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/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.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/ParameterAttributes.h"
44 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
45 : TargetLowering(TM) {
46 Subtarget = &TM.getSubtarget<X86Subtarget>();
47 X86ScalarSSEf64 = Subtarget->hasSSE2();
48 X86ScalarSSEf32 = Subtarget->hasSSE1();
49 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
53 RegInfo = TM.getRegisterInfo();
55 // Set up the TargetLowering object.
57 // X86 is weird, it always uses i8 for shift amounts and setcc results.
58 setShiftAmountType(MVT::i8);
59 setSetCCResultType(MVT::i8);
60 setSetCCResultContents(ZeroOrOneSetCCResult);
61 setSchedulingPreference(SchedulingForRegPressure);
62 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
63 setStackPointerRegisterToSaveRestore(X86StackPtr);
65 if (Subtarget->isTargetDarwin()) {
66 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
67 setUseUnderscoreSetJmp(false);
68 setUseUnderscoreLongJmp(false);
69 } else if (Subtarget->isTargetMingw()) {
70 // MS runtime is weird: it exports _setjmp, but longjmp!
71 setUseUnderscoreSetJmp(true);
72 setUseUnderscoreLongJmp(false);
74 setUseUnderscoreSetJmp(true);
75 setUseUnderscoreLongJmp(true);
78 // Set up the register classes.
79 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
80 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
81 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
82 if (Subtarget->is64Bit())
83 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
85 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
87 // We don't accept any truncstore of integer registers.
88 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
89 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
90 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
91 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
92 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
93 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
95 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
97 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
99 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
101 if (Subtarget->is64Bit()) {
102 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
103 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
106 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
107 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
109 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
112 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
114 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
116 // SSE has no i16 to fp conversion, only i32
117 if (X86ScalarSSEf32) {
118 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
119 // f32 and f64 cases are Legal, f80 case is not
120 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
122 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
123 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
126 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
127 // are Legal, f80 is custom lowered.
128 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
129 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
131 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
133 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
134 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
136 if (X86ScalarSSEf32) {
137 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
138 // f32 and f64 cases are Legal, f80 case is not
139 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
141 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
142 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
145 // Handle FP_TO_UINT by promoting the destination to a larger signed
147 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
148 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
149 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
151 if (Subtarget->is64Bit()) {
152 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
153 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
155 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
156 // Expand FP_TO_UINT into a select.
157 // FIXME: We would like to use a Custom expander here eventually to do
158 // the optimal thing for SSE vs. the default expansion in the legalizer.
159 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
161 // With SSE3 we can use fisttpll to convert to a signed i64.
162 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
165 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
166 if (!X86ScalarSSEf64) {
167 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
168 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
171 // Scalar integer multiply, multiply-high, divide, and remainder are
172 // lowered to use operations that produce two results, to match the
173 // available instructions. This exposes the two-result form to trivial
174 // CSE, which is able to combine x/y and x%y into a single instruction,
175 // for example. The single-result multiply instructions are introduced
176 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
178 setOperationAction(ISD::MUL , MVT::i8 , Expand);
179 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
180 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
181 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
182 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
183 setOperationAction(ISD::SREM , MVT::i8 , Expand);
184 setOperationAction(ISD::UREM , MVT::i8 , Expand);
185 setOperationAction(ISD::MUL , MVT::i16 , Expand);
186 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
187 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
188 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
189 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
190 setOperationAction(ISD::SREM , MVT::i16 , Expand);
191 setOperationAction(ISD::UREM , MVT::i16 , Expand);
192 setOperationAction(ISD::MUL , MVT::i32 , Expand);
193 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
194 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
195 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
196 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
197 setOperationAction(ISD::SREM , MVT::i32 , Expand);
198 setOperationAction(ISD::UREM , MVT::i32 , Expand);
199 setOperationAction(ISD::MUL , MVT::i64 , Expand);
200 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
201 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
202 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
203 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
204 setOperationAction(ISD::SREM , MVT::i64 , Expand);
205 setOperationAction(ISD::UREM , MVT::i64 , Expand);
207 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
208 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
209 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
210 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
211 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
212 if (Subtarget->is64Bit())
213 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
214 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
216 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
217 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
218 setOperationAction(ISD::FREM , MVT::f64 , Expand);
219 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
221 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
222 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
223 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
224 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
225 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
226 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
227 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
228 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
229 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
230 if (Subtarget->is64Bit()) {
231 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
232 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
233 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
236 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
237 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
239 // These should be promoted to a larger select which is supported.
240 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
241 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
242 // X86 wants to expand cmov itself.
243 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
244 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
245 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
246 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
247 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
248 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
249 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
250 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
251 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
252 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
253 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
256 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
258 // X86 ret instruction may pop stack.
259 setOperationAction(ISD::RET , MVT::Other, Custom);
260 if (!Subtarget->is64Bit())
261 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
264 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
265 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
266 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
267 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
268 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
269 if (Subtarget->is64Bit()) {
270 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
271 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
272 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
273 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
275 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
276 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
277 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
278 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
279 // X86 wants to expand memset / memcpy itself.
280 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
281 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
283 // Use the default ISD::LOCATION, ISD::DECLARE expansion.
284 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
285 // FIXME - use subtarget debug flags
286 if (!Subtarget->isTargetDarwin() &&
287 !Subtarget->isTargetELF() &&
288 !Subtarget->isTargetCygMing())
289 setOperationAction(ISD::LABEL, MVT::Other, Expand);
291 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
292 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
293 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
294 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
295 if (Subtarget->is64Bit()) {
297 setExceptionPointerRegister(X86::RAX);
298 setExceptionSelectorRegister(X86::RDX);
300 setExceptionPointerRegister(X86::EAX);
301 setExceptionSelectorRegister(X86::EDX);
303 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
305 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
307 setOperationAction(ISD::TRAP, MVT::Other, Legal);
309 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
310 setOperationAction(ISD::VASTART , MVT::Other, Custom);
311 setOperationAction(ISD::VAARG , MVT::Other, Expand);
312 setOperationAction(ISD::VAEND , MVT::Other, Expand);
313 if (Subtarget->is64Bit())
314 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
316 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
318 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
319 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
320 if (Subtarget->is64Bit())
321 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
322 if (Subtarget->isTargetCygMing())
323 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
325 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
327 if (X86ScalarSSEf64) {
328 // f32 and f64 use SSE.
329 // Set up the FP register classes.
330 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
331 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
333 // Use ANDPD to simulate FABS.
334 setOperationAction(ISD::FABS , MVT::f64, Custom);
335 setOperationAction(ISD::FABS , MVT::f32, Custom);
337 // Use XORP to simulate FNEG.
338 setOperationAction(ISD::FNEG , MVT::f64, Custom);
339 setOperationAction(ISD::FNEG , MVT::f32, Custom);
341 // Use ANDPD and ORPD to simulate FCOPYSIGN.
342 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
343 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
345 // We don't support sin/cos/fmod
346 setOperationAction(ISD::FSIN , MVT::f64, Expand);
347 setOperationAction(ISD::FCOS , MVT::f64, Expand);
348 setOperationAction(ISD::FREM , MVT::f64, Expand);
349 setOperationAction(ISD::FSIN , MVT::f32, Expand);
350 setOperationAction(ISD::FCOS , MVT::f32, Expand);
351 setOperationAction(ISD::FREM , MVT::f32, Expand);
353 // Expand FP immediates into loads from the stack, except for the special
355 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
356 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
357 addLegalFPImmediate(APFloat(+0.0)); // xorpd
358 addLegalFPImmediate(APFloat(+0.0f)); // xorps
360 // Floating truncations from f80 and extensions to f80 go through memory.
361 // If optimizing, we lie about this though and handle it in
362 // InstructionSelectPreprocess so that dagcombine2 can hack on these.
364 setConvertAction(MVT::f32, MVT::f80, Expand);
365 setConvertAction(MVT::f64, MVT::f80, Expand);
366 setConvertAction(MVT::f80, MVT::f32, Expand);
367 setConvertAction(MVT::f80, MVT::f64, Expand);
369 } else if (X86ScalarSSEf32) {
370 // Use SSE for f32, x87 for f64.
371 // Set up the FP register classes.
372 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
373 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
375 // Use ANDPS to simulate FABS.
376 setOperationAction(ISD::FABS , MVT::f32, Custom);
378 // Use XORP to simulate FNEG.
379 setOperationAction(ISD::FNEG , MVT::f32, Custom);
381 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
383 // Use ANDPS and ORPS to simulate FCOPYSIGN.
384 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
385 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
387 // We don't support sin/cos/fmod
388 setOperationAction(ISD::FSIN , MVT::f32, Expand);
389 setOperationAction(ISD::FCOS , MVT::f32, Expand);
390 setOperationAction(ISD::FREM , MVT::f32, Expand);
392 // Expand FP immediates into loads from the stack, except for the special
394 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
395 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
396 addLegalFPImmediate(APFloat(+0.0f)); // xorps
397 addLegalFPImmediate(APFloat(+0.0)); // FLD0
398 addLegalFPImmediate(APFloat(+1.0)); // FLD1
399 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
400 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
402 // SSE <-> X87 conversions go through memory. If optimizing, we lie about
403 // this though and handle it in InstructionSelectPreprocess so that
404 // dagcombine2 can hack on these.
406 setConvertAction(MVT::f32, MVT::f64, Expand);
407 setConvertAction(MVT::f32, MVT::f80, Expand);
408 setConvertAction(MVT::f80, MVT::f32, Expand);
409 setConvertAction(MVT::f64, MVT::f32, Expand);
410 // And x87->x87 truncations also.
411 setConvertAction(MVT::f80, MVT::f64, Expand);
415 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
416 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
419 // f32 and f64 in x87.
420 // Set up the FP register classes.
421 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
422 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
424 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
425 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
426 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
427 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
429 // Floating truncations go through memory. If optimizing, we lie about
430 // this though and handle it in InstructionSelectPreprocess so that
431 // dagcombine2 can hack on these.
433 setConvertAction(MVT::f80, MVT::f32, Expand);
434 setConvertAction(MVT::f64, MVT::f32, Expand);
435 setConvertAction(MVT::f80, MVT::f64, Expand);
439 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
440 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
443 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
444 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
445 addLegalFPImmediate(APFloat(+0.0)); // FLD0
446 addLegalFPImmediate(APFloat(+1.0)); // FLD1
447 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
448 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
449 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
450 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
451 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
452 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
455 // Long double always uses X87.
456 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
457 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
458 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
460 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
461 APFloat TmpFlt(+0.0);
462 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
463 addLegalFPImmediate(TmpFlt); // FLD0
465 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
466 APFloat TmpFlt2(+1.0);
467 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
468 addLegalFPImmediate(TmpFlt2); // FLD1
469 TmpFlt2.changeSign();
470 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
474 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
475 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
478 // Always use a library call for pow.
479 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
480 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
481 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
483 // First set operation action for all vector types to expand. Then we
484 // will selectively turn on ones that can be effectively codegen'd.
485 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
486 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
487 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
488 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
489 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
490 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
491 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
492 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
493 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
494 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
495 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
496 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
497 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
498 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
499 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
500 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
501 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
502 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
503 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
504 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
505 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
506 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
507 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
508 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
509 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
510 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
511 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
512 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
513 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
514 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
515 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
516 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
517 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
518 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
519 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
520 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
521 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
522 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
523 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
526 if (Subtarget->hasMMX()) {
527 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
528 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
529 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
530 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
532 // FIXME: add MMX packed arithmetics
534 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
535 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
536 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
537 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
539 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
540 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
541 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
542 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
544 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
545 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
547 setOperationAction(ISD::AND, MVT::v8i8, Promote);
548 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
549 setOperationAction(ISD::AND, MVT::v4i16, Promote);
550 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
551 setOperationAction(ISD::AND, MVT::v2i32, Promote);
552 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
553 setOperationAction(ISD::AND, MVT::v1i64, Legal);
555 setOperationAction(ISD::OR, MVT::v8i8, Promote);
556 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
557 setOperationAction(ISD::OR, MVT::v4i16, Promote);
558 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
559 setOperationAction(ISD::OR, MVT::v2i32, Promote);
560 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
561 setOperationAction(ISD::OR, MVT::v1i64, Legal);
563 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
564 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
565 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
566 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
567 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
568 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
569 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
571 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
572 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
573 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
574 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
575 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
576 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
577 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
579 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
580 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
581 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
582 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
584 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
585 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
586 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
587 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
589 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
590 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
591 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
592 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
595 if (Subtarget->hasSSE1()) {
596 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
598 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
599 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
600 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
601 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
602 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
603 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
604 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
605 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
606 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
607 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
608 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
611 if (Subtarget->hasSSE2()) {
612 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
613 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
614 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
615 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
616 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
618 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
619 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
620 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
621 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
622 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
623 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
624 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
625 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
626 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
627 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
628 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
629 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
630 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
631 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
632 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
634 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
635 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
636 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
637 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
638 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
639 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
641 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
642 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
643 // Do not attempt to custom lower non-power-of-2 vectors
644 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
646 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
647 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
648 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
650 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
654 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
655 if (Subtarget->is64Bit())
656 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
658 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
659 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
660 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
661 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
662 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
663 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
664 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
665 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
666 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
667 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
668 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
669 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
672 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
674 // Custom lower v2i64 and v2f64 selects.
675 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
676 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
677 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
678 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
681 // We want to custom lower some of our intrinsics.
682 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
684 // We have target-specific dag combine patterns for the following nodes:
685 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
686 setTargetDAGCombine(ISD::SELECT);
688 computeRegisterProperties();
690 // FIXME: These should be based on subtarget info. Plus, the values should
691 // be smaller when we are in optimizing for size mode.
692 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
693 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
694 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
695 allowUnalignedMemoryAccesses = true; // x86 supports it!
698 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
699 /// the desired ByVal argument alignment.
700 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
703 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
704 if (VTy->getBitWidth() == 128)
706 else if (VTy->getBitWidth() == 64)
709 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
710 unsigned EltAlign = 0;
711 getMaxByValAlign(ATy->getElementType(), EltAlign);
712 if (EltAlign > MaxAlign)
714 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
715 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
716 unsigned EltAlign = 0;
717 getMaxByValAlign(STy->getElementType(i), EltAlign);
718 if (EltAlign > MaxAlign)
727 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
728 /// function arguments in the caller parameter area. For X86, aggregates
729 /// that contains are placed at 16-byte boundaries while the rest are at
730 /// 4-byte boundaries.
731 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
732 if (Subtarget->is64Bit())
733 return getTargetData()->getABITypeAlignment(Ty);
735 getMaxByValAlign(Ty, Align);
739 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
741 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
742 SelectionDAG &DAG) const {
743 if (usesGlobalOffsetTable())
744 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
745 if (!Subtarget->isPICStyleRIPRel())
746 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
750 //===----------------------------------------------------------------------===//
751 // Return Value Calling Convention Implementation
752 //===----------------------------------------------------------------------===//
754 #include "X86GenCallingConv.inc"
756 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
757 /// exists skip possible ISD:TokenFactor.
758 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
759 if (Chain.getOpcode() == X86ISD::TAILCALL) {
761 } else if (Chain.getOpcode() == ISD::TokenFactor) {
762 if (Chain.getNumOperands() &&
763 Chain.getOperand(0).getOpcode() == X86ISD::TAILCALL)
764 return Chain.getOperand(0);
769 /// LowerRET - Lower an ISD::RET node.
770 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
771 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
773 SmallVector<CCValAssign, 16> RVLocs;
774 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
775 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
776 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
777 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
779 // If this is the first return lowered for this function, add the regs to the
780 // liveout set for the function.
781 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
782 for (unsigned i = 0; i != RVLocs.size(); ++i)
783 if (RVLocs[i].isRegLoc())
784 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
786 SDOperand Chain = Op.getOperand(0);
788 // Handle tail call return.
789 Chain = GetPossiblePreceedingTailCall(Chain);
790 if (Chain.getOpcode() == X86ISD::TAILCALL) {
791 SDOperand TailCall = Chain;
792 SDOperand TargetAddress = TailCall.getOperand(1);
793 SDOperand StackAdjustment = TailCall.getOperand(2);
794 assert(((TargetAddress.getOpcode() == ISD::Register &&
795 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
796 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
797 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
798 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
799 "Expecting an global address, external symbol, or register");
800 assert(StackAdjustment.getOpcode() == ISD::Constant &&
801 "Expecting a const value");
803 SmallVector<SDOperand,8> Operands;
804 Operands.push_back(Chain.getOperand(0));
805 Operands.push_back(TargetAddress);
806 Operands.push_back(StackAdjustment);
807 // Copy registers used by the call. Last operand is a flag so it is not
809 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
810 Operands.push_back(Chain.getOperand(i));
812 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
819 // Copy the result values into the output registers.
820 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
821 RVLocs[0].getLocReg() != X86::ST0) {
822 for (unsigned i = 0; i != RVLocs.size(); ++i) {
823 CCValAssign &VA = RVLocs[i];
824 assert(VA.isRegLoc() && "Can only return in registers!");
825 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
827 Flag = Chain.getValue(1);
830 // We need to handle a destination of ST0 specially, because it isn't really
832 SDOperand Value = Op.getOperand(1);
834 // an XMM register onto the fp-stack. Do this with an FP_EXTEND to f80.
835 // This will get legalized into a load/store if it can't get optimized away.
836 if (isScalarFPTypeInSSEReg(RVLocs[0].getValVT()))
837 Value = DAG.getNode(ISD::FP_EXTEND, MVT::f80, Value);
839 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
840 SDOperand Ops[] = { Chain, Value };
841 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
842 Flag = Chain.getValue(1);
845 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
847 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
849 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
853 /// LowerCallResult - Lower the result values of an ISD::CALL into the
854 /// appropriate copies out of appropriate physical registers. This assumes that
855 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
856 /// being lowered. The returns a SDNode with the same number of values as the
858 SDNode *X86TargetLowering::
859 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
860 unsigned CallingConv, SelectionDAG &DAG) {
862 // Assign locations to each value returned by this call.
863 SmallVector<CCValAssign, 16> RVLocs;
864 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
865 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
866 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
868 SmallVector<SDOperand, 8> ResultVals;
870 // Copy all of the result registers out of their specified physreg.
871 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
872 for (unsigned i = 0; i != RVLocs.size(); ++i) {
873 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
874 RVLocs[i].getValVT(), InFlag).getValue(1);
875 InFlag = Chain.getValue(2);
876 ResultVals.push_back(Chain.getValue(0));
879 // Copies from the FP stack are special, as ST0 isn't a valid register
880 // before the fp stackifier runs.
882 // Copy ST0 into an RFP register with FP_GET_RESULT. If this will end up
883 // in an SSE register, copy it out as F80 and do a truncate, otherwise use
884 // the specified value type.
885 MVT::ValueType GetResultTy = RVLocs[0].getValVT();
886 if (isScalarFPTypeInSSEReg(GetResultTy))
887 GetResultTy = MVT::f80;
888 SDVTList Tys = DAG.getVTList(GetResultTy, MVT::Other, MVT::Flag);
890 SDOperand GROps[] = { Chain, InFlag };
891 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
892 Chain = RetVal.getValue(1);
893 InFlag = RetVal.getValue(2);
895 // If we want the result in an SSE register, use an FP_TRUNCATE to get it
897 if (GetResultTy != RVLocs[0].getValVT())
898 RetVal = DAG.getNode(ISD::FP_ROUND, RVLocs[0].getValVT(), RetVal,
899 // This truncation won't change the value.
900 DAG.getIntPtrConstant(1));
902 ResultVals.push_back(RetVal);
905 // Merge everything together with a MERGE_VALUES node.
906 ResultVals.push_back(Chain);
907 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
908 &ResultVals[0], ResultVals.size()).Val;
911 /// LowerCallResultToTwo64BitRegs - Lower the result values of an x86-64
912 /// ISD::CALL where the results are known to be in two 64-bit registers,
913 /// e.g. XMM0 and XMM1. This simplify store the two values back to the
914 /// fixed stack slot allocated for StructRet.
915 SDNode *X86TargetLowering::
916 LowerCallResultToTwo64BitRegs(SDOperand Chain, SDOperand InFlag,
917 SDNode *TheCall, unsigned Reg1, unsigned Reg2,
918 MVT::ValueType VT, SelectionDAG &DAG) {
919 SDOperand RetVal1 = DAG.getCopyFromReg(Chain, Reg1, VT, InFlag);
920 Chain = RetVal1.getValue(1);
921 InFlag = RetVal1.getValue(2);
922 SDOperand RetVal2 = DAG.getCopyFromReg(Chain, Reg2, VT, InFlag);
923 Chain = RetVal2.getValue(1);
924 InFlag = RetVal2.getValue(2);
925 SDOperand FIN = TheCall->getOperand(5);
926 Chain = DAG.getStore(Chain, RetVal1, FIN, NULL, 0);
927 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
928 Chain = DAG.getStore(Chain, RetVal2, FIN, NULL, 0);
932 /// LowerCallResultToTwoX87Regs - Lower the result values of an x86-64 ISD::CALL
933 /// where the results are known to be in ST0 and ST1.
934 SDNode *X86TargetLowering::
935 LowerCallResultToTwoX87Regs(SDOperand Chain, SDOperand InFlag,
936 SDNode *TheCall, SelectionDAG &DAG) {
937 SmallVector<SDOperand, 8> ResultVals;
938 const MVT::ValueType VTs[] = { MVT::f80, MVT::f80, MVT::Other, MVT::Flag };
939 SDVTList Tys = DAG.getVTList(VTs, 4);
940 SDOperand Ops[] = { Chain, InFlag };
941 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT2, Tys, Ops, 2);
942 Chain = RetVal.getValue(2);
943 SDOperand FIN = TheCall->getOperand(5);
944 Chain = DAG.getStore(Chain, RetVal.getValue(1), FIN, NULL, 0);
945 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(16));
946 Chain = DAG.getStore(Chain, RetVal, FIN, NULL, 0);
950 //===----------------------------------------------------------------------===//
951 // C & StdCall & Fast Calling Convention implementation
952 //===----------------------------------------------------------------------===//
953 // StdCall calling convention seems to be standard for many Windows' API
954 // routines and around. It differs from C calling convention just a little:
955 // callee should clean up the stack, not caller. Symbols should be also
956 // decorated in some fancy way :) It doesn't support any vector arguments.
957 // For info on fast calling convention see Fast Calling Convention (tail call)
958 // implementation LowerX86_32FastCCCallTo.
960 /// AddLiveIn - This helper function adds the specified physical register to the
961 /// MachineFunction as a live in value. It also creates a corresponding virtual
963 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
964 const TargetRegisterClass *RC) {
965 assert(RC->contains(PReg) && "Not the correct regclass!");
966 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
967 MF.getRegInfo().addLiveIn(PReg, VReg);
971 // Determines whether a CALL node uses struct return semantics.
972 static bool CallIsStructReturn(SDOperand Op) {
973 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
977 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(6));
978 return Flags->getValue() & ISD::ParamFlags::StructReturn;
981 // Determines whether a FORMAL_ARGUMENTS node uses struct return semantics.
982 static bool ArgsAreStructReturn(SDOperand Op) {
983 unsigned NumArgs = Op.Val->getNumValues() - 1;
987 ConstantSDNode *Flags = cast<ConstantSDNode>(Op.getOperand(3));
988 return Flags->getValue() & ISD::ParamFlags::StructReturn;
991 // Determines whether a CALL or FORMAL_ARGUMENTS node requires the callee to pop
992 // its own arguments. Callee pop is necessary to support tail calls.
993 bool X86TargetLowering::IsCalleePop(SDOperand Op) {
994 bool IsVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
998 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1001 case CallingConv::X86_StdCall:
1002 return !Subtarget->is64Bit();
1003 case CallingConv::X86_FastCall:
1004 return !Subtarget->is64Bit();
1005 case CallingConv::Fast:
1006 return PerformTailCallOpt;
1010 // Selects the correct CCAssignFn for a CALL or FORMAL_ARGUMENTS node.
1011 CCAssignFn *X86TargetLowering::CCAssignFnForNode(SDOperand Op) const {
1012 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1014 if (Subtarget->is64Bit())
1015 if (CC == CallingConv::Fast && PerformTailCallOpt)
1016 return CC_X86_64_TailCall;
1020 if (CC == CallingConv::X86_FastCall)
1021 return CC_X86_32_FastCall;
1022 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1023 return CC_X86_32_TailCall;
1028 // Selects the appropriate decoration to apply to a MachineFunction containing a
1029 // given FORMAL_ARGUMENTS node.
1031 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDOperand Op) {
1032 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1033 if (CC == CallingConv::X86_FastCall)
1035 else if (CC == CallingConv::X86_StdCall)
1041 // IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could possibly
1042 // be overwritten when lowering the outgoing arguments in a tail call. Currently
1043 // the implementation of this call is very conservative and assumes all
1044 // arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with virtual
1045 // registers would be overwritten by direct lowering.
1046 // Possible improvement:
1047 // Check FORMAL_ARGUMENTS corresponding MERGE_VALUES for CopyFromReg nodes
1048 // indicating inreg passed arguments which also need not be lowered to a safe
1050 static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op) {
1051 RegisterSDNode * OpReg = NULL;
1052 if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
1053 (Op.getOpcode()== ISD::CopyFromReg &&
1054 (OpReg = cast<RegisterSDNode>(Op.getOperand(1))) &&
1055 OpReg->getReg() >= MRegisterInfo::FirstVirtualRegister))
1060 // CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1061 // by "Src" to address "Dst" with size and alignment information specified by
1062 // the specific parameter attribute. The copy will be passed as a byval function
1065 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1066 unsigned Flags, SelectionDAG &DAG) {
1067 unsigned Align = 1 <<
1068 ((Flags & ISD::ParamFlags::ByValAlign) >> ISD::ParamFlags::ByValAlignOffs);
1069 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1070 ISD::ParamFlags::ByValSizeOffs;
1071 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1072 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1073 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1074 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1077 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
1078 const CCValAssign &VA,
1079 MachineFrameInfo *MFI,
1080 SDOperand Root, unsigned i) {
1081 // Create the nodes corresponding to a load from this parameter slot.
1082 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
1083 bool isByVal = Flags & ISD::ParamFlags::ByVal;
1085 // FIXME: For now, all byval parameter objects are marked mutable. This
1086 // can be changed with more analysis.
1087 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1088 VA.getLocMemOffset(), !isByVal);
1089 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1092 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
1096 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
1097 MachineFunction &MF = DAG.getMachineFunction();
1098 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1100 const Function* Fn = MF.getFunction();
1101 if (Fn->hasExternalLinkage() &&
1102 Subtarget->isTargetCygMing() &&
1103 Fn->getName() == "main")
1104 FuncInfo->setForceFramePointer(true);
1106 // Decorate the function name.
1107 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1109 MachineFrameInfo *MFI = MF.getFrameInfo();
1110 SDOperand Root = Op.getOperand(0);
1111 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1112 unsigned CC = MF.getFunction()->getCallingConv();
1113 bool Is64Bit = Subtarget->is64Bit();
1115 assert(!(isVarArg && CC == CallingConv::Fast) &&
1116 "Var args not supported with calling convention fastcc");
1118 // Assign locations to all of the incoming arguments.
1119 SmallVector<CCValAssign, 16> ArgLocs;
1120 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1121 CCInfo.AnalyzeFormalArguments(Op.Val, CCAssignFnForNode(Op));
1123 SmallVector<SDOperand, 8> ArgValues;
1124 unsigned LastVal = ~0U;
1125 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1126 CCValAssign &VA = ArgLocs[i];
1127 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1129 assert(VA.getValNo() != LastVal &&
1130 "Don't support value assigned to multiple locs yet");
1131 LastVal = VA.getValNo();
1133 if (VA.isRegLoc()) {
1134 MVT::ValueType RegVT = VA.getLocVT();
1135 TargetRegisterClass *RC;
1136 if (RegVT == MVT::i32)
1137 RC = X86::GR32RegisterClass;
1138 else if (Is64Bit && RegVT == MVT::i64)
1139 RC = X86::GR64RegisterClass;
1140 else if (Is64Bit && RegVT == MVT::f32)
1141 RC = X86::FR32RegisterClass;
1142 else if (Is64Bit && RegVT == MVT::f64)
1143 RC = X86::FR64RegisterClass;
1145 assert(MVT::isVector(RegVT));
1146 if (Is64Bit && MVT::getSizeInBits(RegVT) == 64) {
1147 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1150 RC = X86::VR128RegisterClass;
1153 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1154 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1156 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1157 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1159 if (VA.getLocInfo() == CCValAssign::SExt)
1160 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1161 DAG.getValueType(VA.getValVT()));
1162 else if (VA.getLocInfo() == CCValAssign::ZExt)
1163 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1164 DAG.getValueType(VA.getValVT()));
1166 if (VA.getLocInfo() != CCValAssign::Full)
1167 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1169 // Handle MMX values passed in GPRs.
1170 if (Is64Bit && RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1171 MVT::getSizeInBits(RegVT) == 64)
1172 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1174 ArgValues.push_back(ArgValue);
1176 assert(VA.isMemLoc());
1177 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1181 unsigned StackSize = CCInfo.getNextStackOffset();
1182 // align stack specially for tail calls
1183 if (CC == CallingConv::Fast)
1184 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1186 // If the function takes variable number of arguments, make a frame index for
1187 // the start of the first vararg value... for expansion of llvm.va_start.
1189 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1190 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1193 static const unsigned GPR64ArgRegs[] = {
1194 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1196 static const unsigned XMMArgRegs[] = {
1197 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1198 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1201 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1202 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1204 // For X86-64, if there are vararg parameters that are passed via
1205 // registers, then we must store them to their spots on the stack so they
1206 // may be loaded by deferencing the result of va_next.
1207 VarArgsGPOffset = NumIntRegs * 8;
1208 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1209 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1211 // Store the integer parameter registers.
1212 SmallVector<SDOperand, 8> MemOps;
1213 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1214 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1215 DAG.getIntPtrConstant(VarArgsGPOffset));
1216 for (; NumIntRegs != 6; ++NumIntRegs) {
1217 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1218 X86::GR64RegisterClass);
1219 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1220 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1221 MemOps.push_back(Store);
1222 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1223 DAG.getIntPtrConstant(8));
1226 // Now store the XMM (fp + vector) parameter registers.
1227 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1228 DAG.getIntPtrConstant(VarArgsFPOffset));
1229 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1230 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1231 X86::VR128RegisterClass);
1232 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1233 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1234 MemOps.push_back(Store);
1235 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1236 DAG.getIntPtrConstant(16));
1238 if (!MemOps.empty())
1239 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1240 &MemOps[0], MemOps.size());
1244 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1245 // arguments and the arguments after the retaddr has been pushed are
1247 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1248 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1249 (StackSize & 7) == 0)
1252 ArgValues.push_back(Root);
1254 // Some CCs need callee pop.
1255 if (IsCalleePop(Op)) {
1256 BytesToPopOnReturn = StackSize; // Callee pops everything.
1257 BytesCallerReserves = 0;
1259 BytesToPopOnReturn = 0; // Callee pops nothing.
1260 // If this is an sret function, the return should pop the hidden pointer.
1261 if (!Is64Bit && ArgsAreStructReturn(Op))
1262 BytesToPopOnReturn = 4;
1263 BytesCallerReserves = StackSize;
1267 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1268 if (CC == CallingConv::X86_FastCall)
1269 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1272 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1274 // Return the new list of results.
1275 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1276 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1280 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1281 const SDOperand &StackPtr,
1282 const CCValAssign &VA,
1285 SDOperand PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
1286 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1287 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1288 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1289 if (Flags & ISD::ParamFlags::ByVal) {
1290 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG);
1292 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1295 /// ClassifyX86_64SRetCallReturn - Classify how to implement a x86-64
1296 /// struct return call to the specified function. X86-64 ABI specifies
1297 /// some SRet calls are actually returned in registers. Since current
1298 /// LLVM cannot represent multi-value calls, they are represent as
1299 /// calls where the results are passed in a hidden struct provided by
1300 /// the caller. This function examines the type of the struct to
1301 /// determine the correct way to implement the call.
1303 X86TargetLowering::ClassifyX86_64SRetCallReturn(const Function *Fn) {
1304 // FIXME: Disabled for now.
1305 return X86::InMemory;
1307 const PointerType *PTy = cast<PointerType>(Fn->arg_begin()->getType());
1308 const Type *RTy = PTy->getElementType();
1309 unsigned Size = getTargetData()->getABITypeSize(RTy);
1310 if (Size != 16 && Size != 32)
1311 return X86::InMemory;
1314 const StructType *STy = dyn_cast<StructType>(RTy);
1315 if (!STy) return X86::InMemory;
1316 if (STy->getNumElements() == 2 &&
1317 STy->getElementType(0) == Type::X86_FP80Ty &&
1318 STy->getElementType(1) == Type::X86_FP80Ty)
1323 for (Type::subtype_iterator I = RTy->subtype_begin(), E = RTy->subtype_end();
1325 const Type *STy = I->get();
1326 if (!STy->isFPOrFPVector()) {
1334 return X86::InGPR64;
1337 void X86TargetLowering::X86_64AnalyzeSRetCallOperands(SDNode *TheCall,
1340 unsigned NumOps = (TheCall->getNumOperands() - 5) / 2;
1341 for (unsigned i = 1; i != NumOps; ++i) {
1342 MVT::ValueType ArgVT = TheCall->getOperand(5+2*i).getValueType();
1343 SDOperand FlagOp = TheCall->getOperand(5+2*i+1);
1344 unsigned ArgFlags =cast<ConstantSDNode>(FlagOp)->getValue();
1345 if (Fn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo)) {
1346 cerr << "Call operand #" << i << " has unhandled type "
1347 << MVT::getValueTypeString(ArgVT) << "\n";
1353 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
1354 MachineFunction &MF = DAG.getMachineFunction();
1355 SDOperand Chain = Op.getOperand(0);
1356 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1357 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1358 bool IsTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0
1359 && CC == CallingConv::Fast && PerformTailCallOpt;
1360 SDOperand Callee = Op.getOperand(4);
1361 bool Is64Bit = Subtarget->is64Bit();
1362 bool IsStructRet = CallIsStructReturn(Op);
1364 assert(!(isVarArg && CC == CallingConv::Fast) &&
1365 "Var args not supported with calling convention fastcc");
1367 // Analyze operands of the call, assigning locations to each operand.
1368 SmallVector<CCValAssign, 16> ArgLocs;
1369 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1370 CCAssignFn *CCFn = CCAssignFnForNode(Op);
1372 X86::X86_64SRet SRetMethod = X86::InMemory;
1373 if (Is64Bit && IsStructRet)
1374 // FIXME: We can't figure out type of the sret structure for indirect
1375 // calls. We need to copy more information from CallSite to the ISD::CALL
1377 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1379 ClassifyX86_64SRetCallReturn(dyn_cast<Function>(G->getGlobal()));
1381 // UGLY HACK! For x86-64, some 128-bit aggregates are returns in a pair of
1382 // registers. Unfortunately, llvm does not support i128 yet so we pretend it's
1384 if (SRetMethod != X86::InMemory)
1385 X86_64AnalyzeSRetCallOperands(Op.Val, CCFn, CCInfo);
1387 CCInfo.AnalyzeCallOperands(Op.Val, CCFn);
1389 // Get a count of how many bytes are to be pushed on the stack.
1390 unsigned NumBytes = CCInfo.getNextStackOffset();
1391 if (CC == CallingConv::Fast)
1392 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1394 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1395 // arguments and the arguments after the retaddr has been pushed are aligned.
1396 if (!Is64Bit && CC == CallingConv::X86_FastCall &&
1397 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows() &&
1398 (NumBytes & 7) == 0)
1403 // Lower arguments at fp - stackoffset + fpdiff.
1404 unsigned NumBytesCallerPushed =
1405 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1406 FPDiff = NumBytesCallerPushed - NumBytes;
1408 // Set the delta of movement of the returnaddr stackslot.
1409 // But only set if delta is greater than previous delta.
1410 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1411 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1414 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes));
1416 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1418 // Adjust the Return address stack slot.
1420 MVT::ValueType VT = Is64Bit ? MVT::i64 : MVT::i32;
1421 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1422 // Load the "old" Return address.
1424 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1425 // Calculate the new stack slot for the return address.
1426 int SlotSize = Is64Bit ? 8 : 4;
1427 int NewReturnAddrFI =
1428 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1429 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1430 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1434 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1435 SmallVector<SDOperand, 8> MemOpChains;
1439 // Walk the register/memloc assignments, inserting copies/loads. For tail
1440 // calls, lower arguments which could otherwise be possibly overwritten to the
1441 // stack slot where they would go on normal function calls.
1442 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1443 CCValAssign &VA = ArgLocs[i];
1444 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1446 // Promote the value if needed.
1447 switch (VA.getLocInfo()) {
1448 default: assert(0 && "Unknown loc info!");
1449 case CCValAssign::Full: break;
1450 case CCValAssign::SExt:
1451 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1453 case CCValAssign::ZExt:
1454 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1456 case CCValAssign::AExt:
1457 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1461 if (VA.isRegLoc()) {
1462 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1464 if (!IsTailCall || IsPossiblyOverwrittenArgumentOfTailCall(Arg)) {
1465 assert(VA.isMemLoc());
1466 if (StackPtr.Val == 0)
1467 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1469 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1475 if (!MemOpChains.empty())
1476 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1477 &MemOpChains[0], MemOpChains.size());
1479 // Build a sequence of copy-to-reg nodes chained together with token chain
1480 // and flag operands which copy the outgoing args into registers.
1482 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1483 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1485 InFlag = Chain.getValue(1);
1489 InFlag = SDOperand(); // ??? Isn't this nuking the preceding loop's output?
1491 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1493 // Does not work with tail call since ebx is not restored correctly by
1494 // tailcaller. TODO: at least for x86 - verify for x86-64
1495 if (!IsTailCall && !Is64Bit &&
1496 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1497 Subtarget->isPICStyleGOT()) {
1498 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1499 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1501 InFlag = Chain.getValue(1);
1504 if (Is64Bit && isVarArg) {
1505 // From AMD64 ABI document:
1506 // For calls that may call functions that use varargs or stdargs
1507 // (prototype-less calls or calls to functions containing ellipsis (...) in
1508 // the declaration) %al is used as hidden argument to specify the number
1509 // of SSE registers used. The contents of %al do not need to match exactly
1510 // the number of registers, but must be an ubound on the number of SSE
1511 // registers used and is in the range 0 - 8 inclusive.
1513 // Count the number of XMM registers allocated.
1514 static const unsigned XMMArgRegs[] = {
1515 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1516 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1518 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1520 Chain = DAG.getCopyToReg(Chain, X86::AL,
1521 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1522 InFlag = Chain.getValue(1);
1525 // For tail calls lower the arguments to the 'real' stack slot.
1527 SmallVector<SDOperand, 8> MemOpChains2;
1530 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1531 CCValAssign &VA = ArgLocs[i];
1532 if (!VA.isRegLoc()) {
1533 assert(VA.isMemLoc());
1534 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1535 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1536 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1537 // Create frame index.
1538 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1539 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1540 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1541 FIN = DAG.getFrameIndex(FI, MVT::i32);
1542 SDOperand Source = Arg;
1543 if (IsPossiblyOverwrittenArgumentOfTailCall(Arg)) {
1544 // Copy from stack slots to stack slot of a tail called function. This
1545 // needs to be done because if we would lower the arguments directly
1546 // to their real stack slot we might end up overwriting each other.
1547 // Get source stack slot.
1548 Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1549 if (StackPtr.Val == 0)
1550 StackPtr = DAG.getCopyFromReg(Chain, X86StackPtr, getPointerTy());
1551 Source = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, Source);
1552 if ((Flags & ISD::ParamFlags::ByVal)==0)
1553 Source = DAG.getLoad(VA.getValVT(), Chain, Source, NULL, 0);
1556 if (Flags & ISD::ParamFlags::ByVal) {
1557 // Copy relative to framepointer.
1558 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1561 // Store relative to framepointer.
1562 MemOpChains2.push_back(DAG.getStore(Chain, Source, FIN, NULL, 0));
1567 if (!MemOpChains2.empty())
1568 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1569 &MemOpChains2[0], MemOpChains2.size());
1571 // Store the return address to the appropriate stack slot.
1573 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1576 // If the callee is a GlobalAddress node (quite common, every direct call is)
1577 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1578 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1579 // We should use extra load for direct calls to dllimported functions in
1581 if ((IsTailCall || !Is64Bit ||
1582 getTargetMachine().getCodeModel() != CodeModel::Large)
1583 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1584 getTargetMachine(), true))
1585 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1586 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1587 if (IsTailCall || !Is64Bit ||
1588 getTargetMachine().getCodeModel() != CodeModel::Large)
1589 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1590 } else if (IsTailCall) {
1591 assert(Callee.getOpcode() == ISD::LOAD &&
1592 "Function destination must be loaded into virtual register");
1593 unsigned Opc = Is64Bit ? X86::R9 : X86::ECX;
1595 Chain = DAG.getCopyToReg(Chain,
1596 DAG.getRegister(Opc, getPointerTy()) ,
1598 Callee = DAG.getRegister(Opc, getPointerTy());
1599 // Add register as live out.
1600 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1603 // Returns a chain & a flag for retval copy to use.
1604 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1605 SmallVector<SDOperand, 8> Ops;
1608 Ops.push_back(Chain);
1609 Ops.push_back(DAG.getIntPtrConstant(NumBytes));
1610 Ops.push_back(DAG.getIntPtrConstant(0));
1612 Ops.push_back(InFlag);
1613 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1614 InFlag = Chain.getValue(1);
1616 // Returns a chain & a flag for retval copy to use.
1617 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1621 Ops.push_back(Chain);
1622 Ops.push_back(Callee);
1625 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1627 // Add an implicit use GOT pointer in EBX.
1628 if (!IsTailCall && !Is64Bit &&
1629 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1630 Subtarget->isPICStyleGOT())
1631 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1633 // Add argument registers to the end of the list so that they are known live
1635 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1636 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1637 RegsToPass[i].second.getValueType()));
1640 Ops.push_back(InFlag);
1643 assert(InFlag.Val &&
1644 "Flag must be set. Depend on flag being set in LowerRET");
1645 Chain = DAG.getNode(X86ISD::TAILCALL,
1646 Op.Val->getVTList(), &Ops[0], Ops.size());
1648 return SDOperand(Chain.Val, Op.ResNo);
1651 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1652 InFlag = Chain.getValue(1);
1654 // Create the CALLSEQ_END node.
1655 unsigned NumBytesForCalleeToPush;
1656 if (IsCalleePop(Op))
1657 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1658 else if (!Is64Bit && IsStructRet)
1659 // If this is is a call to a struct-return function, the callee
1660 // pops the hidden struct pointer, so we have to push it back.
1661 // This is common for Darwin/X86, Linux & Mingw32 targets.
1662 NumBytesForCalleeToPush = 4;
1664 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1666 // Returns a flag for retval copy to use.
1667 Chain = DAG.getCALLSEQ_END(Chain,
1668 DAG.getIntPtrConstant(NumBytes),
1669 DAG.getIntPtrConstant(NumBytesForCalleeToPush),
1671 InFlag = Chain.getValue(1);
1673 // Handle result values, copying them out of physregs into vregs that we
1675 switch (SRetMethod) {
1677 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1679 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1681 MVT::i64, DAG), Op.ResNo);
1683 return SDOperand(LowerCallResultToTwo64BitRegs(Chain, InFlag, Op.Val,
1684 X86::XMM0, X86::XMM1,
1685 MVT::f64, DAG), Op.ResNo);
1687 return SDOperand(LowerCallResultToTwoX87Regs(Chain, InFlag, Op.Val, DAG),
1693 //===----------------------------------------------------------------------===//
1694 // Fast Calling Convention (tail call) implementation
1695 //===----------------------------------------------------------------------===//
1697 // Like std call, callee cleans arguments, convention except that ECX is
1698 // reserved for storing the tail called function address. Only 2 registers are
1699 // free for argument passing (inreg). Tail call optimization is performed
1701 // * tailcallopt is enabled
1702 // * caller/callee are fastcc
1703 // * elf/pic is disabled OR
1704 // * elf/pic enabled + callee is in module + callee has
1705 // visibility protected or hidden
1706 // To keep the stack aligned according to platform abi the function
1707 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1708 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1709 // If a tail called function callee has more arguments than the caller the
1710 // caller needs to make sure that there is room to move the RETADDR to. This is
1711 // achieved by reserving an area the size of the argument delta right after the
1712 // original REtADDR, but before the saved framepointer or the spilled registers
1713 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1725 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1726 /// for a 16 byte align requirement.
1727 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1728 SelectionDAG& DAG) {
1729 if (PerformTailCallOpt) {
1730 MachineFunction &MF = DAG.getMachineFunction();
1731 const TargetMachine &TM = MF.getTarget();
1732 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1733 unsigned StackAlignment = TFI.getStackAlignment();
1734 uint64_t AlignMask = StackAlignment - 1;
1735 int64_t Offset = StackSize;
1736 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1737 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1738 // Number smaller than 12 so just add the difference.
1739 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1741 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1742 Offset = ((~AlignMask) & Offset) + StackAlignment +
1743 (StackAlignment-SlotSize);
1750 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1751 /// following the call is a return. A function is eligible if caller/callee
1752 /// calling conventions match, currently only fastcc supports tail calls, and
1753 /// the function CALL is immediatly followed by a RET.
1754 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1756 SelectionDAG& DAG) const {
1757 if (!PerformTailCallOpt)
1760 // Check whether CALL node immediatly preceeds the RET node and whether the
1761 // return uses the result of the node or is a void return.
1762 unsigned NumOps = Ret.getNumOperands();
1764 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1765 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1767 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1768 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1769 MachineFunction &MF = DAG.getMachineFunction();
1770 unsigned CallerCC = MF.getFunction()->getCallingConv();
1771 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1772 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1773 SDOperand Callee = Call.getOperand(4);
1774 // On elf/pic %ebx needs to be livein.
1775 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1776 !Subtarget->isPICStyleGOT())
1779 // Can only do local tail calls with PIC.
1780 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1781 return G->getGlobal()->hasHiddenVisibility()
1782 || G->getGlobal()->hasProtectedVisibility();
1789 //===----------------------------------------------------------------------===//
1790 // Other Lowering Hooks
1791 //===----------------------------------------------------------------------===//
1794 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1795 MachineFunction &MF = DAG.getMachineFunction();
1796 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1797 int ReturnAddrIndex = FuncInfo->getRAIndex();
1799 if (ReturnAddrIndex == 0) {
1800 // Set up a frame object for the return address.
1801 if (Subtarget->is64Bit())
1802 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1804 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1806 FuncInfo->setRAIndex(ReturnAddrIndex);
1809 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1814 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1815 /// specific condition code. It returns a false if it cannot do a direct
1816 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1818 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1819 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1820 SelectionDAG &DAG) {
1821 X86CC = X86::COND_INVALID;
1823 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1824 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1825 // X > -1 -> X == 0, jump !sign.
1826 RHS = DAG.getConstant(0, RHS.getValueType());
1827 X86CC = X86::COND_NS;
1829 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1830 // X < 0 -> X == 0, jump on sign.
1831 X86CC = X86::COND_S;
1833 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
1835 RHS = DAG.getConstant(0, RHS.getValueType());
1836 X86CC = X86::COND_LE;
1841 switch (SetCCOpcode) {
1843 case ISD::SETEQ: X86CC = X86::COND_E; break;
1844 case ISD::SETGT: X86CC = X86::COND_G; break;
1845 case ISD::SETGE: X86CC = X86::COND_GE; break;
1846 case ISD::SETLT: X86CC = X86::COND_L; break;
1847 case ISD::SETLE: X86CC = X86::COND_LE; break;
1848 case ISD::SETNE: X86CC = X86::COND_NE; break;
1849 case ISD::SETULT: X86CC = X86::COND_B; break;
1850 case ISD::SETUGT: X86CC = X86::COND_A; break;
1851 case ISD::SETULE: X86CC = X86::COND_BE; break;
1852 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1855 // On a floating point condition, the flags are set as follows:
1857 // 0 | 0 | 0 | X > Y
1858 // 0 | 0 | 1 | X < Y
1859 // 1 | 0 | 0 | X == Y
1860 // 1 | 1 | 1 | unordered
1862 switch (SetCCOpcode) {
1865 case ISD::SETEQ: X86CC = X86::COND_E; break;
1866 case ISD::SETOLT: Flip = true; // Fallthrough
1868 case ISD::SETGT: X86CC = X86::COND_A; break;
1869 case ISD::SETOLE: Flip = true; // Fallthrough
1871 case ISD::SETGE: X86CC = X86::COND_AE; break;
1872 case ISD::SETUGT: Flip = true; // Fallthrough
1874 case ISD::SETLT: X86CC = X86::COND_B; break;
1875 case ISD::SETUGE: Flip = true; // Fallthrough
1877 case ISD::SETLE: X86CC = X86::COND_BE; break;
1879 case ISD::SETNE: X86CC = X86::COND_NE; break;
1880 case ISD::SETUO: X86CC = X86::COND_P; break;
1881 case ISD::SETO: X86CC = X86::COND_NP; break;
1884 std::swap(LHS, RHS);
1887 return X86CC != X86::COND_INVALID;
1890 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1891 /// code. Current x86 isa includes the following FP cmov instructions:
1892 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1893 static bool hasFPCMov(unsigned X86CC) {
1909 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1910 /// true if Op is undef or if its value falls within the specified range (L, H].
1911 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1912 if (Op.getOpcode() == ISD::UNDEF)
1915 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1916 return (Val >= Low && Val < Hi);
1919 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1920 /// true if Op is undef or if its value equal to the specified value.
1921 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1922 if (Op.getOpcode() == ISD::UNDEF)
1924 return cast<ConstantSDNode>(Op)->getValue() == Val;
1927 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1928 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1929 bool X86::isPSHUFDMask(SDNode *N) {
1930 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1932 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
1935 // Check if the value doesn't reference the second vector.
1936 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1937 SDOperand Arg = N->getOperand(i);
1938 if (Arg.getOpcode() == ISD::UNDEF) continue;
1939 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1940 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
1947 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1948 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1949 bool X86::isPSHUFHWMask(SDNode *N) {
1950 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1952 if (N->getNumOperands() != 8)
1955 // Lower quadword copied in order.
1956 for (unsigned i = 0; i != 4; ++i) {
1957 SDOperand Arg = N->getOperand(i);
1958 if (Arg.getOpcode() == ISD::UNDEF) continue;
1959 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1960 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1964 // Upper quadword shuffled.
1965 for (unsigned i = 4; i != 8; ++i) {
1966 SDOperand Arg = N->getOperand(i);
1967 if (Arg.getOpcode() == ISD::UNDEF) continue;
1968 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1969 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1970 if (Val < 4 || Val > 7)
1977 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1978 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1979 bool X86::isPSHUFLWMask(SDNode *N) {
1980 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1982 if (N->getNumOperands() != 8)
1985 // Upper quadword copied in order.
1986 for (unsigned i = 4; i != 8; ++i)
1987 if (!isUndefOrEqual(N->getOperand(i), i))
1990 // Lower quadword shuffled.
1991 for (unsigned i = 0; i != 4; ++i)
1992 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1998 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1999 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2000 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2001 if (NumElems != 2 && NumElems != 4) return false;
2003 unsigned Half = NumElems / 2;
2004 for (unsigned i = 0; i < Half; ++i)
2005 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2007 for (unsigned i = Half; i < NumElems; ++i)
2008 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2014 bool X86::isSHUFPMask(SDNode *N) {
2015 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2016 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2019 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2020 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2021 /// half elements to come from vector 1 (which would equal the dest.) and
2022 /// the upper half to come from vector 2.
2023 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2024 if (NumOps != 2 && NumOps != 4) return false;
2026 unsigned Half = NumOps / 2;
2027 for (unsigned i = 0; i < Half; ++i)
2028 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2030 for (unsigned i = Half; i < NumOps; ++i)
2031 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2036 static bool isCommutedSHUFP(SDNode *N) {
2037 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2038 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2041 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2042 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2043 bool X86::isMOVHLPSMask(SDNode *N) {
2044 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2046 if (N->getNumOperands() != 4)
2049 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2050 return isUndefOrEqual(N->getOperand(0), 6) &&
2051 isUndefOrEqual(N->getOperand(1), 7) &&
2052 isUndefOrEqual(N->getOperand(2), 2) &&
2053 isUndefOrEqual(N->getOperand(3), 3);
2056 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2057 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2059 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2060 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2062 if (N->getNumOperands() != 4)
2065 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2066 return isUndefOrEqual(N->getOperand(0), 2) &&
2067 isUndefOrEqual(N->getOperand(1), 3) &&
2068 isUndefOrEqual(N->getOperand(2), 2) &&
2069 isUndefOrEqual(N->getOperand(3), 3);
2072 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2073 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2074 bool X86::isMOVLPMask(SDNode *N) {
2075 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2077 unsigned NumElems = N->getNumOperands();
2078 if (NumElems != 2 && NumElems != 4)
2081 for (unsigned i = 0; i < NumElems/2; ++i)
2082 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2085 for (unsigned i = NumElems/2; i < NumElems; ++i)
2086 if (!isUndefOrEqual(N->getOperand(i), i))
2092 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2093 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2095 bool X86::isMOVHPMask(SDNode *N) {
2096 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2098 unsigned NumElems = N->getNumOperands();
2099 if (NumElems != 2 && NumElems != 4)
2102 for (unsigned i = 0; i < NumElems/2; ++i)
2103 if (!isUndefOrEqual(N->getOperand(i), i))
2106 for (unsigned i = 0; i < NumElems/2; ++i) {
2107 SDOperand Arg = N->getOperand(i + NumElems/2);
2108 if (!isUndefOrEqual(Arg, i + NumElems))
2115 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2116 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2117 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2118 bool V2IsSplat = false) {
2119 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2122 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2123 SDOperand BitI = Elts[i];
2124 SDOperand BitI1 = Elts[i+1];
2125 if (!isUndefOrEqual(BitI, j))
2128 if (isUndefOrEqual(BitI1, NumElts))
2131 if (!isUndefOrEqual(BitI1, j + NumElts))
2139 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2140 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2141 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2144 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2145 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2146 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2147 bool V2IsSplat = false) {
2148 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2151 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2152 SDOperand BitI = Elts[i];
2153 SDOperand BitI1 = Elts[i+1];
2154 if (!isUndefOrEqual(BitI, j + NumElts/2))
2157 if (isUndefOrEqual(BitI1, NumElts))
2160 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2168 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2169 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2170 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2173 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2174 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2176 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2177 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2179 unsigned NumElems = N->getNumOperands();
2180 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2183 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2184 SDOperand BitI = N->getOperand(i);
2185 SDOperand BitI1 = N->getOperand(i+1);
2187 if (!isUndefOrEqual(BitI, j))
2189 if (!isUndefOrEqual(BitI1, j))
2196 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2197 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2199 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2200 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2202 unsigned NumElems = N->getNumOperands();
2203 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2206 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2207 SDOperand BitI = N->getOperand(i);
2208 SDOperand BitI1 = N->getOperand(i + 1);
2210 if (!isUndefOrEqual(BitI, j))
2212 if (!isUndefOrEqual(BitI1, j))
2219 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2220 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2221 /// MOVSD, and MOVD, i.e. setting the lowest element.
2222 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2223 if (NumElts != 2 && NumElts != 4)
2226 if (!isUndefOrEqual(Elts[0], NumElts))
2229 for (unsigned i = 1; i < NumElts; ++i) {
2230 if (!isUndefOrEqual(Elts[i], i))
2237 bool X86::isMOVLMask(SDNode *N) {
2238 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2239 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2242 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2243 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2244 /// element of vector 2 and the other elements to come from vector 1 in order.
2245 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2246 bool V2IsSplat = false,
2247 bool V2IsUndef = false) {
2248 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2251 if (!isUndefOrEqual(Ops[0], 0))
2254 for (unsigned i = 1; i < NumOps; ++i) {
2255 SDOperand Arg = Ops[i];
2256 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2257 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2258 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2265 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2266 bool V2IsUndef = false) {
2267 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2268 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2269 V2IsSplat, V2IsUndef);
2272 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2273 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2274 bool X86::isMOVSHDUPMask(SDNode *N) {
2275 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2277 if (N->getNumOperands() != 4)
2280 // Expect 1, 1, 3, 3
2281 for (unsigned i = 0; i < 2; ++i) {
2282 SDOperand Arg = N->getOperand(i);
2283 if (Arg.getOpcode() == ISD::UNDEF) continue;
2284 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2285 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2286 if (Val != 1) return false;
2290 for (unsigned i = 2; i < 4; ++i) {
2291 SDOperand Arg = N->getOperand(i);
2292 if (Arg.getOpcode() == ISD::UNDEF) continue;
2293 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2294 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2295 if (Val != 3) return false;
2299 // Don't use movshdup if it can be done with a shufps.
2303 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2304 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2305 bool X86::isMOVSLDUPMask(SDNode *N) {
2306 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2308 if (N->getNumOperands() != 4)
2311 // Expect 0, 0, 2, 2
2312 for (unsigned i = 0; i < 2; ++i) {
2313 SDOperand Arg = N->getOperand(i);
2314 if (Arg.getOpcode() == ISD::UNDEF) continue;
2315 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2316 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2317 if (Val != 0) return false;
2321 for (unsigned i = 2; i < 4; ++i) {
2322 SDOperand Arg = N->getOperand(i);
2323 if (Arg.getOpcode() == ISD::UNDEF) continue;
2324 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2325 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2326 if (Val != 2) return false;
2330 // Don't use movshdup if it can be done with a shufps.
2334 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2335 /// specifies a identity operation on the LHS or RHS.
2336 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2337 unsigned NumElems = N->getNumOperands();
2338 for (unsigned i = 0; i < NumElems; ++i)
2339 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2344 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2345 /// a splat of a single element.
2346 static bool isSplatMask(SDNode *N) {
2347 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2349 // This is a splat operation if each element of the permute is the same, and
2350 // if the value doesn't reference the second vector.
2351 unsigned NumElems = N->getNumOperands();
2352 SDOperand ElementBase;
2354 for (; i != NumElems; ++i) {
2355 SDOperand Elt = N->getOperand(i);
2356 if (isa<ConstantSDNode>(Elt)) {
2362 if (!ElementBase.Val)
2365 for (; i != NumElems; ++i) {
2366 SDOperand Arg = N->getOperand(i);
2367 if (Arg.getOpcode() == ISD::UNDEF) continue;
2368 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2369 if (Arg != ElementBase) return false;
2372 // Make sure it is a splat of the first vector operand.
2373 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2376 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2377 /// a splat of a single element and it's a 2 or 4 element mask.
2378 bool X86::isSplatMask(SDNode *N) {
2379 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2381 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2382 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2384 return ::isSplatMask(N);
2387 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2388 /// specifies a splat of zero element.
2389 bool X86::isSplatLoMask(SDNode *N) {
2390 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2392 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2393 if (!isUndefOrEqual(N->getOperand(i), 0))
2398 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2399 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2401 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2402 unsigned NumOperands = N->getNumOperands();
2403 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2405 for (unsigned i = 0; i < NumOperands; ++i) {
2407 SDOperand Arg = N->getOperand(NumOperands-i-1);
2408 if (Arg.getOpcode() != ISD::UNDEF)
2409 Val = cast<ConstantSDNode>(Arg)->getValue();
2410 if (Val >= NumOperands) Val -= NumOperands;
2412 if (i != NumOperands - 1)
2419 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2420 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2422 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2424 // 8 nodes, but we only care about the last 4.
2425 for (unsigned i = 7; i >= 4; --i) {
2427 SDOperand Arg = N->getOperand(i);
2428 if (Arg.getOpcode() != ISD::UNDEF)
2429 Val = cast<ConstantSDNode>(Arg)->getValue();
2438 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2439 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2441 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2443 // 8 nodes, but we only care about the first 4.
2444 for (int i = 3; i >= 0; --i) {
2446 SDOperand Arg = N->getOperand(i);
2447 if (Arg.getOpcode() != ISD::UNDEF)
2448 Val = cast<ConstantSDNode>(Arg)->getValue();
2457 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2458 /// specifies a 8 element shuffle that can be broken into a pair of
2459 /// PSHUFHW and PSHUFLW.
2460 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2461 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2463 if (N->getNumOperands() != 8)
2466 // Lower quadword shuffled.
2467 for (unsigned i = 0; i != 4; ++i) {
2468 SDOperand Arg = N->getOperand(i);
2469 if (Arg.getOpcode() == ISD::UNDEF) continue;
2470 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2471 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2476 // Upper quadword shuffled.
2477 for (unsigned i = 4; i != 8; ++i) {
2478 SDOperand Arg = N->getOperand(i);
2479 if (Arg.getOpcode() == ISD::UNDEF) continue;
2480 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2481 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2482 if (Val < 4 || Val > 7)
2489 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2490 /// values in ther permute mask.
2491 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2492 SDOperand &V2, SDOperand &Mask,
2493 SelectionDAG &DAG) {
2494 MVT::ValueType VT = Op.getValueType();
2495 MVT::ValueType MaskVT = Mask.getValueType();
2496 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2497 unsigned NumElems = Mask.getNumOperands();
2498 SmallVector<SDOperand, 8> MaskVec;
2500 for (unsigned i = 0; i != NumElems; ++i) {
2501 SDOperand Arg = Mask.getOperand(i);
2502 if (Arg.getOpcode() == ISD::UNDEF) {
2503 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2506 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2507 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2509 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2511 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2515 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2516 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2519 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2520 /// the two vector operands have swapped position.
2522 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2523 MVT::ValueType MaskVT = Mask.getValueType();
2524 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2525 unsigned NumElems = Mask.getNumOperands();
2526 SmallVector<SDOperand, 8> MaskVec;
2527 for (unsigned i = 0; i != NumElems; ++i) {
2528 SDOperand Arg = Mask.getOperand(i);
2529 if (Arg.getOpcode() == ISD::UNDEF) {
2530 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2533 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2534 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2536 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2538 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2540 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2544 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2545 /// match movhlps. The lower half elements should come from upper half of
2546 /// V1 (and in order), and the upper half elements should come from the upper
2547 /// half of V2 (and in order).
2548 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2549 unsigned NumElems = Mask->getNumOperands();
2552 for (unsigned i = 0, e = 2; i != e; ++i)
2553 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2555 for (unsigned i = 2; i != 4; ++i)
2556 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2561 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2562 /// is promoted to a vector.
2563 static inline bool isScalarLoadToVector(SDNode *N) {
2564 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2565 N = N->getOperand(0).Val;
2566 return ISD::isNON_EXTLoad(N);
2571 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2572 /// match movlp{s|d}. The lower half elements should come from lower half of
2573 /// V1 (and in order), and the upper half elements should come from the upper
2574 /// half of V2 (and in order). And since V1 will become the source of the
2575 /// MOVLP, it must be either a vector load or a scalar load to vector.
2576 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2577 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2579 // Is V2 is a vector load, don't do this transformation. We will try to use
2580 // load folding shufps op.
2581 if (ISD::isNON_EXTLoad(V2))
2584 unsigned NumElems = Mask->getNumOperands();
2585 if (NumElems != 2 && NumElems != 4)
2587 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2588 if (!isUndefOrEqual(Mask->getOperand(i), i))
2590 for (unsigned i = NumElems/2; i != NumElems; ++i)
2591 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2596 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2598 static bool isSplatVector(SDNode *N) {
2599 if (N->getOpcode() != ISD::BUILD_VECTOR)
2602 SDOperand SplatValue = N->getOperand(0);
2603 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2604 if (N->getOperand(i) != SplatValue)
2609 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2611 static bool isUndefShuffle(SDNode *N) {
2612 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2615 SDOperand V1 = N->getOperand(0);
2616 SDOperand V2 = N->getOperand(1);
2617 SDOperand Mask = N->getOperand(2);
2618 unsigned NumElems = Mask.getNumOperands();
2619 for (unsigned i = 0; i != NumElems; ++i) {
2620 SDOperand Arg = Mask.getOperand(i);
2621 if (Arg.getOpcode() != ISD::UNDEF) {
2622 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2623 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2625 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2632 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2634 static inline bool isZeroNode(SDOperand Elt) {
2635 return ((isa<ConstantSDNode>(Elt) &&
2636 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2637 (isa<ConstantFPSDNode>(Elt) &&
2638 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2641 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2642 /// to an zero vector.
2643 static bool isZeroShuffle(SDNode *N) {
2644 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2647 SDOperand V1 = N->getOperand(0);
2648 SDOperand V2 = N->getOperand(1);
2649 SDOperand Mask = N->getOperand(2);
2650 unsigned NumElems = Mask.getNumOperands();
2651 for (unsigned i = 0; i != NumElems; ++i) {
2652 SDOperand Arg = Mask.getOperand(i);
2653 if (Arg.getOpcode() == ISD::UNDEF)
2656 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2657 if (Idx < NumElems) {
2658 unsigned Opc = V1.Val->getOpcode();
2659 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2661 if (Opc != ISD::BUILD_VECTOR ||
2662 !isZeroNode(V1.Val->getOperand(Idx)))
2664 } else if (Idx >= NumElems) {
2665 unsigned Opc = V2.Val->getOpcode();
2666 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2668 if (Opc != ISD::BUILD_VECTOR ||
2669 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2676 /// getZeroVector - Returns a vector of specified type with all zero elements.
2678 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2679 assert(MVT::isVector(VT) && "Expected a vector type");
2681 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2682 // type. This ensures they get CSE'd.
2683 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2685 if (MVT::getSizeInBits(VT) == 64) // MMX
2686 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2688 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2689 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2692 /// getOnesVector - Returns a vector of specified type with all bits set.
2694 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2695 assert(MVT::isVector(VT) && "Expected a vector type");
2697 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2698 // type. This ensures they get CSE'd.
2699 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2701 if (MVT::getSizeInBits(VT) == 64) // MMX
2702 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2704 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2705 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2709 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2710 /// that point to V2 points to its first element.
2711 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2712 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2714 bool Changed = false;
2715 SmallVector<SDOperand, 8> MaskVec;
2716 unsigned NumElems = Mask.getNumOperands();
2717 for (unsigned i = 0; i != NumElems; ++i) {
2718 SDOperand Arg = Mask.getOperand(i);
2719 if (Arg.getOpcode() != ISD::UNDEF) {
2720 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2721 if (Val > NumElems) {
2722 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2726 MaskVec.push_back(Arg);
2730 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2731 &MaskVec[0], MaskVec.size());
2735 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2736 /// operation of specified width.
2737 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2738 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2739 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2741 SmallVector<SDOperand, 8> MaskVec;
2742 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2743 for (unsigned i = 1; i != NumElems; ++i)
2744 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2745 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2748 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2749 /// of specified width.
2750 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2751 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2752 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2753 SmallVector<SDOperand, 8> MaskVec;
2754 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2755 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2756 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2758 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2761 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2762 /// of specified width.
2763 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2764 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2765 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2766 unsigned Half = NumElems/2;
2767 SmallVector<SDOperand, 8> MaskVec;
2768 for (unsigned i = 0; i != Half; ++i) {
2769 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2770 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2772 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2775 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2777 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2778 SDOperand V1 = Op.getOperand(0);
2779 SDOperand Mask = Op.getOperand(2);
2780 MVT::ValueType VT = Op.getValueType();
2781 unsigned NumElems = Mask.getNumOperands();
2782 Mask = getUnpacklMask(NumElems, DAG);
2783 while (NumElems != 4) {
2784 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2787 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2789 Mask = getZeroVector(MVT::v4i32, DAG);
2790 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2791 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2792 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2795 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2796 /// vector of zero or undef vector. This produces a shuffle where the low
2797 /// element of V2 is swizzled into the zero/undef vector, landing at element
2798 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2799 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2800 unsigned NumElems, unsigned Idx,
2801 bool isZero, SelectionDAG &DAG) {
2802 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2803 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2804 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2805 SmallVector<SDOperand, 16> MaskVec;
2806 for (unsigned i = 0; i != NumElems; ++i)
2807 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
2808 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
2810 MaskVec.push_back(DAG.getConstant(i, EVT));
2811 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2812 &MaskVec[0], MaskVec.size());
2813 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2816 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2818 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2819 unsigned NumNonZero, unsigned NumZero,
2820 SelectionDAG &DAG, TargetLowering &TLI) {
2826 for (unsigned i = 0; i < 16; ++i) {
2827 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2828 if (ThisIsNonZero && First) {
2830 V = getZeroVector(MVT::v8i16, DAG);
2832 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2837 SDOperand ThisElt(0, 0), LastElt(0, 0);
2838 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2839 if (LastIsNonZero) {
2840 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2842 if (ThisIsNonZero) {
2843 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2844 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2845 ThisElt, DAG.getConstant(8, MVT::i8));
2847 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2852 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2853 DAG.getIntPtrConstant(i/2));
2857 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2860 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2862 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2863 unsigned NumNonZero, unsigned NumZero,
2864 SelectionDAG &DAG, TargetLowering &TLI) {
2870 for (unsigned i = 0; i < 8; ++i) {
2871 bool isNonZero = (NonZeros & (1 << i)) != 0;
2875 V = getZeroVector(MVT::v8i16, DAG);
2877 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2880 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2881 DAG.getIntPtrConstant(i));
2889 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2890 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
2891 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
2892 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
2893 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
2894 // eliminated on x86-32 hosts.
2895 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
2898 if (ISD::isBuildVectorAllOnes(Op.Val))
2899 return getOnesVector(Op.getValueType(), DAG);
2900 return getZeroVector(Op.getValueType(), DAG);
2903 MVT::ValueType VT = Op.getValueType();
2904 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2905 unsigned EVTBits = MVT::getSizeInBits(EVT);
2907 unsigned NumElems = Op.getNumOperands();
2908 unsigned NumZero = 0;
2909 unsigned NumNonZero = 0;
2910 unsigned NonZeros = 0;
2911 bool HasNonImms = false;
2912 SmallSet<SDOperand, 8> Values;
2913 for (unsigned i = 0; i < NumElems; ++i) {
2914 SDOperand Elt = Op.getOperand(i);
2915 if (Elt.getOpcode() == ISD::UNDEF)
2918 if (Elt.getOpcode() != ISD::Constant &&
2919 Elt.getOpcode() != ISD::ConstantFP)
2921 if (isZeroNode(Elt))
2924 NonZeros |= (1 << i);
2929 if (NumNonZero == 0) {
2930 // All undef vector. Return an UNDEF. All zero vectors were handled above.
2931 return DAG.getNode(ISD::UNDEF, VT);
2934 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2935 if (Values.size() == 1)
2938 // Special case for single non-zero element.
2939 if (NumNonZero == 1 && NumElems <= 4) {
2940 unsigned Idx = CountTrailingZeros_32(NonZeros);
2941 SDOperand Item = Op.getOperand(Idx);
2942 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2944 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2945 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2947 else if (!HasNonImms) // Otherwise, it's better to do a constpool load.
2950 if (EVTBits == 32) {
2951 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2952 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2954 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2955 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2956 SmallVector<SDOperand, 8> MaskVec;
2957 for (unsigned i = 0; i < NumElems; i++)
2958 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2959 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2960 &MaskVec[0], MaskVec.size());
2961 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2962 DAG.getNode(ISD::UNDEF, VT), Mask);
2966 // A vector full of immediates; various special cases are already
2967 // handled, so this is best done with a single constant-pool load.
2971 // Let legalizer expand 2-wide build_vectors.
2975 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2976 if (EVTBits == 8 && NumElems == 16) {
2977 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2979 if (V.Val) return V;
2982 if (EVTBits == 16 && NumElems == 8) {
2983 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2985 if (V.Val) return V;
2988 // If element VT is == 32 bits, turn it into a number of shuffles.
2989 SmallVector<SDOperand, 8> V;
2991 if (NumElems == 4 && NumZero > 0) {
2992 for (unsigned i = 0; i < 4; ++i) {
2993 bool isZero = !(NonZeros & (1 << i));
2995 V[i] = getZeroVector(VT, DAG);
2997 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3000 for (unsigned i = 0; i < 2; ++i) {
3001 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3004 V[i] = V[i*2]; // Must be a zero vector.
3007 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3008 getMOVLMask(NumElems, DAG));
3011 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3012 getMOVLMask(NumElems, DAG));
3015 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3016 getUnpacklMask(NumElems, DAG));
3021 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3022 // clears the upper bits.
3023 // FIXME: we can do the same for v4f32 case when we know both parts of
3024 // the lower half come from scalar_to_vector (loadf32). We should do
3025 // that in post legalizer dag combiner with target specific hooks.
3026 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3028 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3029 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3030 SmallVector<SDOperand, 8> MaskVec;
3031 bool Reverse = (NonZeros & 0x3) == 2;
3032 for (unsigned i = 0; i < 2; ++i)
3034 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3036 MaskVec.push_back(DAG.getConstant(i, EVT));
3037 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3038 for (unsigned i = 0; i < 2; ++i)
3040 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3042 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3043 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3044 &MaskVec[0], MaskVec.size());
3045 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3048 if (Values.size() > 2) {
3049 // Expand into a number of unpckl*.
3051 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3052 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3053 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3054 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3055 for (unsigned i = 0; i < NumElems; ++i)
3056 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3058 while (NumElems != 0) {
3059 for (unsigned i = 0; i < NumElems; ++i)
3060 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3071 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3072 SDOperand PermMask, SelectionDAG &DAG,
3073 TargetLowering &TLI) {
3075 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3076 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3077 MVT::ValueType PtrVT = TLI.getPointerTy();
3078 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3079 PermMask.Val->op_end());
3081 // First record which half of which vector the low elements come from.
3082 SmallVector<unsigned, 4> LowQuad(4);
3083 for (unsigned i = 0; i < 4; ++i) {
3084 SDOperand Elt = MaskElts[i];
3085 if (Elt.getOpcode() == ISD::UNDEF)
3087 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3088 int QuadIdx = EltIdx / 4;
3091 int BestLowQuad = -1;
3092 unsigned MaxQuad = 1;
3093 for (unsigned i = 0; i < 4; ++i) {
3094 if (LowQuad[i] > MaxQuad) {
3096 MaxQuad = LowQuad[i];
3100 // Record which half of which vector the high elements come from.
3101 SmallVector<unsigned, 4> HighQuad(4);
3102 for (unsigned i = 4; i < 8; ++i) {
3103 SDOperand Elt = MaskElts[i];
3104 if (Elt.getOpcode() == ISD::UNDEF)
3106 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3107 int QuadIdx = EltIdx / 4;
3108 ++HighQuad[QuadIdx];
3110 int BestHighQuad = -1;
3112 for (unsigned i = 0; i < 4; ++i) {
3113 if (HighQuad[i] > MaxQuad) {
3115 MaxQuad = HighQuad[i];
3119 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3120 if (BestLowQuad != -1 || BestHighQuad != -1) {
3121 // First sort the 4 chunks in order using shufpd.
3122 SmallVector<SDOperand, 8> MaskVec;
3123 if (BestLowQuad != -1)
3124 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3126 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3127 if (BestHighQuad != -1)
3128 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3130 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3131 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3132 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3133 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3134 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3135 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3137 // Now sort high and low parts separately.
3138 BitVector InOrder(8);
3139 if (BestLowQuad != -1) {
3140 // Sort lower half in order using PSHUFLW.
3142 bool AnyOutOrder = false;
3143 for (unsigned i = 0; i != 4; ++i) {
3144 SDOperand Elt = MaskElts[i];
3145 if (Elt.getOpcode() == ISD::UNDEF) {
3146 MaskVec.push_back(Elt);
3149 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3152 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3153 // If this element is in the right place after this shuffle, then
3155 if ((int)(EltIdx / 4) == BestLowQuad)
3160 for (unsigned i = 4; i != 8; ++i)
3161 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3162 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3163 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3167 if (BestHighQuad != -1) {
3168 // Sort high half in order using PSHUFHW if possible.
3170 for (unsigned i = 0; i != 4; ++i)
3171 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3172 bool AnyOutOrder = false;
3173 for (unsigned i = 4; i != 8; ++i) {
3174 SDOperand Elt = MaskElts[i];
3175 if (Elt.getOpcode() == ISD::UNDEF) {
3176 MaskVec.push_back(Elt);
3179 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3182 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3183 // If this element is in the right place after this shuffle, then
3185 if ((int)(EltIdx / 4) == BestHighQuad)
3190 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3191 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3195 // The other elements are put in the right place using pextrw and pinsrw.
3196 for (unsigned i = 0; i != 8; ++i) {
3199 SDOperand Elt = MaskElts[i];
3200 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3203 SDOperand ExtOp = (EltIdx < 8)
3204 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3205 DAG.getConstant(EltIdx, PtrVT))
3206 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3207 DAG.getConstant(EltIdx - 8, PtrVT));
3208 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3209 DAG.getConstant(i, PtrVT));
3214 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3215 ///as few as possible.
3216 // First, let's find out how many elements are already in the right order.
3217 unsigned V1InOrder = 0;
3218 unsigned V1FromV1 = 0;
3219 unsigned V2InOrder = 0;
3220 unsigned V2FromV2 = 0;
3221 SmallVector<SDOperand, 8> V1Elts;
3222 SmallVector<SDOperand, 8> V2Elts;
3223 for (unsigned i = 0; i < 8; ++i) {
3224 SDOperand Elt = MaskElts[i];
3225 if (Elt.getOpcode() == ISD::UNDEF) {
3226 V1Elts.push_back(Elt);
3227 V2Elts.push_back(Elt);
3232 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3234 V1Elts.push_back(Elt);
3235 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3237 } else if (EltIdx == i+8) {
3238 V1Elts.push_back(Elt);
3239 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3241 } else if (EltIdx < 8) {
3242 V1Elts.push_back(Elt);
3245 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3250 if (V2InOrder > V1InOrder) {
3251 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3253 std::swap(V1Elts, V2Elts);
3254 std::swap(V1FromV1, V2FromV2);
3257 if ((V1FromV1 + V1InOrder) != 8) {
3258 // Some elements are from V2.
3260 // If there are elements that are from V1 but out of place,
3261 // then first sort them in place
3262 SmallVector<SDOperand, 8> MaskVec;
3263 for (unsigned i = 0; i < 8; ++i) {
3264 SDOperand Elt = V1Elts[i];
3265 if (Elt.getOpcode() == ISD::UNDEF) {
3266 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3269 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3271 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3273 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3275 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3276 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3280 for (unsigned i = 0; i < 8; ++i) {
3281 SDOperand Elt = V1Elts[i];
3282 if (Elt.getOpcode() == ISD::UNDEF)
3284 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3287 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3288 DAG.getConstant(EltIdx - 8, PtrVT));
3289 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3290 DAG.getConstant(i, PtrVT));
3294 // All elements are from V1.
3296 for (unsigned i = 0; i < 8; ++i) {
3297 SDOperand Elt = V1Elts[i];
3298 if (Elt.getOpcode() == ISD::UNDEF)
3300 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3301 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3302 DAG.getConstant(EltIdx, PtrVT));
3303 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3304 DAG.getConstant(i, PtrVT));
3310 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3311 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3312 /// done when every pair / quad of shuffle mask elements point to elements in
3313 /// the right sequence. e.g.
3314 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3316 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3318 SDOperand PermMask, SelectionDAG &DAG,
3319 TargetLowering &TLI) {
3320 unsigned NumElems = PermMask.getNumOperands();
3321 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3322 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3323 MVT::ValueType NewVT = MaskVT;
3325 case MVT::v4f32: NewVT = MVT::v2f64; break;
3326 case MVT::v4i32: NewVT = MVT::v2i64; break;
3327 case MVT::v8i16: NewVT = MVT::v4i32; break;
3328 case MVT::v16i8: NewVT = MVT::v4i32; break;
3329 default: assert(false && "Unexpected!");
3333 if (MVT::isInteger(VT))
3337 unsigned Scale = NumElems / NewWidth;
3338 SmallVector<SDOperand, 8> MaskVec;
3339 for (unsigned i = 0; i < NumElems; i += Scale) {
3340 unsigned StartIdx = ~0U;
3341 for (unsigned j = 0; j < Scale; ++j) {
3342 SDOperand Elt = PermMask.getOperand(i+j);
3343 if (Elt.getOpcode() == ISD::UNDEF)
3345 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3346 if (StartIdx == ~0U)
3347 StartIdx = EltIdx - (EltIdx % Scale);
3348 if (EltIdx != StartIdx + j)
3351 if (StartIdx == ~0U)
3352 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3354 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3357 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3358 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3359 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3360 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3361 &MaskVec[0], MaskVec.size()));
3365 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3366 SDOperand V1 = Op.getOperand(0);
3367 SDOperand V2 = Op.getOperand(1);
3368 SDOperand PermMask = Op.getOperand(2);
3369 MVT::ValueType VT = Op.getValueType();
3370 unsigned NumElems = PermMask.getNumOperands();
3371 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3372 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3373 bool V1IsSplat = false;
3374 bool V2IsSplat = false;
3376 if (isUndefShuffle(Op.Val))
3377 return DAG.getNode(ISD::UNDEF, VT);
3379 if (isZeroShuffle(Op.Val))
3380 return getZeroVector(VT, DAG);
3382 if (isIdentityMask(PermMask.Val))
3384 else if (isIdentityMask(PermMask.Val, true))
3387 if (isSplatMask(PermMask.Val)) {
3388 if (NumElems <= 4) return Op;
3389 // Promote it to a v4i32 splat.
3390 return PromoteSplat(Op, DAG);
3393 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3395 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3396 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3398 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3399 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3400 // FIXME: Figure out a cleaner way to do this.
3401 // Try to make use of movq to zero out the top part.
3402 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3403 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3405 SDOperand NewV1 = NewOp.getOperand(0);
3406 SDOperand NewV2 = NewOp.getOperand(1);
3407 SDOperand NewMask = NewOp.getOperand(2);
3408 if (isCommutedMOVL(NewMask.Val, true, false)) {
3409 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3410 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3411 NewV1, NewV2, getMOVLMask(2, DAG));
3412 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3415 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3416 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3417 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3418 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3422 if (X86::isMOVLMask(PermMask.Val))
3423 return (V1IsUndef) ? V2 : Op;
3425 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3426 X86::isMOVSLDUPMask(PermMask.Val) ||
3427 X86::isMOVHLPSMask(PermMask.Val) ||
3428 X86::isMOVHPMask(PermMask.Val) ||
3429 X86::isMOVLPMask(PermMask.Val))
3432 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3433 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3434 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3436 bool Commuted = false;
3437 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3438 // 1,1,1,1 -> v8i16 though.
3439 V1IsSplat = isSplatVector(V1.Val);
3440 V2IsSplat = isSplatVector(V2.Val);
3442 // Canonicalize the splat or undef, if present, to be on the RHS.
3443 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3444 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3445 std::swap(V1IsSplat, V2IsSplat);
3446 std::swap(V1IsUndef, V2IsUndef);
3450 // FIXME: Figure out a cleaner way to do this.
3451 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3452 if (V2IsUndef) return V1;
3453 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3455 // V2 is a splat, so the mask may be malformed. That is, it may point
3456 // to any V2 element. The instruction selectior won't like this. Get
3457 // a corrected mask and commute to form a proper MOVS{S|D}.
3458 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3459 if (NewMask.Val != PermMask.Val)
3460 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3465 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3466 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3467 X86::isUNPCKLMask(PermMask.Val) ||
3468 X86::isUNPCKHMask(PermMask.Val))
3472 // Normalize mask so all entries that point to V2 points to its first
3473 // element then try to match unpck{h|l} again. If match, return a
3474 // new vector_shuffle with the corrected mask.
3475 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3476 if (NewMask.Val != PermMask.Val) {
3477 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3478 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3479 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3480 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3481 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3482 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3487 // Normalize the node to match x86 shuffle ops if needed
3488 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3489 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3492 // Commute is back and try unpck* again.
3493 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3494 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3495 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3496 X86::isUNPCKLMask(PermMask.Val) ||
3497 X86::isUNPCKHMask(PermMask.Val))
3501 // If VT is integer, try PSHUF* first, then SHUFP*.
3502 if (MVT::isInteger(VT)) {
3503 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3504 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3505 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3506 X86::isPSHUFDMask(PermMask.Val)) ||
3507 X86::isPSHUFHWMask(PermMask.Val) ||
3508 X86::isPSHUFLWMask(PermMask.Val)) {
3509 if (V2.getOpcode() != ISD::UNDEF)
3510 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3511 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3515 if (X86::isSHUFPMask(PermMask.Val) &&
3516 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3519 // Floating point cases in the other order.
3520 if (X86::isSHUFPMask(PermMask.Val))
3522 if (X86::isPSHUFDMask(PermMask.Val) ||
3523 X86::isPSHUFHWMask(PermMask.Val) ||
3524 X86::isPSHUFLWMask(PermMask.Val)) {
3525 if (V2.getOpcode() != ISD::UNDEF)
3526 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3527 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3532 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3533 if (VT == MVT::v8i16) {
3534 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3539 // Handle all 4 wide cases with a number of shuffles.
3540 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3541 // Don't do this for MMX.
3542 MVT::ValueType MaskVT = PermMask.getValueType();
3543 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3544 SmallVector<std::pair<int, int>, 8> Locs;
3545 Locs.reserve(NumElems);
3546 SmallVector<SDOperand, 8> Mask1(NumElems,
3547 DAG.getNode(ISD::UNDEF, MaskEVT));
3548 SmallVector<SDOperand, 8> Mask2(NumElems,
3549 DAG.getNode(ISD::UNDEF, MaskEVT));
3552 // If no more than two elements come from either vector. This can be
3553 // implemented with two shuffles. First shuffle gather the elements.
3554 // The second shuffle, which takes the first shuffle as both of its
3555 // vector operands, put the elements into the right order.
3556 for (unsigned i = 0; i != NumElems; ++i) {
3557 SDOperand Elt = PermMask.getOperand(i);
3558 if (Elt.getOpcode() == ISD::UNDEF) {
3559 Locs[i] = std::make_pair(-1, -1);
3561 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3562 if (Val < NumElems) {
3563 Locs[i] = std::make_pair(0, NumLo);
3567 Locs[i] = std::make_pair(1, NumHi);
3568 if (2+NumHi < NumElems)
3569 Mask1[2+NumHi] = Elt;
3574 if (NumLo <= 2 && NumHi <= 2) {
3575 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3576 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3577 &Mask1[0], Mask1.size()));
3578 for (unsigned i = 0; i != NumElems; ++i) {
3579 if (Locs[i].first == -1)
3582 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3583 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3584 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3588 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3589 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3590 &Mask2[0], Mask2.size()));
3593 // Break it into (shuffle shuffle_hi, shuffle_lo).
3595 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3596 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3597 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3598 unsigned MaskIdx = 0;
3600 unsigned HiIdx = NumElems/2;
3601 for (unsigned i = 0; i != NumElems; ++i) {
3602 if (i == NumElems/2) {
3608 SDOperand Elt = PermMask.getOperand(i);
3609 if (Elt.getOpcode() == ISD::UNDEF) {
3610 Locs[i] = std::make_pair(-1, -1);
3611 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3612 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3613 (*MaskPtr)[LoIdx] = Elt;
3616 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3617 (*MaskPtr)[HiIdx] = Elt;
3622 SDOperand LoShuffle =
3623 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3624 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3625 &LoMask[0], LoMask.size()));
3626 SDOperand HiShuffle =
3627 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3628 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3629 &HiMask[0], HiMask.size()));
3630 SmallVector<SDOperand, 8> MaskOps;
3631 for (unsigned i = 0; i != NumElems; ++i) {
3632 if (Locs[i].first == -1) {
3633 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3635 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3636 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3639 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3640 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3641 &MaskOps[0], MaskOps.size()));
3648 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3649 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3652 MVT::ValueType VT = Op.getValueType();
3653 // TODO: handle v16i8.
3654 if (MVT::getSizeInBits(VT) == 16) {
3655 SDOperand Vec = Op.getOperand(0);
3656 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3658 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3659 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3660 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3662 // Transform it so it match pextrw which produces a 32-bit result.
3663 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3664 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3665 Op.getOperand(0), Op.getOperand(1));
3666 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3667 DAG.getValueType(VT));
3668 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3669 } else if (MVT::getSizeInBits(VT) == 32) {
3670 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3673 // SHUFPS the element to the lowest double word, then movss.
3674 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3675 SmallVector<SDOperand, 8> IdxVec;
3677 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3679 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3681 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3683 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3684 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3685 &IdxVec[0], IdxVec.size());
3686 SDOperand Vec = Op.getOperand(0);
3687 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3688 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3689 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3690 DAG.getIntPtrConstant(0));
3691 } else if (MVT::getSizeInBits(VT) == 64) {
3692 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3696 // UNPCKHPD the element to the lowest double word, then movsd.
3697 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3698 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3699 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3700 SmallVector<SDOperand, 8> IdxVec;
3701 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3703 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3704 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3705 &IdxVec[0], IdxVec.size());
3706 SDOperand Vec = Op.getOperand(0);
3707 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3708 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3709 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3710 DAG.getIntPtrConstant(0));
3717 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3718 MVT::ValueType VT = Op.getValueType();
3719 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3723 SDOperand N0 = Op.getOperand(0);
3724 SDOperand N1 = Op.getOperand(1);
3725 SDOperand N2 = Op.getOperand(2);
3727 if (MVT::getSizeInBits(EVT) == 16) {
3728 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3729 // as its second argument.
3730 if (N1.getValueType() != MVT::i32)
3731 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3732 if (N2.getValueType() != MVT::i32)
3733 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
3734 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3740 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3741 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3742 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3745 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3746 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3747 // one of the above mentioned nodes. It has to be wrapped because otherwise
3748 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3749 // be used to form addressing mode. These wrapped nodes will be selected
3752 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3753 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3754 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3756 CP->getAlignment());
3757 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3758 // With PIC, the address is actually $g + Offset.
3759 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3760 !Subtarget->isPICStyleRIPRel()) {
3761 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3762 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3770 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3771 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3772 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3773 // If it's a debug information descriptor, don't mess with it.
3774 if (DAG.isVerifiedDebugInfoDesc(Op))
3776 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3777 // With PIC, the address is actually $g + Offset.
3778 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3779 !Subtarget->isPICStyleRIPRel()) {
3780 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3781 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3785 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3786 // load the value at address GV, not the value of GV itself. This means that
3787 // the GlobalAddress must be in the base or index register of the address, not
3788 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3789 // The same applies for external symbols during PIC codegen
3790 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3791 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3796 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3798 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3799 const MVT::ValueType PtrVT) {
3801 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3802 DAG.getNode(X86ISD::GlobalBaseReg,
3804 InFlag = Chain.getValue(1);
3806 // emit leal symbol@TLSGD(,%ebx,1), %eax
3807 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3808 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3809 GA->getValueType(0),
3811 SDOperand Ops[] = { Chain, TGA, InFlag };
3812 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3813 InFlag = Result.getValue(2);
3814 Chain = Result.getValue(1);
3816 // call ___tls_get_addr. This function receives its argument in
3817 // the register EAX.
3818 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3819 InFlag = Chain.getValue(1);
3821 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3822 SDOperand Ops1[] = { Chain,
3823 DAG.getTargetExternalSymbol("___tls_get_addr",
3825 DAG.getRegister(X86::EAX, PtrVT),
3826 DAG.getRegister(X86::EBX, PtrVT),
3828 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3829 InFlag = Chain.getValue(1);
3831 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3834 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3835 // "local exec" model.
3837 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3838 const MVT::ValueType PtrVT) {
3839 // Get the Thread Pointer
3840 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3841 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3843 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3844 GA->getValueType(0),
3846 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3848 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3849 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3851 // The address of the thread local variable is the add of the thread
3852 // pointer with the offset of the variable.
3853 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3857 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3858 // TODO: implement the "local dynamic" model
3859 // TODO: implement the "initial exec"model for pic executables
3860 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3861 "TLS not implemented for non-ELF and 64-bit targets");
3862 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3863 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3864 // otherwise use the "Local Exec"TLS Model
3865 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3866 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3868 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3872 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3873 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3874 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3875 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3876 // With PIC, the address is actually $g + Offset.
3877 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3878 !Subtarget->isPICStyleRIPRel()) {
3879 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3880 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3887 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3888 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3889 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3890 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3891 // With PIC, the address is actually $g + Offset.
3892 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3893 !Subtarget->isPICStyleRIPRel()) {
3894 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3895 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3902 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3903 /// take a 2 x i32 value to shift plus a shift amount.
3904 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3905 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3906 "Not an i64 shift!");
3907 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3908 SDOperand ShOpLo = Op.getOperand(0);
3909 SDOperand ShOpHi = Op.getOperand(1);
3910 SDOperand ShAmt = Op.getOperand(2);
3911 SDOperand Tmp1 = isSRA ?
3912 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3913 DAG.getConstant(0, MVT::i32);
3915 SDOperand Tmp2, Tmp3;
3916 if (Op.getOpcode() == ISD::SHL_PARTS) {
3917 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3918 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3920 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3921 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3924 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3925 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3926 DAG.getConstant(32, MVT::i8));
3927 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3928 AndNode, DAG.getConstant(0, MVT::i8));
3931 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3932 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3933 SmallVector<SDOperand, 4> Ops;
3934 if (Op.getOpcode() == ISD::SHL_PARTS) {
3935 Ops.push_back(Tmp2);
3936 Ops.push_back(Tmp3);
3938 Ops.push_back(Cond);
3939 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3942 Ops.push_back(Tmp3);
3943 Ops.push_back(Tmp1);
3945 Ops.push_back(Cond);
3946 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3948 Ops.push_back(Tmp2);
3949 Ops.push_back(Tmp3);
3951 Ops.push_back(Cond);
3952 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3955 Ops.push_back(Tmp3);
3956 Ops.push_back(Tmp1);
3958 Ops.push_back(Cond);
3959 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3962 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3966 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3969 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3970 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3971 Op.getOperand(0).getValueType() >= MVT::i16 &&
3972 "Unknown SINT_TO_FP to lower!");
3975 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3976 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3977 MachineFunction &MF = DAG.getMachineFunction();
3978 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3979 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3980 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3981 StackSlot, NULL, 0);
3983 // These are really Legal; caller falls through into that case.
3984 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
3986 if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
3987 Subtarget->is64Bit())
3992 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
3994 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3996 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3997 SmallVector<SDOperand, 8> Ops;
3998 Ops.push_back(Chain);
3999 Ops.push_back(StackSlot);
4000 Ops.push_back(DAG.getValueType(SrcVT));
4001 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
4002 Tys, &Ops[0], Ops.size());
4005 Chain = Result.getValue(1);
4006 SDOperand InFlag = Result.getValue(2);
4008 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4009 // shouldn't be necessary except that RFP cannot be live across
4010 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4011 MachineFunction &MF = DAG.getMachineFunction();
4012 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4013 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4014 Tys = DAG.getVTList(MVT::Other);
4015 SmallVector<SDOperand, 8> Ops;
4016 Ops.push_back(Chain);
4017 Ops.push_back(Result);
4018 Ops.push_back(StackSlot);
4019 Ops.push_back(DAG.getValueType(Op.getValueType()));
4020 Ops.push_back(InFlag);
4021 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4022 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
4028 std::pair<SDOperand,SDOperand> X86TargetLowering::
4029 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4030 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4031 "Unknown FP_TO_SINT to lower!");
4033 // These are really Legal.
4034 if (Op.getValueType() == MVT::i32 &&
4035 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4036 return std::make_pair(SDOperand(), SDOperand());
4037 if (Subtarget->is64Bit() &&
4038 Op.getValueType() == MVT::i64 &&
4039 Op.getOperand(0).getValueType() != MVT::f80)
4040 return std::make_pair(SDOperand(), SDOperand());
4042 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4044 MachineFunction &MF = DAG.getMachineFunction();
4045 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4046 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4047 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4049 switch (Op.getValueType()) {
4050 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4051 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4052 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4053 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4056 SDOperand Chain = DAG.getEntryNode();
4057 SDOperand Value = Op.getOperand(0);
4058 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4059 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4060 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
4061 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4063 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4065 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4066 Chain = Value.getValue(1);
4067 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4068 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4071 // Build the FP_TO_INT*_IN_MEM
4072 SDOperand Ops[] = { Chain, Value, StackSlot };
4073 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4075 return std::make_pair(FIST, StackSlot);
4078 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4079 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4080 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4081 if (FIST.Val == 0) return SDOperand();
4084 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4087 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4088 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4089 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4090 if (FIST.Val == 0) return 0;
4092 // Return an i64 load from the stack slot.
4093 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4095 // Use a MERGE_VALUES node to drop the chain result value.
4096 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4099 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4100 MVT::ValueType VT = Op.getValueType();
4101 MVT::ValueType EltVT = VT;
4102 if (MVT::isVector(VT))
4103 EltVT = MVT::getVectorElementType(VT);
4104 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4105 std::vector<Constant*> CV;
4106 if (EltVT == MVT::f64) {
4107 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4111 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4117 Constant *C = ConstantVector::get(CV);
4118 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4119 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4121 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4124 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4125 MVT::ValueType VT = Op.getValueType();
4126 MVT::ValueType EltVT = VT;
4127 unsigned EltNum = 1;
4128 if (MVT::isVector(VT)) {
4129 EltVT = MVT::getVectorElementType(VT);
4130 EltNum = MVT::getVectorNumElements(VT);
4132 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4133 std::vector<Constant*> CV;
4134 if (EltVT == MVT::f64) {
4135 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4139 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4145 Constant *C = ConstantVector::get(CV);
4146 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4147 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4149 if (MVT::isVector(VT)) {
4150 return DAG.getNode(ISD::BIT_CONVERT, VT,
4151 DAG.getNode(ISD::XOR, MVT::v2i64,
4152 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4153 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4155 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4159 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4160 SDOperand Op0 = Op.getOperand(0);
4161 SDOperand Op1 = Op.getOperand(1);
4162 MVT::ValueType VT = Op.getValueType();
4163 MVT::ValueType SrcVT = Op1.getValueType();
4164 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4166 // If second operand is smaller, extend it first.
4167 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4168 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4170 SrcTy = MVT::getTypeForValueType(SrcVT);
4172 // And if it is bigger, shrink it first.
4173 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4174 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
4176 SrcTy = MVT::getTypeForValueType(SrcVT);
4179 // At this point the operands and the result should have the same
4180 // type, and that won't be f80 since that is not custom lowered.
4182 // First get the sign bit of second operand.
4183 std::vector<Constant*> CV;
4184 if (SrcVT == MVT::f64) {
4185 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4186 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4188 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4189 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4190 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4191 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4193 Constant *C = ConstantVector::get(CV);
4194 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4195 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4197 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4199 // Shift sign bit right or left if the two operands have different types.
4200 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4201 // Op0 is MVT::f32, Op1 is MVT::f64.
4202 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4203 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4204 DAG.getConstant(32, MVT::i32));
4205 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4206 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4207 DAG.getIntPtrConstant(0));
4210 // Clear first operand sign bit.
4212 if (VT == MVT::f64) {
4213 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4214 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4216 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4217 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4218 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4219 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4221 C = ConstantVector::get(CV);
4222 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4223 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4225 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4227 // Or the value with the sign bit.
4228 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4231 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4232 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4234 SDOperand Op0 = Op.getOperand(0);
4235 SDOperand Op1 = Op.getOperand(1);
4236 SDOperand CC = Op.getOperand(2);
4237 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4238 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4241 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4243 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4244 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4245 DAG.getConstant(X86CC, MVT::i8), Cond);
4248 assert(isFP && "Illegal integer SetCC!");
4250 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4251 switch (SetCCOpcode) {
4252 default: assert(false && "Illegal floating point SetCC!");
4253 case ISD::SETOEQ: { // !PF & ZF
4254 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4255 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4256 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4257 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4258 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4260 case ISD::SETUNE: { // PF | !ZF
4261 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4262 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4263 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4264 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4265 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4271 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4272 bool addTest = true;
4273 SDOperand Cond = Op.getOperand(0);
4276 if (Cond.getOpcode() == ISD::SETCC)
4277 Cond = LowerSETCC(Cond, DAG);
4279 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4280 // setting operand in place of the X86ISD::SETCC.
4281 if (Cond.getOpcode() == X86ISD::SETCC) {
4282 CC = Cond.getOperand(0);
4284 SDOperand Cmp = Cond.getOperand(1);
4285 unsigned Opc = Cmp.getOpcode();
4286 MVT::ValueType VT = Op.getValueType();
4288 bool IllegalFPCMov = false;
4289 if (MVT::isFloatingPoint(VT) && !MVT::isVector(VT) &&
4290 !isScalarFPTypeInSSEReg(VT)) // FPStack?
4291 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4293 if ((Opc == X86ISD::CMP ||
4294 Opc == X86ISD::COMI ||
4295 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4302 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4303 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4306 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4308 SmallVector<SDOperand, 4> Ops;
4309 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4310 // condition is true.
4311 Ops.push_back(Op.getOperand(2));
4312 Ops.push_back(Op.getOperand(1));
4314 Ops.push_back(Cond);
4315 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4318 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4319 bool addTest = true;
4320 SDOperand Chain = Op.getOperand(0);
4321 SDOperand Cond = Op.getOperand(1);
4322 SDOperand Dest = Op.getOperand(2);
4325 if (Cond.getOpcode() == ISD::SETCC)
4326 Cond = LowerSETCC(Cond, DAG);
4328 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4329 // setting operand in place of the X86ISD::SETCC.
4330 if (Cond.getOpcode() == X86ISD::SETCC) {
4331 CC = Cond.getOperand(0);
4333 SDOperand Cmp = Cond.getOperand(1);
4334 unsigned Opc = Cmp.getOpcode();
4335 if (Opc == X86ISD::CMP ||
4336 Opc == X86ISD::COMI ||
4337 Opc == X86ISD::UCOMI) {
4344 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4345 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4347 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4348 Chain, Op.getOperand(2), CC, Cond);
4352 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4353 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4354 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4355 // that the guard pages used by the OS virtual memory manager are allocated in
4356 // correct sequence.
4358 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4359 SelectionDAG &DAG) {
4360 assert(Subtarget->isTargetCygMing() &&
4361 "This should be used only on Cygwin/Mingw targets");
4364 SDOperand Chain = Op.getOperand(0);
4365 SDOperand Size = Op.getOperand(1);
4366 // FIXME: Ensure alignment here
4370 MVT::ValueType IntPtr = getPointerTy();
4371 MVT::ValueType SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
4373 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4374 Flag = Chain.getValue(1);
4376 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4377 SDOperand Ops[] = { Chain,
4378 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4379 DAG.getRegister(X86::EAX, IntPtr),
4381 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4382 Flag = Chain.getValue(1);
4384 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4386 std::vector<MVT::ValueType> Tys;
4387 Tys.push_back(SPTy);
4388 Tys.push_back(MVT::Other);
4389 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4390 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4393 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4394 SDOperand InFlag(0, 0);
4395 SDOperand Chain = Op.getOperand(0);
4397 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4398 if (Align == 0) Align = 1;
4400 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4401 // If not DWORD aligned or size is more than the threshold, call memset.
4402 // The libc version is likely to be faster for these cases. It can use the
4403 // address value and run time information about the CPU.
4404 if ((Align & 3) != 0 ||
4405 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4406 MVT::ValueType IntPtr = getPointerTy();
4407 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4408 TargetLowering::ArgListTy Args;
4409 TargetLowering::ArgListEntry Entry;
4410 Entry.Node = Op.getOperand(1);
4411 Entry.Ty = IntPtrTy;
4412 Args.push_back(Entry);
4413 // Extend the unsigned i8 argument to be an int value for the call.
4414 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4415 Entry.Ty = IntPtrTy;
4416 Args.push_back(Entry);
4417 Entry.Node = Op.getOperand(3);
4418 Args.push_back(Entry);
4419 std::pair<SDOperand,SDOperand> CallResult =
4420 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4421 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4422 return CallResult.second;
4427 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4428 unsigned BytesLeft = 0;
4429 bool TwoRepStos = false;
4432 uint64_t Val = ValC->getValue() & 255;
4434 // If the value is a constant, then we can potentially use larger sets.
4435 switch (Align & 3) {
4436 case 2: // WORD aligned
4439 Val = (Val << 8) | Val;
4441 case 0: // DWORD aligned
4444 Val = (Val << 8) | Val;
4445 Val = (Val << 16) | Val;
4446 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4449 Val = (Val << 32) | Val;
4452 default: // Byte aligned
4455 Count = Op.getOperand(3);
4459 if (AVT > MVT::i8) {
4461 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4462 Count = DAG.getIntPtrConstant(I->getValue() / UBytes);
4463 BytesLeft = I->getValue() % UBytes;
4465 assert(AVT >= MVT::i32 &&
4466 "Do not use rep;stos if not at least DWORD aligned");
4467 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4468 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4473 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4475 InFlag = Chain.getValue(1);
4478 Count = Op.getOperand(3);
4479 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4480 InFlag = Chain.getValue(1);
4483 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4485 InFlag = Chain.getValue(1);
4486 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4487 Op.getOperand(1), InFlag);
4488 InFlag = Chain.getValue(1);
4490 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4491 SmallVector<SDOperand, 8> Ops;
4492 Ops.push_back(Chain);
4493 Ops.push_back(DAG.getValueType(AVT));
4494 Ops.push_back(InFlag);
4495 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4498 InFlag = Chain.getValue(1);
4499 Count = Op.getOperand(3);
4500 MVT::ValueType CVT = Count.getValueType();
4501 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4502 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4503 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4505 InFlag = Chain.getValue(1);
4506 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4508 Ops.push_back(Chain);
4509 Ops.push_back(DAG.getValueType(MVT::i8));
4510 Ops.push_back(InFlag);
4511 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4512 } else if (BytesLeft) {
4513 // Issue stores for the last 1 - 7 bytes.
4515 unsigned Val = ValC->getValue() & 255;
4516 unsigned Offset = I->getValue() - BytesLeft;
4517 SDOperand DstAddr = Op.getOperand(1);
4518 MVT::ValueType AddrVT = DstAddr.getValueType();
4519 if (BytesLeft >= 4) {
4520 Val = (Val << 8) | Val;
4521 Val = (Val << 16) | Val;
4522 Value = DAG.getConstant(Val, MVT::i32);
4523 Chain = DAG.getStore(Chain, Value,
4524 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4525 DAG.getConstant(Offset, AddrVT)),
4530 if (BytesLeft >= 2) {
4531 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4532 Chain = DAG.getStore(Chain, Value,
4533 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4534 DAG.getConstant(Offset, AddrVT)),
4539 if (BytesLeft == 1) {
4540 Value = DAG.getConstant(Val, MVT::i8);
4541 Chain = DAG.getStore(Chain, Value,
4542 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4543 DAG.getConstant(Offset, AddrVT)),
4551 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4556 SelectionDAG &DAG) {
4558 unsigned BytesLeft = 0;
4559 switch (Align & 3) {
4560 case 2: // WORD aligned
4563 case 0: // DWORD aligned
4565 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4568 default: // Byte aligned
4573 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4574 SDOperand Count = DAG.getIntPtrConstant(Size / UBytes);
4575 BytesLeft = Size % UBytes;
4577 SDOperand InFlag(0, 0);
4578 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4580 InFlag = Chain.getValue(1);
4581 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4583 InFlag = Chain.getValue(1);
4584 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4586 InFlag = Chain.getValue(1);
4588 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4589 SmallVector<SDOperand, 8> Ops;
4590 Ops.push_back(Chain);
4591 Ops.push_back(DAG.getValueType(AVT));
4592 Ops.push_back(InFlag);
4593 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4596 // Issue loads and stores for the last 1 - 7 bytes.
4597 unsigned Offset = Size - BytesLeft;
4598 SDOperand DstAddr = Dest;
4599 MVT::ValueType DstVT = DstAddr.getValueType();
4600 SDOperand SrcAddr = Source;
4601 MVT::ValueType SrcVT = SrcAddr.getValueType();
4603 if (BytesLeft >= 4) {
4604 Value = DAG.getLoad(MVT::i32, Chain,
4605 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4606 DAG.getConstant(Offset, SrcVT)),
4608 Chain = Value.getValue(1);
4609 Chain = DAG.getStore(Chain, Value,
4610 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4611 DAG.getConstant(Offset, DstVT)),
4616 if (BytesLeft >= 2) {
4617 Value = DAG.getLoad(MVT::i16, Chain,
4618 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4619 DAG.getConstant(Offset, SrcVT)),
4621 Chain = Value.getValue(1);
4622 Chain = DAG.getStore(Chain, Value,
4623 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4624 DAG.getConstant(Offset, DstVT)),
4630 if (BytesLeft == 1) {
4631 Value = DAG.getLoad(MVT::i8, Chain,
4632 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4633 DAG.getConstant(Offset, SrcVT)),
4635 Chain = Value.getValue(1);
4636 Chain = DAG.getStore(Chain, Value,
4637 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4638 DAG.getConstant(Offset, DstVT)),
4646 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4647 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4648 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4649 SDOperand TheChain = N->getOperand(0);
4650 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4651 if (Subtarget->is64Bit()) {
4652 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4653 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4654 MVT::i64, rax.getValue(2));
4655 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4656 DAG.getConstant(32, MVT::i8));
4658 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4661 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4662 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4665 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4666 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4667 MVT::i32, eax.getValue(2));
4668 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4669 SDOperand Ops[] = { eax, edx };
4670 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
4672 // Use a MERGE_VALUES to return the value and chain.
4673 Ops[1] = edx.getValue(1);
4674 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4675 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4678 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4679 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4681 if (!Subtarget->is64Bit()) {
4682 // vastart just stores the address of the VarArgsFrameIndex slot into the
4683 // memory location argument.
4684 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4685 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4690 // gp_offset (0 - 6 * 8)
4691 // fp_offset (48 - 48 + 8 * 16)
4692 // overflow_arg_area (point to parameters coming in memory).
4694 SmallVector<SDOperand, 8> MemOps;
4695 SDOperand FIN = Op.getOperand(1);
4697 SDOperand Store = DAG.getStore(Op.getOperand(0),
4698 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4699 FIN, SV->getValue(), SV->getOffset());
4700 MemOps.push_back(Store);
4703 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4704 Store = DAG.getStore(Op.getOperand(0),
4705 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4706 FIN, SV->getValue(), SV->getOffset());
4707 MemOps.push_back(Store);
4709 // Store ptr to overflow_arg_area
4710 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(4));
4711 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4712 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4714 MemOps.push_back(Store);
4716 // Store ptr to reg_save_area.
4717 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, DAG.getIntPtrConstant(8));
4718 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4719 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4721 MemOps.push_back(Store);
4722 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4725 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4726 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4727 SDOperand Chain = Op.getOperand(0);
4728 SDOperand DstPtr = Op.getOperand(1);
4729 SDOperand SrcPtr = Op.getOperand(2);
4730 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4731 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4733 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4734 SrcSV->getValue(), SrcSV->getOffset());
4735 Chain = SrcPtr.getValue(1);
4736 for (unsigned i = 0; i < 3; ++i) {
4737 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4738 SrcSV->getValue(), SrcSV->getOffset());
4739 Chain = Val.getValue(1);
4740 Chain = DAG.getStore(Chain, Val, DstPtr,
4741 DstSV->getValue(), DstSV->getOffset());
4744 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4745 DAG.getIntPtrConstant(8));
4746 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4747 DAG.getIntPtrConstant(8));
4753 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4754 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4756 default: return SDOperand(); // Don't custom lower most intrinsics.
4757 // Comparison intrinsics.
4758 case Intrinsic::x86_sse_comieq_ss:
4759 case Intrinsic::x86_sse_comilt_ss:
4760 case Intrinsic::x86_sse_comile_ss:
4761 case Intrinsic::x86_sse_comigt_ss:
4762 case Intrinsic::x86_sse_comige_ss:
4763 case Intrinsic::x86_sse_comineq_ss:
4764 case Intrinsic::x86_sse_ucomieq_ss:
4765 case Intrinsic::x86_sse_ucomilt_ss:
4766 case Intrinsic::x86_sse_ucomile_ss:
4767 case Intrinsic::x86_sse_ucomigt_ss:
4768 case Intrinsic::x86_sse_ucomige_ss:
4769 case Intrinsic::x86_sse_ucomineq_ss:
4770 case Intrinsic::x86_sse2_comieq_sd:
4771 case Intrinsic::x86_sse2_comilt_sd:
4772 case Intrinsic::x86_sse2_comile_sd:
4773 case Intrinsic::x86_sse2_comigt_sd:
4774 case Intrinsic::x86_sse2_comige_sd:
4775 case Intrinsic::x86_sse2_comineq_sd:
4776 case Intrinsic::x86_sse2_ucomieq_sd:
4777 case Intrinsic::x86_sse2_ucomilt_sd:
4778 case Intrinsic::x86_sse2_ucomile_sd:
4779 case Intrinsic::x86_sse2_ucomigt_sd:
4780 case Intrinsic::x86_sse2_ucomige_sd:
4781 case Intrinsic::x86_sse2_ucomineq_sd: {
4783 ISD::CondCode CC = ISD::SETCC_INVALID;
4786 case Intrinsic::x86_sse_comieq_ss:
4787 case Intrinsic::x86_sse2_comieq_sd:
4791 case Intrinsic::x86_sse_comilt_ss:
4792 case Intrinsic::x86_sse2_comilt_sd:
4796 case Intrinsic::x86_sse_comile_ss:
4797 case Intrinsic::x86_sse2_comile_sd:
4801 case Intrinsic::x86_sse_comigt_ss:
4802 case Intrinsic::x86_sse2_comigt_sd:
4806 case Intrinsic::x86_sse_comige_ss:
4807 case Intrinsic::x86_sse2_comige_sd:
4811 case Intrinsic::x86_sse_comineq_ss:
4812 case Intrinsic::x86_sse2_comineq_sd:
4816 case Intrinsic::x86_sse_ucomieq_ss:
4817 case Intrinsic::x86_sse2_ucomieq_sd:
4818 Opc = X86ISD::UCOMI;
4821 case Intrinsic::x86_sse_ucomilt_ss:
4822 case Intrinsic::x86_sse2_ucomilt_sd:
4823 Opc = X86ISD::UCOMI;
4826 case Intrinsic::x86_sse_ucomile_ss:
4827 case Intrinsic::x86_sse2_ucomile_sd:
4828 Opc = X86ISD::UCOMI;
4831 case Intrinsic::x86_sse_ucomigt_ss:
4832 case Intrinsic::x86_sse2_ucomigt_sd:
4833 Opc = X86ISD::UCOMI;
4836 case Intrinsic::x86_sse_ucomige_ss:
4837 case Intrinsic::x86_sse2_ucomige_sd:
4838 Opc = X86ISD::UCOMI;
4841 case Intrinsic::x86_sse_ucomineq_ss:
4842 case Intrinsic::x86_sse2_ucomineq_sd:
4843 Opc = X86ISD::UCOMI;
4849 SDOperand LHS = Op.getOperand(1);
4850 SDOperand RHS = Op.getOperand(2);
4851 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4853 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4854 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4855 DAG.getConstant(X86CC, MVT::i8), Cond);
4856 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4861 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4862 // Depths > 0 not supported yet!
4863 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4866 // Just load the return address
4867 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4868 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4871 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4872 // Depths > 0 not supported yet!
4873 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4876 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4877 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4878 DAG.getIntPtrConstant(4));
4881 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4882 SelectionDAG &DAG) {
4883 // Is not yet supported on x86-64
4884 if (Subtarget->is64Bit())
4887 return DAG.getIntPtrConstant(8);
4890 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4892 assert(!Subtarget->is64Bit() &&
4893 "Lowering of eh_return builtin is not supported yet on x86-64");
4895 MachineFunction &MF = DAG.getMachineFunction();
4896 SDOperand Chain = Op.getOperand(0);
4897 SDOperand Offset = Op.getOperand(1);
4898 SDOperand Handler = Op.getOperand(2);
4900 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4903 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4904 DAG.getIntPtrConstant(-4UL));
4905 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4906 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4907 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4908 MF.getRegInfo().addLiveOut(X86::ECX);
4910 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4911 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4914 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4915 SelectionDAG &DAG) {
4916 SDOperand Root = Op.getOperand(0);
4917 SDOperand Trmp = Op.getOperand(1); // trampoline
4918 SDOperand FPtr = Op.getOperand(2); // nested function
4919 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4921 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4923 const X86InstrInfo *TII =
4924 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4926 if (Subtarget->is64Bit()) {
4927 SDOperand OutChains[6];
4929 // Large code-model.
4931 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
4932 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
4934 const unsigned char N86R10 =
4935 ((X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R10);
4936 const unsigned char N86R11 =
4937 ((X86RegisterInfo*)RegInfo)->getX86RegNum(X86::R11);
4939 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
4941 // Load the pointer to the nested function into R11.
4942 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
4943 SDOperand Addr = Trmp;
4944 OutChains[0] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4945 TrmpSV->getValue(), TrmpSV->getOffset());
4947 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(2, MVT::i64));
4948 OutChains[1] = DAG.getStore(Root, FPtr, Addr, TrmpSV->getValue(),
4949 TrmpSV->getOffset() + 2, false, 2);
4951 // Load the 'nest' parameter value into R10.
4952 // R10 is specified in X86CallingConv.td
4953 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
4954 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(10, MVT::i64));
4955 OutChains[2] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4956 TrmpSV->getValue(), TrmpSV->getOffset() + 10);
4958 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(12, MVT::i64));
4959 OutChains[3] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4960 TrmpSV->getOffset() + 12, false, 2);
4962 // Jump to the nested function.
4963 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
4964 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(20, MVT::i64));
4965 OutChains[4] = DAG.getStore(Root, DAG.getConstant(OpCode, MVT::i16), Addr,
4966 TrmpSV->getValue(), TrmpSV->getOffset() + 20);
4968 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
4969 Addr = DAG.getNode(ISD::ADD, MVT::i64, Trmp, DAG.getConstant(22, MVT::i64));
4970 OutChains[5] = DAG.getStore(Root, DAG.getConstant(ModRM, MVT::i8), Addr,
4971 TrmpSV->getValue(), TrmpSV->getOffset() + 22);
4974 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 6) };
4975 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4977 const Function *Func =
4978 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4979 unsigned CC = Func->getCallingConv();
4984 assert(0 && "Unsupported calling convention");
4985 case CallingConv::C:
4986 case CallingConv::X86_StdCall: {
4987 // Pass 'nest' parameter in ECX.
4988 // Must be kept in sync with X86CallingConv.td
4991 // Check that ECX wasn't needed by an 'inreg' parameter.
4992 const FunctionType *FTy = Func->getFunctionType();
4993 const ParamAttrsList *Attrs = Func->getParamAttrs();
4995 if (Attrs && !Func->isVarArg()) {
4996 unsigned InRegCount = 0;
4999 for (FunctionType::param_iterator I = FTy->param_begin(),
5000 E = FTy->param_end(); I != E; ++I, ++Idx)
5001 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
5002 // FIXME: should only count parameters that are lowered to integers.
5003 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5005 if (InRegCount > 2) {
5006 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5012 case CallingConv::X86_FastCall:
5013 // Pass 'nest' parameter in EAX.
5014 // Must be kept in sync with X86CallingConv.td
5019 SDOperand OutChains[4];
5020 SDOperand Addr, Disp;
5022 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5023 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5025 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5026 const unsigned char N86Reg =
5027 ((X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5028 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5029 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
5031 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5032 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
5033 TrmpSV->getOffset() + 1, false, 1);
5035 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5036 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5037 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5038 TrmpSV->getValue() + 5, TrmpSV->getOffset());
5040 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5041 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
5042 TrmpSV->getOffset() + 6, false, 1);
5045 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5046 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5050 SDOperand X86TargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
5052 The rounding mode is in bits 11:10 of FPSR, and has the following
5059 FLT_ROUNDS, on the other hand, expects the following:
5066 To perform the conversion, we do:
5067 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5070 MachineFunction &MF = DAG.getMachineFunction();
5071 const TargetMachine &TM = MF.getTarget();
5072 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5073 unsigned StackAlignment = TFI.getStackAlignment();
5074 MVT::ValueType VT = Op.getValueType();
5076 // Save FP Control Word to stack slot
5077 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5078 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5080 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5081 DAG.getEntryNode(), StackSlot);
5083 // Load FP Control Word from stack slot
5084 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5086 // Transform as necessary
5088 DAG.getNode(ISD::SRL, MVT::i16,
5089 DAG.getNode(ISD::AND, MVT::i16,
5090 CWD, DAG.getConstant(0x800, MVT::i16)),
5091 DAG.getConstant(11, MVT::i8));
5093 DAG.getNode(ISD::SRL, MVT::i16,
5094 DAG.getNode(ISD::AND, MVT::i16,
5095 CWD, DAG.getConstant(0x400, MVT::i16)),
5096 DAG.getConstant(9, MVT::i8));
5099 DAG.getNode(ISD::AND, MVT::i16,
5100 DAG.getNode(ISD::ADD, MVT::i16,
5101 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5102 DAG.getConstant(1, MVT::i16)),
5103 DAG.getConstant(3, MVT::i16));
5106 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5107 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5110 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5111 MVT::ValueType VT = Op.getValueType();
5112 MVT::ValueType OpVT = VT;
5113 unsigned NumBits = MVT::getSizeInBits(VT);
5115 Op = Op.getOperand(0);
5116 if (VT == MVT::i8) {
5117 // Zero extend to i32 since there is not an i8 bsr.
5119 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5122 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5123 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5124 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5126 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5127 SmallVector<SDOperand, 4> Ops;
5129 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5130 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5131 Ops.push_back(Op.getValue(1));
5132 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5134 // Finally xor with NumBits-1.
5135 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5138 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5142 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5143 MVT::ValueType VT = Op.getValueType();
5144 MVT::ValueType OpVT = VT;
5145 unsigned NumBits = MVT::getSizeInBits(VT);
5147 Op = Op.getOperand(0);
5148 if (VT == MVT::i8) {
5150 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5153 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5154 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5155 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5157 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5158 SmallVector<SDOperand, 4> Ops;
5160 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5161 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5162 Ops.push_back(Op.getValue(1));
5163 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5166 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5170 /// LowerOperation - Provide custom lowering hooks for some operations.
5172 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5173 switch (Op.getOpcode()) {
5174 default: assert(0 && "Should not custom lower this!");
5175 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5176 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5177 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5178 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5179 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5180 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5181 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5182 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5183 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5184 case ISD::SHL_PARTS:
5185 case ISD::SRA_PARTS:
5186 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5187 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5188 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5189 case ISD::FABS: return LowerFABS(Op, DAG);
5190 case ISD::FNEG: return LowerFNEG(Op, DAG);
5191 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5192 case ISD::SETCC: return LowerSETCC(Op, DAG);
5193 case ISD::SELECT: return LowerSELECT(Op, DAG);
5194 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5195 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5196 case ISD::CALL: return LowerCALL(Op, DAG);
5197 case ISD::RET: return LowerRET(Op, DAG);
5198 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5199 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5200 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5201 case ISD::VASTART: return LowerVASTART(Op, DAG);
5202 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5203 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5204 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5205 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5206 case ISD::FRAME_TO_ARGS_OFFSET:
5207 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5208 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5209 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5210 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5211 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5212 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5213 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5215 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5216 case ISD::READCYCLECOUNTER:
5217 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5221 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5222 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5223 switch (N->getOpcode()) {
5224 default: assert(0 && "Should not custom lower this!");
5225 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5226 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5230 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5232 default: return NULL;
5233 case X86ISD::BSF: return "X86ISD::BSF";
5234 case X86ISD::BSR: return "X86ISD::BSR";
5235 case X86ISD::SHLD: return "X86ISD::SHLD";
5236 case X86ISD::SHRD: return "X86ISD::SHRD";
5237 case X86ISD::FAND: return "X86ISD::FAND";
5238 case X86ISD::FOR: return "X86ISD::FOR";
5239 case X86ISD::FXOR: return "X86ISD::FXOR";
5240 case X86ISD::FSRL: return "X86ISD::FSRL";
5241 case X86ISD::FILD: return "X86ISD::FILD";
5242 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5243 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5244 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5245 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5246 case X86ISD::FLD: return "X86ISD::FLD";
5247 case X86ISD::FST: return "X86ISD::FST";
5248 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5249 case X86ISD::FP_GET_RESULT2: return "X86ISD::FP_GET_RESULT2";
5250 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5251 case X86ISD::CALL: return "X86ISD::CALL";
5252 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5253 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5254 case X86ISD::CMP: return "X86ISD::CMP";
5255 case X86ISD::COMI: return "X86ISD::COMI";
5256 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5257 case X86ISD::SETCC: return "X86ISD::SETCC";
5258 case X86ISD::CMOV: return "X86ISD::CMOV";
5259 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5260 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5261 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5262 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5263 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5264 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5265 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5266 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5267 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5268 case X86ISD::FMAX: return "X86ISD::FMAX";
5269 case X86ISD::FMIN: return "X86ISD::FMIN";
5270 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5271 case X86ISD::FRCP: return "X86ISD::FRCP";
5272 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5273 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5274 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5275 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5276 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5280 // isLegalAddressingMode - Return true if the addressing mode represented
5281 // by AM is legal for this target, for a load/store of the specified type.
5282 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5283 const Type *Ty) const {
5284 // X86 supports extremely general addressing modes.
5286 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5287 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5291 // We can only fold this if we don't need an extra load.
5292 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5295 // X86-64 only supports addr of globals in small code model.
5296 if (Subtarget->is64Bit()) {
5297 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5299 // If lower 4G is not available, then we must use rip-relative addressing.
5300 if (AM.BaseOffs || AM.Scale > 1)
5311 // These scales always work.
5316 // These scales are formed with basereg+scalereg. Only accept if there is
5321 default: // Other stuff never works.
5329 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5330 if (!Ty1->isInteger() || !Ty2->isInteger())
5332 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5333 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5334 if (NumBits1 <= NumBits2)
5336 return Subtarget->is64Bit() || NumBits1 < 64;
5339 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5340 MVT::ValueType VT2) const {
5341 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5343 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5344 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5345 if (NumBits1 <= NumBits2)
5347 return Subtarget->is64Bit() || NumBits1 < 64;
5350 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5351 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5352 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5353 /// are assumed to be legal.
5355 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5356 // Only do shuffles on 128-bit vector types for now.
5357 if (MVT::getSizeInBits(VT) == 64) return false;
5358 return (Mask.Val->getNumOperands() <= 4 ||
5359 isIdentityMask(Mask.Val) ||
5360 isIdentityMask(Mask.Val, true) ||
5361 isSplatMask(Mask.Val) ||
5362 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5363 X86::isUNPCKLMask(Mask.Val) ||
5364 X86::isUNPCKHMask(Mask.Val) ||
5365 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5366 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5369 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5371 SelectionDAG &DAG) const {
5372 unsigned NumElts = BVOps.size();
5373 // Only do shuffles on 128-bit vector types for now.
5374 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5375 if (NumElts == 2) return true;
5377 return (isMOVLMask(&BVOps[0], 4) ||
5378 isCommutedMOVL(&BVOps[0], 4, true) ||
5379 isSHUFPMask(&BVOps[0], 4) ||
5380 isCommutedSHUFP(&BVOps[0], 4));
5385 //===----------------------------------------------------------------------===//
5386 // X86 Scheduler Hooks
5387 //===----------------------------------------------------------------------===//
5390 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5391 MachineBasicBlock *BB) {
5392 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5393 switch (MI->getOpcode()) {
5394 default: assert(false && "Unexpected instr type to insert");
5395 case X86::CMOV_FR32:
5396 case X86::CMOV_FR64:
5397 case X86::CMOV_V4F32:
5398 case X86::CMOV_V2F64:
5399 case X86::CMOV_V2I64: {
5400 // To "insert" a SELECT_CC instruction, we actually have to insert the
5401 // diamond control-flow pattern. The incoming instruction knows the
5402 // destination vreg to set, the condition code register to branch on, the
5403 // true/false values to select between, and a branch opcode to use.
5404 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5405 ilist<MachineBasicBlock>::iterator It = BB;
5411 // cmpTY ccX, r1, r2
5413 // fallthrough --> copy0MBB
5414 MachineBasicBlock *thisMBB = BB;
5415 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5416 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5418 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5419 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5420 MachineFunction *F = BB->getParent();
5421 F->getBasicBlockList().insert(It, copy0MBB);
5422 F->getBasicBlockList().insert(It, sinkMBB);
5423 // Update machine-CFG edges by first adding all successors of the current
5424 // block to the new block which will contain the Phi node for the select.
5425 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5426 e = BB->succ_end(); i != e; ++i)
5427 sinkMBB->addSuccessor(*i);
5428 // Next, remove all successors of the current block, and add the true
5429 // and fallthrough blocks as its successors.
5430 while(!BB->succ_empty())
5431 BB->removeSuccessor(BB->succ_begin());
5432 BB->addSuccessor(copy0MBB);
5433 BB->addSuccessor(sinkMBB);
5436 // %FalseValue = ...
5437 // # fallthrough to sinkMBB
5440 // Update machine-CFG edges
5441 BB->addSuccessor(sinkMBB);
5444 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5447 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5448 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5449 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5451 delete MI; // The pseudo instruction is gone now.
5455 case X86::FP32_TO_INT16_IN_MEM:
5456 case X86::FP32_TO_INT32_IN_MEM:
5457 case X86::FP32_TO_INT64_IN_MEM:
5458 case X86::FP64_TO_INT16_IN_MEM:
5459 case X86::FP64_TO_INT32_IN_MEM:
5460 case X86::FP64_TO_INT64_IN_MEM:
5461 case X86::FP80_TO_INT16_IN_MEM:
5462 case X86::FP80_TO_INT32_IN_MEM:
5463 case X86::FP80_TO_INT64_IN_MEM: {
5464 // Change the floating point control register to use "round towards zero"
5465 // mode when truncating to an integer value.
5466 MachineFunction *F = BB->getParent();
5467 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5468 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5470 // Load the old value of the high byte of the control word...
5472 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
5473 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5475 // Set the high part to be round to zero...
5476 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5479 // Reload the modified control word now...
5480 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5482 // Restore the memory image of control word to original value
5483 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5486 // Get the X86 opcode to use.
5488 switch (MI->getOpcode()) {
5489 default: assert(0 && "illegal opcode!");
5490 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5491 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5492 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5493 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5494 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5495 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5496 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5497 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5498 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5502 MachineOperand &Op = MI->getOperand(0);
5503 if (Op.isRegister()) {
5504 AM.BaseType = X86AddressMode::RegBase;
5505 AM.Base.Reg = Op.getReg();
5507 AM.BaseType = X86AddressMode::FrameIndexBase;
5508 AM.Base.FrameIndex = Op.getIndex();
5510 Op = MI->getOperand(1);
5511 if (Op.isImmediate())
5512 AM.Scale = Op.getImm();
5513 Op = MI->getOperand(2);
5514 if (Op.isImmediate())
5515 AM.IndexReg = Op.getImm();
5516 Op = MI->getOperand(3);
5517 if (Op.isGlobalAddress()) {
5518 AM.GV = Op.getGlobal();
5520 AM.Disp = Op.getImm();
5522 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5523 .addReg(MI->getOperand(4).getReg());
5525 // Reload the original control word now.
5526 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5528 delete MI; // The pseudo instruction is gone now.
5534 //===----------------------------------------------------------------------===//
5535 // X86 Optimization Hooks
5536 //===----------------------------------------------------------------------===//
5538 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5540 uint64_t &KnownZero,
5542 const SelectionDAG &DAG,
5543 unsigned Depth) const {
5544 unsigned Opc = Op.getOpcode();
5545 assert((Opc >= ISD::BUILTIN_OP_END ||
5546 Opc == ISD::INTRINSIC_WO_CHAIN ||
5547 Opc == ISD::INTRINSIC_W_CHAIN ||
5548 Opc == ISD::INTRINSIC_VOID) &&
5549 "Should use MaskedValueIsZero if you don't know whether Op"
5550 " is a target node!");
5552 KnownZero = KnownOne = 0; // Don't know anything.
5556 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5561 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5562 /// element of the result of the vector shuffle.
5563 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5564 MVT::ValueType VT = N->getValueType(0);
5565 SDOperand PermMask = N->getOperand(2);
5566 unsigned NumElems = PermMask.getNumOperands();
5567 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5569 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5571 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5572 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5573 SDOperand Idx = PermMask.getOperand(i);
5574 if (Idx.getOpcode() == ISD::UNDEF)
5575 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5576 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5581 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5582 /// node is a GlobalAddress + an offset.
5583 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5584 unsigned Opc = N->getOpcode();
5585 if (Opc == X86ISD::Wrapper) {
5586 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5587 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5590 } else if (Opc == ISD::ADD) {
5591 SDOperand N1 = N->getOperand(0);
5592 SDOperand N2 = N->getOperand(1);
5593 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5594 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5596 Offset += V->getSignExtended();
5599 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5600 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5602 Offset += V->getSignExtended();
5610 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5612 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5613 MachineFrameInfo *MFI) {
5614 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5617 SDOperand Loc = N->getOperand(1);
5618 SDOperand BaseLoc = Base->getOperand(1);
5619 if (Loc.getOpcode() == ISD::FrameIndex) {
5620 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5622 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5623 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5624 int FS = MFI->getObjectSize(FI);
5625 int BFS = MFI->getObjectSize(BFI);
5626 if (FS != BFS || FS != Size) return false;
5627 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5629 GlobalValue *GV1 = NULL;
5630 GlobalValue *GV2 = NULL;
5631 int64_t Offset1 = 0;
5632 int64_t Offset2 = 0;
5633 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5634 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5635 if (isGA1 && isGA2 && GV1 == GV2)
5636 return Offset1 == (Offset2 + Dist*Size);
5642 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5643 const X86Subtarget *Subtarget) {
5646 if (isGAPlusOffset(Base, GV, Offset))
5647 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5648 // DAG combine handles the stack object case.
5653 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5654 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5655 /// if the load addresses are consecutive, non-overlapping, and in the right
5657 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5658 const X86Subtarget *Subtarget) {
5659 MachineFunction &MF = DAG.getMachineFunction();
5660 MachineFrameInfo *MFI = MF.getFrameInfo();
5661 MVT::ValueType VT = N->getValueType(0);
5662 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5663 SDOperand PermMask = N->getOperand(2);
5664 int NumElems = (int)PermMask.getNumOperands();
5665 SDNode *Base = NULL;
5666 for (int i = 0; i < NumElems; ++i) {
5667 SDOperand Idx = PermMask.getOperand(i);
5668 if (Idx.getOpcode() == ISD::UNDEF) {
5669 if (!Base) return SDOperand();
5672 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5673 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5677 else if (!isConsecutiveLoad(Arg.Val, Base,
5678 i, MVT::getSizeInBits(EVT)/8,MFI))
5683 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5684 LoadSDNode *LD = cast<LoadSDNode>(Base);
5686 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5687 LD->getSrcValueOffset(), LD->isVolatile());
5689 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5690 LD->getSrcValueOffset(), LD->isVolatile(),
5691 LD->getAlignment());
5695 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5696 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5697 const X86Subtarget *Subtarget) {
5698 SDOperand Cond = N->getOperand(0);
5700 // If we have SSE[12] support, try to form min/max nodes.
5701 if (Subtarget->hasSSE2() &&
5702 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5703 if (Cond.getOpcode() == ISD::SETCC) {
5704 // Get the LHS/RHS of the select.
5705 SDOperand LHS = N->getOperand(1);
5706 SDOperand RHS = N->getOperand(2);
5707 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5709 unsigned Opcode = 0;
5710 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5713 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5716 if (!UnsafeFPMath) break;
5718 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5720 Opcode = X86ISD::FMIN;
5723 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5726 if (!UnsafeFPMath) break;
5728 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5730 Opcode = X86ISD::FMAX;
5733 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5736 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5739 if (!UnsafeFPMath) break;
5741 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5743 Opcode = X86ISD::FMIN;
5746 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5749 if (!UnsafeFPMath) break;
5751 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5753 Opcode = X86ISD::FMAX;
5759 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5767 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
5768 /// X86ISD::FXOR nodes.
5769 static SDOperand PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
5770 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
5771 // F[X]OR(0.0, x) -> x
5772 // F[X]OR(x, 0.0) -> x
5773 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
5774 if (C->getValueAPF().isPosZero())
5775 return N->getOperand(1);
5776 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
5777 if (C->getValueAPF().isPosZero())
5778 return N->getOperand(0);
5782 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
5783 static SDOperand PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
5784 // FAND(0.0, x) -> 0.0
5785 // FAND(x, 0.0) -> 0.0
5786 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
5787 if (C->getValueAPF().isPosZero())
5788 return N->getOperand(0);
5789 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
5790 if (C->getValueAPF().isPosZero())
5791 return N->getOperand(1);
5796 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5797 DAGCombinerInfo &DCI) const {
5798 SelectionDAG &DAG = DCI.DAG;
5799 switch (N->getOpcode()) {
5801 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, Subtarget);
5802 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
5804 case X86ISD::FOR: return PerformFORCombine(N, DAG);
5805 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
5811 //===----------------------------------------------------------------------===//
5812 // X86 Inline Assembly Support
5813 //===----------------------------------------------------------------------===//
5815 /// getConstraintType - Given a constraint letter, return the type of
5816 /// constraint it is for this target.
5817 X86TargetLowering::ConstraintType
5818 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5819 if (Constraint.size() == 1) {
5820 switch (Constraint[0]) {
5829 return C_RegisterClass;
5834 return TargetLowering::getConstraintType(Constraint);
5837 /// LowerXConstraint - try to replace an X constraint, which matches anything,
5838 /// with another that has more specific requirements based on the type of the
5839 /// corresponding operand.
5840 void X86TargetLowering::lowerXConstraint(MVT::ValueType ConstraintVT,
5841 std::string& s) const {
5842 if (MVT::isFloatingPoint(ConstraintVT)) {
5843 if (Subtarget->hasSSE2())
5845 else if (Subtarget->hasSSE1())
5850 return TargetLowering::lowerXConstraint(ConstraintVT, s);
5853 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5854 /// vector. If it is invalid, don't add anything to Ops.
5855 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5857 std::vector<SDOperand>&Ops,
5858 SelectionDAG &DAG) {
5859 SDOperand Result(0, 0);
5861 switch (Constraint) {
5864 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5865 if (C->getValue() <= 31) {
5866 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5872 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5873 if (C->getValue() <= 255) {
5874 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5880 // Literal immediates are always ok.
5881 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5882 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5886 // If we are in non-pic codegen mode, we allow the address of a global (with
5887 // an optional displacement) to be used with 'i'.
5888 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5891 // Match either (GA) or (GA+C)
5893 Offset = GA->getOffset();
5894 } else if (Op.getOpcode() == ISD::ADD) {
5895 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5896 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5898 Offset = GA->getOffset()+C->getValue();
5900 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5901 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5903 Offset = GA->getOffset()+C->getValue();
5910 // If addressing this global requires a load (e.g. in PIC mode), we can't
5912 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5916 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5922 // Otherwise, not valid for this mode.
5928 Ops.push_back(Result);
5931 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5934 std::vector<unsigned> X86TargetLowering::
5935 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5936 MVT::ValueType VT) const {
5937 if (Constraint.size() == 1) {
5938 // FIXME: not handling fp-stack yet!
5939 switch (Constraint[0]) { // GCC X86 Constraint Letters
5940 default: break; // Unknown constraint letter
5941 case 'A': // EAX/EDX
5942 if (VT == MVT::i32 || VT == MVT::i64)
5943 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5945 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5948 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5949 else if (VT == MVT::i16)
5950 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5951 else if (VT == MVT::i8)
5952 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5953 else if (VT == MVT::i64)
5954 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
5959 return std::vector<unsigned>();
5962 std::pair<unsigned, const TargetRegisterClass*>
5963 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5964 MVT::ValueType VT) const {
5965 // First, see if this is a constraint that directly corresponds to an LLVM
5967 if (Constraint.size() == 1) {
5968 // GCC Constraint Letters
5969 switch (Constraint[0]) {
5971 case 'r': // GENERAL_REGS
5972 case 'R': // LEGACY_REGS
5973 case 'l': // INDEX_REGS
5974 if (VT == MVT::i64 && Subtarget->is64Bit())
5975 return std::make_pair(0U, X86::GR64RegisterClass);
5977 return std::make_pair(0U, X86::GR32RegisterClass);
5978 else if (VT == MVT::i16)
5979 return std::make_pair(0U, X86::GR16RegisterClass);
5980 else if (VT == MVT::i8)
5981 return std::make_pair(0U, X86::GR8RegisterClass);
5983 case 'y': // MMX_REGS if MMX allowed.
5984 if (!Subtarget->hasMMX()) break;
5985 return std::make_pair(0U, X86::VR64RegisterClass);
5987 case 'Y': // SSE_REGS if SSE2 allowed
5988 if (!Subtarget->hasSSE2()) break;
5990 case 'x': // SSE_REGS if SSE1 allowed
5991 if (!Subtarget->hasSSE1()) break;
5995 // Scalar SSE types.
5998 return std::make_pair(0U, X86::FR32RegisterClass);
6001 return std::make_pair(0U, X86::FR64RegisterClass);
6009 return std::make_pair(0U, X86::VR128RegisterClass);
6015 // Use the default implementation in TargetLowering to convert the register
6016 // constraint into a member of a register class.
6017 std::pair<unsigned, const TargetRegisterClass*> Res;
6018 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6020 // Not found as a standard register?
6021 if (Res.second == 0) {
6022 // GCC calls "st(0)" just plain "st".
6023 if (StringsEqualNoCase("{st}", Constraint)) {
6024 Res.first = X86::ST0;
6025 Res.second = X86::RFP80RegisterClass;
6031 // Otherwise, check to see if this is a register class of the wrong value
6032 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6033 // turn into {ax},{dx}.
6034 if (Res.second->hasType(VT))
6035 return Res; // Correct type already, nothing to do.
6037 // All of the single-register GCC register classes map their values onto
6038 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6039 // really want an 8-bit or 32-bit register, map to the appropriate register
6040 // class and return the appropriate register.
6041 if (Res.second != X86::GR16RegisterClass)
6044 if (VT == MVT::i8) {
6045 unsigned DestReg = 0;
6046 switch (Res.first) {
6048 case X86::AX: DestReg = X86::AL; break;
6049 case X86::DX: DestReg = X86::DL; break;
6050 case X86::CX: DestReg = X86::CL; break;
6051 case X86::BX: DestReg = X86::BL; break;
6054 Res.first = DestReg;
6055 Res.second = Res.second = X86::GR8RegisterClass;
6057 } else if (VT == MVT::i32) {
6058 unsigned DestReg = 0;
6059 switch (Res.first) {
6061 case X86::AX: DestReg = X86::EAX; break;
6062 case X86::DX: DestReg = X86::EDX; break;
6063 case X86::CX: DestReg = X86::ECX; break;
6064 case X86::BX: DestReg = X86::EBX; break;
6065 case X86::SI: DestReg = X86::ESI; break;
6066 case X86::DI: DestReg = X86::EDI; break;
6067 case X86::BP: DestReg = X86::EBP; break;
6068 case X86::SP: DestReg = X86::ESP; break;
6071 Res.first = DestReg;
6072 Res.second = Res.second = X86::GR32RegisterClass;
6074 } else if (VT == MVT::i64) {
6075 unsigned DestReg = 0;
6076 switch (Res.first) {
6078 case X86::AX: DestReg = X86::RAX; break;
6079 case X86::DX: DestReg = X86::RDX; break;
6080 case X86::CX: DestReg = X86::RCX; break;
6081 case X86::BX: DestReg = X86::RBX; break;
6082 case X86::SI: DestReg = X86::RSI; break;
6083 case X86::DI: DestReg = X86::RDI; break;
6084 case X86::BP: DestReg = X86::RBP; break;
6085 case X86::SP: DestReg = X86::RSP; break;
6088 Res.first = DestReg;
6089 Res.second = Res.second = X86::GR64RegisterClass;