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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/StringExtras.h"
41 #include "llvm/Support/CommandLine.h"
45 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
47 // Forward declarations.
48 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
51 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
52 : TargetLowering(TM) {
53 Subtarget = &TM.getSubtarget<X86Subtarget>();
54 X86ScalarSSEf64 = Subtarget->hasSSE2();
55 X86ScalarSSEf32 = Subtarget->hasSSE1();
56 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
60 RegInfo = TM.getRegisterInfo();
63 // Set up the TargetLowering object.
65 // X86 is weird, it always uses i8 for shift amounts and setcc results.
66 setShiftAmountType(MVT::i8);
67 setBooleanContents(ZeroOrOneBooleanContent);
68 setSchedulingPreference(SchedulingForRegPressure);
69 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
70 setStackPointerRegisterToSaveRestore(X86StackPtr);
72 if (Subtarget->isTargetDarwin()) {
73 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
74 setUseUnderscoreSetJmp(false);
75 setUseUnderscoreLongJmp(false);
76 } else if (Subtarget->isTargetMingw()) {
77 // MS runtime is weird: it exports _setjmp, but longjmp!
78 setUseUnderscoreSetJmp(true);
79 setUseUnderscoreLongJmp(false);
81 setUseUnderscoreSetJmp(true);
82 setUseUnderscoreLongJmp(true);
85 // Set up the register classes.
86 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
87 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
88 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
89 if (Subtarget->is64Bit())
90 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
92 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
94 // We don't accept any truncstore of integer registers.
95 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
96 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
97 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
98 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
99 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
100 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
102 // SETOEQ and SETUNE require checking two conditions.
103 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
104 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
105 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
106 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
107 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
108 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
110 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
112 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
113 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
114 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
116 if (Subtarget->is64Bit()) {
117 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
118 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
120 if (!UseSoftFloat && !NoImplicitFloat && X86ScalarSSEf64) {
121 // We have an impenetrably clever algorithm for ui64->double only.
122 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
124 // We have faster algorithm for ui32->single only.
125 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
127 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
131 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
133 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
134 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
136 if (!UseSoftFloat && !NoImplicitFloat) {
137 // SSE has no i16 to fp conversion, only i32
138 if (X86ScalarSSEf32) {
139 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
140 // f32 and f64 cases are Legal, f80 case is not
141 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
143 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
144 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
148 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
151 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
152 // are Legal, f80 is custom lowered.
153 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
154 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
156 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
158 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
159 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
166 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
167 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
170 // Handle FP_TO_UINT by promoting the destination to a larger signed
172 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
173 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
174 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
176 if (Subtarget->is64Bit()) {
177 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
178 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
180 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
181 // Expand FP_TO_UINT into a select.
182 // FIXME: We would like to use a Custom expander here eventually to do
183 // the optimal thing for SSE vs. the default expansion in the legalizer.
184 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
186 // With SSE3 we can use fisttpll to convert to a signed i64.
187 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
190 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
191 if (!X86ScalarSSEf64) {
192 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
193 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
196 // Scalar integer divide and remainder are lowered to use operations that
197 // produce two results, to match the available instructions. This exposes
198 // the two-result form to trivial CSE, which is able to combine x/y and x%y
199 // into a single instruction.
201 // Scalar integer multiply-high is also lowered to use two-result
202 // operations, to match the available instructions. However, plain multiply
203 // (low) operations are left as Legal, as there are single-result
204 // instructions for this in x86. Using the two-result multiply instructions
205 // when both high and low results are needed must be arranged by dagcombine.
206 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
207 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
208 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
209 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
210 setOperationAction(ISD::SREM , MVT::i8 , Expand);
211 setOperationAction(ISD::UREM , MVT::i8 , Expand);
212 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
213 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
214 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
215 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
216 setOperationAction(ISD::SREM , MVT::i16 , Expand);
217 setOperationAction(ISD::UREM , MVT::i16 , Expand);
218 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
219 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
220 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
221 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
222 setOperationAction(ISD::SREM , MVT::i32 , Expand);
223 setOperationAction(ISD::UREM , MVT::i32 , Expand);
224 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
225 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
226 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
227 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
228 setOperationAction(ISD::SREM , MVT::i64 , Expand);
229 setOperationAction(ISD::UREM , MVT::i64 , Expand);
231 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
232 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
233 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
234 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
235 if (Subtarget->is64Bit())
236 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
237 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
238 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
239 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
240 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
241 setOperationAction(ISD::FREM , MVT::f32 , Expand);
242 setOperationAction(ISD::FREM , MVT::f64 , Expand);
243 setOperationAction(ISD::FREM , MVT::f80 , Expand);
244 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
246 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
247 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
248 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
249 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
250 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
251 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
252 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
253 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
254 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
257 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
258 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
261 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
262 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
264 // These should be promoted to a larger select which is supported.
265 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
266 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
267 // X86 wants to expand cmov itself.
268 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
269 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
270 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
271 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
272 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
273 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
274 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
275 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
276 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
277 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
278 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
279 if (Subtarget->is64Bit()) {
280 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
281 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
283 // X86 ret instruction may pop stack.
284 setOperationAction(ISD::RET , MVT::Other, Custom);
285 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
288 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
289 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
290 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
291 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
292 if (Subtarget->is64Bit())
293 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
294 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
295 if (Subtarget->is64Bit()) {
296 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
297 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
298 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
299 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
301 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
302 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
303 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
304 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
305 if (Subtarget->is64Bit()) {
306 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
307 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
308 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
311 if (Subtarget->hasSSE1())
312 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
314 if (!Subtarget->hasSSE2())
315 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
317 // Expand certain atomics
318 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
319 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
320 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
321 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
323 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
328 if (!Subtarget->is64Bit()) {
329 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
330 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
331 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
332 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
333 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
334 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
335 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
338 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
339 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
340 // FIXME - use subtarget debug flags
341 if (!Subtarget->isTargetDarwin() &&
342 !Subtarget->isTargetELF() &&
343 !Subtarget->isTargetCygMing()) {
344 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
345 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
348 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
349 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
350 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
351 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
352 if (Subtarget->is64Bit()) {
353 setExceptionPointerRegister(X86::RAX);
354 setExceptionSelectorRegister(X86::RDX);
356 setExceptionPointerRegister(X86::EAX);
357 setExceptionSelectorRegister(X86::EDX);
359 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
360 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
362 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
364 setOperationAction(ISD::TRAP, MVT::Other, Legal);
366 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
367 setOperationAction(ISD::VASTART , MVT::Other, Custom);
368 setOperationAction(ISD::VAEND , MVT::Other, Expand);
369 if (Subtarget->is64Bit()) {
370 setOperationAction(ISD::VAARG , MVT::Other, Custom);
371 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
373 setOperationAction(ISD::VAARG , MVT::Other, Expand);
374 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
377 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
378 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
379 if (Subtarget->is64Bit())
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
381 if (Subtarget->isTargetCygMing())
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
384 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
386 if (!UseSoftFloat && X86ScalarSSEf64) {
387 // f32 and f64 use SSE.
388 // Set up the FP register classes.
389 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
390 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
392 // Use ANDPD to simulate FABS.
393 setOperationAction(ISD::FABS , MVT::f64, Custom);
394 setOperationAction(ISD::FABS , MVT::f32, Custom);
396 // Use XORP to simulate FNEG.
397 setOperationAction(ISD::FNEG , MVT::f64, Custom);
398 setOperationAction(ISD::FNEG , MVT::f32, Custom);
400 // Use ANDPD and ORPD to simulate FCOPYSIGN.
401 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
402 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
404 // We don't support sin/cos/fmod
405 setOperationAction(ISD::FSIN , MVT::f64, Expand);
406 setOperationAction(ISD::FCOS , MVT::f64, Expand);
407 setOperationAction(ISD::FSIN , MVT::f32, Expand);
408 setOperationAction(ISD::FCOS , MVT::f32, Expand);
410 // Expand FP immediates into loads from the stack, except for the special
412 addLegalFPImmediate(APFloat(+0.0)); // xorpd
413 addLegalFPImmediate(APFloat(+0.0f)); // xorps
415 // Floating truncations from f80 and extensions to f80 go through memory.
416 // If optimizing, we lie about this though and handle it in
417 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
419 setConvertAction(MVT::f32, MVT::f80, Expand);
420 setConvertAction(MVT::f64, MVT::f80, Expand);
421 setConvertAction(MVT::f80, MVT::f32, Expand);
422 setConvertAction(MVT::f80, MVT::f64, Expand);
424 } else if (!UseSoftFloat && X86ScalarSSEf32) {
425 // Use SSE for f32, x87 for f64.
426 // Set up the FP register classes.
427 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
428 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
430 // Use ANDPS to simulate FABS.
431 setOperationAction(ISD::FABS , MVT::f32, Custom);
433 // Use XORP to simulate FNEG.
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
438 // Use ANDPS and ORPS to simulate FCOPYSIGN.
439 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
440 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
442 // We don't support sin/cos/fmod
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Special cases we handle for FP constants.
447 addLegalFPImmediate(APFloat(+0.0f)); // xorps
448 addLegalFPImmediate(APFloat(+0.0)); // FLD0
449 addLegalFPImmediate(APFloat(+1.0)); // FLD1
450 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
451 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
453 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
454 // this though and handle it in InstructionSelectPreprocess so that
455 // dagcombine2 can hack on these.
457 setConvertAction(MVT::f32, MVT::f64, Expand);
458 setConvertAction(MVT::f32, MVT::f80, Expand);
459 setConvertAction(MVT::f80, MVT::f32, Expand);
460 setConvertAction(MVT::f64, MVT::f32, Expand);
461 // And x87->x87 truncations also.
462 setConvertAction(MVT::f80, MVT::f64, Expand);
466 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
467 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
469 } else if (!UseSoftFloat) {
470 // f32 and f64 in x87.
471 // Set up the FP register classes.
472 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
473 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
475 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
476 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
477 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
478 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
480 // Floating truncations go through memory. If optimizing, we lie about
481 // this though and handle it in InstructionSelectPreprocess so that
482 // dagcombine2 can hack on these.
484 setConvertAction(MVT::f80, MVT::f32, Expand);
485 setConvertAction(MVT::f64, MVT::f32, Expand);
486 setConvertAction(MVT::f80, MVT::f64, Expand);
490 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
491 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
493 addLegalFPImmediate(APFloat(+0.0)); // FLD0
494 addLegalFPImmediate(APFloat(+1.0)); // FLD1
495 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
496 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
497 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
503 // Long double always uses X87.
505 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
506 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
507 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
510 APFloat TmpFlt(+0.0);
511 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
513 addLegalFPImmediate(TmpFlt); // FLD0
515 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
516 APFloat TmpFlt2(+1.0);
517 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
519 addLegalFPImmediate(TmpFlt2); // FLD1
520 TmpFlt2.changeSign();
521 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
525 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
526 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
530 // Always use a library call for pow.
531 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
532 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
533 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
535 setOperationAction(ISD::FLOG, MVT::f80, Expand);
536 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
537 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
538 setOperationAction(ISD::FEXP, MVT::f80, Expand);
539 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
541 // First set operation action for all vector types to either promote
542 // (for widening) or expand (for scalarization). Then we will selectively
543 // turn on ones that can be effectively codegen'd.
544 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
545 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
546 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
561 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
591 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
592 // with -msoft-float, disable use of MMX as well.
593 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
594 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
595 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
596 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
597 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
598 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
600 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
601 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
602 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
603 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
605 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
606 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
607 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
608 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
610 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
611 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
613 setOperationAction(ISD::AND, MVT::v8i8, Promote);
614 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
615 setOperationAction(ISD::AND, MVT::v4i16, Promote);
616 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
617 setOperationAction(ISD::AND, MVT::v2i32, Promote);
618 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
619 setOperationAction(ISD::AND, MVT::v1i64, Legal);
621 setOperationAction(ISD::OR, MVT::v8i8, Promote);
622 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
623 setOperationAction(ISD::OR, MVT::v4i16, Promote);
624 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
625 setOperationAction(ISD::OR, MVT::v2i32, Promote);
626 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
627 setOperationAction(ISD::OR, MVT::v1i64, Legal);
629 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
630 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
631 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
632 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
633 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
634 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
635 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
637 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
638 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
639 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
640 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
641 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
642 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
643 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
644 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
645 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
647 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
648 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
649 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
656 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
658 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
659 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
660 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
661 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
663 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
665 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
666 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
667 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
668 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
669 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
670 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
673 if (!UseSoftFloat && Subtarget->hasSSE1()) {
674 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
676 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
677 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
678 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
679 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
680 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
681 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
682 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
683 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
685 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
686 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
690 if (!UseSoftFloat && Subtarget->hasSSE2()) {
691 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
693 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
694 // registers cannot be used even for integer operations.
695 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
696 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
697 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
698 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
700 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
701 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
702 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
703 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
704 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
705 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
706 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
707 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
708 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
709 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
710 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
711 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
712 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
713 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
714 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
715 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
719 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
720 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
722 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
723 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
724 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
725 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
726 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
728 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
729 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
730 MVT VT = (MVT::SimpleValueType)i;
731 // Do not attempt to custom lower non-power-of-2 vectors
732 if (!isPowerOf2_32(VT.getVectorNumElements()))
734 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
735 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
736 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
739 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
740 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
741 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
742 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
743 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
744 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
746 if (Subtarget->is64Bit()) {
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
748 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
751 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
752 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
753 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
754 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
755 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
756 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
757 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
758 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
759 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
760 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
761 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
762 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
765 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
767 // Custom lower v2i64 and v2f64 selects.
768 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
769 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
770 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
771 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
775 if (Subtarget->hasSSE41()) {
776 // FIXME: Do we need to handle scalar-to-vector here?
777 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
779 // i8 and i16 vectors are custom , because the source register and source
780 // source memory operand types are not the same width. f32 vectors are
781 // custom since the immediate controlling the insert encodes additional
783 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
789 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
793 if (Subtarget->is64Bit()) {
794 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
799 if (Subtarget->hasSSE42()) {
800 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
803 // We want to custom lower some of our intrinsics.
804 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
806 // Add/Sub/Mul with overflow operations are custom lowered.
807 setOperationAction(ISD::SADDO, MVT::i32, Custom);
808 setOperationAction(ISD::SADDO, MVT::i64, Custom);
809 setOperationAction(ISD::UADDO, MVT::i32, Custom);
810 setOperationAction(ISD::UADDO, MVT::i64, Custom);
811 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
812 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
813 setOperationAction(ISD::USUBO, MVT::i32, Custom);
814 setOperationAction(ISD::USUBO, MVT::i64, Custom);
815 setOperationAction(ISD::SMULO, MVT::i32, Custom);
816 setOperationAction(ISD::SMULO, MVT::i64, Custom);
817 setOperationAction(ISD::UMULO, MVT::i32, Custom);
818 setOperationAction(ISD::UMULO, MVT::i64, Custom);
820 if (!Subtarget->is64Bit()) {
821 // These libcalls are not available in 32-bit.
822 setLibcallName(RTLIB::SHL_I128, 0);
823 setLibcallName(RTLIB::SRL_I128, 0);
824 setLibcallName(RTLIB::SRA_I128, 0);
827 // We have target-specific dag combine patterns for the following nodes:
828 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
829 setTargetDAGCombine(ISD::BUILD_VECTOR);
830 setTargetDAGCombine(ISD::SELECT);
831 setTargetDAGCombine(ISD::SHL);
832 setTargetDAGCombine(ISD::SRA);
833 setTargetDAGCombine(ISD::SRL);
834 setTargetDAGCombine(ISD::STORE);
835 if (Subtarget->is64Bit())
836 setTargetDAGCombine(ISD::MUL);
838 computeRegisterProperties();
840 // FIXME: These should be based on subtarget info. Plus, the values should
841 // be smaller when we are in optimizing for size mode.
842 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
843 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
844 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
845 allowUnalignedMemoryAccesses = true; // x86 supports it!
846 setPrefLoopAlignment(16);
850 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
855 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
856 /// the desired ByVal argument alignment.
857 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
860 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
861 if (VTy->getBitWidth() == 128)
863 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
864 unsigned EltAlign = 0;
865 getMaxByValAlign(ATy->getElementType(), EltAlign);
866 if (EltAlign > MaxAlign)
868 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
869 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
870 unsigned EltAlign = 0;
871 getMaxByValAlign(STy->getElementType(i), EltAlign);
872 if (EltAlign > MaxAlign)
881 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
882 /// function arguments in the caller parameter area. For X86, aggregates
883 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
884 /// are at 4-byte boundaries.
885 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
886 if (Subtarget->is64Bit()) {
887 // Max of 8 and alignment of type.
888 unsigned TyAlign = TD->getABITypeAlignment(Ty);
895 if (Subtarget->hasSSE1())
896 getMaxByValAlign(Ty, Align);
900 /// getOptimalMemOpType - Returns the target specific optimal type for load
901 /// and store operations as a result of memset, memcpy, and memmove
902 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
905 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
906 bool isSrcConst, bool isSrcStr) const {
907 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
908 // linux. This is because the stack realignment code can't handle certain
909 // cases like PR2962. This should be removed when PR2962 is fixed.
910 if (!NoImplicitFloat && Subtarget->getStackAlignment() >= 16) {
911 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
913 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
916 if (Subtarget->is64Bit() && Size >= 8)
921 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
923 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
924 SelectionDAG &DAG) const {
925 if (usesGlobalOffsetTable())
926 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
927 if (!Subtarget->isPICStyleRIPRel())
928 // This doesn't have DebugLoc associated with it, but is not really the
929 // same as a Register.
930 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
935 //===----------------------------------------------------------------------===//
936 // Return Value Calling Convention Implementation
937 //===----------------------------------------------------------------------===//
939 #include "X86GenCallingConv.inc"
941 /// LowerRET - Lower an ISD::RET node.
942 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
943 DebugLoc dl = Op.getDebugLoc();
944 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
946 SmallVector<CCValAssign, 16> RVLocs;
947 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
948 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
949 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
950 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
952 // If this is the first return lowered for this function, add the regs to the
953 // liveout set for the function.
954 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
955 for (unsigned i = 0; i != RVLocs.size(); ++i)
956 if (RVLocs[i].isRegLoc())
957 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
959 SDValue Chain = Op.getOperand(0);
961 // Handle tail call return.
962 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
963 if (Chain.getOpcode() == X86ISD::TAILCALL) {
964 SDValue TailCall = Chain;
965 SDValue TargetAddress = TailCall.getOperand(1);
966 SDValue StackAdjustment = TailCall.getOperand(2);
967 assert(((TargetAddress.getOpcode() == ISD::Register &&
968 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
969 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
970 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
971 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
972 "Expecting an global address, external symbol, or register");
973 assert(StackAdjustment.getOpcode() == ISD::Constant &&
974 "Expecting a const value");
976 SmallVector<SDValue,8> Operands;
977 Operands.push_back(Chain.getOperand(0));
978 Operands.push_back(TargetAddress);
979 Operands.push_back(StackAdjustment);
980 // Copy registers used by the call. Last operand is a flag so it is not
982 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
983 Operands.push_back(Chain.getOperand(i));
985 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
992 SmallVector<SDValue, 6> RetOps;
993 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
994 // Operand #1 = Bytes To Pop
995 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
997 // Copy the result values into the output registers.
998 for (unsigned i = 0; i != RVLocs.size(); ++i) {
999 CCValAssign &VA = RVLocs[i];
1000 assert(VA.isRegLoc() && "Can only return in registers!");
1001 SDValue ValToCopy = Op.getOperand(i*2+1);
1003 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1004 // the RET instruction and handled by the FP Stackifier.
1005 if (VA.getLocReg() == X86::ST0 ||
1006 VA.getLocReg() == X86::ST1) {
1007 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1008 // change the value to the FP stack register class.
1009 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1010 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1011 RetOps.push_back(ValToCopy);
1012 // Don't emit a copytoreg.
1016 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1017 // which is returned in RAX / RDX.
1018 if (Subtarget->is64Bit()) {
1019 MVT ValVT = ValToCopy.getValueType();
1020 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1021 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1022 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1023 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1027 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1028 Flag = Chain.getValue(1);
1031 // The x86-64 ABI for returning structs by value requires that we copy
1032 // the sret argument into %rax for the return. We saved the argument into
1033 // a virtual register in the entry block, so now we copy the value out
1035 if (Subtarget->is64Bit() &&
1036 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1037 MachineFunction &MF = DAG.getMachineFunction();
1038 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1039 unsigned Reg = FuncInfo->getSRetReturnReg();
1041 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1042 FuncInfo->setSRetReturnReg(Reg);
1044 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1046 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1047 Flag = Chain.getValue(1);
1050 RetOps[0] = Chain; // Update chain.
1052 // Add the flag if we have it.
1054 RetOps.push_back(Flag);
1056 return DAG.getNode(X86ISD::RET_FLAG, dl,
1057 MVT::Other, &RetOps[0], RetOps.size());
1061 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1062 /// appropriate copies out of appropriate physical registers. This assumes that
1063 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1064 /// being lowered. The returns a SDNode with the same number of values as the
1066 SDNode *X86TargetLowering::
1067 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1068 unsigned CallingConv, SelectionDAG &DAG) {
1070 DebugLoc dl = TheCall->getDebugLoc();
1071 // Assign locations to each value returned by this call.
1072 SmallVector<CCValAssign, 16> RVLocs;
1073 bool isVarArg = TheCall->isVarArg();
1074 bool Is64Bit = Subtarget->is64Bit();
1075 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1076 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1078 SmallVector<SDValue, 8> ResultVals;
1080 // Copy all of the result registers out of their specified physreg.
1081 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1082 CCValAssign &VA = RVLocs[i];
1083 MVT CopyVT = VA.getValVT();
1085 // If this is x86-64, and we disabled SSE, we can't return FP values
1086 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1087 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1088 cerr << "SSE register return with SSE disabled\n";
1092 // If this is a call to a function that returns an fp value on the floating
1093 // point stack, but where we prefer to use the value in xmm registers, copy
1094 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1095 if ((VA.getLocReg() == X86::ST0 ||
1096 VA.getLocReg() == X86::ST1) &&
1097 isScalarFPTypeInSSEReg(VA.getValVT())) {
1102 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1103 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1104 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1105 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1106 MVT::v2i64, InFlag).getValue(1);
1107 Val = Chain.getValue(0);
1108 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1109 Val, DAG.getConstant(0, MVT::i64));
1111 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1112 MVT::i64, InFlag).getValue(1);
1113 Val = Chain.getValue(0);
1115 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1117 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1118 CopyVT, InFlag).getValue(1);
1119 Val = Chain.getValue(0);
1121 InFlag = Chain.getValue(2);
1123 if (CopyVT != VA.getValVT()) {
1124 // Round the F80 the right size, which also moves to the appropriate xmm
1126 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1127 // This truncation won't change the value.
1128 DAG.getIntPtrConstant(1));
1131 ResultVals.push_back(Val);
1134 // Merge everything together with a MERGE_VALUES node.
1135 ResultVals.push_back(Chain);
1136 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1137 &ResultVals[0], ResultVals.size()).getNode();
1141 //===----------------------------------------------------------------------===//
1142 // C & StdCall & Fast Calling Convention implementation
1143 //===----------------------------------------------------------------------===//
1144 // StdCall calling convention seems to be standard for many Windows' API
1145 // routines and around. It differs from C calling convention just a little:
1146 // callee should clean up the stack, not caller. Symbols should be also
1147 // decorated in some fancy way :) It doesn't support any vector arguments.
1148 // For info on fast calling convention see Fast Calling Convention (tail call)
1149 // implementation LowerX86_32FastCCCallTo.
1151 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1153 static bool CallIsStructReturn(CallSDNode *TheCall) {
1154 unsigned NumOps = TheCall->getNumArgs();
1158 return TheCall->getArgFlags(0).isSRet();
1161 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1162 /// return semantics.
1163 static bool ArgsAreStructReturn(SDValue Op) {
1164 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1168 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1171 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1172 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1174 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1178 switch (CallingConv) {
1181 case CallingConv::X86_StdCall:
1182 return !Subtarget->is64Bit();
1183 case CallingConv::X86_FastCall:
1184 return !Subtarget->is64Bit();
1185 case CallingConv::Fast:
1186 return PerformTailCallOpt;
1190 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1191 /// given CallingConvention value.
1192 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1193 if (Subtarget->is64Bit()) {
1194 if (Subtarget->isTargetWin64())
1195 return CC_X86_Win64_C;
1196 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1197 return CC_X86_64_TailCall;
1202 if (CC == CallingConv::X86_FastCall)
1203 return CC_X86_32_FastCall;
1204 else if (CC == CallingConv::Fast)
1205 return CC_X86_32_FastCC;
1210 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1211 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1213 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1214 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1215 if (CC == CallingConv::X86_FastCall)
1217 else if (CC == CallingConv::X86_StdCall)
1223 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1224 /// in a register before calling.
1225 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1226 return !IsTailCall && !Is64Bit &&
1227 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1228 Subtarget->isPICStyleGOT();
1231 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1232 /// address to be loaded in a register.
1234 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1235 return !Is64Bit && IsTailCall &&
1236 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1237 Subtarget->isPICStyleGOT();
1240 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1241 /// by "Src" to address "Dst" with size and alignment information specified by
1242 /// the specific parameter attribute. The copy will be passed as a byval
1243 /// function parameter.
1245 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1246 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1248 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1249 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1250 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1253 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1254 const CCValAssign &VA,
1255 MachineFrameInfo *MFI,
1257 SDValue Root, unsigned i) {
1258 // Create the nodes corresponding to a load from this parameter slot.
1259 ISD::ArgFlagsTy Flags =
1260 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1261 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1262 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1264 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1265 // changed with more analysis.
1266 // In case of tail call optimization mark all arguments mutable. Since they
1267 // could be overwritten by lowering of arguments in case of a tail call.
1268 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1269 VA.getLocMemOffset(), isImmutable);
1270 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1271 if (Flags.isByVal())
1273 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1274 PseudoSourceValue::getFixedStack(FI), 0);
1278 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1279 MachineFunction &MF = DAG.getMachineFunction();
1280 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1281 DebugLoc dl = Op.getDebugLoc();
1283 const Function* Fn = MF.getFunction();
1284 if (Fn->hasExternalLinkage() &&
1285 Subtarget->isTargetCygMing() &&
1286 Fn->getName() == "main")
1287 FuncInfo->setForceFramePointer(true);
1289 // Decorate the function name.
1290 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1292 MachineFrameInfo *MFI = MF.getFrameInfo();
1293 SDValue Root = Op.getOperand(0);
1294 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1295 unsigned CC = MF.getFunction()->getCallingConv();
1296 bool Is64Bit = Subtarget->is64Bit();
1297 bool IsWin64 = Subtarget->isTargetWin64();
1299 assert(!(isVarArg && CC == CallingConv::Fast) &&
1300 "Var args not supported with calling convention fastcc");
1302 // Assign locations to all of the incoming arguments.
1303 SmallVector<CCValAssign, 16> ArgLocs;
1304 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1305 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1307 SmallVector<SDValue, 8> ArgValues;
1308 unsigned LastVal = ~0U;
1309 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1310 CCValAssign &VA = ArgLocs[i];
1311 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1313 assert(VA.getValNo() != LastVal &&
1314 "Don't support value assigned to multiple locs yet");
1315 LastVal = VA.getValNo();
1317 if (VA.isRegLoc()) {
1318 MVT RegVT = VA.getLocVT();
1319 TargetRegisterClass *RC = NULL;
1320 if (RegVT == MVT::i32)
1321 RC = X86::GR32RegisterClass;
1322 else if (Is64Bit && RegVT == MVT::i64)
1323 RC = X86::GR64RegisterClass;
1324 else if (RegVT == MVT::f32)
1325 RC = X86::FR32RegisterClass;
1326 else if (RegVT == MVT::f64)
1327 RC = X86::FR64RegisterClass;
1328 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1329 RC = X86::VR128RegisterClass;
1330 else if (RegVT.isVector()) {
1331 assert(RegVT.getSizeInBits() == 64);
1333 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1335 // Darwin calling convention passes MMX values in either GPRs or
1336 // XMMs in x86-64. Other targets pass them in memory.
1337 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1338 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1341 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1346 assert(0 && "Unknown argument type!");
1349 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1350 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1352 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1353 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1355 if (VA.getLocInfo() == CCValAssign::SExt)
1356 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1357 DAG.getValueType(VA.getValVT()));
1358 else if (VA.getLocInfo() == CCValAssign::ZExt)
1359 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1360 DAG.getValueType(VA.getValVT()));
1362 if (VA.getLocInfo() != CCValAssign::Full)
1363 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1365 // Handle MMX values passed in GPRs.
1366 if (Is64Bit && RegVT != VA.getLocVT()) {
1367 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1368 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1369 else if (RC == X86::VR128RegisterClass) {
1370 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1371 ArgValue, DAG.getConstant(0, MVT::i64));
1372 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1376 ArgValues.push_back(ArgValue);
1378 assert(VA.isMemLoc());
1379 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1383 // The x86-64 ABI for returning structs by value requires that we copy
1384 // the sret argument into %rax for the return. Save the argument into
1385 // a virtual register so that we can access it from the return points.
1386 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1387 MachineFunction &MF = DAG.getMachineFunction();
1388 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1389 unsigned Reg = FuncInfo->getSRetReturnReg();
1391 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1392 FuncInfo->setSRetReturnReg(Reg);
1394 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1395 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1398 unsigned StackSize = CCInfo.getNextStackOffset();
1399 // align stack specially for tail calls
1400 if (PerformTailCallOpt && CC == CallingConv::Fast)
1401 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1403 // If the function takes variable number of arguments, make a frame index for
1404 // the start of the first vararg value... for expansion of llvm.va_start.
1406 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1407 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1410 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1412 // FIXME: We should really autogenerate these arrays
1413 static const unsigned GPR64ArgRegsWin64[] = {
1414 X86::RCX, X86::RDX, X86::R8, X86::R9
1416 static const unsigned XMMArgRegsWin64[] = {
1417 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1419 static const unsigned GPR64ArgRegs64Bit[] = {
1420 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1422 static const unsigned XMMArgRegs64Bit[] = {
1423 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1424 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1426 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1429 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1430 GPR64ArgRegs = GPR64ArgRegsWin64;
1431 XMMArgRegs = XMMArgRegsWin64;
1433 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1434 GPR64ArgRegs = GPR64ArgRegs64Bit;
1435 XMMArgRegs = XMMArgRegs64Bit;
1437 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1439 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1442 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1443 "SSE register cannot be used when SSE is disabled!");
1444 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloat) &&
1445 "SSE register cannot be used when SSE is disabled!");
1446 if (UseSoftFloat || NoImplicitFloat || !Subtarget->hasSSE1())
1447 // Kernel mode asks for SSE to be disabled, so don't push them
1449 TotalNumXMMRegs = 0;
1451 // For X86-64, if there are vararg parameters that are passed via
1452 // registers, then we must store them to their spots on the stack so they
1453 // may be loaded by deferencing the result of va_next.
1454 VarArgsGPOffset = NumIntRegs * 8;
1455 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1456 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1457 TotalNumXMMRegs * 16, 16);
1459 // Store the integer parameter registers.
1460 SmallVector<SDValue, 8> MemOps;
1461 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1462 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1463 DAG.getIntPtrConstant(VarArgsGPOffset));
1464 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1465 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1466 X86::GR64RegisterClass);
1467 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1469 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1470 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1471 MemOps.push_back(Store);
1472 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1473 DAG.getIntPtrConstant(8));
1476 // Now store the XMM (fp + vector) parameter registers.
1477 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1478 DAG.getIntPtrConstant(VarArgsFPOffset));
1479 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1480 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1481 X86::VR128RegisterClass);
1482 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1484 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1485 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1486 MemOps.push_back(Store);
1487 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1488 DAG.getIntPtrConstant(16));
1490 if (!MemOps.empty())
1491 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1492 &MemOps[0], MemOps.size());
1496 ArgValues.push_back(Root);
1498 // Some CCs need callee pop.
1499 if (IsCalleePop(isVarArg, CC)) {
1500 BytesToPopOnReturn = StackSize; // Callee pops everything.
1501 BytesCallerReserves = 0;
1503 BytesToPopOnReturn = 0; // Callee pops nothing.
1504 // If this is an sret function, the return should pop the hidden pointer.
1505 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1506 BytesToPopOnReturn = 4;
1507 BytesCallerReserves = StackSize;
1511 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1512 if (CC == CallingConv::X86_FastCall)
1513 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1516 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1518 // Return the new list of results.
1519 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1520 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1524 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1525 const SDValue &StackPtr,
1526 const CCValAssign &VA,
1528 SDValue Arg, ISD::ArgFlagsTy Flags) {
1529 DebugLoc dl = TheCall->getDebugLoc();
1530 unsigned LocMemOffset = VA.getLocMemOffset();
1531 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1532 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1533 if (Flags.isByVal()) {
1534 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1536 return DAG.getStore(Chain, dl, Arg, PtrOff,
1537 PseudoSourceValue::getStack(), LocMemOffset);
1540 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1541 /// optimization is performed and it is required.
1543 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1544 SDValue &OutRetAddr,
1550 if (!IsTailCall || FPDiff==0) return Chain;
1552 // Adjust the Return address stack slot.
1553 MVT VT = getPointerTy();
1554 OutRetAddr = getReturnAddressFrameIndex(DAG);
1556 // Load the "old" Return address.
1557 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1558 return SDValue(OutRetAddr.getNode(), 1);
1561 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1562 /// optimization is performed and it is required (FPDiff!=0).
1564 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1565 SDValue Chain, SDValue RetAddrFrIdx,
1566 bool Is64Bit, int FPDiff, DebugLoc dl) {
1567 // Store the return address to the appropriate stack slot.
1568 if (!FPDiff) return Chain;
1569 // Calculate the new stack slot for the return address.
1570 int SlotSize = Is64Bit ? 8 : 4;
1571 int NewReturnAddrFI =
1572 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1573 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1574 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1575 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1576 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1580 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1581 MachineFunction &MF = DAG.getMachineFunction();
1582 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1583 SDValue Chain = TheCall->getChain();
1584 unsigned CC = TheCall->getCallingConv();
1585 bool isVarArg = TheCall->isVarArg();
1586 bool IsTailCall = TheCall->isTailCall() &&
1587 CC == CallingConv::Fast && PerformTailCallOpt;
1588 SDValue Callee = TheCall->getCallee();
1589 bool Is64Bit = Subtarget->is64Bit();
1590 bool IsStructRet = CallIsStructReturn(TheCall);
1591 DebugLoc dl = TheCall->getDebugLoc();
1593 assert(!(isVarArg && CC == CallingConv::Fast) &&
1594 "Var args not supported with calling convention fastcc");
1596 // Analyze operands of the call, assigning locations to each operand.
1597 SmallVector<CCValAssign, 16> ArgLocs;
1598 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1599 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1601 // Get a count of how many bytes are to be pushed on the stack.
1602 unsigned NumBytes = CCInfo.getNextStackOffset();
1603 if (PerformTailCallOpt && CC == CallingConv::Fast)
1604 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1608 // Lower arguments at fp - stackoffset + fpdiff.
1609 unsigned NumBytesCallerPushed =
1610 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1611 FPDiff = NumBytesCallerPushed - NumBytes;
1613 // Set the delta of movement of the returnaddr stackslot.
1614 // But only set if delta is greater than previous delta.
1615 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1616 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1619 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1621 SDValue RetAddrFrIdx;
1622 // Load return adress for tail calls.
1623 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1626 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1627 SmallVector<SDValue, 8> MemOpChains;
1630 // Walk the register/memloc assignments, inserting copies/loads. In the case
1631 // of tail call optimization arguments are handle later.
1632 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1633 CCValAssign &VA = ArgLocs[i];
1634 SDValue Arg = TheCall->getArg(i);
1635 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1636 bool isByVal = Flags.isByVal();
1638 // Promote the value if needed.
1639 switch (VA.getLocInfo()) {
1640 default: assert(0 && "Unknown loc info!");
1641 case CCValAssign::Full: break;
1642 case CCValAssign::SExt:
1643 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1645 case CCValAssign::ZExt:
1646 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1648 case CCValAssign::AExt:
1649 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1653 if (VA.isRegLoc()) {
1655 MVT RegVT = VA.getLocVT();
1656 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1657 switch (VA.getLocReg()) {
1660 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1662 // Special case: passing MMX values in GPR registers.
1663 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1666 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1667 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1668 // Special case: passing MMX values in XMM registers.
1669 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1670 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1671 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1676 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1678 if (!IsTailCall || (IsTailCall && isByVal)) {
1679 assert(VA.isMemLoc());
1680 if (StackPtr.getNode() == 0)
1681 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1683 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1684 Chain, Arg, Flags));
1689 if (!MemOpChains.empty())
1690 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1691 &MemOpChains[0], MemOpChains.size());
1693 // Build a sequence of copy-to-reg nodes chained together with token chain
1694 // and flag operands which copy the outgoing args into registers.
1696 // Tail call byval lowering might overwrite argument registers so in case of
1697 // tail call optimization the copies to registers are lowered later.
1699 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1700 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1701 RegsToPass[i].second, InFlag);
1702 InFlag = Chain.getValue(1);
1705 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1707 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1708 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1709 DAG.getNode(X86ISD::GlobalBaseReg,
1710 DebugLoc::getUnknownLoc(),
1713 InFlag = Chain.getValue(1);
1715 // If we are tail calling and generating PIC/GOT style code load the address
1716 // of the callee into ecx. The value in ecx is used as target of the tail
1717 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1718 // calls on PIC/GOT architectures. Normally we would just put the address of
1719 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1720 // restored (since ebx is callee saved) before jumping to the target@PLT.
1721 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1722 // Note: The actual moving to ecx is done further down.
1723 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1724 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1725 !G->getGlobal()->hasProtectedVisibility())
1726 Callee = LowerGlobalAddress(Callee, DAG);
1727 else if (isa<ExternalSymbolSDNode>(Callee))
1728 Callee = LowerExternalSymbol(Callee,DAG);
1731 if (Is64Bit && isVarArg) {
1732 // From AMD64 ABI document:
1733 // For calls that may call functions that use varargs or stdargs
1734 // (prototype-less calls or calls to functions containing ellipsis (...) in
1735 // the declaration) %al is used as hidden argument to specify the number
1736 // of SSE registers used. The contents of %al do not need to match exactly
1737 // the number of registers, but must be an ubound on the number of SSE
1738 // registers used and is in the range 0 - 8 inclusive.
1740 // FIXME: Verify this on Win64
1741 // Count the number of XMM registers allocated.
1742 static const unsigned XMMArgRegs[] = {
1743 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1744 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1746 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1747 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1748 && "SSE registers cannot be used when SSE is disabled");
1750 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1751 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1752 InFlag = Chain.getValue(1);
1756 // For tail calls lower the arguments to the 'real' stack slot.
1758 SmallVector<SDValue, 8> MemOpChains2;
1761 // Do not flag preceeding copytoreg stuff together with the following stuff.
1763 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1764 CCValAssign &VA = ArgLocs[i];
1765 if (!VA.isRegLoc()) {
1766 assert(VA.isMemLoc());
1767 SDValue Arg = TheCall->getArg(i);
1768 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1769 // Create frame index.
1770 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1771 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1772 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1773 FIN = DAG.getFrameIndex(FI, getPointerTy());
1775 if (Flags.isByVal()) {
1776 // Copy relative to framepointer.
1777 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1778 if (StackPtr.getNode() == 0)
1779 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1781 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1783 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1786 // Store relative to framepointer.
1787 MemOpChains2.push_back(
1788 DAG.getStore(Chain, dl, Arg, FIN,
1789 PseudoSourceValue::getFixedStack(FI), 0));
1794 if (!MemOpChains2.empty())
1795 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1796 &MemOpChains2[0], MemOpChains2.size());
1798 // Copy arguments to their registers.
1799 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1800 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1801 RegsToPass[i].second, InFlag);
1802 InFlag = Chain.getValue(1);
1806 // Store the return address to the appropriate stack slot.
1807 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1811 // If the callee is a GlobalAddress node (quite common, every direct call is)
1812 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1813 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1814 // We should use extra load for direct calls to dllimported functions in
1816 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1817 getTargetMachine(), true))
1818 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1820 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1821 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1822 } else if (IsTailCall) {
1823 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1825 Chain = DAG.getCopyToReg(Chain, dl,
1826 DAG.getRegister(Opc, getPointerTy()),
1828 Callee = DAG.getRegister(Opc, getPointerTy());
1829 // Add register as live out.
1830 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1833 // Returns a chain & a flag for retval copy to use.
1834 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1835 SmallVector<SDValue, 8> Ops;
1838 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1839 DAG.getIntPtrConstant(0, true), InFlag);
1840 InFlag = Chain.getValue(1);
1842 // Returns a chain & a flag for retval copy to use.
1843 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1847 Ops.push_back(Chain);
1848 Ops.push_back(Callee);
1851 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1853 // Add argument registers to the end of the list so that they are known live
1855 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1856 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1857 RegsToPass[i].second.getValueType()));
1859 // Add an implicit use GOT pointer in EBX.
1860 if (!IsTailCall && !Is64Bit &&
1861 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1862 Subtarget->isPICStyleGOT())
1863 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1865 // Add an implicit use of AL for x86 vararg functions.
1866 if (Is64Bit && isVarArg)
1867 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1869 if (InFlag.getNode())
1870 Ops.push_back(InFlag);
1873 assert(InFlag.getNode() &&
1874 "Flag must be set. Depend on flag being set in LowerRET");
1875 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1876 TheCall->getVTList(), &Ops[0], Ops.size());
1878 return SDValue(Chain.getNode(), Op.getResNo());
1881 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1882 InFlag = Chain.getValue(1);
1884 // Create the CALLSEQ_END node.
1885 unsigned NumBytesForCalleeToPush;
1886 if (IsCalleePop(isVarArg, CC))
1887 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1888 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1889 // If this is is a call to a struct-return function, the callee
1890 // pops the hidden struct pointer, so we have to push it back.
1891 // This is common for Darwin/X86, Linux & Mingw32 targets.
1892 NumBytesForCalleeToPush = 4;
1894 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1896 // Returns a flag for retval copy to use.
1897 Chain = DAG.getCALLSEQ_END(Chain,
1898 DAG.getIntPtrConstant(NumBytes, true),
1899 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1902 InFlag = Chain.getValue(1);
1904 // Handle result values, copying them out of physregs into vregs that we
1906 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1911 //===----------------------------------------------------------------------===//
1912 // Fast Calling Convention (tail call) implementation
1913 //===----------------------------------------------------------------------===//
1915 // Like std call, callee cleans arguments, convention except that ECX is
1916 // reserved for storing the tail called function address. Only 2 registers are
1917 // free for argument passing (inreg). Tail call optimization is performed
1919 // * tailcallopt is enabled
1920 // * caller/callee are fastcc
1921 // On X86_64 architecture with GOT-style position independent code only local
1922 // (within module) calls are supported at the moment.
1923 // To keep the stack aligned according to platform abi the function
1924 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1925 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1926 // If a tail called function callee has more arguments than the caller the
1927 // caller needs to make sure that there is room to move the RETADDR to. This is
1928 // achieved by reserving an area the size of the argument delta right after the
1929 // original REtADDR, but before the saved framepointer or the spilled registers
1930 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1942 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1943 /// for a 16 byte align requirement.
1944 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1945 SelectionDAG& DAG) {
1946 MachineFunction &MF = DAG.getMachineFunction();
1947 const TargetMachine &TM = MF.getTarget();
1948 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1949 unsigned StackAlignment = TFI.getStackAlignment();
1950 uint64_t AlignMask = StackAlignment - 1;
1951 int64_t Offset = StackSize;
1952 uint64_t SlotSize = TD->getPointerSize();
1953 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1954 // Number smaller than 12 so just add the difference.
1955 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1957 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1958 Offset = ((~AlignMask) & Offset) + StackAlignment +
1959 (StackAlignment-SlotSize);
1964 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1965 /// following the call is a return. A function is eligible if caller/callee
1966 /// calling conventions match, currently only fastcc supports tail calls, and
1967 /// the function CALL is immediatly followed by a RET.
1968 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1970 SelectionDAG& DAG) const {
1971 if (!PerformTailCallOpt)
1974 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1975 MachineFunction &MF = DAG.getMachineFunction();
1976 unsigned CallerCC = MF.getFunction()->getCallingConv();
1977 unsigned CalleeCC= TheCall->getCallingConv();
1978 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1979 SDValue Callee = TheCall->getCallee();
1980 // On x86/32Bit PIC/GOT tail calls are supported.
1981 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1982 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1985 // Can only do local tail calls (in same module, hidden or protected) on
1986 // x86_64 PIC/GOT at the moment.
1987 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1988 return G->getGlobal()->hasHiddenVisibility()
1989 || G->getGlobal()->hasProtectedVisibility();
1997 X86TargetLowering::createFastISel(MachineFunction &mf,
1998 MachineModuleInfo *mmo,
2000 DenseMap<const Value *, unsigned> &vm,
2001 DenseMap<const BasicBlock *,
2002 MachineBasicBlock *> &bm,
2003 DenseMap<const AllocaInst *, int> &am
2005 , SmallSet<Instruction*, 8> &cil
2008 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2016 //===----------------------------------------------------------------------===//
2017 // Other Lowering Hooks
2018 //===----------------------------------------------------------------------===//
2021 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2022 MachineFunction &MF = DAG.getMachineFunction();
2023 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2024 int ReturnAddrIndex = FuncInfo->getRAIndex();
2026 if (ReturnAddrIndex == 0) {
2027 // Set up a frame object for the return address.
2028 uint64_t SlotSize = TD->getPointerSize();
2029 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2030 FuncInfo->setRAIndex(ReturnAddrIndex);
2033 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2037 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2038 /// specific condition code, returning the condition code and the LHS/RHS of the
2039 /// comparison to make.
2040 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2041 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2043 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2044 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2045 // X > -1 -> X == 0, jump !sign.
2046 RHS = DAG.getConstant(0, RHS.getValueType());
2047 return X86::COND_NS;
2048 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2049 // X < 0 -> X == 0, jump on sign.
2051 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2053 RHS = DAG.getConstant(0, RHS.getValueType());
2054 return X86::COND_LE;
2058 switch (SetCCOpcode) {
2059 default: assert(0 && "Invalid integer condition!");
2060 case ISD::SETEQ: return X86::COND_E;
2061 case ISD::SETGT: return X86::COND_G;
2062 case ISD::SETGE: return X86::COND_GE;
2063 case ISD::SETLT: return X86::COND_L;
2064 case ISD::SETLE: return X86::COND_LE;
2065 case ISD::SETNE: return X86::COND_NE;
2066 case ISD::SETULT: return X86::COND_B;
2067 case ISD::SETUGT: return X86::COND_A;
2068 case ISD::SETULE: return X86::COND_BE;
2069 case ISD::SETUGE: return X86::COND_AE;
2073 // First determine if it is required or is profitable to flip the operands.
2075 // If LHS is a foldable load, but RHS is not, flip the condition.
2076 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2077 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2078 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2079 std::swap(LHS, RHS);
2082 switch (SetCCOpcode) {
2088 std::swap(LHS, RHS);
2092 // On a floating point condition, the flags are set as follows:
2094 // 0 | 0 | 0 | X > Y
2095 // 0 | 0 | 1 | X < Y
2096 // 1 | 0 | 0 | X == Y
2097 // 1 | 1 | 1 | unordered
2098 switch (SetCCOpcode) {
2099 default: assert(0 && "Condcode should be pre-legalized away");
2101 case ISD::SETEQ: return X86::COND_E;
2102 case ISD::SETOLT: // flipped
2104 case ISD::SETGT: return X86::COND_A;
2105 case ISD::SETOLE: // flipped
2107 case ISD::SETGE: return X86::COND_AE;
2108 case ISD::SETUGT: // flipped
2110 case ISD::SETLT: return X86::COND_B;
2111 case ISD::SETUGE: // flipped
2113 case ISD::SETLE: return X86::COND_BE;
2115 case ISD::SETNE: return X86::COND_NE;
2116 case ISD::SETUO: return X86::COND_P;
2117 case ISD::SETO: return X86::COND_NP;
2121 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2122 /// code. Current x86 isa includes the following FP cmov instructions:
2123 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2124 static bool hasFPCMov(unsigned X86CC) {
2140 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2141 /// the specified range (L, H].
2142 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2143 return (Val < 0) || (Val >= Low && Val < Hi);
2146 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2147 /// specified value.
2148 static bool isUndefOrEqual(int Val, int CmpVal) {
2149 if (Val < 0 || Val == CmpVal)
2154 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2155 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2156 /// the second operand.
2157 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2158 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2159 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2160 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2161 return (Mask[0] < 2 && Mask[1] < 2);
2165 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2166 SmallVector<int, 8> M;
2168 return ::isPSHUFDMask(M, N->getValueType(0));
2171 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2172 /// is suitable for input to PSHUFHW.
2173 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2174 if (VT != MVT::v8i16)
2177 // Lower quadword copied in order or undef.
2178 for (int i = 0; i != 4; ++i)
2179 if (Mask[i] >= 0 && Mask[i] != i)
2182 // Upper quadword shuffled.
2183 for (int i = 4; i != 8; ++i)
2184 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2190 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2191 SmallVector<int, 8> M;
2193 return ::isPSHUFHWMask(M, N->getValueType(0));
2196 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2197 /// is suitable for input to PSHUFLW.
2198 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2199 if (VT != MVT::v8i16)
2202 // Upper quadword copied in order.
2203 for (int i = 4; i != 8; ++i)
2204 if (Mask[i] >= 0 && Mask[i] != i)
2207 // Lower quadword shuffled.
2208 for (int i = 0; i != 4; ++i)
2215 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2216 SmallVector<int, 8> M;
2218 return ::isPSHUFLWMask(M, N->getValueType(0));
2221 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2222 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2223 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2224 int NumElems = VT.getVectorNumElements();
2225 if (NumElems != 2 && NumElems != 4)
2228 int Half = NumElems / 2;
2229 for (int i = 0; i < Half; ++i)
2230 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2232 for (int i = Half; i < NumElems; ++i)
2233 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2239 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2240 SmallVector<int, 8> M;
2242 return ::isSHUFPMask(M, N->getValueType(0));
2245 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2246 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2247 /// half elements to come from vector 1 (which would equal the dest.) and
2248 /// the upper half to come from vector 2.
2249 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2250 int NumElems = VT.getVectorNumElements();
2252 if (NumElems != 2 && NumElems != 4)
2255 int Half = NumElems / 2;
2256 for (int i = 0; i < Half; ++i)
2257 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2259 for (int i = Half; i < NumElems; ++i)
2260 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2265 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2266 SmallVector<int, 8> M;
2268 return isCommutedSHUFPMask(M, N->getValueType(0));
2271 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2272 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2273 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2274 if (N->getValueType(0).getVectorNumElements() != 4)
2277 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2278 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2279 isUndefOrEqual(N->getMaskElt(1), 7) &&
2280 isUndefOrEqual(N->getMaskElt(2), 2) &&
2281 isUndefOrEqual(N->getMaskElt(3), 3);
2284 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2285 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2286 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2287 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2289 if (NumElems != 2 && NumElems != 4)
2292 for (unsigned i = 0; i < NumElems/2; ++i)
2293 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2296 for (unsigned i = NumElems/2; i < NumElems; ++i)
2297 if (!isUndefOrEqual(N->getMaskElt(i), i))
2303 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2304 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2306 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2307 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2309 if (NumElems != 2 && NumElems != 4)
2312 for (unsigned i = 0; i < NumElems/2; ++i)
2313 if (!isUndefOrEqual(N->getMaskElt(i), i))
2316 for (unsigned i = 0; i < NumElems/2; ++i)
2317 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2323 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2324 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2326 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2327 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2332 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2333 isUndefOrEqual(N->getMaskElt(1), 3) &&
2334 isUndefOrEqual(N->getMaskElt(2), 2) &&
2335 isUndefOrEqual(N->getMaskElt(3), 3);
2338 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2339 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2340 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2341 bool V2IsSplat = false) {
2342 int NumElts = VT.getVectorNumElements();
2343 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2346 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2348 int BitI1 = Mask[i+1];
2349 if (!isUndefOrEqual(BitI, j))
2352 if (!isUndefOrEqual(BitI1, NumElts))
2355 if (!isUndefOrEqual(BitI1, j + NumElts))
2362 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2363 SmallVector<int, 8> M;
2365 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2368 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2369 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2370 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2371 bool V2IsSplat = false) {
2372 int NumElts = VT.getVectorNumElements();
2373 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2376 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2378 int BitI1 = Mask[i+1];
2379 if (!isUndefOrEqual(BitI, j + NumElts/2))
2382 if (isUndefOrEqual(BitI1, NumElts))
2385 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2392 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2393 SmallVector<int, 8> M;
2395 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2398 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2399 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2401 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2402 int NumElems = VT.getVectorNumElements();
2403 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2406 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2408 int BitI1 = Mask[i+1];
2409 if (!isUndefOrEqual(BitI, j))
2411 if (!isUndefOrEqual(BitI1, j))
2417 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2418 SmallVector<int, 8> M;
2420 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2423 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2424 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2426 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2427 int NumElems = VT.getVectorNumElements();
2428 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2431 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2433 int BitI1 = Mask[i+1];
2434 if (!isUndefOrEqual(BitI, j))
2436 if (!isUndefOrEqual(BitI1, j))
2442 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2443 SmallVector<int, 8> M;
2445 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2448 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2449 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2450 /// MOVSD, and MOVD, i.e. setting the lowest element.
2451 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2452 int NumElts = VT.getVectorNumElements();
2453 if (NumElts != 2 && NumElts != 4)
2456 if (!isUndefOrEqual(Mask[0], NumElts))
2459 for (int i = 1; i < NumElts; ++i)
2460 if (!isUndefOrEqual(Mask[i], i))
2466 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2467 SmallVector<int, 8> M;
2469 return ::isMOVLMask(M, N->getValueType(0));
2472 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2473 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2474 /// element of vector 2 and the other elements to come from vector 1 in order.
2475 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2476 bool V2IsSplat = false, bool V2IsUndef = false) {
2477 int NumOps = VT.getVectorNumElements();
2478 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2481 if (!isUndefOrEqual(Mask[0], 0))
2484 for (int i = 1; i < NumOps; ++i)
2485 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2486 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2487 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2493 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2494 bool V2IsUndef = false) {
2495 SmallVector<int, 8> M;
2497 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2500 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2501 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2502 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2503 if (N->getValueType(0).getVectorNumElements() != 4)
2506 // Expect 1, 1, 3, 3
2507 for (unsigned i = 0; i < 2; ++i) {
2508 int Elt = N->getMaskElt(i);
2509 if (Elt >= 0 && Elt != 1)
2514 for (unsigned i = 2; i < 4; ++i) {
2515 int Elt = N->getMaskElt(i);
2516 if (Elt >= 0 && Elt != 3)
2521 // Don't use movshdup if it can be done with a shufps.
2522 // FIXME: verify that matching u, u, 3, 3 is what we want.
2526 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2527 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2528 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2529 if (N->getValueType(0).getVectorNumElements() != 4)
2532 // Expect 0, 0, 2, 2
2533 for (unsigned i = 0; i < 2; ++i)
2534 if (N->getMaskElt(i) > 0)
2538 for (unsigned i = 2; i < 4; ++i) {
2539 int Elt = N->getMaskElt(i);
2540 if (Elt >= 0 && Elt != 2)
2545 // Don't use movsldup if it can be done with a shufps.
2549 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2550 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2551 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2552 int e = N->getValueType(0).getVectorNumElements() / 2;
2554 for (int i = 0; i < e; ++i)
2555 if (!isUndefOrEqual(N->getMaskElt(i), i))
2557 for (int i = 0; i < e; ++i)
2558 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2563 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2564 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2566 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2567 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2568 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2570 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2572 for (int i = 0; i < NumOperands; ++i) {
2573 int Val = SVOp->getMaskElt(NumOperands-i-1);
2574 if (Val < 0) Val = 0;
2575 if (Val >= NumOperands) Val -= NumOperands;
2577 if (i != NumOperands - 1)
2583 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2584 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2586 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2587 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2589 // 8 nodes, but we only care about the last 4.
2590 for (unsigned i = 7; i >= 4; --i) {
2591 int Val = SVOp->getMaskElt(i);
2600 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2601 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2603 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2604 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2606 // 8 nodes, but we only care about the first 4.
2607 for (int i = 3; i >= 0; --i) {
2608 int Val = SVOp->getMaskElt(i);
2617 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2618 /// their permute mask.
2619 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2620 SelectionDAG &DAG) {
2621 MVT VT = SVOp->getValueType(0);
2622 unsigned NumElems = VT.getVectorNumElements();
2623 SmallVector<int, 8> MaskVec;
2625 for (unsigned i = 0; i != NumElems; ++i) {
2626 int idx = SVOp->getMaskElt(i);
2628 MaskVec.push_back(idx);
2629 else if (idx < (int)NumElems)
2630 MaskVec.push_back(idx + NumElems);
2632 MaskVec.push_back(idx - NumElems);
2634 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2635 SVOp->getOperand(0), &MaskVec[0]);
2638 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2639 /// the two vector operands have swapped position.
2640 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2641 unsigned NumElems = VT.getVectorNumElements();
2642 for (unsigned i = 0; i != NumElems; ++i) {
2646 else if (idx < (int)NumElems)
2647 Mask[i] = idx + NumElems;
2649 Mask[i] = idx - NumElems;
2653 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2654 /// match movhlps. The lower half elements should come from upper half of
2655 /// V1 (and in order), and the upper half elements should come from the upper
2656 /// half of V2 (and in order).
2657 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2658 if (Op->getValueType(0).getVectorNumElements() != 4)
2660 for (unsigned i = 0, e = 2; i != e; ++i)
2661 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2663 for (unsigned i = 2; i != 4; ++i)
2664 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2669 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2670 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2672 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2673 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2675 N = N->getOperand(0).getNode();
2676 if (!ISD::isNON_EXTLoad(N))
2679 *LD = cast<LoadSDNode>(N);
2683 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2684 /// match movlp{s|d}. The lower half elements should come from lower half of
2685 /// V1 (and in order), and the upper half elements should come from the upper
2686 /// half of V2 (and in order). And since V1 will become the source of the
2687 /// MOVLP, it must be either a vector load or a scalar load to vector.
2688 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2689 ShuffleVectorSDNode *Op) {
2690 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2692 // Is V2 is a vector load, don't do this transformation. We will try to use
2693 // load folding shufps op.
2694 if (ISD::isNON_EXTLoad(V2))
2697 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2699 if (NumElems != 2 && NumElems != 4)
2701 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2702 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2704 for (unsigned i = NumElems/2; i != NumElems; ++i)
2705 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2710 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2712 static bool isSplatVector(SDNode *N) {
2713 if (N->getOpcode() != ISD::BUILD_VECTOR)
2716 SDValue SplatValue = N->getOperand(0);
2717 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2718 if (N->getOperand(i) != SplatValue)
2723 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2725 static inline bool isZeroNode(SDValue Elt) {
2726 return ((isa<ConstantSDNode>(Elt) &&
2727 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2728 (isa<ConstantFPSDNode>(Elt) &&
2729 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2732 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2733 /// to an zero vector.
2734 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2735 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2736 SDValue V1 = N->getOperand(0);
2737 SDValue V2 = N->getOperand(1);
2738 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2739 for (unsigned i = 0; i != NumElems; ++i) {
2740 int Idx = N->getMaskElt(i);
2741 if (Idx >= (int)NumElems) {
2742 unsigned Opc = V2.getOpcode();
2743 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2745 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2747 } else if (Idx >= 0) {
2748 unsigned Opc = V1.getOpcode();
2749 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2751 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2758 /// getZeroVector - Returns a vector of specified type with all zero elements.
2760 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2762 assert(VT.isVector() && "Expected a vector type");
2764 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2765 // type. This ensures they get CSE'd.
2767 if (VT.getSizeInBits() == 64) { // MMX
2768 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2769 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2770 } else if (HasSSE2) { // SSE2
2771 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2772 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2774 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2775 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2777 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2780 /// getOnesVector - Returns a vector of specified type with all bits set.
2782 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2783 assert(VT.isVector() && "Expected a vector type");
2785 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2786 // type. This ensures they get CSE'd.
2787 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2789 if (VT.getSizeInBits() == 64) // MMX
2790 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2792 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2793 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2797 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2798 /// that point to V2 points to its first element.
2799 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2800 MVT VT = SVOp->getValueType(0);
2801 unsigned NumElems = VT.getVectorNumElements();
2803 bool Changed = false;
2804 SmallVector<int, 8> MaskVec;
2805 SVOp->getMask(MaskVec);
2807 for (unsigned i = 0; i != NumElems; ++i) {
2808 if (MaskVec[i] > (int)NumElems) {
2809 MaskVec[i] = NumElems;
2814 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2815 SVOp->getOperand(1), &MaskVec[0]);
2816 return SDValue(SVOp, 0);
2819 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2820 /// operation of specified width.
2821 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2823 unsigned NumElems = VT.getVectorNumElements();
2824 SmallVector<int, 8> Mask;
2825 Mask.push_back(NumElems);
2826 for (unsigned i = 1; i != NumElems; ++i)
2828 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2831 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2832 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2834 unsigned NumElems = VT.getVectorNumElements();
2835 SmallVector<int, 8> Mask;
2836 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2838 Mask.push_back(i + NumElems);
2840 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2843 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2844 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2846 unsigned NumElems = VT.getVectorNumElements();
2847 unsigned Half = NumElems/2;
2848 SmallVector<int, 8> Mask;
2849 for (unsigned i = 0; i != Half; ++i) {
2850 Mask.push_back(i + Half);
2851 Mask.push_back(i + NumElems + Half);
2853 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2856 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2857 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2859 if (SV->getValueType(0).getVectorNumElements() <= 4)
2860 return SDValue(SV, 0);
2862 MVT PVT = MVT::v4f32;
2863 MVT VT = SV->getValueType(0);
2864 DebugLoc dl = SV->getDebugLoc();
2865 SDValue V1 = SV->getOperand(0);
2866 int NumElems = VT.getVectorNumElements();
2867 int EltNo = SV->getSplatIndex();
2869 // unpack elements to the correct location
2870 while (NumElems > 4) {
2871 if (EltNo < NumElems/2) {
2872 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2874 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2875 EltNo -= NumElems/2;
2880 // Perform the splat.
2881 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2882 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2883 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2884 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2887 /// isVectorLoad - Returns true if the node is a vector load, a scalar
2888 /// load that's promoted to vector, or a load bitcasted.
2889 static bool isVectorLoad(SDValue Op) {
2890 assert(Op.getValueType().isVector() && "Expected a vector type");
2891 if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
2892 Op.getOpcode() == ISD::BIT_CONVERT) {
2893 return isa<LoadSDNode>(Op.getOperand(0));
2895 return isa<LoadSDNode>(Op);
2899 /// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
2901 static SDValue CanonicalizeMovddup(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2903 // If we have sse3 and shuffle has more than one use or input is a load, then
2904 // use movddup. Otherwise, use movlhps.
2905 SDValue V1 = SV->getOperand(0);
2907 bool UseMovddup = HasSSE3 && (!SV->hasOneUse() || isVectorLoad(V1));
2908 MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
2909 MVT VT = SV->getValueType(0);
2911 return SDValue(SV, 0);
2913 DebugLoc dl = SV->getDebugLoc();
2914 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2915 if (PVT.getVectorNumElements() == 2) {
2916 int Mask[2] = { 0, 0 };
2917 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), Mask);
2919 int Mask[4] = { 0, 1, 0, 1 };
2920 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), Mask);
2922 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2925 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2926 /// vector of zero or undef vector. This produces a shuffle where the low
2927 /// element of V2 is swizzled into the zero/undef vector, landing at element
2928 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2929 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2930 bool isZero, bool HasSSE2,
2931 SelectionDAG &DAG) {
2932 MVT VT = V2.getValueType();
2934 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
2935 unsigned NumElems = VT.getVectorNumElements();
2936 SmallVector<int, 16> MaskVec;
2937 for (unsigned i = 0; i != NumElems; ++i)
2938 // If this is the insertion idx, put the low elt of V2 here.
2939 MaskVec.push_back(i == Idx ? NumElems : i);
2940 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
2943 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2944 /// a shuffle that is zero.
2946 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
2947 bool Low, SelectionDAG &DAG) {
2948 unsigned NumZeros = 0;
2949 for (int i = 0; i < NumElems; ++i) {
2950 unsigned Index = Low ? i : NumElems-i-1;
2951 int Idx = SVOp->getMaskElt(Index);
2956 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
2957 if (Elt.getNode() && isZeroNode(Elt))
2965 /// isVectorShift - Returns true if the shuffle can be implemented as a
2966 /// logical left or right shift of a vector.
2967 /// FIXME: split into pslldqi, psrldqi, palignr variants.
2968 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
2969 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
2970 int NumElems = SVOp->getValueType(0).getVectorNumElements();
2973 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
2976 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
2980 bool SeenV1 = false;
2981 bool SeenV2 = false;
2982 for (int i = NumZeros; i < NumElems; ++i) {
2983 int Val = isLeft ? (i - NumZeros) : i;
2984 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
2996 if (SeenV1 && SeenV2)
2999 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3005 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3007 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3008 unsigned NumNonZero, unsigned NumZero,
3009 SelectionDAG &DAG, TargetLowering &TLI) {
3013 DebugLoc dl = Op.getDebugLoc();
3016 for (unsigned i = 0; i < 16; ++i) {
3017 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3018 if (ThisIsNonZero && First) {
3020 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3022 V = DAG.getUNDEF(MVT::v8i16);
3027 SDValue ThisElt(0, 0), LastElt(0, 0);
3028 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3029 if (LastIsNonZero) {
3030 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3031 MVT::i16, Op.getOperand(i-1));
3033 if (ThisIsNonZero) {
3034 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3035 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3036 ThisElt, DAG.getConstant(8, MVT::i8));
3038 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3042 if (ThisElt.getNode())
3043 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3044 DAG.getIntPtrConstant(i/2));
3048 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3051 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3053 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3054 unsigned NumNonZero, unsigned NumZero,
3055 SelectionDAG &DAG, TargetLowering &TLI) {
3059 DebugLoc dl = Op.getDebugLoc();
3062 for (unsigned i = 0; i < 8; ++i) {
3063 bool isNonZero = (NonZeros & (1 << i)) != 0;
3067 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3069 V = DAG.getUNDEF(MVT::v8i16);
3072 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3073 MVT::v8i16, V, Op.getOperand(i),
3074 DAG.getIntPtrConstant(i));
3081 /// getVShift - Return a vector logical shift node.
3083 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3084 unsigned NumBits, SelectionDAG &DAG,
3085 const TargetLowering &TLI, DebugLoc dl) {
3086 bool isMMX = VT.getSizeInBits() == 64;
3087 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3088 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3089 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3090 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3091 DAG.getNode(Opc, dl, ShVT, SrcOp,
3092 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3096 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3097 DebugLoc dl = Op.getDebugLoc();
3098 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3099 if (ISD::isBuildVectorAllZeros(Op.getNode())
3100 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3101 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3102 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3103 // eliminated on x86-32 hosts.
3104 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3107 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3108 return getOnesVector(Op.getValueType(), DAG, dl);
3109 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3112 MVT VT = Op.getValueType();
3113 MVT EVT = VT.getVectorElementType();
3114 unsigned EVTBits = EVT.getSizeInBits();
3116 unsigned NumElems = Op.getNumOperands();
3117 unsigned NumZero = 0;
3118 unsigned NumNonZero = 0;
3119 unsigned NonZeros = 0;
3120 bool IsAllConstants = true;
3121 SmallSet<SDValue, 8> Values;
3122 for (unsigned i = 0; i < NumElems; ++i) {
3123 SDValue Elt = Op.getOperand(i);
3124 if (Elt.getOpcode() == ISD::UNDEF)
3127 if (Elt.getOpcode() != ISD::Constant &&
3128 Elt.getOpcode() != ISD::ConstantFP)
3129 IsAllConstants = false;
3130 if (isZeroNode(Elt))
3133 NonZeros |= (1 << i);
3138 if (NumNonZero == 0) {
3139 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3140 return DAG.getUNDEF(VT);
3143 // Special case for single non-zero, non-undef, element.
3144 if (NumNonZero == 1 && NumElems <= 4) {
3145 unsigned Idx = CountTrailingZeros_32(NonZeros);
3146 SDValue Item = Op.getOperand(Idx);
3148 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3149 // the value are obviously zero, truncate the value to i32 and do the
3150 // insertion that way. Only do this if the value is non-constant or if the
3151 // value is a constant being inserted into element 0. It is cheaper to do
3152 // a constant pool load than it is to do a movd + shuffle.
3153 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3154 (!IsAllConstants || Idx == 0)) {
3155 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3156 // Handle MMX and SSE both.
3157 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3158 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3160 // Truncate the value (which may itself be a constant) to i32, and
3161 // convert it to a vector with movd (S2V+shuffle to zero extend).
3162 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3163 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3164 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3165 Subtarget->hasSSE2(), DAG);
3167 // Now we have our 32-bit value zero extended in the low element of
3168 // a vector. If Idx != 0, swizzle it into place.
3170 SmallVector<int, 4> Mask;
3171 Mask.push_back(Idx);
3172 for (unsigned i = 1; i != VecElts; ++i)
3174 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3175 DAG.getUNDEF(Item.getValueType()),
3178 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3182 // If we have a constant or non-constant insertion into the low element of
3183 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3184 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3185 // depending on what the source datatype is. Because we can only get here
3186 // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
3188 // Don't do this for i64 values on x86-32.
3189 (EVT != MVT::i64 || Subtarget->is64Bit())) {
3190 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3191 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3192 return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3193 Subtarget->hasSSE2(), DAG);
3196 // Is it a vector logical left shift?
3197 if (NumElems == 2 && Idx == 1 &&
3198 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3199 unsigned NumBits = VT.getSizeInBits();
3200 return getVShift(true, VT,
3201 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3202 VT, Op.getOperand(1)),
3203 NumBits/2, DAG, *this, dl);
3206 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3209 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3210 // is a non-constant being inserted into an element other than the low one,
3211 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3212 // movd/movss) to move this into the low element, then shuffle it into
3214 if (EVTBits == 32) {
3215 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3217 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3218 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3219 Subtarget->hasSSE2(), DAG);
3220 SmallVector<int, 8> MaskVec;
3221 for (unsigned i = 0; i < NumElems; i++)
3222 MaskVec.push_back(i == Idx ? 0 : 1);
3223 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3227 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3228 if (Values.size() == 1)
3231 // A vector full of immediates; various special cases are already
3232 // handled, so this is best done with a single constant-pool load.
3236 // Let legalizer expand 2-wide build_vectors.
3237 if (EVTBits == 64) {
3238 if (NumNonZero == 1) {
3239 // One half is zero or undef.
3240 unsigned Idx = CountTrailingZeros_32(NonZeros);
3241 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3242 Op.getOperand(Idx));
3243 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3244 Subtarget->hasSSE2(), DAG);
3249 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3250 if (EVTBits == 8 && NumElems == 16) {
3251 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3253 if (V.getNode()) return V;
3256 if (EVTBits == 16 && NumElems == 8) {
3257 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3259 if (V.getNode()) return V;
3262 // If element VT is == 32 bits, turn it into a number of shuffles.
3263 SmallVector<SDValue, 8> V;
3265 if (NumElems == 4 && NumZero > 0) {
3266 for (unsigned i = 0; i < 4; ++i) {
3267 bool isZero = !(NonZeros & (1 << i));
3269 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3271 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3274 for (unsigned i = 0; i < 2; ++i) {
3275 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3278 V[i] = V[i*2]; // Must be a zero vector.
3281 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3284 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3287 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3292 SmallVector<int, 8> MaskVec;
3293 bool Reverse = (NonZeros & 0x3) == 2;
3294 for (unsigned i = 0; i < 2; ++i)
3295 MaskVec.push_back(Reverse ? 1-i : i);
3296 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3297 for (unsigned i = 0; i < 2; ++i)
3298 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3299 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3302 if (Values.size() > 2) {
3303 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3304 // values to be inserted is equal to the number of elements, in which case
3305 // use the unpack code below in the hopes of matching the consecutive elts
3306 // load merge pattern for shuffles.
3307 // FIXME: We could probably just check that here directly.
3308 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3309 getSubtarget()->hasSSE41()) {
3310 V[0] = DAG.getUNDEF(VT);
3311 for (unsigned i = 0; i < NumElems; ++i)
3312 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3313 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3314 Op.getOperand(i), DAG.getIntPtrConstant(i));
3317 // Expand into a number of unpckl*.
3319 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3320 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3321 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3322 for (unsigned i = 0; i < NumElems; ++i)
3323 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3325 while (NumElems != 0) {
3326 for (unsigned i = 0; i < NumElems; ++i)
3327 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3336 // v8i16 shuffles - Prefer shuffles in the following order:
3337 // 1. [all] pshuflw, pshufhw, optional move
3338 // 2. [ssse3] 1 x pshufb
3339 // 3. [ssse3] 2 x pshufb + 1 x por
3340 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3342 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3343 SelectionDAG &DAG, X86TargetLowering &TLI) {
3344 SDValue V1 = SVOp->getOperand(0);
3345 SDValue V2 = SVOp->getOperand(1);
3346 DebugLoc dl = SVOp->getDebugLoc();
3347 SmallVector<int, 8> MaskVals;
3349 // Determine if more than 1 of the words in each of the low and high quadwords
3350 // of the result come from the same quadword of one of the two inputs. Undef
3351 // mask values count as coming from any quadword, for better codegen.
3352 SmallVector<unsigned, 4> LoQuad(4);
3353 SmallVector<unsigned, 4> HiQuad(4);
3354 BitVector InputQuads(4);
3355 for (unsigned i = 0; i < 8; ++i) {
3356 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3357 int EltIdx = SVOp->getMaskElt(i);
3358 MaskVals.push_back(EltIdx);
3367 InputQuads.set(EltIdx / 4);
3370 int BestLoQuad = -1;
3371 unsigned MaxQuad = 1;
3372 for (unsigned i = 0; i < 4; ++i) {
3373 if (LoQuad[i] > MaxQuad) {
3375 MaxQuad = LoQuad[i];
3379 int BestHiQuad = -1;
3381 for (unsigned i = 0; i < 4; ++i) {
3382 if (HiQuad[i] > MaxQuad) {
3384 MaxQuad = HiQuad[i];
3388 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3389 // of the two input vectors, shuffle them into one input vector so only a
3390 // single pshufb instruction is necessary. If There are more than 2 input
3391 // quads, disable the next transformation since it does not help SSSE3.
3392 bool V1Used = InputQuads[0] || InputQuads[1];
3393 bool V2Used = InputQuads[2] || InputQuads[3];
3394 if (TLI.getSubtarget()->hasSSSE3()) {
3395 if (InputQuads.count() == 2 && V1Used && V2Used) {
3396 BestLoQuad = InputQuads.find_first();
3397 BestHiQuad = InputQuads.find_next(BestLoQuad);
3399 if (InputQuads.count() > 2) {
3405 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3406 // the shuffle mask. If a quad is scored as -1, that means that it contains
3407 // words from all 4 input quadwords.
3409 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3410 SmallVector<int, 8> MaskV;
3411 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3412 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3413 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3414 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3415 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3416 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3418 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3419 // source words for the shuffle, to aid later transformations.
3420 bool AllWordsInNewV = true;
3421 bool InOrder[2] = { true, true };
3422 for (unsigned i = 0; i != 8; ++i) {
3423 int idx = MaskVals[i];
3425 InOrder[i/4] = false;
3426 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3428 AllWordsInNewV = false;
3432 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3433 if (AllWordsInNewV) {
3434 for (int i = 0; i != 8; ++i) {
3435 int idx = MaskVals[i];
3438 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3439 if ((idx != i) && idx < 4)
3441 if ((idx != i) && idx > 3)
3450 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3451 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3452 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3453 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3454 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3458 // If we have SSSE3, and all words of the result are from 1 input vector,
3459 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3460 // is present, fall back to case 4.
3461 if (TLI.getSubtarget()->hasSSSE3()) {
3462 SmallVector<SDValue,16> pshufbMask;
3464 // If we have elements from both input vectors, set the high bit of the
3465 // shuffle mask element to zero out elements that come from V2 in the V1
3466 // mask, and elements that come from V1 in the V2 mask, so that the two
3467 // results can be OR'd together.
3468 bool TwoInputs = V1Used && V2Used;
3469 for (unsigned i = 0; i != 8; ++i) {
3470 int EltIdx = MaskVals[i] * 2;
3471 if (TwoInputs && (EltIdx >= 16)) {
3472 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3473 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3476 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3477 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3479 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3480 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3481 DAG.getNode(ISD::BUILD_VECTOR, dl,
3482 MVT::v16i8, &pshufbMask[0], 16));
3484 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3486 // Calculate the shuffle mask for the second input, shuffle it, and
3487 // OR it with the first shuffled input.
3489 for (unsigned i = 0; i != 8; ++i) {
3490 int EltIdx = MaskVals[i] * 2;
3492 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3493 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3496 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3497 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3499 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3500 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3501 DAG.getNode(ISD::BUILD_VECTOR, dl,
3502 MVT::v16i8, &pshufbMask[0], 16));
3503 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3504 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3507 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3508 // and update MaskVals with new element order.
3509 BitVector InOrder(8);
3510 if (BestLoQuad >= 0) {
3511 SmallVector<int, 8> MaskV;
3512 for (int i = 0; i != 4; ++i) {
3513 int idx = MaskVals[i];
3515 MaskV.push_back(-1);
3517 } else if ((idx / 4) == BestLoQuad) {
3518 MaskV.push_back(idx & 3);
3521 MaskV.push_back(-1);
3524 for (unsigned i = 4; i != 8; ++i)
3526 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3530 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3531 // and update MaskVals with the new element order.
3532 if (BestHiQuad >= 0) {
3533 SmallVector<int, 8> MaskV;
3534 for (unsigned i = 0; i != 4; ++i)
3536 for (unsigned i = 4; i != 8; ++i) {
3537 int idx = MaskVals[i];
3539 MaskV.push_back(-1);
3541 } else if ((idx / 4) == BestHiQuad) {
3542 MaskV.push_back((idx & 3) + 4);
3545 MaskV.push_back(-1);
3548 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3552 // In case BestHi & BestLo were both -1, which means each quadword has a word
3553 // from each of the four input quadwords, calculate the InOrder bitvector now
3554 // before falling through to the insert/extract cleanup.
3555 if (BestLoQuad == -1 && BestHiQuad == -1) {
3557 for (int i = 0; i != 8; ++i)
3558 if (MaskVals[i] < 0 || MaskVals[i] == i)
3562 // The other elements are put in the right place using pextrw and pinsrw.
3563 for (unsigned i = 0; i != 8; ++i) {
3566 int EltIdx = MaskVals[i];
3569 SDValue ExtOp = (EltIdx < 8)
3570 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3571 DAG.getIntPtrConstant(EltIdx))
3572 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3573 DAG.getIntPtrConstant(EltIdx - 8));
3574 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3575 DAG.getIntPtrConstant(i));
3580 // v16i8 shuffles - Prefer shuffles in the following order:
3581 // 1. [ssse3] 1 x pshufb
3582 // 2. [ssse3] 2 x pshufb + 1 x por
3583 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3585 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3586 SelectionDAG &DAG, X86TargetLowering &TLI) {
3587 SDValue V1 = SVOp->getOperand(0);
3588 SDValue V2 = SVOp->getOperand(1);
3589 DebugLoc dl = SVOp->getDebugLoc();
3590 SmallVector<int, 16> MaskVals;
3591 SVOp->getMask(MaskVals);
3593 // If we have SSSE3, case 1 is generated when all result bytes come from
3594 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3595 // present, fall back to case 3.
3596 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3599 for (unsigned i = 0; i < 16; ++i) {
3600 int EltIdx = MaskVals[i];
3609 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3610 if (TLI.getSubtarget()->hasSSSE3()) {
3611 SmallVector<SDValue,16> pshufbMask;
3613 // If all result elements are from one input vector, then only translate
3614 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3616 // Otherwise, we have elements from both input vectors, and must zero out
3617 // elements that come from V2 in the first mask, and V1 in the second mask
3618 // so that we can OR them together.
3619 bool TwoInputs = !(V1Only || V2Only);
3620 for (unsigned i = 0; i != 16; ++i) {
3621 int EltIdx = MaskVals[i];
3622 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3623 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3626 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3628 // If all the elements are from V2, assign it to V1 and return after
3629 // building the first pshufb.
3632 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3633 DAG.getNode(ISD::BUILD_VECTOR, dl,
3634 MVT::v16i8, &pshufbMask[0], 16));
3638 // Calculate the shuffle mask for the second input, shuffle it, and
3639 // OR it with the first shuffled input.
3641 for (unsigned i = 0; i != 16; ++i) {
3642 int EltIdx = MaskVals[i];
3644 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3647 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3649 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3650 DAG.getNode(ISD::BUILD_VECTOR, dl,
3651 MVT::v16i8, &pshufbMask[0], 16));
3652 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3655 // No SSSE3 - Calculate in place words and then fix all out of place words
3656 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3657 // the 16 different words that comprise the two doublequadword input vectors.
3658 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3659 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3660 SDValue NewV = V2Only ? V2 : V1;
3661 for (int i = 0; i != 8; ++i) {
3662 int Elt0 = MaskVals[i*2];
3663 int Elt1 = MaskVals[i*2+1];
3665 // This word of the result is all undef, skip it.
3666 if (Elt0 < 0 && Elt1 < 0)
3669 // This word of the result is already in the correct place, skip it.
3670 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3672 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3675 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3676 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3679 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3680 // using a single extract together, load it and store it.
3681 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3682 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3683 DAG.getIntPtrConstant(Elt1 / 2));
3684 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3685 DAG.getIntPtrConstant(i));
3689 // If Elt1 is defined, extract it from the appropriate source. If the
3690 // source byte is not also odd, shift the extracted word left 8 bits
3691 // otherwise clear the bottom 8 bits if we need to do an or.
3693 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3694 DAG.getIntPtrConstant(Elt1 / 2));
3695 if ((Elt1 & 1) == 0)
3696 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3697 DAG.getConstant(8, TLI.getShiftAmountTy()));
3699 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3700 DAG.getConstant(0xFF00, MVT::i16));
3702 // If Elt0 is defined, extract it from the appropriate source. If the
3703 // source byte is not also even, shift the extracted word right 8 bits. If
3704 // Elt1 was also defined, OR the extracted values together before
3705 // inserting them in the result.
3707 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3708 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3709 if ((Elt0 & 1) != 0)
3710 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3711 DAG.getConstant(8, TLI.getShiftAmountTy()));
3713 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3714 DAG.getConstant(0x00FF, MVT::i16));
3715 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3718 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3719 DAG.getIntPtrConstant(i));
3721 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3724 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3725 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3726 /// done when every pair / quad of shuffle mask elements point to elements in
3727 /// the right sequence. e.g.
3728 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3730 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3732 TargetLowering &TLI, DebugLoc dl) {
3733 MVT VT = SVOp->getValueType(0);
3734 SDValue V1 = SVOp->getOperand(0);
3735 SDValue V2 = SVOp->getOperand(1);
3736 unsigned NumElems = VT.getVectorNumElements();
3737 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3738 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3739 MVT MaskEltVT = MaskVT.getVectorElementType();
3741 switch (VT.getSimpleVT()) {
3742 default: assert(false && "Unexpected!");
3743 case MVT::v4f32: NewVT = MVT::v2f64; break;
3744 case MVT::v4i32: NewVT = MVT::v2i64; break;
3745 case MVT::v8i16: NewVT = MVT::v4i32; break;
3746 case MVT::v16i8: NewVT = MVT::v4i32; break;
3749 if (NewWidth == 2) {
3755 int Scale = NumElems / NewWidth;
3756 SmallVector<int, 8> MaskVec;
3757 for (unsigned i = 0; i < NumElems; i += Scale) {
3759 for (int j = 0; j < Scale; ++j) {
3760 int EltIdx = SVOp->getMaskElt(i+j);
3764 StartIdx = EltIdx - (EltIdx % Scale);
3765 if (EltIdx != StartIdx + j)
3769 MaskVec.push_back(-1);
3771 MaskVec.push_back(StartIdx / Scale);
3774 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3775 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3776 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3779 /// getVZextMovL - Return a zero-extending vector move low node.
3781 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3782 SDValue SrcOp, SelectionDAG &DAG,
3783 const X86Subtarget *Subtarget, DebugLoc dl) {
3784 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3785 LoadSDNode *LD = NULL;
3786 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3787 LD = dyn_cast<LoadSDNode>(SrcOp);
3789 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3791 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3792 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3793 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3794 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3795 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3797 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3798 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3799 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3800 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3808 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3809 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3810 DAG.getNode(ISD::BIT_CONVERT, dl,
3814 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3817 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3818 SDValue V1 = SVOp->getOperand(0);
3819 SDValue V2 = SVOp->getOperand(1);
3820 DebugLoc dl = SVOp->getDebugLoc();
3821 MVT VT = SVOp->getValueType(0);
3823 SmallVector<std::pair<int, int>, 8> Locs;
3825 SmallVector<int, 8> Mask1(4U, -1);
3826 SmallVector<int, 8> PermMask;
3827 SVOp->getMask(PermMask);
3831 for (unsigned i = 0; i != 4; ++i) {
3832 int Idx = PermMask[i];
3834 Locs[i] = std::make_pair(-1, -1);
3836 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3838 Locs[i] = std::make_pair(0, NumLo);
3842 Locs[i] = std::make_pair(1, NumHi);
3844 Mask1[2+NumHi] = Idx;
3850 if (NumLo <= 2 && NumHi <= 2) {
3851 // If no more than two elements come from either vector. This can be
3852 // implemented with two shuffles. First shuffle gather the elements.
3853 // The second shuffle, which takes the first shuffle as both of its
3854 // vector operands, put the elements into the right order.
3855 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3857 SmallVector<int, 8> Mask2(4U, -1);
3859 for (unsigned i = 0; i != 4; ++i) {
3860 if (Locs[i].first == -1)
3863 unsigned Idx = (i < 2) ? 0 : 4;
3864 Idx += Locs[i].first * 2 + Locs[i].second;
3869 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3870 } else if (NumLo == 3 || NumHi == 3) {
3871 // Otherwise, we must have three elements from one vector, call it X, and
3872 // one element from the other, call it Y. First, use a shufps to build an
3873 // intermediate vector with the one element from Y and the element from X
3874 // that will be in the same half in the final destination (the indexes don't
3875 // matter). Then, use a shufps to build the final vector, taking the half
3876 // containing the element from Y from the intermediate, and the other half
3879 // Normalize it so the 3 elements come from V1.
3880 CommuteVectorShuffleMask(PermMask, VT);
3884 // Find the element from V2.
3886 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3887 int Val = PermMask[HiIndex];
3894 Mask1[0] = PermMask[HiIndex];
3896 Mask1[2] = PermMask[HiIndex^1];
3898 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3901 Mask1[0] = PermMask[0];
3902 Mask1[1] = PermMask[1];
3903 Mask1[2] = HiIndex & 1 ? 6 : 4;
3904 Mask1[3] = HiIndex & 1 ? 4 : 6;
3905 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3907 Mask1[0] = HiIndex & 1 ? 2 : 0;
3908 Mask1[1] = HiIndex & 1 ? 0 : 2;
3909 Mask1[2] = PermMask[2];
3910 Mask1[3] = PermMask[3];
3915 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3919 // Break it into (shuffle shuffle_hi, shuffle_lo).
3921 SmallVector<int,8> LoMask(4U, -1);
3922 SmallVector<int,8> HiMask(4U, -1);
3924 SmallVector<int,8> *MaskPtr = &LoMask;
3925 unsigned MaskIdx = 0;
3928 for (unsigned i = 0; i != 4; ++i) {
3935 int Idx = PermMask[i];
3937 Locs[i] = std::make_pair(-1, -1);
3938 } else if (Idx < 4) {
3939 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3940 (*MaskPtr)[LoIdx] = Idx;
3943 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3944 (*MaskPtr)[HiIdx] = Idx;
3949 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
3950 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
3951 SmallVector<int, 8> MaskOps;
3952 for (unsigned i = 0; i != 4; ++i) {
3953 if (Locs[i].first == -1) {
3954 MaskOps.push_back(-1);
3956 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3957 MaskOps.push_back(Idx);
3960 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
3964 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3965 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3966 SDValue V1 = Op.getOperand(0);
3967 SDValue V2 = Op.getOperand(1);
3968 MVT VT = Op.getValueType();
3969 DebugLoc dl = Op.getDebugLoc();
3970 unsigned NumElems = VT.getVectorNumElements();
3971 bool isMMX = VT.getSizeInBits() == 64;
3972 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3973 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3974 bool V1IsSplat = false;
3975 bool V2IsSplat = false;
3977 if (isZeroShuffle(SVOp))
3978 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3980 // Canonicalize movddup shuffles.
3981 if (V2IsUndef && Subtarget->hasSSE2() && VT.getSizeInBits() == 128 &&
3982 X86::isMOVDDUPMask(SVOp))
3983 return CanonicalizeMovddup(SVOp, DAG, Subtarget->hasSSE3());
3985 // Promote splats to v4f32.
3986 if (SVOp->isSplat()) {
3987 if (isMMX || NumElems < 4)
3989 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
3992 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3994 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3995 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3996 if (NewOp.getNode())
3997 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3998 LowerVECTOR_SHUFFLE(NewOp, DAG));
3999 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4000 // FIXME: Figure out a cleaner way to do this.
4001 // Try to make use of movq to zero out the top part.
4002 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4003 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4004 if (NewOp.getNode()) {
4005 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4006 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4007 DAG, Subtarget, dl);
4009 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4010 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4011 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4012 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4013 DAG, Subtarget, dl);
4017 if (X86::isPSHUFDMask(SVOp))
4020 // Check if this can be converted into a logical shift.
4021 bool isLeft = false;
4024 bool isShift = getSubtarget()->hasSSE2() &&
4025 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4026 if (isShift && ShVal.hasOneUse()) {
4027 // If the shifted value has multiple uses, it may be cheaper to use
4028 // v_set0 + movlhps or movhlps, etc.
4029 MVT EVT = VT.getVectorElementType();
4030 ShAmt *= EVT.getSizeInBits();
4031 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4034 if (X86::isMOVLMask(SVOp)) {
4037 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4038 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4043 // FIXME: fold these into legal mask.
4044 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4045 X86::isMOVSLDUPMask(SVOp) ||
4046 X86::isMOVHLPSMask(SVOp) ||
4047 X86::isMOVHPMask(SVOp) ||
4048 X86::isMOVLPMask(SVOp)))
4051 if (ShouldXformToMOVHLPS(SVOp) ||
4052 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4053 return CommuteVectorShuffle(SVOp, DAG);
4056 // No better options. Use a vshl / vsrl.
4057 MVT EVT = VT.getVectorElementType();
4058 ShAmt *= EVT.getSizeInBits();
4059 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4062 bool Commuted = false;
4063 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4064 // 1,1,1,1 -> v8i16 though.
4065 V1IsSplat = isSplatVector(V1.getNode());
4066 V2IsSplat = isSplatVector(V2.getNode());
4068 // Canonicalize the splat or undef, if present, to be on the RHS.
4069 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4070 Op = CommuteVectorShuffle(SVOp, DAG);
4071 SVOp = cast<ShuffleVectorSDNode>(Op);
4072 V1 = SVOp->getOperand(0);
4073 V2 = SVOp->getOperand(1);
4074 std::swap(V1IsSplat, V2IsSplat);
4075 std::swap(V1IsUndef, V2IsUndef);
4079 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4080 // Shuffling low element of v1 into undef, just return v1.
4083 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4084 // the instruction selector will not match, so get a canonical MOVL with
4085 // swapped operands to undo the commute.
4086 return getMOVL(DAG, dl, VT, V2, V1);
4089 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4090 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4091 X86::isUNPCKLMask(SVOp) ||
4092 X86::isUNPCKHMask(SVOp))
4096 // Normalize mask so all entries that point to V2 points to its first
4097 // element then try to match unpck{h|l} again. If match, return a
4098 // new vector_shuffle with the corrected mask.
4099 SDValue NewMask = NormalizeMask(SVOp, DAG);
4100 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4101 if (NSVOp != SVOp) {
4102 if (X86::isUNPCKLMask(NSVOp, true)) {
4104 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4111 // Commute is back and try unpck* again.
4112 // FIXME: this seems wrong.
4113 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4114 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4115 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4116 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4117 X86::isUNPCKLMask(NewSVOp) ||
4118 X86::isUNPCKHMask(NewSVOp))
4122 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4124 // Normalize the node to match x86 shuffle ops if needed
4125 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4126 return CommuteVectorShuffle(SVOp, DAG);
4128 // Check for legal shuffle and return?
4129 SmallVector<int, 16> PermMask;
4130 SVOp->getMask(PermMask);
4131 if (isShuffleMaskLegal(PermMask, VT))
4134 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4135 if (VT == MVT::v8i16) {
4136 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4137 if (NewOp.getNode())
4141 if (VT == MVT::v16i8) {
4142 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4143 if (NewOp.getNode())
4147 // Handle all 4 wide cases with a number of shuffles except for MMX.
4148 if (NumElems == 4 && !isMMX)
4149 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4155 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4156 SelectionDAG &DAG) {
4157 MVT VT = Op.getValueType();
4158 DebugLoc dl = Op.getDebugLoc();
4159 if (VT.getSizeInBits() == 8) {
4160 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4161 Op.getOperand(0), Op.getOperand(1));
4162 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4163 DAG.getValueType(VT));
4164 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4165 } else if (VT.getSizeInBits() == 16) {
4166 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4167 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4169 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4170 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4171 DAG.getNode(ISD::BIT_CONVERT, dl,
4175 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4176 Op.getOperand(0), Op.getOperand(1));
4177 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4178 DAG.getValueType(VT));
4179 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4180 } else if (VT == MVT::f32) {
4181 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4182 // the result back to FR32 register. It's only worth matching if the
4183 // result has a single use which is a store or a bitcast to i32. And in
4184 // the case of a store, it's not worth it if the index is a constant 0,
4185 // because a MOVSSmr can be used instead, which is smaller and faster.
4186 if (!Op.hasOneUse())
4188 SDNode *User = *Op.getNode()->use_begin();
4189 if ((User->getOpcode() != ISD::STORE ||
4190 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4191 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4192 (User->getOpcode() != ISD::BIT_CONVERT ||
4193 User->getValueType(0) != MVT::i32))
4195 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4196 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4199 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4200 } else if (VT == MVT::i32) {
4201 // ExtractPS works with constant index.
4202 if (isa<ConstantSDNode>(Op.getOperand(1)))
4210 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4211 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4214 if (Subtarget->hasSSE41()) {
4215 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4220 MVT VT = Op.getValueType();
4221 DebugLoc dl = Op.getDebugLoc();
4222 // TODO: handle v16i8.
4223 if (VT.getSizeInBits() == 16) {
4224 SDValue Vec = Op.getOperand(0);
4225 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4227 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4228 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4229 DAG.getNode(ISD::BIT_CONVERT, dl,
4232 // Transform it so it match pextrw which produces a 32-bit result.
4233 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4234 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4235 Op.getOperand(0), Op.getOperand(1));
4236 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4237 DAG.getValueType(VT));
4238 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4239 } else if (VT.getSizeInBits() == 32) {
4240 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4244 // SHUFPS the element to the lowest double word, then movss.
4245 int Mask[4] = { Idx, -1, -1, -1 };
4246 MVT VVT = Op.getOperand(0).getValueType();
4247 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4248 DAG.getUNDEF(VVT), Mask);
4249 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4250 DAG.getIntPtrConstant(0));
4251 } else if (VT.getSizeInBits() == 64) {
4252 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4253 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4254 // to match extract_elt for f64.
4255 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4259 // UNPCKHPD the element to the lowest double word, then movsd.
4260 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4261 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4262 int Mask[2] = { 1, -1 };
4263 MVT VVT = Op.getOperand(0).getValueType();
4264 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4265 DAG.getUNDEF(VVT), Mask);
4266 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4267 DAG.getIntPtrConstant(0));
4274 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4275 MVT VT = Op.getValueType();
4276 MVT EVT = VT.getVectorElementType();
4277 DebugLoc dl = Op.getDebugLoc();
4279 SDValue N0 = Op.getOperand(0);
4280 SDValue N1 = Op.getOperand(1);
4281 SDValue N2 = Op.getOperand(2);
4283 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4284 isa<ConstantSDNode>(N2)) {
4285 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4287 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4289 if (N1.getValueType() != MVT::i32)
4290 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4291 if (N2.getValueType() != MVT::i32)
4292 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4293 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4294 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4295 // Bits [7:6] of the constant are the source select. This will always be
4296 // zero here. The DAG Combiner may combine an extract_elt index into these
4297 // bits. For example (insert (extract, 3), 2) could be matched by putting
4298 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4299 // Bits [5:4] of the constant are the destination select. This is the
4300 // value of the incoming immediate.
4301 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4302 // combine either bitwise AND or insert of float 0.0 to set these bits.
4303 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4304 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4305 } else if (EVT == MVT::i32) {
4306 // InsertPS works with constant index.
4307 if (isa<ConstantSDNode>(N2))
4314 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4315 MVT VT = Op.getValueType();
4316 MVT EVT = VT.getVectorElementType();
4318 if (Subtarget->hasSSE41())
4319 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4324 DebugLoc dl = Op.getDebugLoc();
4325 SDValue N0 = Op.getOperand(0);
4326 SDValue N1 = Op.getOperand(1);
4327 SDValue N2 = Op.getOperand(2);
4329 if (EVT.getSizeInBits() == 16) {
4330 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4331 // as its second argument.
4332 if (N1.getValueType() != MVT::i32)
4333 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4334 if (N2.getValueType() != MVT::i32)
4335 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4336 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4342 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4343 DebugLoc dl = Op.getDebugLoc();
4344 if (Op.getValueType() == MVT::v2f32)
4345 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4346 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4347 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4348 Op.getOperand(0))));
4350 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4351 MVT VT = MVT::v2i32;
4352 switch (Op.getValueType().getSimpleVT()) {
4359 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4360 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4363 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4364 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4365 // one of the above mentioned nodes. It has to be wrapped because otherwise
4366 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4367 // be used to form addressing mode. These wrapped nodes will be selected
4370 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4371 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4372 // FIXME there isn't really any debug info here, should come from the parent
4373 DebugLoc dl = CP->getDebugLoc();
4374 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4375 CP->getAlignment());
4376 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4377 // With PIC, the address is actually $g + Offset.
4378 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4379 !Subtarget->isPICStyleRIPRel()) {
4380 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4381 DAG.getNode(X86ISD::GlobalBaseReg,
4382 DebugLoc::getUnknownLoc(),
4391 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4393 SelectionDAG &DAG) const {
4394 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4395 bool ExtraLoadRequired =
4396 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4398 // Create the TargetGlobalAddress node, folding in the constant
4399 // offset if it is legal.
4401 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4402 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4405 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4406 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4408 // With PIC, the address is actually $g + Offset.
4409 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4410 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4411 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4415 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4416 // load the value at address GV, not the value of GV itself. This means that
4417 // the GlobalAddress must be in the base or index register of the address, not
4418 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4419 // The same applies for external symbols during PIC codegen
4420 if (ExtraLoadRequired)
4421 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4422 PseudoSourceValue::getGOT(), 0);
4424 // If there was a non-zero offset that we didn't fold, create an explicit
4427 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4428 DAG.getConstant(Offset, getPointerTy()));
4434 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4435 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4436 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4437 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4441 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4442 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg) {
4443 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4444 DebugLoc dl = GA->getDebugLoc();
4445 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4446 GA->getValueType(0),
4449 SDValue Ops[] = { Chain, TGA, *InFlag };
4450 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4452 SDValue Ops[] = { Chain, TGA };
4453 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4455 SDValue Flag = Chain.getValue(1);
4456 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4459 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4461 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4464 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4465 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4466 DAG.getNode(X86ISD::GlobalBaseReg,
4467 DebugLoc::getUnknownLoc(),
4469 InFlag = Chain.getValue(1);
4471 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX);
4474 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4476 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4478 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX);
4481 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4482 // "local exec" model.
4483 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4484 const MVT PtrVT, TLSModel::Model model,
4486 DebugLoc dl = GA->getDebugLoc();
4487 // Get the Thread Pointer
4488 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4489 DebugLoc::getUnknownLoc(), PtrVT,
4490 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4493 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4496 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4498 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4499 GA->getValueType(0),
4501 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4503 if (model == TLSModel::InitialExec)
4504 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4505 PseudoSourceValue::getGOT(), 0);
4507 // The address of the thread local variable is the add of the thread
4508 // pointer with the offset of the variable.
4509 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4513 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4514 // TODO: implement the "local dynamic" model
4515 // TODO: implement the "initial exec"model for pic executables
4516 assert(Subtarget->isTargetELF() &&
4517 "TLS not implemented for non-ELF targets");
4518 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4519 GlobalValue *GV = GA->getGlobal();
4520 TLSModel::Model model =
4521 getTLSModel (GV, getTargetMachine().getRelocationModel());
4522 if (Subtarget->is64Bit()) {
4524 case TLSModel::GeneralDynamic:
4525 case TLSModel::LocalDynamic: // not implemented
4526 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4528 case TLSModel::InitialExec:
4529 case TLSModel::LocalExec:
4530 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, true);
4534 case TLSModel::GeneralDynamic:
4535 case TLSModel::LocalDynamic: // not implemented
4536 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4538 case TLSModel::InitialExec:
4539 case TLSModel::LocalExec:
4540 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, false);
4543 assert(0 && "Unreachable");
4548 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4549 // FIXME there isn't really any debug info here
4550 DebugLoc dl = Op.getDebugLoc();
4551 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4552 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4553 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4554 // With PIC, the address is actually $g + Offset.
4555 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4556 !Subtarget->isPICStyleRIPRel()) {
4557 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4558 DAG.getNode(X86ISD::GlobalBaseReg,
4559 DebugLoc::getUnknownLoc(),
4567 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4568 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4569 // FIXME there isn't really any debug into here
4570 DebugLoc dl = JT->getDebugLoc();
4571 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4572 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4573 // With PIC, the address is actually $g + Offset.
4574 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4575 !Subtarget->isPICStyleRIPRel()) {
4576 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4577 DAG.getNode(X86ISD::GlobalBaseReg,
4578 DebugLoc::getUnknownLoc(),
4586 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4587 /// take a 2 x i32 value to shift plus a shift amount.
4588 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4589 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4590 MVT VT = Op.getValueType();
4591 unsigned VTBits = VT.getSizeInBits();
4592 DebugLoc dl = Op.getDebugLoc();
4593 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4594 SDValue ShOpLo = Op.getOperand(0);
4595 SDValue ShOpHi = Op.getOperand(1);
4596 SDValue ShAmt = Op.getOperand(2);
4597 SDValue Tmp1 = isSRA ?
4598 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4599 DAG.getConstant(VTBits - 1, MVT::i8)) :
4600 DAG.getConstant(0, VT);
4603 if (Op.getOpcode() == ISD::SHL_PARTS) {
4604 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4605 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4607 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4608 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4611 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4612 DAG.getConstant(VTBits, MVT::i8));
4613 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4614 AndNode, DAG.getConstant(0, MVT::i8));
4617 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4618 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4619 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4621 if (Op.getOpcode() == ISD::SHL_PARTS) {
4622 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4623 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4625 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4626 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4629 SDValue Ops[2] = { Lo, Hi };
4630 return DAG.getMergeValues(Ops, 2, dl);
4633 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4634 MVT SrcVT = Op.getOperand(0).getValueType();
4635 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4636 "Unknown SINT_TO_FP to lower!");
4638 // These are really Legal; caller falls through into that case.
4639 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4641 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
4642 Subtarget->is64Bit())
4645 DebugLoc dl = Op.getDebugLoc();
4646 unsigned Size = SrcVT.getSizeInBits()/8;
4647 MachineFunction &MF = DAG.getMachineFunction();
4648 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4649 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4650 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4652 PseudoSourceValue::getFixedStack(SSFI), 0);
4656 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4658 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4660 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4661 SmallVector<SDValue, 8> Ops;
4662 Ops.push_back(Chain);
4663 Ops.push_back(StackSlot);
4664 Ops.push_back(DAG.getValueType(SrcVT));
4665 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4666 Tys, &Ops[0], Ops.size());
4669 Chain = Result.getValue(1);
4670 SDValue InFlag = Result.getValue(2);
4672 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4673 // shouldn't be necessary except that RFP cannot be live across
4674 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4675 MachineFunction &MF = DAG.getMachineFunction();
4676 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4677 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4678 Tys = DAG.getVTList(MVT::Other);
4679 SmallVector<SDValue, 8> Ops;
4680 Ops.push_back(Chain);
4681 Ops.push_back(Result);
4682 Ops.push_back(StackSlot);
4683 Ops.push_back(DAG.getValueType(Op.getValueType()));
4684 Ops.push_back(InFlag);
4685 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4686 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4687 PseudoSourceValue::getFixedStack(SSFI), 0);
4693 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4694 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4695 // This algorithm is not obvious. Here it is in C code, more or less:
4697 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4698 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4699 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4701 // Copy ints to xmm registers.
4702 __m128i xh = _mm_cvtsi32_si128( hi );
4703 __m128i xl = _mm_cvtsi32_si128( lo );
4705 // Combine into low half of a single xmm register.
4706 __m128i x = _mm_unpacklo_epi32( xh, xl );
4710 // Merge in appropriate exponents to give the integer bits the right
4712 x = _mm_unpacklo_epi32( x, exp );
4714 // Subtract away the biases to deal with the IEEE-754 double precision
4716 d = _mm_sub_pd( (__m128d) x, bias );
4718 // All conversions up to here are exact. The correctly rounded result is
4719 // calculated using the current rounding mode using the following
4721 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4722 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4723 // store doesn't really need to be here (except
4724 // maybe to zero the other double)
4729 DebugLoc dl = Op.getDebugLoc();
4731 // Build some magic constants.
4732 std::vector<Constant*> CV0;
4733 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4734 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4735 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4736 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4737 Constant *C0 = ConstantVector::get(CV0);
4738 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4740 std::vector<Constant*> CV1;
4741 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4742 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4743 Constant *C1 = ConstantVector::get(CV1);
4744 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4746 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4747 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4749 DAG.getIntPtrConstant(1)));
4750 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4751 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4753 DAG.getIntPtrConstant(0)));
4754 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4755 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4756 PseudoSourceValue::getConstantPool(), 0,
4758 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4759 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4760 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4761 PseudoSourceValue::getConstantPool(), 0,
4763 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4765 // Add the halves; easiest way is to swap them into another reg first.
4766 int ShufMask[2] = { 1, -1 };
4767 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4768 DAG.getUNDEF(MVT::v2f64), ShufMask);
4769 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4770 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4771 DAG.getIntPtrConstant(0));
4774 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4775 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4776 DebugLoc dl = Op.getDebugLoc();
4777 // FP constant to bias correct the final result.
4778 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4781 // Load the 32-bit value into an XMM register.
4782 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4783 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4785 DAG.getIntPtrConstant(0)));
4787 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4788 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4789 DAG.getIntPtrConstant(0));
4791 // Or the load with the bias.
4792 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4793 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4794 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4796 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4797 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4798 MVT::v2f64, Bias)));
4799 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4800 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4801 DAG.getIntPtrConstant(0));
4803 // Subtract the bias.
4804 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4806 // Handle final rounding.
4807 MVT DestVT = Op.getValueType();
4809 if (DestVT.bitsLT(MVT::f64)) {
4810 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4811 DAG.getIntPtrConstant(0));
4812 } else if (DestVT.bitsGT(MVT::f64)) {
4813 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4816 // Handle final rounding.
4820 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4821 SDValue N0 = Op.getOperand(0);
4822 DebugLoc dl = Op.getDebugLoc();
4824 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4825 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4826 // the optimization here.
4827 if (DAG.SignBitIsZero(N0))
4828 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4830 MVT SrcVT = N0.getValueType();
4831 if (SrcVT == MVT::i64) {
4832 // We only handle SSE2 f64 target here; caller can handle the rest.
4833 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4836 return LowerUINT_TO_FP_i64(Op, DAG);
4837 } else if (SrcVT == MVT::i32) {
4838 return LowerUINT_TO_FP_i32(Op, DAG);
4841 assert(0 && "Unknown UINT_TO_FP to lower!");
4845 std::pair<SDValue,SDValue> X86TargetLowering::
4846 FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
4847 DebugLoc dl = Op.getDebugLoc();
4848 assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
4849 Op.getValueType().getSimpleVT() >= MVT::i16 &&
4850 "Unknown FP_TO_SINT to lower!");
4852 // These are really Legal.
4853 if (Op.getValueType() == MVT::i32 &&
4854 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4855 return std::make_pair(SDValue(), SDValue());
4856 if (Subtarget->is64Bit() &&
4857 Op.getValueType() == MVT::i64 &&
4858 Op.getOperand(0).getValueType() != MVT::f80)
4859 return std::make_pair(SDValue(), SDValue());
4861 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4863 MachineFunction &MF = DAG.getMachineFunction();
4864 unsigned MemSize = Op.getValueType().getSizeInBits()/8;
4865 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4866 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4868 switch (Op.getValueType().getSimpleVT()) {
4869 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4870 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4871 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4872 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4875 SDValue Chain = DAG.getEntryNode();
4876 SDValue Value = Op.getOperand(0);
4877 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4878 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4879 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
4880 PseudoSourceValue::getFixedStack(SSFI), 0);
4881 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4883 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4885 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
4886 Chain = Value.getValue(1);
4887 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4888 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4891 // Build the FP_TO_INT*_IN_MEM
4892 SDValue Ops[] = { Chain, Value, StackSlot };
4893 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
4895 return std::make_pair(FIST, StackSlot);
4898 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4899 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
4900 SDValue FIST = Vals.first, StackSlot = Vals.second;
4901 if (FIST.getNode() == 0) return SDValue();
4904 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4905 FIST, StackSlot, NULL, 0);
4908 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4909 DebugLoc dl = Op.getDebugLoc();
4910 MVT VT = Op.getValueType();
4913 EltVT = VT.getVectorElementType();
4914 std::vector<Constant*> CV;
4915 if (EltVT == MVT::f64) {
4916 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4920 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4926 Constant *C = ConstantVector::get(CV);
4927 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4928 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4929 PseudoSourceValue::getConstantPool(), 0,
4931 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
4934 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4935 DebugLoc dl = Op.getDebugLoc();
4936 MVT VT = Op.getValueType();
4938 unsigned EltNum = 1;
4939 if (VT.isVector()) {
4940 EltVT = VT.getVectorElementType();
4941 EltNum = VT.getVectorNumElements();
4943 std::vector<Constant*> CV;
4944 if (EltVT == MVT::f64) {
4945 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4949 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4955 Constant *C = ConstantVector::get(CV);
4956 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4957 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4958 PseudoSourceValue::getConstantPool(), 0,
4960 if (VT.isVector()) {
4961 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4962 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
4963 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4965 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
4967 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
4971 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4972 SDValue Op0 = Op.getOperand(0);
4973 SDValue Op1 = Op.getOperand(1);
4974 DebugLoc dl = Op.getDebugLoc();
4975 MVT VT = Op.getValueType();
4976 MVT SrcVT = Op1.getValueType();
4978 // If second operand is smaller, extend it first.
4979 if (SrcVT.bitsLT(VT)) {
4980 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
4983 // And if it is bigger, shrink it first.
4984 if (SrcVT.bitsGT(VT)) {
4985 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
4989 // At this point the operands and the result should have the same
4990 // type, and that won't be f80 since that is not custom lowered.
4992 // First get the sign bit of second operand.
4993 std::vector<Constant*> CV;
4994 if (SrcVT == MVT::f64) {
4995 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4996 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4998 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4999 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5000 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5001 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5003 Constant *C = ConstantVector::get(CV);
5004 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5005 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5006 PseudoSourceValue::getConstantPool(), 0,
5008 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5010 // Shift sign bit right or left if the two operands have different types.
5011 if (SrcVT.bitsGT(VT)) {
5012 // Op0 is MVT::f32, Op1 is MVT::f64.
5013 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5014 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5015 DAG.getConstant(32, MVT::i32));
5016 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5017 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5018 DAG.getIntPtrConstant(0));
5021 // Clear first operand sign bit.
5023 if (VT == MVT::f64) {
5024 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5025 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5027 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5028 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5029 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5030 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5032 C = ConstantVector::get(CV);
5033 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5034 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5035 PseudoSourceValue::getConstantPool(), 0,
5037 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5039 // Or the value with the sign bit.
5040 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5043 /// Emit nodes that will be selected as "test Op0,Op0", or something
5045 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5046 SelectionDAG &DAG) {
5047 DebugLoc dl = Op.getDebugLoc();
5049 // CF and OF aren't always set the way we want. Determine which
5050 // of these we need.
5051 bool NeedCF = false;
5052 bool NeedOF = false;
5054 case X86::COND_A: case X86::COND_AE:
5055 case X86::COND_B: case X86::COND_BE:
5058 case X86::COND_G: case X86::COND_GE:
5059 case X86::COND_L: case X86::COND_LE:
5060 case X86::COND_O: case X86::COND_NO:
5066 // See if we can use the EFLAGS value from the operand instead of
5067 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5068 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5069 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5070 unsigned Opcode = 0;
5071 unsigned NumOperands = 0;
5072 switch (Op.getNode()->getOpcode()) {
5074 // Due to an isel shortcoming, be conservative if this add is likely to
5075 // be selected as part of a load-modify-store instruction. When the root
5076 // node in a match is a store, isel doesn't know how to remap non-chain
5077 // non-flag uses of other nodes in the match, such as the ADD in this
5078 // case. This leads to the ADD being left around and reselected, with
5079 // the result being two adds in the output.
5080 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5081 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5082 if (UI->getOpcode() == ISD::STORE)
5084 if (ConstantSDNode *C =
5085 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5086 // An add of one will be selected as an INC.
5087 if (C->getAPIntValue() == 1) {
5088 Opcode = X86ISD::INC;
5092 // An add of negative one (subtract of one) will be selected as a DEC.
5093 if (C->getAPIntValue().isAllOnesValue()) {
5094 Opcode = X86ISD::DEC;
5099 // Otherwise use a regular EFLAGS-setting add.
5100 Opcode = X86ISD::ADD;
5104 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5105 // likely to be selected as part of a load-modify-store instruction.
5106 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5107 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5108 if (UI->getOpcode() == ISD::STORE)
5110 // Otherwise use a regular EFLAGS-setting sub.
5111 Opcode = X86ISD::SUB;
5118 return SDValue(Op.getNode(), 1);
5124 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5125 SmallVector<SDValue, 4> Ops;
5126 for (unsigned i = 0; i != NumOperands; ++i)
5127 Ops.push_back(Op.getOperand(i));
5128 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5129 DAG.ReplaceAllUsesWith(Op, New);
5130 return SDValue(New.getNode(), 1);
5134 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5135 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5136 DAG.getConstant(0, Op.getValueType()));
5139 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5141 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5142 SelectionDAG &DAG) {
5143 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5144 if (C->getAPIntValue() == 0)
5145 return EmitTest(Op0, X86CC, DAG);
5147 DebugLoc dl = Op0.getDebugLoc();
5148 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5151 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5152 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5153 SDValue Op0 = Op.getOperand(0);
5154 SDValue Op1 = Op.getOperand(1);
5155 DebugLoc dl = Op.getDebugLoc();
5156 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5158 // Lower (X & (1 << N)) == 0 to BT(X, N).
5159 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5160 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5161 if (Op0.getOpcode() == ISD::AND &&
5163 Op1.getOpcode() == ISD::Constant &&
5164 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5165 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5167 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5168 if (ConstantSDNode *Op010C =
5169 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5170 if (Op010C->getZExtValue() == 1) {
5171 LHS = Op0.getOperand(0);
5172 RHS = Op0.getOperand(1).getOperand(1);
5174 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5175 if (ConstantSDNode *Op000C =
5176 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5177 if (Op000C->getZExtValue() == 1) {
5178 LHS = Op0.getOperand(1);
5179 RHS = Op0.getOperand(0).getOperand(1);
5181 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5182 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5183 SDValue AndLHS = Op0.getOperand(0);
5184 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5185 LHS = AndLHS.getOperand(0);
5186 RHS = AndLHS.getOperand(1);
5190 if (LHS.getNode()) {
5191 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5192 // instruction. Since the shift amount is in-range-or-undefined, we know
5193 // that doing a bittest on the i16 value is ok. We extend to i32 because
5194 // the encoding for the i16 version is larger than the i32 version.
5195 if (LHS.getValueType() == MVT::i8)
5196 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5198 // If the operand types disagree, extend the shift amount to match. Since
5199 // BT ignores high bits (like shifts) we can use anyextend.
5200 if (LHS.getValueType() != RHS.getValueType())
5201 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5203 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5204 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5205 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5206 DAG.getConstant(Cond, MVT::i8), BT);
5210 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5211 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5213 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5214 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5215 DAG.getConstant(X86CC, MVT::i8), Cond);
5218 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5220 SDValue Op0 = Op.getOperand(0);
5221 SDValue Op1 = Op.getOperand(1);
5222 SDValue CC = Op.getOperand(2);
5223 MVT VT = Op.getValueType();
5224 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5225 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5226 DebugLoc dl = Op.getDebugLoc();
5230 MVT VT0 = Op0.getValueType();
5231 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5232 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5235 switch (SetCCOpcode) {
5238 case ISD::SETEQ: SSECC = 0; break;
5240 case ISD::SETGT: Swap = true; // Fallthrough
5242 case ISD::SETOLT: SSECC = 1; break;
5244 case ISD::SETGE: Swap = true; // Fallthrough
5246 case ISD::SETOLE: SSECC = 2; break;
5247 case ISD::SETUO: SSECC = 3; break;
5249 case ISD::SETNE: SSECC = 4; break;
5250 case ISD::SETULE: Swap = true;
5251 case ISD::SETUGE: SSECC = 5; break;
5252 case ISD::SETULT: Swap = true;
5253 case ISD::SETUGT: SSECC = 6; break;
5254 case ISD::SETO: SSECC = 7; break;
5257 std::swap(Op0, Op1);
5259 // In the two special cases we can't handle, emit two comparisons.
5261 if (SetCCOpcode == ISD::SETUEQ) {
5263 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5264 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5265 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5267 else if (SetCCOpcode == ISD::SETONE) {
5269 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5270 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5271 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5273 assert(0 && "Illegal FP comparison");
5275 // Handle all other FP comparisons here.
5276 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5279 // We are handling one of the integer comparisons here. Since SSE only has
5280 // GT and EQ comparisons for integer, swapping operands and multiple
5281 // operations may be required for some comparisons.
5282 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5283 bool Swap = false, Invert = false, FlipSigns = false;
5285 switch (VT.getSimpleVT()) {
5287 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5288 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5289 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5290 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5293 switch (SetCCOpcode) {
5295 case ISD::SETNE: Invert = true;
5296 case ISD::SETEQ: Opc = EQOpc; break;
5297 case ISD::SETLT: Swap = true;
5298 case ISD::SETGT: Opc = GTOpc; break;
5299 case ISD::SETGE: Swap = true;
5300 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5301 case ISD::SETULT: Swap = true;
5302 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5303 case ISD::SETUGE: Swap = true;
5304 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5307 std::swap(Op0, Op1);
5309 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5310 // bits of the inputs before performing those operations.
5312 MVT EltVT = VT.getVectorElementType();
5313 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5315 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5316 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5318 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5319 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5322 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5324 // If the logical-not of the result is required, perform that now.
5326 Result = DAG.getNOT(dl, Result, VT);
5331 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5332 static bool isX86LogicalCmp(SDValue Op) {
5333 unsigned Opc = Op.getNode()->getOpcode();
5334 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5336 if (Op.getResNo() == 1 &&
5337 (Opc == X86ISD::ADD ||
5338 Opc == X86ISD::SUB ||
5339 Opc == X86ISD::SMUL ||
5340 Opc == X86ISD::UMUL ||
5341 Opc == X86ISD::INC ||
5342 Opc == X86ISD::DEC))
5348 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5349 bool addTest = true;
5350 SDValue Cond = Op.getOperand(0);
5351 DebugLoc dl = Op.getDebugLoc();
5354 if (Cond.getOpcode() == ISD::SETCC)
5355 Cond = LowerSETCC(Cond, DAG);
5357 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5358 // setting operand in place of the X86ISD::SETCC.
5359 if (Cond.getOpcode() == X86ISD::SETCC) {
5360 CC = Cond.getOperand(0);
5362 SDValue Cmp = Cond.getOperand(1);
5363 unsigned Opc = Cmp.getOpcode();
5364 MVT VT = Op.getValueType();
5366 bool IllegalFPCMov = false;
5367 if (VT.isFloatingPoint() && !VT.isVector() &&
5368 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5369 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5371 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5372 Opc == X86ISD::BT) { // FIXME
5379 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5380 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5383 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5384 SmallVector<SDValue, 4> Ops;
5385 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5386 // condition is true.
5387 Ops.push_back(Op.getOperand(2));
5388 Ops.push_back(Op.getOperand(1));
5390 Ops.push_back(Cond);
5391 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5394 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5395 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5396 // from the AND / OR.
5397 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5398 Opc = Op.getOpcode();
5399 if (Opc != ISD::OR && Opc != ISD::AND)
5401 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5402 Op.getOperand(0).hasOneUse() &&
5403 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5404 Op.getOperand(1).hasOneUse());
5407 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5408 // 1 and that the SETCC node has a single use.
5409 static bool isXor1OfSetCC(SDValue Op) {
5410 if (Op.getOpcode() != ISD::XOR)
5412 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5413 if (N1C && N1C->getAPIntValue() == 1) {
5414 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5415 Op.getOperand(0).hasOneUse();
5420 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5421 bool addTest = true;
5422 SDValue Chain = Op.getOperand(0);
5423 SDValue Cond = Op.getOperand(1);
5424 SDValue Dest = Op.getOperand(2);
5425 DebugLoc dl = Op.getDebugLoc();
5428 if (Cond.getOpcode() == ISD::SETCC)
5429 Cond = LowerSETCC(Cond, DAG);
5431 // FIXME: LowerXALUO doesn't handle these!!
5432 else if (Cond.getOpcode() == X86ISD::ADD ||
5433 Cond.getOpcode() == X86ISD::SUB ||
5434 Cond.getOpcode() == X86ISD::SMUL ||
5435 Cond.getOpcode() == X86ISD::UMUL)
5436 Cond = LowerXALUO(Cond, DAG);
5439 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5440 // setting operand in place of the X86ISD::SETCC.
5441 if (Cond.getOpcode() == X86ISD::SETCC) {
5442 CC = Cond.getOperand(0);
5444 SDValue Cmp = Cond.getOperand(1);
5445 unsigned Opc = Cmp.getOpcode();
5446 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5447 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5451 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5455 // These can only come from an arithmetic instruction with overflow,
5456 // e.g. SADDO, UADDO.
5457 Cond = Cond.getNode()->getOperand(1);
5464 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5465 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5466 if (CondOpc == ISD::OR) {
5467 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5468 // two branches instead of an explicit OR instruction with a
5470 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5471 isX86LogicalCmp(Cmp)) {
5472 CC = Cond.getOperand(0).getOperand(0);
5473 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5474 Chain, Dest, CC, Cmp);
5475 CC = Cond.getOperand(1).getOperand(0);
5479 } else { // ISD::AND
5480 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5481 // two branches instead of an explicit AND instruction with a
5482 // separate test. However, we only do this if this block doesn't
5483 // have a fall-through edge, because this requires an explicit
5484 // jmp when the condition is false.
5485 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5486 isX86LogicalCmp(Cmp) &&
5487 Op.getNode()->hasOneUse()) {
5488 X86::CondCode CCode =
5489 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5490 CCode = X86::GetOppositeBranchCondition(CCode);
5491 CC = DAG.getConstant(CCode, MVT::i8);
5492 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5493 // Look for an unconditional branch following this conditional branch.
5494 // We need this because we need to reverse the successors in order
5495 // to implement FCMP_OEQ.
5496 if (User.getOpcode() == ISD::BR) {
5497 SDValue FalseBB = User.getOperand(1);
5499 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5500 assert(NewBR == User);
5503 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5504 Chain, Dest, CC, Cmp);
5505 X86::CondCode CCode =
5506 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5507 CCode = X86::GetOppositeBranchCondition(CCode);
5508 CC = DAG.getConstant(CCode, MVT::i8);
5514 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5515 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5516 // It should be transformed during dag combiner except when the condition
5517 // is set by a arithmetics with overflow node.
5518 X86::CondCode CCode =
5519 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5520 CCode = X86::GetOppositeBranchCondition(CCode);
5521 CC = DAG.getConstant(CCode, MVT::i8);
5522 Cond = Cond.getOperand(0).getOperand(1);
5528 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5529 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5531 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5532 Chain, Dest, CC, Cond);
5536 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5537 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5538 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5539 // that the guard pages used by the OS virtual memory manager are allocated in
5540 // correct sequence.
5542 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5543 SelectionDAG &DAG) {
5544 assert(Subtarget->isTargetCygMing() &&
5545 "This should be used only on Cygwin/Mingw targets");
5546 DebugLoc dl = Op.getDebugLoc();
5549 SDValue Chain = Op.getOperand(0);
5550 SDValue Size = Op.getOperand(1);
5551 // FIXME: Ensure alignment here
5555 MVT IntPtr = getPointerTy();
5556 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5558 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5560 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5561 Flag = Chain.getValue(1);
5563 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5564 SDValue Ops[] = { Chain,
5565 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5566 DAG.getRegister(X86::EAX, IntPtr),
5567 DAG.getRegister(X86StackPtr, SPTy),
5569 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5570 Flag = Chain.getValue(1);
5572 Chain = DAG.getCALLSEQ_END(Chain,
5573 DAG.getIntPtrConstant(0, true),
5574 DAG.getIntPtrConstant(0, true),
5577 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5579 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5580 return DAG.getMergeValues(Ops1, 2, dl);
5584 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5586 SDValue Dst, SDValue Src,
5587 SDValue Size, unsigned Align,
5589 uint64_t DstSVOff) {
5590 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5592 // If not DWORD aligned or size is more than the threshold, call the library.
5593 // The libc version is likely to be faster for these cases. It can use the
5594 // address value and run time information about the CPU.
5595 if ((Align & 3) != 0 ||
5597 ConstantSize->getZExtValue() >
5598 getSubtarget()->getMaxInlineSizeThreshold()) {
5599 SDValue InFlag(0, 0);
5601 // Check to see if there is a specialized entry-point for memory zeroing.
5602 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5604 if (const char *bzeroEntry = V &&
5605 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5606 MVT IntPtr = getPointerTy();
5607 const Type *IntPtrTy = TD->getIntPtrType();
5608 TargetLowering::ArgListTy Args;
5609 TargetLowering::ArgListEntry Entry;
5611 Entry.Ty = IntPtrTy;
5612 Args.push_back(Entry);
5614 Args.push_back(Entry);
5615 std::pair<SDValue,SDValue> CallResult =
5616 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5617 CallingConv::C, false,
5618 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5619 return CallResult.second;
5622 // Otherwise have the target-independent code call memset.
5626 uint64_t SizeVal = ConstantSize->getZExtValue();
5627 SDValue InFlag(0, 0);
5630 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5631 unsigned BytesLeft = 0;
5632 bool TwoRepStos = false;
5635 uint64_t Val = ValC->getZExtValue() & 255;
5637 // If the value is a constant, then we can potentially use larger sets.
5638 switch (Align & 3) {
5639 case 2: // WORD aligned
5642 Val = (Val << 8) | Val;
5644 case 0: // DWORD aligned
5647 Val = (Val << 8) | Val;
5648 Val = (Val << 16) | Val;
5649 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5652 Val = (Val << 32) | Val;
5655 default: // Byte aligned
5658 Count = DAG.getIntPtrConstant(SizeVal);
5662 if (AVT.bitsGT(MVT::i8)) {
5663 unsigned UBytes = AVT.getSizeInBits() / 8;
5664 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5665 BytesLeft = SizeVal % UBytes;
5668 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5670 InFlag = Chain.getValue(1);
5673 Count = DAG.getIntPtrConstant(SizeVal);
5674 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5675 InFlag = Chain.getValue(1);
5678 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5681 InFlag = Chain.getValue(1);
5682 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5685 InFlag = Chain.getValue(1);
5687 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5688 SmallVector<SDValue, 8> Ops;
5689 Ops.push_back(Chain);
5690 Ops.push_back(DAG.getValueType(AVT));
5691 Ops.push_back(InFlag);
5692 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5695 InFlag = Chain.getValue(1);
5697 MVT CVT = Count.getValueType();
5698 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5699 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5700 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5703 InFlag = Chain.getValue(1);
5704 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5706 Ops.push_back(Chain);
5707 Ops.push_back(DAG.getValueType(MVT::i8));
5708 Ops.push_back(InFlag);
5709 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5710 } else if (BytesLeft) {
5711 // Handle the last 1 - 7 bytes.
5712 unsigned Offset = SizeVal - BytesLeft;
5713 MVT AddrVT = Dst.getValueType();
5714 MVT SizeVT = Size.getValueType();
5716 Chain = DAG.getMemset(Chain, dl,
5717 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5718 DAG.getConstant(Offset, AddrVT)),
5720 DAG.getConstant(BytesLeft, SizeVT),
5721 Align, DstSV, DstSVOff + Offset);
5724 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5729 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5730 SDValue Chain, SDValue Dst, SDValue Src,
5731 SDValue Size, unsigned Align,
5733 const Value *DstSV, uint64_t DstSVOff,
5734 const Value *SrcSV, uint64_t SrcSVOff) {
5735 // This requires the copy size to be a constant, preferrably
5736 // within a subtarget-specific limit.
5737 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5740 uint64_t SizeVal = ConstantSize->getZExtValue();
5741 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5744 /// If not DWORD aligned, call the library.
5745 if ((Align & 3) != 0)
5750 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5753 unsigned UBytes = AVT.getSizeInBits() / 8;
5754 unsigned CountVal = SizeVal / UBytes;
5755 SDValue Count = DAG.getIntPtrConstant(CountVal);
5756 unsigned BytesLeft = SizeVal % UBytes;
5758 SDValue InFlag(0, 0);
5759 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5762 InFlag = Chain.getValue(1);
5763 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5766 InFlag = Chain.getValue(1);
5767 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5770 InFlag = Chain.getValue(1);
5772 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5773 SmallVector<SDValue, 8> Ops;
5774 Ops.push_back(Chain);
5775 Ops.push_back(DAG.getValueType(AVT));
5776 Ops.push_back(InFlag);
5777 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5779 SmallVector<SDValue, 4> Results;
5780 Results.push_back(RepMovs);
5782 // Handle the last 1 - 7 bytes.
5783 unsigned Offset = SizeVal - BytesLeft;
5784 MVT DstVT = Dst.getValueType();
5785 MVT SrcVT = Src.getValueType();
5786 MVT SizeVT = Size.getValueType();
5787 Results.push_back(DAG.getMemcpy(Chain, dl,
5788 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5789 DAG.getConstant(Offset, DstVT)),
5790 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5791 DAG.getConstant(Offset, SrcVT)),
5792 DAG.getConstant(BytesLeft, SizeVT),
5793 Align, AlwaysInline,
5794 DstSV, DstSVOff + Offset,
5795 SrcSV, SrcSVOff + Offset));
5798 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5799 &Results[0], Results.size());
5802 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5803 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5804 DebugLoc dl = Op.getDebugLoc();
5806 if (!Subtarget->is64Bit()) {
5807 // vastart just stores the address of the VarArgsFrameIndex slot into the
5808 // memory location argument.
5809 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5810 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
5814 // gp_offset (0 - 6 * 8)
5815 // fp_offset (48 - 48 + 8 * 16)
5816 // overflow_arg_area (point to parameters coming in memory).
5818 SmallVector<SDValue, 8> MemOps;
5819 SDValue FIN = Op.getOperand(1);
5821 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
5822 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5824 MemOps.push_back(Store);
5827 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5828 FIN, DAG.getIntPtrConstant(4));
5829 Store = DAG.getStore(Op.getOperand(0), dl,
5830 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5832 MemOps.push_back(Store);
5834 // Store ptr to overflow_arg_area
5835 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5836 FIN, DAG.getIntPtrConstant(4));
5837 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5838 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
5839 MemOps.push_back(Store);
5841 // Store ptr to reg_save_area.
5842 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5843 FIN, DAG.getIntPtrConstant(8));
5844 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5845 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
5846 MemOps.push_back(Store);
5847 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5848 &MemOps[0], MemOps.size());
5851 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5852 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5853 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5854 SDValue Chain = Op.getOperand(0);
5855 SDValue SrcPtr = Op.getOperand(1);
5856 SDValue SrcSV = Op.getOperand(2);
5858 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5863 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5864 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5865 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5866 SDValue Chain = Op.getOperand(0);
5867 SDValue DstPtr = Op.getOperand(1);
5868 SDValue SrcPtr = Op.getOperand(2);
5869 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5870 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5871 DebugLoc dl = Op.getDebugLoc();
5873 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
5874 DAG.getIntPtrConstant(24), 8, false,
5875 DstSV, 0, SrcSV, 0);
5879 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5880 DebugLoc dl = Op.getDebugLoc();
5881 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5883 default: return SDValue(); // Don't custom lower most intrinsics.
5884 // Comparison intrinsics.
5885 case Intrinsic::x86_sse_comieq_ss:
5886 case Intrinsic::x86_sse_comilt_ss:
5887 case Intrinsic::x86_sse_comile_ss:
5888 case Intrinsic::x86_sse_comigt_ss:
5889 case Intrinsic::x86_sse_comige_ss:
5890 case Intrinsic::x86_sse_comineq_ss:
5891 case Intrinsic::x86_sse_ucomieq_ss:
5892 case Intrinsic::x86_sse_ucomilt_ss:
5893 case Intrinsic::x86_sse_ucomile_ss:
5894 case Intrinsic::x86_sse_ucomigt_ss:
5895 case Intrinsic::x86_sse_ucomige_ss:
5896 case Intrinsic::x86_sse_ucomineq_ss:
5897 case Intrinsic::x86_sse2_comieq_sd:
5898 case Intrinsic::x86_sse2_comilt_sd:
5899 case Intrinsic::x86_sse2_comile_sd:
5900 case Intrinsic::x86_sse2_comigt_sd:
5901 case Intrinsic::x86_sse2_comige_sd:
5902 case Intrinsic::x86_sse2_comineq_sd:
5903 case Intrinsic::x86_sse2_ucomieq_sd:
5904 case Intrinsic::x86_sse2_ucomilt_sd:
5905 case Intrinsic::x86_sse2_ucomile_sd:
5906 case Intrinsic::x86_sse2_ucomigt_sd:
5907 case Intrinsic::x86_sse2_ucomige_sd:
5908 case Intrinsic::x86_sse2_ucomineq_sd: {
5910 ISD::CondCode CC = ISD::SETCC_INVALID;
5913 case Intrinsic::x86_sse_comieq_ss:
5914 case Intrinsic::x86_sse2_comieq_sd:
5918 case Intrinsic::x86_sse_comilt_ss:
5919 case Intrinsic::x86_sse2_comilt_sd:
5923 case Intrinsic::x86_sse_comile_ss:
5924 case Intrinsic::x86_sse2_comile_sd:
5928 case Intrinsic::x86_sse_comigt_ss:
5929 case Intrinsic::x86_sse2_comigt_sd:
5933 case Intrinsic::x86_sse_comige_ss:
5934 case Intrinsic::x86_sse2_comige_sd:
5938 case Intrinsic::x86_sse_comineq_ss:
5939 case Intrinsic::x86_sse2_comineq_sd:
5943 case Intrinsic::x86_sse_ucomieq_ss:
5944 case Intrinsic::x86_sse2_ucomieq_sd:
5945 Opc = X86ISD::UCOMI;
5948 case Intrinsic::x86_sse_ucomilt_ss:
5949 case Intrinsic::x86_sse2_ucomilt_sd:
5950 Opc = X86ISD::UCOMI;
5953 case Intrinsic::x86_sse_ucomile_ss:
5954 case Intrinsic::x86_sse2_ucomile_sd:
5955 Opc = X86ISD::UCOMI;
5958 case Intrinsic::x86_sse_ucomigt_ss:
5959 case Intrinsic::x86_sse2_ucomigt_sd:
5960 Opc = X86ISD::UCOMI;
5963 case Intrinsic::x86_sse_ucomige_ss:
5964 case Intrinsic::x86_sse2_ucomige_sd:
5965 Opc = X86ISD::UCOMI;
5968 case Intrinsic::x86_sse_ucomineq_ss:
5969 case Intrinsic::x86_sse2_ucomineq_sd:
5970 Opc = X86ISD::UCOMI;
5975 SDValue LHS = Op.getOperand(1);
5976 SDValue RHS = Op.getOperand(2);
5977 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
5978 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
5979 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5980 DAG.getConstant(X86CC, MVT::i8), Cond);
5981 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
5984 // Fix vector shift instructions where the last operand is a non-immediate
5986 case Intrinsic::x86_sse2_pslli_w:
5987 case Intrinsic::x86_sse2_pslli_d:
5988 case Intrinsic::x86_sse2_pslli_q:
5989 case Intrinsic::x86_sse2_psrli_w:
5990 case Intrinsic::x86_sse2_psrli_d:
5991 case Intrinsic::x86_sse2_psrli_q:
5992 case Intrinsic::x86_sse2_psrai_w:
5993 case Intrinsic::x86_sse2_psrai_d:
5994 case Intrinsic::x86_mmx_pslli_w:
5995 case Intrinsic::x86_mmx_pslli_d:
5996 case Intrinsic::x86_mmx_pslli_q:
5997 case Intrinsic::x86_mmx_psrli_w:
5998 case Intrinsic::x86_mmx_psrli_d:
5999 case Intrinsic::x86_mmx_psrli_q:
6000 case Intrinsic::x86_mmx_psrai_w:
6001 case Intrinsic::x86_mmx_psrai_d: {
6002 SDValue ShAmt = Op.getOperand(2);
6003 if (isa<ConstantSDNode>(ShAmt))
6006 unsigned NewIntNo = 0;
6007 MVT ShAmtVT = MVT::v4i32;
6009 case Intrinsic::x86_sse2_pslli_w:
6010 NewIntNo = Intrinsic::x86_sse2_psll_w;
6012 case Intrinsic::x86_sse2_pslli_d:
6013 NewIntNo = Intrinsic::x86_sse2_psll_d;
6015 case Intrinsic::x86_sse2_pslli_q:
6016 NewIntNo = Intrinsic::x86_sse2_psll_q;
6018 case Intrinsic::x86_sse2_psrli_w:
6019 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6021 case Intrinsic::x86_sse2_psrli_d:
6022 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6024 case Intrinsic::x86_sse2_psrli_q:
6025 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6027 case Intrinsic::x86_sse2_psrai_w:
6028 NewIntNo = Intrinsic::x86_sse2_psra_w;
6030 case Intrinsic::x86_sse2_psrai_d:
6031 NewIntNo = Intrinsic::x86_sse2_psra_d;
6034 ShAmtVT = MVT::v2i32;
6036 case Intrinsic::x86_mmx_pslli_w:
6037 NewIntNo = Intrinsic::x86_mmx_psll_w;
6039 case Intrinsic::x86_mmx_pslli_d:
6040 NewIntNo = Intrinsic::x86_mmx_psll_d;
6042 case Intrinsic::x86_mmx_pslli_q:
6043 NewIntNo = Intrinsic::x86_mmx_psll_q;
6045 case Intrinsic::x86_mmx_psrli_w:
6046 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6048 case Intrinsic::x86_mmx_psrli_d:
6049 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6051 case Intrinsic::x86_mmx_psrli_q:
6052 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6054 case Intrinsic::x86_mmx_psrai_w:
6055 NewIntNo = Intrinsic::x86_mmx_psra_w;
6057 case Intrinsic::x86_mmx_psrai_d:
6058 NewIntNo = Intrinsic::x86_mmx_psra_d;
6060 default: abort(); // Can't reach here.
6065 MVT VT = Op.getValueType();
6066 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6067 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6068 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6069 DAG.getConstant(NewIntNo, MVT::i32),
6070 Op.getOperand(1), ShAmt);
6075 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6076 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6077 DebugLoc dl = Op.getDebugLoc();
6080 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6082 DAG.getConstant(TD->getPointerSize(),
6083 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6084 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6085 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6090 // Just load the return address.
6091 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6092 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6093 RetAddrFI, NULL, 0);
6096 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6097 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6098 MFI->setFrameAddressIsTaken(true);
6099 MVT VT = Op.getValueType();
6100 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6101 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6102 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6103 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6105 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6109 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6110 SelectionDAG &DAG) {
6111 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6114 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6116 MachineFunction &MF = DAG.getMachineFunction();
6117 SDValue Chain = Op.getOperand(0);
6118 SDValue Offset = Op.getOperand(1);
6119 SDValue Handler = Op.getOperand(2);
6120 DebugLoc dl = Op.getDebugLoc();
6122 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6124 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6126 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6127 DAG.getIntPtrConstant(-TD->getPointerSize()));
6128 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6129 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6130 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6131 MF.getRegInfo().addLiveOut(StoreAddrReg);
6133 return DAG.getNode(X86ISD::EH_RETURN, dl,
6135 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6138 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6139 SelectionDAG &DAG) {
6140 SDValue Root = Op.getOperand(0);
6141 SDValue Trmp = Op.getOperand(1); // trampoline
6142 SDValue FPtr = Op.getOperand(2); // nested function
6143 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6144 DebugLoc dl = Op.getDebugLoc();
6146 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6148 const X86InstrInfo *TII =
6149 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6151 if (Subtarget->is64Bit()) {
6152 SDValue OutChains[6];
6154 // Large code-model.
6156 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6157 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6159 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6160 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6162 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6164 // Load the pointer to the nested function into R11.
6165 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6166 SDValue Addr = Trmp;
6167 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6170 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6171 DAG.getConstant(2, MVT::i64));
6172 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6174 // Load the 'nest' parameter value into R10.
6175 // R10 is specified in X86CallingConv.td
6176 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6177 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6178 DAG.getConstant(10, MVT::i64));
6179 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6180 Addr, TrmpAddr, 10);
6182 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6183 DAG.getConstant(12, MVT::i64));
6184 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6186 // Jump to the nested function.
6187 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6188 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6189 DAG.getConstant(20, MVT::i64));
6190 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6191 Addr, TrmpAddr, 20);
6193 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6194 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6195 DAG.getConstant(22, MVT::i64));
6196 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6200 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6201 return DAG.getMergeValues(Ops, 2, dl);
6203 const Function *Func =
6204 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6205 unsigned CC = Func->getCallingConv();
6210 assert(0 && "Unsupported calling convention");
6211 case CallingConv::C:
6212 case CallingConv::X86_StdCall: {
6213 // Pass 'nest' parameter in ECX.
6214 // Must be kept in sync with X86CallingConv.td
6217 // Check that ECX wasn't needed by an 'inreg' parameter.
6218 const FunctionType *FTy = Func->getFunctionType();
6219 const AttrListPtr &Attrs = Func->getAttributes();
6221 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6222 unsigned InRegCount = 0;
6225 for (FunctionType::param_iterator I = FTy->param_begin(),
6226 E = FTy->param_end(); I != E; ++I, ++Idx)
6227 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6228 // FIXME: should only count parameters that are lowered to integers.
6229 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6231 if (InRegCount > 2) {
6232 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6238 case CallingConv::X86_FastCall:
6239 case CallingConv::Fast:
6240 // Pass 'nest' parameter in EAX.
6241 // Must be kept in sync with X86CallingConv.td
6246 SDValue OutChains[4];
6249 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6250 DAG.getConstant(10, MVT::i32));
6251 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6253 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6254 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6255 OutChains[0] = DAG.getStore(Root, dl,
6256 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6259 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6260 DAG.getConstant(1, MVT::i32));
6261 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6263 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6265 DAG.getConstant(5, MVT::i32));
6266 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6267 TrmpAddr, 5, false, 1);
6269 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6270 DAG.getConstant(6, MVT::i32));
6271 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6274 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6275 return DAG.getMergeValues(Ops, 2, dl);
6279 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6281 The rounding mode is in bits 11:10 of FPSR, and has the following
6288 FLT_ROUNDS, on the other hand, expects the following:
6295 To perform the conversion, we do:
6296 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6299 MachineFunction &MF = DAG.getMachineFunction();
6300 const TargetMachine &TM = MF.getTarget();
6301 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6302 unsigned StackAlignment = TFI.getStackAlignment();
6303 MVT VT = Op.getValueType();
6304 DebugLoc dl = Op.getDebugLoc();
6306 // Save FP Control Word to stack slot
6307 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6308 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6310 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6311 DAG.getEntryNode(), StackSlot);
6313 // Load FP Control Word from stack slot
6314 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6316 // Transform as necessary
6318 DAG.getNode(ISD::SRL, dl, MVT::i16,
6319 DAG.getNode(ISD::AND, dl, MVT::i16,
6320 CWD, DAG.getConstant(0x800, MVT::i16)),
6321 DAG.getConstant(11, MVT::i8));
6323 DAG.getNode(ISD::SRL, dl, MVT::i16,
6324 DAG.getNode(ISD::AND, dl, MVT::i16,
6325 CWD, DAG.getConstant(0x400, MVT::i16)),
6326 DAG.getConstant(9, MVT::i8));
6329 DAG.getNode(ISD::AND, dl, MVT::i16,
6330 DAG.getNode(ISD::ADD, dl, MVT::i16,
6331 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6332 DAG.getConstant(1, MVT::i16)),
6333 DAG.getConstant(3, MVT::i16));
6336 return DAG.getNode((VT.getSizeInBits() < 16 ?
6337 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6340 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6341 MVT VT = Op.getValueType();
6343 unsigned NumBits = VT.getSizeInBits();
6344 DebugLoc dl = Op.getDebugLoc();
6346 Op = Op.getOperand(0);
6347 if (VT == MVT::i8) {
6348 // Zero extend to i32 since there is not an i8 bsr.
6350 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6353 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6354 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6355 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6357 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6358 SmallVector<SDValue, 4> Ops;
6360 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6361 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6362 Ops.push_back(Op.getValue(1));
6363 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6365 // Finally xor with NumBits-1.
6366 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6369 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6373 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6374 MVT VT = Op.getValueType();
6376 unsigned NumBits = VT.getSizeInBits();
6377 DebugLoc dl = Op.getDebugLoc();
6379 Op = Op.getOperand(0);
6380 if (VT == MVT::i8) {
6382 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6385 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6386 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6387 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6389 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6390 SmallVector<SDValue, 4> Ops;
6392 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6393 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6394 Ops.push_back(Op.getValue(1));
6395 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6398 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6402 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6403 MVT VT = Op.getValueType();
6404 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6405 DebugLoc dl = Op.getDebugLoc();
6407 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6408 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6409 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6410 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6411 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6413 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6414 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6415 // return AloBlo + AloBhi + AhiBlo;
6417 SDValue A = Op.getOperand(0);
6418 SDValue B = Op.getOperand(1);
6420 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6421 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6422 A, DAG.getConstant(32, MVT::i32));
6423 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6424 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6425 B, DAG.getConstant(32, MVT::i32));
6426 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6427 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6429 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6430 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6432 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6433 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6435 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6436 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6437 AloBhi, DAG.getConstant(32, MVT::i32));
6438 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6439 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6440 AhiBlo, DAG.getConstant(32, MVT::i32));
6441 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6442 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6447 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6448 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6449 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6450 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6451 // has only one use.
6452 SDNode *N = Op.getNode();
6453 SDValue LHS = N->getOperand(0);
6454 SDValue RHS = N->getOperand(1);
6455 unsigned BaseOp = 0;
6457 DebugLoc dl = Op.getDebugLoc();
6459 switch (Op.getOpcode()) {
6460 default: assert(0 && "Unknown ovf instruction!");
6462 // A subtract of one will be selected as a INC. Note that INC doesn't
6463 // set CF, so we can't do this for UADDO.
6464 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6465 if (C->getAPIntValue() == 1) {
6466 BaseOp = X86ISD::INC;
6470 BaseOp = X86ISD::ADD;
6474 BaseOp = X86ISD::ADD;
6478 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6479 // set CF, so we can't do this for USUBO.
6480 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6481 if (C->getAPIntValue() == 1) {
6482 BaseOp = X86ISD::DEC;
6486 BaseOp = X86ISD::SUB;
6490 BaseOp = X86ISD::SUB;
6494 BaseOp = X86ISD::SMUL;
6498 BaseOp = X86ISD::UMUL;
6503 // Also sets EFLAGS.
6504 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6505 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6508 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6509 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6511 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6515 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6516 MVT T = Op.getValueType();
6517 DebugLoc dl = Op.getDebugLoc();
6520 switch(T.getSimpleVT()) {
6522 assert(false && "Invalid value type!");
6523 case MVT::i8: Reg = X86::AL; size = 1; break;
6524 case MVT::i16: Reg = X86::AX; size = 2; break;
6525 case MVT::i32: Reg = X86::EAX; size = 4; break;
6527 assert(Subtarget->is64Bit() && "Node not type legal!");
6528 Reg = X86::RAX; size = 8;
6531 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6532 Op.getOperand(2), SDValue());
6533 SDValue Ops[] = { cpIn.getValue(0),
6536 DAG.getTargetConstant(size, MVT::i8),
6538 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6539 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6541 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6545 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6546 SelectionDAG &DAG) {
6547 assert(Subtarget->is64Bit() && "Result not type legalized?");
6548 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6549 SDValue TheChain = Op.getOperand(0);
6550 DebugLoc dl = Op.getDebugLoc();
6551 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6552 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6553 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6555 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6556 DAG.getConstant(32, MVT::i8));
6558 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6561 return DAG.getMergeValues(Ops, 2, dl);
6564 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6565 SDNode *Node = Op.getNode();
6566 DebugLoc dl = Node->getDebugLoc();
6567 MVT T = Node->getValueType(0);
6568 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6569 DAG.getConstant(0, T), Node->getOperand(2));
6570 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6571 cast<AtomicSDNode>(Node)->getMemoryVT(),
6572 Node->getOperand(0),
6573 Node->getOperand(1), negOp,
6574 cast<AtomicSDNode>(Node)->getSrcValue(),
6575 cast<AtomicSDNode>(Node)->getAlignment());
6578 /// LowerOperation - Provide custom lowering hooks for some operations.
6580 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6581 switch (Op.getOpcode()) {
6582 default: assert(0 && "Should not custom lower this!");
6583 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6584 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6585 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6586 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6587 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6588 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6589 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6590 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6591 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6592 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6593 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6594 case ISD::SHL_PARTS:
6595 case ISD::SRA_PARTS:
6596 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6597 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6598 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6599 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6600 case ISD::FABS: return LowerFABS(Op, DAG);
6601 case ISD::FNEG: return LowerFNEG(Op, DAG);
6602 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6603 case ISD::SETCC: return LowerSETCC(Op, DAG);
6604 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6605 case ISD::SELECT: return LowerSELECT(Op, DAG);
6606 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6607 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6608 case ISD::CALL: return LowerCALL(Op, DAG);
6609 case ISD::RET: return LowerRET(Op, DAG);
6610 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6611 case ISD::VASTART: return LowerVASTART(Op, DAG);
6612 case ISD::VAARG: return LowerVAARG(Op, DAG);
6613 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6614 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6615 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6616 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6617 case ISD::FRAME_TO_ARGS_OFFSET:
6618 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6619 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6620 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6621 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6622 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6623 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6624 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6625 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6631 case ISD::UMULO: return LowerXALUO(Op, DAG);
6632 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6636 void X86TargetLowering::
6637 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6638 SelectionDAG &DAG, unsigned NewOp) {
6639 MVT T = Node->getValueType(0);
6640 DebugLoc dl = Node->getDebugLoc();
6641 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6643 SDValue Chain = Node->getOperand(0);
6644 SDValue In1 = Node->getOperand(1);
6645 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6646 Node->getOperand(2), DAG.getIntPtrConstant(0));
6647 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6648 Node->getOperand(2), DAG.getIntPtrConstant(1));
6649 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6650 // have a MemOperand. Pass the info through as a normal operand.
6651 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6652 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6653 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6654 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6655 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6656 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6657 Results.push_back(Result.getValue(2));
6660 /// ReplaceNodeResults - Replace a node with an illegal result type
6661 /// with a new node built out of custom code.
6662 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6663 SmallVectorImpl<SDValue>&Results,
6664 SelectionDAG &DAG) {
6665 DebugLoc dl = N->getDebugLoc();
6666 switch (N->getOpcode()) {
6668 assert(false && "Do not know how to custom type legalize this operation!");
6670 case ISD::FP_TO_SINT: {
6671 std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
6672 SDValue FIST = Vals.first, StackSlot = Vals.second;
6673 if (FIST.getNode() != 0) {
6674 MVT VT = N->getValueType(0);
6675 // Return a load from the stack slot.
6676 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6680 case ISD::READCYCLECOUNTER: {
6681 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6682 SDValue TheChain = N->getOperand(0);
6683 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6684 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6686 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6688 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6689 SDValue Ops[] = { eax, edx };
6690 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6691 Results.push_back(edx.getValue(1));
6694 case ISD::ATOMIC_CMP_SWAP: {
6695 MVT T = N->getValueType(0);
6696 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6697 SDValue cpInL, cpInH;
6698 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6699 DAG.getConstant(0, MVT::i32));
6700 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6701 DAG.getConstant(1, MVT::i32));
6702 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6703 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6705 SDValue swapInL, swapInH;
6706 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6707 DAG.getConstant(0, MVT::i32));
6708 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6709 DAG.getConstant(1, MVT::i32));
6710 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6712 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6713 swapInL.getValue(1));
6714 SDValue Ops[] = { swapInH.getValue(0),
6716 swapInH.getValue(1) };
6717 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6718 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6719 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6720 MVT::i32, Result.getValue(1));
6721 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6722 MVT::i32, cpOutL.getValue(2));
6723 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6724 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6725 Results.push_back(cpOutH.getValue(1));
6728 case ISD::ATOMIC_LOAD_ADD:
6729 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6731 case ISD::ATOMIC_LOAD_AND:
6732 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6734 case ISD::ATOMIC_LOAD_NAND:
6735 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6737 case ISD::ATOMIC_LOAD_OR:
6738 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6740 case ISD::ATOMIC_LOAD_SUB:
6741 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6743 case ISD::ATOMIC_LOAD_XOR:
6744 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6746 case ISD::ATOMIC_SWAP:
6747 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6752 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6754 default: return NULL;
6755 case X86ISD::BSF: return "X86ISD::BSF";
6756 case X86ISD::BSR: return "X86ISD::BSR";
6757 case X86ISD::SHLD: return "X86ISD::SHLD";
6758 case X86ISD::SHRD: return "X86ISD::SHRD";
6759 case X86ISD::FAND: return "X86ISD::FAND";
6760 case X86ISD::FOR: return "X86ISD::FOR";
6761 case X86ISD::FXOR: return "X86ISD::FXOR";
6762 case X86ISD::FSRL: return "X86ISD::FSRL";
6763 case X86ISD::FILD: return "X86ISD::FILD";
6764 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6765 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6766 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6767 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6768 case X86ISD::FLD: return "X86ISD::FLD";
6769 case X86ISD::FST: return "X86ISD::FST";
6770 case X86ISD::CALL: return "X86ISD::CALL";
6771 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6772 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6773 case X86ISD::BT: return "X86ISD::BT";
6774 case X86ISD::CMP: return "X86ISD::CMP";
6775 case X86ISD::COMI: return "X86ISD::COMI";
6776 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6777 case X86ISD::SETCC: return "X86ISD::SETCC";
6778 case X86ISD::CMOV: return "X86ISD::CMOV";
6779 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6780 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6781 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6782 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6783 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6784 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6785 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6786 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6787 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6788 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6789 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6790 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6791 case X86ISD::FMAX: return "X86ISD::FMAX";
6792 case X86ISD::FMIN: return "X86ISD::FMIN";
6793 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6794 case X86ISD::FRCP: return "X86ISD::FRCP";
6795 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6796 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
6797 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6798 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6799 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6800 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6801 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6802 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6803 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6804 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6805 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6806 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6807 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6808 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6809 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6810 case X86ISD::VSHL: return "X86ISD::VSHL";
6811 case X86ISD::VSRL: return "X86ISD::VSRL";
6812 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6813 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6814 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6815 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6816 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6817 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6818 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6819 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6820 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6821 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6822 case X86ISD::ADD: return "X86ISD::ADD";
6823 case X86ISD::SUB: return "X86ISD::SUB";
6824 case X86ISD::SMUL: return "X86ISD::SMUL";
6825 case X86ISD::UMUL: return "X86ISD::UMUL";
6826 case X86ISD::INC: return "X86ISD::INC";
6827 case X86ISD::DEC: return "X86ISD::DEC";
6828 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
6832 // isLegalAddressingMode - Return true if the addressing mode represented
6833 // by AM is legal for this target, for a load/store of the specified type.
6834 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6835 const Type *Ty) const {
6836 // X86 supports extremely general addressing modes.
6838 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6839 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6843 // We can only fold this if we don't need an extra load.
6844 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6846 // If BaseGV requires a register, we cannot also have a BaseReg.
6847 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6851 // X86-64 only supports addr of globals in small code model.
6852 if (Subtarget->is64Bit()) {
6853 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6855 // If lower 4G is not available, then we must use rip-relative addressing.
6856 if (AM.BaseOffs || AM.Scale > 1)
6867 // These scales always work.
6872 // These scales are formed with basereg+scalereg. Only accept if there is
6877 default: // Other stuff never works.
6885 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6886 if (!Ty1->isInteger() || !Ty2->isInteger())
6888 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6889 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6890 if (NumBits1 <= NumBits2)
6892 return Subtarget->is64Bit() || NumBits1 < 64;
6895 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6896 if (!VT1.isInteger() || !VT2.isInteger())
6898 unsigned NumBits1 = VT1.getSizeInBits();
6899 unsigned NumBits2 = VT2.getSizeInBits();
6900 if (NumBits1 <= NumBits2)
6902 return Subtarget->is64Bit() || NumBits1 < 64;
6905 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
6906 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6907 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
6910 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
6911 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6912 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
6915 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6916 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6917 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6918 /// are assumed to be legal.
6920 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6922 // Only do shuffles on 128-bit vector types for now.
6923 if (VT.getSizeInBits() == 64)
6926 // FIXME: pshufb, blends, palignr, shifts.
6927 return (VT.getVectorNumElements() == 2 ||
6928 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
6929 isMOVLMask(M, VT) ||
6930 isSHUFPMask(M, VT) ||
6931 isPSHUFDMask(M, VT) ||
6932 isPSHUFHWMask(M, VT) ||
6933 isPSHUFLWMask(M, VT) ||
6934 isUNPCKLMask(M, VT) ||
6935 isUNPCKHMask(M, VT) ||
6936 isUNPCKL_v_undef_Mask(M, VT) ||
6937 isUNPCKH_v_undef_Mask(M, VT));
6941 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
6943 unsigned NumElts = VT.getVectorNumElements();
6944 // FIXME: This collection of masks seems suspect.
6947 if (NumElts == 4 && VT.getSizeInBits() == 128) {
6948 return (isMOVLMask(Mask, VT) ||
6949 isCommutedMOVLMask(Mask, VT, true) ||
6950 isSHUFPMask(Mask, VT) ||
6951 isCommutedSHUFPMask(Mask, VT));
6956 //===----------------------------------------------------------------------===//
6957 // X86 Scheduler Hooks
6958 //===----------------------------------------------------------------------===//
6960 // private utility function
6962 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6963 MachineBasicBlock *MBB,
6971 TargetRegisterClass *RC,
6972 bool invSrc) const {
6973 // For the atomic bitwise operator, we generate
6976 // ld t1 = [bitinstr.addr]
6977 // op t2 = t1, [bitinstr.val]
6979 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6981 // fallthrough -->nextMBB
6982 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6983 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6984 MachineFunction::iterator MBBIter = MBB;
6987 /// First build the CFG
6988 MachineFunction *F = MBB->getParent();
6989 MachineBasicBlock *thisMBB = MBB;
6990 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6991 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6992 F->insert(MBBIter, newMBB);
6993 F->insert(MBBIter, nextMBB);
6995 // Move all successors to thisMBB to nextMBB
6996 nextMBB->transferSuccessors(thisMBB);
6998 // Update thisMBB to fall through to newMBB
6999 thisMBB->addSuccessor(newMBB);
7001 // newMBB jumps to itself and fall through to nextMBB
7002 newMBB->addSuccessor(nextMBB);
7003 newMBB->addSuccessor(newMBB);
7005 // Insert instructions into newMBB based on incoming instruction
7006 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7007 "unexpected number of operands");
7008 DebugLoc dl = bInstr->getDebugLoc();
7009 MachineOperand& destOper = bInstr->getOperand(0);
7010 MachineOperand* argOpers[2 + X86AddrNumOperands];
7011 int numArgs = bInstr->getNumOperands() - 1;
7012 for (int i=0; i < numArgs; ++i)
7013 argOpers[i] = &bInstr->getOperand(i+1);
7015 // x86 address has 4 operands: base, index, scale, and displacement
7016 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7017 int valArgIndx = lastAddrIndx + 1;
7019 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7020 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7021 for (int i=0; i <= lastAddrIndx; ++i)
7022 (*MIB).addOperand(*argOpers[i]);
7024 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7026 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7031 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7032 assert((argOpers[valArgIndx]->isReg() ||
7033 argOpers[valArgIndx]->isImm()) &&
7035 if (argOpers[valArgIndx]->isReg())
7036 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7038 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7040 (*MIB).addOperand(*argOpers[valArgIndx]);
7042 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7045 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7046 for (int i=0; i <= lastAddrIndx; ++i)
7047 (*MIB).addOperand(*argOpers[i]);
7049 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7050 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7052 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7056 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7058 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7062 // private utility function: 64 bit atomics on 32 bit host.
7064 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7065 MachineBasicBlock *MBB,
7070 bool invSrc) const {
7071 // For the atomic bitwise operator, we generate
7072 // thisMBB (instructions are in pairs, except cmpxchg8b)
7073 // ld t1,t2 = [bitinstr.addr]
7075 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7076 // op t5, t6 <- out1, out2, [bitinstr.val]
7077 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7078 // mov ECX, EBX <- t5, t6
7079 // mov EAX, EDX <- t1, t2
7080 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7081 // mov t3, t4 <- EAX, EDX
7083 // result in out1, out2
7084 // fallthrough -->nextMBB
7086 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7087 const unsigned LoadOpc = X86::MOV32rm;
7088 const unsigned copyOpc = X86::MOV32rr;
7089 const unsigned NotOpc = X86::NOT32r;
7090 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7091 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7092 MachineFunction::iterator MBBIter = MBB;
7095 /// First build the CFG
7096 MachineFunction *F = MBB->getParent();
7097 MachineBasicBlock *thisMBB = MBB;
7098 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7099 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7100 F->insert(MBBIter, newMBB);
7101 F->insert(MBBIter, nextMBB);
7103 // Move all successors to thisMBB to nextMBB
7104 nextMBB->transferSuccessors(thisMBB);
7106 // Update thisMBB to fall through to newMBB
7107 thisMBB->addSuccessor(newMBB);
7109 // newMBB jumps to itself and fall through to nextMBB
7110 newMBB->addSuccessor(nextMBB);
7111 newMBB->addSuccessor(newMBB);
7113 DebugLoc dl = bInstr->getDebugLoc();
7114 // Insert instructions into newMBB based on incoming instruction
7115 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7116 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7117 "unexpected number of operands");
7118 MachineOperand& dest1Oper = bInstr->getOperand(0);
7119 MachineOperand& dest2Oper = bInstr->getOperand(1);
7120 MachineOperand* argOpers[2 + X86AddrNumOperands];
7121 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7122 argOpers[i] = &bInstr->getOperand(i+2);
7124 // x86 address has 4 operands: base, index, scale, and displacement
7125 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7127 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7128 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7129 for (int i=0; i <= lastAddrIndx; ++i)
7130 (*MIB).addOperand(*argOpers[i]);
7131 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7132 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7133 // add 4 to displacement.
7134 for (int i=0; i <= lastAddrIndx-2; ++i)
7135 (*MIB).addOperand(*argOpers[i]);
7136 MachineOperand newOp3 = *(argOpers[3]);
7138 newOp3.setImm(newOp3.getImm()+4);
7140 newOp3.setOffset(newOp3.getOffset()+4);
7141 (*MIB).addOperand(newOp3);
7142 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7144 // t3/4 are defined later, at the bottom of the loop
7145 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7146 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7147 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7148 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7149 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7150 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7152 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7153 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7155 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7156 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7162 int valArgIndx = lastAddrIndx + 1;
7163 assert((argOpers[valArgIndx]->isReg() ||
7164 argOpers[valArgIndx]->isImm()) &&
7166 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7167 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7168 if (argOpers[valArgIndx]->isReg())
7169 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7171 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7172 if (regOpcL != X86::MOV32rr)
7174 (*MIB).addOperand(*argOpers[valArgIndx]);
7175 assert(argOpers[valArgIndx + 1]->isReg() ==
7176 argOpers[valArgIndx]->isReg());
7177 assert(argOpers[valArgIndx + 1]->isImm() ==
7178 argOpers[valArgIndx]->isImm());
7179 if (argOpers[valArgIndx + 1]->isReg())
7180 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7182 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7183 if (regOpcH != X86::MOV32rr)
7185 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7187 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7189 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7192 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7194 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7197 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7198 for (int i=0; i <= lastAddrIndx; ++i)
7199 (*MIB).addOperand(*argOpers[i]);
7201 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7202 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7204 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7205 MIB.addReg(X86::EAX);
7206 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7207 MIB.addReg(X86::EDX);
7210 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7212 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7216 // private utility function
7218 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7219 MachineBasicBlock *MBB,
7220 unsigned cmovOpc) const {
7221 // For the atomic min/max operator, we generate
7224 // ld t1 = [min/max.addr]
7225 // mov t2 = [min/max.val]
7227 // cmov[cond] t2 = t1
7229 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7231 // fallthrough -->nextMBB
7233 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7234 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7235 MachineFunction::iterator MBBIter = MBB;
7238 /// First build the CFG
7239 MachineFunction *F = MBB->getParent();
7240 MachineBasicBlock *thisMBB = MBB;
7241 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7242 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7243 F->insert(MBBIter, newMBB);
7244 F->insert(MBBIter, nextMBB);
7246 // Move all successors to thisMBB to nextMBB
7247 nextMBB->transferSuccessors(thisMBB);
7249 // Update thisMBB to fall through to newMBB
7250 thisMBB->addSuccessor(newMBB);
7252 // newMBB jumps to newMBB and fall through to nextMBB
7253 newMBB->addSuccessor(nextMBB);
7254 newMBB->addSuccessor(newMBB);
7256 DebugLoc dl = mInstr->getDebugLoc();
7257 // Insert instructions into newMBB based on incoming instruction
7258 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7259 "unexpected number of operands");
7260 MachineOperand& destOper = mInstr->getOperand(0);
7261 MachineOperand* argOpers[2 + X86AddrNumOperands];
7262 int numArgs = mInstr->getNumOperands() - 1;
7263 for (int i=0; i < numArgs; ++i)
7264 argOpers[i] = &mInstr->getOperand(i+1);
7266 // x86 address has 4 operands: base, index, scale, and displacement
7267 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7268 int valArgIndx = lastAddrIndx + 1;
7270 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7271 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7272 for (int i=0; i <= lastAddrIndx; ++i)
7273 (*MIB).addOperand(*argOpers[i]);
7275 // We only support register and immediate values
7276 assert((argOpers[valArgIndx]->isReg() ||
7277 argOpers[valArgIndx]->isImm()) &&
7280 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7281 if (argOpers[valArgIndx]->isReg())
7282 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7284 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7285 (*MIB).addOperand(*argOpers[valArgIndx]);
7287 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7290 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7295 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7296 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7300 // Cmp and exchange if none has modified the memory location
7301 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7302 for (int i=0; i <= lastAddrIndx; ++i)
7303 (*MIB).addOperand(*argOpers[i]);
7305 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7306 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7308 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7309 MIB.addReg(X86::EAX);
7312 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7314 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7320 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7321 MachineBasicBlock *BB) const {
7322 DebugLoc dl = MI->getDebugLoc();
7323 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7324 switch (MI->getOpcode()) {
7325 default: assert(false && "Unexpected instr type to insert");
7326 case X86::CMOV_V1I64:
7327 case X86::CMOV_FR32:
7328 case X86::CMOV_FR64:
7329 case X86::CMOV_V4F32:
7330 case X86::CMOV_V2F64:
7331 case X86::CMOV_V2I64: {
7332 // To "insert" a SELECT_CC instruction, we actually have to insert the
7333 // diamond control-flow pattern. The incoming instruction knows the
7334 // destination vreg to set, the condition code register to branch on, the
7335 // true/false values to select between, and a branch opcode to use.
7336 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7337 MachineFunction::iterator It = BB;
7343 // cmpTY ccX, r1, r2
7345 // fallthrough --> copy0MBB
7346 MachineBasicBlock *thisMBB = BB;
7347 MachineFunction *F = BB->getParent();
7348 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7349 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7351 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7352 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7353 F->insert(It, copy0MBB);
7354 F->insert(It, sinkMBB);
7355 // Update machine-CFG edges by transferring all successors of the current
7356 // block to the new block which will contain the Phi node for the select.
7357 sinkMBB->transferSuccessors(BB);
7359 // Add the true and fallthrough blocks as its successors.
7360 BB->addSuccessor(copy0MBB);
7361 BB->addSuccessor(sinkMBB);
7364 // %FalseValue = ...
7365 // # fallthrough to sinkMBB
7368 // Update machine-CFG edges
7369 BB->addSuccessor(sinkMBB);
7372 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7375 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7376 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7377 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7379 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7383 case X86::FP32_TO_INT16_IN_MEM:
7384 case X86::FP32_TO_INT32_IN_MEM:
7385 case X86::FP32_TO_INT64_IN_MEM:
7386 case X86::FP64_TO_INT16_IN_MEM:
7387 case X86::FP64_TO_INT32_IN_MEM:
7388 case X86::FP64_TO_INT64_IN_MEM:
7389 case X86::FP80_TO_INT16_IN_MEM:
7390 case X86::FP80_TO_INT32_IN_MEM:
7391 case X86::FP80_TO_INT64_IN_MEM: {
7392 // Change the floating point control register to use "round towards zero"
7393 // mode when truncating to an integer value.
7394 MachineFunction *F = BB->getParent();
7395 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7396 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7398 // Load the old value of the high byte of the control word...
7400 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7401 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7404 // Set the high part to be round to zero...
7405 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7408 // Reload the modified control word now...
7409 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7411 // Restore the memory image of control word to original value
7412 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7415 // Get the X86 opcode to use.
7417 switch (MI->getOpcode()) {
7418 default: assert(0 && "illegal opcode!");
7419 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7420 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7421 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7422 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7423 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7424 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7425 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7426 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7427 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7431 MachineOperand &Op = MI->getOperand(0);
7433 AM.BaseType = X86AddressMode::RegBase;
7434 AM.Base.Reg = Op.getReg();
7436 AM.BaseType = X86AddressMode::FrameIndexBase;
7437 AM.Base.FrameIndex = Op.getIndex();
7439 Op = MI->getOperand(1);
7441 AM.Scale = Op.getImm();
7442 Op = MI->getOperand(2);
7444 AM.IndexReg = Op.getImm();
7445 Op = MI->getOperand(3);
7446 if (Op.isGlobal()) {
7447 AM.GV = Op.getGlobal();
7449 AM.Disp = Op.getImm();
7451 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7452 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7454 // Reload the original control word now.
7455 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7457 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7460 case X86::ATOMAND32:
7461 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7462 X86::AND32ri, X86::MOV32rm,
7463 X86::LCMPXCHG32, X86::MOV32rr,
7464 X86::NOT32r, X86::EAX,
7465 X86::GR32RegisterClass);
7467 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7468 X86::OR32ri, X86::MOV32rm,
7469 X86::LCMPXCHG32, X86::MOV32rr,
7470 X86::NOT32r, X86::EAX,
7471 X86::GR32RegisterClass);
7472 case X86::ATOMXOR32:
7473 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7474 X86::XOR32ri, X86::MOV32rm,
7475 X86::LCMPXCHG32, X86::MOV32rr,
7476 X86::NOT32r, X86::EAX,
7477 X86::GR32RegisterClass);
7478 case X86::ATOMNAND32:
7479 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7480 X86::AND32ri, X86::MOV32rm,
7481 X86::LCMPXCHG32, X86::MOV32rr,
7482 X86::NOT32r, X86::EAX,
7483 X86::GR32RegisterClass, true);
7484 case X86::ATOMMIN32:
7485 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7486 case X86::ATOMMAX32:
7487 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7488 case X86::ATOMUMIN32:
7489 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7490 case X86::ATOMUMAX32:
7491 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7493 case X86::ATOMAND16:
7494 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7495 X86::AND16ri, X86::MOV16rm,
7496 X86::LCMPXCHG16, X86::MOV16rr,
7497 X86::NOT16r, X86::AX,
7498 X86::GR16RegisterClass);
7500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7501 X86::OR16ri, X86::MOV16rm,
7502 X86::LCMPXCHG16, X86::MOV16rr,
7503 X86::NOT16r, X86::AX,
7504 X86::GR16RegisterClass);
7505 case X86::ATOMXOR16:
7506 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7507 X86::XOR16ri, X86::MOV16rm,
7508 X86::LCMPXCHG16, X86::MOV16rr,
7509 X86::NOT16r, X86::AX,
7510 X86::GR16RegisterClass);
7511 case X86::ATOMNAND16:
7512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7513 X86::AND16ri, X86::MOV16rm,
7514 X86::LCMPXCHG16, X86::MOV16rr,
7515 X86::NOT16r, X86::AX,
7516 X86::GR16RegisterClass, true);
7517 case X86::ATOMMIN16:
7518 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7519 case X86::ATOMMAX16:
7520 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7521 case X86::ATOMUMIN16:
7522 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7523 case X86::ATOMUMAX16:
7524 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7527 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7528 X86::AND8ri, X86::MOV8rm,
7529 X86::LCMPXCHG8, X86::MOV8rr,
7530 X86::NOT8r, X86::AL,
7531 X86::GR8RegisterClass);
7533 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7534 X86::OR8ri, X86::MOV8rm,
7535 X86::LCMPXCHG8, X86::MOV8rr,
7536 X86::NOT8r, X86::AL,
7537 X86::GR8RegisterClass);
7539 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7540 X86::XOR8ri, X86::MOV8rm,
7541 X86::LCMPXCHG8, X86::MOV8rr,
7542 X86::NOT8r, X86::AL,
7543 X86::GR8RegisterClass);
7544 case X86::ATOMNAND8:
7545 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7546 X86::AND8ri, X86::MOV8rm,
7547 X86::LCMPXCHG8, X86::MOV8rr,
7548 X86::NOT8r, X86::AL,
7549 X86::GR8RegisterClass, true);
7550 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7551 // This group is for 64-bit host.
7552 case X86::ATOMAND64:
7553 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7554 X86::AND64ri32, X86::MOV64rm,
7555 X86::LCMPXCHG64, X86::MOV64rr,
7556 X86::NOT64r, X86::RAX,
7557 X86::GR64RegisterClass);
7559 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7560 X86::OR64ri32, X86::MOV64rm,
7561 X86::LCMPXCHG64, X86::MOV64rr,
7562 X86::NOT64r, X86::RAX,
7563 X86::GR64RegisterClass);
7564 case X86::ATOMXOR64:
7565 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7566 X86::XOR64ri32, X86::MOV64rm,
7567 X86::LCMPXCHG64, X86::MOV64rr,
7568 X86::NOT64r, X86::RAX,
7569 X86::GR64RegisterClass);
7570 case X86::ATOMNAND64:
7571 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7572 X86::AND64ri32, X86::MOV64rm,
7573 X86::LCMPXCHG64, X86::MOV64rr,
7574 X86::NOT64r, X86::RAX,
7575 X86::GR64RegisterClass, true);
7576 case X86::ATOMMIN64:
7577 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7578 case X86::ATOMMAX64:
7579 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7580 case X86::ATOMUMIN64:
7581 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7582 case X86::ATOMUMAX64:
7583 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7585 // This group does 64-bit operations on a 32-bit host.
7586 case X86::ATOMAND6432:
7587 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7588 X86::AND32rr, X86::AND32rr,
7589 X86::AND32ri, X86::AND32ri,
7591 case X86::ATOMOR6432:
7592 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7593 X86::OR32rr, X86::OR32rr,
7594 X86::OR32ri, X86::OR32ri,
7596 case X86::ATOMXOR6432:
7597 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7598 X86::XOR32rr, X86::XOR32rr,
7599 X86::XOR32ri, X86::XOR32ri,
7601 case X86::ATOMNAND6432:
7602 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7603 X86::AND32rr, X86::AND32rr,
7604 X86::AND32ri, X86::AND32ri,
7606 case X86::ATOMADD6432:
7607 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7608 X86::ADD32rr, X86::ADC32rr,
7609 X86::ADD32ri, X86::ADC32ri,
7611 case X86::ATOMSUB6432:
7612 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7613 X86::SUB32rr, X86::SBB32rr,
7614 X86::SUB32ri, X86::SBB32ri,
7616 case X86::ATOMSWAP6432:
7617 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7618 X86::MOV32rr, X86::MOV32rr,
7619 X86::MOV32ri, X86::MOV32ri,
7624 //===----------------------------------------------------------------------===//
7625 // X86 Optimization Hooks
7626 //===----------------------------------------------------------------------===//
7628 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7632 const SelectionDAG &DAG,
7633 unsigned Depth) const {
7634 unsigned Opc = Op.getOpcode();
7635 assert((Opc >= ISD::BUILTIN_OP_END ||
7636 Opc == ISD::INTRINSIC_WO_CHAIN ||
7637 Opc == ISD::INTRINSIC_W_CHAIN ||
7638 Opc == ISD::INTRINSIC_VOID) &&
7639 "Should use MaskedValueIsZero if you don't know whether Op"
7640 " is a target node!");
7642 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7651 // These nodes' second result is a boolean.
7652 if (Op.getResNo() == 0)
7656 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7657 Mask.getBitWidth() - 1);
7662 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7663 /// node is a GlobalAddress + offset.
7664 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7665 GlobalValue* &GA, int64_t &Offset) const{
7666 if (N->getOpcode() == X86ISD::Wrapper) {
7667 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7668 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7669 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7673 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7676 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7677 const TargetLowering &TLI) {
7680 if (TLI.isGAPlusOffset(Base, GV, Offset))
7681 return (GV->getAlignment() >= N && (Offset % N) == 0);
7682 // DAG combine handles the stack object case.
7686 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7687 MVT EVT, SDNode *&Base,
7688 SelectionDAG &DAG, MachineFrameInfo *MFI,
7689 const TargetLowering &TLI) {
7691 for (unsigned i = 0; i < NumElems; ++i) {
7692 if (N->getMaskElt(i) < 0) {
7698 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7699 if (!Elt.getNode() ||
7700 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7703 Base = Elt.getNode();
7704 if (Base->getOpcode() == ISD::UNDEF)
7708 if (Elt.getOpcode() == ISD::UNDEF)
7711 if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
7712 EVT.getSizeInBits()/8, i, MFI))
7718 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7719 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7720 /// if the load addresses are consecutive, non-overlapping, and in the right
7721 /// order. In the case of v2i64, it will see if it can rewrite the
7722 /// shuffle to be an appropriate build vector so it can take advantage of
7723 // performBuildVectorCombine.
7724 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7725 const TargetLowering &TLI) {
7726 DebugLoc dl = N->getDebugLoc();
7727 MVT VT = N->getValueType(0);
7728 MVT EVT = VT.getVectorElementType();
7729 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7730 unsigned NumElems = VT.getVectorNumElements();
7732 // For x86-32 machines, if we see an insert and then a shuffle in a v2i64
7733 // where the upper half is 0, it is advantageous to rewrite it as a build
7734 // vector of (0, val) so it can use movq.
7735 if (VT == MVT::v2i64) {
7737 In[0] = N->getOperand(0);
7738 In[1] = N->getOperand(1);
7739 int Idx0 = SVN->getMaskElt(0);
7740 int Idx1 = SVN->getMaskElt(1);
7741 // FIXME: can we take advantage of undef index?
7742 if (Idx0 >= 0 && Idx1 >= 0 &&
7743 In[Idx0/2].getOpcode() == ISD::INSERT_VECTOR_ELT &&
7744 In[Idx1/2].getOpcode() == ISD::BUILD_VECTOR) {
7745 ConstantSDNode* InsertVecIdx =
7746 dyn_cast<ConstantSDNode>(In[Idx0/2].getOperand(2));
7748 InsertVecIdx->getZExtValue() == (unsigned)(Idx0 % 2) &&
7749 isZeroNode(In[Idx1/2].getOperand(Idx1 % 2))) {
7750 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
7751 In[Idx0/2].getOperand(1),
7752 In[Idx1/2].getOperand(Idx1 % 2));
7757 // Try to combine a vector_shuffle into a 128-bit load.
7758 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7759 SDNode *Base = NULL;
7760 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, Base, DAG, MFI, TLI))
7763 LoadSDNode *LD = cast<LoadSDNode>(Base);
7764 if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
7765 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7766 LD->getSrcValue(), LD->getSrcValueOffset(),
7768 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7769 LD->getSrcValue(), LD->getSrcValueOffset(),
7770 LD->isVolatile(), LD->getAlignment());
7773 /// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
7774 static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
7775 TargetLowering::DAGCombinerInfo &DCI,
7776 const X86Subtarget *Subtarget,
7777 const TargetLowering &TLI) {
7778 unsigned NumOps = N->getNumOperands();
7779 DebugLoc dl = N->getDebugLoc();
7781 // Ignore single operand BUILD_VECTOR.
7785 MVT VT = N->getValueType(0);
7786 MVT EVT = VT.getVectorElementType();
7787 if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
7788 // We are looking for load i64 and zero extend. We want to transform
7789 // it before legalizer has a chance to expand it. Also look for i64
7790 // BUILD_PAIR bit casted to f64.
7792 // This must be an insertion into a zero vector.
7793 SDValue HighElt = N->getOperand(1);
7794 if (!isZeroNode(HighElt))
7797 // Value must be a load.
7798 SDNode *Base = N->getOperand(0).getNode();
7799 if (!isa<LoadSDNode>(Base)) {
7800 if (Base->getOpcode() != ISD::BIT_CONVERT)
7802 Base = Base->getOperand(0).getNode();
7803 if (!isa<LoadSDNode>(Base))
7807 // Transform it into VZEXT_LOAD addr.
7808 LoadSDNode *LD = cast<LoadSDNode>(Base);
7810 // Load must not be an extload.
7811 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
7814 // Load type should legal type so we don't have to legalize it.
7815 if (!TLI.isTypeLegal(VT))
7818 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
7819 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7820 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7821 TargetLowering::TargetLoweringOpt TLO(DAG);
7822 TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
7823 DCI.CommitTargetLoweringOpt(TLO);
7827 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7828 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7829 const X86Subtarget *Subtarget) {
7830 DebugLoc DL = N->getDebugLoc();
7831 SDValue Cond = N->getOperand(0);
7832 // Get the LHS/RHS of the select.
7833 SDValue LHS = N->getOperand(1);
7834 SDValue RHS = N->getOperand(2);
7836 // If we have SSE[12] support, try to form min/max nodes.
7837 if (Subtarget->hasSSE2() &&
7838 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7839 Cond.getOpcode() == ISD::SETCC) {
7840 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7842 unsigned Opcode = 0;
7843 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7846 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7849 if (!UnsafeFPMath) break;
7851 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7853 Opcode = X86ISD::FMIN;
7856 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7859 if (!UnsafeFPMath) break;
7861 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7863 Opcode = X86ISD::FMAX;
7866 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7869 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7872 if (!UnsafeFPMath) break;
7874 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7876 Opcode = X86ISD::FMIN;
7879 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7882 if (!UnsafeFPMath) break;
7884 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7886 Opcode = X86ISD::FMAX;
7892 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
7895 // If this is a select between two integer constants, try to do some
7897 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
7898 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
7899 // Don't do this for crazy integer types.
7900 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
7901 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
7902 // so that TrueC (the true value) is larger than FalseC.
7903 bool NeedsCondInvert = false;
7905 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
7906 // Efficiently invertible.
7907 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
7908 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
7909 isa<ConstantSDNode>(Cond.getOperand(1))))) {
7910 NeedsCondInvert = true;
7911 std::swap(TrueC, FalseC);
7914 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
7915 if (FalseC->getAPIntValue() == 0 &&
7916 TrueC->getAPIntValue().isPowerOf2()) {
7917 if (NeedsCondInvert) // Invert the condition if needed.
7918 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7919 DAG.getConstant(1, Cond.getValueType()));
7921 // Zero extend the condition if needed.
7922 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
7924 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
7925 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
7926 DAG.getConstant(ShAmt, MVT::i8));
7929 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
7930 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
7931 if (NeedsCondInvert) // Invert the condition if needed.
7932 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7933 DAG.getConstant(1, Cond.getValueType()));
7935 // Zero extend the condition if needed.
7936 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
7937 FalseC->getValueType(0), Cond);
7938 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7939 SDValue(FalseC, 0));
7942 // Optimize cases that will turn into an LEA instruction. This requires
7943 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
7944 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
7945 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
7946 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
7948 bool isFastMultiplier = false;
7950 switch ((unsigned char)Diff) {
7952 case 1: // result = add base, cond
7953 case 2: // result = lea base( , cond*2)
7954 case 3: // result = lea base(cond, cond*2)
7955 case 4: // result = lea base( , cond*4)
7956 case 5: // result = lea base(cond, cond*4)
7957 case 8: // result = lea base( , cond*8)
7958 case 9: // result = lea base(cond, cond*8)
7959 isFastMultiplier = true;
7964 if (isFastMultiplier) {
7965 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
7966 if (NeedsCondInvert) // Invert the condition if needed.
7967 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7968 DAG.getConstant(1, Cond.getValueType()));
7970 // Zero extend the condition if needed.
7971 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
7973 // Scale the condition by the difference.
7975 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
7976 DAG.getConstant(Diff, Cond.getValueType()));
7978 // Add the base if non-zero.
7979 if (FalseC->getAPIntValue() != 0)
7980 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7981 SDValue(FalseC, 0));
7991 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
7992 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
7993 TargetLowering::DAGCombinerInfo &DCI) {
7994 DebugLoc DL = N->getDebugLoc();
7996 // If the flag operand isn't dead, don't touch this CMOV.
7997 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8000 // If this is a select between two integer constants, try to do some
8001 // optimizations. Note that the operands are ordered the opposite of SELECT
8003 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8004 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8005 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8006 // larger than FalseC (the false value).
8007 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8009 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8010 CC = X86::GetOppositeBranchCondition(CC);
8011 std::swap(TrueC, FalseC);
8014 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8015 // This is efficient for any integer data type (including i8/i16) and
8017 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8018 SDValue Cond = N->getOperand(3);
8019 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8020 DAG.getConstant(CC, MVT::i8), Cond);
8022 // Zero extend the condition if needed.
8023 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8025 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8026 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8027 DAG.getConstant(ShAmt, MVT::i8));
8028 if (N->getNumValues() == 2) // Dead flag value?
8029 return DCI.CombineTo(N, Cond, SDValue());
8033 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8034 // for any integer data type, including i8/i16.
8035 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8036 SDValue Cond = N->getOperand(3);
8037 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8038 DAG.getConstant(CC, MVT::i8), Cond);
8040 // Zero extend the condition if needed.
8041 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8042 FalseC->getValueType(0), Cond);
8043 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8044 SDValue(FalseC, 0));
8046 if (N->getNumValues() == 2) // Dead flag value?
8047 return DCI.CombineTo(N, Cond, SDValue());
8051 // Optimize cases that will turn into an LEA instruction. This requires
8052 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8053 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8054 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8055 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8057 bool isFastMultiplier = false;
8059 switch ((unsigned char)Diff) {
8061 case 1: // result = add base, cond
8062 case 2: // result = lea base( , cond*2)
8063 case 3: // result = lea base(cond, cond*2)
8064 case 4: // result = lea base( , cond*4)
8065 case 5: // result = lea base(cond, cond*4)
8066 case 8: // result = lea base( , cond*8)
8067 case 9: // result = lea base(cond, cond*8)
8068 isFastMultiplier = true;
8073 if (isFastMultiplier) {
8074 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8075 SDValue Cond = N->getOperand(3);
8076 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8077 DAG.getConstant(CC, MVT::i8), Cond);
8078 // Zero extend the condition if needed.
8079 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8081 // Scale the condition by the difference.
8083 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8084 DAG.getConstant(Diff, Cond.getValueType()));
8086 // Add the base if non-zero.
8087 if (FalseC->getAPIntValue() != 0)
8088 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8089 SDValue(FalseC, 0));
8090 if (N->getNumValues() == 2) // Dead flag value?
8091 return DCI.CombineTo(N, Cond, SDValue());
8101 /// PerformMulCombine - Optimize a single multiply with constant into two
8102 /// in order to implement it with two cheaper instructions, e.g.
8103 /// LEA + SHL, LEA + LEA.
8104 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8105 TargetLowering::DAGCombinerInfo &DCI) {
8106 if (DAG.getMachineFunction().
8107 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8110 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8113 MVT VT = N->getValueType(0);
8117 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8120 uint64_t MulAmt = C->getZExtValue();
8121 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8124 uint64_t MulAmt1 = 0;
8125 uint64_t MulAmt2 = 0;
8126 if ((MulAmt % 9) == 0) {
8128 MulAmt2 = MulAmt / 9;
8129 } else if ((MulAmt % 5) == 0) {
8131 MulAmt2 = MulAmt / 5;
8132 } else if ((MulAmt % 3) == 0) {
8134 MulAmt2 = MulAmt / 3;
8137 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8138 DebugLoc DL = N->getDebugLoc();
8140 if (isPowerOf2_64(MulAmt2) &&
8141 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8142 // If second multiplifer is pow2, issue it first. We want the multiply by
8143 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8145 std::swap(MulAmt1, MulAmt2);
8148 if (isPowerOf2_64(MulAmt1))
8149 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8150 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8152 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8153 DAG.getConstant(MulAmt1, VT));
8155 if (isPowerOf2_64(MulAmt2))
8156 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8157 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8159 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8160 DAG.getConstant(MulAmt2, VT));
8162 // Do not add new nodes to DAG combiner worklist.
8163 DCI.CombineTo(N, NewMul, false);
8169 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8171 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8172 const X86Subtarget *Subtarget) {
8173 // On X86 with SSE2 support, we can transform this to a vector shift if
8174 // all elements are shifted by the same amount. We can't do this in legalize
8175 // because the a constant vector is typically transformed to a constant pool
8176 // so we have no knowledge of the shift amount.
8177 if (!Subtarget->hasSSE2())
8180 MVT VT = N->getValueType(0);
8181 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8184 SDValue ShAmtOp = N->getOperand(1);
8185 MVT EltVT = VT.getVectorElementType();
8186 DebugLoc DL = N->getDebugLoc();
8188 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8189 unsigned NumElts = VT.getVectorNumElements();
8191 for (; i != NumElts; ++i) {
8192 SDValue Arg = ShAmtOp.getOperand(i);
8193 if (Arg.getOpcode() == ISD::UNDEF) continue;
8197 for (; i != NumElts; ++i) {
8198 SDValue Arg = ShAmtOp.getOperand(i);
8199 if (Arg.getOpcode() == ISD::UNDEF) continue;
8200 if (Arg != BaseShAmt) {
8204 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8205 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8206 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8207 DAG.getIntPtrConstant(0));
8211 if (EltVT.bitsGT(MVT::i32))
8212 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8213 else if (EltVT.bitsLT(MVT::i32))
8214 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8216 // The shift amount is identical so we can do a vector shift.
8217 SDValue ValOp = N->getOperand(0);
8218 switch (N->getOpcode()) {
8220 assert(0 && "Unknown shift opcode!");
8223 if (VT == MVT::v2i64)
8224 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8225 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8227 if (VT == MVT::v4i32)
8228 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8229 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8231 if (VT == MVT::v8i16)
8232 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8233 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8237 if (VT == MVT::v4i32)
8238 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8239 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8241 if (VT == MVT::v8i16)
8242 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8243 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8247 if (VT == MVT::v2i64)
8248 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8249 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8251 if (VT == MVT::v4i32)
8252 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8253 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8255 if (VT == MVT::v8i16)
8256 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8257 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8264 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8265 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8266 const X86Subtarget *Subtarget) {
8267 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8268 // the FP state in cases where an emms may be missing.
8269 // A preferable solution to the general problem is to figure out the right
8270 // places to insert EMMS. This qualifies as a quick hack.
8272 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8273 StoreSDNode *St = cast<StoreSDNode>(N);
8274 MVT VT = St->getValue().getValueType();
8275 if (VT.getSizeInBits() != 64)
8278 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloat && Subtarget->hasSSE2();
8279 if ((VT.isVector() ||
8280 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8281 isa<LoadSDNode>(St->getValue()) &&
8282 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8283 St->getChain().hasOneUse() && !St->isVolatile()) {
8284 SDNode* LdVal = St->getValue().getNode();
8286 int TokenFactorIndex = -1;
8287 SmallVector<SDValue, 8> Ops;
8288 SDNode* ChainVal = St->getChain().getNode();
8289 // Must be a store of a load. We currently handle two cases: the load
8290 // is a direct child, and it's under an intervening TokenFactor. It is
8291 // possible to dig deeper under nested TokenFactors.
8292 if (ChainVal == LdVal)
8293 Ld = cast<LoadSDNode>(St->getChain());
8294 else if (St->getValue().hasOneUse() &&
8295 ChainVal->getOpcode() == ISD::TokenFactor) {
8296 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8297 if (ChainVal->getOperand(i).getNode() == LdVal) {
8298 TokenFactorIndex = i;
8299 Ld = cast<LoadSDNode>(St->getValue());
8301 Ops.push_back(ChainVal->getOperand(i));
8305 if (!Ld || !ISD::isNormalLoad(Ld))
8308 // If this is not the MMX case, i.e. we are just turning i64 load/store
8309 // into f64 load/store, avoid the transformation if there are multiple
8310 // uses of the loaded value.
8311 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8314 DebugLoc LdDL = Ld->getDebugLoc();
8315 DebugLoc StDL = N->getDebugLoc();
8316 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8317 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8319 if (Subtarget->is64Bit() || F64IsLegal) {
8320 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8321 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8322 Ld->getBasePtr(), Ld->getSrcValue(),
8323 Ld->getSrcValueOffset(), Ld->isVolatile(),
8324 Ld->getAlignment());
8325 SDValue NewChain = NewLd.getValue(1);
8326 if (TokenFactorIndex != -1) {
8327 Ops.push_back(NewChain);
8328 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8331 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8332 St->getSrcValue(), St->getSrcValueOffset(),
8333 St->isVolatile(), St->getAlignment());
8336 // Otherwise, lower to two pairs of 32-bit loads / stores.
8337 SDValue LoAddr = Ld->getBasePtr();
8338 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8339 DAG.getConstant(4, MVT::i32));
8341 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8342 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8343 Ld->isVolatile(), Ld->getAlignment());
8344 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8345 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8347 MinAlign(Ld->getAlignment(), 4));
8349 SDValue NewChain = LoLd.getValue(1);
8350 if (TokenFactorIndex != -1) {
8351 Ops.push_back(LoLd);
8352 Ops.push_back(HiLd);
8353 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8357 LoAddr = St->getBasePtr();
8358 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8359 DAG.getConstant(4, MVT::i32));
8361 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8362 St->getSrcValue(), St->getSrcValueOffset(),
8363 St->isVolatile(), St->getAlignment());
8364 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8366 St->getSrcValueOffset() + 4,
8368 MinAlign(St->getAlignment(), 4));
8369 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8374 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8375 /// X86ISD::FXOR nodes.
8376 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8377 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8378 // F[X]OR(0.0, x) -> x
8379 // F[X]OR(x, 0.0) -> x
8380 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8381 if (C->getValueAPF().isPosZero())
8382 return N->getOperand(1);
8383 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8384 if (C->getValueAPF().isPosZero())
8385 return N->getOperand(0);
8389 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8390 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8391 // FAND(0.0, x) -> 0.0
8392 // FAND(x, 0.0) -> 0.0
8393 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8394 if (C->getValueAPF().isPosZero())
8395 return N->getOperand(0);
8396 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8397 if (C->getValueAPF().isPosZero())
8398 return N->getOperand(1);
8402 static SDValue PerformBTCombine(SDNode *N,
8404 TargetLowering::DAGCombinerInfo &DCI) {
8405 // BT ignores high bits in the bit index operand.
8406 SDValue Op1 = N->getOperand(1);
8407 if (Op1.hasOneUse()) {
8408 unsigned BitWidth = Op1.getValueSizeInBits();
8409 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8410 APInt KnownZero, KnownOne;
8411 TargetLowering::TargetLoweringOpt TLO(DAG);
8412 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8413 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8414 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8415 DCI.CommitTargetLoweringOpt(TLO);
8420 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8421 DAGCombinerInfo &DCI) const {
8422 SelectionDAG &DAG = DCI.DAG;
8423 switch (N->getOpcode()) {
8425 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8426 case ISD::BUILD_VECTOR:
8427 return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
8428 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8429 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8430 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8433 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8434 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8436 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8437 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8438 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8444 //===----------------------------------------------------------------------===//
8445 // X86 Inline Assembly Support
8446 //===----------------------------------------------------------------------===//
8448 /// getConstraintType - Given a constraint letter, return the type of
8449 /// constraint it is for this target.
8450 X86TargetLowering::ConstraintType
8451 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8452 if (Constraint.size() == 1) {
8453 switch (Constraint[0]) {
8465 return C_RegisterClass;
8473 return TargetLowering::getConstraintType(Constraint);
8476 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8477 /// with another that has more specific requirements based on the type of the
8478 /// corresponding operand.
8479 const char *X86TargetLowering::
8480 LowerXConstraint(MVT ConstraintVT) const {
8481 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8482 // 'f' like normal targets.
8483 if (ConstraintVT.isFloatingPoint()) {
8484 if (Subtarget->hasSSE2())
8486 if (Subtarget->hasSSE1())
8490 return TargetLowering::LowerXConstraint(ConstraintVT);
8493 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8494 /// vector. If it is invalid, don't add anything to Ops.
8495 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8498 std::vector<SDValue>&Ops,
8499 SelectionDAG &DAG) const {
8500 SDValue Result(0, 0);
8502 switch (Constraint) {
8505 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8506 if (C->getZExtValue() <= 31) {
8507 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8513 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8514 if (C->getZExtValue() <= 63) {
8515 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8521 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8522 if (C->getZExtValue() <= 255) {
8523 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8529 // 32-bit signed value
8530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8531 const ConstantInt *CI = C->getConstantIntValue();
8532 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8533 // Widen to 64 bits here to get it sign extended.
8534 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8537 // FIXME gcc accepts some relocatable values here too, but only in certain
8538 // memory models; it's complicated.
8543 // 32-bit unsigned value
8544 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8545 const ConstantInt *CI = C->getConstantIntValue();
8546 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8547 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8551 // FIXME gcc accepts some relocatable values here too, but only in certain
8552 // memory models; it's complicated.
8556 // Literal immediates are always ok.
8557 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8558 // Widen to 64 bits here to get it sign extended.
8559 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8563 // If we are in non-pic codegen mode, we allow the address of a global (with
8564 // an optional displacement) to be used with 'i'.
8565 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
8568 // Match either (GA) or (GA+C)
8570 Offset = GA->getOffset();
8571 } else if (Op.getOpcode() == ISD::ADD) {
8572 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8573 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8575 Offset = GA->getOffset()+C->getZExtValue();
8577 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8578 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
8580 Offset = GA->getOffset()+C->getZExtValue();
8588 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
8591 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8597 // Otherwise, not valid for this mode.
8602 if (Result.getNode()) {
8603 Ops.push_back(Result);
8606 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8610 std::vector<unsigned> X86TargetLowering::
8611 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8613 if (Constraint.size() == 1) {
8614 // FIXME: not handling fp-stack yet!
8615 switch (Constraint[0]) { // GCC X86 Constraint Letters
8616 default: break; // Unknown constraint letter
8617 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8620 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8621 else if (VT == MVT::i16)
8622 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8623 else if (VT == MVT::i8)
8624 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8625 else if (VT == MVT::i64)
8626 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8631 return std::vector<unsigned>();
8634 std::pair<unsigned, const TargetRegisterClass*>
8635 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8637 // First, see if this is a constraint that directly corresponds to an LLVM
8639 if (Constraint.size() == 1) {
8640 // GCC Constraint Letters
8641 switch (Constraint[0]) {
8643 case 'r': // GENERAL_REGS
8644 case 'R': // LEGACY_REGS
8645 case 'l': // INDEX_REGS
8647 return std::make_pair(0U, X86::GR8RegisterClass);
8649 return std::make_pair(0U, X86::GR16RegisterClass);
8650 if (VT == MVT::i32 || !Subtarget->is64Bit())
8651 return std::make_pair(0U, X86::GR32RegisterClass);
8652 return std::make_pair(0U, X86::GR64RegisterClass);
8653 case 'f': // FP Stack registers.
8654 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8655 // value to the correct fpstack register class.
8656 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8657 return std::make_pair(0U, X86::RFP32RegisterClass);
8658 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8659 return std::make_pair(0U, X86::RFP64RegisterClass);
8660 return std::make_pair(0U, X86::RFP80RegisterClass);
8661 case 'y': // MMX_REGS if MMX allowed.
8662 if (!Subtarget->hasMMX()) break;
8663 return std::make_pair(0U, X86::VR64RegisterClass);
8664 case 'Y': // SSE_REGS if SSE2 allowed
8665 if (!Subtarget->hasSSE2()) break;
8667 case 'x': // SSE_REGS if SSE1 allowed
8668 if (!Subtarget->hasSSE1()) break;
8670 switch (VT.getSimpleVT()) {
8672 // Scalar SSE types.
8675 return std::make_pair(0U, X86::FR32RegisterClass);
8678 return std::make_pair(0U, X86::FR64RegisterClass);
8686 return std::make_pair(0U, X86::VR128RegisterClass);
8692 // Use the default implementation in TargetLowering to convert the register
8693 // constraint into a member of a register class.
8694 std::pair<unsigned, const TargetRegisterClass*> Res;
8695 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8697 // Not found as a standard register?
8698 if (Res.second == 0) {
8699 // GCC calls "st(0)" just plain "st".
8700 if (StringsEqualNoCase("{st}", Constraint)) {
8701 Res.first = X86::ST0;
8702 Res.second = X86::RFP80RegisterClass;
8704 // 'A' means EAX + EDX.
8705 if (Constraint == "A") {
8706 Res.first = X86::EAX;
8707 Res.second = X86::GRADRegisterClass;
8712 // Otherwise, check to see if this is a register class of the wrong value
8713 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8714 // turn into {ax},{dx}.
8715 if (Res.second->hasType(VT))
8716 return Res; // Correct type already, nothing to do.
8718 // All of the single-register GCC register classes map their values onto
8719 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8720 // really want an 8-bit or 32-bit register, map to the appropriate register
8721 // class and return the appropriate register.
8722 if (Res.second == X86::GR16RegisterClass) {
8723 if (VT == MVT::i8) {
8724 unsigned DestReg = 0;
8725 switch (Res.first) {
8727 case X86::AX: DestReg = X86::AL; break;
8728 case X86::DX: DestReg = X86::DL; break;
8729 case X86::CX: DestReg = X86::CL; break;
8730 case X86::BX: DestReg = X86::BL; break;
8733 Res.first = DestReg;
8734 Res.second = X86::GR8RegisterClass;
8736 } else if (VT == MVT::i32) {
8737 unsigned DestReg = 0;
8738 switch (Res.first) {
8740 case X86::AX: DestReg = X86::EAX; break;
8741 case X86::DX: DestReg = X86::EDX; break;
8742 case X86::CX: DestReg = X86::ECX; break;
8743 case X86::BX: DestReg = X86::EBX; break;
8744 case X86::SI: DestReg = X86::ESI; break;
8745 case X86::DI: DestReg = X86::EDI; break;
8746 case X86::BP: DestReg = X86::EBP; break;
8747 case X86::SP: DestReg = X86::ESP; break;
8750 Res.first = DestReg;
8751 Res.second = X86::GR32RegisterClass;
8753 } else if (VT == MVT::i64) {
8754 unsigned DestReg = 0;
8755 switch (Res.first) {
8757 case X86::AX: DestReg = X86::RAX; break;
8758 case X86::DX: DestReg = X86::RDX; break;
8759 case X86::CX: DestReg = X86::RCX; break;
8760 case X86::BX: DestReg = X86::RBX; break;
8761 case X86::SI: DestReg = X86::RSI; break;
8762 case X86::DI: DestReg = X86::RDI; break;
8763 case X86::BP: DestReg = X86::RBP; break;
8764 case X86::SP: DestReg = X86::RSP; break;
8767 Res.first = DestReg;
8768 Res.second = X86::GR64RegisterClass;
8771 } else if (Res.second == X86::FR32RegisterClass ||
8772 Res.second == X86::FR64RegisterClass ||
8773 Res.second == X86::VR128RegisterClass) {
8774 // Handle references to XMM physical registers that got mapped into the
8775 // wrong class. This can happen with constraints like {xmm0} where the
8776 // target independent register mapper will just pick the first match it can
8777 // find, ignoring the required type.
8779 Res.second = X86::FR32RegisterClass;
8780 else if (VT == MVT::f64)
8781 Res.second = X86::FR64RegisterClass;
8782 else if (X86::VR128RegisterClass->hasType(VT))
8783 Res.second = X86::VR128RegisterClass;
8789 //===----------------------------------------------------------------------===//
8790 // X86 Widen vector type
8791 //===----------------------------------------------------------------------===//
8793 /// getWidenVectorType: given a vector type, returns the type to widen
8794 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8795 /// If there is no vector type that we want to widen to, returns MVT::Other
8796 /// When and where to widen is target dependent based on the cost of
8797 /// scalarizing vs using the wider vector type.
8799 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8800 assert(VT.isVector());
8801 if (isTypeLegal(VT))
8804 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8805 // type based on element type. This would speed up our search (though
8806 // it may not be worth it since the size of the list is relatively
8808 MVT EltVT = VT.getVectorElementType();
8809 unsigned NElts = VT.getVectorNumElements();
8811 // On X86, it make sense to widen any vector wider than 1
8815 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8816 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8817 MVT SVT = (MVT::SimpleValueType)nVT;
8819 if (isTypeLegal(SVT) &&
8820 SVT.getVectorElementType() == EltVT &&
8821 SVT.getVectorNumElements() > NElts)