i16,
7314 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7315 DAG.getConstant(1, MVT::i16)),
7316 DAG.getConstant(3, MVT::i16));
7319 return DAG.getNode((VT.getSizeInBits() < 16 ?
7320 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7323 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7324 EVT VT = Op.getValueType();
7326 unsigned NumBits = VT.getSizeInBits();
7327 DebugLoc dl = Op.getDebugLoc();
7329 Op = Op.getOperand(0);
7330 if (VT == MVT::i8) {
7331 // Zero extend to i32 since there is not an i8 bsr.
7333 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7336 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7337 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7338 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7340 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7343 DAG.getConstant(NumBits+NumBits-1, OpVT),
7344 DAG.getConstant(X86::COND_E, MVT::i8),
7347 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7349 // Finally xor with NumBits-1.
7350 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7353 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7357 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7358 EVT VT = Op.getValueType();
7360 unsigned NumBits = VT.getSizeInBits();
7361 DebugLoc dl = Op.getDebugLoc();
7363 Op = Op.getOperand(0);
7364 if (VT == MVT::i8) {
7366 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7369 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7370 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7371 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7373 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7376 DAG.getConstant(NumBits, OpVT),
7377 DAG.getConstant(X86::COND_E, MVT::i8),
7380 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7383 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7387 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7388 EVT VT = Op.getValueType();
7389 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7390 DebugLoc dl = Op.getDebugLoc();
7392 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7393 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7394 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7395 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7396 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7398 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7399 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7400 // return AloBlo + AloBhi + AhiBlo;
7402 SDValue A = Op.getOperand(0);
7403 SDValue B = Op.getOperand(1);
7405 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7406 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7407 A, DAG.getConstant(32, MVT::i32));
7408 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7409 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7410 B, DAG.getConstant(32, MVT::i32));
7411 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7412 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7414 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7415 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7417 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7418 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7420 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7421 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7422 AloBhi, DAG.getConstant(32, MVT::i32));
7423 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7424 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7425 AhiBlo, DAG.getConstant(32, MVT::i32));
7426 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7427 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7432 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7433 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7434 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7435 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7436 // has only one use.
7437 SDNode *N = Op.getNode();
7438 SDValue LHS = N->getOperand(0);
7439 SDValue RHS = N->getOperand(1);
7440 unsigned BaseOp = 0;
7442 DebugLoc dl = Op.getDebugLoc();
7444 switch (Op.getOpcode()) {
7445 default: llvm_unreachable("Unknown ovf instruction!");
7447 // A subtract of one will be selected as a INC. Note that INC doesn't
7448 // set CF, so we can't do this for UADDO.
7449 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7450 if (C->getAPIntValue() == 1) {
7451 BaseOp = X86ISD::INC;
7455 BaseOp = X86ISD::ADD;
7459 BaseOp = X86ISD::ADD;
7463 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7464 // set CF, so we can't do this for USUBO.
7465 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7466 if (C->getAPIntValue() == 1) {
7467 BaseOp = X86ISD::DEC;
7471 BaseOp = X86ISD::SUB;
7475 BaseOp = X86ISD::SUB;
7479 BaseOp = X86ISD::SMUL;
7483 BaseOp = X86ISD::UMUL;
7488 // Also sets EFLAGS.
7489 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7490 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7493 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7494 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7496 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7500 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7501 EVT T = Op.getValueType();
7502 DebugLoc dl = Op.getDebugLoc();
7505 switch(T.getSimpleVT().SimpleTy) {
7507 assert(false && "Invalid value type!");
7508 case MVT::i8: Reg = X86::AL; size = 1; break;
7509 case MVT::i16: Reg = X86::AX; size = 2; break;
7510 case MVT::i32: Reg = X86::EAX; size = 4; break;
7512 assert(Subtarget->is64Bit() && "Node not type legal!");
7513 Reg = X86::RAX; size = 8;
7516 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7517 Op.getOperand(2), SDValue());
7518 SDValue Ops[] = { cpIn.getValue(0),
7521 DAG.getTargetConstant(size, MVT::i8),
7523 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7524 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7526 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7530 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7531 SelectionDAG &DAG) const {
7532 assert(Subtarget->is64Bit() && "Result not type legalized?");
7533 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7534 SDValue TheChain = Op.getOperand(0);
7535 DebugLoc dl = Op.getDebugLoc();
7536 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7537 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7538 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7540 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7541 DAG.getConstant(32, MVT::i8));
7543 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7546 return DAG.getMergeValues(Ops, 2, dl);
7549 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7550 SelectionDAG &DAG) const {
7551 EVT SrcVT = Op.getOperand(0).getValueType();
7552 EVT DstVT = Op.getValueType();
7553 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7554 Subtarget->hasMMX() && !DisableMMX) &&
7555 "Unexpected custom BIT_CONVERT");
7556 assert((DstVT == MVT::i64 ||
7557 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7558 "Unexpected custom BIT_CONVERT");
7559 // i64 <=> MMX conversions are Legal.
7560 if (SrcVT==MVT::i64 && DstVT.isVector())
7562 if (DstVT==MVT::i64 && SrcVT.isVector())
7564 // MMX <=> MMX conversions are Legal.
7565 if (SrcVT.isVector() && DstVT.isVector())
7567 // All other conversions need to be expanded.
7570 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7571 SDNode *Node = Op.getNode();
7572 DebugLoc dl = Node->getDebugLoc();
7573 EVT T = Node->getValueType(0);
7574 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7575 DAG.getConstant(0, T), Node->getOperand(2));
7576 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7577 cast<AtomicSDNode>(Node)->getMemoryVT(),
7578 Node->getOperand(0),
7579 Node->getOperand(1), negOp,
7580 cast<AtomicSDNode>(Node)->getSrcValue(),
7581 cast<AtomicSDNode>(Node)->getAlignment());
7584 /// LowerOperation - Provide custom lowering hooks for some operations.
7586 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7587 switch (Op.getOpcode()) {
7588 default: llvm_unreachable("Should not custom lower this!");
7589 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7590 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7591 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7592 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7593 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7594 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7595 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7596 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7597 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7598 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7599 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7600 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7601 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7602 case ISD::SHL_PARTS:
7603 case ISD::SRA_PARTS:
7604 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7605 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7606 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7607 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7608 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7609 case ISD::FABS: return LowerFABS(Op, DAG);
7610 case ISD::FNEG: return LowerFNEG(Op, DAG);
7611 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7612 case ISD::SETCC: return LowerSETCC(Op, DAG);
7613 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7614 case ISD::SELECT: return LowerSELECT(Op, DAG);
7615 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7616 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7617 case ISD::VASTART: return LowerVASTART(Op, DAG);
7618 case ISD::VAARG: return LowerVAARG(Op, DAG);
7619 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7620 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7621 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7622 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7623 case ISD::FRAME_TO_ARGS_OFFSET:
7624 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7625 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7626 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7627 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7628 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7629 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7630 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7631 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7637 case ISD::UMULO: return LowerXALUO(Op, DAG);
7638 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7639 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7643 void X86TargetLowering::
7644 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7645 SelectionDAG &DAG, unsigned NewOp) const {
7646 EVT T = Node->getValueType(0);
7647 DebugLoc dl = Node->getDebugLoc();
7648 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7650 SDValue Chain = Node->getOperand(0);
7651 SDValue In1 = Node->getOperand(1);
7652 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7653 Node->getOperand(2), DAG.getIntPtrConstant(0));
7654 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7655 Node->getOperand(2), DAG.getIntPtrConstant(1));
7656 SDValue Ops[] = { Chain, In1, In2L, In2H };
7657 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7659 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7660 cast<MemSDNode>(Node)->getMemOperand());
7661 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7662 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7663 Results.push_back(Result.getValue(2));
7666 /// ReplaceNodeResults - Replace a node with an illegal result type
7667 /// with a new node built out of custom code.
7668 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7669 SmallVectorImpl<SDValue>&Results,
7670 SelectionDAG &DAG) const {
7671 DebugLoc dl = N->getDebugLoc();
7672 switch (N->getOpcode()) {
7674 assert(false && "Do not know how to custom type legalize this operation!");
7676 case ISD::FP_TO_SINT: {
7677 std::pair<SDValue,SDValue> Vals =
7678 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7679 SDValue FIST = Vals.first, StackSlot = Vals.second;
7680 if (FIST.getNode() != 0) {
7681 EVT VT = N->getValueType(0);
7682 // Return a load from the stack slot.
7683 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7688 case ISD::READCYCLECOUNTER: {
7689 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7690 SDValue TheChain = N->getOperand(0);
7691 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7692 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7694 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7696 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7697 SDValue Ops[] = { eax, edx };
7698 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7699 Results.push_back(edx.getValue(1));
7702 case ISD::ATOMIC_CMP_SWAP: {
7703 EVT T = N->getValueType(0);
7704 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7705 SDValue cpInL, cpInH;
7706 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7707 DAG.getConstant(0, MVT::i32));
7708 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7709 DAG.getConstant(1, MVT::i32));
7710 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7711 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7713 SDValue swapInL, swapInH;
7714 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7715 DAG.getConstant(0, MVT::i32));
7716 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7717 DAG.getConstant(1, MVT::i32));
7718 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7720 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7721 swapInL.getValue(1));
7722 SDValue Ops[] = { swapInH.getValue(0),
7724 swapInH.getValue(1) };
7725 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7726 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7727 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7728 MVT::i32, Result.getValue(1));
7729 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7730 MVT::i32, cpOutL.getValue(2));
7731 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7732 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7733 Results.push_back(cpOutH.getValue(1));
7736 case ISD::ATOMIC_LOAD_ADD:
7737 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7739 case ISD::ATOMIC_LOAD_AND:
7740 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7742 case ISD::ATOMIC_LOAD_NAND:
7743 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7745 case ISD::ATOMIC_LOAD_OR:
7746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7748 case ISD::ATOMIC_LOAD_SUB:
7749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7751 case ISD::ATOMIC_LOAD_XOR:
7752 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7754 case ISD::ATOMIC_SWAP:
7755 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7760 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7762 default: return NULL;
7763 case X86ISD::BSF: return "X86ISD::BSF";
7764 case X86ISD::BSR: return "X86ISD::BSR";
7765 case X86ISD::SHLD: return "X86ISD::SHLD";
7766 case X86ISD::SHRD: return "X86ISD::SHRD";
7767 case X86ISD::FAND: return "X86ISD::FAND";
7768 case X86ISD::FOR: return "X86ISD::FOR";
7769 case X86ISD::FXOR: return "X86ISD::FXOR";
7770 case X86ISD::FSRL: return "X86ISD::FSRL";
7771 case X86ISD::FILD: return "X86ISD::FILD";
7772 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7773 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7774 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7775 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7776 case X86ISD::FLD: return "X86ISD::FLD";
7777 case X86ISD::FST: return "X86ISD::FST";
7778 case X86ISD::CALL: return "X86ISD::CALL";
7779 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7780 case X86ISD::BT: return "X86ISD::BT";
7781 case X86ISD::CMP: return "X86ISD::CMP";
7782 case X86ISD::COMI: return "X86ISD::COMI";
7783 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7784 case X86ISD::SETCC: return "X86ISD::SETCC";
7785 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7786 case X86ISD::CMOV: return "X86ISD::CMOV";
7787 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7788 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7789 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7790 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7791 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7792 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7793 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7794 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7795 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7796 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7797 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7798 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7799 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7800 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7801 case X86ISD::FMAX: return "X86ISD::FMAX";
7802 case X86ISD::FMIN: return "X86ISD::FMIN";
7803 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7804 case X86ISD::FRCP: return "X86ISD::FRCP";
7805 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7806 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
7807 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7808 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7809 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7810 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7811 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7812 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7813 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7814 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7815 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7816 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7817 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7818 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7819 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7820 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7821 case X86ISD::VSHL: return "X86ISD::VSHL";
7822 case X86ISD::VSRL: return "X86ISD::VSRL";
7823 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7824 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7825 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7826 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7827 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7828 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7829 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7830 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7831 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7832 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7833 case X86ISD::ADD: return "X86ISD::ADD";
7834 case X86ISD::SUB: return "X86ISD::SUB";
7835 case X86ISD::SMUL: return "X86ISD::SMUL";
7836 case X86ISD::UMUL: return "X86ISD::UMUL";
7837 case X86ISD::INC: return "X86ISD::INC";
7838 case X86ISD::DEC: return "X86ISD::DEC";
7839 case X86ISD::OR: return "X86ISD::OR";
7840 case X86ISD::XOR: return "X86ISD::XOR";
7841 case X86ISD::AND: return "X86ISD::AND";
7842 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7843 case X86ISD::PTEST: return "X86ISD::PTEST";
7844 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7845 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7849 // isLegalAddressingMode - Return true if the addressing mode represented
7850 // by AM is legal for this target, for a load/store of the specified type.
7851 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7852 const Type *Ty) const {
7853 // X86 supports extremely general addressing modes.
7854 CodeModel::Model M = getTargetMachine().getCodeModel();
7856 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7857 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7862 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7864 // If a reference to this global requires an extra load, we can't fold it.
7865 if (isGlobalStubReference(GVFlags))
7868 // If BaseGV requires a register for the PIC base, we cannot also have a
7869 // BaseReg specified.
7870 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7873 // If lower 4G is not available, then we must use rip-relative addressing.
7874 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7884 // These scales always work.
7889 // These scales are formed with basereg+scalereg. Only accept if there is
7894 default: // Other stuff never works.
7902 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7903 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7905 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7906 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7907 if (NumBits1 <= NumBits2)
7912 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7913 if (!VT1.isInteger() || !VT2.isInteger())
7915 unsigned NumBits1 = VT1.getSizeInBits();
7916 unsigned NumBits2 = VT2.getSizeInBits();
7917 if (NumBits1 <= NumBits2)
7922 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7923 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7924 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7927 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7928 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7929 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7932 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7933 // i16 instructions are longer (0x66 prefix) and potentially slower.
7934 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7937 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7938 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7939 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7940 /// are assumed to be legal.
7942 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7944 // Very little shuffling can be done for 64-bit vectors right now.
7945 if (VT.getSizeInBits() == 64)
7946 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
7948 // FIXME: pshufb, blends, shifts.
7949 return (VT.getVectorNumElements() == 2 ||
7950 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7951 isMOVLMask(M, VT) ||
7952 isSHUFPMask(M, VT) ||
7953 isPSHUFDMask(M, VT) ||
7954 isPSHUFHWMask(M, VT) ||
7955 isPSHUFLWMask(M, VT) ||
7956 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7957 isUNPCKLMask(M, VT) ||
7958 isUNPCKHMask(M, VT) ||
7959 isUNPCKL_v_undef_Mask(M, VT) ||
7960 isUNPCKH_v_undef_Mask(M, VT));
7964 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7966 unsigned NumElts = VT.getVectorNumElements();
7967 // FIXME: This collection of masks seems suspect.
7970 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7971 return (isMOVLMask(Mask, VT) ||
7972 isCommutedMOVLMask(Mask, VT, true) ||
7973 isSHUFPMask(Mask, VT) ||
7974 isCommutedSHUFPMask(Mask, VT));
7979 //===----------------------------------------------------------------------===//
7980 // X86 Scheduler Hooks
7981 //===----------------------------------------------------------------------===//
7983 // private utility function
7985 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7986 MachineBasicBlock *MBB,
7994 TargetRegisterClass *RC,
7995 bool invSrc) const {
7996 // For the atomic bitwise operator, we generate
7999 // ld t1 = [bitinstr.addr]
8000 // op t2 = t1, [bitinstr.val]
8002 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8004 // fallthrough -->nextMBB
8005 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8006 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8007 MachineFunction::iterator MBBIter = MBB;
8010 /// First build the CFG
8011 MachineFunction *F = MBB->getParent();
8012 MachineBasicBlock *thisMBB = MBB;
8013 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8014 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8015 F->insert(MBBIter, newMBB);
8016 F->insert(MBBIter, nextMBB);
8018 // Move all successors to thisMBB to nextMBB
8019 nextMBB->transferSuccessors(thisMBB);
8021 // Update thisMBB to fall through to newMBB
8022 thisMBB->addSuccessor(newMBB);
8024 // newMBB jumps to itself and fall through to nextMBB
8025 newMBB->addSuccessor(nextMBB);
8026 newMBB->addSuccessor(newMBB);
8028 // Insert instructions into newMBB based on incoming instruction
8029 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8030 "unexpected number of operands");
8031 DebugLoc dl = bInstr->getDebugLoc();
8032 MachineOperand& destOper = bInstr->getOperand(0);
8033 MachineOperand* argOpers[2 + X86AddrNumOperands];
8034 int numArgs = bInstr->getNumOperands() - 1;
8035 for (int i=0; i < numArgs; ++i)
8036 argOpers[i] = &bInstr->getOperand(i+1);
8038 // x86 address has 4 operands: base, index, scale, and displacement
8039 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8040 int valArgIndx = lastAddrIndx + 1;
8042 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8043 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8044 for (int i=0; i <= lastAddrIndx; ++i)
8045 (*MIB).addOperand(*argOpers[i]);
8047 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8049 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8054 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8055 assert((argOpers[valArgIndx]->isReg() ||
8056 argOpers[valArgIndx]->isImm()) &&
8058 if (argOpers[valArgIndx]->isReg())
8059 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8061 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8063 (*MIB).addOperand(*argOpers[valArgIndx]);
8065 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
8068 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8069 for (int i=0; i <= lastAddrIndx; ++i)
8070 (*MIB).addOperand(*argOpers[i]);
8072 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8073 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8074 bInstr->memoperands_end());
8076 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
8080 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8082 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8086 // private utility function: 64 bit atomics on 32 bit host.
8088 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8089 MachineBasicBlock *MBB,
8094 bool invSrc) const {
8095 // For the atomic bitwise operator, we generate
8096 // thisMBB (instructions are in pairs, except cmpxchg8b)
8097 // ld t1,t2 = [bitinstr.addr]
8099 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8100 // op t5, t6 <- out1, out2, [bitinstr.val]
8101 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8102 // mov ECX, EBX <- t5, t6
8103 // mov EAX, EDX <- t1, t2
8104 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8105 // mov t3, t4 <- EAX, EDX
8107 // result in out1, out2
8108 // fallthrough -->nextMBB
8110 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8111 const unsigned LoadOpc = X86::MOV32rm;
8112 const unsigned copyOpc = X86::MOV32rr;
8113 const unsigned NotOpc = X86::NOT32r;
8114 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8115 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8116 MachineFunction::iterator MBBIter = MBB;
8119 /// First build the CFG
8120 MachineFunction *F = MBB->getParent();
8121 MachineBasicBlock *thisMBB = MBB;
8122 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8123 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8124 F->insert(MBBIter, newMBB);
8125 F->insert(MBBIter, nextMBB);
8127 // Move all successors to thisMBB to nextMBB
8128 nextMBB->transferSuccessors(thisMBB);
8130 // Update thisMBB to fall through to newMBB
8131 thisMBB->addSuccessor(newMBB);
8133 // newMBB jumps to itself and fall through to nextMBB
8134 newMBB->addSuccessor(nextMBB);
8135 newMBB->addSuccessor(newMBB);
8137 DebugLoc dl = bInstr->getDebugLoc();
8138 // Insert instructions into newMBB based on incoming instruction
8139 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8140 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8141 "unexpected number of operands");
8142 MachineOperand& dest1Oper = bInstr->getOperand(0);
8143 MachineOperand& dest2Oper = bInstr->getOperand(1);
8144 MachineOperand* argOpers[2 + X86AddrNumOperands];
8145 for (int i=0; i < 2 + X86AddrNumOperands; ++i) {
8146 argOpers[i] = &bInstr->getOperand(i+2);
8148 // We use some of the operands multiple times, so conservatively just
8149 // clear any kill flags that might be present.
8150 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8151 argOpers[i]->setIsKill(false);
8154 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8155 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8157 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8158 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8159 for (int i=0; i <= lastAddrIndx; ++i)
8160 (*MIB).addOperand(*argOpers[i]);
8161 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8162 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8163 // add 4 to displacement.
8164 for (int i=0; i <= lastAddrIndx-2; ++i)
8165 (*MIB).addOperand(*argOpers[i]);
8166 MachineOperand newOp3 = *(argOpers[3]);
8168 newOp3.setImm(newOp3.getImm()+4);
8170 newOp3.setOffset(newOp3.getOffset()+4);
8171 (*MIB).addOperand(newOp3);
8172 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8174 // t3/4 are defined later, at the bottom of the loop
8175 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8176 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8177 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8178 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8179 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8180 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8182 // The subsequent operations should be using the destination registers of
8183 //the PHI instructions.
8185 t1 = F->getRegInfo().createVirtualRegister(RC);
8186 t2 = F->getRegInfo().createVirtualRegister(RC);
8187 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8188 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8190 t1 = dest1Oper.getReg();
8191 t2 = dest2Oper.getReg();
8194 int valArgIndx = lastAddrIndx + 1;
8195 assert((argOpers[valArgIndx]->isReg() ||
8196 argOpers[valArgIndx]->isImm()) &&
8198 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8199 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8200 if (argOpers[valArgIndx]->isReg())
8201 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8203 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8204 if (regOpcL != X86::MOV32rr)
8206 (*MIB).addOperand(*argOpers[valArgIndx]);
8207 assert(argOpers[valArgIndx + 1]->isReg() ==
8208 argOpers[valArgIndx]->isReg());
8209 assert(argOpers[valArgIndx + 1]->isImm() ==
8210 argOpers[valArgIndx]->isImm());
8211 if (argOpers[valArgIndx + 1]->isReg())
8212 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8214 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8215 if (regOpcH != X86::MOV32rr)
8217 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8219 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8221 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8224 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8226 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8229 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8230 for (int i=0; i <= lastAddrIndx; ++i)
8231 (*MIB).addOperand(*argOpers[i]);
8233 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8234 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8235 bInstr->memoperands_end());
8237 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8238 MIB.addReg(X86::EAX);
8239 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8240 MIB.addReg(X86::EDX);
8243 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8245 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8249 // private utility function
8251 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8252 MachineBasicBlock *MBB,
8253 unsigned cmovOpc) const {
8254 // For the atomic min/max operator, we generate
8257 // ld t1 = [min/max.addr]
8258 // mov t2 = [min/max.val]
8260 // cmov[cond] t2 = t1
8262 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8264 // fallthrough -->nextMBB
8266 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8267 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8268 MachineFunction::iterator MBBIter = MBB;
8271 /// First build the CFG
8272 MachineFunction *F = MBB->getParent();
8273 MachineBasicBlock *thisMBB = MBB;
8274 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8275 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8276 F->insert(MBBIter, newMBB);
8277 F->insert(MBBIter, nextMBB);
8279 // Move all successors of thisMBB to nextMBB
8280 nextMBB->transferSuccessors(thisMBB);
8282 // Update thisMBB to fall through to newMBB
8283 thisMBB->addSuccessor(newMBB);
8285 // newMBB jumps to newMBB and fall through to nextMBB
8286 newMBB->addSuccessor(nextMBB);
8287 newMBB->addSuccessor(newMBB);
8289 DebugLoc dl = mInstr->getDebugLoc();
8290 // Insert instructions into newMBB based on incoming instruction
8291 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8292 "unexpected number of operands");
8293 MachineOperand& destOper = mInstr->getOperand(0);
8294 MachineOperand* argOpers[2 + X86AddrNumOperands];
8295 int numArgs = mInstr->getNumOperands() - 1;
8296 for (int i=0; i < numArgs; ++i)
8297 argOpers[i] = &mInstr->getOperand(i+1);
8299 // x86 address has 4 operands: base, index, scale, and displacement
8300 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8301 int valArgIndx = lastAddrIndx + 1;
8303 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8304 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8305 for (int i=0; i <= lastAddrIndx; ++i)
8306 (*MIB).addOperand(*argOpers[i]);
8308 // We only support register and immediate values
8309 assert((argOpers[valArgIndx]->isReg() ||
8310 argOpers[valArgIndx]->isImm()) &&
8313 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8314 if (argOpers[valArgIndx]->isReg())
8315 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8317 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8318 (*MIB).addOperand(*argOpers[valArgIndx]);
8320 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8323 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8328 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8329 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8333 // Cmp and exchange if none has modified the memory location
8334 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8335 for (int i=0; i <= lastAddrIndx; ++i)
8336 (*MIB).addOperand(*argOpers[i]);
8338 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8339 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8340 mInstr->memoperands_end());
8342 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8343 MIB.addReg(X86::EAX);
8346 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8348 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8352 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8353 // all of this code can be replaced with that in the .td file.
8355 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8356 unsigned numArgs, bool memArg) const {
8358 MachineFunction *F = BB->getParent();
8359 DebugLoc dl = MI->getDebugLoc();
8360 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8364 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8366 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8368 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8370 for (unsigned i = 0; i < numArgs; ++i) {
8371 MachineOperand &Op = MI->getOperand(i+1);
8373 if (!(Op.isReg() && Op.isImplicit()))
8377 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8380 F->DeleteMachineInstr(MI);
8386 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8388 MachineBasicBlock *MBB) const {
8389 // Emit code to save XMM registers to the stack. The ABI says that the
8390 // number of registers to save is given in %al, so it's theoretically
8391 // possible to do an indirect jump trick to avoid saving all of them,
8392 // however this code takes a simpler approach and just executes all
8393 // of the stores if %al is non-zero. It's less code, and it's probably
8394 // easier on the hardware branch predictor, and stores aren't all that
8395 // expensive anyway.
8397 // Create the new basic blocks. One block contains all the XMM stores,
8398 // and one block is the final destination regardless of whether any
8399 // stores were performed.
8400 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8401 MachineFunction *F = MBB->getParent();
8402 MachineFunction::iterator MBBIter = MBB;
8404 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8405 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8406 F->insert(MBBIter, XMMSaveMBB);
8407 F->insert(MBBIter, EndMBB);
8410 // Move any original successors of MBB to the end block.
8411 EndMBB->transferSuccessors(MBB);
8412 // The original block will now fall through to the XMM save block.
8413 MBB->addSuccessor(XMMSaveMBB);
8414 // The XMMSaveMBB will fall through to the end block.
8415 XMMSaveMBB->addSuccessor(EndMBB);
8417 // Now add the instructions.
8418 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8419 DebugLoc DL = MI->getDebugLoc();
8421 unsigned CountReg = MI->getOperand(0).getReg();
8422 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8423 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8425 if (!Subtarget->isTargetWin64()) {
8426 // If %al is 0, branch around the XMM save block.
8427 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8428 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8429 MBB->addSuccessor(EndMBB);
8432 // In the XMM save block, save all the XMM argument registers.
8433 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8434 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8435 MachineMemOperand *MMO =
8436 F->getMachineMemOperand(
8437 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8438 MachineMemOperand::MOStore, Offset,
8439 /*Size=*/16, /*Align=*/16);
8440 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8441 .addFrameIndex(RegSaveFrameIndex)
8442 .addImm(/*Scale=*/1)
8443 .addReg(/*IndexReg=*/0)
8444 .addImm(/*Disp=*/Offset)
8445 .addReg(/*Segment=*/0)
8446 .addReg(MI->getOperand(i).getReg())
8447 .addMemOperand(MMO);
8450 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8456 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8457 MachineBasicBlock *BB) const {
8458 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8459 DebugLoc DL = MI->getDebugLoc();
8461 // To "insert" a SELECT_CC instruction, we actually have to insert the
8462 // diamond control-flow pattern. The incoming instruction knows the
8463 // destination vreg to set, the condition code register to branch on, the
8464 // true/false values to select between, and a branch opcode to use.
8465 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8466 MachineFunction::iterator It = BB;
8472 // cmpTY ccX, r1, r2
8474 // fallthrough --> copy0MBB
8475 MachineBasicBlock *thisMBB = BB;
8476 MachineFunction *F = BB->getParent();
8477 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8478 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8480 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8482 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8483 F->insert(It, copy0MBB);
8484 F->insert(It, sinkMBB);
8486 // Update machine-CFG edges by first adding all successors of the current
8487 // block to the new block which will contain the Phi node for the select.
8488 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8489 E = BB->succ_end(); I != E; ++I)
8490 sinkMBB->addSuccessor(*I);
8492 // Next, remove all successors of the current block, and add the true
8493 // and fallthrough blocks as its successors.
8494 while (!BB->succ_empty())
8495 BB->removeSuccessor(BB->succ_begin());
8497 // Add the true and fallthrough blocks as its successors.
8498 BB->addSuccessor(copy0MBB);
8499 BB->addSuccessor(sinkMBB);
8501 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8502 // live into the sink and copy blocks.
8503 const MachineFunction *MF = BB->getParent();
8504 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8505 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8506 const MachineInstr *Term = BB->getFirstTerminator();
8508 for (unsigned I = 0, E = Term->getNumOperands(); I != E; ++I) {
8509 const MachineOperand &MO = Term->getOperand(I);
8510 if (!MO.isReg() || MO.isKill() || MO.isDead()) continue;
8511 unsigned Reg = MO.getReg();
8512 if (Reg != X86::EFLAGS) continue;
8513 copy0MBB->addLiveIn(Reg);
8514 sinkMBB->addLiveIn(Reg);
8518 // %FalseValue = ...
8519 // # fallthrough to sinkMBB
8520 copy0MBB->addSuccessor(sinkMBB);
8523 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8525 BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8526 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8527 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8529 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8534 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8535 MachineBasicBlock *BB) const {
8536 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8537 DebugLoc DL = MI->getDebugLoc();
8538 MachineFunction *F = BB->getParent();
8540 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8541 // non-trivial part is impdef of ESP.
8542 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8545 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8546 .addExternalSymbol("_alloca")
8547 .addReg(X86::EAX, RegState::Implicit)
8548 .addReg(X86::ESP, RegState::Implicit)
8549 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8550 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8552 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8557 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8558 MachineBasicBlock *BB) const {
8559 // This is pretty easy. We're taking the value that we received from
8560 // our load from the relocation, sticking it in either RDI (x86-64)
8561 // or EAX and doing an indirect call. The return value will then
8562 // be in the normal return register.
8563 const X86InstrInfo *TII
8564 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8565 DebugLoc DL = MI->getDebugLoc();
8566 MachineFunction *F = BB->getParent();
8568 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8570 if (Subtarget->is64Bit()) {
8571 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rm), X86::RDI)
8573 .addImm(0).addReg(0)
8574 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8575 MI->getOperand(3).getTargetFlags())
8577 MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
8578 addDirectMem(MIB, X86::RDI).addReg(0);
8579 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8580 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8582 .addImm(0).addReg(0)
8583 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8584 MI->getOperand(3).getTargetFlags())
8586 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8587 addDirectMem(MIB, X86::EAX).addReg(0);
8589 MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
8590 .addReg(TII->getGlobalBaseReg(F))
8591 .addImm(0).addReg(0)
8592 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8593 MI->getOperand(3).getTargetFlags())
8595 MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
8596 addDirectMem(MIB, X86::EAX).addReg(0);
8599 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8604 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8605 MachineBasicBlock *BB) const {
8606 switch (MI->getOpcode()) {
8607 default: assert(false && "Unexpected instr type to insert");
8608 case X86::MINGW_ALLOCA:
8609 return EmitLoweredMingwAlloca(MI, BB);
8610 case X86::TLSCall_32:
8611 case X86::TLSCall_64:
8612 return EmitLoweredTLSCall(MI, BB);
8614 case X86::CMOV_V1I64:
8615 case X86::CMOV_FR32:
8616 case X86::CMOV_FR64:
8617 case X86::CMOV_V4F32:
8618 case X86::CMOV_V2F64:
8619 case X86::CMOV_V2I64:
8620 case X86::CMOV_GR16:
8621 case X86::CMOV_GR32:
8622 case X86::CMOV_RFP32:
8623 case X86::CMOV_RFP64:
8624 case X86::CMOV_RFP80:
8625 return EmitLoweredSelect(MI, BB);
8627 case X86::FP32_TO_INT16_IN_MEM:
8628 case X86::FP32_TO_INT32_IN_MEM:
8629 case X86::FP32_TO_INT64_IN_MEM:
8630 case X86::FP64_TO_INT16_IN_MEM:
8631 case X86::FP64_TO_INT32_IN_MEM:
8632 case X86::FP64_TO_INT64_IN_MEM:
8633 case X86::FP80_TO_INT16_IN_MEM:
8634 case X86::FP80_TO_INT32_IN_MEM:
8635 case X86::FP80_TO_INT64_IN_MEM: {
8636 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8637 DebugLoc DL = MI->getDebugLoc();
8639 // Change the floating point control register to use "round towards zero"
8640 // mode when truncating to an integer value.
8641 MachineFunction *F = BB->getParent();
8642 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8643 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8645 // Load the old value of the high byte of the control word...
8647 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8648 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8651 // Set the high part to be round to zero...
8652 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8655 // Reload the modified control word now...
8656 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8658 // Restore the memory image of control word to original value
8659 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8662 // Get the X86 opcode to use.
8664 switch (MI->getOpcode()) {
8665 default: llvm_unreachable("illegal opcode!");
8666 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8667 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8668 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8669 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8670 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8671 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8672 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8673 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8674 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8678 MachineOperand &Op = MI->getOperand(0);
8680 AM.BaseType = X86AddressMode::RegBase;
8681 AM.Base.Reg = Op.getReg();
8683 AM.BaseType = X86AddressMode::FrameIndexBase;
8684 AM.Base.FrameIndex = Op.getIndex();
8686 Op = MI->getOperand(1);
8688 AM.Scale = Op.getImm();
8689 Op = MI->getOperand(2);
8691 AM.IndexReg = Op.getImm();
8692 Op = MI->getOperand(3);
8693 if (Op.isGlobal()) {
8694 AM.GV = Op.getGlobal();
8696 AM.Disp = Op.getImm();
8698 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8699 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8701 // Reload the original control word now.
8702 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8704 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8707 // String/text processing lowering.
8708 case X86::PCMPISTRM128REG:
8709 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8710 case X86::PCMPISTRM128MEM:
8711 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8712 case X86::PCMPESTRM128REG:
8713 return EmitPCMP(MI, BB, 5, false /* in mem */);
8714 case X86::PCMPESTRM128MEM:
8715 return EmitPCMP(MI, BB, 5, true /* in mem */);
8718 case X86::ATOMAND32:
8719 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8720 X86::AND32ri, X86::MOV32rm,
8721 X86::LCMPXCHG32, X86::MOV32rr,
8722 X86::NOT32r, X86::EAX,
8723 X86::GR32RegisterClass);
8725 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8726 X86::OR32ri, X86::MOV32rm,
8727 X86::LCMPXCHG32, X86::MOV32rr,
8728 X86::NOT32r, X86::EAX,
8729 X86::GR32RegisterClass);
8730 case X86::ATOMXOR32:
8731 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8732 X86::XOR32ri, X86::MOV32rm,
8733 X86::LCMPXCHG32, X86::MOV32rr,
8734 X86::NOT32r, X86::EAX,
8735 X86::GR32RegisterClass);
8736 case X86::ATOMNAND32:
8737 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8738 X86::AND32ri, X86::MOV32rm,
8739 X86::LCMPXCHG32, X86::MOV32rr,
8740 X86::NOT32r, X86::EAX,
8741 X86::GR32RegisterClass, true);
8742 case X86::ATOMMIN32:
8743 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8744 case X86::ATOMMAX32:
8745 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8746 case X86::ATOMUMIN32:
8747 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8748 case X86::ATOMUMAX32:
8749 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8751 case X86::ATOMAND16:
8752 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8753 X86::AND16ri, X86::MOV16rm,
8754 X86::LCMPXCHG16, X86::MOV16rr,
8755 X86::NOT16r, X86::AX,
8756 X86::GR16RegisterClass);
8758 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8759 X86::OR16ri, X86::MOV16rm,
8760 X86::LCMPXCHG16, X86::MOV16rr,
8761 X86::NOT16r, X86::AX,
8762 X86::GR16RegisterClass);
8763 case X86::ATOMXOR16:
8764 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8765 X86::XOR16ri, X86::MOV16rm,
8766 X86::LCMPXCHG16, X86::MOV16rr,
8767 X86::NOT16r, X86::AX,
8768 X86::GR16RegisterClass);
8769 case X86::ATOMNAND16:
8770 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8771 X86::AND16ri, X86::MOV16rm,
8772 X86::LCMPXCHG16, X86::MOV16rr,
8773 X86::NOT16r, X86::AX,
8774 X86::GR16RegisterClass, true);
8775 case X86::ATOMMIN16:
8776 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8777 case X86::ATOMMAX16:
8778 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8779 case X86::ATOMUMIN16:
8780 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8781 case X86::ATOMUMAX16:
8782 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8785 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8786 X86::AND8ri, X86::MOV8rm,
8787 X86::LCMPXCHG8, X86::MOV8rr,
8788 X86::NOT8r, X86::AL,
8789 X86::GR8RegisterClass);
8791 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8792 X86::OR8ri, X86::MOV8rm,
8793 X86::LCMPXCHG8, X86::MOV8rr,
8794 X86::NOT8r, X86::AL,
8795 X86::GR8RegisterClass);
8797 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8798 X86::XOR8ri, X86::MOV8rm,
8799 X86::LCMPXCHG8, X86::MOV8rr,
8800 X86::NOT8r, X86::AL,
8801 X86::GR8RegisterClass);
8802 case X86::ATOMNAND8:
8803 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8804 X86::AND8ri, X86::MOV8rm,
8805 X86::LCMPXCHG8, X86::MOV8rr,
8806 X86::NOT8r, X86::AL,
8807 X86::GR8RegisterClass, true);
8808 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8809 // This group is for 64-bit host.
8810 case X86::ATOMAND64:
8811 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8812 X86::AND64ri32, X86::MOV64rm,
8813 X86::LCMPXCHG64, X86::MOV64rr,
8814 X86::NOT64r, X86::RAX,
8815 X86::GR64RegisterClass);
8817 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8818 X86::OR64ri32, X86::MOV64rm,
8819 X86::LCMPXCHG64, X86::MOV64rr,
8820 X86::NOT64r, X86::RAX,
8821 X86::GR64RegisterClass);
8822 case X86::ATOMXOR64:
8823 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8824 X86::XOR64ri32, X86::MOV64rm,
8825 X86::LCMPXCHG64, X86::MOV64rr,
8826 X86::NOT64r, X86::RAX,
8827 X86::GR64RegisterClass);
8828 case X86::ATOMNAND64:
8829 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8830 X86::AND64ri32, X86::MOV64rm,
8831 X86::LCMPXCHG64, X86::MOV64rr,
8832 X86::NOT64r, X86::RAX,
8833 X86::GR64RegisterClass, true);
8834 case X86::ATOMMIN64:
8835 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8836 case X86::ATOMMAX64:
8837 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8838 case X86::ATOMUMIN64:
8839 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8840 case X86::ATOMUMAX64:
8841 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8843 // This group does 64-bit operations on a 32-bit host.
8844 case X86::ATOMAND6432:
8845 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8846 X86::AND32rr, X86::AND32rr,
8847 X86::AND32ri, X86::AND32ri,
8849 case X86::ATOMOR6432:
8850 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8851 X86::OR32rr, X86::OR32rr,
8852 X86::OR32ri, X86::OR32ri,
8854 case X86::ATOMXOR6432:
8855 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8856 X86::XOR32rr, X86::XOR32rr,
8857 X86::XOR32ri, X86::XOR32ri,
8859 case X86::ATOMNAND6432:
8860 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8861 X86::AND32rr, X86::AND32rr,
8862 X86::AND32ri, X86::AND32ri,
8864 case X86::ATOMADD6432:
8865 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8866 X86::ADD32rr, X86::ADC32rr,
8867 X86::ADD32ri, X86::ADC32ri,
8869 case X86::ATOMSUB6432:
8870 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8871 X86::SUB32rr, X86::SBB32rr,
8872 X86::SUB32ri, X86::SBB32ri,
8874 case X86::ATOMSWAP6432:
8875 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8876 X86::MOV32rr, X86::MOV32rr,
8877 X86::MOV32ri, X86::MOV32ri,
8879 case X86::VASTART_SAVE_XMM_REGS:
8880 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8884 //===----------------------------------------------------------------------===//
8885 // X86 Optimization Hooks
8886 //===----------------------------------------------------------------------===//
8888 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8892 const SelectionDAG &DAG,
8893 unsigned Depth) const {
8894 unsigned Opc = Op.getOpcode();
8895 assert((Opc >= ISD::BUILTIN_OP_END ||
8896 Opc == ISD::INTRINSIC_WO_CHAIN ||
8897 Opc == ISD::INTRINSIC_W_CHAIN ||
8898 Opc == ISD::INTRINSIC_VOID) &&
8899 "Should use MaskedValueIsZero if you don't know whether Op"
8900 " is a target node!");
8902 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8914 // These nodes' second result is a boolean.
8915 if (Op.getResNo() == 0)
8919 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8920 Mask.getBitWidth() - 1);
8925 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8926 /// node is a GlobalAddress + offset.
8927 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8928 const GlobalValue* &GA,
8929 int64_t &Offset) const {
8930 if (N->getOpcode() == X86ISD::Wrapper) {
8931 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8932 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8933 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8937 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8940 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8941 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8942 /// if the load addresses are consecutive, non-overlapping, and in the right
8944 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8945 const TargetLowering &TLI) {
8946 DebugLoc dl = N->getDebugLoc();
8947 EVT VT = N->getValueType(0);
8948 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8950 if (VT.getSizeInBits() != 128)
8953 SmallVector<SDValue, 16> Elts;
8954 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
8955 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
8957 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
8960 /// PerformShuffleCombine - Detect vector gather/scatter index generation
8961 /// and convert it from being a bunch of shuffles and extracts to a simple
8962 /// store and scalar loads to extract the elements.
8963 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
8964 const TargetLowering &TLI) {
8965 SDValue InputVector = N->getOperand(0);
8967 // Only operate on vectors of 4 elements, where the alternative shuffling
8968 // gets to be more expensive.
8969 if (InputVector.getValueType() != MVT::v4i32)
8972 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
8973 // single use which is a sign-extend or zero-extend, and all elements are
8975 SmallVector<SDNode *, 4> Uses;
8976 unsigned ExtractedElements = 0;
8977 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
8978 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
8979 if (UI.getUse().getResNo() != InputVector.getResNo())
8982 SDNode *Extract = *UI;
8983 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8986 if (Extract->getValueType(0) != MVT::i32)
8988 if (!Extract->hasOneUse())
8990 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
8991 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
8993 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
8996 // Record which element was extracted.
8997 ExtractedElements |=
8998 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9000 Uses.push_back(Extract);
9003 // If not all the elements were used, this may not be worthwhile.
9004 if (ExtractedElements != 15)
9007 // Ok, we've now decided to do the transformation.
9008 DebugLoc dl = InputVector.getDebugLoc();
9010 // Store the value to a temporary stack slot.
9011 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9012 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL, 0,
9015 // Replace each use (extract) with a load of the appropriate element.
9016 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9017 UE = Uses.end(); UI != UE; ++UI) {
9018 SDNode *Extract = *UI;
9020 // Compute the element's address.
9021 SDValue Idx = Extract->getOperand(1);
9023 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9024 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9025 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9027 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), OffsetVal, StackPtr);
9030 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch, ScalarAddr,
9031 NULL, 0, false, false, 0);
9033 // Replace the exact with the load.
9034 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9037 // The replacement was made in place; don't return anything.
9041 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9042 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9043 const X86Subtarget *Subtarget) {
9044 DebugLoc DL = N->getDebugLoc();
9045 SDValue Cond = N->getOperand(0);
9046 // Get the LHS/RHS of the select.
9047 SDValue LHS = N->getOperand(1);
9048 SDValue RHS = N->getOperand(2);
9050 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9051 // instructions match the semantics of the common C idiom x<y?x:y but not
9052 // x<=y?x:y, because of how they handle negative zero (which can be
9053 // ignored in unsafe-math mode).
9054 if (Subtarget->hasSSE2() &&
9055 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9056 Cond.getOpcode() == ISD::SETCC) {
9057 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9059 unsigned Opcode = 0;
9060 // Check for x CC y ? x : y.
9061 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9062 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9066 // Converting this to a min would handle NaNs incorrectly, and swapping
9067 // the operands would cause it to handle comparisons between positive
9068 // and negative zero incorrectly.
9069 if (!FiniteOnlyFPMath() &&
9070 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9071 if (!UnsafeFPMath &&
9072 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9074 std::swap(LHS, RHS);
9076 Opcode = X86ISD::FMIN;
9079 // Converting this to a min would handle comparisons between positive
9080 // and negative zero incorrectly.
9081 if (!UnsafeFPMath &&
9082 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9084 Opcode = X86ISD::FMIN;
9087 // Converting this to a min would handle both negative zeros and NaNs
9088 // incorrectly, but we can swap the operands to fix both.
9089 std::swap(LHS, RHS);
9093 Opcode = X86ISD::FMIN;
9097 // Converting this to a max would handle comparisons between positive
9098 // and negative zero incorrectly.
9099 if (!UnsafeFPMath &&
9100 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9102 Opcode = X86ISD::FMAX;
9105 // Converting this to a max would handle NaNs incorrectly, and swapping
9106 // the operands would cause it to handle comparisons between positive
9107 // and negative zero incorrectly.
9108 if (!FiniteOnlyFPMath() &&
9109 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
9110 if (!UnsafeFPMath &&
9111 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9113 std::swap(LHS, RHS);
9115 Opcode = X86ISD::FMAX;
9118 // Converting this to a max would handle both negative zeros and NaNs
9119 // incorrectly, but we can swap the operands to fix both.
9120 std::swap(LHS, RHS);
9124 Opcode = X86ISD::FMAX;
9127 // Check for x CC y ? y : x -- a min/max with reversed arms.
9128 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9129 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9133 // Converting this to a min would handle comparisons between positive
9134 // and negative zero incorrectly, and swapping the operands would
9135 // cause it to handle NaNs incorrectly.
9136 if (!UnsafeFPMath &&
9137 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9138 if (!FiniteOnlyFPMath() &&
9139 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9141 std::swap(LHS, RHS);
9143 Opcode = X86ISD::FMIN;
9146 // Converting this to a min would handle NaNs incorrectly.
9147 if (!UnsafeFPMath &&
9148 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9150 Opcode = X86ISD::FMIN;
9153 // Converting this to a min would handle both negative zeros and NaNs
9154 // incorrectly, but we can swap the operands to fix both.
9155 std::swap(LHS, RHS);
9159 Opcode = X86ISD::FMIN;
9163 // Converting this to a max would handle NaNs incorrectly.
9164 if (!FiniteOnlyFPMath() &&
9165 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9167 Opcode = X86ISD::FMAX;
9170 // Converting this to a max would handle comparisons between positive
9171 // and negative zero incorrectly, and swapping the operands would
9172 // cause it to handle NaNs incorrectly.
9173 if (!UnsafeFPMath &&
9174 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9175 if (!FiniteOnlyFPMath() &&
9176 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9178 std::swap(LHS, RHS);
9180 Opcode = X86ISD::FMAX;
9183 // Converting this to a max would handle both negative zeros and NaNs
9184 // incorrectly, but we can swap the operands to fix both.
9185 std::swap(LHS, RHS);
9189 Opcode = X86ISD::FMAX;
9195 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9198 // If this is a select between two integer constants, try to do some
9200 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9201 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9202 // Don't do this for crazy integer types.
9203 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9204 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9205 // so that TrueC (the true value) is larger than FalseC.
9206 bool NeedsCondInvert = false;
9208 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9209 // Efficiently invertible.
9210 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9211 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9212 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9213 NeedsCondInvert = true;
9214 std::swap(TrueC, FalseC);
9217 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9218 if (FalseC->getAPIntValue() == 0 &&
9219 TrueC->getAPIntValue().isPowerOf2()) {
9220 if (NeedsCondInvert) // Invert the condition if needed.
9221 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9222 DAG.getConstant(1, Cond.getValueType()));
9224 // Zero extend the condition if needed.
9225 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9227 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9228 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9229 DAG.getConstant(ShAmt, MVT::i8));
9232 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9233 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9234 if (NeedsCondInvert) // Invert the condition if needed.
9235 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9236 DAG.getConstant(1, Cond.getValueType()));
9238 // Zero extend the condition if needed.
9239 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9240 FalseC->getValueType(0), Cond);
9241 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9242 SDValue(FalseC, 0));
9245 // Optimize cases that will turn into an LEA instruction. This requires
9246 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9247 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9248 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9249 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9251 bool isFastMultiplier = false;
9253 switch ((unsigned char)Diff) {
9255 case 1: // result = add base, cond
9256 case 2: // result = lea base( , cond*2)
9257 case 3: // result = lea base(cond, cond*2)
9258 case 4: // result = lea base( , cond*4)
9259 case 5: // result = lea base(cond, cond*4)
9260 case 8: // result = lea base( , cond*8)
9261 case 9: // result = lea base(cond, cond*8)
9262 isFastMultiplier = true;
9267 if (isFastMultiplier) {
9268 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9269 if (NeedsCondInvert) // Invert the condition if needed.
9270 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9271 DAG.getConstant(1, Cond.getValueType()));
9273 // Zero extend the condition if needed.
9274 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9276 // Scale the condition by the difference.
9278 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9279 DAG.getConstant(Diff, Cond.getValueType()));
9281 // Add the base if non-zero.
9282 if (FalseC->getAPIntValue() != 0)
9283 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9284 SDValue(FalseC, 0));
9294 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9295 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9296 TargetLowering::DAGCombinerInfo &DCI) {
9297 DebugLoc DL = N->getDebugLoc();
9299 // If the flag operand isn't dead, don't touch this CMOV.
9300 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9303 // If this is a select between two integer constants, try to do some
9304 // optimizations. Note that the operands are ordered the opposite of SELECT
9306 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9307 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9308 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9309 // larger than FalseC (the false value).
9310 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9312 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9313 CC = X86::GetOppositeBranchCondition(CC);
9314 std::swap(TrueC, FalseC);
9317 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9318 // This is efficient for any integer data type (including i8/i16) and
9320 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9321 SDValue Cond = N->getOperand(3);
9322 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9323 DAG.getConstant(CC, MVT::i8), Cond);
9325 // Zero extend the condition if needed.
9326 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9328 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9329 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9330 DAG.getConstant(ShAmt, MVT::i8));
9331 if (N->getNumValues() == 2) // Dead flag value?
9332 return DCI.CombineTo(N, Cond, SDValue());
9336 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9337 // for any integer data type, including i8/i16.
9338 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9339 SDValue Cond = N->getOperand(3);
9340 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9341 DAG.getConstant(CC, MVT::i8), Cond);
9343 // Zero extend the condition if needed.
9344 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9345 FalseC->getValueType(0), Cond);
9346 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9347 SDValue(FalseC, 0));
9349 if (N->getNumValues() == 2) // Dead flag value?
9350 return DCI.CombineTo(N, Cond, SDValue());
9354 // Optimize cases that will turn into an LEA instruction. This requires
9355 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9356 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9357 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9358 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9360 bool isFastMultiplier = false;
9362 switch ((unsigned char)Diff) {
9364 case 1: // result = add base, cond
9365 case 2: // result = lea base( , cond*2)
9366 case 3: // result = lea base(cond, cond*2)
9367 case 4: // result = lea base( , cond*4)
9368 case 5: // result = lea base(cond, cond*4)
9369 case 8: // result = lea base( , cond*8)
9370 case 9: // result = lea base(cond, cond*8)
9371 isFastMultiplier = true;
9376 if (isFastMultiplier) {
9377 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9378 SDValue Cond = N->getOperand(3);
9379 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9380 DAG.getConstant(CC, MVT::i8), Cond);
9381 // Zero extend the condition if needed.
9382 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9384 // Scale the condition by the difference.
9386 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9387 DAG.getConstant(Diff, Cond.getValueType()));
9389 // Add the base if non-zero.
9390 if (FalseC->getAPIntValue() != 0)
9391 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9392 SDValue(FalseC, 0));
9393 if (N->getNumValues() == 2) // Dead flag value?
9394 return DCI.CombineTo(N, Cond, SDValue());
9404 /// PerformMulCombine - Optimize a single multiply with constant into two
9405 /// in order to implement it with two cheaper instructions, e.g.
9406 /// LEA + SHL, LEA + LEA.
9407 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9408 TargetLowering::DAGCombinerInfo &DCI) {
9409 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9412 EVT VT = N->getValueType(0);
9416 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9419 uint64_t MulAmt = C->getZExtValue();
9420 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9423 uint64_t MulAmt1 = 0;
9424 uint64_t MulAmt2 = 0;
9425 if ((MulAmt % 9) == 0) {
9427 MulAmt2 = MulAmt / 9;
9428 } else if ((MulAmt % 5) == 0) {
9430 MulAmt2 = MulAmt / 5;
9431 } else if ((MulAmt % 3) == 0) {
9433 MulAmt2 = MulAmt / 3;
9436 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9437 DebugLoc DL = N->getDebugLoc();
9439 if (isPowerOf2_64(MulAmt2) &&
9440 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9441 // If second multiplifer is pow2, issue it first. We want the multiply by
9442 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9444 std::swap(MulAmt1, MulAmt2);
9447 if (isPowerOf2_64(MulAmt1))
9448 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9449 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9451 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9452 DAG.getConstant(MulAmt1, VT));
9454 if (isPowerOf2_64(MulAmt2))
9455 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9456 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9458 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9459 DAG.getConstant(MulAmt2, VT));
9461 // Do not add new nodes to DAG combiner worklist.
9462 DCI.CombineTo(N, NewMul, false);
9467 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9468 SDValue N0 = N->getOperand(0);
9469 SDValue N1 = N->getOperand(1);
9470 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9471 EVT VT = N0.getValueType();
9473 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9474 // since the result of setcc_c is all zero's or all ones.
9475 if (N1C && N0.getOpcode() == ISD::AND &&
9476 N0.getOperand(1).getOpcode() == ISD::Constant) {
9477 SDValue N00 = N0.getOperand(0);
9478 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9479 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9480 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9481 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9482 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9483 APInt ShAmt = N1C->getAPIntValue();
9484 Mask = Mask.shl(ShAmt);
9486 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9487 N00, DAG.getConstant(Mask, VT));
9494 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9496 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9497 const X86Subtarget *Subtarget) {
9498 EVT VT = N->getValueType(0);
9499 if (!VT.isVector() && VT.isInteger() &&
9500 N->getOpcode() == ISD::SHL)
9501 return PerformSHLCombine(N, DAG);
9503 // On X86 with SSE2 support, we can transform this to a vector shift if
9504 // all elements are shifted by the same amount. We can't do this in legalize
9505 // because the a constant vector is typically transformed to a constant pool
9506 // so we have no knowledge of the shift amount.
9507 if (!Subtarget->hasSSE2())
9510 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9513 SDValue ShAmtOp = N->getOperand(1);
9514 EVT EltVT = VT.getVectorElementType();
9515 DebugLoc DL = N->getDebugLoc();
9516 SDValue BaseShAmt = SDValue();
9517 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9518 unsigned NumElts = VT.getVectorNumElements();
9520 for (; i != NumElts; ++i) {
9521 SDValue Arg = ShAmtOp.getOperand(i);
9522 if (Arg.getOpcode() == ISD::UNDEF) continue;
9526 for (; i != NumElts; ++i) {
9527 SDValue Arg = ShAmtOp.getOperand(i);
9528 if (Arg.getOpcode() == ISD::UNDEF) continue;
9529 if (Arg != BaseShAmt) {
9533 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9534 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9535 SDValue InVec = ShAmtOp.getOperand(0);
9536 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9537 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9539 for (; i != NumElts; ++i) {
9540 SDValue Arg = InVec.getOperand(i);
9541 if (Arg.getOpcode() == ISD::UNDEF) continue;
9545 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9546 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9547 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9548 if (C->getZExtValue() == SplatIdx)
9549 BaseShAmt = InVec.getOperand(1);
9552 if (BaseShAmt.getNode() == 0)
9553 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9554 DAG.getIntPtrConstant(0));
9558 // The shift amount is an i32.
9559 if (EltVT.bitsGT(MVT::i32))
9560 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9561 else if (EltVT.bitsLT(MVT::i32))
9562 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9564 // The shift amount is identical so we can do a vector shift.
9565 SDValue ValOp = N->getOperand(0);
9566 switch (N->getOpcode()) {
9568 llvm_unreachable("Unknown shift opcode!");
9571 if (VT == MVT::v2i64)
9572 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9573 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9575 if (VT == MVT::v4i32)
9576 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9577 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9579 if (VT == MVT::v8i16)
9580 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9581 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9585 if (VT == MVT::v4i32)
9586 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9587 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9589 if (VT == MVT::v8i16)
9590 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9591 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9595 if (VT == MVT::v2i64)
9596 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9597 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9599 if (VT == MVT::v4i32)
9600 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9601 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9603 if (VT == MVT::v8i16)
9604 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9605 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9612 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9613 TargetLowering::DAGCombinerInfo &DCI,
9614 const X86Subtarget *Subtarget) {
9615 if (DCI.isBeforeLegalizeOps())
9618 EVT VT = N->getValueType(0);
9619 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9622 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9623 SDValue N0 = N->getOperand(0);
9624 SDValue N1 = N->getOperand(1);
9625 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9627 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9629 if (!N0.hasOneUse() || !N1.hasOneUse())
9632 SDValue ShAmt0 = N0.getOperand(1);
9633 if (ShAmt0.getValueType() != MVT::i8)
9635 SDValue ShAmt1 = N1.getOperand(1);
9636 if (ShAmt1.getValueType() != MVT::i8)
9638 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9639 ShAmt0 = ShAmt0.getOperand(0);
9640 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9641 ShAmt1 = ShAmt1.getOperand(0);
9643 DebugLoc DL = N->getDebugLoc();
9644 unsigned Opc = X86ISD::SHLD;
9645 SDValue Op0 = N0.getOperand(0);
9646 SDValue Op1 = N1.getOperand(0);
9647 if (ShAmt0.getOpcode() == ISD::SUB) {
9649 std::swap(Op0, Op1);
9650 std::swap(ShAmt0, ShAmt1);
9653 unsigned Bits = VT.getSizeInBits();
9654 if (ShAmt1.getOpcode() == ISD::SUB) {
9655 SDValue Sum = ShAmt1.getOperand(0);
9656 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9657 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9658 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9659 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9660 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9661 return DAG.getNode(Opc, DL, VT,
9663 DAG.getNode(ISD::TRUNCATE, DL,
9666 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9667 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9669 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9670 return DAG.getNode(Opc, DL, VT,
9671 N0.getOperand(0), N1.getOperand(0),
9672 DAG.getNode(ISD::TRUNCATE, DL,
9679 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9680 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9681 const X86Subtarget *Subtarget) {
9682 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9683 // the FP state in cases where an emms may be missing.
9684 // A preferable solution to the general problem is to figure out the right
9685 // places to insert EMMS. This qualifies as a quick hack.
9687 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9688 StoreSDNode *St = cast<StoreSDNode>(N);
9689 EVT VT = St->getValue().getValueType();
9690 if (VT.getSizeInBits() != 64)
9693 const Function *F = DAG.getMachineFunction().getFunction();
9694 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9695 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9696 && Subtarget->hasSSE2();
9697 if ((VT.isVector() ||
9698 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9699 isa<LoadSDNode>(St->getValue()) &&
9700 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9701 St->getChain().hasOneUse() && !St->isVolatile()) {
9702 SDNode* LdVal = St->getValue().getNode();
9704 int TokenFactorIndex = -1;
9705 SmallVector<SDValue, 8> Ops;
9706 SDNode* ChainVal = St->getChain().getNode();
9707 // Must be a store of a load. We currently handle two cases: the load
9708 // is a direct child, and it's under an intervening TokenFactor. It is
9709 // possible to dig deeper under nested TokenFactors.
9710 if (ChainVal == LdVal)
9711 Ld = cast<LoadSDNode>(St->getChain());
9712 else if (St->getValue().hasOneUse() &&
9713 ChainVal->getOpcode() == ISD::TokenFactor) {
9714 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9715 if (ChainVal->getOperand(i).getNode() == LdVal) {
9716 TokenFactorIndex = i;
9717 Ld = cast<LoadSDNode>(St->getValue());
9719 Ops.push_back(ChainVal->getOperand(i));
9723 if (!Ld || !ISD::isNormalLoad(Ld))
9726 // If this is not the MMX case, i.e. we are just turning i64 load/store
9727 // into f64 load/store, avoid the transformation if there are multiple
9728 // uses of the loaded value.
9729 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9732 DebugLoc LdDL = Ld->getDebugLoc();
9733 DebugLoc StDL = N->getDebugLoc();
9734 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9735 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9737 if (Subtarget->is64Bit() || F64IsLegal) {
9738 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9739 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9740 Ld->getBasePtr(), Ld->getSrcValue(),
9741 Ld->getSrcValueOffset(), Ld->isVolatile(),
9742 Ld->isNonTemporal(), Ld->getAlignment());
9743 SDValue NewChain = NewLd.getValue(1);
9744 if (TokenFactorIndex != -1) {
9745 Ops.push_back(NewChain);
9746 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9749 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9750 St->getSrcValue(), St->getSrcValueOffset(),
9751 St->isVolatile(), St->isNonTemporal(),
9752 St->getAlignment());
9755 // Otherwise, lower to two pairs of 32-bit loads / stores.
9756 SDValue LoAddr = Ld->getBasePtr();
9757 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9758 DAG.getConstant(4, MVT::i32));
9760 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9761 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9762 Ld->isVolatile(), Ld->isNonTemporal(),
9763 Ld->getAlignment());
9764 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9765 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9766 Ld->isVolatile(), Ld->isNonTemporal(),
9767 MinAlign(Ld->getAlignment(), 4));
9769 SDValue NewChain = LoLd.getValue(1);
9770 if (TokenFactorIndex != -1) {
9771 Ops.push_back(LoLd);
9772 Ops.push_back(HiLd);
9773 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9777 LoAddr = St->getBasePtr();
9778 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9779 DAG.getConstant(4, MVT::i32));
9781 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9782 St->getSrcValue(), St->getSrcValueOffset(),
9783 St->isVolatile(), St->isNonTemporal(),
9784 St->getAlignment());
9785 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9787 St->getSrcValueOffset() + 4,
9789 St->isNonTemporal(),
9790 MinAlign(St->getAlignment(), 4));
9791 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9796 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9797 /// X86ISD::FXOR nodes.
9798 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9799 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9800 // F[X]OR(0.0, x) -> x
9801 // F[X]OR(x, 0.0) -> x
9802 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9803 if (C->getValueAPF().isPosZero())
9804 return N->getOperand(1);
9805 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9806 if (C->getValueAPF().isPosZero())
9807 return N->getOperand(0);
9811 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9812 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9813 // FAND(0.0, x) -> 0.0
9814 // FAND(x, 0.0) -> 0.0
9815 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9816 if (C->getValueAPF().isPosZero())
9817 return N->getOperand(0);
9818 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9819 if (C->getValueAPF().isPosZero())
9820 return N->getOperand(1);
9824 static SDValue PerformBTCombine(SDNode *N,
9826 TargetLowering::DAGCombinerInfo &DCI) {
9827 // BT ignores high bits in the bit index operand.
9828 SDValue Op1 = N->getOperand(1);
9829 if (Op1.hasOneUse()) {
9830 unsigned BitWidth = Op1.getValueSizeInBits();
9831 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9832 APInt KnownZero, KnownOne;
9833 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
9834 !DCI.isBeforeLegalizeOps());
9835 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9836 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9837 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9838 DCI.CommitTargetLoweringOpt(TLO);
9843 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9844 SDValue Op = N->getOperand(0);
9845 if (Op.getOpcode() == ISD::BIT_CONVERT)
9846 Op = Op.getOperand(0);
9847 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9848 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9849 VT.getVectorElementType().getSizeInBits() ==
9850 OpVT.getVectorElementType().getSizeInBits()) {
9851 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9856 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9857 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9858 // (and (i32 x86isd::setcc_carry), 1)
9859 // This eliminates the zext. This transformation is necessary because
9860 // ISD::SETCC is always legalized to i8.
9861 DebugLoc dl = N->getDebugLoc();
9862 SDValue N0 = N->getOperand(0);
9863 EVT VT = N->getValueType(0);
9864 if (N0.getOpcode() == ISD::AND &&
9866 N0.getOperand(0).hasOneUse()) {
9867 SDValue N00 = N0.getOperand(0);
9868 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9870 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9871 if (!C || C->getZExtValue() != 1)
9873 return DAG.getNode(ISD::AND, dl, VT,
9874 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9875 N00.getOperand(0), N00.getOperand(1)),
9876 DAG.getConstant(1, VT));
9882 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9883 DAGCombinerInfo &DCI) const {
9884 SelectionDAG &DAG = DCI.DAG;
9885 switch (N->getOpcode()) {
9887 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9888 case ISD::EXTRACT_VECTOR_ELT:
9889 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
9890 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9891 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9892 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9895 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9896 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
9897 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9899 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9900 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9901 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9902 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9903 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9909 /// isTypeDesirableForOp - Return true if the target has native support for
9910 /// the specified value type and it is 'desirable' to use the type for the
9911 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
9912 /// instruction encodings are longer and some i16 instructions are slow.
9913 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
9914 if (!isTypeLegal(VT))
9923 case ISD::SIGN_EXTEND:
9924 case ISD::ZERO_EXTEND:
9925 case ISD::ANY_EXTEND:
9938 static bool MayFoldLoad(SDValue Op) {
9939 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
9942 static bool MayFoldIntoStore(SDValue Op) {
9943 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
9946 /// IsDesirableToPromoteOp - This method query the target whether it is
9947 /// beneficial for dag combiner to promote the specified node. If true, it
9948 /// should return the desired promotion type by reference.
9949 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
9950 EVT VT = Op.getValueType();
9954 bool Promote = false;
9955 bool Commute = false;
9956 switch (Op.getOpcode()) {
9959 LoadSDNode *LD = cast<LoadSDNode>(Op);
9960 // If the non-extending load has a single use and it's not live out, then it
9962 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
9964 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9965 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9966 // The only case where we'd want to promote LOAD (rather then it being
9967 // promoted as an operand is when it's only use is liveout.
9968 if (UI->getOpcode() != ISD::CopyToReg)
9975 case ISD::SIGN_EXTEND:
9976 case ISD::ZERO_EXTEND:
9977 case ISD::ANY_EXTEND:
9982 SDValue N0 = Op.getOperand(0);
9983 // Look out for (store (shl (load), x)).
9984 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
9997 SDValue N0 = Op.getOperand(0);
9998 SDValue N1 = Op.getOperand(1);
9999 if (!Commute && MayFoldLoad(N1))
10001 // Avoid disabling potential load folding opportunities.
10002 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10004 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10014 //===----------------------------------------------------------------------===//
10015 // X86 Inline Assembly Support
10016 //===----------------------------------------------------------------------===//
10018 static bool LowerToBSwap(CallInst *CI) {
10019 // FIXME: this should verify that we are targetting a 486 or better. If not,
10020 // we will turn this bswap into something that will be lowered to logical ops
10021 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10022 // so don't worry about this.
10024 // Verify this is a simple bswap.
10025 if (CI->getNumOperands() != 2 ||
10026 CI->getType() != CI->getArgOperand(0)->getType() ||
10027 !CI->getType()->isIntegerTy())
10030 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10031 if (!Ty || Ty->getBitWidth() % 16 != 0)
10034 // Okay, we can do this xform, do so now.
10035 const Type *Tys[] = { Ty };
10036 Module *M = CI->getParent()->getParent()->getParent();
10037 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10039 Value *Op = CI->getArgOperand(0);
10040 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10042 CI->replaceAllUsesWith(Op);
10043 CI->eraseFromParent();
10047 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10048 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10049 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10051 std::string AsmStr = IA->getAsmString();
10053 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10054 SmallVector<StringRef, 4> AsmPieces;
10055 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10057 switch (AsmPieces.size()) {
10058 default: return false;
10060 AsmStr = AsmPieces[0];
10062 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10065 if (AsmPieces.size() == 2 &&
10066 (AsmPieces[0] == "bswap" ||
10067 AsmPieces[0] == "bswapq" ||
10068 AsmPieces[0] == "bswapl") &&
10069 (AsmPieces[1] == "$0" ||
10070 AsmPieces[1] == "${0:q}")) {
10071 // No need to check constraints, nothing other than the equivalent of
10072 // "=r,0" would be valid here.
10073 return LowerToBSwap(CI);
10075 // rorw $$8, ${0:w} --> llvm.bswap.i16
10076 if (CI->getType()->isIntegerTy(16) &&
10077 AsmPieces.size() == 3 &&
10078 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10079 AsmPieces[1] == "$$8," &&
10080 AsmPieces[2] == "${0:w}" &&
10081 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10083 const std::string &Constraints = IA->getConstraintString();
10084 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10085 std::sort(AsmPieces.begin(), AsmPieces.end());
10086 if (AsmPieces.size() == 4 &&
10087 AsmPieces[0] == "~{cc}" &&
10088 AsmPieces[1] == "~{dirflag}" &&
10089 AsmPieces[2] == "~{flags}" &&
10090 AsmPieces[3] == "~{fpsr}") {
10091 return LowerToBSwap(CI);
10096 if (CI->getType()->isIntegerTy(64) &&
10097 Constraints.size() >= 2 &&
10098 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10099 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10100 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10101 SmallVector<StringRef, 4> Words;
10102 SplitString(AsmPieces[0], Words, " \t");
10103 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10105 SplitString(AsmPieces[1], Words, " \t");
10106 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10108 SplitString(AsmPieces[2], Words, " \t,");
10109 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10110 Words[2] == "%edx") {
10111 return LowerToBSwap(CI);
10123 /// getConstraintType - Given a constraint letter, return the type of
10124 /// constraint it is for this target.
10125 X86TargetLowering::ConstraintType
10126 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10127 if (Constraint.size() == 1) {
10128 switch (Constraint[0]) {
10140 return C_RegisterClass;
10148 return TargetLowering::getConstraintType(Constraint);
10151 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10152 /// with another that has more specific requirements based on the type of the
10153 /// corresponding operand.
10154 const char *X86TargetLowering::
10155 LowerXConstraint(EVT ConstraintVT) const {
10156 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10157 // 'f' like normal targets.
10158 if (ConstraintVT.isFloatingPoint()) {
10159 if (Subtarget->hasSSE2())
10161 if (Subtarget->hasSSE1())
10165 return TargetLowering::LowerXConstraint(ConstraintVT);
10168 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10169 /// vector. If it is invalid, don't add anything to Ops.
10170 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10172 std::vector<SDValue>&Ops,
10173 SelectionDAG &DAG) const {
10174 SDValue Result(0, 0);
10176 switch (Constraint) {
10179 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10180 if (C->getZExtValue() <= 31) {
10181 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10187 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10188 if (C->getZExtValue() <= 63) {
10189 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10195 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10196 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10197 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10204 if (C->getZExtValue() <= 255) {
10205 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10211 // 32-bit signed value
10212 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10213 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10214 C->getSExtValue())) {
10215 // Widen to 64 bits here to get it sign extended.
10216 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10219 // FIXME gcc accepts some relocatable values here too, but only in certain
10220 // memory models; it's complicated.
10225 // 32-bit unsigned value
10226 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10227 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10228 C->getZExtValue())) {
10229 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10233 // FIXME gcc accepts some relocatable values here too, but only in certain
10234 // memory models; it's complicated.
10238 // Literal immediates are always ok.
10239 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10240 // Widen to 64 bits here to get it sign extended.
10241 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10245 // In any sort of PIC mode addresses need to be computed at runtime by
10246 // adding in a register or some sort of table lookup. These can't
10247 // be used as immediates.
10248 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC() ||
10249 Subtarget->isPICStyleRIPRel())
10252 // If we are in non-pic codegen mode, we allow the address of a global (with
10253 // an optional displacement) to be used with 'i'.
10254 GlobalAddressSDNode *GA = 0;
10255 int64_t Offset = 0;
10257 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10259 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10260 Offset += GA->getOffset();
10262 } else if (Op.getOpcode() == ISD::ADD) {
10263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10264 Offset += C->getZExtValue();
10265 Op = Op.getOperand(0);
10268 } else if (Op.getOpcode() == ISD::SUB) {
10269 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10270 Offset += -C->getZExtValue();
10271 Op = Op.getOperand(0);
10276 // Otherwise, this isn't something we can handle, reject it.
10280 const GlobalValue *GV = GA->getGlobal();
10281 // If we require an extra load to get this address, as in PIC mode, we
10282 // can't accept it.
10283 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10284 getTargetMachine())))
10287 Result = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10292 if (Result.getNode()) {
10293 Ops.push_back(Result);
10296 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10299 std::vector<unsigned> X86TargetLowering::
10300 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10302 if (Constraint.size() == 1) {
10303 // FIXME: not handling fp-stack yet!
10304 switch (Constraint[0]) { // GCC X86 Constraint Letters
10305 default: break; // Unknown constraint letter
10306 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10307 if (Subtarget->is64Bit()) {
10308 if (VT == MVT::i32)
10309 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10310 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10311 X86::R10D,X86::R11D,X86::R12D,
10312 X86::R13D,X86::R14D,X86::R15D,
10313 X86::EBP, X86::ESP, 0);
10314 else if (VT == MVT::i16)
10315 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10316 X86::SI, X86::DI, X86::R8W,X86::R9W,
10317 X86::R10W,X86::R11W,X86::R12W,
10318 X86::R13W,X86::R14W,X86::R15W,
10319 X86::BP, X86::SP, 0);
10320 else if (VT == MVT::i8)
10321 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10322 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10323 X86::R10B,X86::R11B,X86::R12B,
10324 X86::R13B,X86::R14B,X86::R15B,
10325 X86::BPL, X86::SPL, 0);
10327 else if (VT == MVT::i64)
10328 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10329 X86::RSI, X86::RDI, X86::R8, X86::R9,
10330 X86::R10, X86::R11, X86::R12,
10331 X86::R13, X86::R14, X86::R15,
10332 X86::RBP, X86::RSP, 0);
10336 // 32-bit fallthrough
10337 case 'Q': // Q_REGS
10338 if (VT == MVT::i32)
10339 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10340 else if (VT == MVT::i16)
10341 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10342 else if (VT == MVT::i8)
10343 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10344 else if (VT == MVT::i64)
10345 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10350 return std::vector<unsigned>();
10353 std::pair<unsigned, const TargetRegisterClass*>
10354 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10356 // First, see if this is a constraint that directly corresponds to an LLVM
10358 if (Constraint.size() == 1) {
10359 // GCC Constraint Letters
10360 switch (Constraint[0]) {
10362 case 'r': // GENERAL_REGS
10363 case 'l': // INDEX_REGS
10365 return std::make_pair(0U, X86::GR8RegisterClass);
10366 if (VT == MVT::i16)
10367 return std::make_pair(0U, X86::GR16RegisterClass);
10368 if (VT == MVT::i32 || !Subtarget->is64Bit())
10369 return std::make_pair(0U, X86::GR32RegisterClass);
10370 return std::make_pair(0U, X86::GR64RegisterClass);
10371 case 'R': // LEGACY_REGS
10373 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10374 if (VT == MVT::i16)
10375 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10376 if (VT == MVT::i32 || !Subtarget->is64Bit())
10377 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10378 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10379 case 'f': // FP Stack registers.
10380 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10381 // value to the correct fpstack register class.
10382 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10383 return std::make_pair(0U, X86::RFP32RegisterClass);
10384 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10385 return std::make_pair(0U, X86::RFP64RegisterClass);
10386 return std::make_pair(0U, X86::RFP80RegisterClass);
10387 case 'y': // MMX_REGS if MMX allowed.
10388 if (!Subtarget->hasMMX()) break;
10389 return std::make_pair(0U, X86::VR64RegisterClass);
10390 case 'Y': // SSE_REGS if SSE2 allowed
10391 if (!Subtarget->hasSSE2()) break;
10393 case 'x': // SSE_REGS if SSE1 allowed
10394 if (!Subtarget->hasSSE1()) break;
10396 switch (VT.getSimpleVT().SimpleTy) {
10398 // Scalar SSE types.
10401 return std::make_pair(0U, X86::FR32RegisterClass);
10404 return std::make_pair(0U, X86::FR64RegisterClass);
10412 return std::make_pair(0U, X86::VR128RegisterClass);
10418 // Use the default implementation in TargetLowering to convert the register
10419 // constraint into a member of a register class.
10420 std::pair<unsigned, const TargetRegisterClass*> Res;
10421 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10423 // Not found as a standard register?
10424 if (Res.second == 0) {
10425 // Map st(0) -> st(7) -> ST0
10426 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10427 tolower(Constraint[1]) == 's' &&
10428 tolower(Constraint[2]) == 't' &&
10429 Constraint[3] == '(' &&
10430 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10431 Constraint[5] == ')' &&
10432 Constraint[6] == '}') {
10434 Res.first = X86::ST0+Constraint[4]-'0';
10435 Res.second = X86::RFP80RegisterClass;
10439 // GCC allows "st(0)" to be called just plain "st".
10440 if (StringRef("{st}").equals_lower(Constraint)) {
10441 Res.first = X86::ST0;
10442 Res.second = X86::RFP80RegisterClass;
10447 if (StringRef("{flags}").equals_lower(Constraint)) {
10448 Res.first = X86::EFLAGS;
10449 Res.second = X86::CCRRegisterClass;
10453 // 'A' means EAX + EDX.
10454 if (Constraint == "A") {
10455 Res.first = X86::EAX;
10456 Res.second = X86::GR32_ADRegisterClass;
10462 // Otherwise, check to see if this is a register class of the wrong value
10463 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10464 // turn into {ax},{dx}.
10465 if (Res.second->hasType(VT))
10466 return Res; // Correct type already, nothing to do.
10468 // All of the single-register GCC register classes map their values onto
10469 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10470 // really want an 8-bit or 32-bit register, map to the appropriate register
10471 // class and return the appropriate register.
10472 if (Res.second == X86::GR16RegisterClass) {
10473 if (VT == MVT::i8) {
10474 unsigned DestReg = 0;
10475 switch (Res.first) {
10477 case X86::AX: DestReg = X86::AL; break;
10478 case X86::DX: DestReg = X86::DL; break;
10479 case X86::CX: DestReg = X86::CL; break;
10480 case X86::BX: DestReg = X86::BL; break;
10483 Res.first = DestReg;
10484 Res.second = X86::GR8RegisterClass;
10486 } else if (VT == MVT::i32) {
10487 unsigned DestReg = 0;
10488 switch (Res.first) {
10490 case X86::AX: DestReg = X86::EAX; break;
10491 case X86::DX: DestReg = X86::EDX; break;
10492 case X86::CX: DestReg = X86::ECX; break;
10493 case X86::BX: DestReg = X86::EBX; break;
10494 case X86::SI: DestReg = X86::ESI; break;
10495 case X86::DI: DestReg = X86::EDI; break;
10496 case X86::BP: DestReg = X86::EBP; break;
10497 case X86::SP: DestReg = X86::ESP; break;
10500 Res.first = DestReg;
10501 Res.second = X86::GR32RegisterClass;
10503 } else if (VT == MVT::i64) {
10504 unsigned DestReg = 0;
10505 switch (Res.first) {
10507 case X86::AX: DestReg = X86::RAX; break;
10508 case X86::DX: DestReg = X86::RDX; break;
10509 case X86::CX: DestReg = X86::RCX; break;
10510 case X86::BX: DestReg = X86::RBX; break;
10511 case X86::SI: DestReg = X86::RSI; break;
10512 case X86::DI: DestReg = X86::RDI; break;
10513 case X86::BP: DestReg = X86::RBP; break;
10514 case X86::SP: DestReg = X86::RSP; break;
10517 Res.first = DestReg;
10518 Res.second = X86::GR64RegisterClass;
10521 } else if (Res.second == X86::FR32RegisterClass ||
10522 Res.second == X86::FR64RegisterClass ||
10523 Res.second == X86::VR128RegisterClass) {
10524 // Handle references to XMM physical registers that got mapped into the
10525 // wrong class. This can happen with constraints like {xmm0} where the
10526 // target independent register mapper will just pick the first match it can
10527 // find, ignoring the required type.
10528 if (VT == MVT::f32)
10529 Res.second = X86::FR32RegisterClass;
10530 else if (VT == MVT::f64)
10531 Res.second = X86::FR64RegisterClass;
10532 else if (X86::VR128RegisterClass->hasType(VT))
10533 Res.second = X86::VR128RegisterClass;