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 "X86ShuffleDecode.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCExpr.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/ADT/BitVector.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/Dwarf.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
54 using namespace dwarf;
56 STATISTIC(NumTailCalls, "Number of tail calls");
59 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
61 // Forward declarations.
62 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
65 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
67 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
69 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
70 if (is64Bit) return new X8664_MachoTargetObjectFile();
71 return new TargetLoweringObjectFileMachO();
72 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
73 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
74 return new X8632_ELFTargetObjectFile(TM);
75 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
76 return new TargetLoweringObjectFileCOFF();
78 llvm_unreachable("unknown subtarget type");
81 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
82 : TargetLowering(TM, createTLOF(TM)) {
83 Subtarget = &TM.getSubtarget<X86Subtarget>();
84 X86ScalarSSEf64 = Subtarget->hasSSE2();
85 X86ScalarSSEf32 = Subtarget->hasSSE1();
86 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
88 RegInfo = TM.getRegisterInfo();
91 // Set up the TargetLowering object.
93 // X86 is weird, it always uses i8 for shift amounts and setcc results.
94 setShiftAmountType(MVT::i8);
95 setBooleanContents(ZeroOrOneBooleanContent);
96 setSchedulingPreference(Sched::RegPressure);
97 setStackPointerRegisterToSaveRestore(X86StackPtr);
99 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygwin()) {
100 // Setup Windows compiler runtime calls.
101 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
102 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
105 if (Subtarget->isTargetDarwin()) {
106 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
107 setUseUnderscoreSetJmp(false);
108 setUseUnderscoreLongJmp(false);
109 } else if (Subtarget->isTargetMingw()) {
110 // MS runtime is weird: it exports _setjmp, but longjmp!
111 setUseUnderscoreSetJmp(true);
112 setUseUnderscoreLongJmp(false);
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(true);
118 // Set up the register classes.
119 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
120 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
121 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
122 if (Subtarget->is64Bit())
123 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
125 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
127 // We don't accept any truncstore of integer registers.
128 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
129 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
130 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
131 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
132 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
133 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
135 // SETOEQ and SETUNE require checking two conditions.
136 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
137 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
138 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
139 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
140 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
141 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
143 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
145 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
147 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
149 if (Subtarget->is64Bit()) {
150 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
151 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
152 } else if (!UseSoftFloat) {
153 // We have an algorithm for SSE2->double, and we turn this into a
154 // 64-bit FILD followed by conditional FADD for other targets.
155 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
156 // We have an algorithm for SSE2, and we turn this into a 64-bit
157 // FILD for other targets.
158 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
161 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
163 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
164 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
167 // SSE has no i16 to fp conversion, only i32
168 if (X86ScalarSSEf32) {
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
170 // f32 and f64 cases are Legal, f80 case is not
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
181 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
182 // are Legal, f80 is custom lowered.
183 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
186 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
188 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
191 if (X86ScalarSSEf32) {
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
193 // f32 and f64 cases are Legal, f80 case is not
194 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
197 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
200 // Handle FP_TO_UINT by promoting the destination to a larger signed
202 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
203 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
206 if (Subtarget->is64Bit()) {
207 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
209 } else if (!UseSoftFloat) {
210 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
211 // Expand FP_TO_UINT into a select.
212 // FIXME: We would like to use a Custom expander here eventually to do
213 // the optimal thing for SSE vs. the default expansion in the legalizer.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
216 // With SSE3 we can use fisttpll to convert to a signed i64; without
217 // SSE, we're stuck with a fistpll.
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
221 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
222 if (!X86ScalarSSEf64) {
223 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
225 if (Subtarget->is64Bit()) {
226 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
227 // Without SSE, i64->f64 goes through memory.
228 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
232 // Scalar integer divide and remainder are lowered to use operations that
233 // produce two results, to match the available instructions. This exposes
234 // the two-result form to trivial CSE, which is able to combine x/y and x%y
235 // into a single instruction.
237 // Scalar integer multiply-high is also lowered to use two-result
238 // operations, to match the available instructions. However, plain multiply
239 // (low) operations are left as Legal, as there are single-result
240 // instructions for this in x86. Using the two-result multiply instructions
241 // when both high and low results are needed must be arranged by dagcombine.
242 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
243 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
244 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
246 setOperationAction(ISD::SREM , MVT::i8 , Expand);
247 setOperationAction(ISD::UREM , MVT::i8 , Expand);
248 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
249 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
250 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
252 setOperationAction(ISD::SREM , MVT::i16 , Expand);
253 setOperationAction(ISD::UREM , MVT::i16 , Expand);
254 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
255 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
256 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
258 setOperationAction(ISD::SREM , MVT::i32 , Expand);
259 setOperationAction(ISD::UREM , MVT::i32 , Expand);
260 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
261 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
262 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
264 setOperationAction(ISD::SREM , MVT::i64 , Expand);
265 setOperationAction(ISD::UREM , MVT::i64 , Expand);
267 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
268 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
269 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
270 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
271 if (Subtarget->is64Bit())
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
275 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
276 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f32 , Expand);
278 setOperationAction(ISD::FREM , MVT::f64 , Expand);
279 setOperationAction(ISD::FREM , MVT::f80 , Expand);
280 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
282 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
283 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
285 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
286 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
287 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
288 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
291 if (Subtarget->is64Bit()) {
292 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
293 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
294 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
297 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
298 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
300 // These should be promoted to a larger select which is supported.
301 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
302 // X86 wants to expand cmov itself.
303 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
304 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
305 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
307 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
308 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
309 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
310 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
311 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
312 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
313 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
314 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
315 if (Subtarget->is64Bit()) {
316 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
317 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
319 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
322 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
323 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
324 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
325 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
326 if (Subtarget->is64Bit())
327 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
328 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
329 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
330 if (Subtarget->is64Bit()) {
331 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
332 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
333 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
334 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
335 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
337 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
338 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
341 if (Subtarget->is64Bit()) {
342 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
343 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
344 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
347 if (Subtarget->hasSSE1())
348 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
350 // We may not have a libcall for MEMBARRIER so we should lower this.
351 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
353 // On X86 and X86-64, atomic operations are lowered to locked instructions.
354 // Locked instructions, in turn, have implicit fence semantics (all memory
355 // operations are flushed before issuing the locked instruction, and they
356 // are not buffered), so we can fold away the common pattern of
357 // fence-atomic-fence.
358 setShouldFoldAtomicFences(true);
360 // Expand certain atomics
361 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
371 if (!Subtarget->is64Bit()) {
372 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
375 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
381 // FIXME - use subtarget debug flags
382 if (!Subtarget->isTargetDarwin() &&
383 !Subtarget->isTargetELF() &&
384 !Subtarget->isTargetCygMing()) {
385 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
388 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
389 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
390 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
391 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
392 if (Subtarget->is64Bit()) {
393 setExceptionPointerRegister(X86::RAX);
394 setExceptionSelectorRegister(X86::RDX);
396 setExceptionPointerRegister(X86::EAX);
397 setExceptionSelectorRegister(X86::EDX);
399 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
400 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
402 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
404 setOperationAction(ISD::TRAP, MVT::Other, Legal);
406 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
407 setOperationAction(ISD::VASTART , MVT::Other, Custom);
408 setOperationAction(ISD::VAEND , MVT::Other, Expand);
409 if (Subtarget->is64Bit()) {
410 setOperationAction(ISD::VAARG , MVT::Other, Custom);
411 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
413 setOperationAction(ISD::VAARG , MVT::Other, Expand);
414 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
417 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
418 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
419 if (Subtarget->is64Bit())
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
421 if (Subtarget->isTargetCygMing())
422 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
424 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
426 if (!UseSoftFloat && X86ScalarSSEf64) {
427 // f32 and f64 use SSE.
428 // Set up the FP register classes.
429 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
430 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
432 // Use ANDPD to simulate FABS.
433 setOperationAction(ISD::FABS , MVT::f64, Custom);
434 setOperationAction(ISD::FABS , MVT::f32, Custom);
436 // Use XORP to simulate FNEG.
437 setOperationAction(ISD::FNEG , MVT::f64, Custom);
438 setOperationAction(ISD::FNEG , MVT::f32, Custom);
440 // Use ANDPD and ORPD to simulate FCOPYSIGN.
441 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
442 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
444 // We don't support sin/cos/fmod
445 setOperationAction(ISD::FSIN , MVT::f64, Expand);
446 setOperationAction(ISD::FCOS , MVT::f64, Expand);
447 setOperationAction(ISD::FSIN , MVT::f32, Expand);
448 setOperationAction(ISD::FCOS , MVT::f32, Expand);
450 // Expand FP immediates into loads from the stack, except for the special
452 addLegalFPImmediate(APFloat(+0.0)); // xorpd
453 addLegalFPImmediate(APFloat(+0.0f)); // xorps
454 } else if (!UseSoftFloat && X86ScalarSSEf32) {
455 // Use SSE for f32, x87 for f64.
456 // Set up the FP register classes.
457 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
458 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
460 // Use ANDPS to simulate FABS.
461 setOperationAction(ISD::FABS , MVT::f32, Custom);
463 // Use XORP to simulate FNEG.
464 setOperationAction(ISD::FNEG , MVT::f32, Custom);
466 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
468 // Use ANDPS and ORPS to simulate FCOPYSIGN.
469 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
472 // We don't support sin/cos/fmod
473 setOperationAction(ISD::FSIN , MVT::f32, Expand);
474 setOperationAction(ISD::FCOS , MVT::f32, Expand);
476 // Special cases we handle for FP constants.
477 addLegalFPImmediate(APFloat(+0.0f)); // xorps
478 addLegalFPImmediate(APFloat(+0.0)); // FLD0
479 addLegalFPImmediate(APFloat(+1.0)); // FLD1
480 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
481 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
484 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
485 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
487 } else if (!UseSoftFloat) {
488 // f32 and f64 in x87.
489 // Set up the FP register classes.
490 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
491 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
493 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
494 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
495 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
496 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
499 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
500 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
502 addLegalFPImmediate(APFloat(+0.0)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
506 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
507 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
508 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
509 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
512 // Long double always uses X87.
514 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
515 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
516 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
519 APFloat TmpFlt(+0.0);
520 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
522 addLegalFPImmediate(TmpFlt); // FLD0
524 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
525 APFloat TmpFlt2(+1.0);
526 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
528 addLegalFPImmediate(TmpFlt2); // FLD1
529 TmpFlt2.changeSign();
530 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
534 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
535 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
539 // Always use a library call for pow.
540 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
541 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
542 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
544 setOperationAction(ISD::FLOG, MVT::f80, Expand);
545 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
546 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
547 setOperationAction(ISD::FEXP, MVT::f80, Expand);
548 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
550 // First set operation action for all vector types to either promote
551 // (for widening) or expand (for scalarization). Then we will selectively
552 // turn on ones that can be effectively codegen'd.
553 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
554 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
555 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
570 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
571 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
604 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
608 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
609 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
610 setTruncStoreAction((MVT::SimpleValueType)VT,
611 (MVT::SimpleValueType)InnerVT, Expand);
612 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
614 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
618 // with -msoft-float, disable use of MMX as well.
619 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
620 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass, false);
621 // No operations on x86mmx supported, everything uses intrinsics.
624 // MMX-sized vectors (other than x86mmx) are expected to be expanded
625 // into smaller operations.
626 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
627 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
628 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
629 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
630 setOperationAction(ISD::AND, MVT::v8i8, Expand);
631 setOperationAction(ISD::AND, MVT::v4i16, Expand);
632 setOperationAction(ISD::AND, MVT::v2i32, Expand);
633 setOperationAction(ISD::AND, MVT::v1i64, Expand);
634 setOperationAction(ISD::OR, MVT::v8i8, Expand);
635 setOperationAction(ISD::OR, MVT::v4i16, Expand);
636 setOperationAction(ISD::OR, MVT::v2i32, Expand);
637 setOperationAction(ISD::OR, MVT::v1i64, Expand);
638 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
639 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
640 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
641 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
642 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
643 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
644 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
645 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
646 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
647 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
648 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
649 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
650 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
651 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Expand);
652 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Expand);
653 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Expand);
654 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Expand);
656 if (!UseSoftFloat && Subtarget->hasSSE1()) {
657 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
659 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
660 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
661 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
662 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
663 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
664 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
665 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
666 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
668 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
669 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
670 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
673 if (!UseSoftFloat && Subtarget->hasSSE2()) {
674 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
676 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
677 // registers cannot be used even for integer operations.
678 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
679 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
680 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
681 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
683 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
684 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
685 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
686 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
687 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
688 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
689 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
690 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
691 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
692 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
693 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
694 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
695 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
696 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
697 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
698 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
703 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
705 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
706 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
707 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
708 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
709 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
711 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
712 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
713 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
714 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
715 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
717 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
718 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
719 EVT VT = (MVT::SimpleValueType)i;
720 // Do not attempt to custom lower non-power-of-2 vectors
721 if (!isPowerOf2_32(VT.getVectorNumElements()))
723 // Do not attempt to custom lower non-128-bit vectors
724 if (!VT.is128BitVector())
726 setOperationAction(ISD::BUILD_VECTOR,
727 VT.getSimpleVT().SimpleTy, Custom);
728 setOperationAction(ISD::VECTOR_SHUFFLE,
729 VT.getSimpleVT().SimpleTy, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
731 VT.getSimpleVT().SimpleTy, Custom);
734 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
735 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
736 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
737 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
739 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
741 if (Subtarget->is64Bit()) {
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
746 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
747 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
748 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
751 // Do not attempt to promote non-128-bit vectors
752 if (!VT.is128BitVector())
755 setOperationAction(ISD::AND, SVT, Promote);
756 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
757 setOperationAction(ISD::OR, SVT, Promote);
758 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
759 setOperationAction(ISD::XOR, SVT, Promote);
760 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
761 setOperationAction(ISD::LOAD, SVT, Promote);
762 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
763 setOperationAction(ISD::SELECT, SVT, Promote);
764 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
767 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
769 // Custom lower v2i64 and v2f64 selects.
770 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
771 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
772 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
773 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
775 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
776 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
779 if (Subtarget->hasSSE41()) {
780 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
781 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
782 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
783 setOperationAction(ISD::FRINT, MVT::f32, Legal);
784 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
785 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
786 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
787 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
788 setOperationAction(ISD::FRINT, MVT::f64, Legal);
789 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
791 // FIXME: Do we need to handle scalar-to-vector here?
792 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
794 // Can turn SHL into an integer multiply.
795 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
796 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
798 // i8 and i16 vectors are custom , because the source register and source
799 // source memory operand types are not the same width. f32 vectors are
800 // custom since the immediate controlling the insert encodes additional
802 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
803 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
804 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
805 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
807 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
808 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
809 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
810 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
812 if (Subtarget->is64Bit()) {
813 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
814 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
818 if (Subtarget->hasSSE42()) {
819 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
822 if (!UseSoftFloat && Subtarget->hasAVX()) {
823 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
824 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
825 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
826 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
827 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
829 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
830 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
831 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
832 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
833 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
834 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
835 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
836 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
837 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
838 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
839 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
840 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
841 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
842 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
843 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
845 // Operations to consider commented out -v16i16 v32i8
846 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
847 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
848 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
849 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
850 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
851 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
852 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
853 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
854 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
855 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
856 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
857 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
858 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
859 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
861 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
862 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
863 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
864 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
866 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
867 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
868 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
869 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
870 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
872 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
873 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
874 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
875 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
876 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
877 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
880 // Not sure we want to do this since there are no 256-bit integer
883 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
884 // This includes 256-bit vectors
885 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
886 EVT VT = (MVT::SimpleValueType)i;
888 // Do not attempt to custom lower non-power-of-2 vectors
889 if (!isPowerOf2_32(VT.getVectorNumElements()))
892 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
893 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
894 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
897 if (Subtarget->is64Bit()) {
898 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
899 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
904 // Not sure we want to do this since there are no 256-bit integer
907 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
908 // Including 256-bit vectors
909 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
910 EVT VT = (MVT::SimpleValueType)i;
912 if (!VT.is256BitVector()) {
915 setOperationAction(ISD::AND, VT, Promote);
916 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
917 setOperationAction(ISD::OR, VT, Promote);
918 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
919 setOperationAction(ISD::XOR, VT, Promote);
920 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
921 setOperationAction(ISD::LOAD, VT, Promote);
922 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
923 setOperationAction(ISD::SELECT, VT, Promote);
924 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
927 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
931 // We want to custom lower some of our intrinsics.
932 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
934 // Add/Sub/Mul with overflow operations are custom lowered.
935 setOperationAction(ISD::SADDO, MVT::i32, Custom);
936 setOperationAction(ISD::UADDO, MVT::i32, Custom);
937 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
938 setOperationAction(ISD::USUBO, MVT::i32, Custom);
939 setOperationAction(ISD::SMULO, MVT::i32, Custom);
941 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
942 // handle type legalization for these operations here.
944 // FIXME: We really should do custom legalization for addition and
945 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
946 // than generic legalization for 64-bit multiplication-with-overflow, though.
947 if (Subtarget->is64Bit()) {
948 setOperationAction(ISD::SADDO, MVT::i64, Custom);
949 setOperationAction(ISD::UADDO, MVT::i64, Custom);
950 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
951 setOperationAction(ISD::USUBO, MVT::i64, Custom);
952 setOperationAction(ISD::SMULO, MVT::i64, Custom);
955 if (!Subtarget->is64Bit()) {
956 // These libcalls are not available in 32-bit.
957 setLibcallName(RTLIB::SHL_I128, 0);
958 setLibcallName(RTLIB::SRL_I128, 0);
959 setLibcallName(RTLIB::SRA_I128, 0);
962 // We have target-specific dag combine patterns for the following nodes:
963 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
964 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
965 setTargetDAGCombine(ISD::BUILD_VECTOR);
966 setTargetDAGCombine(ISD::SELECT);
967 setTargetDAGCombine(ISD::SHL);
968 setTargetDAGCombine(ISD::SRA);
969 setTargetDAGCombine(ISD::SRL);
970 setTargetDAGCombine(ISD::OR);
971 setTargetDAGCombine(ISD::STORE);
972 setTargetDAGCombine(ISD::ZERO_EXTEND);
973 if (Subtarget->is64Bit())
974 setTargetDAGCombine(ISD::MUL);
976 computeRegisterProperties();
978 // FIXME: These should be based on subtarget info. Plus, the values should
979 // be smaller when we are in optimizing for size mode.
980 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
981 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
982 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
983 setPrefLoopAlignment(16);
984 benefitFromCodePlacementOpt = true;
988 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
993 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
994 /// the desired ByVal argument alignment.
995 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
998 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
999 if (VTy->getBitWidth() == 128)
1001 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1002 unsigned EltAlign = 0;
1003 getMaxByValAlign(ATy->getElementType(), EltAlign);
1004 if (EltAlign > MaxAlign)
1005 MaxAlign = EltAlign;
1006 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1007 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1008 unsigned EltAlign = 0;
1009 getMaxByValAlign(STy->getElementType(i), EltAlign);
1010 if (EltAlign > MaxAlign)
1011 MaxAlign = EltAlign;
1019 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1020 /// function arguments in the caller parameter area. For X86, aggregates
1021 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1022 /// are at 4-byte boundaries.
1023 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1024 if (Subtarget->is64Bit()) {
1025 // Max of 8 and alignment of type.
1026 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1033 if (Subtarget->hasSSE1())
1034 getMaxByValAlign(Ty, Align);
1038 /// getOptimalMemOpType - Returns the target specific optimal type for load
1039 /// and store operations as a result of memset, memcpy, and memmove
1040 /// lowering. If DstAlign is zero that means it's safe to destination
1041 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1042 /// means there isn't a need to check it against alignment requirement,
1043 /// probably because the source does not need to be loaded. If
1044 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1045 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1046 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1047 /// constant so it does not need to be loaded.
1048 /// It returns EVT::Other if the type should be determined using generic
1049 /// target-independent logic.
1051 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1052 unsigned DstAlign, unsigned SrcAlign,
1053 bool NonScalarIntSafe,
1055 MachineFunction &MF) const {
1056 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1057 // linux. This is because the stack realignment code can't handle certain
1058 // cases like PR2962. This should be removed when PR2962 is fixed.
1059 const Function *F = MF.getFunction();
1060 if (NonScalarIntSafe &&
1061 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1063 (Subtarget->isUnalignedMemAccessFast() ||
1064 ((DstAlign == 0 || DstAlign >= 16) &&
1065 (SrcAlign == 0 || SrcAlign >= 16))) &&
1066 Subtarget->getStackAlignment() >= 16) {
1067 if (Subtarget->hasSSE2())
1069 if (Subtarget->hasSSE1())
1071 } else if (!MemcpyStrSrc && Size >= 8 &&
1072 !Subtarget->is64Bit() &&
1073 Subtarget->getStackAlignment() >= 8 &&
1074 Subtarget->hasSSE2()) {
1075 // Do not use f64 to lower memcpy if source is string constant. It's
1076 // better to use i32 to avoid the loads.
1080 if (Subtarget->is64Bit() && Size >= 8)
1085 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1086 /// current function. The returned value is a member of the
1087 /// MachineJumpTableInfo::JTEntryKind enum.
1088 unsigned X86TargetLowering::getJumpTableEncoding() const {
1089 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1091 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1092 Subtarget->isPICStyleGOT())
1093 return MachineJumpTableInfo::EK_Custom32;
1095 // Otherwise, use the normal jump table encoding heuristics.
1096 return TargetLowering::getJumpTableEncoding();
1099 /// getPICBaseSymbol - Return the X86-32 PIC base.
1101 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1102 MCContext &Ctx) const {
1103 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1104 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1105 Twine(MF->getFunctionNumber())+"$pb");
1110 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1111 const MachineBasicBlock *MBB,
1112 unsigned uid,MCContext &Ctx) const{
1113 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1114 Subtarget->isPICStyleGOT());
1115 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1117 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1118 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1121 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1123 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1124 SelectionDAG &DAG) const {
1125 if (!Subtarget->is64Bit())
1126 // This doesn't have DebugLoc associated with it, but is not really the
1127 // same as a Register.
1128 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1132 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1133 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1135 const MCExpr *X86TargetLowering::
1136 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1137 MCContext &Ctx) const {
1138 // X86-64 uses RIP relative addressing based on the jump table label.
1139 if (Subtarget->isPICStyleRIPRel())
1140 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1142 // Otherwise, the reference is relative to the PIC base.
1143 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1146 /// getFunctionAlignment - Return the Log2 alignment of this function.
1147 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1148 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1151 std::pair<const TargetRegisterClass*, uint8_t>
1152 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1153 const TargetRegisterClass *RRC = 0;
1155 switch (VT.getSimpleVT().SimpleTy) {
1157 return TargetLowering::findRepresentativeClass(VT);
1158 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1159 RRC = (Subtarget->is64Bit()
1160 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1163 RRC = X86::VR64RegisterClass;
1165 case MVT::f32: case MVT::f64:
1166 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1167 case MVT::v4f32: case MVT::v2f64:
1168 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1170 RRC = X86::VR128RegisterClass;
1173 return std::make_pair(RRC, Cost);
1177 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1178 MachineFunction &MF) const {
1179 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1180 switch (RC->getID()) {
1183 case X86::GR32RegClassID:
1185 case X86::GR64RegClassID:
1187 case X86::VR128RegClassID:
1188 return Subtarget->is64Bit() ? 10 : 4;
1189 case X86::VR64RegClassID:
1194 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1195 unsigned &Offset) const {
1196 if (!Subtarget->isTargetLinux())
1199 if (Subtarget->is64Bit()) {
1200 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1202 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1215 //===----------------------------------------------------------------------===//
1216 // Return Value Calling Convention Implementation
1217 //===----------------------------------------------------------------------===//
1219 #include "X86GenCallingConv.inc"
1222 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1223 const SmallVectorImpl<ISD::OutputArg> &Outs,
1224 LLVMContext &Context) const {
1225 SmallVector<CCValAssign, 16> RVLocs;
1226 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1228 return CCInfo.CheckReturn(Outs, RetCC_X86);
1232 X86TargetLowering::LowerReturn(SDValue Chain,
1233 CallingConv::ID CallConv, bool isVarArg,
1234 const SmallVectorImpl<ISD::OutputArg> &Outs,
1235 const SmallVectorImpl<SDValue> &OutVals,
1236 DebugLoc dl, SelectionDAG &DAG) const {
1237 MachineFunction &MF = DAG.getMachineFunction();
1238 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1240 SmallVector<CCValAssign, 16> RVLocs;
1241 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1242 RVLocs, *DAG.getContext());
1243 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1245 // Add the regs to the liveout set for the function.
1246 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1247 for (unsigned i = 0; i != RVLocs.size(); ++i)
1248 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1249 MRI.addLiveOut(RVLocs[i].getLocReg());
1253 SmallVector<SDValue, 6> RetOps;
1254 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1255 // Operand #1 = Bytes To Pop
1256 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1259 // Copy the result values into the output registers.
1260 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1261 CCValAssign &VA = RVLocs[i];
1262 assert(VA.isRegLoc() && "Can only return in registers!");
1263 SDValue ValToCopy = OutVals[i];
1264 EVT ValVT = ValToCopy.getValueType();
1266 // If this is x86-64, and we disabled SSE, we can't return FP values,
1267 // or SSE or MMX vectors.
1268 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1269 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1270 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1271 report_fatal_error("SSE register return with SSE disabled");
1273 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1274 // llvm-gcc has never done it right and no one has noticed, so this
1275 // should be OK for now.
1276 if (ValVT == MVT::f64 &&
1277 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1278 report_fatal_error("SSE2 register return with SSE2 disabled");
1280 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1281 // the RET instruction and handled by the FP Stackifier.
1282 if (VA.getLocReg() == X86::ST0 ||
1283 VA.getLocReg() == X86::ST1) {
1284 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1285 // change the value to the FP stack register class.
1286 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1287 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1288 RetOps.push_back(ValToCopy);
1289 // Don't emit a copytoreg.
1293 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1294 // which is returned in RAX / RDX.
1295 if (Subtarget->is64Bit()) {
1296 if (ValVT == MVT::x86mmx) {
1297 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1298 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1299 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1301 // If we don't have SSE2 available, convert to v4f32 so the generated
1302 // register is legal.
1303 if (!Subtarget->hasSSE2())
1304 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1309 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1310 Flag = Chain.getValue(1);
1313 // The x86-64 ABI for returning structs by value requires that we copy
1314 // the sret argument into %rax for the return. We saved the argument into
1315 // a virtual register in the entry block, so now we copy the value out
1317 if (Subtarget->is64Bit() &&
1318 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1319 MachineFunction &MF = DAG.getMachineFunction();
1320 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1321 unsigned Reg = FuncInfo->getSRetReturnReg();
1323 "SRetReturnReg should have been set in LowerFormalArguments().");
1324 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1326 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1327 Flag = Chain.getValue(1);
1329 // RAX now acts like a return value.
1330 MRI.addLiveOut(X86::RAX);
1333 RetOps[0] = Chain; // Update chain.
1335 // Add the flag if we have it.
1337 RetOps.push_back(Flag);
1339 return DAG.getNode(X86ISD::RET_FLAG, dl,
1340 MVT::Other, &RetOps[0], RetOps.size());
1343 /// LowerCallResult - Lower the result values of a call into the
1344 /// appropriate copies out of appropriate physical registers.
1347 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1348 CallingConv::ID CallConv, bool isVarArg,
1349 const SmallVectorImpl<ISD::InputArg> &Ins,
1350 DebugLoc dl, SelectionDAG &DAG,
1351 SmallVectorImpl<SDValue> &InVals) const {
1353 // Assign locations to each value returned by this call.
1354 SmallVector<CCValAssign, 16> RVLocs;
1355 bool Is64Bit = Subtarget->is64Bit();
1356 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1357 RVLocs, *DAG.getContext());
1358 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1360 // Copy all of the result registers out of their specified physreg.
1361 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1362 CCValAssign &VA = RVLocs[i];
1363 EVT CopyVT = VA.getValVT();
1365 // If this is x86-64, and we disabled SSE, we can't return FP values
1366 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1367 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1368 report_fatal_error("SSE register return with SSE disabled");
1373 // If this is a call to a function that returns an fp value on the floating
1374 // point stack, we must guarantee the the value is popped from the stack, so
1375 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1376 // if the return value is not used. We use the FpGET_ST0 instructions
1378 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1379 // If we prefer to use the value in xmm registers, copy it out as f80 and
1380 // use a truncate to move it from fp stack reg to xmm reg.
1381 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1382 bool isST0 = VA.getLocReg() == X86::ST0;
1384 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1385 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1386 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1387 SDValue Ops[] = { Chain, InFlag };
1388 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1390 Val = Chain.getValue(0);
1392 // Round the f80 to the right size, which also moves it to the appropriate
1394 if (CopyVT != VA.getValVT())
1395 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1396 // This truncation won't change the value.
1397 DAG.getIntPtrConstant(1));
1398 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1399 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1400 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1401 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1402 MVT::v2i64, InFlag).getValue(1);
1403 Val = Chain.getValue(0);
1404 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1405 Val, DAG.getConstant(0, MVT::i64));
1407 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1408 MVT::i64, InFlag).getValue(1);
1409 Val = Chain.getValue(0);
1411 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1413 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1414 CopyVT, InFlag).getValue(1);
1415 Val = Chain.getValue(0);
1417 InFlag = Chain.getValue(2);
1418 InVals.push_back(Val);
1425 //===----------------------------------------------------------------------===//
1426 // C & StdCall & Fast Calling Convention implementation
1427 //===----------------------------------------------------------------------===//
1428 // StdCall calling convention seems to be standard for many Windows' API
1429 // routines and around. It differs from C calling convention just a little:
1430 // callee should clean up the stack, not caller. Symbols should be also
1431 // decorated in some fancy way :) It doesn't support any vector arguments.
1432 // For info on fast calling convention see Fast Calling Convention (tail call)
1433 // implementation LowerX86_32FastCCCallTo.
1435 /// CallIsStructReturn - Determines whether a call uses struct return
1437 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1441 return Outs[0].Flags.isSRet();
1444 /// ArgsAreStructReturn - Determines whether a function uses struct
1445 /// return semantics.
1447 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1451 return Ins[0].Flags.isSRet();
1454 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1455 /// given CallingConvention value.
1456 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1457 if (Subtarget->is64Bit()) {
1458 if (CC == CallingConv::GHC)
1459 return CC_X86_64_GHC;
1460 else if (Subtarget->isTargetWin64())
1461 return CC_X86_Win64_C;
1466 if (CC == CallingConv::X86_FastCall)
1467 return CC_X86_32_FastCall;
1468 else if (CC == CallingConv::X86_ThisCall)
1469 return CC_X86_32_ThisCall;
1470 else if (CC == CallingConv::Fast)
1471 return CC_X86_32_FastCC;
1472 else if (CC == CallingConv::GHC)
1473 return CC_X86_32_GHC;
1478 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1479 /// by "Src" to address "Dst" with size and alignment information specified by
1480 /// the specific parameter attribute. The copy will be passed as a byval
1481 /// function parameter.
1483 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1484 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1486 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1488 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1489 /*isVolatile*/false, /*AlwaysInline=*/true,
1490 MachinePointerInfo(), MachinePointerInfo());
1493 /// IsTailCallConvention - Return true if the calling convention is one that
1494 /// supports tail call optimization.
1495 static bool IsTailCallConvention(CallingConv::ID CC) {
1496 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1499 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1500 /// a tailcall target by changing its ABI.
1501 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1502 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1506 X86TargetLowering::LowerMemArgument(SDValue Chain,
1507 CallingConv::ID CallConv,
1508 const SmallVectorImpl<ISD::InputArg> &Ins,
1509 DebugLoc dl, SelectionDAG &DAG,
1510 const CCValAssign &VA,
1511 MachineFrameInfo *MFI,
1513 // Create the nodes corresponding to a load from this parameter slot.
1514 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1515 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1516 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1519 // If value is passed by pointer we have address passed instead of the value
1521 if (VA.getLocInfo() == CCValAssign::Indirect)
1522 ValVT = VA.getLocVT();
1524 ValVT = VA.getValVT();
1526 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1527 // changed with more analysis.
1528 // In case of tail call optimization mark all arguments mutable. Since they
1529 // could be overwritten by lowering of arguments in case of a tail call.
1530 if (Flags.isByVal()) {
1531 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1532 VA.getLocMemOffset(), isImmutable);
1533 return DAG.getFrameIndex(FI, getPointerTy());
1535 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1536 VA.getLocMemOffset(), isImmutable);
1537 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1538 return DAG.getLoad(ValVT, dl, Chain, FIN,
1539 MachinePointerInfo::getFixedStack(FI),
1545 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1546 CallingConv::ID CallConv,
1548 const SmallVectorImpl<ISD::InputArg> &Ins,
1551 SmallVectorImpl<SDValue> &InVals)
1553 MachineFunction &MF = DAG.getMachineFunction();
1554 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1556 const Function* Fn = MF.getFunction();
1557 if (Fn->hasExternalLinkage() &&
1558 Subtarget->isTargetCygMing() &&
1559 Fn->getName() == "main")
1560 FuncInfo->setForceFramePointer(true);
1562 MachineFrameInfo *MFI = MF.getFrameInfo();
1563 bool Is64Bit = Subtarget->is64Bit();
1564 bool IsWin64 = Subtarget->isTargetWin64();
1566 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1567 "Var args not supported with calling convention fastcc or ghc");
1569 // Assign locations to all of the incoming arguments.
1570 SmallVector<CCValAssign, 16> ArgLocs;
1571 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1572 ArgLocs, *DAG.getContext());
1573 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1575 unsigned LastVal = ~0U;
1577 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1578 CCValAssign &VA = ArgLocs[i];
1579 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1581 assert(VA.getValNo() != LastVal &&
1582 "Don't support value assigned to multiple locs yet");
1583 LastVal = VA.getValNo();
1585 if (VA.isRegLoc()) {
1586 EVT RegVT = VA.getLocVT();
1587 TargetRegisterClass *RC = NULL;
1588 if (RegVT == MVT::i32)
1589 RC = X86::GR32RegisterClass;
1590 else if (Is64Bit && RegVT == MVT::i64)
1591 RC = X86::GR64RegisterClass;
1592 else if (RegVT == MVT::f32)
1593 RC = X86::FR32RegisterClass;
1594 else if (RegVT == MVT::f64)
1595 RC = X86::FR64RegisterClass;
1596 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1597 RC = X86::VR256RegisterClass;
1598 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1599 RC = X86::VR128RegisterClass;
1600 else if (RegVT == MVT::x86mmx)
1601 RC = X86::VR64RegisterClass;
1603 llvm_unreachable("Unknown argument type!");
1605 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1606 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1608 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1609 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1611 if (VA.getLocInfo() == CCValAssign::SExt)
1612 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1613 DAG.getValueType(VA.getValVT()));
1614 else if (VA.getLocInfo() == CCValAssign::ZExt)
1615 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1616 DAG.getValueType(VA.getValVT()));
1617 else if (VA.getLocInfo() == CCValAssign::BCvt)
1618 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1620 if (VA.isExtInLoc()) {
1621 // Handle MMX values passed in XMM regs.
1622 if (RegVT.isVector()) {
1623 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1626 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1629 assert(VA.isMemLoc());
1630 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1633 // If value is passed via pointer - do a load.
1634 if (VA.getLocInfo() == CCValAssign::Indirect)
1635 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1636 MachinePointerInfo(), false, false, 0);
1638 InVals.push_back(ArgValue);
1641 // The x86-64 ABI for returning structs by value requires that we copy
1642 // the sret argument into %rax for the return. Save the argument into
1643 // a virtual register so that we can access it from the return points.
1644 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1645 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1646 unsigned Reg = FuncInfo->getSRetReturnReg();
1648 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1649 FuncInfo->setSRetReturnReg(Reg);
1651 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1652 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1655 unsigned StackSize = CCInfo.getNextStackOffset();
1656 // Align stack specially for tail calls.
1657 if (FuncIsMadeTailCallSafe(CallConv))
1658 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1660 // If the function takes variable number of arguments, make a frame index for
1661 // the start of the first vararg value... for expansion of llvm.va_start.
1663 if (!IsWin64 && (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1664 CallConv != CallingConv::X86_ThisCall))) {
1665 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1668 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1670 // FIXME: We should really autogenerate these arrays
1671 static const unsigned GPR64ArgRegsWin64[] = {
1672 X86::RCX, X86::RDX, X86::R8, X86::R9
1674 static const unsigned GPR64ArgRegs64Bit[] = {
1675 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1677 static const unsigned XMMArgRegs64Bit[] = {
1678 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1679 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1681 const unsigned *GPR64ArgRegs;
1682 unsigned NumXMMRegs = 0;
1685 // The XMM registers which might contain var arg parameters are shadowed
1686 // in their paired GPR. So we only need to save the GPR to their home
1688 TotalNumIntRegs = 4;
1689 GPR64ArgRegs = GPR64ArgRegsWin64;
1691 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1692 GPR64ArgRegs = GPR64ArgRegs64Bit;
1694 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1696 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1699 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1700 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1701 "SSE register cannot be used when SSE is disabled!");
1702 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1703 "SSE register cannot be used when SSE is disabled!");
1704 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1705 // Kernel mode asks for SSE to be disabled, so don't push them
1707 TotalNumXMMRegs = 0;
1710 const TargetFrameInfo &TFI = *getTargetMachine().getFrameInfo();
1711 // Get to the caller-allocated home save location. Add 8 to account
1712 // for the return address.
1713 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1714 FuncInfo->setRegSaveFrameIndex(
1715 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1716 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1718 // For X86-64, if there are vararg parameters that are passed via
1719 // registers, then we must store them to their spots on the stack so they
1720 // may be loaded by deferencing the result of va_next.
1721 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1722 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1723 FuncInfo->setRegSaveFrameIndex(
1724 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1728 // Store the integer parameter registers.
1729 SmallVector<SDValue, 8> MemOps;
1730 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1732 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1733 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1734 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1735 DAG.getIntPtrConstant(Offset));
1736 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1737 X86::GR64RegisterClass);
1738 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1740 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1741 MachinePointerInfo::getFixedStack(
1742 FuncInfo->getRegSaveFrameIndex(), Offset),
1744 MemOps.push_back(Store);
1748 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1749 // Now store the XMM (fp + vector) parameter registers.
1750 SmallVector<SDValue, 11> SaveXMMOps;
1751 SaveXMMOps.push_back(Chain);
1753 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1754 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1755 SaveXMMOps.push_back(ALVal);
1757 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1758 FuncInfo->getRegSaveFrameIndex()));
1759 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1760 FuncInfo->getVarArgsFPOffset()));
1762 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1763 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1764 X86::VR128RegisterClass);
1765 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1766 SaveXMMOps.push_back(Val);
1768 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1770 &SaveXMMOps[0], SaveXMMOps.size()));
1773 if (!MemOps.empty())
1774 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1775 &MemOps[0], MemOps.size());
1779 // Some CCs need callee pop.
1780 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1781 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1783 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1784 // If this is an sret function, the return should pop the hidden pointer.
1785 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1786 FuncInfo->setBytesToPopOnReturn(4);
1790 // RegSaveFrameIndex is X86-64 only.
1791 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1792 if (CallConv == CallingConv::X86_FastCall ||
1793 CallConv == CallingConv::X86_ThisCall)
1794 // fastcc functions can't have varargs.
1795 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1802 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1803 SDValue StackPtr, SDValue Arg,
1804 DebugLoc dl, SelectionDAG &DAG,
1805 const CCValAssign &VA,
1806 ISD::ArgFlagsTy Flags) const {
1807 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1808 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1809 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1810 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1811 if (Flags.isByVal())
1812 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1814 return DAG.getStore(Chain, dl, Arg, PtrOff,
1815 MachinePointerInfo::getStack(LocMemOffset),
1819 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1820 /// optimization is performed and it is required.
1822 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1823 SDValue &OutRetAddr, SDValue Chain,
1824 bool IsTailCall, bool Is64Bit,
1825 int FPDiff, DebugLoc dl) const {
1826 // Adjust the Return address stack slot.
1827 EVT VT = getPointerTy();
1828 OutRetAddr = getReturnAddressFrameIndex(DAG);
1830 // Load the "old" Return address.
1831 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1833 return SDValue(OutRetAddr.getNode(), 1);
1836 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1837 /// optimization is performed and it is required (FPDiff!=0).
1839 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1840 SDValue Chain, SDValue RetAddrFrIdx,
1841 bool Is64Bit, int FPDiff, DebugLoc dl) {
1842 // Store the return address to the appropriate stack slot.
1843 if (!FPDiff) return Chain;
1844 // Calculate the new stack slot for the return address.
1845 int SlotSize = Is64Bit ? 8 : 4;
1846 int NewReturnAddrFI =
1847 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1848 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1849 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1850 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1851 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1857 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1858 CallingConv::ID CallConv, bool isVarArg,
1860 const SmallVectorImpl<ISD::OutputArg> &Outs,
1861 const SmallVectorImpl<SDValue> &OutVals,
1862 const SmallVectorImpl<ISD::InputArg> &Ins,
1863 DebugLoc dl, SelectionDAG &DAG,
1864 SmallVectorImpl<SDValue> &InVals) const {
1865 MachineFunction &MF = DAG.getMachineFunction();
1866 bool Is64Bit = Subtarget->is64Bit();
1867 bool IsStructRet = CallIsStructReturn(Outs);
1868 bool IsSibcall = false;
1871 // Check if it's really possible to do a tail call.
1872 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1873 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1874 Outs, OutVals, Ins, DAG);
1876 // Sibcalls are automatically detected tailcalls which do not require
1878 if (!GuaranteedTailCallOpt && isTailCall)
1885 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1886 "Var args not supported with calling convention fastcc or ghc");
1888 // Analyze operands of the call, assigning locations to each operand.
1889 SmallVector<CCValAssign, 16> ArgLocs;
1890 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1891 ArgLocs, *DAG.getContext());
1892 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1894 // Get a count of how many bytes are to be pushed on the stack.
1895 unsigned NumBytes = CCInfo.getNextStackOffset();
1897 // This is a sibcall. The memory operands are available in caller's
1898 // own caller's stack.
1900 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1901 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1904 if (isTailCall && !IsSibcall) {
1905 // Lower arguments at fp - stackoffset + fpdiff.
1906 unsigned NumBytesCallerPushed =
1907 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1908 FPDiff = NumBytesCallerPushed - NumBytes;
1910 // Set the delta of movement of the returnaddr stackslot.
1911 // But only set if delta is greater than previous delta.
1912 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1913 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1917 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1919 SDValue RetAddrFrIdx;
1920 // Load return adress for tail calls.
1921 if (isTailCall && FPDiff)
1922 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1923 Is64Bit, FPDiff, dl);
1925 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1926 SmallVector<SDValue, 8> MemOpChains;
1929 // Walk the register/memloc assignments, inserting copies/loads. In the case
1930 // of tail call optimization arguments are handle later.
1931 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1932 CCValAssign &VA = ArgLocs[i];
1933 EVT RegVT = VA.getLocVT();
1934 SDValue Arg = OutVals[i];
1935 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1936 bool isByVal = Flags.isByVal();
1938 // Promote the value if needed.
1939 switch (VA.getLocInfo()) {
1940 default: llvm_unreachable("Unknown loc info!");
1941 case CCValAssign::Full: break;
1942 case CCValAssign::SExt:
1943 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1945 case CCValAssign::ZExt:
1946 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1948 case CCValAssign::AExt:
1949 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1950 // Special case: passing MMX values in XMM registers.
1951 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1952 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1953 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1955 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1957 case CCValAssign::BCvt:
1958 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1960 case CCValAssign::Indirect: {
1961 // Store the argument.
1962 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1963 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1964 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1965 MachinePointerInfo::getFixedStack(FI),
1972 if (VA.isRegLoc()) {
1973 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1974 if (isVarArg && Subtarget->isTargetWin64()) {
1975 // Win64 ABI requires argument XMM reg to be copied to the corresponding
1976 // shadow reg if callee is a varargs function.
1977 unsigned ShadowReg = 0;
1978 switch (VA.getLocReg()) {
1979 case X86::XMM0: ShadowReg = X86::RCX; break;
1980 case X86::XMM1: ShadowReg = X86::RDX; break;
1981 case X86::XMM2: ShadowReg = X86::R8; break;
1982 case X86::XMM3: ShadowReg = X86::R9; break;
1985 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
1987 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1988 assert(VA.isMemLoc());
1989 if (StackPtr.getNode() == 0)
1990 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1991 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1992 dl, DAG, VA, Flags));
1996 if (!MemOpChains.empty())
1997 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1998 &MemOpChains[0], MemOpChains.size());
2000 // Build a sequence of copy-to-reg nodes chained together with token chain
2001 // and flag operands which copy the outgoing args into registers.
2003 // Tail call byval lowering might overwrite argument registers so in case of
2004 // tail call optimization the copies to registers are lowered later.
2006 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2007 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2008 RegsToPass[i].second, InFlag);
2009 InFlag = Chain.getValue(1);
2012 if (Subtarget->isPICStyleGOT()) {
2013 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2016 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2017 DAG.getNode(X86ISD::GlobalBaseReg,
2018 DebugLoc(), getPointerTy()),
2020 InFlag = Chain.getValue(1);
2022 // If we are tail calling and generating PIC/GOT style code load the
2023 // address of the callee into ECX. The value in ecx is used as target of
2024 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2025 // for tail calls on PIC/GOT architectures. Normally we would just put the
2026 // address of GOT into ebx and then call target@PLT. But for tail calls
2027 // ebx would be restored (since ebx is callee saved) before jumping to the
2030 // Note: The actual moving to ECX is done further down.
2031 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2032 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2033 !G->getGlobal()->hasProtectedVisibility())
2034 Callee = LowerGlobalAddress(Callee, DAG);
2035 else if (isa<ExternalSymbolSDNode>(Callee))
2036 Callee = LowerExternalSymbol(Callee, DAG);
2040 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2041 // From AMD64 ABI document:
2042 // For calls that may call functions that use varargs or stdargs
2043 // (prototype-less calls or calls to functions containing ellipsis (...) in
2044 // the declaration) %al is used as hidden argument to specify the number
2045 // of SSE registers used. The contents of %al do not need to match exactly
2046 // the number of registers, but must be an ubound on the number of SSE
2047 // registers used and is in the range 0 - 8 inclusive.
2049 // Count the number of XMM registers allocated.
2050 static const unsigned XMMArgRegs[] = {
2051 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2052 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2054 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2055 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2056 && "SSE registers cannot be used when SSE is disabled");
2058 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2059 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2060 InFlag = Chain.getValue(1);
2064 // For tail calls lower the arguments to the 'real' stack slot.
2066 // Force all the incoming stack arguments to be loaded from the stack
2067 // before any new outgoing arguments are stored to the stack, because the
2068 // outgoing stack slots may alias the incoming argument stack slots, and
2069 // the alias isn't otherwise explicit. This is slightly more conservative
2070 // than necessary, because it means that each store effectively depends
2071 // on every argument instead of just those arguments it would clobber.
2072 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2074 SmallVector<SDValue, 8> MemOpChains2;
2077 // Do not flag preceeding copytoreg stuff together with the following stuff.
2079 if (GuaranteedTailCallOpt) {
2080 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2081 CCValAssign &VA = ArgLocs[i];
2084 assert(VA.isMemLoc());
2085 SDValue Arg = OutVals[i];
2086 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2087 // Create frame index.
2088 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2089 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2090 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2091 FIN = DAG.getFrameIndex(FI, getPointerTy());
2093 if (Flags.isByVal()) {
2094 // Copy relative to framepointer.
2095 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2096 if (StackPtr.getNode() == 0)
2097 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2099 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2101 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2105 // Store relative to framepointer.
2106 MemOpChains2.push_back(
2107 DAG.getStore(ArgChain, dl, Arg, FIN,
2108 MachinePointerInfo::getFixedStack(FI),
2114 if (!MemOpChains2.empty())
2115 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2116 &MemOpChains2[0], MemOpChains2.size());
2118 // Copy arguments to their registers.
2119 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2120 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2121 RegsToPass[i].second, InFlag);
2122 InFlag = Chain.getValue(1);
2126 // Store the return address to the appropriate stack slot.
2127 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2131 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2132 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2133 // In the 64-bit large code model, we have to make all calls
2134 // through a register, since the call instruction's 32-bit
2135 // pc-relative offset may not be large enough to hold the whole
2137 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2138 // If the callee is a GlobalAddress node (quite common, every direct call
2139 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2142 // We should use extra load for direct calls to dllimported functions in
2144 const GlobalValue *GV = G->getGlobal();
2145 if (!GV->hasDLLImportLinkage()) {
2146 unsigned char OpFlags = 0;
2148 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2149 // external symbols most go through the PLT in PIC mode. If the symbol
2150 // has hidden or protected visibility, or if it is static or local, then
2151 // we don't need to use the PLT - we can directly call it.
2152 if (Subtarget->isTargetELF() &&
2153 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2154 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2155 OpFlags = X86II::MO_PLT;
2156 } else if (Subtarget->isPICStyleStubAny() &&
2157 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2158 Subtarget->getDarwinVers() < 9) {
2159 // PC-relative references to external symbols should go through $stub,
2160 // unless we're building with the leopard linker or later, which
2161 // automatically synthesizes these stubs.
2162 OpFlags = X86II::MO_DARWIN_STUB;
2165 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2166 G->getOffset(), OpFlags);
2168 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2169 unsigned char OpFlags = 0;
2171 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2172 // symbols should go through the PLT.
2173 if (Subtarget->isTargetELF() &&
2174 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2175 OpFlags = X86II::MO_PLT;
2176 } else if (Subtarget->isPICStyleStubAny() &&
2177 Subtarget->getDarwinVers() < 9) {
2178 // PC-relative references to external symbols should go through $stub,
2179 // unless we're building with the leopard linker or later, which
2180 // automatically synthesizes these stubs.
2181 OpFlags = X86II::MO_DARWIN_STUB;
2184 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2188 // Returns a chain & a flag for retval copy to use.
2189 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2190 SmallVector<SDValue, 8> Ops;
2192 if (!IsSibcall && isTailCall) {
2193 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2194 DAG.getIntPtrConstant(0, true), InFlag);
2195 InFlag = Chain.getValue(1);
2198 Ops.push_back(Chain);
2199 Ops.push_back(Callee);
2202 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2204 // Add argument registers to the end of the list so that they are known live
2206 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2207 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2208 RegsToPass[i].second.getValueType()));
2210 // Add an implicit use GOT pointer in EBX.
2211 if (!isTailCall && Subtarget->isPICStyleGOT())
2212 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2214 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2215 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2216 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2218 if (InFlag.getNode())
2219 Ops.push_back(InFlag);
2223 //// If this is the first return lowered for this function, add the regs
2224 //// to the liveout set for the function.
2225 // This isn't right, although it's probably harmless on x86; liveouts
2226 // should be computed from returns not tail calls. Consider a void
2227 // function making a tail call to a function returning int.
2228 return DAG.getNode(X86ISD::TC_RETURN, dl,
2229 NodeTys, &Ops[0], Ops.size());
2232 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2233 InFlag = Chain.getValue(1);
2235 // Create the CALLSEQ_END node.
2236 unsigned NumBytesForCalleeToPush;
2237 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2238 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2239 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2240 // If this is a call to a struct-return function, the callee
2241 // pops the hidden struct pointer, so we have to push it back.
2242 // This is common for Darwin/X86, Linux & Mingw32 targets.
2243 NumBytesForCalleeToPush = 4;
2245 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2247 // Returns a flag for retval copy to use.
2249 Chain = DAG.getCALLSEQ_END(Chain,
2250 DAG.getIntPtrConstant(NumBytes, true),
2251 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2254 InFlag = Chain.getValue(1);
2257 // Handle result values, copying them out of physregs into vregs that we
2259 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2260 Ins, dl, DAG, InVals);
2264 //===----------------------------------------------------------------------===//
2265 // Fast Calling Convention (tail call) implementation
2266 //===----------------------------------------------------------------------===//
2268 // Like std call, callee cleans arguments, convention except that ECX is
2269 // reserved for storing the tail called function address. Only 2 registers are
2270 // free for argument passing (inreg). Tail call optimization is performed
2272 // * tailcallopt is enabled
2273 // * caller/callee are fastcc
2274 // On X86_64 architecture with GOT-style position independent code only local
2275 // (within module) calls are supported at the moment.
2276 // To keep the stack aligned according to platform abi the function
2277 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2278 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2279 // If a tail called function callee has more arguments than the caller the
2280 // caller needs to make sure that there is room to move the RETADDR to. This is
2281 // achieved by reserving an area the size of the argument delta right after the
2282 // original REtADDR, but before the saved framepointer or the spilled registers
2283 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2295 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2296 /// for a 16 byte align requirement.
2298 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2299 SelectionDAG& DAG) const {
2300 MachineFunction &MF = DAG.getMachineFunction();
2301 const TargetMachine &TM = MF.getTarget();
2302 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2303 unsigned StackAlignment = TFI.getStackAlignment();
2304 uint64_t AlignMask = StackAlignment - 1;
2305 int64_t Offset = StackSize;
2306 uint64_t SlotSize = TD->getPointerSize();
2307 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2308 // Number smaller than 12 so just add the difference.
2309 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2311 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2312 Offset = ((~AlignMask) & Offset) + StackAlignment +
2313 (StackAlignment-SlotSize);
2318 /// MatchingStackOffset - Return true if the given stack call argument is
2319 /// already available in the same position (relatively) of the caller's
2320 /// incoming argument stack.
2322 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2323 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2324 const X86InstrInfo *TII) {
2325 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2327 if (Arg.getOpcode() == ISD::CopyFromReg) {
2328 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2329 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2331 MachineInstr *Def = MRI->getVRegDef(VR);
2334 if (!Flags.isByVal()) {
2335 if (!TII->isLoadFromStackSlot(Def, FI))
2338 unsigned Opcode = Def->getOpcode();
2339 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2340 Def->getOperand(1).isFI()) {
2341 FI = Def->getOperand(1).getIndex();
2342 Bytes = Flags.getByValSize();
2346 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2347 if (Flags.isByVal())
2348 // ByVal argument is passed in as a pointer but it's now being
2349 // dereferenced. e.g.
2350 // define @foo(%struct.X* %A) {
2351 // tail call @bar(%struct.X* byval %A)
2354 SDValue Ptr = Ld->getBasePtr();
2355 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2358 FI = FINode->getIndex();
2362 assert(FI != INT_MAX);
2363 if (!MFI->isFixedObjectIndex(FI))
2365 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2368 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2369 /// for tail call optimization. Targets which want to do tail call
2370 /// optimization should implement this function.
2372 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2373 CallingConv::ID CalleeCC,
2375 bool isCalleeStructRet,
2376 bool isCallerStructRet,
2377 const SmallVectorImpl<ISD::OutputArg> &Outs,
2378 const SmallVectorImpl<SDValue> &OutVals,
2379 const SmallVectorImpl<ISD::InputArg> &Ins,
2380 SelectionDAG& DAG) const {
2381 if (!IsTailCallConvention(CalleeCC) &&
2382 CalleeCC != CallingConv::C)
2385 // If -tailcallopt is specified, make fastcc functions tail-callable.
2386 const MachineFunction &MF = DAG.getMachineFunction();
2387 const Function *CallerF = DAG.getMachineFunction().getFunction();
2388 CallingConv::ID CallerCC = CallerF->getCallingConv();
2389 bool CCMatch = CallerCC == CalleeCC;
2391 if (GuaranteedTailCallOpt) {
2392 if (IsTailCallConvention(CalleeCC) && CCMatch)
2397 // Look for obvious safe cases to perform tail call optimization that do not
2398 // require ABI changes. This is what gcc calls sibcall.
2400 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2401 // emit a special epilogue.
2402 if (RegInfo->needsStackRealignment(MF))
2405 // Do not sibcall optimize vararg calls unless the call site is not passing
2407 if (isVarArg && !Outs.empty())
2410 // Also avoid sibcall optimization if either caller or callee uses struct
2411 // return semantics.
2412 if (isCalleeStructRet || isCallerStructRet)
2415 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2416 // Therefore if it's not used by the call it is not safe to optimize this into
2418 bool Unused = false;
2419 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2426 SmallVector<CCValAssign, 16> RVLocs;
2427 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2428 RVLocs, *DAG.getContext());
2429 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2430 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2431 CCValAssign &VA = RVLocs[i];
2432 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2437 // If the calling conventions do not match, then we'd better make sure the
2438 // results are returned in the same way as what the caller expects.
2440 SmallVector<CCValAssign, 16> RVLocs1;
2441 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2442 RVLocs1, *DAG.getContext());
2443 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2445 SmallVector<CCValAssign, 16> RVLocs2;
2446 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2447 RVLocs2, *DAG.getContext());
2448 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2450 if (RVLocs1.size() != RVLocs2.size())
2452 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2453 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2455 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2457 if (RVLocs1[i].isRegLoc()) {
2458 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2461 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2467 // If the callee takes no arguments then go on to check the results of the
2469 if (!Outs.empty()) {
2470 // Check if stack adjustment is needed. For now, do not do this if any
2471 // argument is passed on the stack.
2472 SmallVector<CCValAssign, 16> ArgLocs;
2473 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2474 ArgLocs, *DAG.getContext());
2475 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2476 if (CCInfo.getNextStackOffset()) {
2477 MachineFunction &MF = DAG.getMachineFunction();
2478 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2480 if (Subtarget->isTargetWin64())
2481 // Win64 ABI has additional complications.
2484 // Check if the arguments are already laid out in the right way as
2485 // the caller's fixed stack objects.
2486 MachineFrameInfo *MFI = MF.getFrameInfo();
2487 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2488 const X86InstrInfo *TII =
2489 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2490 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2491 CCValAssign &VA = ArgLocs[i];
2492 SDValue Arg = OutVals[i];
2493 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2494 if (VA.getLocInfo() == CCValAssign::Indirect)
2496 if (!VA.isRegLoc()) {
2497 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2504 // If the tailcall address may be in a register, then make sure it's
2505 // possible to register allocate for it. In 32-bit, the call address can
2506 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2507 // callee-saved registers are restored. These happen to be the same
2508 // registers used to pass 'inreg' arguments so watch out for those.
2509 if (!Subtarget->is64Bit() &&
2510 !isa<GlobalAddressSDNode>(Callee) &&
2511 !isa<ExternalSymbolSDNode>(Callee)) {
2512 unsigned NumInRegs = 0;
2513 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2514 CCValAssign &VA = ArgLocs[i];
2517 unsigned Reg = VA.getLocReg();
2520 case X86::EAX: case X86::EDX: case X86::ECX:
2521 if (++NumInRegs == 3)
2533 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2534 return X86::createFastISel(funcInfo);
2538 //===----------------------------------------------------------------------===//
2539 // Other Lowering Hooks
2540 //===----------------------------------------------------------------------===//
2542 static bool MayFoldLoad(SDValue Op) {
2543 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2546 static bool MayFoldIntoStore(SDValue Op) {
2547 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2550 static bool isTargetShuffle(unsigned Opcode) {
2552 default: return false;
2553 case X86ISD::PSHUFD:
2554 case X86ISD::PSHUFHW:
2555 case X86ISD::PSHUFLW:
2556 case X86ISD::SHUFPD:
2557 case X86ISD::PALIGN:
2558 case X86ISD::SHUFPS:
2559 case X86ISD::MOVLHPS:
2560 case X86ISD::MOVLHPD:
2561 case X86ISD::MOVHLPS:
2562 case X86ISD::MOVLPS:
2563 case X86ISD::MOVLPD:
2564 case X86ISD::MOVSHDUP:
2565 case X86ISD::MOVSLDUP:
2566 case X86ISD::MOVDDUP:
2569 case X86ISD::UNPCKLPS:
2570 case X86ISD::UNPCKLPD:
2571 case X86ISD::PUNPCKLWD:
2572 case X86ISD::PUNPCKLBW:
2573 case X86ISD::PUNPCKLDQ:
2574 case X86ISD::PUNPCKLQDQ:
2575 case X86ISD::UNPCKHPS:
2576 case X86ISD::UNPCKHPD:
2577 case X86ISD::PUNPCKHWD:
2578 case X86ISD::PUNPCKHBW:
2579 case X86ISD::PUNPCKHDQ:
2580 case X86ISD::PUNPCKHQDQ:
2586 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2587 SDValue V1, SelectionDAG &DAG) {
2589 default: llvm_unreachable("Unknown x86 shuffle node");
2590 case X86ISD::MOVSHDUP:
2591 case X86ISD::MOVSLDUP:
2592 case X86ISD::MOVDDUP:
2593 return DAG.getNode(Opc, dl, VT, V1);
2599 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2600 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2602 default: llvm_unreachable("Unknown x86 shuffle node");
2603 case X86ISD::PSHUFD:
2604 case X86ISD::PSHUFHW:
2605 case X86ISD::PSHUFLW:
2606 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2612 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2613 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2615 default: llvm_unreachable("Unknown x86 shuffle node");
2616 case X86ISD::PALIGN:
2617 case X86ISD::SHUFPD:
2618 case X86ISD::SHUFPS:
2619 return DAG.getNode(Opc, dl, VT, V1, V2,
2620 DAG.getConstant(TargetMask, MVT::i8));
2625 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2626 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2628 default: llvm_unreachable("Unknown x86 shuffle node");
2629 case X86ISD::MOVLHPS:
2630 case X86ISD::MOVLHPD:
2631 case X86ISD::MOVHLPS:
2632 case X86ISD::MOVLPS:
2633 case X86ISD::MOVLPD:
2636 case X86ISD::UNPCKLPS:
2637 case X86ISD::UNPCKLPD:
2638 case X86ISD::PUNPCKLWD:
2639 case X86ISD::PUNPCKLBW:
2640 case X86ISD::PUNPCKLDQ:
2641 case X86ISD::PUNPCKLQDQ:
2642 case X86ISD::UNPCKHPS:
2643 case X86ISD::UNPCKHPD:
2644 case X86ISD::PUNPCKHWD:
2645 case X86ISD::PUNPCKHBW:
2646 case X86ISD::PUNPCKHDQ:
2647 case X86ISD::PUNPCKHQDQ:
2648 return DAG.getNode(Opc, dl, VT, V1, V2);
2653 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2654 MachineFunction &MF = DAG.getMachineFunction();
2655 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2656 int ReturnAddrIndex = FuncInfo->getRAIndex();
2658 if (ReturnAddrIndex == 0) {
2659 // Set up a frame object for the return address.
2660 uint64_t SlotSize = TD->getPointerSize();
2661 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2663 FuncInfo->setRAIndex(ReturnAddrIndex);
2666 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2670 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2671 bool hasSymbolicDisplacement) {
2672 // Offset should fit into 32 bit immediate field.
2673 if (!isInt<32>(Offset))
2676 // If we don't have a symbolic displacement - we don't have any extra
2678 if (!hasSymbolicDisplacement)
2681 // FIXME: Some tweaks might be needed for medium code model.
2682 if (M != CodeModel::Small && M != CodeModel::Kernel)
2685 // For small code model we assume that latest object is 16MB before end of 31
2686 // bits boundary. We may also accept pretty large negative constants knowing
2687 // that all objects are in the positive half of address space.
2688 if (M == CodeModel::Small && Offset < 16*1024*1024)
2691 // For kernel code model we know that all object resist in the negative half
2692 // of 32bits address space. We may not accept negative offsets, since they may
2693 // be just off and we may accept pretty large positive ones.
2694 if (M == CodeModel::Kernel && Offset > 0)
2700 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2701 /// specific condition code, returning the condition code and the LHS/RHS of the
2702 /// comparison to make.
2703 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2704 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2706 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2707 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2708 // X > -1 -> X == 0, jump !sign.
2709 RHS = DAG.getConstant(0, RHS.getValueType());
2710 return X86::COND_NS;
2711 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2712 // X < 0 -> X == 0, jump on sign.
2714 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2716 RHS = DAG.getConstant(0, RHS.getValueType());
2717 return X86::COND_LE;
2721 switch (SetCCOpcode) {
2722 default: llvm_unreachable("Invalid integer condition!");
2723 case ISD::SETEQ: return X86::COND_E;
2724 case ISD::SETGT: return X86::COND_G;
2725 case ISD::SETGE: return X86::COND_GE;
2726 case ISD::SETLT: return X86::COND_L;
2727 case ISD::SETLE: return X86::COND_LE;
2728 case ISD::SETNE: return X86::COND_NE;
2729 case ISD::SETULT: return X86::COND_B;
2730 case ISD::SETUGT: return X86::COND_A;
2731 case ISD::SETULE: return X86::COND_BE;
2732 case ISD::SETUGE: return X86::COND_AE;
2736 // First determine if it is required or is profitable to flip the operands.
2738 // If LHS is a foldable load, but RHS is not, flip the condition.
2739 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2740 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2741 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2742 std::swap(LHS, RHS);
2745 switch (SetCCOpcode) {
2751 std::swap(LHS, RHS);
2755 // On a floating point condition, the flags are set as follows:
2757 // 0 | 0 | 0 | X > Y
2758 // 0 | 0 | 1 | X < Y
2759 // 1 | 0 | 0 | X == Y
2760 // 1 | 1 | 1 | unordered
2761 switch (SetCCOpcode) {
2762 default: llvm_unreachable("Condcode should be pre-legalized away");
2764 case ISD::SETEQ: return X86::COND_E;
2765 case ISD::SETOLT: // flipped
2767 case ISD::SETGT: return X86::COND_A;
2768 case ISD::SETOLE: // flipped
2770 case ISD::SETGE: return X86::COND_AE;
2771 case ISD::SETUGT: // flipped
2773 case ISD::SETLT: return X86::COND_B;
2774 case ISD::SETUGE: // flipped
2776 case ISD::SETLE: return X86::COND_BE;
2778 case ISD::SETNE: return X86::COND_NE;
2779 case ISD::SETUO: return X86::COND_P;
2780 case ISD::SETO: return X86::COND_NP;
2782 case ISD::SETUNE: return X86::COND_INVALID;
2786 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2787 /// code. Current x86 isa includes the following FP cmov instructions:
2788 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2789 static bool hasFPCMov(unsigned X86CC) {
2805 /// isFPImmLegal - Returns true if the target can instruction select the
2806 /// specified FP immediate natively. If false, the legalizer will
2807 /// materialize the FP immediate as a load from a constant pool.
2808 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2809 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2810 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2816 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2817 /// the specified range (L, H].
2818 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2819 return (Val < 0) || (Val >= Low && Val < Hi);
2822 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2823 /// specified value.
2824 static bool isUndefOrEqual(int Val, int CmpVal) {
2825 if (Val < 0 || Val == CmpVal)
2830 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2831 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2832 /// the second operand.
2833 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2834 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2835 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2836 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2837 return (Mask[0] < 2 && Mask[1] < 2);
2841 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2842 SmallVector<int, 8> M;
2844 return ::isPSHUFDMask(M, N->getValueType(0));
2847 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2848 /// is suitable for input to PSHUFHW.
2849 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2850 if (VT != MVT::v8i16)
2853 // Lower quadword copied in order or undef.
2854 for (int i = 0; i != 4; ++i)
2855 if (Mask[i] >= 0 && Mask[i] != i)
2858 // Upper quadword shuffled.
2859 for (int i = 4; i != 8; ++i)
2860 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2866 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2867 SmallVector<int, 8> M;
2869 return ::isPSHUFHWMask(M, N->getValueType(0));
2872 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2873 /// is suitable for input to PSHUFLW.
2874 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2875 if (VT != MVT::v8i16)
2878 // Upper quadword copied in order.
2879 for (int i = 4; i != 8; ++i)
2880 if (Mask[i] >= 0 && Mask[i] != i)
2883 // Lower quadword shuffled.
2884 for (int i = 0; i != 4; ++i)
2891 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2892 SmallVector<int, 8> M;
2894 return ::isPSHUFLWMask(M, N->getValueType(0));
2897 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2898 /// is suitable for input to PALIGNR.
2899 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2901 int i, e = VT.getVectorNumElements();
2903 // Do not handle v2i64 / v2f64 shuffles with palignr.
2904 if (e < 4 || !hasSSSE3)
2907 for (i = 0; i != e; ++i)
2911 // All undef, not a palignr.
2915 // Determine if it's ok to perform a palignr with only the LHS, since we
2916 // don't have access to the actual shuffle elements to see if RHS is undef.
2917 bool Unary = Mask[i] < (int)e;
2918 bool NeedsUnary = false;
2920 int s = Mask[i] - i;
2922 // Check the rest of the elements to see if they are consecutive.
2923 for (++i; i != e; ++i) {
2928 Unary = Unary && (m < (int)e);
2929 NeedsUnary = NeedsUnary || (m < s);
2931 if (NeedsUnary && !Unary)
2933 if (Unary && m != ((s+i) & (e-1)))
2935 if (!Unary && m != (s+i))
2941 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2942 SmallVector<int, 8> M;
2944 return ::isPALIGNRMask(M, N->getValueType(0), true);
2947 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2948 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2949 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2950 int NumElems = VT.getVectorNumElements();
2951 if (NumElems != 2 && NumElems != 4)
2954 int Half = NumElems / 2;
2955 for (int i = 0; i < Half; ++i)
2956 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2958 for (int i = Half; i < NumElems; ++i)
2959 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2965 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2966 SmallVector<int, 8> M;
2968 return ::isSHUFPMask(M, N->getValueType(0));
2971 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2972 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2973 /// half elements to come from vector 1 (which would equal the dest.) and
2974 /// the upper half to come from vector 2.
2975 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2976 int NumElems = VT.getVectorNumElements();
2978 if (NumElems != 2 && NumElems != 4)
2981 int Half = NumElems / 2;
2982 for (int i = 0; i < Half; ++i)
2983 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2985 for (int i = Half; i < NumElems; ++i)
2986 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2991 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2992 SmallVector<int, 8> M;
2994 return isCommutedSHUFPMask(M, N->getValueType(0));
2997 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2998 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2999 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3000 if (N->getValueType(0).getVectorNumElements() != 4)
3003 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3004 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3005 isUndefOrEqual(N->getMaskElt(1), 7) &&
3006 isUndefOrEqual(N->getMaskElt(2), 2) &&
3007 isUndefOrEqual(N->getMaskElt(3), 3);
3010 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3011 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3013 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3014 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3019 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3020 isUndefOrEqual(N->getMaskElt(1), 3) &&
3021 isUndefOrEqual(N->getMaskElt(2), 2) &&
3022 isUndefOrEqual(N->getMaskElt(3), 3);
3025 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3026 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3027 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3028 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3030 if (NumElems != 2 && NumElems != 4)
3033 for (unsigned i = 0; i < NumElems/2; ++i)
3034 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3037 for (unsigned i = NumElems/2; i < NumElems; ++i)
3038 if (!isUndefOrEqual(N->getMaskElt(i), i))
3044 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3045 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3046 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3047 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3049 if (NumElems != 2 && NumElems != 4)
3052 for (unsigned i = 0; i < NumElems/2; ++i)
3053 if (!isUndefOrEqual(N->getMaskElt(i), i))
3056 for (unsigned i = 0; i < NumElems/2; ++i)
3057 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3063 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3064 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3065 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3066 bool V2IsSplat = false) {
3067 int NumElts = VT.getVectorNumElements();
3068 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3071 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3073 int BitI1 = Mask[i+1];
3074 if (!isUndefOrEqual(BitI, j))
3077 if (!isUndefOrEqual(BitI1, NumElts))
3080 if (!isUndefOrEqual(BitI1, j + NumElts))
3087 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3088 SmallVector<int, 8> M;
3090 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3093 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3094 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3095 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3096 bool V2IsSplat = false) {
3097 int NumElts = VT.getVectorNumElements();
3098 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3101 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3103 int BitI1 = Mask[i+1];
3104 if (!isUndefOrEqual(BitI, j + NumElts/2))
3107 if (isUndefOrEqual(BitI1, NumElts))
3110 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3117 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3118 SmallVector<int, 8> M;
3120 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3123 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3124 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3126 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3127 int NumElems = VT.getVectorNumElements();
3128 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3131 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3133 int BitI1 = Mask[i+1];
3134 if (!isUndefOrEqual(BitI, j))
3136 if (!isUndefOrEqual(BitI1, j))
3142 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3143 SmallVector<int, 8> M;
3145 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3148 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3149 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3151 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3152 int NumElems = VT.getVectorNumElements();
3153 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3156 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3158 int BitI1 = Mask[i+1];
3159 if (!isUndefOrEqual(BitI, j))
3161 if (!isUndefOrEqual(BitI1, j))
3167 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3168 SmallVector<int, 8> M;
3170 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3173 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3174 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3175 /// MOVSD, and MOVD, i.e. setting the lowest element.
3176 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3177 if (VT.getVectorElementType().getSizeInBits() < 32)
3180 int NumElts = VT.getVectorNumElements();
3182 if (!isUndefOrEqual(Mask[0], NumElts))
3185 for (int i = 1; i < NumElts; ++i)
3186 if (!isUndefOrEqual(Mask[i], i))
3192 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3193 SmallVector<int, 8> M;
3195 return ::isMOVLMask(M, N->getValueType(0));
3198 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3199 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3200 /// element of vector 2 and the other elements to come from vector 1 in order.
3201 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3202 bool V2IsSplat = false, bool V2IsUndef = false) {
3203 int NumOps = VT.getVectorNumElements();
3204 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3207 if (!isUndefOrEqual(Mask[0], 0))
3210 for (int i = 1; i < NumOps; ++i)
3211 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3212 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3213 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3219 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3220 bool V2IsUndef = false) {
3221 SmallVector<int, 8> M;
3223 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3226 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3227 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3228 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3229 if (N->getValueType(0).getVectorNumElements() != 4)
3232 // Expect 1, 1, 3, 3
3233 for (unsigned i = 0; i < 2; ++i) {
3234 int Elt = N->getMaskElt(i);
3235 if (Elt >= 0 && Elt != 1)
3240 for (unsigned i = 2; i < 4; ++i) {
3241 int Elt = N->getMaskElt(i);
3242 if (Elt >= 0 && Elt != 3)
3247 // Don't use movshdup if it can be done with a shufps.
3248 // FIXME: verify that matching u, u, 3, 3 is what we want.
3252 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3253 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3254 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3255 if (N->getValueType(0).getVectorNumElements() != 4)
3258 // Expect 0, 0, 2, 2
3259 for (unsigned i = 0; i < 2; ++i)
3260 if (N->getMaskElt(i) > 0)
3264 for (unsigned i = 2; i < 4; ++i) {
3265 int Elt = N->getMaskElt(i);
3266 if (Elt >= 0 && Elt != 2)
3271 // Don't use movsldup if it can be done with a shufps.
3275 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3276 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3277 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3278 int e = N->getValueType(0).getVectorNumElements() / 2;
3280 for (int i = 0; i < e; ++i)
3281 if (!isUndefOrEqual(N->getMaskElt(i), i))
3283 for (int i = 0; i < e; ++i)
3284 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3289 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3290 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3291 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3292 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3293 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3295 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3297 for (int i = 0; i < NumOperands; ++i) {
3298 int Val = SVOp->getMaskElt(NumOperands-i-1);
3299 if (Val < 0) Val = 0;
3300 if (Val >= NumOperands) Val -= NumOperands;
3302 if (i != NumOperands - 1)
3308 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3309 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3310 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3311 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3313 // 8 nodes, but we only care about the last 4.
3314 for (unsigned i = 7; i >= 4; --i) {
3315 int Val = SVOp->getMaskElt(i);
3324 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3325 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3326 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3327 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3329 // 8 nodes, but we only care about the first 4.
3330 for (int i = 3; i >= 0; --i) {
3331 int Val = SVOp->getMaskElt(i);
3340 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3341 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3342 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3343 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3344 EVT VVT = N->getValueType(0);
3345 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3349 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3350 Val = SVOp->getMaskElt(i);
3354 return (Val - i) * EltSize;
3357 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3359 bool X86::isZeroNode(SDValue Elt) {
3360 return ((isa<ConstantSDNode>(Elt) &&
3361 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3362 (isa<ConstantFPSDNode>(Elt) &&
3363 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3366 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3367 /// their permute mask.
3368 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3369 SelectionDAG &DAG) {
3370 EVT VT = SVOp->getValueType(0);
3371 unsigned NumElems = VT.getVectorNumElements();
3372 SmallVector<int, 8> MaskVec;
3374 for (unsigned i = 0; i != NumElems; ++i) {
3375 int idx = SVOp->getMaskElt(i);
3377 MaskVec.push_back(idx);
3378 else if (idx < (int)NumElems)
3379 MaskVec.push_back(idx + NumElems);
3381 MaskVec.push_back(idx - NumElems);
3383 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3384 SVOp->getOperand(0), &MaskVec[0]);
3387 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3388 /// the two vector operands have swapped position.
3389 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3390 unsigned NumElems = VT.getVectorNumElements();
3391 for (unsigned i = 0; i != NumElems; ++i) {
3395 else if (idx < (int)NumElems)
3396 Mask[i] = idx + NumElems;
3398 Mask[i] = idx - NumElems;
3402 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3403 /// match movhlps. The lower half elements should come from upper half of
3404 /// V1 (and in order), and the upper half elements should come from the upper
3405 /// half of V2 (and in order).
3406 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3407 if (Op->getValueType(0).getVectorNumElements() != 4)
3409 for (unsigned i = 0, e = 2; i != e; ++i)
3410 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3412 for (unsigned i = 2; i != 4; ++i)
3413 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3418 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3419 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3421 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3422 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3424 N = N->getOperand(0).getNode();
3425 if (!ISD::isNON_EXTLoad(N))
3428 *LD = cast<LoadSDNode>(N);
3432 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3433 /// match movlp{s|d}. The lower half elements should come from lower half of
3434 /// V1 (and in order), and the upper half elements should come from the upper
3435 /// half of V2 (and in order). And since V1 will become the source of the
3436 /// MOVLP, it must be either a vector load or a scalar load to vector.
3437 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3438 ShuffleVectorSDNode *Op) {
3439 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3441 // Is V2 is a vector load, don't do this transformation. We will try to use
3442 // load folding shufps op.
3443 if (ISD::isNON_EXTLoad(V2))
3446 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3448 if (NumElems != 2 && NumElems != 4)
3450 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3451 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3453 for (unsigned i = NumElems/2; i != NumElems; ++i)
3454 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3459 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3461 static bool isSplatVector(SDNode *N) {
3462 if (N->getOpcode() != ISD::BUILD_VECTOR)
3465 SDValue SplatValue = N->getOperand(0);
3466 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3467 if (N->getOperand(i) != SplatValue)
3472 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3473 /// to an zero vector.
3474 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3475 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3476 SDValue V1 = N->getOperand(0);
3477 SDValue V2 = N->getOperand(1);
3478 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3479 for (unsigned i = 0; i != NumElems; ++i) {
3480 int Idx = N->getMaskElt(i);
3481 if (Idx >= (int)NumElems) {
3482 unsigned Opc = V2.getOpcode();
3483 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3485 if (Opc != ISD::BUILD_VECTOR ||
3486 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3488 } else if (Idx >= 0) {
3489 unsigned Opc = V1.getOpcode();
3490 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3492 if (Opc != ISD::BUILD_VECTOR ||
3493 !X86::isZeroNode(V1.getOperand(Idx)))
3500 /// getZeroVector - Returns a vector of specified type with all zero elements.
3502 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3504 assert(VT.isVector() && "Expected a vector type");
3506 // Always build SSE zero vectors as <4 x i32> bitcasted
3507 // to their dest type. This ensures they get CSE'd.
3509 if (VT.getSizeInBits() == 128) { // SSE
3510 if (HasSSE2) { // SSE2
3511 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3512 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3514 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3515 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3517 } else if (VT.getSizeInBits() == 256) { // AVX
3518 // 256-bit logic and arithmetic instructions in AVX are
3519 // all floating-point, no support for integer ops. Default
3520 // to emitting fp zeroed vectors then.
3521 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3522 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3523 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3525 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3528 /// getOnesVector - Returns a vector of specified type with all bits set.
3530 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3531 assert(VT.isVector() && "Expected a vector type");
3533 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3534 // type. This ensures they get CSE'd.
3535 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3537 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3538 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3542 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3543 /// that point to V2 points to its first element.
3544 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3545 EVT VT = SVOp->getValueType(0);
3546 unsigned NumElems = VT.getVectorNumElements();
3548 bool Changed = false;
3549 SmallVector<int, 8> MaskVec;
3550 SVOp->getMask(MaskVec);
3552 for (unsigned i = 0; i != NumElems; ++i) {
3553 if (MaskVec[i] > (int)NumElems) {
3554 MaskVec[i] = NumElems;
3559 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3560 SVOp->getOperand(1), &MaskVec[0]);
3561 return SDValue(SVOp, 0);
3564 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3565 /// operation of specified width.
3566 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3568 unsigned NumElems = VT.getVectorNumElements();
3569 SmallVector<int, 8> Mask;
3570 Mask.push_back(NumElems);
3571 for (unsigned i = 1; i != NumElems; ++i)
3573 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3576 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3577 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3579 unsigned NumElems = VT.getVectorNumElements();
3580 SmallVector<int, 8> Mask;
3581 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3583 Mask.push_back(i + NumElems);
3585 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3588 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3589 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3591 unsigned NumElems = VT.getVectorNumElements();
3592 unsigned Half = NumElems/2;
3593 SmallVector<int, 8> Mask;
3594 for (unsigned i = 0; i != Half; ++i) {
3595 Mask.push_back(i + Half);
3596 Mask.push_back(i + NumElems + Half);
3598 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3601 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3602 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3603 EVT PVT = MVT::v4f32;
3604 EVT VT = SV->getValueType(0);
3605 DebugLoc dl = SV->getDebugLoc();
3606 SDValue V1 = SV->getOperand(0);
3607 int NumElems = VT.getVectorNumElements();
3608 int EltNo = SV->getSplatIndex();
3610 // unpack elements to the correct location
3611 while (NumElems > 4) {
3612 if (EltNo < NumElems/2) {
3613 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3615 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3616 EltNo -= NumElems/2;
3621 // Perform the splat.
3622 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3623 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3624 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3625 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3628 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3629 /// vector of zero or undef vector. This produces a shuffle where the low
3630 /// element of V2 is swizzled into the zero/undef vector, landing at element
3631 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3632 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3633 bool isZero, bool HasSSE2,
3634 SelectionDAG &DAG) {
3635 EVT VT = V2.getValueType();
3637 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3638 unsigned NumElems = VT.getVectorNumElements();
3639 SmallVector<int, 16> MaskVec;
3640 for (unsigned i = 0; i != NumElems; ++i)
3641 // If this is the insertion idx, put the low elt of V2 here.
3642 MaskVec.push_back(i == Idx ? NumElems : i);
3643 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3646 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3647 /// element of the result of the vector shuffle.
3648 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3651 return SDValue(); // Limit search depth.
3653 SDValue V = SDValue(N, 0);
3654 EVT VT = V.getValueType();
3655 unsigned Opcode = V.getOpcode();
3657 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3658 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3659 Index = SV->getMaskElt(Index);
3662 return DAG.getUNDEF(VT.getVectorElementType());
3664 int NumElems = VT.getVectorNumElements();
3665 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3666 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3669 // Recurse into target specific vector shuffles to find scalars.
3670 if (isTargetShuffle(Opcode)) {
3671 int NumElems = VT.getVectorNumElements();
3672 SmallVector<unsigned, 16> ShuffleMask;
3676 case X86ISD::SHUFPS:
3677 case X86ISD::SHUFPD:
3678 ImmN = N->getOperand(N->getNumOperands()-1);
3679 DecodeSHUFPSMask(NumElems,
3680 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3683 case X86ISD::PUNPCKHBW:
3684 case X86ISD::PUNPCKHWD:
3685 case X86ISD::PUNPCKHDQ:
3686 case X86ISD::PUNPCKHQDQ:
3687 DecodePUNPCKHMask(NumElems, ShuffleMask);
3689 case X86ISD::UNPCKHPS:
3690 case X86ISD::UNPCKHPD:
3691 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3693 case X86ISD::PUNPCKLBW:
3694 case X86ISD::PUNPCKLWD:
3695 case X86ISD::PUNPCKLDQ:
3696 case X86ISD::PUNPCKLQDQ:
3697 DecodePUNPCKLMask(NumElems, ShuffleMask);
3699 case X86ISD::UNPCKLPS:
3700 case X86ISD::UNPCKLPD:
3701 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3703 case X86ISD::MOVHLPS:
3704 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3706 case X86ISD::MOVLHPS:
3707 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3709 case X86ISD::PSHUFD:
3710 ImmN = N->getOperand(N->getNumOperands()-1);
3711 DecodePSHUFMask(NumElems,
3712 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3715 case X86ISD::PSHUFHW:
3716 ImmN = N->getOperand(N->getNumOperands()-1);
3717 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3720 case X86ISD::PSHUFLW:
3721 ImmN = N->getOperand(N->getNumOperands()-1);
3722 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3726 case X86ISD::MOVSD: {
3727 // The index 0 always comes from the first element of the second source,
3728 // this is why MOVSS and MOVSD are used in the first place. The other
3729 // elements come from the other positions of the first source vector.
3730 unsigned OpNum = (Index == 0) ? 1 : 0;
3731 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3735 assert("not implemented for target shuffle node");
3739 Index = ShuffleMask[Index];
3741 return DAG.getUNDEF(VT.getVectorElementType());
3743 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3744 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3748 // Actual nodes that may contain scalar elements
3749 if (Opcode == ISD::BIT_CONVERT) {
3750 V = V.getOperand(0);
3751 EVT SrcVT = V.getValueType();
3752 unsigned NumElems = VT.getVectorNumElements();
3754 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3758 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3759 return (Index == 0) ? V.getOperand(0)
3760 : DAG.getUNDEF(VT.getVectorElementType());
3762 if (V.getOpcode() == ISD::BUILD_VECTOR)
3763 return V.getOperand(Index);
3768 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3769 /// shuffle operation which come from a consecutively from a zero. The
3770 /// search can start in two diferent directions, from left or right.
3772 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3773 bool ZerosFromLeft, SelectionDAG &DAG) {
3776 while (i < NumElems) {
3777 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3778 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3779 if (!(Elt.getNode() &&
3780 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3788 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3789 /// MaskE correspond consecutively to elements from one of the vector operands,
3790 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3792 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3793 int OpIdx, int NumElems, unsigned &OpNum) {
3794 bool SeenV1 = false;
3795 bool SeenV2 = false;
3797 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3798 int Idx = SVOp->getMaskElt(i);
3799 // Ignore undef indicies
3808 // Only accept consecutive elements from the same vector
3809 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3813 OpNum = SeenV1 ? 0 : 1;
3817 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3818 /// logical left shift of a vector.
3819 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3820 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3821 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3822 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3823 false /* check zeros from right */, DAG);
3829 // Considering the elements in the mask that are not consecutive zeros,
3830 // check if they consecutively come from only one of the source vectors.
3832 // V1 = {X, A, B, C} 0
3834 // vector_shuffle V1, V2 <1, 2, 3, X>
3836 if (!isShuffleMaskConsecutive(SVOp,
3837 0, // Mask Start Index
3838 NumElems-NumZeros-1, // Mask End Index
3839 NumZeros, // Where to start looking in the src vector
3840 NumElems, // Number of elements in vector
3841 OpSrc)) // Which source operand ?
3846 ShVal = SVOp->getOperand(OpSrc);
3850 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3851 /// logical left shift of a vector.
3852 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3853 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3854 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3855 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3856 true /* check zeros from left */, DAG);
3862 // Considering the elements in the mask that are not consecutive zeros,
3863 // check if they consecutively come from only one of the source vectors.
3865 // 0 { A, B, X, X } = V2
3867 // vector_shuffle V1, V2 <X, X, 4, 5>
3869 if (!isShuffleMaskConsecutive(SVOp,
3870 NumZeros, // Mask Start Index
3871 NumElems-1, // Mask End Index
3872 0, // Where to start looking in the src vector
3873 NumElems, // Number of elements in vector
3874 OpSrc)) // Which source operand ?
3879 ShVal = SVOp->getOperand(OpSrc);
3883 /// isVectorShift - Returns true if the shuffle can be implemented as a
3884 /// logical left or right shift of a vector.
3885 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3886 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3887 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3888 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3894 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3896 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3897 unsigned NumNonZero, unsigned NumZero,
3899 const TargetLowering &TLI) {
3903 DebugLoc dl = Op.getDebugLoc();
3906 for (unsigned i = 0; i < 16; ++i) {
3907 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3908 if (ThisIsNonZero && First) {
3910 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3912 V = DAG.getUNDEF(MVT::v8i16);
3917 SDValue ThisElt(0, 0), LastElt(0, 0);
3918 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3919 if (LastIsNonZero) {
3920 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3921 MVT::i16, Op.getOperand(i-1));
3923 if (ThisIsNonZero) {
3924 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3925 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3926 ThisElt, DAG.getConstant(8, MVT::i8));
3928 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3932 if (ThisElt.getNode())
3933 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3934 DAG.getIntPtrConstant(i/2));
3938 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3941 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3943 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3944 unsigned NumNonZero, unsigned NumZero,
3946 const TargetLowering &TLI) {
3950 DebugLoc dl = Op.getDebugLoc();
3953 for (unsigned i = 0; i < 8; ++i) {
3954 bool isNonZero = (NonZeros & (1 << i)) != 0;
3958 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3960 V = DAG.getUNDEF(MVT::v8i16);
3963 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3964 MVT::v8i16, V, Op.getOperand(i),
3965 DAG.getIntPtrConstant(i));
3972 /// getVShift - Return a vector logical shift node.
3974 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3975 unsigned NumBits, SelectionDAG &DAG,
3976 const TargetLowering &TLI, DebugLoc dl) {
3977 EVT ShVT = MVT::v2i64;
3978 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3979 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3980 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3981 DAG.getNode(Opc, dl, ShVT, SrcOp,
3982 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3986 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3987 SelectionDAG &DAG) const {
3989 // Check if the scalar load can be widened into a vector load. And if
3990 // the address is "base + cst" see if the cst can be "absorbed" into
3991 // the shuffle mask.
3992 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3993 SDValue Ptr = LD->getBasePtr();
3994 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3996 EVT PVT = LD->getValueType(0);
3997 if (PVT != MVT::i32 && PVT != MVT::f32)
4002 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4003 FI = FINode->getIndex();
4005 } else if (Ptr.getOpcode() == ISD::ADD &&
4006 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4007 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4008 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4009 Offset = Ptr.getConstantOperandVal(1);
4010 Ptr = Ptr.getOperand(0);
4015 SDValue Chain = LD->getChain();
4016 // Make sure the stack object alignment is at least 16.
4017 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4018 if (DAG.InferPtrAlignment(Ptr) < 16) {
4019 if (MFI->isFixedObjectIndex(FI)) {
4020 // Can't change the alignment. FIXME: It's possible to compute
4021 // the exact stack offset and reference FI + adjust offset instead.
4022 // If someone *really* cares about this. That's the way to implement it.
4025 MFI->setObjectAlignment(FI, 16);
4029 // (Offset % 16) must be multiple of 4. Then address is then
4030 // Ptr + (Offset & ~15).
4033 if ((Offset % 16) & 3)
4035 int64_t StartOffset = Offset & ~15;
4037 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4038 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4040 int EltNo = (Offset - StartOffset) >> 2;
4041 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4042 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4043 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4044 LD->getPointerInfo().getWithOffset(StartOffset),
4046 // Canonicalize it to a v4i32 shuffle.
4047 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4048 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4049 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4050 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4056 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4057 /// vector of type 'VT', see if the elements can be replaced by a single large
4058 /// load which has the same value as a build_vector whose operands are 'elts'.
4060 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4062 /// FIXME: we'd also like to handle the case where the last elements are zero
4063 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4064 /// There's even a handy isZeroNode for that purpose.
4065 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4066 DebugLoc &DL, SelectionDAG &DAG) {
4067 EVT EltVT = VT.getVectorElementType();
4068 unsigned NumElems = Elts.size();
4070 LoadSDNode *LDBase = NULL;
4071 unsigned LastLoadedElt = -1U;
4073 // For each element in the initializer, see if we've found a load or an undef.
4074 // If we don't find an initial load element, or later load elements are
4075 // non-consecutive, bail out.
4076 for (unsigned i = 0; i < NumElems; ++i) {
4077 SDValue Elt = Elts[i];
4079 if (!Elt.getNode() ||
4080 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4083 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4085 LDBase = cast<LoadSDNode>(Elt.getNode());
4089 if (Elt.getOpcode() == ISD::UNDEF)
4092 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4093 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4098 // If we have found an entire vector of loads and undefs, then return a large
4099 // load of the entire vector width starting at the base pointer. If we found
4100 // consecutive loads for the low half, generate a vzext_load node.
4101 if (LastLoadedElt == NumElems - 1) {
4102 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4103 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4104 LDBase->getPointerInfo(),
4105 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4106 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4107 LDBase->getPointerInfo(),
4108 LDBase->isVolatile(), LDBase->isNonTemporal(),
4109 LDBase->getAlignment());
4110 } else if (NumElems == 4 && LastLoadedElt == 1) {
4111 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4112 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4113 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4115 LDBase->getMemOperand());
4116 return DAG.getNode(ISD::BIT_CONVERT, DL, VT, ResNode);
4122 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4123 DebugLoc dl = Op.getDebugLoc();
4124 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4125 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4126 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4127 // is present, so AllOnes is ignored.
4128 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4129 (Op.getValueType().getSizeInBits() != 256 &&
4130 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4131 // Canonicalize this to <4 x i32> (SSE) to
4132 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4133 // eliminated on x86-32 hosts.
4134 if (Op.getValueType() == MVT::v4i32)
4137 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4138 return getOnesVector(Op.getValueType(), DAG, dl);
4139 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4142 EVT VT = Op.getValueType();
4143 EVT ExtVT = VT.getVectorElementType();
4144 unsigned EVTBits = ExtVT.getSizeInBits();
4146 unsigned NumElems = Op.getNumOperands();
4147 unsigned NumZero = 0;
4148 unsigned NumNonZero = 0;
4149 unsigned NonZeros = 0;
4150 bool IsAllConstants = true;
4151 SmallSet<SDValue, 8> Values;
4152 for (unsigned i = 0; i < NumElems; ++i) {
4153 SDValue Elt = Op.getOperand(i);
4154 if (Elt.getOpcode() == ISD::UNDEF)
4157 if (Elt.getOpcode() != ISD::Constant &&
4158 Elt.getOpcode() != ISD::ConstantFP)
4159 IsAllConstants = false;
4160 if (X86::isZeroNode(Elt))
4163 NonZeros |= (1 << i);
4168 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4169 if (NumNonZero == 0)
4170 return DAG.getUNDEF(VT);
4172 // Special case for single non-zero, non-undef, element.
4173 if (NumNonZero == 1) {
4174 unsigned Idx = CountTrailingZeros_32(NonZeros);
4175 SDValue Item = Op.getOperand(Idx);
4177 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4178 // the value are obviously zero, truncate the value to i32 and do the
4179 // insertion that way. Only do this if the value is non-constant or if the
4180 // value is a constant being inserted into element 0. It is cheaper to do
4181 // a constant pool load than it is to do a movd + shuffle.
4182 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4183 (!IsAllConstants || Idx == 0)) {
4184 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4186 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4187 EVT VecVT = MVT::v4i32;
4188 unsigned VecElts = 4;
4190 // Truncate the value (which may itself be a constant) to i32, and
4191 // convert it to a vector with movd (S2V+shuffle to zero extend).
4192 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4193 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4194 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4195 Subtarget->hasSSE2(), DAG);
4197 // Now we have our 32-bit value zero extended in the low element of
4198 // a vector. If Idx != 0, swizzle it into place.
4200 SmallVector<int, 4> Mask;
4201 Mask.push_back(Idx);
4202 for (unsigned i = 1; i != VecElts; ++i)
4204 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4205 DAG.getUNDEF(Item.getValueType()),
4208 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4212 // If we have a constant or non-constant insertion into the low element of
4213 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4214 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4215 // depending on what the source datatype is.
4218 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4219 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4220 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4221 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4222 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4223 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4225 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4226 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4227 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4228 EVT MiddleVT = MVT::v4i32;
4229 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4230 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4231 Subtarget->hasSSE2(), DAG);
4232 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4236 // Is it a vector logical left shift?
4237 if (NumElems == 2 && Idx == 1 &&
4238 X86::isZeroNode(Op.getOperand(0)) &&
4239 !X86::isZeroNode(Op.getOperand(1))) {
4240 unsigned NumBits = VT.getSizeInBits();
4241 return getVShift(true, VT,
4242 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4243 VT, Op.getOperand(1)),
4244 NumBits/2, DAG, *this, dl);
4247 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4250 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4251 // is a non-constant being inserted into an element other than the low one,
4252 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4253 // movd/movss) to move this into the low element, then shuffle it into
4255 if (EVTBits == 32) {
4256 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4258 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4259 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4260 Subtarget->hasSSE2(), DAG);
4261 SmallVector<int, 8> MaskVec;
4262 for (unsigned i = 0; i < NumElems; i++)
4263 MaskVec.push_back(i == Idx ? 0 : 1);
4264 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4268 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4269 if (Values.size() == 1) {
4270 if (EVTBits == 32) {
4271 // Instead of a shuffle like this:
4272 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4273 // Check if it's possible to issue this instead.
4274 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4275 unsigned Idx = CountTrailingZeros_32(NonZeros);
4276 SDValue Item = Op.getOperand(Idx);
4277 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4278 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4283 // A vector full of immediates; various special cases are already
4284 // handled, so this is best done with a single constant-pool load.
4288 // Let legalizer expand 2-wide build_vectors.
4289 if (EVTBits == 64) {
4290 if (NumNonZero == 1) {
4291 // One half is zero or undef.
4292 unsigned Idx = CountTrailingZeros_32(NonZeros);
4293 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4294 Op.getOperand(Idx));
4295 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4296 Subtarget->hasSSE2(), DAG);
4301 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4302 if (EVTBits == 8 && NumElems == 16) {
4303 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4305 if (V.getNode()) return V;
4308 if (EVTBits == 16 && NumElems == 8) {
4309 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4311 if (V.getNode()) return V;
4314 // If element VT is == 32 bits, turn it into a number of shuffles.
4315 SmallVector<SDValue, 8> V;
4317 if (NumElems == 4 && NumZero > 0) {
4318 for (unsigned i = 0; i < 4; ++i) {
4319 bool isZero = !(NonZeros & (1 << i));
4321 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4323 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4326 for (unsigned i = 0; i < 2; ++i) {
4327 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4330 V[i] = V[i*2]; // Must be a zero vector.
4333 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4336 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4339 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4344 SmallVector<int, 8> MaskVec;
4345 bool Reverse = (NonZeros & 0x3) == 2;
4346 for (unsigned i = 0; i < 2; ++i)
4347 MaskVec.push_back(Reverse ? 1-i : i);
4348 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4349 for (unsigned i = 0; i < 2; ++i)
4350 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4351 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4354 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4355 // Check for a build vector of consecutive loads.
4356 for (unsigned i = 0; i < NumElems; ++i)
4357 V[i] = Op.getOperand(i);
4359 // Check for elements which are consecutive loads.
4360 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4364 // For SSE 4.1, use insertps to put the high elements into the low element.
4365 if (getSubtarget()->hasSSE41()) {
4367 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4368 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4370 Result = DAG.getUNDEF(VT);
4372 for (unsigned i = 1; i < NumElems; ++i) {
4373 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4374 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4375 Op.getOperand(i), DAG.getIntPtrConstant(i));
4380 // Otherwise, expand into a number of unpckl*, start by extending each of
4381 // our (non-undef) elements to the full vector width with the element in the
4382 // bottom slot of the vector (which generates no code for SSE).
4383 for (unsigned i = 0; i < NumElems; ++i) {
4384 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4385 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4387 V[i] = DAG.getUNDEF(VT);
4390 // Next, we iteratively mix elements, e.g. for v4f32:
4391 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4392 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4393 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4394 unsigned EltStride = NumElems >> 1;
4395 while (EltStride != 0) {
4396 for (unsigned i = 0; i < EltStride; ++i) {
4397 // If V[i+EltStride] is undef and this is the first round of mixing,
4398 // then it is safe to just drop this shuffle: V[i] is already in the
4399 // right place, the one element (since it's the first round) being
4400 // inserted as undef can be dropped. This isn't safe for successive
4401 // rounds because they will permute elements within both vectors.
4402 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4403 EltStride == NumElems/2)
4406 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4416 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4417 // We support concatenate two MMX registers and place them in a MMX
4418 // register. This is better than doing a stack convert.
4419 DebugLoc dl = Op.getDebugLoc();
4420 EVT ResVT = Op.getValueType();
4421 assert(Op.getNumOperands() == 2);
4422 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4423 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4425 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4426 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4427 InVec = Op.getOperand(1);
4428 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4429 unsigned NumElts = ResVT.getVectorNumElements();
4430 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4431 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4432 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4434 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4435 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4436 Mask[0] = 0; Mask[1] = 2;
4437 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4439 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4442 // v8i16 shuffles - Prefer shuffles in the following order:
4443 // 1. [all] pshuflw, pshufhw, optional move
4444 // 2. [ssse3] 1 x pshufb
4445 // 3. [ssse3] 2 x pshufb + 1 x por
4446 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4448 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4449 SelectionDAG &DAG) const {
4450 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4451 SDValue V1 = SVOp->getOperand(0);
4452 SDValue V2 = SVOp->getOperand(1);
4453 DebugLoc dl = SVOp->getDebugLoc();
4454 SmallVector<int, 8> MaskVals;
4456 // Determine if more than 1 of the words in each of the low and high quadwords
4457 // of the result come from the same quadword of one of the two inputs. Undef
4458 // mask values count as coming from any quadword, for better codegen.
4459 SmallVector<unsigned, 4> LoQuad(4);
4460 SmallVector<unsigned, 4> HiQuad(4);
4461 BitVector InputQuads(4);
4462 for (unsigned i = 0; i < 8; ++i) {
4463 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4464 int EltIdx = SVOp->getMaskElt(i);
4465 MaskVals.push_back(EltIdx);
4474 InputQuads.set(EltIdx / 4);
4477 int BestLoQuad = -1;
4478 unsigned MaxQuad = 1;
4479 for (unsigned i = 0; i < 4; ++i) {
4480 if (LoQuad[i] > MaxQuad) {
4482 MaxQuad = LoQuad[i];
4486 int BestHiQuad = -1;
4488 for (unsigned i = 0; i < 4; ++i) {
4489 if (HiQuad[i] > MaxQuad) {
4491 MaxQuad = HiQuad[i];
4495 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4496 // of the two input vectors, shuffle them into one input vector so only a
4497 // single pshufb instruction is necessary. If There are more than 2 input
4498 // quads, disable the next transformation since it does not help SSSE3.
4499 bool V1Used = InputQuads[0] || InputQuads[1];
4500 bool V2Used = InputQuads[2] || InputQuads[3];
4501 if (Subtarget->hasSSSE3()) {
4502 if (InputQuads.count() == 2 && V1Used && V2Used) {
4503 BestLoQuad = InputQuads.find_first();
4504 BestHiQuad = InputQuads.find_next(BestLoQuad);
4506 if (InputQuads.count() > 2) {
4512 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4513 // the shuffle mask. If a quad is scored as -1, that means that it contains
4514 // words from all 4 input quadwords.
4516 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4517 SmallVector<int, 8> MaskV;
4518 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4519 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4520 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4521 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4522 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4523 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4525 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4526 // source words for the shuffle, to aid later transformations.
4527 bool AllWordsInNewV = true;
4528 bool InOrder[2] = { true, true };
4529 for (unsigned i = 0; i != 8; ++i) {
4530 int idx = MaskVals[i];
4532 InOrder[i/4] = false;
4533 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4535 AllWordsInNewV = false;
4539 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4540 if (AllWordsInNewV) {
4541 for (int i = 0; i != 8; ++i) {
4542 int idx = MaskVals[i];
4545 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4546 if ((idx != i) && idx < 4)
4548 if ((idx != i) && idx > 3)
4557 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4558 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4559 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4560 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4561 unsigned TargetMask = 0;
4562 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4563 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4564 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4565 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4566 V1 = NewV.getOperand(0);
4567 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4571 // If we have SSSE3, and all words of the result are from 1 input vector,
4572 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4573 // is present, fall back to case 4.
4574 if (Subtarget->hasSSSE3()) {
4575 SmallVector<SDValue,16> pshufbMask;
4577 // If we have elements from both input vectors, set the high bit of the
4578 // shuffle mask element to zero out elements that come from V2 in the V1
4579 // mask, and elements that come from V1 in the V2 mask, so that the two
4580 // results can be OR'd together.
4581 bool TwoInputs = V1Used && V2Used;
4582 for (unsigned i = 0; i != 8; ++i) {
4583 int EltIdx = MaskVals[i] * 2;
4584 if (TwoInputs && (EltIdx >= 16)) {
4585 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4586 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4589 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4590 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4592 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4593 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4594 DAG.getNode(ISD::BUILD_VECTOR, dl,
4595 MVT::v16i8, &pshufbMask[0], 16));
4597 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4599 // Calculate the shuffle mask for the second input, shuffle it, and
4600 // OR it with the first shuffled input.
4602 for (unsigned i = 0; i != 8; ++i) {
4603 int EltIdx = MaskVals[i] * 2;
4605 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4606 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4609 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4610 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4612 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4613 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4614 DAG.getNode(ISD::BUILD_VECTOR, dl,
4615 MVT::v16i8, &pshufbMask[0], 16));
4616 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4617 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4620 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4621 // and update MaskVals with new element order.
4622 BitVector InOrder(8);
4623 if (BestLoQuad >= 0) {
4624 SmallVector<int, 8> MaskV;
4625 for (int i = 0; i != 4; ++i) {
4626 int idx = MaskVals[i];
4628 MaskV.push_back(-1);
4630 } else if ((idx / 4) == BestLoQuad) {
4631 MaskV.push_back(idx & 3);
4634 MaskV.push_back(-1);
4637 for (unsigned i = 4; i != 8; ++i)
4639 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4642 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4643 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4645 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4649 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4650 // and update MaskVals with the new element order.
4651 if (BestHiQuad >= 0) {
4652 SmallVector<int, 8> MaskV;
4653 for (unsigned i = 0; i != 4; ++i)
4655 for (unsigned i = 4; i != 8; ++i) {
4656 int idx = MaskVals[i];
4658 MaskV.push_back(-1);
4660 } else if ((idx / 4) == BestHiQuad) {
4661 MaskV.push_back((idx & 3) + 4);
4664 MaskV.push_back(-1);
4667 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4670 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4671 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4673 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4677 // In case BestHi & BestLo were both -1, which means each quadword has a word
4678 // from each of the four input quadwords, calculate the InOrder bitvector now
4679 // before falling through to the insert/extract cleanup.
4680 if (BestLoQuad == -1 && BestHiQuad == -1) {
4682 for (int i = 0; i != 8; ++i)
4683 if (MaskVals[i] < 0 || MaskVals[i] == i)
4687 // The other elements are put in the right place using pextrw and pinsrw.
4688 for (unsigned i = 0; i != 8; ++i) {
4691 int EltIdx = MaskVals[i];
4694 SDValue ExtOp = (EltIdx < 8)
4695 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4696 DAG.getIntPtrConstant(EltIdx))
4697 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4698 DAG.getIntPtrConstant(EltIdx - 8));
4699 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4700 DAG.getIntPtrConstant(i));
4705 // v16i8 shuffles - Prefer shuffles in the following order:
4706 // 1. [ssse3] 1 x pshufb
4707 // 2. [ssse3] 2 x pshufb + 1 x por
4708 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4710 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4712 const X86TargetLowering &TLI) {
4713 SDValue V1 = SVOp->getOperand(0);
4714 SDValue V2 = SVOp->getOperand(1);
4715 DebugLoc dl = SVOp->getDebugLoc();
4716 SmallVector<int, 16> MaskVals;
4717 SVOp->getMask(MaskVals);
4719 // If we have SSSE3, case 1 is generated when all result bytes come from
4720 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4721 // present, fall back to case 3.
4722 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4725 for (unsigned i = 0; i < 16; ++i) {
4726 int EltIdx = MaskVals[i];
4735 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4736 if (TLI.getSubtarget()->hasSSSE3()) {
4737 SmallVector<SDValue,16> pshufbMask;
4739 // If all result elements are from one input vector, then only translate
4740 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4742 // Otherwise, we have elements from both input vectors, and must zero out
4743 // elements that come from V2 in the first mask, and V1 in the second mask
4744 // so that we can OR them together.
4745 bool TwoInputs = !(V1Only || V2Only);
4746 for (unsigned i = 0; i != 16; ++i) {
4747 int EltIdx = MaskVals[i];
4748 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4749 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4752 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4754 // If all the elements are from V2, assign it to V1 and return after
4755 // building the first pshufb.
4758 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4759 DAG.getNode(ISD::BUILD_VECTOR, dl,
4760 MVT::v16i8, &pshufbMask[0], 16));
4764 // Calculate the shuffle mask for the second input, shuffle it, and
4765 // OR it with the first shuffled input.
4767 for (unsigned i = 0; i != 16; ++i) {
4768 int EltIdx = MaskVals[i];
4770 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4773 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4775 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4776 DAG.getNode(ISD::BUILD_VECTOR, dl,
4777 MVT::v16i8, &pshufbMask[0], 16));
4778 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4781 // No SSSE3 - Calculate in place words and then fix all out of place words
4782 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4783 // the 16 different words that comprise the two doublequadword input vectors.
4784 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4785 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4786 SDValue NewV = V2Only ? V2 : V1;
4787 for (int i = 0; i != 8; ++i) {
4788 int Elt0 = MaskVals[i*2];
4789 int Elt1 = MaskVals[i*2+1];
4791 // This word of the result is all undef, skip it.
4792 if (Elt0 < 0 && Elt1 < 0)
4795 // This word of the result is already in the correct place, skip it.
4796 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4798 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4801 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4802 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4805 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4806 // using a single extract together, load it and store it.
4807 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4808 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4809 DAG.getIntPtrConstant(Elt1 / 2));
4810 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4811 DAG.getIntPtrConstant(i));
4815 // If Elt1 is defined, extract it from the appropriate source. If the
4816 // source byte is not also odd, shift the extracted word left 8 bits
4817 // otherwise clear the bottom 8 bits if we need to do an or.
4819 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4820 DAG.getIntPtrConstant(Elt1 / 2));
4821 if ((Elt1 & 1) == 0)
4822 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4823 DAG.getConstant(8, TLI.getShiftAmountTy()));
4825 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4826 DAG.getConstant(0xFF00, MVT::i16));
4828 // If Elt0 is defined, extract it from the appropriate source. If the
4829 // source byte is not also even, shift the extracted word right 8 bits. If
4830 // Elt1 was also defined, OR the extracted values together before
4831 // inserting them in the result.
4833 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4834 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4835 if ((Elt0 & 1) != 0)
4836 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4837 DAG.getConstant(8, TLI.getShiftAmountTy()));
4839 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4840 DAG.getConstant(0x00FF, MVT::i16));
4841 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4844 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4845 DAG.getIntPtrConstant(i));
4847 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4850 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4851 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
4852 /// done when every pair / quad of shuffle mask elements point to elements in
4853 /// the right sequence. e.g.
4854 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
4856 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4857 SelectionDAG &DAG, DebugLoc dl) {
4858 EVT VT = SVOp->getValueType(0);
4859 SDValue V1 = SVOp->getOperand(0);
4860 SDValue V2 = SVOp->getOperand(1);
4861 unsigned NumElems = VT.getVectorNumElements();
4862 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4864 switch (VT.getSimpleVT().SimpleTy) {
4865 default: assert(false && "Unexpected!");
4866 case MVT::v4f32: NewVT = MVT::v2f64; break;
4867 case MVT::v4i32: NewVT = MVT::v2i64; break;
4868 case MVT::v8i16: NewVT = MVT::v4i32; break;
4869 case MVT::v16i8: NewVT = MVT::v4i32; break;
4872 int Scale = NumElems / NewWidth;
4873 SmallVector<int, 8> MaskVec;
4874 for (unsigned i = 0; i < NumElems; i += Scale) {
4876 for (int j = 0; j < Scale; ++j) {
4877 int EltIdx = SVOp->getMaskElt(i+j);
4881 StartIdx = EltIdx - (EltIdx % Scale);
4882 if (EltIdx != StartIdx + j)
4886 MaskVec.push_back(-1);
4888 MaskVec.push_back(StartIdx / Scale);
4891 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4892 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4893 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4896 /// getVZextMovL - Return a zero-extending vector move low node.
4898 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4899 SDValue SrcOp, SelectionDAG &DAG,
4900 const X86Subtarget *Subtarget, DebugLoc dl) {
4901 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4902 LoadSDNode *LD = NULL;
4903 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4904 LD = dyn_cast<LoadSDNode>(SrcOp);
4906 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4908 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4909 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4910 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4911 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4912 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4914 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4915 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4916 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4917 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4925 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4926 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4927 DAG.getNode(ISD::BIT_CONVERT, dl,
4931 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4934 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4935 SDValue V1 = SVOp->getOperand(0);
4936 SDValue V2 = SVOp->getOperand(1);
4937 DebugLoc dl = SVOp->getDebugLoc();
4938 EVT VT = SVOp->getValueType(0);
4940 SmallVector<std::pair<int, int>, 8> Locs;
4942 SmallVector<int, 8> Mask1(4U, -1);
4943 SmallVector<int, 8> PermMask;
4944 SVOp->getMask(PermMask);
4948 for (unsigned i = 0; i != 4; ++i) {
4949 int Idx = PermMask[i];
4951 Locs[i] = std::make_pair(-1, -1);
4953 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4955 Locs[i] = std::make_pair(0, NumLo);
4959 Locs[i] = std::make_pair(1, NumHi);
4961 Mask1[2+NumHi] = Idx;
4967 if (NumLo <= 2 && NumHi <= 2) {
4968 // If no more than two elements come from either vector. This can be
4969 // implemented with two shuffles. First shuffle gather the elements.
4970 // The second shuffle, which takes the first shuffle as both of its
4971 // vector operands, put the elements into the right order.
4972 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4974 SmallVector<int, 8> Mask2(4U, -1);
4976 for (unsigned i = 0; i != 4; ++i) {
4977 if (Locs[i].first == -1)
4980 unsigned Idx = (i < 2) ? 0 : 4;
4981 Idx += Locs[i].first * 2 + Locs[i].second;
4986 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4987 } else if (NumLo == 3 || NumHi == 3) {
4988 // Otherwise, we must have three elements from one vector, call it X, and
4989 // one element from the other, call it Y. First, use a shufps to build an
4990 // intermediate vector with the one element from Y and the element from X
4991 // that will be in the same half in the final destination (the indexes don't
4992 // matter). Then, use a shufps to build the final vector, taking the half
4993 // containing the element from Y from the intermediate, and the other half
4996 // Normalize it so the 3 elements come from V1.
4997 CommuteVectorShuffleMask(PermMask, VT);
5001 // Find the element from V2.
5003 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5004 int Val = PermMask[HiIndex];
5011 Mask1[0] = PermMask[HiIndex];
5013 Mask1[2] = PermMask[HiIndex^1];
5015 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5018 Mask1[0] = PermMask[0];
5019 Mask1[1] = PermMask[1];
5020 Mask1[2] = HiIndex & 1 ? 6 : 4;
5021 Mask1[3] = HiIndex & 1 ? 4 : 6;
5022 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5024 Mask1[0] = HiIndex & 1 ? 2 : 0;
5025 Mask1[1] = HiIndex & 1 ? 0 : 2;
5026 Mask1[2] = PermMask[2];
5027 Mask1[3] = PermMask[3];
5032 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5036 // Break it into (shuffle shuffle_hi, shuffle_lo).
5038 SmallVector<int,8> LoMask(4U, -1);
5039 SmallVector<int,8> HiMask(4U, -1);
5041 SmallVector<int,8> *MaskPtr = &LoMask;
5042 unsigned MaskIdx = 0;
5045 for (unsigned i = 0; i != 4; ++i) {
5052 int Idx = PermMask[i];
5054 Locs[i] = std::make_pair(-1, -1);
5055 } else if (Idx < 4) {
5056 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5057 (*MaskPtr)[LoIdx] = Idx;
5060 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5061 (*MaskPtr)[HiIdx] = Idx;
5066 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5067 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5068 SmallVector<int, 8> MaskOps;
5069 for (unsigned i = 0; i != 4; ++i) {
5070 if (Locs[i].first == -1) {
5071 MaskOps.push_back(-1);
5073 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5074 MaskOps.push_back(Idx);
5077 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5080 static bool MayFoldVectorLoad(SDValue V) {
5081 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5082 V = V.getOperand(0);
5083 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5084 V = V.getOperand(0);
5090 // FIXME: the version above should always be used. Since there's
5091 // a bug where several vector shuffles can't be folded because the
5092 // DAG is not updated during lowering and a node claims to have two
5093 // uses while it only has one, use this version, and let isel match
5094 // another instruction if the load really happens to have more than
5095 // one use. Remove this version after this bug get fixed.
5096 // rdar://8434668, PR8156
5097 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5098 if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
5099 V = V.getOperand(0);
5100 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5101 V = V.getOperand(0);
5102 if (ISD::isNormalLoad(V.getNode()))
5107 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5108 /// a vector extract, and if both can be later optimized into a single load.
5109 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5110 /// here because otherwise a target specific shuffle node is going to be
5111 /// emitted for this shuffle, and the optimization not done.
5112 /// FIXME: This is probably not the best approach, but fix the problem
5113 /// until the right path is decided.
5115 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5116 const TargetLowering &TLI) {
5117 EVT VT = V.getValueType();
5118 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5120 // Be sure that the vector shuffle is present in a pattern like this:
5121 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5125 SDNode *N = *V.getNode()->use_begin();
5126 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5129 SDValue EltNo = N->getOperand(1);
5130 if (!isa<ConstantSDNode>(EltNo))
5133 // If the bit convert changed the number of elements, it is unsafe
5134 // to examine the mask.
5135 bool HasShuffleIntoBitcast = false;
5136 if (V.getOpcode() == ISD::BIT_CONVERT) {
5137 EVT SrcVT = V.getOperand(0).getValueType();
5138 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5140 V = V.getOperand(0);
5141 HasShuffleIntoBitcast = true;
5144 // Select the input vector, guarding against out of range extract vector.
5145 unsigned NumElems = VT.getVectorNumElements();
5146 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5147 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5148 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5150 // Skip one more bit_convert if necessary
5151 if (V.getOpcode() == ISD::BIT_CONVERT)
5152 V = V.getOperand(0);
5154 if (ISD::isNormalLoad(V.getNode())) {
5155 // Is the original load suitable?
5156 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5158 // FIXME: avoid the multi-use bug that is preventing lots of
5159 // of foldings to be detected, this is still wrong of course, but
5160 // give the temporary desired behavior, and if it happens that
5161 // the load has real more uses, during isel it will not fold, and
5162 // will generate poor code.
5163 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5166 if (!HasShuffleIntoBitcast)
5169 // If there's a bitcast before the shuffle, check if the load type and
5170 // alignment is valid.
5171 unsigned Align = LN0->getAlignment();
5173 TLI.getTargetData()->getABITypeAlignment(
5174 VT.getTypeForEVT(*DAG.getContext()));
5176 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5184 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5185 EVT VT = Op.getValueType();
5187 // Canonizalize to v2f64.
5188 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, V1);
5189 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5190 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5195 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5197 SDValue V1 = Op.getOperand(0);
5198 SDValue V2 = Op.getOperand(1);
5199 EVT VT = Op.getValueType();
5201 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5203 if (HasSSE2 && VT == MVT::v2f64)
5204 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5207 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5211 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5212 SDValue V1 = Op.getOperand(0);
5213 SDValue V2 = Op.getOperand(1);
5214 EVT VT = Op.getValueType();
5216 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5217 "unsupported shuffle type");
5219 if (V2.getOpcode() == ISD::UNDEF)
5223 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5227 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5228 SDValue V1 = Op.getOperand(0);
5229 SDValue V2 = Op.getOperand(1);
5230 EVT VT = Op.getValueType();
5231 unsigned NumElems = VT.getVectorNumElements();
5233 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5234 // operand of these instructions is only memory, so check if there's a
5235 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5237 bool CanFoldLoad = false;
5239 // Trivial case, when V2 comes from a load.
5240 if (MayFoldVectorLoad(V2))
5243 // When V1 is a load, it can be folded later into a store in isel, example:
5244 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5246 // (MOVLPSmr addr:$src1, VR128:$src2)
5247 // So, recognize this potential and also use MOVLPS or MOVLPD
5248 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5252 if (HasSSE2 && NumElems == 2)
5253 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5256 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5259 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5260 // movl and movlp will both match v2i64, but v2i64 is never matched by
5261 // movl earlier because we make it strict to avoid messing with the movlp load
5262 // folding logic (see the code above getMOVLP call). Match it here then,
5263 // this is horrible, but will stay like this until we move all shuffle
5264 // matching to x86 specific nodes. Note that for the 1st condition all
5265 // types are matched with movsd.
5266 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5267 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5269 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5272 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5274 // Invert the operand order and use SHUFPS to match it.
5275 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5276 X86::getShuffleSHUFImmediate(SVOp), DAG);
5279 static inline unsigned getUNPCKLOpcode(EVT VT) {
5280 switch(VT.getSimpleVT().SimpleTy) {
5281 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5282 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5283 case MVT::v4f32: return X86ISD::UNPCKLPS;
5284 case MVT::v2f64: return X86ISD::UNPCKLPD;
5285 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5286 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5288 llvm_unreachable("Unknow type for unpckl");
5293 static inline unsigned getUNPCKHOpcode(EVT VT) {
5294 switch(VT.getSimpleVT().SimpleTy) {
5295 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5296 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5297 case MVT::v4f32: return X86ISD::UNPCKHPS;
5298 case MVT::v2f64: return X86ISD::UNPCKHPD;
5299 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5300 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5302 llvm_unreachable("Unknow type for unpckh");
5308 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5309 const TargetLowering &TLI,
5310 const X86Subtarget *Subtarget) {
5311 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5312 EVT VT = Op.getValueType();
5313 DebugLoc dl = Op.getDebugLoc();
5314 SDValue V1 = Op.getOperand(0);
5315 SDValue V2 = Op.getOperand(1);
5317 if (isZeroShuffle(SVOp))
5318 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5320 // Handle splat operations
5321 if (SVOp->isSplat()) {
5322 // Special case, this is the only place now where it's
5323 // allowed to return a vector_shuffle operation without
5324 // using a target specific node, because *hopefully* it
5325 // will be optimized away by the dag combiner.
5326 if (VT.getVectorNumElements() <= 4 &&
5327 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5330 // Handle splats by matching through known masks
5331 if (VT.getVectorNumElements() <= 4)
5334 // Canonicalize all of the remaining to v4f32.
5335 return PromoteSplat(SVOp, DAG);
5338 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5340 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5341 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5342 if (NewOp.getNode())
5343 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, NewOp);
5344 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5345 // FIXME: Figure out a cleaner way to do this.
5346 // Try to make use of movq to zero out the top part.
5347 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5348 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5349 if (NewOp.getNode()) {
5350 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5351 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5352 DAG, Subtarget, dl);
5354 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5355 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5356 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5357 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5358 DAG, Subtarget, dl);
5365 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5366 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5367 SDValue V1 = Op.getOperand(0);
5368 SDValue V2 = Op.getOperand(1);
5369 EVT VT = Op.getValueType();
5370 DebugLoc dl = Op.getDebugLoc();
5371 unsigned NumElems = VT.getVectorNumElements();
5372 bool isMMX = VT.getSizeInBits() == 64;
5373 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5374 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5375 bool V1IsSplat = false;
5376 bool V2IsSplat = false;
5377 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5378 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5379 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5380 MachineFunction &MF = DAG.getMachineFunction();
5381 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5383 // Shuffle operations on MMX not supported.
5387 // Vector shuffle lowering takes 3 steps:
5389 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5390 // narrowing and commutation of operands should be handled.
5391 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5393 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5394 // so the shuffle can be broken into other shuffles and the legalizer can
5395 // try the lowering again.
5397 // The general ideia is that no vector_shuffle operation should be left to
5398 // be matched during isel, all of them must be converted to a target specific
5401 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5402 // narrowing and commutation of operands should be handled. The actual code
5403 // doesn't include all of those, work in progress...
5404 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5405 if (NewOp.getNode())
5408 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5409 // unpckh_undef). Only use pshufd if speed is more important than size.
5410 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5411 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5412 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5413 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5414 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5415 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5417 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5418 RelaxedMayFoldVectorLoad(V1))
5419 return getMOVDDup(Op, dl, V1, DAG);
5421 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5422 return getMOVHighToLow(Op, dl, DAG);
5424 // Use to match splats
5425 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5426 (VT == MVT::v2f64 || VT == MVT::v2i64))
5427 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5429 if (X86::isPSHUFDMask(SVOp)) {
5430 // The actual implementation will match the mask in the if above and then
5431 // during isel it can match several different instructions, not only pshufd
5432 // as its name says, sad but true, emulate the behavior for now...
5433 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5434 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5436 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5438 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5439 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5441 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5442 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5445 if (VT == MVT::v4f32)
5446 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5450 // Check if this can be converted into a logical shift.
5451 bool isLeft = false;
5454 bool isShift = getSubtarget()->hasSSE2() &&
5455 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5456 if (isShift && ShVal.hasOneUse()) {
5457 // If the shifted value has multiple uses, it may be cheaper to use
5458 // v_set0 + movlhps or movhlps, etc.
5459 EVT EltVT = VT.getVectorElementType();
5460 ShAmt *= EltVT.getSizeInBits();
5461 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5464 if (X86::isMOVLMask(SVOp)) {
5467 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5468 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5469 if (!X86::isMOVLPMask(SVOp)) {
5470 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5471 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5473 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5474 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5478 // FIXME: fold these into legal mask.
5479 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5480 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5482 if (X86::isMOVHLPSMask(SVOp))
5483 return getMOVHighToLow(Op, dl, DAG);
5485 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5486 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5488 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5489 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5491 if (X86::isMOVLPMask(SVOp))
5492 return getMOVLP(Op, dl, DAG, HasSSE2);
5494 if (ShouldXformToMOVHLPS(SVOp) ||
5495 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5496 return CommuteVectorShuffle(SVOp, DAG);
5499 // No better options. Use a vshl / vsrl.
5500 EVT EltVT = VT.getVectorElementType();
5501 ShAmt *= EltVT.getSizeInBits();
5502 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5505 bool Commuted = false;
5506 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5507 // 1,1,1,1 -> v8i16 though.
5508 V1IsSplat = isSplatVector(V1.getNode());
5509 V2IsSplat = isSplatVector(V2.getNode());
5511 // Canonicalize the splat or undef, if present, to be on the RHS.
5512 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5513 Op = CommuteVectorShuffle(SVOp, DAG);
5514 SVOp = cast<ShuffleVectorSDNode>(Op);
5515 V1 = SVOp->getOperand(0);
5516 V2 = SVOp->getOperand(1);
5517 std::swap(V1IsSplat, V2IsSplat);
5518 std::swap(V1IsUndef, V2IsUndef);
5522 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5523 // Shuffling low element of v1 into undef, just return v1.
5526 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5527 // the instruction selector will not match, so get a canonical MOVL with
5528 // swapped operands to undo the commute.
5529 return getMOVL(DAG, dl, VT, V2, V1);
5532 if (X86::isUNPCKLMask(SVOp))
5533 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5535 if (X86::isUNPCKHMask(SVOp))
5536 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5539 // Normalize mask so all entries that point to V2 points to its first
5540 // element then try to match unpck{h|l} again. If match, return a
5541 // new vector_shuffle with the corrected mask.
5542 SDValue NewMask = NormalizeMask(SVOp, DAG);
5543 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5544 if (NSVOp != SVOp) {
5545 if (X86::isUNPCKLMask(NSVOp, true)) {
5547 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5554 // Commute is back and try unpck* again.
5555 // FIXME: this seems wrong.
5556 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5557 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5559 if (X86::isUNPCKLMask(NewSVOp))
5560 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5562 if (X86::isUNPCKHMask(NewSVOp))
5563 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5566 // Normalize the node to match x86 shuffle ops if needed
5567 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5568 return CommuteVectorShuffle(SVOp, DAG);
5570 // The checks below are all present in isShuffleMaskLegal, but they are
5571 // inlined here right now to enable us to directly emit target specific
5572 // nodes, and remove one by one until they don't return Op anymore.
5573 SmallVector<int, 16> M;
5576 if (isPALIGNRMask(M, VT, HasSSSE3))
5577 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5578 X86::getShufflePALIGNRImmediate(SVOp),
5581 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5582 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5583 if (VT == MVT::v2f64)
5584 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5585 if (VT == MVT::v2i64)
5586 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5589 if (isPSHUFHWMask(M, VT))
5590 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5591 X86::getShufflePSHUFHWImmediate(SVOp),
5594 if (isPSHUFLWMask(M, VT))
5595 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5596 X86::getShufflePSHUFLWImmediate(SVOp),
5599 if (isSHUFPMask(M, VT)) {
5600 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5601 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5602 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5604 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5605 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5609 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5610 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5611 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5612 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5613 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5614 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5616 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5617 if (VT == MVT::v8i16) {
5618 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5619 if (NewOp.getNode())
5623 if (VT == MVT::v16i8) {
5624 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5625 if (NewOp.getNode())
5629 // Handle all 4 wide cases with a number of shuffles.
5631 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5637 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5638 SelectionDAG &DAG) const {
5639 EVT VT = Op.getValueType();
5640 DebugLoc dl = Op.getDebugLoc();
5641 if (VT.getSizeInBits() == 8) {
5642 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5643 Op.getOperand(0), Op.getOperand(1));
5644 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5645 DAG.getValueType(VT));
5646 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5647 } else if (VT.getSizeInBits() == 16) {
5648 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5649 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5651 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5652 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5653 DAG.getNode(ISD::BIT_CONVERT, dl,
5657 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5658 Op.getOperand(0), Op.getOperand(1));
5659 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5660 DAG.getValueType(VT));
5661 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5662 } else if (VT == MVT::f32) {
5663 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5664 // the result back to FR32 register. It's only worth matching if the
5665 // result has a single use which is a store or a bitcast to i32. And in
5666 // the case of a store, it's not worth it if the index is a constant 0,
5667 // because a MOVSSmr can be used instead, which is smaller and faster.
5668 if (!Op.hasOneUse())
5670 SDNode *User = *Op.getNode()->use_begin();
5671 if ((User->getOpcode() != ISD::STORE ||
5672 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5673 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5674 (User->getOpcode() != ISD::BIT_CONVERT ||
5675 User->getValueType(0) != MVT::i32))
5677 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5678 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5681 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5682 } else if (VT == MVT::i32) {
5683 // ExtractPS works with constant index.
5684 if (isa<ConstantSDNode>(Op.getOperand(1)))
5692 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5693 SelectionDAG &DAG) const {
5694 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5697 if (Subtarget->hasSSE41()) {
5698 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5703 EVT VT = Op.getValueType();
5704 DebugLoc dl = Op.getDebugLoc();
5705 // TODO: handle v16i8.
5706 if (VT.getSizeInBits() == 16) {
5707 SDValue Vec = Op.getOperand(0);
5708 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5710 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5711 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5712 DAG.getNode(ISD::BIT_CONVERT, dl,
5715 // Transform it so it match pextrw which produces a 32-bit result.
5716 EVT EltVT = MVT::i32;
5717 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5718 Op.getOperand(0), Op.getOperand(1));
5719 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5720 DAG.getValueType(VT));
5721 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5722 } else if (VT.getSizeInBits() == 32) {
5723 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5727 // SHUFPS the element to the lowest double word, then movss.
5728 int Mask[4] = { Idx, -1, -1, -1 };
5729 EVT VVT = Op.getOperand(0).getValueType();
5730 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5731 DAG.getUNDEF(VVT), Mask);
5732 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5733 DAG.getIntPtrConstant(0));
5734 } else if (VT.getSizeInBits() == 64) {
5735 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5736 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5737 // to match extract_elt for f64.
5738 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5742 // UNPCKHPD the element to the lowest double word, then movsd.
5743 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5744 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5745 int Mask[2] = { 1, -1 };
5746 EVT VVT = Op.getOperand(0).getValueType();
5747 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5748 DAG.getUNDEF(VVT), Mask);
5749 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5750 DAG.getIntPtrConstant(0));
5757 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5758 SelectionDAG &DAG) const {
5759 EVT VT = Op.getValueType();
5760 EVT EltVT = VT.getVectorElementType();
5761 DebugLoc dl = Op.getDebugLoc();
5763 SDValue N0 = Op.getOperand(0);
5764 SDValue N1 = Op.getOperand(1);
5765 SDValue N2 = Op.getOperand(2);
5767 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5768 isa<ConstantSDNode>(N2)) {
5770 if (VT == MVT::v8i16)
5771 Opc = X86ISD::PINSRW;
5772 else if (VT == MVT::v16i8)
5773 Opc = X86ISD::PINSRB;
5775 Opc = X86ISD::PINSRB;
5777 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5779 if (N1.getValueType() != MVT::i32)
5780 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5781 if (N2.getValueType() != MVT::i32)
5782 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5783 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5784 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5785 // Bits [7:6] of the constant are the source select. This will always be
5786 // zero here. The DAG Combiner may combine an extract_elt index into these
5787 // bits. For example (insert (extract, 3), 2) could be matched by putting
5788 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5789 // Bits [5:4] of the constant are the destination select. This is the
5790 // value of the incoming immediate.
5791 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5792 // combine either bitwise AND or insert of float 0.0 to set these bits.
5793 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5794 // Create this as a scalar to vector..
5795 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5796 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5797 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5798 // PINSR* works with constant index.
5805 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5806 EVT VT = Op.getValueType();
5807 EVT EltVT = VT.getVectorElementType();
5809 if (Subtarget->hasSSE41())
5810 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5812 if (EltVT == MVT::i8)
5815 DebugLoc dl = Op.getDebugLoc();
5816 SDValue N0 = Op.getOperand(0);
5817 SDValue N1 = Op.getOperand(1);
5818 SDValue N2 = Op.getOperand(2);
5820 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5821 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5822 // as its second argument.
5823 if (N1.getValueType() != MVT::i32)
5824 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5825 if (N2.getValueType() != MVT::i32)
5826 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5827 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
5833 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5834 DebugLoc dl = Op.getDebugLoc();
5836 if (Op.getValueType() == MVT::v1i64 &&
5837 Op.getOperand(0).getValueType() == MVT::i64)
5838 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5840 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5841 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
5842 "Expected an SSE type!");
5843 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5844 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
5847 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5848 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5849 // one of the above mentioned nodes. It has to be wrapped because otherwise
5850 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5851 // be used to form addressing mode. These wrapped nodes will be selected
5854 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5855 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5857 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5859 unsigned char OpFlag = 0;
5860 unsigned WrapperKind = X86ISD::Wrapper;
5861 CodeModel::Model M = getTargetMachine().getCodeModel();
5863 if (Subtarget->isPICStyleRIPRel() &&
5864 (M == CodeModel::Small || M == CodeModel::Kernel))
5865 WrapperKind = X86ISD::WrapperRIP;
5866 else if (Subtarget->isPICStyleGOT())
5867 OpFlag = X86II::MO_GOTOFF;
5868 else if (Subtarget->isPICStyleStubPIC())
5869 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5871 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5873 CP->getOffset(), OpFlag);
5874 DebugLoc DL = CP->getDebugLoc();
5875 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5876 // With PIC, the address is actually $g + Offset.
5878 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5879 DAG.getNode(X86ISD::GlobalBaseReg,
5880 DebugLoc(), getPointerTy()),
5887 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5888 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5890 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5892 unsigned char OpFlag = 0;
5893 unsigned WrapperKind = X86ISD::Wrapper;
5894 CodeModel::Model M = getTargetMachine().getCodeModel();
5896 if (Subtarget->isPICStyleRIPRel() &&
5897 (M == CodeModel::Small || M == CodeModel::Kernel))
5898 WrapperKind = X86ISD::WrapperRIP;
5899 else if (Subtarget->isPICStyleGOT())
5900 OpFlag = X86II::MO_GOTOFF;
5901 else if (Subtarget->isPICStyleStubPIC())
5902 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5904 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5906 DebugLoc DL = JT->getDebugLoc();
5907 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5909 // With PIC, the address is actually $g + Offset.
5911 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5912 DAG.getNode(X86ISD::GlobalBaseReg,
5913 DebugLoc(), getPointerTy()),
5921 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5922 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5924 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5926 unsigned char OpFlag = 0;
5927 unsigned WrapperKind = X86ISD::Wrapper;
5928 CodeModel::Model M = getTargetMachine().getCodeModel();
5930 if (Subtarget->isPICStyleRIPRel() &&
5931 (M == CodeModel::Small || M == CodeModel::Kernel))
5932 WrapperKind = X86ISD::WrapperRIP;
5933 else if (Subtarget->isPICStyleGOT())
5934 OpFlag = X86II::MO_GOTOFF;
5935 else if (Subtarget->isPICStyleStubPIC())
5936 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5938 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5940 DebugLoc DL = Op.getDebugLoc();
5941 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5944 // With PIC, the address is actually $g + Offset.
5945 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5946 !Subtarget->is64Bit()) {
5947 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5948 DAG.getNode(X86ISD::GlobalBaseReg,
5949 DebugLoc(), getPointerTy()),
5957 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5958 // Create the TargetBlockAddressAddress node.
5959 unsigned char OpFlags =
5960 Subtarget->ClassifyBlockAddressReference();
5961 CodeModel::Model M = getTargetMachine().getCodeModel();
5962 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5963 DebugLoc dl = Op.getDebugLoc();
5964 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5965 /*isTarget=*/true, OpFlags);
5967 if (Subtarget->isPICStyleRIPRel() &&
5968 (M == CodeModel::Small || M == CodeModel::Kernel))
5969 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5971 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5973 // With PIC, the address is actually $g + Offset.
5974 if (isGlobalRelativeToPICBase(OpFlags)) {
5975 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5976 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5984 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5986 SelectionDAG &DAG) const {
5987 // Create the TargetGlobalAddress node, folding in the constant
5988 // offset if it is legal.
5989 unsigned char OpFlags =
5990 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5991 CodeModel::Model M = getTargetMachine().getCodeModel();
5993 if (OpFlags == X86II::MO_NO_FLAG &&
5994 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5995 // A direct static reference to a global.
5996 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5999 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6002 if (Subtarget->isPICStyleRIPRel() &&
6003 (M == CodeModel::Small || M == CodeModel::Kernel))
6004 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6006 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6008 // With PIC, the address is actually $g + Offset.
6009 if (isGlobalRelativeToPICBase(OpFlags)) {
6010 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6011 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6015 // For globals that require a load from a stub to get the address, emit the
6017 if (isGlobalStubReference(OpFlags))
6018 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6019 MachinePointerInfo::getGOT(), false, false, 0);
6021 // If there was a non-zero offset that we didn't fold, create an explicit
6024 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6025 DAG.getConstant(Offset, getPointerTy()));
6031 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6032 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6033 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6034 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6038 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6039 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6040 unsigned char OperandFlags) {
6041 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6042 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6043 DebugLoc dl = GA->getDebugLoc();
6044 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6045 GA->getValueType(0),
6049 SDValue Ops[] = { Chain, TGA, *InFlag };
6050 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6052 SDValue Ops[] = { Chain, TGA };
6053 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6056 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6057 MFI->setAdjustsStack(true);
6059 SDValue Flag = Chain.getValue(1);
6060 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6063 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6065 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6068 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6069 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6070 DAG.getNode(X86ISD::GlobalBaseReg,
6071 DebugLoc(), PtrVT), InFlag);
6072 InFlag = Chain.getValue(1);
6074 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6077 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6079 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6081 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6082 X86::RAX, X86II::MO_TLSGD);
6085 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6086 // "local exec" model.
6087 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6088 const EVT PtrVT, TLSModel::Model model,
6090 DebugLoc dl = GA->getDebugLoc();
6092 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6093 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6094 is64Bit ? 257 : 256));
6096 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6097 DAG.getIntPtrConstant(0),
6098 MachinePointerInfo(Ptr), false, false, 0);
6100 unsigned char OperandFlags = 0;
6101 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6103 unsigned WrapperKind = X86ISD::Wrapper;
6104 if (model == TLSModel::LocalExec) {
6105 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6106 } else if (is64Bit) {
6107 assert(model == TLSModel::InitialExec);
6108 OperandFlags = X86II::MO_GOTTPOFF;
6109 WrapperKind = X86ISD::WrapperRIP;
6111 assert(model == TLSModel::InitialExec);
6112 OperandFlags = X86II::MO_INDNTPOFF;
6115 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6117 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6118 GA->getValueType(0),
6119 GA->getOffset(), OperandFlags);
6120 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6122 if (model == TLSModel::InitialExec)
6123 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6124 MachinePointerInfo::getGOT(), false, false, 0);
6126 // The address of the thread local variable is the add of the thread
6127 // pointer with the offset of the variable.
6128 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6132 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6134 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6135 const GlobalValue *GV = GA->getGlobal();
6137 if (Subtarget->isTargetELF()) {
6138 // TODO: implement the "local dynamic" model
6139 // TODO: implement the "initial exec"model for pic executables
6141 // If GV is an alias then use the aliasee for determining
6142 // thread-localness.
6143 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6144 GV = GA->resolveAliasedGlobal(false);
6146 TLSModel::Model model
6147 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6150 case TLSModel::GeneralDynamic:
6151 case TLSModel::LocalDynamic: // not implemented
6152 if (Subtarget->is64Bit())
6153 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6154 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6156 case TLSModel::InitialExec:
6157 case TLSModel::LocalExec:
6158 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6159 Subtarget->is64Bit());
6161 } else if (Subtarget->isTargetDarwin()) {
6162 // Darwin only has one model of TLS. Lower to that.
6163 unsigned char OpFlag = 0;
6164 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6165 X86ISD::WrapperRIP : X86ISD::Wrapper;
6167 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6169 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6170 !Subtarget->is64Bit();
6172 OpFlag = X86II::MO_TLVP_PIC_BASE;
6174 OpFlag = X86II::MO_TLVP;
6175 DebugLoc DL = Op.getDebugLoc();
6176 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6178 GA->getOffset(), OpFlag);
6179 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6181 // With PIC32, the address is actually $g + Offset.
6183 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6184 DAG.getNode(X86ISD::GlobalBaseReg,
6185 DebugLoc(), getPointerTy()),
6188 // Lowering the machine isd will make sure everything is in the right
6190 SDValue Args[] = { Offset };
6191 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
6193 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6194 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6195 MFI->setAdjustsStack(true);
6197 // And our return value (tls address) is in the standard call return value
6199 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6200 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6204 "TLS not implemented for this target.");
6206 llvm_unreachable("Unreachable");
6211 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6212 /// take a 2 x i32 value to shift plus a shift amount.
6213 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6214 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6215 EVT VT = Op.getValueType();
6216 unsigned VTBits = VT.getSizeInBits();
6217 DebugLoc dl = Op.getDebugLoc();
6218 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6219 SDValue ShOpLo = Op.getOperand(0);
6220 SDValue ShOpHi = Op.getOperand(1);
6221 SDValue ShAmt = Op.getOperand(2);
6222 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6223 DAG.getConstant(VTBits - 1, MVT::i8))
6224 : DAG.getConstant(0, VT);
6227 if (Op.getOpcode() == ISD::SHL_PARTS) {
6228 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6229 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6231 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6232 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6235 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6236 DAG.getConstant(VTBits, MVT::i8));
6237 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6238 AndNode, DAG.getConstant(0, MVT::i8));
6241 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6242 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6243 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6245 if (Op.getOpcode() == ISD::SHL_PARTS) {
6246 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6247 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6249 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6250 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6253 SDValue Ops[2] = { Lo, Hi };
6254 return DAG.getMergeValues(Ops, 2, dl);
6257 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6258 SelectionDAG &DAG) const {
6259 EVT SrcVT = Op.getOperand(0).getValueType();
6261 if (SrcVT.isVector())
6264 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6265 "Unknown SINT_TO_FP to lower!");
6267 // These are really Legal; return the operand so the caller accepts it as
6269 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6271 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6272 Subtarget->is64Bit()) {
6276 DebugLoc dl = Op.getDebugLoc();
6277 unsigned Size = SrcVT.getSizeInBits()/8;
6278 MachineFunction &MF = DAG.getMachineFunction();
6279 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6280 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6281 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6283 MachinePointerInfo::getFixedStack(SSFI),
6285 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6288 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6290 SelectionDAG &DAG) const {
6292 DebugLoc DL = Op.getDebugLoc();
6294 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6296 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6298 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6300 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6302 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6303 MachineMemOperand *MMO =
6304 DAG.getMachineFunction()
6305 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6306 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6308 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6309 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6311 Tys, Ops, array_lengthof(Ops),
6315 Chain = Result.getValue(1);
6316 SDValue InFlag = Result.getValue(2);
6318 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6319 // shouldn't be necessary except that RFP cannot be live across
6320 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6321 MachineFunction &MF = DAG.getMachineFunction();
6322 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6323 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6324 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6325 Tys = DAG.getVTList(MVT::Other);
6327 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6329 MachineMemOperand *MMO =
6330 DAG.getMachineFunction()
6331 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6332 MachineMemOperand::MOStore, SSFISize, SSFISize);
6334 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6335 Ops, array_lengthof(Ops),
6336 Op.getValueType(), MMO);
6337 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6338 MachinePointerInfo::getFixedStack(SSFI),
6345 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6346 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6347 SelectionDAG &DAG) const {
6348 // This algorithm is not obvious. Here it is in C code, more or less:
6350 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6351 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6352 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6354 // Copy ints to xmm registers.
6355 __m128i xh = _mm_cvtsi32_si128( hi );
6356 __m128i xl = _mm_cvtsi32_si128( lo );
6358 // Combine into low half of a single xmm register.
6359 __m128i x = _mm_unpacklo_epi32( xh, xl );
6363 // Merge in appropriate exponents to give the integer bits the right
6365 x = _mm_unpacklo_epi32( x, exp );
6367 // Subtract away the biases to deal with the IEEE-754 double precision
6369 d = _mm_sub_pd( (__m128d) x, bias );
6371 // All conversions up to here are exact. The correctly rounded result is
6372 // calculated using the current rounding mode using the following
6374 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6375 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6376 // store doesn't really need to be here (except
6377 // maybe to zero the other double)
6382 DebugLoc dl = Op.getDebugLoc();
6383 LLVMContext *Context = DAG.getContext();
6385 // Build some magic constants.
6386 std::vector<Constant*> CV0;
6387 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6388 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6389 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6390 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6391 Constant *C0 = ConstantVector::get(CV0);
6392 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6394 std::vector<Constant*> CV1;
6396 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6398 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6399 Constant *C1 = ConstantVector::get(CV1);
6400 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6402 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6403 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6405 DAG.getIntPtrConstant(1)));
6406 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6407 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6409 DAG.getIntPtrConstant(0)));
6410 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6411 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6412 MachinePointerInfo::getConstantPool(),
6414 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6415 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6416 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6417 MachinePointerInfo::getConstantPool(),
6419 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6421 // Add the halves; easiest way is to swap them into another reg first.
6422 int ShufMask[2] = { 1, -1 };
6423 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6424 DAG.getUNDEF(MVT::v2f64), ShufMask);
6425 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6426 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6427 DAG.getIntPtrConstant(0));
6430 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6431 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6432 SelectionDAG &DAG) const {
6433 DebugLoc dl = Op.getDebugLoc();
6434 // FP constant to bias correct the final result.
6435 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6438 // Load the 32-bit value into an XMM register.
6439 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6440 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6442 DAG.getIntPtrConstant(0)));
6444 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6445 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6446 DAG.getIntPtrConstant(0));
6448 // Or the load with the bias.
6449 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6450 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6451 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6453 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6454 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6455 MVT::v2f64, Bias)));
6456 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6457 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6458 DAG.getIntPtrConstant(0));
6460 // Subtract the bias.
6461 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6463 // Handle final rounding.
6464 EVT DestVT = Op.getValueType();
6466 if (DestVT.bitsLT(MVT::f64)) {
6467 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6468 DAG.getIntPtrConstant(0));
6469 } else if (DestVT.bitsGT(MVT::f64)) {
6470 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6473 // Handle final rounding.
6477 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6478 SelectionDAG &DAG) const {
6479 SDValue N0 = Op.getOperand(0);
6480 DebugLoc dl = Op.getDebugLoc();
6482 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6483 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6484 // the optimization here.
6485 if (DAG.SignBitIsZero(N0))
6486 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6488 EVT SrcVT = N0.getValueType();
6489 EVT DstVT = Op.getValueType();
6490 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6491 return LowerUINT_TO_FP_i64(Op, DAG);
6492 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6493 return LowerUINT_TO_FP_i32(Op, DAG);
6495 // Make a 64-bit buffer, and use it to build an FILD.
6496 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6497 if (SrcVT == MVT::i32) {
6498 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6499 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6500 getPointerTy(), StackSlot, WordOff);
6501 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6502 StackSlot, MachinePointerInfo(),
6504 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6505 OffsetSlot, MachinePointerInfo(),
6507 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6511 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6512 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6513 StackSlot, MachinePointerInfo(),
6515 // For i64 source, we need to add the appropriate power of 2 if the input
6516 // was negative. This is the same as the optimization in
6517 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6518 // we must be careful to do the computation in x87 extended precision, not
6519 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6520 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6521 MachineMemOperand *MMO =
6522 DAG.getMachineFunction()
6523 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6524 MachineMemOperand::MOLoad, 8, 8);
6526 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6527 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6528 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
6531 APInt FF(32, 0x5F800000ULL);
6533 // Check whether the sign bit is set.
6534 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6535 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6538 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6539 SDValue FudgePtr = DAG.getConstantPool(
6540 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6543 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6544 SDValue Zero = DAG.getIntPtrConstant(0);
6545 SDValue Four = DAG.getIntPtrConstant(4);
6546 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6548 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6550 // Load the value out, extending it from f32 to f80.
6551 // FIXME: Avoid the extend by constructing the right constant pool?
6552 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6553 FudgePtr, MachinePointerInfo::getConstantPool(),
6554 MVT::f32, false, false, 4);
6555 // Extend everything to 80 bits to force it to be done on x87.
6556 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6557 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6560 std::pair<SDValue,SDValue> X86TargetLowering::
6561 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6562 DebugLoc DL = Op.getDebugLoc();
6564 EVT DstTy = Op.getValueType();
6567 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6571 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6572 DstTy.getSimpleVT() >= MVT::i16 &&
6573 "Unknown FP_TO_SINT to lower!");
6575 // These are really Legal.
6576 if (DstTy == MVT::i32 &&
6577 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6578 return std::make_pair(SDValue(), SDValue());
6579 if (Subtarget->is64Bit() &&
6580 DstTy == MVT::i64 &&
6581 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6582 return std::make_pair(SDValue(), SDValue());
6584 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6586 MachineFunction &MF = DAG.getMachineFunction();
6587 unsigned MemSize = DstTy.getSizeInBits()/8;
6588 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6589 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6594 switch (DstTy.getSimpleVT().SimpleTy) {
6595 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6596 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6597 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6598 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6601 SDValue Chain = DAG.getEntryNode();
6602 SDValue Value = Op.getOperand(0);
6603 EVT TheVT = Op.getOperand(0).getValueType();
6604 if (isScalarFPTypeInSSEReg(TheVT)) {
6605 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6606 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
6607 MachinePointerInfo::getFixedStack(SSFI),
6609 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6611 Chain, StackSlot, DAG.getValueType(TheVT)
6614 MachineMemOperand *MMO =
6615 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6616 MachineMemOperand::MOLoad, MemSize, MemSize);
6617 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
6619 Chain = Value.getValue(1);
6620 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6621 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6624 MachineMemOperand *MMO =
6625 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6626 MachineMemOperand::MOStore, MemSize, MemSize);
6628 // Build the FP_TO_INT*_IN_MEM
6629 SDValue Ops[] = { Chain, Value, StackSlot };
6630 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
6631 Ops, 3, DstTy, MMO);
6633 return std::make_pair(FIST, StackSlot);
6636 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6637 SelectionDAG &DAG) const {
6638 if (Op.getValueType().isVector())
6641 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6642 SDValue FIST = Vals.first, StackSlot = Vals.second;
6643 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6644 if (FIST.getNode() == 0) return Op;
6647 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6648 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6651 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6652 SelectionDAG &DAG) const {
6653 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6654 SDValue FIST = Vals.first, StackSlot = Vals.second;
6655 assert(FIST.getNode() && "Unexpected failure");
6658 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6659 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6662 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6663 SelectionDAG &DAG) const {
6664 LLVMContext *Context = DAG.getContext();
6665 DebugLoc dl = Op.getDebugLoc();
6666 EVT VT = Op.getValueType();
6669 EltVT = VT.getVectorElementType();
6670 std::vector<Constant*> CV;
6671 if (EltVT == MVT::f64) {
6672 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6676 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6682 Constant *C = ConstantVector::get(CV);
6683 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6684 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6685 MachinePointerInfo::getConstantPool(),
6687 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6690 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6691 LLVMContext *Context = DAG.getContext();
6692 DebugLoc dl = Op.getDebugLoc();
6693 EVT VT = Op.getValueType();
6696 EltVT = VT.getVectorElementType();
6697 std::vector<Constant*> CV;
6698 if (EltVT == MVT::f64) {
6699 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6703 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6709 Constant *C = ConstantVector::get(CV);
6710 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6711 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6712 MachinePointerInfo::getConstantPool(),
6714 if (VT.isVector()) {
6715 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6716 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6717 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6719 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6721 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6725 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6726 LLVMContext *Context = DAG.getContext();
6727 SDValue Op0 = Op.getOperand(0);
6728 SDValue Op1 = Op.getOperand(1);
6729 DebugLoc dl = Op.getDebugLoc();
6730 EVT VT = Op.getValueType();
6731 EVT SrcVT = Op1.getValueType();
6733 // If second operand is smaller, extend it first.
6734 if (SrcVT.bitsLT(VT)) {
6735 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6738 // And if it is bigger, shrink it first.
6739 if (SrcVT.bitsGT(VT)) {
6740 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6744 // At this point the operands and the result should have the same
6745 // type, and that won't be f80 since that is not custom lowered.
6747 // First get the sign bit of second operand.
6748 std::vector<Constant*> CV;
6749 if (SrcVT == MVT::f64) {
6750 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6751 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6753 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6754 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6755 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6756 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6758 Constant *C = ConstantVector::get(CV);
6759 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6760 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6761 MachinePointerInfo::getConstantPool(),
6763 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6765 // Shift sign bit right or left if the two operands have different types.
6766 if (SrcVT.bitsGT(VT)) {
6767 // Op0 is MVT::f32, Op1 is MVT::f64.
6768 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6769 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6770 DAG.getConstant(32, MVT::i32));
6771 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6772 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6773 DAG.getIntPtrConstant(0));
6776 // Clear first operand sign bit.
6778 if (VT == MVT::f64) {
6779 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6780 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6782 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6783 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6784 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6785 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6787 C = ConstantVector::get(CV);
6788 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6789 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6790 MachinePointerInfo::getConstantPool(),
6792 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6794 // Or the value with the sign bit.
6795 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6798 /// Emit nodes that will be selected as "test Op0,Op0", or something
6800 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6801 SelectionDAG &DAG) const {
6802 DebugLoc dl = Op.getDebugLoc();
6804 // CF and OF aren't always set the way we want. Determine which
6805 // of these we need.
6806 bool NeedCF = false;
6807 bool NeedOF = false;
6810 case X86::COND_A: case X86::COND_AE:
6811 case X86::COND_B: case X86::COND_BE:
6814 case X86::COND_G: case X86::COND_GE:
6815 case X86::COND_L: case X86::COND_LE:
6816 case X86::COND_O: case X86::COND_NO:
6821 // See if we can use the EFLAGS value from the operand instead of
6822 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6823 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6824 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6825 // Emit a CMP with 0, which is the TEST pattern.
6826 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6827 DAG.getConstant(0, Op.getValueType()));
6829 unsigned Opcode = 0;
6830 unsigned NumOperands = 0;
6831 switch (Op.getNode()->getOpcode()) {
6833 // Due to an isel shortcoming, be conservative if this add is likely to be
6834 // selected as part of a load-modify-store instruction. When the root node
6835 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6836 // uses of other nodes in the match, such as the ADD in this case. This
6837 // leads to the ADD being left around and reselected, with the result being
6838 // two adds in the output. Alas, even if none our users are stores, that
6839 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6840 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6841 // climbing the DAG back to the root, and it doesn't seem to be worth the
6843 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6844 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6845 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6848 if (ConstantSDNode *C =
6849 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6850 // An add of one will be selected as an INC.
6851 if (C->getAPIntValue() == 1) {
6852 Opcode = X86ISD::INC;
6857 // An add of negative one (subtract of one) will be selected as a DEC.
6858 if (C->getAPIntValue().isAllOnesValue()) {
6859 Opcode = X86ISD::DEC;
6865 // Otherwise use a regular EFLAGS-setting add.
6866 Opcode = X86ISD::ADD;
6870 // If the primary and result isn't used, don't bother using X86ISD::AND,
6871 // because a TEST instruction will be better.
6872 bool NonFlagUse = false;
6873 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6874 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6876 unsigned UOpNo = UI.getOperandNo();
6877 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6878 // Look pass truncate.
6879 UOpNo = User->use_begin().getOperandNo();
6880 User = *User->use_begin();
6883 if (User->getOpcode() != ISD::BRCOND &&
6884 User->getOpcode() != ISD::SETCC &&
6885 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6898 // Due to the ISEL shortcoming noted above, be conservative if this op is
6899 // likely to be selected as part of a load-modify-store instruction.
6900 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6901 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6902 if (UI->getOpcode() == ISD::STORE)
6905 // Otherwise use a regular EFLAGS-setting instruction.
6906 switch (Op.getNode()->getOpcode()) {
6907 default: llvm_unreachable("unexpected operator!");
6908 case ISD::SUB: Opcode = X86ISD::SUB; break;
6909 case ISD::OR: Opcode = X86ISD::OR; break;
6910 case ISD::XOR: Opcode = X86ISD::XOR; break;
6911 case ISD::AND: Opcode = X86ISD::AND; break;
6923 return SDValue(Op.getNode(), 1);
6930 // Emit a CMP with 0, which is the TEST pattern.
6931 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6932 DAG.getConstant(0, Op.getValueType()));
6934 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6935 SmallVector<SDValue, 4> Ops;
6936 for (unsigned i = 0; i != NumOperands; ++i)
6937 Ops.push_back(Op.getOperand(i));
6939 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6940 DAG.ReplaceAllUsesWith(Op, New);
6941 return SDValue(New.getNode(), 1);
6944 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6946 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6947 SelectionDAG &DAG) const {
6948 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6949 if (C->getAPIntValue() == 0)
6950 return EmitTest(Op0, X86CC, DAG);
6952 DebugLoc dl = Op0.getDebugLoc();
6953 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6956 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6957 /// if it's possible.
6958 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6959 DebugLoc dl, SelectionDAG &DAG) const {
6960 SDValue Op0 = And.getOperand(0);
6961 SDValue Op1 = And.getOperand(1);
6962 if (Op0.getOpcode() == ISD::TRUNCATE)
6963 Op0 = Op0.getOperand(0);
6964 if (Op1.getOpcode() == ISD::TRUNCATE)
6965 Op1 = Op1.getOperand(0);
6968 if (Op1.getOpcode() == ISD::SHL)
6969 std::swap(Op0, Op1);
6970 if (Op0.getOpcode() == ISD::SHL) {
6971 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6972 if (And00C->getZExtValue() == 1) {
6973 // If we looked past a truncate, check that it's only truncating away
6975 unsigned BitWidth = Op0.getValueSizeInBits();
6976 unsigned AndBitWidth = And.getValueSizeInBits();
6977 if (BitWidth > AndBitWidth) {
6978 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6979 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6980 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6984 RHS = Op0.getOperand(1);
6986 } else if (Op1.getOpcode() == ISD::Constant) {
6987 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6988 SDValue AndLHS = Op0;
6989 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6990 LHS = AndLHS.getOperand(0);
6991 RHS = AndLHS.getOperand(1);
6995 if (LHS.getNode()) {
6996 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6997 // instruction. Since the shift amount is in-range-or-undefined, we know
6998 // that doing a bittest on the i32 value is ok. We extend to i32 because
6999 // the encoding for the i16 version is larger than the i32 version.
7000 // Also promote i16 to i32 for performance / code size reason.
7001 if (LHS.getValueType() == MVT::i8 ||
7002 LHS.getValueType() == MVT::i16)
7003 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7005 // If the operand types disagree, extend the shift amount to match. Since
7006 // BT ignores high bits (like shifts) we can use anyextend.
7007 if (LHS.getValueType() != RHS.getValueType())
7008 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7010 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7011 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7012 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7013 DAG.getConstant(Cond, MVT::i8), BT);
7019 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7020 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7021 SDValue Op0 = Op.getOperand(0);
7022 SDValue Op1 = Op.getOperand(1);
7023 DebugLoc dl = Op.getDebugLoc();
7024 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7026 // Optimize to BT if possible.
7027 // Lower (X & (1 << N)) == 0 to BT(X, N).
7028 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7029 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7030 if (Op0.getOpcode() == ISD::AND &&
7032 Op1.getOpcode() == ISD::Constant &&
7033 cast<ConstantSDNode>(Op1)->isNullValue() &&
7034 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7035 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7036 if (NewSetCC.getNode())
7040 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
7041 if (Op0.getOpcode() == X86ISD::SETCC &&
7042 Op1.getOpcode() == ISD::Constant &&
7043 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7044 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7045 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7046 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7047 bool Invert = (CC == ISD::SETNE) ^
7048 cast<ConstantSDNode>(Op1)->isNullValue();
7050 CCode = X86::GetOppositeBranchCondition(CCode);
7051 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7052 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7055 bool isFP = Op1.getValueType().isFloatingPoint();
7056 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7057 if (X86CC == X86::COND_INVALID)
7060 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
7062 // Use sbb x, x to materialize carry bit into a GPR.
7063 if (X86CC == X86::COND_B)
7064 return DAG.getNode(ISD::AND, dl, MVT::i8,
7065 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
7066 DAG.getConstant(X86CC, MVT::i8), Cond),
7067 DAG.getConstant(1, MVT::i8));
7069 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7070 DAG.getConstant(X86CC, MVT::i8), Cond);
7073 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7075 SDValue Op0 = Op.getOperand(0);
7076 SDValue Op1 = Op.getOperand(1);
7077 SDValue CC = Op.getOperand(2);
7078 EVT VT = Op.getValueType();
7079 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7080 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7081 DebugLoc dl = Op.getDebugLoc();
7085 EVT VT0 = Op0.getValueType();
7086 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7087 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7090 switch (SetCCOpcode) {
7093 case ISD::SETEQ: SSECC = 0; break;
7095 case ISD::SETGT: Swap = true; // Fallthrough
7097 case ISD::SETOLT: SSECC = 1; break;
7099 case ISD::SETGE: Swap = true; // Fallthrough
7101 case ISD::SETOLE: SSECC = 2; break;
7102 case ISD::SETUO: SSECC = 3; break;
7104 case ISD::SETNE: SSECC = 4; break;
7105 case ISD::SETULE: Swap = true;
7106 case ISD::SETUGE: SSECC = 5; break;
7107 case ISD::SETULT: Swap = true;
7108 case ISD::SETUGT: SSECC = 6; break;
7109 case ISD::SETO: SSECC = 7; break;
7112 std::swap(Op0, Op1);
7114 // In the two special cases we can't handle, emit two comparisons.
7116 if (SetCCOpcode == ISD::SETUEQ) {
7118 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7119 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7120 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7122 else if (SetCCOpcode == ISD::SETONE) {
7124 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7125 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7126 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7128 llvm_unreachable("Illegal FP comparison");
7130 // Handle all other FP comparisons here.
7131 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7134 // We are handling one of the integer comparisons here. Since SSE only has
7135 // GT and EQ comparisons for integer, swapping operands and multiple
7136 // operations may be required for some comparisons.
7137 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7138 bool Swap = false, Invert = false, FlipSigns = false;
7140 switch (VT.getSimpleVT().SimpleTy) {
7142 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7143 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7144 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7145 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7148 switch (SetCCOpcode) {
7150 case ISD::SETNE: Invert = true;
7151 case ISD::SETEQ: Opc = EQOpc; break;
7152 case ISD::SETLT: Swap = true;
7153 case ISD::SETGT: Opc = GTOpc; break;
7154 case ISD::SETGE: Swap = true;
7155 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7156 case ISD::SETULT: Swap = true;
7157 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7158 case ISD::SETUGE: Swap = true;
7159 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7162 std::swap(Op0, Op1);
7164 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7165 // bits of the inputs before performing those operations.
7167 EVT EltVT = VT.getVectorElementType();
7168 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7170 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7171 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7173 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7174 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7177 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7179 // If the logical-not of the result is required, perform that now.
7181 Result = DAG.getNOT(dl, Result, VT);
7186 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7187 static bool isX86LogicalCmp(SDValue Op) {
7188 unsigned Opc = Op.getNode()->getOpcode();
7189 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7191 if (Op.getResNo() == 1 &&
7192 (Opc == X86ISD::ADD ||
7193 Opc == X86ISD::SUB ||
7194 Opc == X86ISD::SMUL ||
7195 Opc == X86ISD::UMUL ||
7196 Opc == X86ISD::INC ||
7197 Opc == X86ISD::DEC ||
7198 Opc == X86ISD::OR ||
7199 Opc == X86ISD::XOR ||
7200 Opc == X86ISD::AND))
7206 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7207 bool addTest = true;
7208 SDValue Cond = Op.getOperand(0);
7209 DebugLoc dl = Op.getDebugLoc();
7212 if (Cond.getOpcode() == ISD::SETCC) {
7213 SDValue NewCond = LowerSETCC(Cond, DAG);
7214 if (NewCond.getNode())
7218 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
7219 SDValue Op1 = Op.getOperand(1);
7220 SDValue Op2 = Op.getOperand(2);
7221 if (Cond.getOpcode() == X86ISD::SETCC &&
7222 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
7223 SDValue Cmp = Cond.getOperand(1);
7224 if (Cmp.getOpcode() == X86ISD::CMP) {
7225 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
7226 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7227 ConstantSDNode *RHSC =
7228 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
7229 if (N1C && N1C->isAllOnesValue() &&
7230 N2C && N2C->isNullValue() &&
7231 RHSC && RHSC->isNullValue()) {
7232 SDValue CmpOp0 = Cmp.getOperand(0);
7233 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7234 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7235 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
7236 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7241 // Look pass (and (setcc_carry (cmp ...)), 1).
7242 if (Cond.getOpcode() == ISD::AND &&
7243 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7244 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7245 if (C && C->getAPIntValue() == 1)
7246 Cond = Cond.getOperand(0);
7249 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7250 // setting operand in place of the X86ISD::SETCC.
7251 if (Cond.getOpcode() == X86ISD::SETCC ||
7252 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7253 CC = Cond.getOperand(0);
7255 SDValue Cmp = Cond.getOperand(1);
7256 unsigned Opc = Cmp.getOpcode();
7257 EVT VT = Op.getValueType();
7259 bool IllegalFPCMov = false;
7260 if (VT.isFloatingPoint() && !VT.isVector() &&
7261 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7262 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7264 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7265 Opc == X86ISD::BT) { // FIXME
7272 // Look pass the truncate.
7273 if (Cond.getOpcode() == ISD::TRUNCATE)
7274 Cond = Cond.getOperand(0);
7276 // We know the result of AND is compared against zero. Try to match
7278 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7279 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7280 if (NewSetCC.getNode()) {
7281 CC = NewSetCC.getOperand(0);
7282 Cond = NewSetCC.getOperand(1);
7289 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7290 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7293 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7294 // condition is true.
7295 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7296 SDValue Ops[] = { Op2, Op1, CC, Cond };
7297 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7300 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7301 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7302 // from the AND / OR.
7303 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7304 Opc = Op.getOpcode();
7305 if (Opc != ISD::OR && Opc != ISD::AND)
7307 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7308 Op.getOperand(0).hasOneUse() &&
7309 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7310 Op.getOperand(1).hasOneUse());
7313 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7314 // 1 and that the SETCC node has a single use.
7315 static bool isXor1OfSetCC(SDValue Op) {
7316 if (Op.getOpcode() != ISD::XOR)
7318 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7319 if (N1C && N1C->getAPIntValue() == 1) {
7320 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7321 Op.getOperand(0).hasOneUse();
7326 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7327 bool addTest = true;
7328 SDValue Chain = Op.getOperand(0);
7329 SDValue Cond = Op.getOperand(1);
7330 SDValue Dest = Op.getOperand(2);
7331 DebugLoc dl = Op.getDebugLoc();
7334 if (Cond.getOpcode() == ISD::SETCC) {
7335 SDValue NewCond = LowerSETCC(Cond, DAG);
7336 if (NewCond.getNode())
7340 // FIXME: LowerXALUO doesn't handle these!!
7341 else if (Cond.getOpcode() == X86ISD::ADD ||
7342 Cond.getOpcode() == X86ISD::SUB ||
7343 Cond.getOpcode() == X86ISD::SMUL ||
7344 Cond.getOpcode() == X86ISD::UMUL)
7345 Cond = LowerXALUO(Cond, DAG);
7348 // Look pass (and (setcc_carry (cmp ...)), 1).
7349 if (Cond.getOpcode() == ISD::AND &&
7350 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7351 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7352 if (C && C->getAPIntValue() == 1)
7353 Cond = Cond.getOperand(0);
7356 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7357 // setting operand in place of the X86ISD::SETCC.
7358 if (Cond.getOpcode() == X86ISD::SETCC ||
7359 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7360 CC = Cond.getOperand(0);
7362 SDValue Cmp = Cond.getOperand(1);
7363 unsigned Opc = Cmp.getOpcode();
7364 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7365 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7369 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7373 // These can only come from an arithmetic instruction with overflow,
7374 // e.g. SADDO, UADDO.
7375 Cond = Cond.getNode()->getOperand(1);
7382 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7383 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7384 if (CondOpc == ISD::OR) {
7385 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7386 // two branches instead of an explicit OR instruction with a
7388 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7389 isX86LogicalCmp(Cmp)) {
7390 CC = Cond.getOperand(0).getOperand(0);
7391 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7392 Chain, Dest, CC, Cmp);
7393 CC = Cond.getOperand(1).getOperand(0);
7397 } else { // ISD::AND
7398 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7399 // two branches instead of an explicit AND instruction with a
7400 // separate test. However, we only do this if this block doesn't
7401 // have a fall-through edge, because this requires an explicit
7402 // jmp when the condition is false.
7403 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7404 isX86LogicalCmp(Cmp) &&
7405 Op.getNode()->hasOneUse()) {
7406 X86::CondCode CCode =
7407 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7408 CCode = X86::GetOppositeBranchCondition(CCode);
7409 CC = DAG.getConstant(CCode, MVT::i8);
7410 SDNode *User = *Op.getNode()->use_begin();
7411 // Look for an unconditional branch following this conditional branch.
7412 // We need this because we need to reverse the successors in order
7413 // to implement FCMP_OEQ.
7414 if (User->getOpcode() == ISD::BR) {
7415 SDValue FalseBB = User->getOperand(1);
7417 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7418 assert(NewBR == User);
7422 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7423 Chain, Dest, CC, Cmp);
7424 X86::CondCode CCode =
7425 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7426 CCode = X86::GetOppositeBranchCondition(CCode);
7427 CC = DAG.getConstant(CCode, MVT::i8);
7433 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7434 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7435 // It should be transformed during dag combiner except when the condition
7436 // is set by a arithmetics with overflow node.
7437 X86::CondCode CCode =
7438 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7439 CCode = X86::GetOppositeBranchCondition(CCode);
7440 CC = DAG.getConstant(CCode, MVT::i8);
7441 Cond = Cond.getOperand(0).getOperand(1);
7447 // Look pass the truncate.
7448 if (Cond.getOpcode() == ISD::TRUNCATE)
7449 Cond = Cond.getOperand(0);
7451 // We know the result of AND is compared against zero. Try to match
7453 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7454 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7455 if (NewSetCC.getNode()) {
7456 CC = NewSetCC.getOperand(0);
7457 Cond = NewSetCC.getOperand(1);
7464 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7465 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7467 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7468 Chain, Dest, CC, Cond);
7472 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7473 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7474 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7475 // that the guard pages used by the OS virtual memory manager are allocated in
7476 // correct sequence.
7478 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7479 SelectionDAG &DAG) const {
7480 assert(Subtarget->isTargetCygMing() &&
7481 "This should be used only on Cygwin/Mingw targets");
7482 DebugLoc dl = Op.getDebugLoc();
7485 SDValue Chain = Op.getOperand(0);
7486 SDValue Size = Op.getOperand(1);
7487 // FIXME: Ensure alignment here
7491 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7493 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7494 Flag = Chain.getValue(1);
7496 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7498 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7499 Flag = Chain.getValue(1);
7501 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7503 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7504 return DAG.getMergeValues(Ops1, 2, dl);
7507 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7508 MachineFunction &MF = DAG.getMachineFunction();
7509 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7511 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7512 DebugLoc DL = Op.getDebugLoc();
7514 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
7515 // vastart just stores the address of the VarArgsFrameIndex slot into the
7516 // memory location argument.
7517 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7519 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7520 MachinePointerInfo(SV), false, false, 0);
7524 // gp_offset (0 - 6 * 8)
7525 // fp_offset (48 - 48 + 8 * 16)
7526 // overflow_arg_area (point to parameters coming in memory).
7528 SmallVector<SDValue, 8> MemOps;
7529 SDValue FIN = Op.getOperand(1);
7531 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7532 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7534 FIN, MachinePointerInfo(SV), false, false, 0);
7535 MemOps.push_back(Store);
7538 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7539 FIN, DAG.getIntPtrConstant(4));
7540 Store = DAG.getStore(Op.getOperand(0), DL,
7541 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7543 FIN, MachinePointerInfo(SV, 4), false, false, 0);
7544 MemOps.push_back(Store);
7546 // Store ptr to overflow_arg_area
7547 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7548 FIN, DAG.getIntPtrConstant(4));
7549 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7551 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
7552 MachinePointerInfo(SV, 8),
7554 MemOps.push_back(Store);
7556 // Store ptr to reg_save_area.
7557 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7558 FIN, DAG.getIntPtrConstant(8));
7559 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7561 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
7562 MachinePointerInfo(SV, 16), false, false, 0);
7563 MemOps.push_back(Store);
7564 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
7565 &MemOps[0], MemOps.size());
7568 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7569 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7570 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7572 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7576 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7577 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7578 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7579 SDValue Chain = Op.getOperand(0);
7580 SDValue DstPtr = Op.getOperand(1);
7581 SDValue SrcPtr = Op.getOperand(2);
7582 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7583 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7584 DebugLoc DL = Op.getDebugLoc();
7586 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
7587 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7589 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
7593 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7594 DebugLoc dl = Op.getDebugLoc();
7595 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7597 default: return SDValue(); // Don't custom lower most intrinsics.
7598 // Comparison intrinsics.
7599 case Intrinsic::x86_sse_comieq_ss:
7600 case Intrinsic::x86_sse_comilt_ss:
7601 case Intrinsic::x86_sse_comile_ss:
7602 case Intrinsic::x86_sse_comigt_ss:
7603 case Intrinsic::x86_sse_comige_ss:
7604 case Intrinsic::x86_sse_comineq_ss:
7605 case Intrinsic::x86_sse_ucomieq_ss:
7606 case Intrinsic::x86_sse_ucomilt_ss:
7607 case Intrinsic::x86_sse_ucomile_ss:
7608 case Intrinsic::x86_sse_ucomigt_ss:
7609 case Intrinsic::x86_sse_ucomige_ss:
7610 case Intrinsic::x86_sse_ucomineq_ss:
7611 case Intrinsic::x86_sse2_comieq_sd:
7612 case Intrinsic::x86_sse2_comilt_sd:
7613 case Intrinsic::x86_sse2_comile_sd:
7614 case Intrinsic::x86_sse2_comigt_sd:
7615 case Intrinsic::x86_sse2_comige_sd:
7616 case Intrinsic::x86_sse2_comineq_sd:
7617 case Intrinsic::x86_sse2_ucomieq_sd:
7618 case Intrinsic::x86_sse2_ucomilt_sd:
7619 case Intrinsic::x86_sse2_ucomile_sd:
7620 case Intrinsic::x86_sse2_ucomigt_sd:
7621 case Intrinsic::x86_sse2_ucomige_sd:
7622 case Intrinsic::x86_sse2_ucomineq_sd: {
7624 ISD::CondCode CC = ISD::SETCC_INVALID;
7627 case Intrinsic::x86_sse_comieq_ss:
7628 case Intrinsic::x86_sse2_comieq_sd:
7632 case Intrinsic::x86_sse_comilt_ss:
7633 case Intrinsic::x86_sse2_comilt_sd:
7637 case Intrinsic::x86_sse_comile_ss:
7638 case Intrinsic::x86_sse2_comile_sd:
7642 case Intrinsic::x86_sse_comigt_ss:
7643 case Intrinsic::x86_sse2_comigt_sd:
7647 case Intrinsic::x86_sse_comige_ss:
7648 case Intrinsic::x86_sse2_comige_sd:
7652 case Intrinsic::x86_sse_comineq_ss:
7653 case Intrinsic::x86_sse2_comineq_sd:
7657 case Intrinsic::x86_sse_ucomieq_ss:
7658 case Intrinsic::x86_sse2_ucomieq_sd:
7659 Opc = X86ISD::UCOMI;
7662 case Intrinsic::x86_sse_ucomilt_ss:
7663 case Intrinsic::x86_sse2_ucomilt_sd:
7664 Opc = X86ISD::UCOMI;
7667 case Intrinsic::x86_sse_ucomile_ss:
7668 case Intrinsic::x86_sse2_ucomile_sd:
7669 Opc = X86ISD::UCOMI;
7672 case Intrinsic::x86_sse_ucomigt_ss:
7673 case Intrinsic::x86_sse2_ucomigt_sd:
7674 Opc = X86ISD::UCOMI;
7677 case Intrinsic::x86_sse_ucomige_ss:
7678 case Intrinsic::x86_sse2_ucomige_sd:
7679 Opc = X86ISD::UCOMI;
7682 case Intrinsic::x86_sse_ucomineq_ss:
7683 case Intrinsic::x86_sse2_ucomineq_sd:
7684 Opc = X86ISD::UCOMI;
7689 SDValue LHS = Op.getOperand(1);
7690 SDValue RHS = Op.getOperand(2);
7691 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7692 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7693 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7694 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7695 DAG.getConstant(X86CC, MVT::i8), Cond);
7696 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7698 // ptest and testp intrinsics. The intrinsic these come from are designed to
7699 // return an integer value, not just an instruction so lower it to the ptest
7700 // or testp pattern and a setcc for the result.
7701 case Intrinsic::x86_sse41_ptestz:
7702 case Intrinsic::x86_sse41_ptestc:
7703 case Intrinsic::x86_sse41_ptestnzc:
7704 case Intrinsic::x86_avx_ptestz_256:
7705 case Intrinsic::x86_avx_ptestc_256:
7706 case Intrinsic::x86_avx_ptestnzc_256:
7707 case Intrinsic::x86_avx_vtestz_ps:
7708 case Intrinsic::x86_avx_vtestc_ps:
7709 case Intrinsic::x86_avx_vtestnzc_ps:
7710 case Intrinsic::x86_avx_vtestz_pd:
7711 case Intrinsic::x86_avx_vtestc_pd:
7712 case Intrinsic::x86_avx_vtestnzc_pd:
7713 case Intrinsic::x86_avx_vtestz_ps_256:
7714 case Intrinsic::x86_avx_vtestc_ps_256:
7715 case Intrinsic::x86_avx_vtestnzc_ps_256:
7716 case Intrinsic::x86_avx_vtestz_pd_256:
7717 case Intrinsic::x86_avx_vtestc_pd_256:
7718 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7719 bool IsTestPacked = false;
7722 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7723 case Intrinsic::x86_avx_vtestz_ps:
7724 case Intrinsic::x86_avx_vtestz_pd:
7725 case Intrinsic::x86_avx_vtestz_ps_256:
7726 case Intrinsic::x86_avx_vtestz_pd_256:
7727 IsTestPacked = true; // Fallthrough
7728 case Intrinsic::x86_sse41_ptestz:
7729 case Intrinsic::x86_avx_ptestz_256:
7731 X86CC = X86::COND_E;
7733 case Intrinsic::x86_avx_vtestc_ps:
7734 case Intrinsic::x86_avx_vtestc_pd:
7735 case Intrinsic::x86_avx_vtestc_ps_256:
7736 case Intrinsic::x86_avx_vtestc_pd_256:
7737 IsTestPacked = true; // Fallthrough
7738 case Intrinsic::x86_sse41_ptestc:
7739 case Intrinsic::x86_avx_ptestc_256:
7741 X86CC = X86::COND_B;
7743 case Intrinsic::x86_avx_vtestnzc_ps:
7744 case Intrinsic::x86_avx_vtestnzc_pd:
7745 case Intrinsic::x86_avx_vtestnzc_ps_256:
7746 case Intrinsic::x86_avx_vtestnzc_pd_256:
7747 IsTestPacked = true; // Fallthrough
7748 case Intrinsic::x86_sse41_ptestnzc:
7749 case Intrinsic::x86_avx_ptestnzc_256:
7751 X86CC = X86::COND_A;
7755 SDValue LHS = Op.getOperand(1);
7756 SDValue RHS = Op.getOperand(2);
7757 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7758 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7759 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7760 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7761 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7764 // Fix vector shift instructions where the last operand is a non-immediate
7766 case Intrinsic::x86_sse2_pslli_w:
7767 case Intrinsic::x86_sse2_pslli_d:
7768 case Intrinsic::x86_sse2_pslli_q:
7769 case Intrinsic::x86_sse2_psrli_w:
7770 case Intrinsic::x86_sse2_psrli_d:
7771 case Intrinsic::x86_sse2_psrli_q:
7772 case Intrinsic::x86_sse2_psrai_w:
7773 case Intrinsic::x86_sse2_psrai_d:
7774 case Intrinsic::x86_mmx_pslli_w:
7775 case Intrinsic::x86_mmx_pslli_d:
7776 case Intrinsic::x86_mmx_pslli_q:
7777 case Intrinsic::x86_mmx_psrli_w:
7778 case Intrinsic::x86_mmx_psrli_d:
7779 case Intrinsic::x86_mmx_psrli_q:
7780 case Intrinsic::x86_mmx_psrai_w:
7781 case Intrinsic::x86_mmx_psrai_d: {
7782 SDValue ShAmt = Op.getOperand(2);
7783 if (isa<ConstantSDNode>(ShAmt))
7786 unsigned NewIntNo = 0;
7787 EVT ShAmtVT = MVT::v4i32;
7789 case Intrinsic::x86_sse2_pslli_w:
7790 NewIntNo = Intrinsic::x86_sse2_psll_w;
7792 case Intrinsic::x86_sse2_pslli_d:
7793 NewIntNo = Intrinsic::x86_sse2_psll_d;
7795 case Intrinsic::x86_sse2_pslli_q:
7796 NewIntNo = Intrinsic::x86_sse2_psll_q;
7798 case Intrinsic::x86_sse2_psrli_w:
7799 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7801 case Intrinsic::x86_sse2_psrli_d:
7802 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7804 case Intrinsic::x86_sse2_psrli_q:
7805 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7807 case Intrinsic::x86_sse2_psrai_w:
7808 NewIntNo = Intrinsic::x86_sse2_psra_w;
7810 case Intrinsic::x86_sse2_psrai_d:
7811 NewIntNo = Intrinsic::x86_sse2_psra_d;
7814 ShAmtVT = MVT::v2i32;
7816 case Intrinsic::x86_mmx_pslli_w:
7817 NewIntNo = Intrinsic::x86_mmx_psll_w;
7819 case Intrinsic::x86_mmx_pslli_d:
7820 NewIntNo = Intrinsic::x86_mmx_psll_d;
7822 case Intrinsic::x86_mmx_pslli_q:
7823 NewIntNo = Intrinsic::x86_mmx_psll_q;
7825 case Intrinsic::x86_mmx_psrli_w:
7826 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7828 case Intrinsic::x86_mmx_psrli_d:
7829 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7831 case Intrinsic::x86_mmx_psrli_q:
7832 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7834 case Intrinsic::x86_mmx_psrai_w:
7835 NewIntNo = Intrinsic::x86_mmx_psra_w;
7837 case Intrinsic::x86_mmx_psrai_d:
7838 NewIntNo = Intrinsic::x86_mmx_psra_d;
7840 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7846 // The vector shift intrinsics with scalars uses 32b shift amounts but
7847 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7851 ShOps[1] = DAG.getConstant(0, MVT::i32);
7852 if (ShAmtVT == MVT::v4i32) {
7853 ShOps[2] = DAG.getUNDEF(MVT::i32);
7854 ShOps[3] = DAG.getUNDEF(MVT::i32);
7855 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7857 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7858 // FIXME this must be lowered to get rid of the invalid type.
7861 EVT VT = Op.getValueType();
7862 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7863 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7864 DAG.getConstant(NewIntNo, MVT::i32),
7865 Op.getOperand(1), ShAmt);
7870 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7871 SelectionDAG &DAG) const {
7872 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7873 MFI->setReturnAddressIsTaken(true);
7875 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7876 DebugLoc dl = Op.getDebugLoc();
7879 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7881 DAG.getConstant(TD->getPointerSize(),
7882 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7883 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7884 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7886 MachinePointerInfo(), false, false, 0);
7889 // Just load the return address.
7890 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7891 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7892 RetAddrFI, MachinePointerInfo(), false, false, 0);
7895 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7896 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7897 MFI->setFrameAddressIsTaken(true);
7899 EVT VT = Op.getValueType();
7900 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7901 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7902 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7903 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7905 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
7906 MachinePointerInfo(),
7911 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7912 SelectionDAG &DAG) const {
7913 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7916 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7917 MachineFunction &MF = DAG.getMachineFunction();
7918 SDValue Chain = Op.getOperand(0);
7919 SDValue Offset = Op.getOperand(1);
7920 SDValue Handler = Op.getOperand(2);
7921 DebugLoc dl = Op.getDebugLoc();
7923 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7924 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7926 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7928 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7929 DAG.getIntPtrConstant(TD->getPointerSize()));
7930 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7931 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
7933 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7934 MF.getRegInfo().addLiveOut(StoreAddrReg);
7936 return DAG.getNode(X86ISD::EH_RETURN, dl,
7938 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7941 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7942 SelectionDAG &DAG) const {
7943 SDValue Root = Op.getOperand(0);
7944 SDValue Trmp = Op.getOperand(1); // trampoline
7945 SDValue FPtr = Op.getOperand(2); // nested function
7946 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7947 DebugLoc dl = Op.getDebugLoc();
7949 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7951 if (Subtarget->is64Bit()) {
7952 SDValue OutChains[6];
7954 // Large code-model.
7955 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7956 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7958 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7959 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7961 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7963 // Load the pointer to the nested function into R11.
7964 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7965 SDValue Addr = Trmp;
7966 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7967 Addr, MachinePointerInfo(TrmpAddr),
7970 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7971 DAG.getConstant(2, MVT::i64));
7972 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
7973 MachinePointerInfo(TrmpAddr, 2),
7976 // Load the 'nest' parameter value into R10.
7977 // R10 is specified in X86CallingConv.td
7978 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7979 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7980 DAG.getConstant(10, MVT::i64));
7981 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7982 Addr, MachinePointerInfo(TrmpAddr, 10),
7985 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7986 DAG.getConstant(12, MVT::i64));
7987 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
7988 MachinePointerInfo(TrmpAddr, 12),
7991 // Jump to the nested function.
7992 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7993 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7994 DAG.getConstant(20, MVT::i64));
7995 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7996 Addr, MachinePointerInfo(TrmpAddr, 20),
7999 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8000 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8001 DAG.getConstant(22, MVT::i64));
8002 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8003 MachinePointerInfo(TrmpAddr, 22),
8007 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8008 return DAG.getMergeValues(Ops, 2, dl);
8010 const Function *Func =
8011 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8012 CallingConv::ID CC = Func->getCallingConv();
8017 llvm_unreachable("Unsupported calling convention");
8018 case CallingConv::C:
8019 case CallingConv::X86_StdCall: {
8020 // Pass 'nest' parameter in ECX.
8021 // Must be kept in sync with X86CallingConv.td
8024 // Check that ECX wasn't needed by an 'inreg' parameter.
8025 const FunctionType *FTy = Func->getFunctionType();
8026 const AttrListPtr &Attrs = Func->getAttributes();
8028 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8029 unsigned InRegCount = 0;
8032 for (FunctionType::param_iterator I = FTy->param_begin(),
8033 E = FTy->param_end(); I != E; ++I, ++Idx)
8034 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8035 // FIXME: should only count parameters that are lowered to integers.
8036 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8038 if (InRegCount > 2) {
8039 report_fatal_error("Nest register in use - reduce number of inreg"
8045 case CallingConv::X86_FastCall:
8046 case CallingConv::X86_ThisCall:
8047 case CallingConv::Fast:
8048 // Pass 'nest' parameter in EAX.
8049 // Must be kept in sync with X86CallingConv.td
8054 SDValue OutChains[4];
8057 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8058 DAG.getConstant(10, MVT::i32));
8059 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8061 // This is storing the opcode for MOV32ri.
8062 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8063 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8064 OutChains[0] = DAG.getStore(Root, dl,
8065 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8066 Trmp, MachinePointerInfo(TrmpAddr),
8069 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8070 DAG.getConstant(1, MVT::i32));
8071 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8072 MachinePointerInfo(TrmpAddr, 1),
8075 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8076 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8077 DAG.getConstant(5, MVT::i32));
8078 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8079 MachinePointerInfo(TrmpAddr, 5),
8082 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8083 DAG.getConstant(6, MVT::i32));
8084 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8085 MachinePointerInfo(TrmpAddr, 6),
8089 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8090 return DAG.getMergeValues(Ops, 2, dl);
8094 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8095 SelectionDAG &DAG) const {
8097 The rounding mode is in bits 11:10 of FPSR, and has the following
8104 FLT_ROUNDS, on the other hand, expects the following:
8111 To perform the conversion, we do:
8112 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8115 MachineFunction &MF = DAG.getMachineFunction();
8116 const TargetMachine &TM = MF.getTarget();
8117 const TargetFrameInfo &TFI = *TM.getFrameInfo();
8118 unsigned StackAlignment = TFI.getStackAlignment();
8119 EVT VT = Op.getValueType();
8120 DebugLoc DL = Op.getDebugLoc();
8122 // Save FP Control Word to stack slot
8123 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8124 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8127 MachineMemOperand *MMO =
8128 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8129 MachineMemOperand::MOStore, 2, 2);
8131 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8132 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8133 DAG.getVTList(MVT::Other),
8134 Ops, 2, MVT::i16, MMO);
8136 // Load FP Control Word from stack slot
8137 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8138 MachinePointerInfo(), false, false, 0);
8140 // Transform as necessary
8142 DAG.getNode(ISD::SRL, DL, MVT::i16,
8143 DAG.getNode(ISD::AND, DL, MVT::i16,
8144 CWD, DAG.getConstant(0x800, MVT::i16)),
8145 DAG.getConstant(11, MVT::i8));
8147 DAG.getNode(ISD::SRL, DL, MVT::i16,
8148 DAG.getNode(ISD::AND, DL, MVT::i16,
8149 CWD, DAG.getConstant(0x400, MVT::i16)),
8150 DAG.getConstant(9, MVT::i8));
8153 DAG.getNode(ISD::AND, DL, MVT::i16,
8154 DAG.getNode(ISD::ADD, DL, MVT::i16,
8155 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8156 DAG.getConstant(1, MVT::i16)),
8157 DAG.getConstant(3, MVT::i16));
8160 return DAG.getNode((VT.getSizeInBits() < 16 ?
8161 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8164 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8165 EVT VT = Op.getValueType();
8167 unsigned NumBits = VT.getSizeInBits();
8168 DebugLoc dl = Op.getDebugLoc();
8170 Op = Op.getOperand(0);
8171 if (VT == MVT::i8) {
8172 // Zero extend to i32 since there is not an i8 bsr.
8174 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8177 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8178 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8179 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8181 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8184 DAG.getConstant(NumBits+NumBits-1, OpVT),
8185 DAG.getConstant(X86::COND_E, MVT::i8),
8188 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8190 // Finally xor with NumBits-1.
8191 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8194 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8198 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8199 EVT VT = Op.getValueType();
8201 unsigned NumBits = VT.getSizeInBits();
8202 DebugLoc dl = Op.getDebugLoc();
8204 Op = Op.getOperand(0);
8205 if (VT == MVT::i8) {
8207 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8210 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8211 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8212 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8214 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8217 DAG.getConstant(NumBits, OpVT),
8218 DAG.getConstant(X86::COND_E, MVT::i8),
8221 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8224 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8228 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8229 EVT VT = Op.getValueType();
8230 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8231 DebugLoc dl = Op.getDebugLoc();
8233 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8234 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8235 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8236 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8237 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8239 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8240 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8241 // return AloBlo + AloBhi + AhiBlo;
8243 SDValue A = Op.getOperand(0);
8244 SDValue B = Op.getOperand(1);
8246 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8247 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8248 A, DAG.getConstant(32, MVT::i32));
8249 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8250 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8251 B, DAG.getConstant(32, MVT::i32));
8252 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8253 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8255 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8256 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8258 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8259 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8261 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8262 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8263 AloBhi, DAG.getConstant(32, MVT::i32));
8264 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8265 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8266 AhiBlo, DAG.getConstant(32, MVT::i32));
8267 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8268 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8272 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8273 EVT VT = Op.getValueType();
8274 DebugLoc dl = Op.getDebugLoc();
8275 SDValue R = Op.getOperand(0);
8277 LLVMContext *Context = DAG.getContext();
8279 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8281 if (VT == MVT::v4i32) {
8282 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8283 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8284 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8286 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8288 std::vector<Constant*> CV(4, CI);
8289 Constant *C = ConstantVector::get(CV);
8290 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8291 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8292 MachinePointerInfo::getConstantPool(),
8295 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8296 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8297 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8298 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8300 if (VT == MVT::v16i8) {
8302 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8303 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8304 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8306 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8307 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8309 std::vector<Constant*> CVM1(16, CM1);
8310 std::vector<Constant*> CVM2(16, CM2);
8311 Constant *C = ConstantVector::get(CVM1);
8312 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8313 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8314 MachinePointerInfo::getConstantPool(),
8317 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8318 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8319 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8320 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8321 DAG.getConstant(4, MVT::i32));
8322 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8323 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8326 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8328 C = ConstantVector::get(CVM2);
8329 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8330 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8331 MachinePointerInfo::getConstantPool(),
8334 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8335 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8336 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8337 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8338 DAG.getConstant(2, MVT::i32));
8339 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8340 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8343 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8345 // return pblendv(r, r+r, a);
8346 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8347 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8348 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8354 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8355 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8356 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8357 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8358 // has only one use.
8359 SDNode *N = Op.getNode();
8360 SDValue LHS = N->getOperand(0);
8361 SDValue RHS = N->getOperand(1);
8362 unsigned BaseOp = 0;
8364 DebugLoc dl = Op.getDebugLoc();
8366 switch (Op.getOpcode()) {
8367 default: llvm_unreachable("Unknown ovf instruction!");
8369 // A subtract of one will be selected as a INC. Note that INC doesn't
8370 // set CF, so we can't do this for UADDO.
8371 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8372 if (C->getAPIntValue() == 1) {
8373 BaseOp = X86ISD::INC;
8377 BaseOp = X86ISD::ADD;
8381 BaseOp = X86ISD::ADD;
8385 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8386 // set CF, so we can't do this for USUBO.
8387 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8388 if (C->getAPIntValue() == 1) {
8389 BaseOp = X86ISD::DEC;
8393 BaseOp = X86ISD::SUB;
8397 BaseOp = X86ISD::SUB;
8401 BaseOp = X86ISD::SMUL;
8405 BaseOp = X86ISD::UMUL;
8410 // Also sets EFLAGS.
8411 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8412 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8415 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8416 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8418 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8422 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8423 DebugLoc dl = Op.getDebugLoc();
8425 if (!Subtarget->hasSSE2()) {
8426 SDValue Chain = Op.getOperand(0);
8427 SDValue Zero = DAG.getConstant(0,
8428 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8430 DAG.getRegister(X86::ESP, MVT::i32), // Base
8431 DAG.getTargetConstant(1, MVT::i8), // Scale
8432 DAG.getRegister(0, MVT::i32), // Index
8433 DAG.getTargetConstant(0, MVT::i32), // Disp
8434 DAG.getRegister(0, MVT::i32), // Segment.
8439 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8440 array_lengthof(Ops));
8441 return SDValue(Res, 0);
8444 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8446 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8448 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8449 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8450 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8451 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8453 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8454 if (!Op1 && !Op2 && !Op3 && Op4)
8455 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8457 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8458 if (Op1 && !Op2 && !Op3 && !Op4)
8459 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8461 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8463 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8466 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8467 EVT T = Op.getValueType();
8468 DebugLoc DL = Op.getDebugLoc();
8471 switch(T.getSimpleVT().SimpleTy) {
8473 assert(false && "Invalid value type!");
8474 case MVT::i8: Reg = X86::AL; size = 1; break;
8475 case MVT::i16: Reg = X86::AX; size = 2; break;
8476 case MVT::i32: Reg = X86::EAX; size = 4; break;
8478 assert(Subtarget->is64Bit() && "Node not type legal!");
8479 Reg = X86::RAX; size = 8;
8482 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
8483 Op.getOperand(2), SDValue());
8484 SDValue Ops[] = { cpIn.getValue(0),
8487 DAG.getTargetConstant(size, MVT::i8),
8489 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8490 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
8491 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
8494 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
8498 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8499 SelectionDAG &DAG) const {
8500 assert(Subtarget->is64Bit() && "Result not type legalized?");
8501 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8502 SDValue TheChain = Op.getOperand(0);
8503 DebugLoc dl = Op.getDebugLoc();
8504 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8505 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8506 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8508 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8509 DAG.getConstant(32, MVT::i8));
8511 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8514 return DAG.getMergeValues(Ops, 2, dl);
8517 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8518 SelectionDAG &DAG) const {
8519 EVT SrcVT = Op.getOperand(0).getValueType();
8520 EVT DstVT = Op.getValueType();
8521 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8522 Subtarget->hasMMX() && !DisableMMX) &&
8523 "Unexpected custom BIT_CONVERT");
8524 assert((DstVT == MVT::i64 ||
8525 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8526 "Unexpected custom BIT_CONVERT");
8527 // i64 <=> MMX conversions are Legal.
8528 if (SrcVT==MVT::i64 && DstVT.isVector())
8530 if (DstVT==MVT::i64 && SrcVT.isVector())
8532 // MMX <=> MMX conversions are Legal.
8533 if (SrcVT.isVector() && DstVT.isVector())
8535 // All other conversions need to be expanded.
8538 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8539 SDNode *Node = Op.getNode();
8540 DebugLoc dl = Node->getDebugLoc();
8541 EVT T = Node->getValueType(0);
8542 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8543 DAG.getConstant(0, T), Node->getOperand(2));
8544 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8545 cast<AtomicSDNode>(Node)->getMemoryVT(),
8546 Node->getOperand(0),
8547 Node->getOperand(1), negOp,
8548 cast<AtomicSDNode>(Node)->getSrcValue(),
8549 cast<AtomicSDNode>(Node)->getAlignment());
8552 /// LowerOperation - Provide custom lowering hooks for some operations.
8554 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8555 switch (Op.getOpcode()) {
8556 default: llvm_unreachable("Should not custom lower this!");
8557 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8558 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8559 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8560 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8561 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8562 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8563 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8564 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8565 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8566 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8567 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8568 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8569 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8570 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8571 case ISD::SHL_PARTS:
8572 case ISD::SRA_PARTS:
8573 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8574 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8575 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8576 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8577 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8578 case ISD::FABS: return LowerFABS(Op, DAG);
8579 case ISD::FNEG: return LowerFNEG(Op, DAG);
8580 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8581 case ISD::SETCC: return LowerSETCC(Op, DAG);
8582 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8583 case ISD::SELECT: return LowerSELECT(Op, DAG);
8584 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8585 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8586 case ISD::VASTART: return LowerVASTART(Op, DAG);
8587 case ISD::VAARG: return LowerVAARG(Op, DAG);
8588 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8589 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8590 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8591 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8592 case ISD::FRAME_TO_ARGS_OFFSET:
8593 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8594 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8595 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8596 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8597 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8598 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8599 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8600 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8601 case ISD::SHL: return LowerSHL(Op, DAG);
8607 case ISD::UMULO: return LowerXALUO(Op, DAG);
8608 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8609 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8613 void X86TargetLowering::
8614 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8615 SelectionDAG &DAG, unsigned NewOp) const {
8616 EVT T = Node->getValueType(0);
8617 DebugLoc dl = Node->getDebugLoc();
8618 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8620 SDValue Chain = Node->getOperand(0);
8621 SDValue In1 = Node->getOperand(1);
8622 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8623 Node->getOperand(2), DAG.getIntPtrConstant(0));
8624 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8625 Node->getOperand(2), DAG.getIntPtrConstant(1));
8626 SDValue Ops[] = { Chain, In1, In2L, In2H };
8627 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8629 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8630 cast<MemSDNode>(Node)->getMemOperand());
8631 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8632 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8633 Results.push_back(Result.getValue(2));
8636 /// ReplaceNodeResults - Replace a node with an illegal result type
8637 /// with a new node built out of custom code.
8638 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8639 SmallVectorImpl<SDValue>&Results,
8640 SelectionDAG &DAG) const {
8641 DebugLoc dl = N->getDebugLoc();
8642 switch (N->getOpcode()) {
8644 assert(false && "Do not know how to custom type legalize this operation!");
8646 case ISD::FP_TO_SINT: {
8647 std::pair<SDValue,SDValue> Vals =
8648 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8649 SDValue FIST = Vals.first, StackSlot = Vals.second;
8650 if (FIST.getNode() != 0) {
8651 EVT VT = N->getValueType(0);
8652 // Return a load from the stack slot.
8653 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
8654 MachinePointerInfo(), false, false, 0));
8658 case ISD::READCYCLECOUNTER: {
8659 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8660 SDValue TheChain = N->getOperand(0);
8661 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8662 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8664 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8666 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8667 SDValue Ops[] = { eax, edx };
8668 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8669 Results.push_back(edx.getValue(1));
8672 case ISD::ATOMIC_CMP_SWAP: {
8673 EVT T = N->getValueType(0);
8674 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8675 SDValue cpInL, cpInH;
8676 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8677 DAG.getConstant(0, MVT::i32));
8678 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8679 DAG.getConstant(1, MVT::i32));
8680 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8681 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8683 SDValue swapInL, swapInH;
8684 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8685 DAG.getConstant(0, MVT::i32));
8686 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8687 DAG.getConstant(1, MVT::i32));
8688 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8690 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8691 swapInL.getValue(1));
8692 SDValue Ops[] = { swapInH.getValue(0),
8694 swapInH.getValue(1) };
8695 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8696 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8697 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8698 MVT::i32, Result.getValue(1));
8699 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8700 MVT::i32, cpOutL.getValue(2));
8701 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8702 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8703 Results.push_back(cpOutH.getValue(1));
8706 case ISD::ATOMIC_LOAD_ADD:
8707 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8709 case ISD::ATOMIC_LOAD_AND:
8710 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8712 case ISD::ATOMIC_LOAD_NAND:
8713 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8715 case ISD::ATOMIC_LOAD_OR:
8716 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8718 case ISD::ATOMIC_LOAD_SUB:
8719 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8721 case ISD::ATOMIC_LOAD_XOR:
8722 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8724 case ISD::ATOMIC_SWAP:
8725 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8730 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8732 default: return NULL;
8733 case X86ISD::BSF: return "X86ISD::BSF";
8734 case X86ISD::BSR: return "X86ISD::BSR";
8735 case X86ISD::SHLD: return "X86ISD::SHLD";
8736 case X86ISD::SHRD: return "X86ISD::SHRD";
8737 case X86ISD::FAND: return "X86ISD::FAND";
8738 case X86ISD::FOR: return "X86ISD::FOR";
8739 case X86ISD::FXOR: return "X86ISD::FXOR";
8740 case X86ISD::FSRL: return "X86ISD::FSRL";
8741 case X86ISD::FILD: return "X86ISD::FILD";
8742 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8743 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8744 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8745 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8746 case X86ISD::FLD: return "X86ISD::FLD";
8747 case X86ISD::FST: return "X86ISD::FST";
8748 case X86ISD::CALL: return "X86ISD::CALL";
8749 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8750 case X86ISD::BT: return "X86ISD::BT";
8751 case X86ISD::CMP: return "X86ISD::CMP";
8752 case X86ISD::COMI: return "X86ISD::COMI";
8753 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8754 case X86ISD::SETCC: return "X86ISD::SETCC";
8755 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8756 case X86ISD::CMOV: return "X86ISD::CMOV";
8757 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8758 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8759 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8760 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8761 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8762 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8763 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8764 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8765 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8766 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8767 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8768 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8769 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8770 case X86ISD::FMAX: return "X86ISD::FMAX";
8771 case X86ISD::FMIN: return "X86ISD::FMIN";
8772 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8773 case X86ISD::FRCP: return "X86ISD::FRCP";
8774 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8775 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8776 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8777 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8778 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8779 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8780 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8781 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8782 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8783 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8784 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8785 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8786 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8787 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8788 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8789 case X86ISD::VSHL: return "X86ISD::VSHL";
8790 case X86ISD::VSRL: return "X86ISD::VSRL";
8791 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8792 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8793 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8794 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8795 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8796 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8797 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8798 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8799 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8800 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8801 case X86ISD::ADD: return "X86ISD::ADD";
8802 case X86ISD::SUB: return "X86ISD::SUB";
8803 case X86ISD::SMUL: return "X86ISD::SMUL";
8804 case X86ISD::UMUL: return "X86ISD::UMUL";
8805 case X86ISD::INC: return "X86ISD::INC";
8806 case X86ISD::DEC: return "X86ISD::DEC";
8807 case X86ISD::OR: return "X86ISD::OR";
8808 case X86ISD::XOR: return "X86ISD::XOR";
8809 case X86ISD::AND: return "X86ISD::AND";
8810 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8811 case X86ISD::PTEST: return "X86ISD::PTEST";
8812 case X86ISD::TESTP: return "X86ISD::TESTP";
8813 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8814 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8815 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8816 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8817 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8818 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8819 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8820 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8821 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8822 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8823 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8824 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8825 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8826 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8827 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8828 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8829 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8830 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8831 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8832 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8833 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8834 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8835 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8836 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8837 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8838 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8839 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8840 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8841 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8842 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8843 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8844 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8845 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8846 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8847 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8851 // isLegalAddressingMode - Return true if the addressing mode represented
8852 // by AM is legal for this target, for a load/store of the specified type.
8853 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8854 const Type *Ty) const {
8855 // X86 supports extremely general addressing modes.
8856 CodeModel::Model M = getTargetMachine().getCodeModel();
8857 Reloc::Model R = getTargetMachine().getRelocationModel();
8859 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8860 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8865 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8867 // If a reference to this global requires an extra load, we can't fold it.
8868 if (isGlobalStubReference(GVFlags))
8871 // If BaseGV requires a register for the PIC base, we cannot also have a
8872 // BaseReg specified.
8873 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8876 // If lower 4G is not available, then we must use rip-relative addressing.
8877 if ((M != CodeModel::Small || R != Reloc::Static) &&
8878 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8888 // These scales always work.
8893 // These scales are formed with basereg+scalereg. Only accept if there is
8898 default: // Other stuff never works.
8906 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8907 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8909 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8910 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8911 if (NumBits1 <= NumBits2)
8916 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8917 if (!VT1.isInteger() || !VT2.isInteger())
8919 unsigned NumBits1 = VT1.getSizeInBits();
8920 unsigned NumBits2 = VT2.getSizeInBits();
8921 if (NumBits1 <= NumBits2)
8926 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8927 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8928 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8931 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8932 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8933 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8936 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8937 // i16 instructions are longer (0x66 prefix) and potentially slower.
8938 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8941 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8942 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8943 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8944 /// are assumed to be legal.
8946 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8948 // Very little shuffling can be done for 64-bit vectors right now.
8949 if (VT.getSizeInBits() == 64)
8950 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8952 // FIXME: pshufb, blends, shifts.
8953 return (VT.getVectorNumElements() == 2 ||
8954 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8955 isMOVLMask(M, VT) ||
8956 isSHUFPMask(M, VT) ||
8957 isPSHUFDMask(M, VT) ||
8958 isPSHUFHWMask(M, VT) ||
8959 isPSHUFLWMask(M, VT) ||
8960 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8961 isUNPCKLMask(M, VT) ||
8962 isUNPCKHMask(M, VT) ||
8963 isUNPCKL_v_undef_Mask(M, VT) ||
8964 isUNPCKH_v_undef_Mask(M, VT));
8968 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8970 unsigned NumElts = VT.getVectorNumElements();
8971 // FIXME: This collection of masks seems suspect.
8974 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8975 return (isMOVLMask(Mask, VT) ||
8976 isCommutedMOVLMask(Mask, VT, true) ||
8977 isSHUFPMask(Mask, VT) ||
8978 isCommutedSHUFPMask(Mask, VT));
8983 //===----------------------------------------------------------------------===//
8984 // X86 Scheduler Hooks
8985 //===----------------------------------------------------------------------===//
8987 // private utility function
8989 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8990 MachineBasicBlock *MBB,
8997 TargetRegisterClass *RC,
8998 bool invSrc) const {
8999 // For the atomic bitwise operator, we generate
9002 // ld t1 = [bitinstr.addr]
9003 // op t2 = t1, [bitinstr.val]
9005 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9007 // fallthrough -->nextMBB
9008 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9009 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9010 MachineFunction::iterator MBBIter = MBB;
9013 /// First build the CFG
9014 MachineFunction *F = MBB->getParent();
9015 MachineBasicBlock *thisMBB = MBB;
9016 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9017 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9018 F->insert(MBBIter, newMBB);
9019 F->insert(MBBIter, nextMBB);
9021 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9022 nextMBB->splice(nextMBB->begin(), thisMBB,
9023 llvm::next(MachineBasicBlock::iterator(bInstr)),
9025 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9027 // Update thisMBB to fall through to newMBB
9028 thisMBB->addSuccessor(newMBB);
9030 // newMBB jumps to itself and fall through to nextMBB
9031 newMBB->addSuccessor(nextMBB);
9032 newMBB->addSuccessor(newMBB);
9034 // Insert instructions into newMBB based on incoming instruction
9035 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9036 "unexpected number of operands");
9037 DebugLoc dl = bInstr->getDebugLoc();
9038 MachineOperand& destOper = bInstr->getOperand(0);
9039 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9040 int numArgs = bInstr->getNumOperands() - 1;
9041 for (int i=0; i < numArgs; ++i)
9042 argOpers[i] = &bInstr->getOperand(i+1);
9044 // x86 address has 4 operands: base, index, scale, and displacement
9045 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9046 int valArgIndx = lastAddrIndx + 1;
9048 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9049 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9050 for (int i=0; i <= lastAddrIndx; ++i)
9051 (*MIB).addOperand(*argOpers[i]);
9053 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9055 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9060 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9061 assert((argOpers[valArgIndx]->isReg() ||
9062 argOpers[valArgIndx]->isImm()) &&
9064 if (argOpers[valArgIndx]->isReg())
9065 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9067 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9069 (*MIB).addOperand(*argOpers[valArgIndx]);
9071 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9074 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9075 for (int i=0; i <= lastAddrIndx; ++i)
9076 (*MIB).addOperand(*argOpers[i]);
9078 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9079 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9080 bInstr->memoperands_end());
9082 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9086 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9088 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9092 // private utility function: 64 bit atomics on 32 bit host.
9094 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9095 MachineBasicBlock *MBB,
9100 bool invSrc) const {
9101 // For the atomic bitwise operator, we generate
9102 // thisMBB (instructions are in pairs, except cmpxchg8b)
9103 // ld t1,t2 = [bitinstr.addr]
9105 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9106 // op t5, t6 <- out1, out2, [bitinstr.val]
9107 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9108 // mov ECX, EBX <- t5, t6
9109 // mov EAX, EDX <- t1, t2
9110 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9111 // mov t3, t4 <- EAX, EDX
9113 // result in out1, out2
9114 // fallthrough -->nextMBB
9116 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9117 const unsigned LoadOpc = X86::MOV32rm;
9118 const unsigned NotOpc = X86::NOT32r;
9119 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9120 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9121 MachineFunction::iterator MBBIter = MBB;
9124 /// First build the CFG
9125 MachineFunction *F = MBB->getParent();
9126 MachineBasicBlock *thisMBB = MBB;
9127 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9128 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9129 F->insert(MBBIter, newMBB);
9130 F->insert(MBBIter, nextMBB);
9132 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9133 nextMBB->splice(nextMBB->begin(), thisMBB,
9134 llvm::next(MachineBasicBlock::iterator(bInstr)),
9136 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9138 // Update thisMBB to fall through to newMBB
9139 thisMBB->addSuccessor(newMBB);
9141 // newMBB jumps to itself and fall through to nextMBB
9142 newMBB->addSuccessor(nextMBB);
9143 newMBB->addSuccessor(newMBB);
9145 DebugLoc dl = bInstr->getDebugLoc();
9146 // Insert instructions into newMBB based on incoming instruction
9147 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9148 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9149 "unexpected number of operands");
9150 MachineOperand& dest1Oper = bInstr->getOperand(0);
9151 MachineOperand& dest2Oper = bInstr->getOperand(1);
9152 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9153 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9154 argOpers[i] = &bInstr->getOperand(i+2);
9156 // We use some of the operands multiple times, so conservatively just
9157 // clear any kill flags that might be present.
9158 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9159 argOpers[i]->setIsKill(false);
9162 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9163 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9165 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9166 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9167 for (int i=0; i <= lastAddrIndx; ++i)
9168 (*MIB).addOperand(*argOpers[i]);
9169 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9170 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9171 // add 4 to displacement.
9172 for (int i=0; i <= lastAddrIndx-2; ++i)
9173 (*MIB).addOperand(*argOpers[i]);
9174 MachineOperand newOp3 = *(argOpers[3]);
9176 newOp3.setImm(newOp3.getImm()+4);
9178 newOp3.setOffset(newOp3.getOffset()+4);
9179 (*MIB).addOperand(newOp3);
9180 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9182 // t3/4 are defined later, at the bottom of the loop
9183 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9184 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9185 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9186 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9187 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9188 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9190 // The subsequent operations should be using the destination registers of
9191 //the PHI instructions.
9193 t1 = F->getRegInfo().createVirtualRegister(RC);
9194 t2 = F->getRegInfo().createVirtualRegister(RC);
9195 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9196 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9198 t1 = dest1Oper.getReg();
9199 t2 = dest2Oper.getReg();
9202 int valArgIndx = lastAddrIndx + 1;
9203 assert((argOpers[valArgIndx]->isReg() ||
9204 argOpers[valArgIndx]->isImm()) &&
9206 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9207 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9208 if (argOpers[valArgIndx]->isReg())
9209 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9211 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9212 if (regOpcL != X86::MOV32rr)
9214 (*MIB).addOperand(*argOpers[valArgIndx]);
9215 assert(argOpers[valArgIndx + 1]->isReg() ==
9216 argOpers[valArgIndx]->isReg());
9217 assert(argOpers[valArgIndx + 1]->isImm() ==
9218 argOpers[valArgIndx]->isImm());
9219 if (argOpers[valArgIndx + 1]->isReg())
9220 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9222 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9223 if (regOpcH != X86::MOV32rr)
9225 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9227 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9229 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9232 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9234 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9237 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9238 for (int i=0; i <= lastAddrIndx; ++i)
9239 (*MIB).addOperand(*argOpers[i]);
9241 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9242 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9243 bInstr->memoperands_end());
9245 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9246 MIB.addReg(X86::EAX);
9247 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9248 MIB.addReg(X86::EDX);
9251 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9253 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9257 // private utility function
9259 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9260 MachineBasicBlock *MBB,
9261 unsigned cmovOpc) const {
9262 // For the atomic min/max operator, we generate
9265 // ld t1 = [min/max.addr]
9266 // mov t2 = [min/max.val]
9268 // cmov[cond] t2 = t1
9270 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9272 // fallthrough -->nextMBB
9274 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9275 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9276 MachineFunction::iterator MBBIter = MBB;
9279 /// First build the CFG
9280 MachineFunction *F = MBB->getParent();
9281 MachineBasicBlock *thisMBB = MBB;
9282 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9283 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9284 F->insert(MBBIter, newMBB);
9285 F->insert(MBBIter, nextMBB);
9287 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9288 nextMBB->splice(nextMBB->begin(), thisMBB,
9289 llvm::next(MachineBasicBlock::iterator(mInstr)),
9291 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9293 // Update thisMBB to fall through to newMBB
9294 thisMBB->addSuccessor(newMBB);
9296 // newMBB jumps to newMBB and fall through to nextMBB
9297 newMBB->addSuccessor(nextMBB);
9298 newMBB->addSuccessor(newMBB);
9300 DebugLoc dl = mInstr->getDebugLoc();
9301 // Insert instructions into newMBB based on incoming instruction
9302 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9303 "unexpected number of operands");
9304 MachineOperand& destOper = mInstr->getOperand(0);
9305 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9306 int numArgs = mInstr->getNumOperands() - 1;
9307 for (int i=0; i < numArgs; ++i)
9308 argOpers[i] = &mInstr->getOperand(i+1);
9310 // x86 address has 4 operands: base, index, scale, and displacement
9311 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9312 int valArgIndx = lastAddrIndx + 1;
9314 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9315 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9316 for (int i=0; i <= lastAddrIndx; ++i)
9317 (*MIB).addOperand(*argOpers[i]);
9319 // We only support register and immediate values
9320 assert((argOpers[valArgIndx]->isReg() ||
9321 argOpers[valArgIndx]->isImm()) &&
9324 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9325 if (argOpers[valArgIndx]->isReg())
9326 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9328 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9329 (*MIB).addOperand(*argOpers[valArgIndx]);
9331 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9334 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9339 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9340 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9344 // Cmp and exchange if none has modified the memory location
9345 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9346 for (int i=0; i <= lastAddrIndx; ++i)
9347 (*MIB).addOperand(*argOpers[i]);
9349 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9350 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9351 mInstr->memoperands_end());
9353 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9354 MIB.addReg(X86::EAX);
9357 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9359 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9363 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9364 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9367 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9368 unsigned numArgs, bool memArg) const {
9370 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9371 "Target must have SSE4.2 or AVX features enabled");
9373 DebugLoc dl = MI->getDebugLoc();
9374 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9378 if (!Subtarget->hasAVX()) {
9380 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9382 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9385 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9387 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9390 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9392 for (unsigned i = 0; i < numArgs; ++i) {
9393 MachineOperand &Op = MI->getOperand(i+1);
9395 if (!(Op.isReg() && Op.isImplicit()))
9399 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9402 MI->eraseFromParent();
9408 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9410 MachineBasicBlock *MBB) const {
9411 // Emit code to save XMM registers to the stack. The ABI says that the
9412 // number of registers to save is given in %al, so it's theoretically
9413 // possible to do an indirect jump trick to avoid saving all of them,
9414 // however this code takes a simpler approach and just executes all
9415 // of the stores if %al is non-zero. It's less code, and it's probably
9416 // easier on the hardware branch predictor, and stores aren't all that
9417 // expensive anyway.
9419 // Create the new basic blocks. One block contains all the XMM stores,
9420 // and one block is the final destination regardless of whether any
9421 // stores were performed.
9422 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9423 MachineFunction *F = MBB->getParent();
9424 MachineFunction::iterator MBBIter = MBB;
9426 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9427 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9428 F->insert(MBBIter, XMMSaveMBB);
9429 F->insert(MBBIter, EndMBB);
9431 // Transfer the remainder of MBB and its successor edges to EndMBB.
9432 EndMBB->splice(EndMBB->begin(), MBB,
9433 llvm::next(MachineBasicBlock::iterator(MI)),
9435 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9437 // The original block will now fall through to the XMM save block.
9438 MBB->addSuccessor(XMMSaveMBB);
9439 // The XMMSaveMBB will fall through to the end block.
9440 XMMSaveMBB->addSuccessor(EndMBB);
9442 // Now add the instructions.
9443 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9444 DebugLoc DL = MI->getDebugLoc();
9446 unsigned CountReg = MI->getOperand(0).getReg();
9447 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9448 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9450 if (!Subtarget->isTargetWin64()) {
9451 // If %al is 0, branch around the XMM save block.
9452 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9453 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9454 MBB->addSuccessor(EndMBB);
9457 // In the XMM save block, save all the XMM argument registers.
9458 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9459 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9460 MachineMemOperand *MMO =
9461 F->getMachineMemOperand(
9462 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
9463 MachineMemOperand::MOStore,
9464 /*Size=*/16, /*Align=*/16);
9465 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9466 .addFrameIndex(RegSaveFrameIndex)
9467 .addImm(/*Scale=*/1)
9468 .addReg(/*IndexReg=*/0)
9469 .addImm(/*Disp=*/Offset)
9470 .addReg(/*Segment=*/0)
9471 .addReg(MI->getOperand(i).getReg())
9472 .addMemOperand(MMO);
9475 MI->eraseFromParent(); // The pseudo instruction is gone now.
9481 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9482 MachineBasicBlock *BB) const {
9483 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9484 DebugLoc DL = MI->getDebugLoc();
9486 // To "insert" a SELECT_CC instruction, we actually have to insert the
9487 // diamond control-flow pattern. The incoming instruction knows the
9488 // destination vreg to set, the condition code register to branch on, the
9489 // true/false values to select between, and a branch opcode to use.
9490 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9491 MachineFunction::iterator It = BB;
9497 // cmpTY ccX, r1, r2
9499 // fallthrough --> copy0MBB
9500 MachineBasicBlock *thisMBB = BB;
9501 MachineFunction *F = BB->getParent();
9502 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9503 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9504 F->insert(It, copy0MBB);
9505 F->insert(It, sinkMBB);
9507 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9508 // live into the sink and copy blocks.
9509 const MachineFunction *MF = BB->getParent();
9510 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9511 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9513 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9514 const MachineOperand &MO = MI->getOperand(I);
9515 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9516 unsigned Reg = MO.getReg();
9517 if (Reg != X86::EFLAGS) continue;
9518 copy0MBB->addLiveIn(Reg);
9519 sinkMBB->addLiveIn(Reg);
9522 // Transfer the remainder of BB and its successor edges to sinkMBB.
9523 sinkMBB->splice(sinkMBB->begin(), BB,
9524 llvm::next(MachineBasicBlock::iterator(MI)),
9526 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9528 // Add the true and fallthrough blocks as its successors.
9529 BB->addSuccessor(copy0MBB);
9530 BB->addSuccessor(sinkMBB);
9532 // Create the conditional branch instruction.
9534 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9535 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9538 // %FalseValue = ...
9539 // # fallthrough to sinkMBB
9540 copy0MBB->addSuccessor(sinkMBB);
9543 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9545 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9546 TII->get(X86::PHI), MI->getOperand(0).getReg())
9547 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9548 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9550 MI->eraseFromParent(); // The pseudo instruction is gone now.
9555 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9556 MachineBasicBlock *BB) const {
9557 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9558 DebugLoc DL = MI->getDebugLoc();
9560 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9561 // non-trivial part is impdef of ESP.
9562 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9565 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9566 .addExternalSymbol("_alloca")
9567 .addReg(X86::EAX, RegState::Implicit)
9568 .addReg(X86::ESP, RegState::Implicit)
9569 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9570 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9571 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9573 MI->eraseFromParent(); // The pseudo instruction is gone now.
9578 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9579 MachineBasicBlock *BB) const {
9580 // This is pretty easy. We're taking the value that we received from
9581 // our load from the relocation, sticking it in either RDI (x86-64)
9582 // or EAX and doing an indirect call. The return value will then
9583 // be in the normal return register.
9584 const X86InstrInfo *TII
9585 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9586 DebugLoc DL = MI->getDebugLoc();
9587 MachineFunction *F = BB->getParent();
9589 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
9590 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9592 if (Subtarget->is64Bit()) {
9593 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9594 TII->get(X86::MOV64rm), X86::RDI)
9596 .addImm(0).addReg(0)
9597 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9598 MI->getOperand(3).getTargetFlags())
9600 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
9601 addDirectMem(MIB, X86::RDI);
9602 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9603 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9604 TII->get(X86::MOV32rm), X86::EAX)
9606 .addImm(0).addReg(0)
9607 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9608 MI->getOperand(3).getTargetFlags())
9610 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9611 addDirectMem(MIB, X86::EAX);
9613 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9614 TII->get(X86::MOV32rm), X86::EAX)
9615 .addReg(TII->getGlobalBaseReg(F))
9616 .addImm(0).addReg(0)
9617 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9618 MI->getOperand(3).getTargetFlags())
9620 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9621 addDirectMem(MIB, X86::EAX);
9624 MI->eraseFromParent(); // The pseudo instruction is gone now.
9629 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9630 MachineBasicBlock *BB) const {
9631 switch (MI->getOpcode()) {
9632 default: assert(false && "Unexpected instr type to insert");
9633 case X86::MINGW_ALLOCA:
9634 return EmitLoweredMingwAlloca(MI, BB);
9635 case X86::TLSCall_32:
9636 case X86::TLSCall_64:
9637 return EmitLoweredTLSCall(MI, BB);
9639 case X86::CMOV_FR32:
9640 case X86::CMOV_FR64:
9641 case X86::CMOV_V4F32:
9642 case X86::CMOV_V2F64:
9643 case X86::CMOV_V2I64:
9644 case X86::CMOV_GR16:
9645 case X86::CMOV_GR32:
9646 case X86::CMOV_RFP32:
9647 case X86::CMOV_RFP64:
9648 case X86::CMOV_RFP80:
9649 return EmitLoweredSelect(MI, BB);
9651 case X86::FP32_TO_INT16_IN_MEM:
9652 case X86::FP32_TO_INT32_IN_MEM:
9653 case X86::FP32_TO_INT64_IN_MEM:
9654 case X86::FP64_TO_INT16_IN_MEM:
9655 case X86::FP64_TO_INT32_IN_MEM:
9656 case X86::FP64_TO_INT64_IN_MEM:
9657 case X86::FP80_TO_INT16_IN_MEM:
9658 case X86::FP80_TO_INT32_IN_MEM:
9659 case X86::FP80_TO_INT64_IN_MEM: {
9660 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9661 DebugLoc DL = MI->getDebugLoc();
9663 // Change the floating point control register to use "round towards zero"
9664 // mode when truncating to an integer value.
9665 MachineFunction *F = BB->getParent();
9666 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9667 addFrameReference(BuildMI(*BB, MI, DL,
9668 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9670 // Load the old value of the high byte of the control word...
9672 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9673 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9676 // Set the high part to be round to zero...
9677 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9680 // Reload the modified control word now...
9681 addFrameReference(BuildMI(*BB, MI, DL,
9682 TII->get(X86::FLDCW16m)), CWFrameIdx);
9684 // Restore the memory image of control word to original value
9685 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9688 // Get the X86 opcode to use.
9690 switch (MI->getOpcode()) {
9691 default: llvm_unreachable("illegal opcode!");
9692 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9693 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9694 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9695 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9696 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9697 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9698 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9699 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9700 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9704 MachineOperand &Op = MI->getOperand(0);
9706 AM.BaseType = X86AddressMode::RegBase;
9707 AM.Base.Reg = Op.getReg();
9709 AM.BaseType = X86AddressMode::FrameIndexBase;
9710 AM.Base.FrameIndex = Op.getIndex();
9712 Op = MI->getOperand(1);
9714 AM.Scale = Op.getImm();
9715 Op = MI->getOperand(2);
9717 AM.IndexReg = Op.getImm();
9718 Op = MI->getOperand(3);
9719 if (Op.isGlobal()) {
9720 AM.GV = Op.getGlobal();
9722 AM.Disp = Op.getImm();
9724 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9725 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9727 // Reload the original control word now.
9728 addFrameReference(BuildMI(*BB, MI, DL,
9729 TII->get(X86::FLDCW16m)), CWFrameIdx);
9731 MI->eraseFromParent(); // The pseudo instruction is gone now.
9734 // String/text processing lowering.
9735 case X86::PCMPISTRM128REG:
9736 case X86::VPCMPISTRM128REG:
9737 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9738 case X86::PCMPISTRM128MEM:
9739 case X86::VPCMPISTRM128MEM:
9740 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9741 case X86::PCMPESTRM128REG:
9742 case X86::VPCMPESTRM128REG:
9743 return EmitPCMP(MI, BB, 5, false /* in mem */);
9744 case X86::PCMPESTRM128MEM:
9745 case X86::VPCMPESTRM128MEM:
9746 return EmitPCMP(MI, BB, 5, true /* in mem */);
9749 case X86::ATOMAND32:
9750 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9751 X86::AND32ri, X86::MOV32rm,
9753 X86::NOT32r, X86::EAX,
9754 X86::GR32RegisterClass);
9756 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9757 X86::OR32ri, X86::MOV32rm,
9759 X86::NOT32r, X86::EAX,
9760 X86::GR32RegisterClass);
9761 case X86::ATOMXOR32:
9762 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9763 X86::XOR32ri, X86::MOV32rm,
9765 X86::NOT32r, X86::EAX,
9766 X86::GR32RegisterClass);
9767 case X86::ATOMNAND32:
9768 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9769 X86::AND32ri, X86::MOV32rm,
9771 X86::NOT32r, X86::EAX,
9772 X86::GR32RegisterClass, true);
9773 case X86::ATOMMIN32:
9774 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9775 case X86::ATOMMAX32:
9776 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9777 case X86::ATOMUMIN32:
9778 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9779 case X86::ATOMUMAX32:
9780 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9782 case X86::ATOMAND16:
9783 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9784 X86::AND16ri, X86::MOV16rm,
9786 X86::NOT16r, X86::AX,
9787 X86::GR16RegisterClass);
9789 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9790 X86::OR16ri, X86::MOV16rm,
9792 X86::NOT16r, X86::AX,
9793 X86::GR16RegisterClass);
9794 case X86::ATOMXOR16:
9795 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9796 X86::XOR16ri, X86::MOV16rm,
9798 X86::NOT16r, X86::AX,
9799 X86::GR16RegisterClass);
9800 case X86::ATOMNAND16:
9801 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9802 X86::AND16ri, X86::MOV16rm,
9804 X86::NOT16r, X86::AX,
9805 X86::GR16RegisterClass, true);
9806 case X86::ATOMMIN16:
9807 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9808 case X86::ATOMMAX16:
9809 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9810 case X86::ATOMUMIN16:
9811 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9812 case X86::ATOMUMAX16:
9813 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9816 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9817 X86::AND8ri, X86::MOV8rm,
9819 X86::NOT8r, X86::AL,
9820 X86::GR8RegisterClass);
9822 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9823 X86::OR8ri, X86::MOV8rm,
9825 X86::NOT8r, X86::AL,
9826 X86::GR8RegisterClass);
9828 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9829 X86::XOR8ri, X86::MOV8rm,
9831 X86::NOT8r, X86::AL,
9832 X86::GR8RegisterClass);
9833 case X86::ATOMNAND8:
9834 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9835 X86::AND8ri, X86::MOV8rm,
9837 X86::NOT8r, X86::AL,
9838 X86::GR8RegisterClass, true);
9839 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9840 // This group is for 64-bit host.
9841 case X86::ATOMAND64:
9842 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9843 X86::AND64ri32, X86::MOV64rm,
9845 X86::NOT64r, X86::RAX,
9846 X86::GR64RegisterClass);
9848 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9849 X86::OR64ri32, X86::MOV64rm,
9851 X86::NOT64r, X86::RAX,
9852 X86::GR64RegisterClass);
9853 case X86::ATOMXOR64:
9854 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9855 X86::XOR64ri32, X86::MOV64rm,
9857 X86::NOT64r, X86::RAX,
9858 X86::GR64RegisterClass);
9859 case X86::ATOMNAND64:
9860 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9861 X86::AND64ri32, X86::MOV64rm,
9863 X86::NOT64r, X86::RAX,
9864 X86::GR64RegisterClass, true);
9865 case X86::ATOMMIN64:
9866 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9867 case X86::ATOMMAX64:
9868 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9869 case X86::ATOMUMIN64:
9870 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9871 case X86::ATOMUMAX64:
9872 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9874 // This group does 64-bit operations on a 32-bit host.
9875 case X86::ATOMAND6432:
9876 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9877 X86::AND32rr, X86::AND32rr,
9878 X86::AND32ri, X86::AND32ri,
9880 case X86::ATOMOR6432:
9881 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9882 X86::OR32rr, X86::OR32rr,
9883 X86::OR32ri, X86::OR32ri,
9885 case X86::ATOMXOR6432:
9886 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9887 X86::XOR32rr, X86::XOR32rr,
9888 X86::XOR32ri, X86::XOR32ri,
9890 case X86::ATOMNAND6432:
9891 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9892 X86::AND32rr, X86::AND32rr,
9893 X86::AND32ri, X86::AND32ri,
9895 case X86::ATOMADD6432:
9896 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9897 X86::ADD32rr, X86::ADC32rr,
9898 X86::ADD32ri, X86::ADC32ri,
9900 case X86::ATOMSUB6432:
9901 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9902 X86::SUB32rr, X86::SBB32rr,
9903 X86::SUB32ri, X86::SBB32ri,
9905 case X86::ATOMSWAP6432:
9906 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9907 X86::MOV32rr, X86::MOV32rr,
9908 X86::MOV32ri, X86::MOV32ri,
9910 case X86::VASTART_SAVE_XMM_REGS:
9911 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9915 //===----------------------------------------------------------------------===//
9916 // X86 Optimization Hooks
9917 //===----------------------------------------------------------------------===//
9919 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9923 const SelectionDAG &DAG,
9924 unsigned Depth) const {
9925 unsigned Opc = Op.getOpcode();
9926 assert((Opc >= ISD::BUILTIN_OP_END ||
9927 Opc == ISD::INTRINSIC_WO_CHAIN ||
9928 Opc == ISD::INTRINSIC_W_CHAIN ||
9929 Opc == ISD::INTRINSIC_VOID) &&
9930 "Should use MaskedValueIsZero if you don't know whether Op"
9931 " is a target node!");
9933 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9945 // These nodes' second result is a boolean.
9946 if (Op.getResNo() == 0)
9950 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9951 Mask.getBitWidth() - 1);
9956 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
9957 unsigned Depth) const {
9958 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
9959 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
9960 return Op.getValueType().getScalarType().getSizeInBits();
9966 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9967 /// node is a GlobalAddress + offset.
9968 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9969 const GlobalValue* &GA,
9970 int64_t &Offset) const {
9971 if (N->getOpcode() == X86ISD::Wrapper) {
9972 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9973 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9974 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9978 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9981 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9982 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9983 /// if the load addresses are consecutive, non-overlapping, and in the right
9985 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9986 const TargetLowering &TLI) {
9987 DebugLoc dl = N->getDebugLoc();
9988 EVT VT = N->getValueType(0);
9990 if (VT.getSizeInBits() != 128)
9993 SmallVector<SDValue, 16> Elts;
9994 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9995 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
9997 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10000 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10001 /// generation and convert it from being a bunch of shuffles and extracts
10002 /// to a simple store and scalar loads to extract the elements.
10003 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10004 const TargetLowering &TLI) {
10005 SDValue InputVector = N->getOperand(0);
10007 // Only operate on vectors of 4 elements, where the alternative shuffling
10008 // gets to be more expensive.
10009 if (InputVector.getValueType() != MVT::v4i32)
10012 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
10013 // single use which is a sign-extend or zero-extend, and all elements are
10015 SmallVector<SDNode *, 4> Uses;
10016 unsigned ExtractedElements = 0;
10017 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
10018 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
10019 if (UI.getUse().getResNo() != InputVector.getResNo())
10022 SDNode *Extract = *UI;
10023 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10026 if (Extract->getValueType(0) != MVT::i32)
10028 if (!Extract->hasOneUse())
10030 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
10031 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
10033 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
10036 // Record which element was extracted.
10037 ExtractedElements |=
10038 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
10040 Uses.push_back(Extract);
10043 // If not all the elements were used, this may not be worthwhile.
10044 if (ExtractedElements != 15)
10047 // Ok, we've now decided to do the transformation.
10048 DebugLoc dl = InputVector.getDebugLoc();
10050 // Store the value to a temporary stack slot.
10051 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
10052 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
10053 MachinePointerInfo(), false, false, 0);
10055 // Replace each use (extract) with a load of the appropriate element.
10056 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
10057 UE = Uses.end(); UI != UE; ++UI) {
10058 SDNode *Extract = *UI;
10060 // Compute the element's address.
10061 SDValue Idx = Extract->getOperand(1);
10063 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
10064 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
10065 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
10067 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
10068 StackPtr, OffsetVal);
10070 // Load the scalar.
10071 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
10072 ScalarAddr, MachinePointerInfo(),
10075 // Replace the exact with the load.
10076 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
10079 // The replacement was made in place; don't return anything.
10083 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
10084 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
10085 const X86Subtarget *Subtarget) {
10086 DebugLoc DL = N->getDebugLoc();
10087 SDValue Cond = N->getOperand(0);
10088 // Get the LHS/RHS of the select.
10089 SDValue LHS = N->getOperand(1);
10090 SDValue RHS = N->getOperand(2);
10092 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
10093 // instructions match the semantics of the common C idiom x<y?x:y but not
10094 // x<=y?x:y, because of how they handle negative zero (which can be
10095 // ignored in unsafe-math mode).
10096 if (Subtarget->hasSSE2() &&
10097 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
10098 Cond.getOpcode() == ISD::SETCC) {
10099 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
10101 unsigned Opcode = 0;
10102 // Check for x CC y ? x : y.
10103 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
10104 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
10108 // Converting this to a min would handle NaNs incorrectly, and swapping
10109 // the operands would cause it to handle comparisons between positive
10110 // and negative zero incorrectly.
10111 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10112 if (!UnsafeFPMath &&
10113 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10115 std::swap(LHS, RHS);
10117 Opcode = X86ISD::FMIN;
10120 // Converting this to a min would handle comparisons between positive
10121 // and negative zero incorrectly.
10122 if (!UnsafeFPMath &&
10123 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
10125 Opcode = X86ISD::FMIN;
10128 // Converting this to a min would handle both negative zeros and NaNs
10129 // incorrectly, but we can swap the operands to fix both.
10130 std::swap(LHS, RHS);
10134 Opcode = X86ISD::FMIN;
10138 // Converting this to a max would handle comparisons between positive
10139 // and negative zero incorrectly.
10140 if (!UnsafeFPMath &&
10141 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
10143 Opcode = X86ISD::FMAX;
10146 // Converting this to a max would handle NaNs incorrectly, and swapping
10147 // the operands would cause it to handle comparisons between positive
10148 // and negative zero incorrectly.
10149 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10150 if (!UnsafeFPMath &&
10151 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10153 std::swap(LHS, RHS);
10155 Opcode = X86ISD::FMAX;
10158 // Converting this to a max would handle both negative zeros and NaNs
10159 // incorrectly, but we can swap the operands to fix both.
10160 std::swap(LHS, RHS);
10164 Opcode = X86ISD::FMAX;
10167 // Check for x CC y ? y : x -- a min/max with reversed arms.
10168 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10169 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10173 // Converting this to a min would handle comparisons between positive
10174 // and negative zero incorrectly, and swapping the operands would
10175 // cause it to handle NaNs incorrectly.
10176 if (!UnsafeFPMath &&
10177 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10178 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10180 std::swap(LHS, RHS);
10182 Opcode = X86ISD::FMIN;
10185 // Converting this to a min would handle NaNs incorrectly.
10186 if (!UnsafeFPMath &&
10187 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10189 Opcode = X86ISD::FMIN;
10192 // Converting this to a min would handle both negative zeros and NaNs
10193 // incorrectly, but we can swap the operands to fix both.
10194 std::swap(LHS, RHS);
10198 Opcode = X86ISD::FMIN;
10202 // Converting this to a max would handle NaNs incorrectly.
10203 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10205 Opcode = X86ISD::FMAX;
10208 // Converting this to a max would handle comparisons between positive
10209 // and negative zero incorrectly, and swapping the operands would
10210 // cause it to handle NaNs incorrectly.
10211 if (!UnsafeFPMath &&
10212 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10213 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10215 std::swap(LHS, RHS);
10217 Opcode = X86ISD::FMAX;
10220 // Converting this to a max would handle both negative zeros and NaNs
10221 // incorrectly, but we can swap the operands to fix both.
10222 std::swap(LHS, RHS);
10226 Opcode = X86ISD::FMAX;
10232 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10235 // If this is a select between two integer constants, try to do some
10237 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10238 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10239 // Don't do this for crazy integer types.
10240 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10241 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10242 // so that TrueC (the true value) is larger than FalseC.
10243 bool NeedsCondInvert = false;
10245 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10246 // Efficiently invertible.
10247 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10248 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10249 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10250 NeedsCondInvert = true;
10251 std::swap(TrueC, FalseC);
10254 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10255 if (FalseC->getAPIntValue() == 0 &&
10256 TrueC->getAPIntValue().isPowerOf2()) {
10257 if (NeedsCondInvert) // Invert the condition if needed.
10258 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10259 DAG.getConstant(1, Cond.getValueType()));
10261 // Zero extend the condition if needed.
10262 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10264 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10265 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10266 DAG.getConstant(ShAmt, MVT::i8));
10269 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10270 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10271 if (NeedsCondInvert) // Invert the condition if needed.
10272 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10273 DAG.getConstant(1, Cond.getValueType()));
10275 // Zero extend the condition if needed.
10276 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10277 FalseC->getValueType(0), Cond);
10278 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10279 SDValue(FalseC, 0));
10282 // Optimize cases that will turn into an LEA instruction. This requires
10283 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10284 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10285 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10286 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10288 bool isFastMultiplier = false;
10290 switch ((unsigned char)Diff) {
10292 case 1: // result = add base, cond
10293 case 2: // result = lea base( , cond*2)
10294 case 3: // result = lea base(cond, cond*2)
10295 case 4: // result = lea base( , cond*4)
10296 case 5: // result = lea base(cond, cond*4)
10297 case 8: // result = lea base( , cond*8)
10298 case 9: // result = lea base(cond, cond*8)
10299 isFastMultiplier = true;
10304 if (isFastMultiplier) {
10305 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10306 if (NeedsCondInvert) // Invert the condition if needed.
10307 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10308 DAG.getConstant(1, Cond.getValueType()));
10310 // Zero extend the condition if needed.
10311 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10313 // Scale the condition by the difference.
10315 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10316 DAG.getConstant(Diff, Cond.getValueType()));
10318 // Add the base if non-zero.
10319 if (FalseC->getAPIntValue() != 0)
10320 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10321 SDValue(FalseC, 0));
10331 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10332 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10333 TargetLowering::DAGCombinerInfo &DCI) {
10334 DebugLoc DL = N->getDebugLoc();
10336 // If the flag operand isn't dead, don't touch this CMOV.
10337 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10340 // If this is a select between two integer constants, try to do some
10341 // optimizations. Note that the operands are ordered the opposite of SELECT
10343 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10344 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10345 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10346 // larger than FalseC (the false value).
10347 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10349 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10350 CC = X86::GetOppositeBranchCondition(CC);
10351 std::swap(TrueC, FalseC);
10354 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10355 // This is efficient for any integer data type (including i8/i16) and
10357 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10358 SDValue Cond = N->getOperand(3);
10359 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10360 DAG.getConstant(CC, MVT::i8), Cond);
10362 // Zero extend the condition if needed.
10363 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10365 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10366 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10367 DAG.getConstant(ShAmt, MVT::i8));
10368 if (N->getNumValues() == 2) // Dead flag value?
10369 return DCI.CombineTo(N, Cond, SDValue());
10373 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10374 // for any integer data type, including i8/i16.
10375 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10376 SDValue Cond = N->getOperand(3);
10377 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10378 DAG.getConstant(CC, MVT::i8), Cond);
10380 // Zero extend the condition if needed.
10381 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10382 FalseC->getValueType(0), Cond);
10383 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10384 SDValue(FalseC, 0));
10386 if (N->getNumValues() == 2) // Dead flag value?
10387 return DCI.CombineTo(N, Cond, SDValue());
10391 // Optimize cases that will turn into an LEA instruction. This requires
10392 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10393 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10394 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10395 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10397 bool isFastMultiplier = false;
10399 switch ((unsigned char)Diff) {
10401 case 1: // result = add base, cond
10402 case 2: // result = lea base( , cond*2)
10403 case 3: // result = lea base(cond, cond*2)
10404 case 4: // result = lea base( , cond*4)
10405 case 5: // result = lea base(cond, cond*4)
10406 case 8: // result = lea base( , cond*8)
10407 case 9: // result = lea base(cond, cond*8)
10408 isFastMultiplier = true;
10413 if (isFastMultiplier) {
10414 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10415 SDValue Cond = N->getOperand(3);
10416 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10417 DAG.getConstant(CC, MVT::i8), Cond);
10418 // Zero extend the condition if needed.
10419 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10421 // Scale the condition by the difference.
10423 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10424 DAG.getConstant(Diff, Cond.getValueType()));
10426 // Add the base if non-zero.
10427 if (FalseC->getAPIntValue() != 0)
10428 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10429 SDValue(FalseC, 0));
10430 if (N->getNumValues() == 2) // Dead flag value?
10431 return DCI.CombineTo(N, Cond, SDValue());
10441 /// PerformMulCombine - Optimize a single multiply with constant into two
10442 /// in order to implement it with two cheaper instructions, e.g.
10443 /// LEA + SHL, LEA + LEA.
10444 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10445 TargetLowering::DAGCombinerInfo &DCI) {
10446 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10449 EVT VT = N->getValueType(0);
10450 if (VT != MVT::i64)
10453 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10456 uint64_t MulAmt = C->getZExtValue();
10457 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10460 uint64_t MulAmt1 = 0;
10461 uint64_t MulAmt2 = 0;
10462 if ((MulAmt % 9) == 0) {
10464 MulAmt2 = MulAmt / 9;
10465 } else if ((MulAmt % 5) == 0) {
10467 MulAmt2 = MulAmt / 5;
10468 } else if ((MulAmt % 3) == 0) {
10470 MulAmt2 = MulAmt / 3;
10473 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10474 DebugLoc DL = N->getDebugLoc();
10476 if (isPowerOf2_64(MulAmt2) &&
10477 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10478 // If second multiplifer is pow2, issue it first. We want the multiply by
10479 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10481 std::swap(MulAmt1, MulAmt2);
10484 if (isPowerOf2_64(MulAmt1))
10485 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10486 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10488 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10489 DAG.getConstant(MulAmt1, VT));
10491 if (isPowerOf2_64(MulAmt2))
10492 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10493 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10495 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10496 DAG.getConstant(MulAmt2, VT));
10498 // Do not add new nodes to DAG combiner worklist.
10499 DCI.CombineTo(N, NewMul, false);
10504 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10505 SDValue N0 = N->getOperand(0);
10506 SDValue N1 = N->getOperand(1);
10507 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10508 EVT VT = N0.getValueType();
10510 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10511 // since the result of setcc_c is all zero's or all ones.
10512 if (N1C && N0.getOpcode() == ISD::AND &&
10513 N0.getOperand(1).getOpcode() == ISD::Constant) {
10514 SDValue N00 = N0.getOperand(0);
10515 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10516 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10517 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10518 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10519 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10520 APInt ShAmt = N1C->getAPIntValue();
10521 Mask = Mask.shl(ShAmt);
10523 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10524 N00, DAG.getConstant(Mask, VT));
10531 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10533 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10534 const X86Subtarget *Subtarget) {
10535 EVT VT = N->getValueType(0);
10536 if (!VT.isVector() && VT.isInteger() &&
10537 N->getOpcode() == ISD::SHL)
10538 return PerformSHLCombine(N, DAG);
10540 // On X86 with SSE2 support, we can transform this to a vector shift if
10541 // all elements are shifted by the same amount. We can't do this in legalize
10542 // because the a constant vector is typically transformed to a constant pool
10543 // so we have no knowledge of the shift amount.
10544 if (!Subtarget->hasSSE2())
10547 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10550 SDValue ShAmtOp = N->getOperand(1);
10551 EVT EltVT = VT.getVectorElementType();
10552 DebugLoc DL = N->getDebugLoc();
10553 SDValue BaseShAmt = SDValue();
10554 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10555 unsigned NumElts = VT.getVectorNumElements();
10557 for (; i != NumElts; ++i) {
10558 SDValue Arg = ShAmtOp.getOperand(i);
10559 if (Arg.getOpcode() == ISD::UNDEF) continue;
10563 for (; i != NumElts; ++i) {
10564 SDValue Arg = ShAmtOp.getOperand(i);
10565 if (Arg.getOpcode() == ISD::UNDEF) continue;
10566 if (Arg != BaseShAmt) {
10570 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10571 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10572 SDValue InVec = ShAmtOp.getOperand(0);
10573 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10574 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10576 for (; i != NumElts; ++i) {
10577 SDValue Arg = InVec.getOperand(i);
10578 if (Arg.getOpcode() == ISD::UNDEF) continue;
10582 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10583 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10584 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10585 if (C->getZExtValue() == SplatIdx)
10586 BaseShAmt = InVec.getOperand(1);
10589 if (BaseShAmt.getNode() == 0)
10590 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10591 DAG.getIntPtrConstant(0));
10595 // The shift amount is an i32.
10596 if (EltVT.bitsGT(MVT::i32))
10597 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10598 else if (EltVT.bitsLT(MVT::i32))
10599 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10601 // The shift amount is identical so we can do a vector shift.
10602 SDValue ValOp = N->getOperand(0);
10603 switch (N->getOpcode()) {
10605 llvm_unreachable("Unknown shift opcode!");
10608 if (VT == MVT::v2i64)
10609 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10610 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10612 if (VT == MVT::v4i32)
10613 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10614 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10616 if (VT == MVT::v8i16)
10617 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10618 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10622 if (VT == MVT::v4i32)
10623 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10624 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10626 if (VT == MVT::v8i16)
10627 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10628 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10632 if (VT == MVT::v2i64)
10633 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10634 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10636 if (VT == MVT::v4i32)
10637 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10638 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10640 if (VT == MVT::v8i16)
10641 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10642 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10649 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10650 TargetLowering::DAGCombinerInfo &DCI,
10651 const X86Subtarget *Subtarget) {
10652 if (DCI.isBeforeLegalizeOps())
10655 EVT VT = N->getValueType(0);
10656 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10659 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10660 SDValue N0 = N->getOperand(0);
10661 SDValue N1 = N->getOperand(1);
10662 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10664 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10666 if (!N0.hasOneUse() || !N1.hasOneUse())
10669 SDValue ShAmt0 = N0.getOperand(1);
10670 if (ShAmt0.getValueType() != MVT::i8)
10672 SDValue ShAmt1 = N1.getOperand(1);
10673 if (ShAmt1.getValueType() != MVT::i8)
10675 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10676 ShAmt0 = ShAmt0.getOperand(0);
10677 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10678 ShAmt1 = ShAmt1.getOperand(0);
10680 DebugLoc DL = N->getDebugLoc();
10681 unsigned Opc = X86ISD::SHLD;
10682 SDValue Op0 = N0.getOperand(0);
10683 SDValue Op1 = N1.getOperand(0);
10684 if (ShAmt0.getOpcode() == ISD::SUB) {
10685 Opc = X86ISD::SHRD;
10686 std::swap(Op0, Op1);
10687 std::swap(ShAmt0, ShAmt1);
10690 unsigned Bits = VT.getSizeInBits();
10691 if (ShAmt1.getOpcode() == ISD::SUB) {
10692 SDValue Sum = ShAmt1.getOperand(0);
10693 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10694 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10695 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10696 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10697 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10698 return DAG.getNode(Opc, DL, VT,
10700 DAG.getNode(ISD::TRUNCATE, DL,
10703 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10704 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10706 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10707 return DAG.getNode(Opc, DL, VT,
10708 N0.getOperand(0), N1.getOperand(0),
10709 DAG.getNode(ISD::TRUNCATE, DL,
10716 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10717 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10718 const X86Subtarget *Subtarget) {
10719 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10720 // the FP state in cases where an emms may be missing.
10721 // A preferable solution to the general problem is to figure out the right
10722 // places to insert EMMS. This qualifies as a quick hack.
10724 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10725 StoreSDNode *St = cast<StoreSDNode>(N);
10726 EVT VT = St->getValue().getValueType();
10727 if (VT.getSizeInBits() != 64)
10730 const Function *F = DAG.getMachineFunction().getFunction();
10731 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10732 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10733 && Subtarget->hasSSE2();
10734 if ((VT.isVector() ||
10735 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10736 isa<LoadSDNode>(St->getValue()) &&
10737 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10738 St->getChain().hasOneUse() && !St->isVolatile()) {
10739 SDNode* LdVal = St->getValue().getNode();
10740 LoadSDNode *Ld = 0;
10741 int TokenFactorIndex = -1;
10742 SmallVector<SDValue, 8> Ops;
10743 SDNode* ChainVal = St->getChain().getNode();
10744 // Must be a store of a load. We currently handle two cases: the load
10745 // is a direct child, and it's under an intervening TokenFactor. It is
10746 // possible to dig deeper under nested TokenFactors.
10747 if (ChainVal == LdVal)
10748 Ld = cast<LoadSDNode>(St->getChain());
10749 else if (St->getValue().hasOneUse() &&
10750 ChainVal->getOpcode() == ISD::TokenFactor) {
10751 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10752 if (ChainVal->getOperand(i).getNode() == LdVal) {
10753 TokenFactorIndex = i;
10754 Ld = cast<LoadSDNode>(St->getValue());
10756 Ops.push_back(ChainVal->getOperand(i));
10760 if (!Ld || !ISD::isNormalLoad(Ld))
10763 // If this is not the MMX case, i.e. we are just turning i64 load/store
10764 // into f64 load/store, avoid the transformation if there are multiple
10765 // uses of the loaded value.
10766 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10769 DebugLoc LdDL = Ld->getDebugLoc();
10770 DebugLoc StDL = N->getDebugLoc();
10771 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10772 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10774 if (Subtarget->is64Bit() || F64IsLegal) {
10775 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10776 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
10777 Ld->getPointerInfo(), Ld->isVolatile(),
10778 Ld->isNonTemporal(), Ld->getAlignment());
10779 SDValue NewChain = NewLd.getValue(1);
10780 if (TokenFactorIndex != -1) {
10781 Ops.push_back(NewChain);
10782 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10785 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10786 St->getPointerInfo(),
10787 St->isVolatile(), St->isNonTemporal(),
10788 St->getAlignment());
10791 // Otherwise, lower to two pairs of 32-bit loads / stores.
10792 SDValue LoAddr = Ld->getBasePtr();
10793 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10794 DAG.getConstant(4, MVT::i32));
10796 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10797 Ld->getPointerInfo(),
10798 Ld->isVolatile(), Ld->isNonTemporal(),
10799 Ld->getAlignment());
10800 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10801 Ld->getPointerInfo().getWithOffset(4),
10802 Ld->isVolatile(), Ld->isNonTemporal(),
10803 MinAlign(Ld->getAlignment(), 4));
10805 SDValue NewChain = LoLd.getValue(1);
10806 if (TokenFactorIndex != -1) {
10807 Ops.push_back(LoLd);
10808 Ops.push_back(HiLd);
10809 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10813 LoAddr = St->getBasePtr();
10814 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10815 DAG.getConstant(4, MVT::i32));
10817 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10818 St->getPointerInfo(),
10819 St->isVolatile(), St->isNonTemporal(),
10820 St->getAlignment());
10821 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10822 St->getPointerInfo().getWithOffset(4),
10824 St->isNonTemporal(),
10825 MinAlign(St->getAlignment(), 4));
10826 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10831 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10832 /// X86ISD::FXOR nodes.
10833 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10834 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10835 // F[X]OR(0.0, x) -> x
10836 // F[X]OR(x, 0.0) -> x
10837 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10838 if (C->getValueAPF().isPosZero())
10839 return N->getOperand(1);
10840 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10841 if (C->getValueAPF().isPosZero())
10842 return N->getOperand(0);
10846 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10847 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10848 // FAND(0.0, x) -> 0.0
10849 // FAND(x, 0.0) -> 0.0
10850 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10851 if (C->getValueAPF().isPosZero())
10852 return N->getOperand(0);
10853 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10854 if (C->getValueAPF().isPosZero())
10855 return N->getOperand(1);
10859 static SDValue PerformBTCombine(SDNode *N,
10861 TargetLowering::DAGCombinerInfo &DCI) {
10862 // BT ignores high bits in the bit index operand.
10863 SDValue Op1 = N->getOperand(1);
10864 if (Op1.hasOneUse()) {
10865 unsigned BitWidth = Op1.getValueSizeInBits();
10866 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10867 APInt KnownZero, KnownOne;
10868 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10869 !DCI.isBeforeLegalizeOps());
10870 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10871 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10872 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10873 DCI.CommitTargetLoweringOpt(TLO);
10878 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10879 SDValue Op = N->getOperand(0);
10880 if (Op.getOpcode() == ISD::BIT_CONVERT)
10881 Op = Op.getOperand(0);
10882 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10883 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10884 VT.getVectorElementType().getSizeInBits() ==
10885 OpVT.getVectorElementType().getSizeInBits()) {
10886 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10891 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10892 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10893 // (and (i32 x86isd::setcc_carry), 1)
10894 // This eliminates the zext. This transformation is necessary because
10895 // ISD::SETCC is always legalized to i8.
10896 DebugLoc dl = N->getDebugLoc();
10897 SDValue N0 = N->getOperand(0);
10898 EVT VT = N->getValueType(0);
10899 if (N0.getOpcode() == ISD::AND &&
10901 N0.getOperand(0).hasOneUse()) {
10902 SDValue N00 = N0.getOperand(0);
10903 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10905 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10906 if (!C || C->getZExtValue() != 1)
10908 return DAG.getNode(ISD::AND, dl, VT,
10909 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10910 N00.getOperand(0), N00.getOperand(1)),
10911 DAG.getConstant(1, VT));
10917 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10918 DAGCombinerInfo &DCI) const {
10919 SelectionDAG &DAG = DCI.DAG;
10920 switch (N->getOpcode()) {
10922 case ISD::EXTRACT_VECTOR_ELT:
10923 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10924 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10925 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10926 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10929 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10930 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10931 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10933 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10934 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10935 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10936 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10937 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10938 case X86ISD::SHUFPS: // Handle all target specific shuffles
10939 case X86ISD::SHUFPD:
10940 case X86ISD::PALIGN:
10941 case X86ISD::PUNPCKHBW:
10942 case X86ISD::PUNPCKHWD:
10943 case X86ISD::PUNPCKHDQ:
10944 case X86ISD::PUNPCKHQDQ:
10945 case X86ISD::UNPCKHPS:
10946 case X86ISD::UNPCKHPD:
10947 case X86ISD::PUNPCKLBW:
10948 case X86ISD::PUNPCKLWD:
10949 case X86ISD::PUNPCKLDQ:
10950 case X86ISD::PUNPCKLQDQ:
10951 case X86ISD::UNPCKLPS:
10952 case X86ISD::UNPCKLPD:
10953 case X86ISD::MOVHLPS:
10954 case X86ISD::MOVLHPS:
10955 case X86ISD::PSHUFD:
10956 case X86ISD::PSHUFHW:
10957 case X86ISD::PSHUFLW:
10958 case X86ISD::MOVSS:
10959 case X86ISD::MOVSD:
10960 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10966 /// isTypeDesirableForOp - Return true if the target has native support for
10967 /// the specified value type and it is 'desirable' to use the type for the
10968 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10969 /// instruction encodings are longer and some i16 instructions are slow.
10970 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10971 if (!isTypeLegal(VT))
10973 if (VT != MVT::i16)
10980 case ISD::SIGN_EXTEND:
10981 case ISD::ZERO_EXTEND:
10982 case ISD::ANY_EXTEND:
10995 /// IsDesirableToPromoteOp - This method query the target whether it is
10996 /// beneficial for dag combiner to promote the specified node. If true, it
10997 /// should return the desired promotion type by reference.
10998 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10999 EVT VT = Op.getValueType();
11000 if (VT != MVT::i16)
11003 bool Promote = false;
11004 bool Commute = false;
11005 switch (Op.getOpcode()) {
11008 LoadSDNode *LD = cast<LoadSDNode>(Op);
11009 // If the non-extending load has a single use and it's not live out, then it
11010 // might be folded.
11011 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
11012 Op.hasOneUse()*/) {
11013 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11014 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
11015 // The only case where we'd want to promote LOAD (rather then it being
11016 // promoted as an operand is when it's only use is liveout.
11017 if (UI->getOpcode() != ISD::CopyToReg)
11024 case ISD::SIGN_EXTEND:
11025 case ISD::ZERO_EXTEND:
11026 case ISD::ANY_EXTEND:
11031 SDValue N0 = Op.getOperand(0);
11032 // Look out for (store (shl (load), x)).
11033 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
11046 SDValue N0 = Op.getOperand(0);
11047 SDValue N1 = Op.getOperand(1);
11048 if (!Commute && MayFoldLoad(N1))
11050 // Avoid disabling potential load folding opportunities.
11051 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
11053 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
11063 //===----------------------------------------------------------------------===//
11064 // X86 Inline Assembly Support
11065 //===----------------------------------------------------------------------===//
11067 static bool LowerToBSwap(CallInst *CI) {
11068 // FIXME: this should verify that we are targetting a 486 or better. If not,
11069 // we will turn this bswap into something that will be lowered to logical ops
11070 // instead of emitting the bswap asm. For now, we don't support 486 or lower
11071 // so don't worry about this.
11073 // Verify this is a simple bswap.
11074 if (CI->getNumArgOperands() != 1 ||
11075 CI->getType() != CI->getArgOperand(0)->getType() ||
11076 !CI->getType()->isIntegerTy())
11079 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11080 if (!Ty || Ty->getBitWidth() % 16 != 0)
11083 // Okay, we can do this xform, do so now.
11084 const Type *Tys[] = { Ty };
11085 Module *M = CI->getParent()->getParent()->getParent();
11086 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
11088 Value *Op = CI->getArgOperand(0);
11089 Op = CallInst::Create(Int, Op, CI->getName(), CI);
11091 CI->replaceAllUsesWith(Op);
11092 CI->eraseFromParent();
11096 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
11097 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11098 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
11100 std::string AsmStr = IA->getAsmString();
11102 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
11103 SmallVector<StringRef, 4> AsmPieces;
11104 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
11106 switch (AsmPieces.size()) {
11107 default: return false;
11109 AsmStr = AsmPieces[0];
11111 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
11114 if (AsmPieces.size() == 2 &&
11115 (AsmPieces[0] == "bswap" ||
11116 AsmPieces[0] == "bswapq" ||
11117 AsmPieces[0] == "bswapl") &&
11118 (AsmPieces[1] == "$0" ||
11119 AsmPieces[1] == "${0:q}")) {
11120 // No need to check constraints, nothing other than the equivalent of
11121 // "=r,0" would be valid here.
11122 return LowerToBSwap(CI);
11124 // rorw $$8, ${0:w} --> llvm.bswap.i16
11125 if (CI->getType()->isIntegerTy(16) &&
11126 AsmPieces.size() == 3 &&
11127 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
11128 AsmPieces[1] == "$$8," &&
11129 AsmPieces[2] == "${0:w}" &&
11130 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
11132 const std::string &Constraints = IA->getConstraintString();
11133 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
11134 std::sort(AsmPieces.begin(), AsmPieces.end());
11135 if (AsmPieces.size() == 4 &&
11136 AsmPieces[0] == "~{cc}" &&
11137 AsmPieces[1] == "~{dirflag}" &&
11138 AsmPieces[2] == "~{flags}" &&
11139 AsmPieces[3] == "~{fpsr}") {
11140 return LowerToBSwap(CI);
11145 if (CI->getType()->isIntegerTy(64) &&
11146 Constraints.size() >= 2 &&
11147 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
11148 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
11149 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
11150 SmallVector<StringRef, 4> Words;
11151 SplitString(AsmPieces[0], Words, " \t");
11152 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
11154 SplitString(AsmPieces[1], Words, " \t");
11155 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
11157 SplitString(AsmPieces[2], Words, " \t,");
11158 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
11159 Words[2] == "%edx") {
11160 return LowerToBSwap(CI);
11172 /// getConstraintType - Given a constraint letter, return the type of
11173 /// constraint it is for this target.
11174 X86TargetLowering::ConstraintType
11175 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
11176 if (Constraint.size() == 1) {
11177 switch (Constraint[0]) {
11189 return C_RegisterClass;
11197 return TargetLowering::getConstraintType(Constraint);
11200 /// Examine constraint type and operand type and determine a weight value,
11201 /// where: -1 = invalid match, and 0 = so-so match to 3 = good match.
11202 /// This object must already have been set up with the operand type
11203 /// and the current alternative constraint selected.
11204 int X86TargetLowering::getSingleConstraintMatchWeight(
11205 AsmOperandInfo &info, const char *constraint) const {
11207 Value *CallOperandVal = info.CallOperandVal;
11208 // If we don't have a value, we can't do a match,
11209 // but allow it at the lowest weight.
11210 if (CallOperandVal == NULL)
11212 // Look at the constraint type.
11213 switch (*constraint) {
11215 return TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11218 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
11219 if (C->getZExtValue() <= 31)
11228 /// LowerXConstraint - try to replace an X constraint, which matches anything,
11229 /// with another that has more specific requirements based on the type of the
11230 /// corresponding operand.
11231 const char *X86TargetLowering::
11232 LowerXConstraint(EVT ConstraintVT) const {
11233 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
11234 // 'f' like normal targets.
11235 if (ConstraintVT.isFloatingPoint()) {
11236 if (Subtarget->hasSSE2())
11238 if (Subtarget->hasSSE1())
11242 return TargetLowering::LowerXConstraint(ConstraintVT);
11245 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11246 /// vector. If it is invalid, don't add anything to Ops.
11247 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11249 std::vector<SDValue>&Ops,
11250 SelectionDAG &DAG) const {
11251 SDValue Result(0, 0);
11253 switch (Constraint) {
11256 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11257 if (C->getZExtValue() <= 31) {
11258 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11264 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11265 if (C->getZExtValue() <= 63) {
11266 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11272 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11273 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
11274 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11280 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11281 if (C->getZExtValue() <= 255) {
11282 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11288 // 32-bit signed value
11289 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11290 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11291 C->getSExtValue())) {
11292 // Widen to 64 bits here to get it sign extended.
11293 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
11296 // FIXME gcc accepts some relocatable values here too, but only in certain
11297 // memory models; it's complicated.
11302 // 32-bit unsigned value
11303 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
11304 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
11305 C->getZExtValue())) {
11306 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
11310 // FIXME gcc accepts some relocatable values here too, but only in certain
11311 // memory models; it's complicated.
11315 // Literal immediates are always ok.
11316 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11317 // Widen to 64 bits here to get it sign extended.
11318 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11322 // In any sort of PIC mode addresses need to be computed at runtime by
11323 // adding in a register or some sort of table lookup. These can't
11324 // be used as immediates.
11325 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11328 // If we are in non-pic codegen mode, we allow the address of a global (with
11329 // an optional displacement) to be used with 'i'.
11330 GlobalAddressSDNode *GA = 0;
11331 int64_t Offset = 0;
11333 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11335 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11336 Offset += GA->getOffset();
11338 } else if (Op.getOpcode() == ISD::ADD) {
11339 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11340 Offset += C->getZExtValue();
11341 Op = Op.getOperand(0);
11344 } else if (Op.getOpcode() == ISD::SUB) {
11345 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11346 Offset += -C->getZExtValue();
11347 Op = Op.getOperand(0);
11352 // Otherwise, this isn't something we can handle, reject it.
11356 const GlobalValue *GV = GA->getGlobal();
11357 // If we require an extra load to get this address, as in PIC mode, we
11358 // can't accept it.
11359 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11360 getTargetMachine())))
11363 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11364 GA->getValueType(0), Offset);
11369 if (Result.getNode()) {
11370 Ops.push_back(Result);
11373 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11376 std::vector<unsigned> X86TargetLowering::
11377 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11379 if (Constraint.size() == 1) {
11380 // FIXME: not handling fp-stack yet!
11381 switch (Constraint[0]) { // GCC X86 Constraint Letters
11382 default: break; // Unknown constraint letter
11383 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11384 if (Subtarget->is64Bit()) {
11385 if (VT == MVT::i32)
11386 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11387 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11388 X86::R10D,X86::R11D,X86::R12D,
11389 X86::R13D,X86::R14D,X86::R15D,
11390 X86::EBP, X86::ESP, 0);
11391 else if (VT == MVT::i16)
11392 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11393 X86::SI, X86::DI, X86::R8W,X86::R9W,
11394 X86::R10W,X86::R11W,X86::R12W,
11395 X86::R13W,X86::R14W,X86::R15W,
11396 X86::BP, X86::SP, 0);
11397 else if (VT == MVT::i8)
11398 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11399 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11400 X86::R10B,X86::R11B,X86::R12B,
11401 X86::R13B,X86::R14B,X86::R15B,
11402 X86::BPL, X86::SPL, 0);
11404 else if (VT == MVT::i64)
11405 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11406 X86::RSI, X86::RDI, X86::R8, X86::R9,
11407 X86::R10, X86::R11, X86::R12,
11408 X86::R13, X86::R14, X86::R15,
11409 X86::RBP, X86::RSP, 0);
11413 // 32-bit fallthrough
11414 case 'Q': // Q_REGS
11415 if (VT == MVT::i32)
11416 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11417 else if (VT == MVT::i16)
11418 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11419 else if (VT == MVT::i8)
11420 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11421 else if (VT == MVT::i64)
11422 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11427 return std::vector<unsigned>();
11430 std::pair<unsigned, const TargetRegisterClass*>
11431 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11433 // First, see if this is a constraint that directly corresponds to an LLVM
11435 if (Constraint.size() == 1) {
11436 // GCC Constraint Letters
11437 switch (Constraint[0]) {
11439 case 'r': // GENERAL_REGS
11440 case 'l': // INDEX_REGS
11442 return std::make_pair(0U, X86::GR8RegisterClass);
11443 if (VT == MVT::i16)
11444 return std::make_pair(0U, X86::GR16RegisterClass);
11445 if (VT == MVT::i32 || !Subtarget->is64Bit())
11446 return std::make_pair(0U, X86::GR32RegisterClass);
11447 return std::make_pair(0U, X86::GR64RegisterClass);
11448 case 'R': // LEGACY_REGS
11450 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11451 if (VT == MVT::i16)
11452 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11453 if (VT == MVT::i32 || !Subtarget->is64Bit())
11454 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11455 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11456 case 'f': // FP Stack registers.
11457 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11458 // value to the correct fpstack register class.
11459 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11460 return std::make_pair(0U, X86::RFP32RegisterClass);
11461 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11462 return std::make_pair(0U, X86::RFP64RegisterClass);
11463 return std::make_pair(0U, X86::RFP80RegisterClass);
11464 case 'y': // MMX_REGS if MMX allowed.
11465 if (!Subtarget->hasMMX()) break;
11466 return std::make_pair(0U, X86::VR64RegisterClass);
11467 case 'Y': // SSE_REGS if SSE2 allowed
11468 if (!Subtarget->hasSSE2()) break;
11470 case 'x': // SSE_REGS if SSE1 allowed
11471 if (!Subtarget->hasSSE1()) break;
11473 switch (VT.getSimpleVT().SimpleTy) {
11475 // Scalar SSE types.
11478 return std::make_pair(0U, X86::FR32RegisterClass);
11481 return std::make_pair(0U, X86::FR64RegisterClass);
11489 return std::make_pair(0U, X86::VR128RegisterClass);
11495 // Use the default implementation in TargetLowering to convert the register
11496 // constraint into a member of a register class.
11497 std::pair<unsigned, const TargetRegisterClass*> Res;
11498 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11500 // Not found as a standard register?
11501 if (Res.second == 0) {
11502 // Map st(0) -> st(7) -> ST0
11503 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11504 tolower(Constraint[1]) == 's' &&
11505 tolower(Constraint[2]) == 't' &&
11506 Constraint[3] == '(' &&
11507 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11508 Constraint[5] == ')' &&
11509 Constraint[6] == '}') {
11511 Res.first = X86::ST0+Constraint[4]-'0';
11512 Res.second = X86::RFP80RegisterClass;
11516 // GCC allows "st(0)" to be called just plain "st".
11517 if (StringRef("{st}").equals_lower(Constraint)) {
11518 Res.first = X86::ST0;
11519 Res.second = X86::RFP80RegisterClass;
11524 if (StringRef("{flags}").equals_lower(Constraint)) {
11525 Res.first = X86::EFLAGS;
11526 Res.second = X86::CCRRegisterClass;
11530 // 'A' means EAX + EDX.
11531 if (Constraint == "A") {
11532 Res.first = X86::EAX;
11533 Res.second = X86::GR32_ADRegisterClass;
11539 // Otherwise, check to see if this is a register class of the wrong value
11540 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11541 // turn into {ax},{dx}.
11542 if (Res.second->hasType(VT))
11543 return Res; // Correct type already, nothing to do.
11545 // All of the single-register GCC register classes map their values onto
11546 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11547 // really want an 8-bit or 32-bit register, map to the appropriate register
11548 // class and return the appropriate register.
11549 if (Res.second == X86::GR16RegisterClass) {
11550 if (VT == MVT::i8) {
11551 unsigned DestReg = 0;
11552 switch (Res.first) {
11554 case X86::AX: DestReg = X86::AL; break;
11555 case X86::DX: DestReg = X86::DL; break;
11556 case X86::CX: DestReg = X86::CL; break;
11557 case X86::BX: DestReg = X86::BL; break;
11560 Res.first = DestReg;
11561 Res.second = X86::GR8RegisterClass;
11563 } else if (VT == MVT::i32) {
11564 unsigned DestReg = 0;
11565 switch (Res.first) {
11567 case X86::AX: DestReg = X86::EAX; break;
11568 case X86::DX: DestReg = X86::EDX; break;
11569 case X86::CX: DestReg = X86::ECX; break;
11570 case X86::BX: DestReg = X86::EBX; break;
11571 case X86::SI: DestReg = X86::ESI; break;
11572 case X86::DI: DestReg = X86::EDI; break;
11573 case X86::BP: DestReg = X86::EBP; break;
11574 case X86::SP: DestReg = X86::ESP; break;
11577 Res.first = DestReg;
11578 Res.second = X86::GR32RegisterClass;
11580 } else if (VT == MVT::i64) {
11581 unsigned DestReg = 0;
11582 switch (Res.first) {
11584 case X86::AX: DestReg = X86::RAX; break;
11585 case X86::DX: DestReg = X86::RDX; break;
11586 case X86::CX: DestReg = X86::RCX; break;
11587 case X86::BX: DestReg = X86::RBX; break;
11588 case X86::SI: DestReg = X86::RSI; break;
11589 case X86::DI: DestReg = X86::RDI; break;
11590 case X86::BP: DestReg = X86::RBP; break;
11591 case X86::SP: DestReg = X86::RSP; break;
11594 Res.first = DestReg;
11595 Res.second = X86::GR64RegisterClass;
11598 } else if (Res.second == X86::FR32RegisterClass ||
11599 Res.second == X86::FR64RegisterClass ||
11600 Res.second == X86::VR128RegisterClass) {
11601 // Handle references to XMM physical registers that got mapped into the
11602 // wrong class. This can happen with constraints like {xmm0} where the
11603 // target independent register mapper will just pick the first match it can
11604 // find, ignoring the required type.
11605 if (VT == MVT::f32)
11606 Res.second = X86::FR32RegisterClass;
11607 else if (VT == MVT::f64)
11608 Res.second = X86::FR64RegisterClass;
11609 else if (X86::VR128RegisterClass->hasType(VT))
11610 Res.second = X86::VR128RegisterClass;