1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source 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/VectorExtras.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/SelectionDAG.h"
33 #include "llvm/CodeGen/SSARegMap.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/StringExtras.h"
39 #include "llvm/ParameterAttributes.h"
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSEf64 = Subtarget->hasSSE2();
46 X86ScalarSSEf32 = Subtarget->hasSSE1();
47 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
50 RegInfo = TM.getRegisterInfo();
52 // Set up the TargetLowering object.
54 // X86 is weird, it always uses i8 for shift amounts and setcc results.
55 setShiftAmountType(MVT::i8);
56 setSetCCResultType(MVT::i8);
57 setSetCCResultContents(ZeroOrOneSetCCResult);
58 setSchedulingPreference(SchedulingForRegPressure);
59 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
60 setStackPointerRegisterToSaveRestore(X86StackPtr);
62 if (Subtarget->isTargetDarwin()) {
63 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
64 setUseUnderscoreSetJmp(false);
65 setUseUnderscoreLongJmp(false);
66 } else if (Subtarget->isTargetMingw()) {
67 // MS runtime is weird: it exports _setjmp, but longjmp!
68 setUseUnderscoreSetJmp(true);
69 setUseUnderscoreLongJmp(false);
71 setUseUnderscoreSetJmp(true);
72 setUseUnderscoreLongJmp(true);
75 // Set up the register classes.
76 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
77 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
78 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
79 if (Subtarget->is64Bit())
80 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
82 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
84 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
86 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
87 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
88 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
90 if (Subtarget->is64Bit()) {
91 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
92 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
95 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
96 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
103 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
104 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
105 // SSE has no i16 to fp conversion, only i32
106 if (X86ScalarSSEf32) {
107 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
108 // f32 and f64 cases are Legal, f80 case is not
109 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
111 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
115 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
116 // are Legal, f80 is custom lowered.
117 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
118 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
120 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
122 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
123 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
125 if (X86ScalarSSEf32) {
126 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
127 // f32 and f64 cases are Legal, f80 case is not
128 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
134 // Handle FP_TO_UINT by promoting the destination to a larger signed
136 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
137 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
138 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
140 if (Subtarget->is64Bit()) {
141 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
142 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
144 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
145 // Expand FP_TO_UINT into a select.
146 // FIXME: We would like to use a Custom expander here eventually to do
147 // the optimal thing for SSE vs. the default expansion in the legalizer.
148 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
150 // With SSE3 we can use fisttpll to convert to a signed i64.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
154 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
155 if (!X86ScalarSSEf64) {
156 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
157 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
160 // Scalar integer multiply, multiply-high, divide, and remainder are
161 // lowered to use operations that produce two results, to match the
162 // available instructions. This exposes the two-result form to trivial
163 // CSE, which is able to combine x/y and x%y into a single instruction,
164 // for example. The single-result multiply instructions are introduced
165 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
167 setOperationAction(ISD::MUL , MVT::i8 , Expand);
168 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
169 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
170 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
171 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
172 setOperationAction(ISD::SREM , MVT::i8 , Expand);
173 setOperationAction(ISD::UREM , MVT::i8 , Expand);
174 setOperationAction(ISD::MUL , MVT::i16 , Expand);
175 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
176 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
177 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
178 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
179 setOperationAction(ISD::SREM , MVT::i16 , Expand);
180 setOperationAction(ISD::UREM , MVT::i16 , Expand);
181 setOperationAction(ISD::MUL , MVT::i32 , Expand);
182 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
183 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
184 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
185 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
186 setOperationAction(ISD::SREM , MVT::i32 , Expand);
187 setOperationAction(ISD::UREM , MVT::i32 , Expand);
188 setOperationAction(ISD::MUL , MVT::i64 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
193 setOperationAction(ISD::SREM , MVT::i64 , Expand);
194 setOperationAction(ISD::UREM , MVT::i64 , Expand);
196 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
197 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
198 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
199 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
200 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
201 if (Subtarget->is64Bit())
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
203 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
204 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
205 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
206 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
207 setOperationAction(ISD::FREM , MVT::f64 , Expand);
209 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
210 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
211 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
212 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
213 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
214 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
215 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
216 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
217 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
220 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
221 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
224 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
225 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
227 // These should be promoted to a larger select which is supported.
228 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
229 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
230 // X86 wants to expand cmov itself.
231 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
232 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
233 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
234 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
235 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
236 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
237 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
238 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
239 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
240 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
241 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
242 if (Subtarget->is64Bit()) {
243 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
244 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
246 // X86 ret instruction may pop stack.
247 setOperationAction(ISD::RET , MVT::Other, Custom);
248 if (!Subtarget->is64Bit())
249 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
252 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
253 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
254 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
255 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
256 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
257 if (Subtarget->is64Bit()) {
258 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
259 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
260 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
261 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
263 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
264 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
265 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
266 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
267 // X86 wants to expand memset / memcpy itself.
268 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
269 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
271 // Use the default ISD::LOCATION expansion.
272 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
273 // FIXME - use subtarget debug flags
274 if (!Subtarget->isTargetDarwin() &&
275 !Subtarget->isTargetELF() &&
276 !Subtarget->isTargetCygMing())
277 setOperationAction(ISD::LABEL, MVT::Other, Expand);
279 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
280 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
281 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
282 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
283 if (Subtarget->is64Bit()) {
285 setExceptionPointerRegister(X86::RAX);
286 setExceptionSelectorRegister(X86::RDX);
288 setExceptionPointerRegister(X86::EAX);
289 setExceptionSelectorRegister(X86::EDX);
291 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
293 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
295 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
296 setOperationAction(ISD::VASTART , MVT::Other, Custom);
297 setOperationAction(ISD::VAARG , MVT::Other, Expand);
298 setOperationAction(ISD::VAEND , MVT::Other, Expand);
299 if (Subtarget->is64Bit())
300 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
302 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
304 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
305 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
306 if (Subtarget->is64Bit())
307 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
308 if (Subtarget->isTargetCygMing())
309 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
311 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
313 if (X86ScalarSSEf64) {
314 // f32 and f64 use SSE.
315 // Set up the FP register classes.
316 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
317 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
319 // Use ANDPD to simulate FABS.
320 setOperationAction(ISD::FABS , MVT::f64, Custom);
321 setOperationAction(ISD::FABS , MVT::f32, Custom);
323 // Use XORP to simulate FNEG.
324 setOperationAction(ISD::FNEG , MVT::f64, Custom);
325 setOperationAction(ISD::FNEG , MVT::f32, Custom);
327 // Use ANDPD and ORPD to simulate FCOPYSIGN.
328 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
329 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
331 // We don't support sin/cos/fmod
332 setOperationAction(ISD::FSIN , MVT::f64, Expand);
333 setOperationAction(ISD::FCOS , MVT::f64, Expand);
334 setOperationAction(ISD::FREM , MVT::f64, Expand);
335 setOperationAction(ISD::FSIN , MVT::f32, Expand);
336 setOperationAction(ISD::FCOS , MVT::f32, Expand);
337 setOperationAction(ISD::FREM , MVT::f32, Expand);
339 // Expand FP immediates into loads from the stack, except for the special
341 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
342 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
343 addLegalFPImmediate(APFloat(+0.0)); // xorpd
344 addLegalFPImmediate(APFloat(+0.0f)); // xorps
346 // Conversions to long double (in X87) go through memory.
347 setConvertAction(MVT::f32, MVT::f80, Expand);
348 setConvertAction(MVT::f64, MVT::f80, Expand);
350 // Conversions from long double (in X87) go through memory.
351 setConvertAction(MVT::f80, MVT::f32, Expand);
352 setConvertAction(MVT::f80, MVT::f64, Expand);
353 } else if (X86ScalarSSEf32) {
354 // Use SSE for f32, x87 for f64.
355 // Set up the FP register classes.
356 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
357 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
359 // Use ANDPS to simulate FABS.
360 setOperationAction(ISD::FABS , MVT::f32, Custom);
362 // Use XORP to simulate FNEG.
363 setOperationAction(ISD::FNEG , MVT::f32, Custom);
365 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
367 // Use ANDPS and ORPS to simulate FCOPYSIGN.
368 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
369 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
371 // We don't support sin/cos/fmod
372 setOperationAction(ISD::FSIN , MVT::f32, Expand);
373 setOperationAction(ISD::FCOS , MVT::f32, Expand);
374 setOperationAction(ISD::FREM , MVT::f32, Expand);
376 // Expand FP immediates into loads from the stack, except for the special
378 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
379 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
380 addLegalFPImmediate(APFloat(+0.0f)); // xorps
381 addLegalFPImmediate(APFloat(+0.0)); // FLD0
382 addLegalFPImmediate(APFloat(+1.0)); // FLD1
383 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
384 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
386 // SSE->x87 conversions go through memory.
387 setConvertAction(MVT::f32, MVT::f64, Expand);
388 setConvertAction(MVT::f32, MVT::f80, Expand);
390 // x87->SSE truncations need to go through memory.
391 setConvertAction(MVT::f80, MVT::f32, Expand);
392 setConvertAction(MVT::f64, MVT::f32, Expand);
393 // And x87->x87 truncations also.
394 setConvertAction(MVT::f80, MVT::f64, Expand);
397 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
398 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
401 // f32 and f64 in x87.
402 // Set up the FP register classes.
403 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
404 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
406 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
407 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
408 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
409 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
411 // Floating truncations need to go through memory.
412 setConvertAction(MVT::f80, MVT::f32, Expand);
413 setConvertAction(MVT::f64, MVT::f32, Expand);
414 setConvertAction(MVT::f80, MVT::f64, Expand);
417 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
418 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
421 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
422 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
423 addLegalFPImmediate(APFloat(+0.0)); // FLD0
424 addLegalFPImmediate(APFloat(+1.0)); // FLD1
425 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
426 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
427 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
428 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
429 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
430 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
433 // Long double always uses X87.
434 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
435 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
436 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
437 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
439 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
440 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
443 // Always use a library call for pow.
444 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
445 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
446 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
448 // First set operation action for all vector types to expand. Then we
449 // will selectively turn on ones that can be effectively codegen'd.
450 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
451 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
452 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
453 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
454 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
455 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
456 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
457 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
458 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
459 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
460 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
461 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
462 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
463 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
464 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
465 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
466 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
467 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
468 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
469 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
470 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
471 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
472 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
473 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
474 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
475 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
476 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
477 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
478 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
479 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
480 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
481 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
482 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
485 if (Subtarget->hasMMX()) {
486 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
487 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
488 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
489 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
491 // FIXME: add MMX packed arithmetics
493 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
494 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
495 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
496 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
498 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
499 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
500 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
501 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
503 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
504 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
506 setOperationAction(ISD::AND, MVT::v8i8, Promote);
507 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
508 setOperationAction(ISD::AND, MVT::v4i16, Promote);
509 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
510 setOperationAction(ISD::AND, MVT::v2i32, Promote);
511 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
512 setOperationAction(ISD::AND, MVT::v1i64, Legal);
514 setOperationAction(ISD::OR, MVT::v8i8, Promote);
515 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
516 setOperationAction(ISD::OR, MVT::v4i16, Promote);
517 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
518 setOperationAction(ISD::OR, MVT::v2i32, Promote);
519 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
520 setOperationAction(ISD::OR, MVT::v1i64, Legal);
522 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
523 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
524 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
525 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
526 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
527 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
528 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
530 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
531 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
532 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
533 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
534 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
535 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
536 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
538 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
539 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
540 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
541 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
543 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
544 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
545 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
546 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
548 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
549 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
550 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
551 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
554 if (Subtarget->hasSSE1()) {
555 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
557 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
558 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
559 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
560 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
561 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
562 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
563 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
564 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
565 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
567 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
570 if (Subtarget->hasSSE2()) {
571 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
572 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
573 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
574 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
575 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
577 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
578 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
579 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
580 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
581 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
582 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
583 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
584 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
585 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
586 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
587 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
588 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
589 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
590 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
591 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
593 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
594 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
595 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
596 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
597 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
598 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
600 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
601 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
602 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
603 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
604 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
606 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
607 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
608 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
609 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
610 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
611 if (Subtarget->is64Bit())
612 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
614 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
615 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
616 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
617 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
618 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
619 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
620 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
621 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
622 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
623 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
624 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
625 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
628 // Custom lower v2i64 and v2f64 selects.
629 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
630 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
631 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
632 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
635 // We want to custom lower some of our intrinsics.
636 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
638 // We have target-specific dag combine patterns for the following nodes:
639 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
640 setTargetDAGCombine(ISD::SELECT);
642 computeRegisterProperties();
644 // FIXME: These should be based on subtarget info. Plus, the values should
645 // be smaller when we are in optimizing for size mode.
646 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
647 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
648 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
649 allowUnalignedMemoryAccesses = true; // x86 supports it!
653 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
655 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
656 SelectionDAG &DAG) const {
657 if (usesGlobalOffsetTable())
658 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
659 if (!Subtarget->isPICStyleRIPRel())
660 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
664 //===----------------------------------------------------------------------===//
665 // Return Value Calling Convention Implementation
666 //===----------------------------------------------------------------------===//
668 #include "X86GenCallingConv.inc"
670 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
671 /// exists skip possible ISD:TokenFactor.
672 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
673 if (Chain.getOpcode()==X86ISD::TAILCALL) {
675 } else if (Chain.getOpcode()==ISD::TokenFactor) {
676 if (Chain.getNumOperands() &&
677 Chain.getOperand(0).getOpcode()==X86ISD::TAILCALL)
678 return Chain.getOperand(0);
683 /// LowerRET - Lower an ISD::RET node.
684 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
685 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
687 SmallVector<CCValAssign, 16> RVLocs;
688 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
689 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
690 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
691 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
693 // If this is the first return lowered for this function, add the regs to the
694 // liveout set for the function.
695 if (DAG.getMachineFunction().liveout_empty()) {
696 for (unsigned i = 0; i != RVLocs.size(); ++i)
697 if (RVLocs[i].isRegLoc())
698 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
700 SDOperand Chain = Op.getOperand(0);
702 // Handle tail call return.
703 Chain = GetPossiblePreceedingTailCall(Chain);
704 if (Chain.getOpcode() == X86ISD::TAILCALL) {
705 SDOperand TailCall = Chain;
706 SDOperand TargetAddress = TailCall.getOperand(1);
707 SDOperand StackAdjustment = TailCall.getOperand(2);
708 assert ( ((TargetAddress.getOpcode() == ISD::Register &&
709 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
710 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
711 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
712 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
713 "Expecting an global address, external symbol, or register");
714 assert( StackAdjustment.getOpcode() == ISD::Constant &&
715 "Expecting a const value");
717 SmallVector<SDOperand,8> Operands;
718 Operands.push_back(Chain.getOperand(0));
719 Operands.push_back(TargetAddress);
720 Operands.push_back(StackAdjustment);
721 // Copy registers used by the call. Last operand is a flag so it is not
723 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
724 Operands.push_back(Chain.getOperand(i));
726 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
733 // Copy the result values into the output registers.
734 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
735 RVLocs[0].getLocReg() != X86::ST0) {
736 for (unsigned i = 0; i != RVLocs.size(); ++i) {
737 CCValAssign &VA = RVLocs[i];
738 assert(VA.isRegLoc() && "Can only return in registers!");
739 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
741 Flag = Chain.getValue(1);
744 // We need to handle a destination of ST0 specially, because it isn't really
746 SDOperand Value = Op.getOperand(1);
748 // If this is an FP return with ScalarSSE, we need to move the value from
749 // an XMM register onto the fp-stack.
750 if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
751 (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
754 // If this is a load into a scalarsse value, don't store the loaded value
755 // back to the stack, only to reload it: just replace the scalar-sse load.
756 if (ISD::isNON_EXTLoad(Value.Val) &&
757 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
758 Chain = Value.getOperand(0);
759 MemLoc = Value.getOperand(1);
761 // Spill the value to memory and reload it into top of stack.
762 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
763 MachineFunction &MF = DAG.getMachineFunction();
764 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
765 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
766 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
768 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
769 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
770 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
771 Chain = Value.getValue(1);
774 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
775 SDOperand Ops[] = { Chain, Value };
776 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
777 Flag = Chain.getValue(1);
780 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
782 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
784 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
788 /// LowerCallResult - Lower the result values of an ISD::CALL into the
789 /// appropriate copies out of appropriate physical registers. This assumes that
790 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
791 /// being lowered. The returns a SDNode with the same number of values as the
793 SDNode *X86TargetLowering::
794 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
795 unsigned CallingConv, SelectionDAG &DAG) {
797 // Assign locations to each value returned by this call.
798 SmallVector<CCValAssign, 16> RVLocs;
799 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
800 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
801 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
804 SmallVector<SDOperand, 8> ResultVals;
806 // Copy all of the result registers out of their specified physreg.
807 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
808 for (unsigned i = 0; i != RVLocs.size(); ++i) {
809 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
810 RVLocs[i].getValVT(), InFlag).getValue(1);
811 InFlag = Chain.getValue(2);
812 ResultVals.push_back(Chain.getValue(0));
815 // Copies from the FP stack are special, as ST0 isn't a valid register
816 // before the fp stackifier runs.
818 // Copy ST0 into an RFP register with FP_GET_RESULT.
819 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
820 SDOperand GROps[] = { Chain, InFlag };
821 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
822 Chain = RetVal.getValue(1);
823 InFlag = RetVal.getValue(2);
825 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
827 if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
828 (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
829 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
830 // shouldn't be necessary except that RFP cannot be live across
831 // multiple blocks. When stackifier is fixed, they can be uncoupled.
832 MachineFunction &MF = DAG.getMachineFunction();
833 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
834 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
836 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
838 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
839 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
840 Chain = RetVal.getValue(1);
842 ResultVals.push_back(RetVal);
845 // Merge everything together with a MERGE_VALUES node.
846 ResultVals.push_back(Chain);
847 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
848 &ResultVals[0], ResultVals.size()).Val;
852 //===----------------------------------------------------------------------===//
853 // C & StdCall & Fast Calling Convention implementation
854 //===----------------------------------------------------------------------===//
855 // StdCall calling convention seems to be standard for many Windows' API
856 // routines and around. It differs from C calling convention just a little:
857 // callee should clean up the stack, not caller. Symbols should be also
858 // decorated in some fancy way :) It doesn't support any vector arguments.
859 // For info on fast calling convention see Fast Calling Convention (tail call)
860 // implementation LowerX86_32FastCCCallTo.
862 /// AddLiveIn - This helper function adds the specified physical register to the
863 /// MachineFunction as a live in value. It also creates a corresponding virtual
865 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
866 const TargetRegisterClass *RC) {
867 assert(RC->contains(PReg) && "Not the correct regclass!");
868 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
869 MF.addLiveIn(PReg, VReg);
873 // align stack arguments according to platform alignment needed for tail calls
874 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG& DAG);
876 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
877 const CCValAssign &VA,
878 MachineFrameInfo *MFI,
879 SDOperand Root, unsigned i) {
880 // Create the nodes corresponding to a load from this parameter slot.
881 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
882 VA.getLocMemOffset());
883 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
885 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
887 if (Flags & ISD::ParamFlags::ByVal)
890 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
893 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
895 unsigned NumArgs = Op.Val->getNumValues() - 1;
896 MachineFunction &MF = DAG.getMachineFunction();
897 MachineFrameInfo *MFI = MF.getFrameInfo();
898 SDOperand Root = Op.getOperand(0);
899 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
900 unsigned CC = MF.getFunction()->getCallingConv();
901 // Assign locations to all of the incoming arguments.
902 SmallVector<CCValAssign, 16> ArgLocs;
903 CCState CCInfo(CC, isVarArg,
904 getTargetMachine(), ArgLocs);
905 // Check for possible tail call calling convention.
906 if (CC == CallingConv::Fast && PerformTailCallOpt)
907 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_TailCall);
909 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
911 SmallVector<SDOperand, 8> ArgValues;
912 unsigned LastVal = ~0U;
913 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
914 CCValAssign &VA = ArgLocs[i];
915 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
917 assert(VA.getValNo() != LastVal &&
918 "Don't support value assigned to multiple locs yet");
919 LastVal = VA.getValNo();
922 MVT::ValueType RegVT = VA.getLocVT();
923 TargetRegisterClass *RC;
924 if (RegVT == MVT::i32)
925 RC = X86::GR32RegisterClass;
927 assert(MVT::isVector(RegVT));
928 RC = X86::VR128RegisterClass;
931 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
932 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
934 // If this is an 8 or 16-bit value, it is really passed promoted to 32
935 // bits. Insert an assert[sz]ext to capture this, then truncate to the
937 if (VA.getLocInfo() == CCValAssign::SExt)
938 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
939 DAG.getValueType(VA.getValVT()));
940 else if (VA.getLocInfo() == CCValAssign::ZExt)
941 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
942 DAG.getValueType(VA.getValVT()));
944 if (VA.getLocInfo() != CCValAssign::Full)
945 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
947 ArgValues.push_back(ArgValue);
949 assert(VA.isMemLoc());
950 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
954 unsigned StackSize = CCInfo.getNextStackOffset();
955 // align stack specially for tail calls
956 if (CC==CallingConv::Fast)
957 StackSize = GetAlignedArgumentStackSize(StackSize,DAG);
959 ArgValues.push_back(Root);
961 // If the function takes variable number of arguments, make a frame index for
962 // the start of the first vararg value... for expansion of llvm.va_start.
964 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
966 // Tail call calling convention (CallingConv::Fast) does not support varargs.
967 assert( !(isVarArg && CC == CallingConv::Fast) &&
968 "CallingConv::Fast does not support varargs.");
970 if (isStdCall && !isVarArg &&
971 (CC==CallingConv::Fast && PerformTailCallOpt || CC!=CallingConv::Fast)) {
972 BytesToPopOnReturn = StackSize; // Callee pops everything..
973 BytesCallerReserves = 0;
975 BytesToPopOnReturn = 0; // Callee pops nothing.
977 // If this is an sret function, the return should pop the hidden pointer.
979 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
980 ISD::ParamFlags::StructReturn))
981 BytesToPopOnReturn = 4;
983 BytesCallerReserves = StackSize;
986 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
988 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
989 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
991 // Return the new list of results.
992 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
993 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
996 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
998 SDOperand Chain = Op.getOperand(0);
999 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1000 SDOperand Callee = Op.getOperand(4);
1001 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1003 // Analyze operands of the call, assigning locations to each operand.
1004 SmallVector<CCValAssign, 16> ArgLocs;
1005 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1006 if(CC==CallingConv::Fast && PerformTailCallOpt)
1007 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1009 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
1011 // Get a count of how many bytes are to be pushed on the stack.
1012 unsigned NumBytes = CCInfo.getNextStackOffset();
1013 if (CC==CallingConv::Fast)
1014 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1016 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1018 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1019 SmallVector<SDOperand, 8> MemOpChains;
1023 // Walk the register/memloc assignments, inserting copies/loads.
1024 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1025 CCValAssign &VA = ArgLocs[i];
1026 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1028 // Promote the value if needed.
1029 switch (VA.getLocInfo()) {
1030 default: assert(0 && "Unknown loc info!");
1031 case CCValAssign::Full: break;
1032 case CCValAssign::SExt:
1033 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1035 case CCValAssign::ZExt:
1036 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1038 case CCValAssign::AExt:
1039 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1043 if (VA.isRegLoc()) {
1044 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1046 assert(VA.isMemLoc());
1047 if (StackPtr.Val == 0)
1048 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1050 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1055 // If the first argument is an sret pointer, remember it.
1056 bool isSRet = NumOps &&
1057 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
1058 ISD::ParamFlags::StructReturn);
1060 if (!MemOpChains.empty())
1061 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1062 &MemOpChains[0], MemOpChains.size());
1064 // Build a sequence of copy-to-reg nodes chained together with token chain
1065 // and flag operands which copy the outgoing args into registers.
1067 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1068 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1070 InFlag = Chain.getValue(1);
1073 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1075 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1076 Subtarget->isPICStyleGOT()) {
1077 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1078 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1080 InFlag = Chain.getValue(1);
1083 // If the callee is a GlobalAddress node (quite common, every direct call is)
1084 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1085 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1086 // We should use extra load for direct calls to dllimported functions in
1088 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1089 getTargetMachine(), true))
1090 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1091 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1092 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1094 // Returns a chain & a flag for retval copy to use.
1095 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1096 SmallVector<SDOperand, 8> Ops;
1097 Ops.push_back(Chain);
1098 Ops.push_back(Callee);
1100 // Add argument registers to the end of the list so that they are known live
1102 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1103 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1104 RegsToPass[i].second.getValueType()));
1106 // Add an implicit use GOT pointer in EBX.
1107 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1108 Subtarget->isPICStyleGOT())
1109 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1112 Ops.push_back(InFlag);
1114 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1115 InFlag = Chain.getValue(1);
1117 // Create the CALLSEQ_END node.
1118 unsigned NumBytesForCalleeToPush = 0;
1120 if (CC == CallingConv::X86_StdCall ||
1121 (CC == CallingConv::Fast && PerformTailCallOpt)) {
1123 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1125 NumBytesForCalleeToPush = NumBytes;
1126 assert(!(isVarArg && CC==CallingConv::Fast) &&
1127 "CallingConv::Fast does not support varargs.");
1129 // If this is is a call to a struct-return function, the callee
1130 // pops the hidden struct pointer, so we have to push it back.
1131 // This is common for Darwin/X86, Linux & Mingw32 targets.
1132 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1135 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1137 Ops.push_back(Chain);
1138 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1139 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
1140 Ops.push_back(InFlag);
1141 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1142 InFlag = Chain.getValue(1);
1144 // Handle result values, copying them out of physregs into vregs that we
1146 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1150 //===----------------------------------------------------------------------===//
1151 // FastCall Calling Convention implementation
1152 //===----------------------------------------------------------------------===//
1154 // The X86 'fastcall' calling convention passes up to two integer arguments in
1155 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1156 // and requires that the callee pop its arguments off the stack (allowing proper
1157 // tail calls), and has the same return value conventions as C calling convs.
1159 // This calling convention always arranges for the callee pop value to be 8n+4
1160 // bytes, which is needed for tail recursion elimination and stack alignment
1163 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
1164 MachineFunction &MF = DAG.getMachineFunction();
1165 MachineFrameInfo *MFI = MF.getFrameInfo();
1166 SDOperand Root = Op.getOperand(0);
1167 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1169 // Assign locations to all of the incoming arguments.
1170 SmallVector<CCValAssign, 16> ArgLocs;
1171 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1172 getTargetMachine(), ArgLocs);
1173 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
1175 SmallVector<SDOperand, 8> ArgValues;
1176 unsigned LastVal = ~0U;
1177 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1178 CCValAssign &VA = ArgLocs[i];
1179 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1181 assert(VA.getValNo() != LastVal &&
1182 "Don't support value assigned to multiple locs yet");
1183 LastVal = VA.getValNo();
1185 if (VA.isRegLoc()) {
1186 MVT::ValueType RegVT = VA.getLocVT();
1187 TargetRegisterClass *RC;
1188 if (RegVT == MVT::i32)
1189 RC = X86::GR32RegisterClass;
1191 assert(MVT::isVector(RegVT));
1192 RC = X86::VR128RegisterClass;
1195 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1196 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1198 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1199 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1201 if (VA.getLocInfo() == CCValAssign::SExt)
1202 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1203 DAG.getValueType(VA.getValVT()));
1204 else if (VA.getLocInfo() == CCValAssign::ZExt)
1205 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1206 DAG.getValueType(VA.getValVT()));
1208 if (VA.getLocInfo() != CCValAssign::Full)
1209 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1211 ArgValues.push_back(ArgValue);
1213 assert(VA.isMemLoc());
1214 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1218 ArgValues.push_back(Root);
1220 unsigned StackSize = CCInfo.getNextStackOffset();
1222 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1223 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1224 // arguments and the arguments after the retaddr has been pushed are
1226 if ((StackSize & 7) == 0)
1230 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1231 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1232 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1233 BytesCallerReserves = 0;
1235 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1236 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1238 // Return the new list of results.
1239 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1240 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1244 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1245 const SDOperand &StackPtr,
1246 const CCValAssign &VA,
1249 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1250 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1251 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1252 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1253 if (Flags & ISD::ParamFlags::ByVal) {
1254 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1255 ISD::ParamFlags::ByValAlignOffs);
1257 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1258 ISD::ParamFlags::ByValSizeOffs;
1260 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1261 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1262 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i1);
1264 return DAG.getMemcpy(Chain, PtrOff, Arg, SizeNode, AlignNode,
1267 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1271 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1273 SDOperand Chain = Op.getOperand(0);
1274 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1275 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1276 SDOperand Callee = Op.getOperand(4);
1278 // Analyze operands of the call, assigning locations to each operand.
1279 SmallVector<CCValAssign, 16> ArgLocs;
1280 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1281 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1283 // Get a count of how many bytes are to be pushed on the stack.
1284 unsigned NumBytes = CCInfo.getNextStackOffset();
1286 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1287 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1288 // arguments and the arguments after the retaddr has been pushed are
1290 if ((NumBytes & 7) == 0)
1294 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1296 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1297 SmallVector<SDOperand, 8> MemOpChains;
1301 // Walk the register/memloc assignments, inserting copies/loads.
1302 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1303 CCValAssign &VA = ArgLocs[i];
1304 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1306 // Promote the value if needed.
1307 switch (VA.getLocInfo()) {
1308 default: assert(0 && "Unknown loc info!");
1309 case CCValAssign::Full: break;
1310 case CCValAssign::SExt:
1311 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1313 case CCValAssign::ZExt:
1314 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1316 case CCValAssign::AExt:
1317 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1321 if (VA.isRegLoc()) {
1322 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1324 assert(VA.isMemLoc());
1325 if (StackPtr.Val == 0)
1326 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1328 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1333 if (!MemOpChains.empty())
1334 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1335 &MemOpChains[0], MemOpChains.size());
1337 // Build a sequence of copy-to-reg nodes chained together with token chain
1338 // and flag operands which copy the outgoing args into registers.
1340 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1341 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1343 InFlag = Chain.getValue(1);
1346 // If the callee is a GlobalAddress node (quite common, every direct call is)
1347 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1348 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1349 // We should use extra load for direct calls to dllimported functions in
1351 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1352 getTargetMachine(), true))
1353 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1354 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1355 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1357 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1359 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1360 Subtarget->isPICStyleGOT()) {
1361 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1362 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1364 InFlag = Chain.getValue(1);
1367 // Returns a chain & a flag for retval copy to use.
1368 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1369 SmallVector<SDOperand, 8> Ops;
1370 Ops.push_back(Chain);
1371 Ops.push_back(Callee);
1373 // Add argument registers to the end of the list so that they are known live
1375 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1376 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1377 RegsToPass[i].second.getValueType()));
1379 // Add an implicit use GOT pointer in EBX.
1380 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1381 Subtarget->isPICStyleGOT())
1382 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1385 Ops.push_back(InFlag);
1387 assert(isTailCall==false && "no tail call here");
1388 Chain = DAG.getNode(X86ISD::CALL,
1389 NodeTys, &Ops[0], Ops.size());
1390 InFlag = Chain.getValue(1);
1392 // Returns a flag for retval copy to use.
1393 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1395 Ops.push_back(Chain);
1396 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1397 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1398 Ops.push_back(InFlag);
1399 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1400 InFlag = Chain.getValue(1);
1402 // Handle result values, copying them out of physregs into vregs that we
1404 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1407 //===----------------------------------------------------------------------===//
1408 // Fast Calling Convention (tail call) implementation
1409 //===----------------------------------------------------------------------===//
1411 // Like std call, callee cleans arguments, convention except that ECX is
1412 // reserved for storing the tail called function address. Only 2 registers are
1413 // free for argument passing (inreg). Tail call optimization is performed
1415 // * tailcallopt is enabled
1416 // * caller/callee are fastcc
1417 // * elf/pic is disabled OR
1418 // * elf/pic enabled + callee is in module + callee has
1419 // visibility protected or hidden
1420 // To keep the stack aligned according to platform abi the function
1421 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1422 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1423 // If a tail called function callee has more arguments than the caller the
1424 // caller needs to make sure that there is room to move the RETADDR to. This is
1425 // achieved by reserving an area the size of the argument delta right after the
1426 // original REtADDR, but before the saved framepointer or the spilled registers
1427 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1439 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1440 /// for a 16 byte align requirement.
1441 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1442 SelectionDAG& DAG) {
1443 if (PerformTailCallOpt) {
1444 MachineFunction &MF = DAG.getMachineFunction();
1445 const TargetMachine &TM = MF.getTarget();
1446 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1447 unsigned StackAlignment = TFI.getStackAlignment();
1448 uint64_t AlignMask = StackAlignment - 1;
1449 int64_t Offset = StackSize;
1450 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1451 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1452 // Number smaller than 12 so just add the difference.
1453 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1455 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1456 Offset = ((~AlignMask) & Offset) + StackAlignment +
1457 (StackAlignment-SlotSize);
1464 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1465 /// following the call is a return. A function is eligible if caller/callee
1466 /// calling conventions match, currently only fastcc supports tail calls, and
1467 /// the function CALL is immediatly followed by a RET.
1468 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1470 SelectionDAG& DAG) const {
1471 if (!PerformTailCallOpt)
1474 // Check whether CALL node immediatly preceeds the RET node and whether the
1475 // return uses the result of the node or is a void return.
1476 unsigned NumOps = Ret.getNumOperands();
1478 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1479 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1481 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1482 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1483 MachineFunction &MF = DAG.getMachineFunction();
1484 unsigned CallerCC = MF.getFunction()->getCallingConv();
1485 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1486 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1487 SDOperand Callee = Call.getOperand(4);
1488 // On elf/pic %ebx needs to be livein.
1489 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1490 !Subtarget->isPICStyleGOT())
1493 // Can only do local tail calls with PIC.
1494 GlobalValue * GV = 0;
1495 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1497 (GV = G->getGlobal()) &&
1498 (GV->hasHiddenVisibility() || GV->hasProtectedVisibility()))
1506 SDOperand X86TargetLowering::LowerX86_TailCallTo(SDOperand Op,
1509 SDOperand Chain = Op.getOperand(0);
1510 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1511 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1512 SDOperand Callee = Op.getOperand(4);
1513 bool is64Bit = Subtarget->is64Bit();
1515 assert(isTailCall && PerformTailCallOpt && "Should only emit tail calls.");
1517 // Analyze operands of the call, assigning locations to each operand.
1518 SmallVector<CCValAssign, 16> ArgLocs;
1519 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1521 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1523 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1526 // Lower arguments at fp - stackoffset + fpdiff.
1527 MachineFunction &MF = DAG.getMachineFunction();
1529 unsigned NumBytesToBePushed =
1530 GetAlignedArgumentStackSize(CCInfo.getNextStackOffset(), DAG);
1532 unsigned NumBytesCallerPushed =
1533 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1534 int FPDiff = NumBytesCallerPushed - NumBytesToBePushed;
1536 // Set the delta of movement of the returnaddr stackslot.
1537 // But only set if delta is greater than previous delta.
1538 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1539 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1542 getCALLSEQ_START(Chain, DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1544 // Adjust the Return address stack slot.
1545 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1547 MVT::ValueType VT = is64Bit ? MVT::i64 : MVT::i32;
1548 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1549 // Load the "old" Return address.
1551 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1552 // Calculate the new stack slot for the return address.
1553 int SlotSize = is64Bit ? 8 : 4;
1554 int NewReturnAddrFI =
1555 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1556 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1557 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1560 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1561 SmallVector<SDOperand, 8> MemOpChains;
1562 SmallVector<SDOperand, 8> MemOpChains2;
1563 SDOperand FramePtr, StackPtr;
1568 // Walk the register/memloc assignments, inserting copies/loads. Lower
1569 // arguments first to the stack slot where they would normally - in case of a
1570 // normal function call - be.
1571 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1572 CCValAssign &VA = ArgLocs[i];
1573 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1575 // Promote the value if needed.
1576 switch (VA.getLocInfo()) {
1577 default: assert(0 && "Unknown loc info!");
1578 case CCValAssign::Full: break;
1579 case CCValAssign::SExt:
1580 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1582 case CCValAssign::ZExt:
1583 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1585 case CCValAssign::AExt:
1586 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1590 if (VA.isRegLoc()) {
1591 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1593 assert(VA.isMemLoc());
1594 if (StackPtr.Val == 0)
1595 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1597 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1602 if (!MemOpChains.empty())
1603 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1604 &MemOpChains[0], MemOpChains.size());
1606 // Build a sequence of copy-to-reg nodes chained together with token chain
1607 // and flag operands which copy the outgoing args into registers.
1609 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1610 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1612 InFlag = Chain.getValue(1);
1614 InFlag = SDOperand();
1616 // Copy from stack slots to stack slot of a tail called function. This needs
1617 // to be done because if we would lower the arguments directly to their real
1618 // stack slot we might end up overwriting each other.
1619 // TODO: To make this more efficient (sometimes saving a store/load) we could
1620 // analyse the arguments and emit this store/load/store sequence only for
1621 // arguments which would be overwritten otherwise.
1622 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1623 CCValAssign &VA = ArgLocs[i];
1624 if (!VA.isRegLoc()) {
1625 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1626 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1628 // Get source stack slot.
1629 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1630 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1631 // Create frame index.
1632 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1633 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1634 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1635 FIN = DAG.getFrameIndex(FI, MVT::i32);
1636 if (Flags & ISD::ParamFlags::ByVal) {
1637 // Copy relative to framepointer.
1638 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1639 ISD::ParamFlags::ByValAlignOffs);
1641 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1642 ISD::ParamFlags::ByValSizeOffs;
1644 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1645 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1646 // Copy relative to framepointer.
1647 MemOpChains2.push_back(DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, FIN,
1648 PtrOff, SizeNode, AlignNode));
1650 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1651 // Store relative to framepointer.
1652 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1657 if (!MemOpChains2.empty())
1658 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1659 &MemOpChains2[0], MemOpChains.size());
1661 // Store the return address to the appropriate stack slot.
1663 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1665 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1667 // Does not work with tail call since ebx is not restored correctly by
1668 // tailcaller. TODO: at least for x86 - verify for x86-64
1670 // If the callee is a GlobalAddress node (quite common, every direct call is)
1671 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1672 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1673 // We should use extra load for direct calls to dllimported functions in
1675 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1676 getTargetMachine(), true))
1677 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1678 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1679 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1681 assert(Callee.getOpcode() == ISD::LOAD &&
1682 "Function destination must be loaded into virtual register");
1683 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1685 Chain = DAG.getCopyToReg(Chain,
1686 DAG.getRegister(Opc, getPointerTy()) ,
1688 Callee = DAG.getRegister(Opc, getPointerTy());
1689 // Add register as live out.
1690 DAG.getMachineFunction().addLiveOut(Opc);
1693 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1694 SmallVector<SDOperand, 8> Ops;
1696 Ops.push_back(Chain);
1697 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1698 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1700 Ops.push_back(InFlag);
1701 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1702 InFlag = Chain.getValue(1);
1704 // Returns a chain & a flag for retval copy to use.
1705 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1707 Ops.push_back(Chain);
1708 Ops.push_back(Callee);
1709 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1710 // Add argument registers to the end of the list so that they are known live
1712 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1713 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1714 RegsToPass[i].second.getValueType()));
1716 Ops.push_back(InFlag);
1717 assert(InFlag.Val &&
1718 "Flag must be set. Depend on flag being set in LowerRET");
1719 Chain = DAG.getNode(X86ISD::TAILCALL,
1720 Op.Val->getVTList(), &Ops[0], Ops.size());
1722 return SDOperand(Chain.Val, Op.ResNo);
1725 //===----------------------------------------------------------------------===//
1726 // X86-64 C Calling Convention implementation
1727 //===----------------------------------------------------------------------===//
1730 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1731 MachineFunction &MF = DAG.getMachineFunction();
1732 MachineFrameInfo *MFI = MF.getFrameInfo();
1733 SDOperand Root = Op.getOperand(0);
1734 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1735 unsigned CC= MF.getFunction()->getCallingConv();
1737 static const unsigned GPR64ArgRegs[] = {
1738 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1740 static const unsigned XMMArgRegs[] = {
1741 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1742 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1746 // Assign locations to all of the incoming arguments.
1747 SmallVector<CCValAssign, 16> ArgLocs;
1748 CCState CCInfo(CC, isVarArg,
1749 getTargetMachine(), ArgLocs);
1750 if (CC == CallingConv::Fast && PerformTailCallOpt)
1751 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1753 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1755 SmallVector<SDOperand, 8> ArgValues;
1756 unsigned LastVal = ~0U;
1757 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1758 CCValAssign &VA = ArgLocs[i];
1759 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1761 assert(VA.getValNo() != LastVal &&
1762 "Don't support value assigned to multiple locs yet");
1763 LastVal = VA.getValNo();
1765 if (VA.isRegLoc()) {
1766 MVT::ValueType RegVT = VA.getLocVT();
1767 TargetRegisterClass *RC;
1768 if (RegVT == MVT::i32)
1769 RC = X86::GR32RegisterClass;
1770 else if (RegVT == MVT::i64)
1771 RC = X86::GR64RegisterClass;
1772 else if (RegVT == MVT::f32)
1773 RC = X86::FR32RegisterClass;
1774 else if (RegVT == MVT::f64)
1775 RC = X86::FR64RegisterClass;
1777 assert(MVT::isVector(RegVT));
1778 if (MVT::getSizeInBits(RegVT) == 64) {
1779 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1782 RC = X86::VR128RegisterClass;
1785 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1786 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1788 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1789 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1791 if (VA.getLocInfo() == CCValAssign::SExt)
1792 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1793 DAG.getValueType(VA.getValVT()));
1794 else if (VA.getLocInfo() == CCValAssign::ZExt)
1795 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1796 DAG.getValueType(VA.getValVT()));
1798 if (VA.getLocInfo() != CCValAssign::Full)
1799 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1801 // Handle MMX values passed in GPRs.
1802 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1803 MVT::getSizeInBits(RegVT) == 64)
1804 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1806 ArgValues.push_back(ArgValue);
1808 assert(VA.isMemLoc());
1809 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1813 unsigned StackSize = CCInfo.getNextStackOffset();
1814 if (CC==CallingConv::Fast)
1815 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1817 // If the function takes variable number of arguments, make a frame index for
1818 // the start of the first vararg value... for expansion of llvm.va_start.
1820 assert(CC!=CallingConv::Fast
1821 && "Var arg not supported with calling convention fastcc");
1822 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1823 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1825 // For X86-64, if there are vararg parameters that are passed via
1826 // registers, then we must store them to their spots on the stack so they
1827 // may be loaded by deferencing the result of va_next.
1828 VarArgsGPOffset = NumIntRegs * 8;
1829 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1830 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1831 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1833 // Store the integer parameter registers.
1834 SmallVector<SDOperand, 8> MemOps;
1835 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1836 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1837 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1838 for (; NumIntRegs != 6; ++NumIntRegs) {
1839 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1840 X86::GR64RegisterClass);
1841 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1842 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1843 MemOps.push_back(Store);
1844 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1845 DAG.getConstant(8, getPointerTy()));
1848 // Now store the XMM (fp + vector) parameter registers.
1849 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1850 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1851 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1852 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1853 X86::VR128RegisterClass);
1854 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1855 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1856 MemOps.push_back(Store);
1857 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1858 DAG.getConstant(16, getPointerTy()));
1860 if (!MemOps.empty())
1861 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1862 &MemOps[0], MemOps.size());
1865 ArgValues.push_back(Root);
1866 // Tail call convention (fastcc) needs callee pop.
1867 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1868 BytesToPopOnReturn = StackSize; // Callee pops everything.
1869 BytesCallerReserves = 0;
1871 BytesToPopOnReturn = 0; // Callee pops nothing.
1872 BytesCallerReserves = StackSize;
1874 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1875 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1877 // Return the new list of results.
1878 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1879 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1883 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1885 SDOperand Chain = Op.getOperand(0);
1886 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1887 SDOperand Callee = Op.getOperand(4);
1889 // Analyze operands of the call, assigning locations to each operand.
1890 SmallVector<CCValAssign, 16> ArgLocs;
1891 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1892 if (CC==CallingConv::Fast && PerformTailCallOpt)
1893 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1895 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1897 // Get a count of how many bytes are to be pushed on the stack.
1898 unsigned NumBytes = CCInfo.getNextStackOffset();
1899 if (CC == CallingConv::Fast)
1900 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1902 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1904 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1905 SmallVector<SDOperand, 8> MemOpChains;
1909 // Walk the register/memloc assignments, inserting copies/loads.
1910 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1911 CCValAssign &VA = ArgLocs[i];
1912 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1914 // Promote the value if needed.
1915 switch (VA.getLocInfo()) {
1916 default: assert(0 && "Unknown loc info!");
1917 case CCValAssign::Full: break;
1918 case CCValAssign::SExt:
1919 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1921 case CCValAssign::ZExt:
1922 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1924 case CCValAssign::AExt:
1925 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1929 if (VA.isRegLoc()) {
1930 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1932 assert(VA.isMemLoc());
1933 if (StackPtr.Val == 0)
1934 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1936 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1941 if (!MemOpChains.empty())
1942 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1943 &MemOpChains[0], MemOpChains.size());
1945 // Build a sequence of copy-to-reg nodes chained together with token chain
1946 // and flag operands which copy the outgoing args into registers.
1948 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1949 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1951 InFlag = Chain.getValue(1);
1955 assert ( CallingConv::Fast != CC &&
1956 "Var args not supported with calling convention fastcc");
1958 // From AMD64 ABI document:
1959 // For calls that may call functions that use varargs or stdargs
1960 // (prototype-less calls or calls to functions containing ellipsis (...) in
1961 // the declaration) %al is used as hidden argument to specify the number
1962 // of SSE registers used. The contents of %al do not need to match exactly
1963 // the number of registers, but must be an ubound on the number of SSE
1964 // registers used and is in the range 0 - 8 inclusive.
1966 // Count the number of XMM registers allocated.
1967 static const unsigned XMMArgRegs[] = {
1968 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1969 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1971 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1973 Chain = DAG.getCopyToReg(Chain, X86::AL,
1974 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1975 InFlag = Chain.getValue(1);
1978 // If the callee is a GlobalAddress node (quite common, every direct call is)
1979 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1980 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1981 // We should use extra load for direct calls to dllimported functions in
1983 if (getTargetMachine().getCodeModel() != CodeModel::Large
1984 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1985 getTargetMachine(), true))
1986 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1987 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1988 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1989 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1991 // Returns a chain & a flag for retval copy to use.
1992 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1993 SmallVector<SDOperand, 8> Ops;
1994 Ops.push_back(Chain);
1995 Ops.push_back(Callee);
1997 // Add argument registers to the end of the list so that they are known live
1999 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2000 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2001 RegsToPass[i].second.getValueType()));
2004 Ops.push_back(InFlag);
2006 Chain = DAG.getNode(X86ISD::CALL,
2007 NodeTys, &Ops[0], Ops.size());
2008 InFlag = Chain.getValue(1);
2009 int NumBytesForCalleeToPush = 0;
2010 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2011 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2013 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2015 // Returns a flag for retval copy to use.
2016 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2018 Ops.push_back(Chain);
2019 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2020 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2021 Ops.push_back(InFlag);
2022 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2023 InFlag = Chain.getValue(1);
2025 // Handle result values, copying them out of physregs into vregs that we
2027 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2031 //===----------------------------------------------------------------------===//
2032 // Other Lowering Hooks
2033 //===----------------------------------------------------------------------===//
2036 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2037 MachineFunction &MF = DAG.getMachineFunction();
2038 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2039 int ReturnAddrIndex = FuncInfo->getRAIndex();
2041 if (ReturnAddrIndex == 0) {
2042 // Set up a frame object for the return address.
2043 if (Subtarget->is64Bit())
2044 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2046 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2048 FuncInfo->setRAIndex(ReturnAddrIndex);
2051 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2056 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2057 /// specific condition code. It returns a false if it cannot do a direct
2058 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2060 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2061 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2062 SelectionDAG &DAG) {
2063 X86CC = X86::COND_INVALID;
2065 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2066 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2067 // X > -1 -> X == 0, jump !sign.
2068 RHS = DAG.getConstant(0, RHS.getValueType());
2069 X86CC = X86::COND_NS;
2071 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2072 // X < 0 -> X == 0, jump on sign.
2073 X86CC = X86::COND_S;
2075 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2077 RHS = DAG.getConstant(0, RHS.getValueType());
2078 X86CC = X86::COND_LE;
2083 switch (SetCCOpcode) {
2085 case ISD::SETEQ: X86CC = X86::COND_E; break;
2086 case ISD::SETGT: X86CC = X86::COND_G; break;
2087 case ISD::SETGE: X86CC = X86::COND_GE; break;
2088 case ISD::SETLT: X86CC = X86::COND_L; break;
2089 case ISD::SETLE: X86CC = X86::COND_LE; break;
2090 case ISD::SETNE: X86CC = X86::COND_NE; break;
2091 case ISD::SETULT: X86CC = X86::COND_B; break;
2092 case ISD::SETUGT: X86CC = X86::COND_A; break;
2093 case ISD::SETULE: X86CC = X86::COND_BE; break;
2094 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2097 // On a floating point condition, the flags are set as follows:
2099 // 0 | 0 | 0 | X > Y
2100 // 0 | 0 | 1 | X < Y
2101 // 1 | 0 | 0 | X == Y
2102 // 1 | 1 | 1 | unordered
2104 switch (SetCCOpcode) {
2107 case ISD::SETEQ: X86CC = X86::COND_E; break;
2108 case ISD::SETOLT: Flip = true; // Fallthrough
2110 case ISD::SETGT: X86CC = X86::COND_A; break;
2111 case ISD::SETOLE: Flip = true; // Fallthrough
2113 case ISD::SETGE: X86CC = X86::COND_AE; break;
2114 case ISD::SETUGT: Flip = true; // Fallthrough
2116 case ISD::SETLT: X86CC = X86::COND_B; break;
2117 case ISD::SETUGE: Flip = true; // Fallthrough
2119 case ISD::SETLE: X86CC = X86::COND_BE; break;
2121 case ISD::SETNE: X86CC = X86::COND_NE; break;
2122 case ISD::SETUO: X86CC = X86::COND_P; break;
2123 case ISD::SETO: X86CC = X86::COND_NP; break;
2126 std::swap(LHS, RHS);
2129 return X86CC != X86::COND_INVALID;
2132 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2133 /// code. Current x86 isa includes the following FP cmov instructions:
2134 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2135 static bool hasFPCMov(unsigned X86CC) {
2151 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2152 /// true if Op is undef or if its value falls within the specified range (L, H].
2153 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2154 if (Op.getOpcode() == ISD::UNDEF)
2157 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2158 return (Val >= Low && Val < Hi);
2161 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2162 /// true if Op is undef or if its value equal to the specified value.
2163 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2164 if (Op.getOpcode() == ISD::UNDEF)
2166 return cast<ConstantSDNode>(Op)->getValue() == Val;
2169 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2170 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2171 bool X86::isPSHUFDMask(SDNode *N) {
2172 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2174 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2177 // Check if the value doesn't reference the second vector.
2178 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2179 SDOperand Arg = N->getOperand(i);
2180 if (Arg.getOpcode() == ISD::UNDEF) continue;
2181 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2182 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2189 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2190 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2191 bool X86::isPSHUFHWMask(SDNode *N) {
2192 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2194 if (N->getNumOperands() != 8)
2197 // Lower quadword copied in order.
2198 for (unsigned i = 0; i != 4; ++i) {
2199 SDOperand Arg = N->getOperand(i);
2200 if (Arg.getOpcode() == ISD::UNDEF) continue;
2201 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2202 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2206 // Upper quadword shuffled.
2207 for (unsigned i = 4; i != 8; ++i) {
2208 SDOperand Arg = N->getOperand(i);
2209 if (Arg.getOpcode() == ISD::UNDEF) continue;
2210 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2211 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2212 if (Val < 4 || Val > 7)
2219 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2220 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2221 bool X86::isPSHUFLWMask(SDNode *N) {
2222 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2224 if (N->getNumOperands() != 8)
2227 // Upper quadword copied in order.
2228 for (unsigned i = 4; i != 8; ++i)
2229 if (!isUndefOrEqual(N->getOperand(i), i))
2232 // Lower quadword shuffled.
2233 for (unsigned i = 0; i != 4; ++i)
2234 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2240 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2241 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2242 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2243 if (NumElems != 2 && NumElems != 4) return false;
2245 unsigned Half = NumElems / 2;
2246 for (unsigned i = 0; i < Half; ++i)
2247 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2249 for (unsigned i = Half; i < NumElems; ++i)
2250 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2256 bool X86::isSHUFPMask(SDNode *N) {
2257 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2258 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2261 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2262 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2263 /// half elements to come from vector 1 (which would equal the dest.) and
2264 /// the upper half to come from vector 2.
2265 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2266 if (NumOps != 2 && NumOps != 4) return false;
2268 unsigned Half = NumOps / 2;
2269 for (unsigned i = 0; i < Half; ++i)
2270 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2272 for (unsigned i = Half; i < NumOps; ++i)
2273 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2278 static bool isCommutedSHUFP(SDNode *N) {
2279 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2280 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2283 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2284 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2285 bool X86::isMOVHLPSMask(SDNode *N) {
2286 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2288 if (N->getNumOperands() != 4)
2291 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2292 return isUndefOrEqual(N->getOperand(0), 6) &&
2293 isUndefOrEqual(N->getOperand(1), 7) &&
2294 isUndefOrEqual(N->getOperand(2), 2) &&
2295 isUndefOrEqual(N->getOperand(3), 3);
2298 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2299 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2301 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2302 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2304 if (N->getNumOperands() != 4)
2307 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2308 return isUndefOrEqual(N->getOperand(0), 2) &&
2309 isUndefOrEqual(N->getOperand(1), 3) &&
2310 isUndefOrEqual(N->getOperand(2), 2) &&
2311 isUndefOrEqual(N->getOperand(3), 3);
2314 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2315 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2316 bool X86::isMOVLPMask(SDNode *N) {
2317 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2319 unsigned NumElems = N->getNumOperands();
2320 if (NumElems != 2 && NumElems != 4)
2323 for (unsigned i = 0; i < NumElems/2; ++i)
2324 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2327 for (unsigned i = NumElems/2; i < NumElems; ++i)
2328 if (!isUndefOrEqual(N->getOperand(i), i))
2334 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2335 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2337 bool X86::isMOVHPMask(SDNode *N) {
2338 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2340 unsigned NumElems = N->getNumOperands();
2341 if (NumElems != 2 && NumElems != 4)
2344 for (unsigned i = 0; i < NumElems/2; ++i)
2345 if (!isUndefOrEqual(N->getOperand(i), i))
2348 for (unsigned i = 0; i < NumElems/2; ++i) {
2349 SDOperand Arg = N->getOperand(i + NumElems/2);
2350 if (!isUndefOrEqual(Arg, i + NumElems))
2357 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2358 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2359 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2360 bool V2IsSplat = false) {
2361 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2364 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2365 SDOperand BitI = Elts[i];
2366 SDOperand BitI1 = Elts[i+1];
2367 if (!isUndefOrEqual(BitI, j))
2370 if (isUndefOrEqual(BitI1, NumElts))
2373 if (!isUndefOrEqual(BitI1, j + NumElts))
2381 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2382 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2383 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2386 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2387 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2388 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2389 bool V2IsSplat = false) {
2390 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2393 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2394 SDOperand BitI = Elts[i];
2395 SDOperand BitI1 = Elts[i+1];
2396 if (!isUndefOrEqual(BitI, j + NumElts/2))
2399 if (isUndefOrEqual(BitI1, NumElts))
2402 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2410 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2411 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2412 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2415 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2416 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2418 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2419 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2421 unsigned NumElems = N->getNumOperands();
2422 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2425 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2426 SDOperand BitI = N->getOperand(i);
2427 SDOperand BitI1 = N->getOperand(i+1);
2429 if (!isUndefOrEqual(BitI, j))
2431 if (!isUndefOrEqual(BitI1, j))
2438 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2439 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2441 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2442 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2444 unsigned NumElems = N->getNumOperands();
2445 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2448 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2449 SDOperand BitI = N->getOperand(i);
2450 SDOperand BitI1 = N->getOperand(i + 1);
2452 if (!isUndefOrEqual(BitI, j))
2454 if (!isUndefOrEqual(BitI1, j))
2461 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2462 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2463 /// MOVSD, and MOVD, i.e. setting the lowest element.
2464 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2465 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2468 if (!isUndefOrEqual(Elts[0], NumElts))
2471 for (unsigned i = 1; i < NumElts; ++i) {
2472 if (!isUndefOrEqual(Elts[i], i))
2479 bool X86::isMOVLMask(SDNode *N) {
2480 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2481 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2484 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2485 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2486 /// element of vector 2 and the other elements to come from vector 1 in order.
2487 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2488 bool V2IsSplat = false,
2489 bool V2IsUndef = false) {
2490 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2493 if (!isUndefOrEqual(Ops[0], 0))
2496 for (unsigned i = 1; i < NumOps; ++i) {
2497 SDOperand Arg = Ops[i];
2498 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2499 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2500 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2507 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2508 bool V2IsUndef = false) {
2509 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2510 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2511 V2IsSplat, V2IsUndef);
2514 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2515 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2516 bool X86::isMOVSHDUPMask(SDNode *N) {
2517 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2519 if (N->getNumOperands() != 4)
2522 // Expect 1, 1, 3, 3
2523 for (unsigned i = 0; i < 2; ++i) {
2524 SDOperand Arg = N->getOperand(i);
2525 if (Arg.getOpcode() == ISD::UNDEF) continue;
2526 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2527 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2528 if (Val != 1) return false;
2532 for (unsigned i = 2; i < 4; ++i) {
2533 SDOperand Arg = N->getOperand(i);
2534 if (Arg.getOpcode() == ISD::UNDEF) continue;
2535 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2536 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2537 if (Val != 3) return false;
2541 // Don't use movshdup if it can be done with a shufps.
2545 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2546 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2547 bool X86::isMOVSLDUPMask(SDNode *N) {
2548 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2550 if (N->getNumOperands() != 4)
2553 // Expect 0, 0, 2, 2
2554 for (unsigned i = 0; i < 2; ++i) {
2555 SDOperand Arg = N->getOperand(i);
2556 if (Arg.getOpcode() == ISD::UNDEF) continue;
2557 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2558 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2559 if (Val != 0) return false;
2563 for (unsigned i = 2; i < 4; ++i) {
2564 SDOperand Arg = N->getOperand(i);
2565 if (Arg.getOpcode() == ISD::UNDEF) continue;
2566 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2567 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2568 if (Val != 2) return false;
2572 // Don't use movshdup if it can be done with a shufps.
2576 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2577 /// specifies a identity operation on the LHS or RHS.
2578 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2579 unsigned NumElems = N->getNumOperands();
2580 for (unsigned i = 0; i < NumElems; ++i)
2581 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2586 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2587 /// a splat of a single element.
2588 static bool isSplatMask(SDNode *N) {
2589 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2591 // This is a splat operation if each element of the permute is the same, and
2592 // if the value doesn't reference the second vector.
2593 unsigned NumElems = N->getNumOperands();
2594 SDOperand ElementBase;
2596 for (; i != NumElems; ++i) {
2597 SDOperand Elt = N->getOperand(i);
2598 if (isa<ConstantSDNode>(Elt)) {
2604 if (!ElementBase.Val)
2607 for (; i != NumElems; ++i) {
2608 SDOperand Arg = N->getOperand(i);
2609 if (Arg.getOpcode() == ISD::UNDEF) continue;
2610 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2611 if (Arg != ElementBase) return false;
2614 // Make sure it is a splat of the first vector operand.
2615 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2618 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2619 /// a splat of a single element and it's a 2 or 4 element mask.
2620 bool X86::isSplatMask(SDNode *N) {
2621 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2623 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2624 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2626 return ::isSplatMask(N);
2629 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2630 /// specifies a splat of zero element.
2631 bool X86::isSplatLoMask(SDNode *N) {
2632 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2634 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2635 if (!isUndefOrEqual(N->getOperand(i), 0))
2640 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2641 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2643 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2644 unsigned NumOperands = N->getNumOperands();
2645 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2647 for (unsigned i = 0; i < NumOperands; ++i) {
2649 SDOperand Arg = N->getOperand(NumOperands-i-1);
2650 if (Arg.getOpcode() != ISD::UNDEF)
2651 Val = cast<ConstantSDNode>(Arg)->getValue();
2652 if (Val >= NumOperands) Val -= NumOperands;
2654 if (i != NumOperands - 1)
2661 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2662 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2664 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2666 // 8 nodes, but we only care about the last 4.
2667 for (unsigned i = 7; i >= 4; --i) {
2669 SDOperand Arg = N->getOperand(i);
2670 if (Arg.getOpcode() != ISD::UNDEF)
2671 Val = cast<ConstantSDNode>(Arg)->getValue();
2680 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2681 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2683 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2685 // 8 nodes, but we only care about the first 4.
2686 for (int i = 3; i >= 0; --i) {
2688 SDOperand Arg = N->getOperand(i);
2689 if (Arg.getOpcode() != ISD::UNDEF)
2690 Val = cast<ConstantSDNode>(Arg)->getValue();
2699 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2700 /// specifies a 8 element shuffle that can be broken into a pair of
2701 /// PSHUFHW and PSHUFLW.
2702 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2703 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2705 if (N->getNumOperands() != 8)
2708 // Lower quadword shuffled.
2709 for (unsigned i = 0; i != 4; ++i) {
2710 SDOperand Arg = N->getOperand(i);
2711 if (Arg.getOpcode() == ISD::UNDEF) continue;
2712 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2713 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2718 // Upper quadword shuffled.
2719 for (unsigned i = 4; i != 8; ++i) {
2720 SDOperand Arg = N->getOperand(i);
2721 if (Arg.getOpcode() == ISD::UNDEF) continue;
2722 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2723 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2724 if (Val < 4 || Val > 7)
2731 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2732 /// values in ther permute mask.
2733 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2734 SDOperand &V2, SDOperand &Mask,
2735 SelectionDAG &DAG) {
2736 MVT::ValueType VT = Op.getValueType();
2737 MVT::ValueType MaskVT = Mask.getValueType();
2738 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2739 unsigned NumElems = Mask.getNumOperands();
2740 SmallVector<SDOperand, 8> MaskVec;
2742 for (unsigned i = 0; i != NumElems; ++i) {
2743 SDOperand Arg = Mask.getOperand(i);
2744 if (Arg.getOpcode() == ISD::UNDEF) {
2745 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2748 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2749 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2751 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2753 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2757 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2758 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2761 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2762 /// match movhlps. The lower half elements should come from upper half of
2763 /// V1 (and in order), and the upper half elements should come from the upper
2764 /// half of V2 (and in order).
2765 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2766 unsigned NumElems = Mask->getNumOperands();
2769 for (unsigned i = 0, e = 2; i != e; ++i)
2770 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2772 for (unsigned i = 2; i != 4; ++i)
2773 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2778 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2779 /// is promoted to a vector.
2780 static inline bool isScalarLoadToVector(SDNode *N) {
2781 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2782 N = N->getOperand(0).Val;
2783 return ISD::isNON_EXTLoad(N);
2788 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2789 /// match movlp{s|d}. The lower half elements should come from lower half of
2790 /// V1 (and in order), and the upper half elements should come from the upper
2791 /// half of V2 (and in order). And since V1 will become the source of the
2792 /// MOVLP, it must be either a vector load or a scalar load to vector.
2793 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2794 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2796 // Is V2 is a vector load, don't do this transformation. We will try to use
2797 // load folding shufps op.
2798 if (ISD::isNON_EXTLoad(V2))
2801 unsigned NumElems = Mask->getNumOperands();
2802 if (NumElems != 2 && NumElems != 4)
2804 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2805 if (!isUndefOrEqual(Mask->getOperand(i), i))
2807 for (unsigned i = NumElems/2; i != NumElems; ++i)
2808 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2813 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2815 static bool isSplatVector(SDNode *N) {
2816 if (N->getOpcode() != ISD::BUILD_VECTOR)
2819 SDOperand SplatValue = N->getOperand(0);
2820 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2821 if (N->getOperand(i) != SplatValue)
2826 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2828 static bool isUndefShuffle(SDNode *N) {
2829 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2832 SDOperand V1 = N->getOperand(0);
2833 SDOperand V2 = N->getOperand(1);
2834 SDOperand Mask = N->getOperand(2);
2835 unsigned NumElems = Mask.getNumOperands();
2836 for (unsigned i = 0; i != NumElems; ++i) {
2837 SDOperand Arg = Mask.getOperand(i);
2838 if (Arg.getOpcode() != ISD::UNDEF) {
2839 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2840 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2842 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2849 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2851 static inline bool isZeroNode(SDOperand Elt) {
2852 return ((isa<ConstantSDNode>(Elt) &&
2853 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2854 (isa<ConstantFPSDNode>(Elt) &&
2855 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2858 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2859 /// to an zero vector.
2860 static bool isZeroShuffle(SDNode *N) {
2861 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2864 SDOperand V1 = N->getOperand(0);
2865 SDOperand V2 = N->getOperand(1);
2866 SDOperand Mask = N->getOperand(2);
2867 unsigned NumElems = Mask.getNumOperands();
2868 for (unsigned i = 0; i != NumElems; ++i) {
2869 SDOperand Arg = Mask.getOperand(i);
2870 if (Arg.getOpcode() != ISD::UNDEF) {
2871 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2872 if (Idx < NumElems) {
2873 unsigned Opc = V1.Val->getOpcode();
2874 if (Opc == ISD::UNDEF)
2876 if (Opc != ISD::BUILD_VECTOR ||
2877 !isZeroNode(V1.Val->getOperand(Idx)))
2879 } else if (Idx >= NumElems) {
2880 unsigned Opc = V2.Val->getOpcode();
2881 if (Opc == ISD::UNDEF)
2883 if (Opc != ISD::BUILD_VECTOR ||
2884 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2892 /// getZeroVector - Returns a vector of specified type with all zero elements.
2894 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2895 assert(MVT::isVector(VT) && "Expected a vector type");
2896 unsigned NumElems = MVT::getVectorNumElements(VT);
2897 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2898 bool isFP = MVT::isFloatingPoint(EVT);
2899 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2900 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2901 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2904 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2905 /// that point to V2 points to its first element.
2906 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2907 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2909 bool Changed = false;
2910 SmallVector<SDOperand, 8> MaskVec;
2911 unsigned NumElems = Mask.getNumOperands();
2912 for (unsigned i = 0; i != NumElems; ++i) {
2913 SDOperand Arg = Mask.getOperand(i);
2914 if (Arg.getOpcode() != ISD::UNDEF) {
2915 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2916 if (Val > NumElems) {
2917 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2921 MaskVec.push_back(Arg);
2925 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2926 &MaskVec[0], MaskVec.size());
2930 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2931 /// operation of specified width.
2932 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2933 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2934 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2936 SmallVector<SDOperand, 8> MaskVec;
2937 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2938 for (unsigned i = 1; i != NumElems; ++i)
2939 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2940 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2943 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2944 /// of specified width.
2945 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2946 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2947 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2948 SmallVector<SDOperand, 8> MaskVec;
2949 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2950 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2951 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2953 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2956 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2957 /// of specified width.
2958 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2959 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2960 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2961 unsigned Half = NumElems/2;
2962 SmallVector<SDOperand, 8> MaskVec;
2963 for (unsigned i = 0; i != Half; ++i) {
2964 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2965 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2967 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2970 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2972 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2973 SDOperand V1 = Op.getOperand(0);
2974 SDOperand Mask = Op.getOperand(2);
2975 MVT::ValueType VT = Op.getValueType();
2976 unsigned NumElems = Mask.getNumOperands();
2977 Mask = getUnpacklMask(NumElems, DAG);
2978 while (NumElems != 4) {
2979 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2982 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2984 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2985 Mask = getZeroVector(MaskVT, DAG);
2986 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2987 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2988 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2991 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2992 /// vector of zero or undef vector.
2993 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2994 unsigned NumElems, unsigned Idx,
2995 bool isZero, SelectionDAG &DAG) {
2996 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2997 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2998 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2999 SDOperand Zero = DAG.getConstant(0, EVT);
3000 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
3001 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
3002 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3003 &MaskVec[0], MaskVec.size());
3004 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3007 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3009 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
3010 unsigned NumNonZero, unsigned NumZero,
3011 SelectionDAG &DAG, TargetLowering &TLI) {
3017 for (unsigned i = 0; i < 16; ++i) {
3018 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3019 if (ThisIsNonZero && First) {
3021 V = getZeroVector(MVT::v8i16, DAG);
3023 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3028 SDOperand ThisElt(0, 0), LastElt(0, 0);
3029 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3030 if (LastIsNonZero) {
3031 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3033 if (ThisIsNonZero) {
3034 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3035 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3036 ThisElt, DAG.getConstant(8, MVT::i8));
3038 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3043 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3044 DAG.getConstant(i/2, TLI.getPointerTy()));
3048 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3051 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3053 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3054 unsigned NumNonZero, unsigned NumZero,
3055 SelectionDAG &DAG, TargetLowering &TLI) {
3061 for (unsigned i = 0; i < 8; ++i) {
3062 bool isNonZero = (NonZeros & (1 << i)) != 0;
3066 V = getZeroVector(MVT::v8i16, DAG);
3068 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3071 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3072 DAG.getConstant(i, TLI.getPointerTy()));
3080 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3081 // All zero's are handled with pxor.
3082 if (ISD::isBuildVectorAllZeros(Op.Val))
3085 // All one's are handled with pcmpeqd.
3086 if (ISD::isBuildVectorAllOnes(Op.Val))
3089 MVT::ValueType VT = Op.getValueType();
3090 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3091 unsigned EVTBits = MVT::getSizeInBits(EVT);
3093 unsigned NumElems = Op.getNumOperands();
3094 unsigned NumZero = 0;
3095 unsigned NumNonZero = 0;
3096 unsigned NonZeros = 0;
3097 unsigned NumNonZeroImms = 0;
3098 std::set<SDOperand> Values;
3099 for (unsigned i = 0; i < NumElems; ++i) {
3100 SDOperand Elt = Op.getOperand(i);
3101 if (Elt.getOpcode() != ISD::UNDEF) {
3103 if (isZeroNode(Elt))
3106 NonZeros |= (1 << i);
3108 if (Elt.getOpcode() == ISD::Constant ||
3109 Elt.getOpcode() == ISD::ConstantFP)
3115 if (NumNonZero == 0) {
3117 // All undef vector. Return an UNDEF.
3118 return DAG.getNode(ISD::UNDEF, VT);
3120 // A mix of zero and undef. Return a zero vector.
3121 return getZeroVector(VT, DAG);
3124 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3125 if (Values.size() == 1)
3128 // Special case for single non-zero element.
3129 if (NumNonZero == 1) {
3130 unsigned Idx = CountTrailingZeros_32(NonZeros);
3131 SDOperand Item = Op.getOperand(Idx);
3132 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3134 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3135 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3138 if (EVTBits == 32) {
3139 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3140 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3142 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3143 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3144 SmallVector<SDOperand, 8> MaskVec;
3145 for (unsigned i = 0; i < NumElems; i++)
3146 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3147 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3148 &MaskVec[0], MaskVec.size());
3149 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3150 DAG.getNode(ISD::UNDEF, VT), Mask);
3154 // A vector full of immediates; various special cases are already
3155 // handled, so this is best done with a single constant-pool load.
3156 if (NumNonZero == NumNonZeroImms)
3159 // Let legalizer expand 2-wide build_vectors.
3163 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3164 if (EVTBits == 8 && NumElems == 16) {
3165 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3167 if (V.Val) return V;
3170 if (EVTBits == 16 && NumElems == 8) {
3171 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3173 if (V.Val) return V;
3176 // If element VT is == 32 bits, turn it into a number of shuffles.
3177 SmallVector<SDOperand, 8> V;
3179 if (NumElems == 4 && NumZero > 0) {
3180 for (unsigned i = 0; i < 4; ++i) {
3181 bool isZero = !(NonZeros & (1 << i));
3183 V[i] = getZeroVector(VT, DAG);
3185 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3188 for (unsigned i = 0; i < 2; ++i) {
3189 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3192 V[i] = V[i*2]; // Must be a zero vector.
3195 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3196 getMOVLMask(NumElems, DAG));
3199 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3200 getMOVLMask(NumElems, DAG));
3203 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3204 getUnpacklMask(NumElems, DAG));
3209 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3210 // clears the upper bits.
3211 // FIXME: we can do the same for v4f32 case when we know both parts of
3212 // the lower half come from scalar_to_vector (loadf32). We should do
3213 // that in post legalizer dag combiner with target specific hooks.
3214 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3216 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3217 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3218 SmallVector<SDOperand, 8> MaskVec;
3219 bool Reverse = (NonZeros & 0x3) == 2;
3220 for (unsigned i = 0; i < 2; ++i)
3222 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3224 MaskVec.push_back(DAG.getConstant(i, EVT));
3225 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3226 for (unsigned i = 0; i < 2; ++i)
3228 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3230 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3231 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3232 &MaskVec[0], MaskVec.size());
3233 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3236 if (Values.size() > 2) {
3237 // Expand into a number of unpckl*.
3239 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3240 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3241 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3242 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3243 for (unsigned i = 0; i < NumElems; ++i)
3244 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3246 while (NumElems != 0) {
3247 for (unsigned i = 0; i < NumElems; ++i)
3248 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3259 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3260 SDOperand V1 = Op.getOperand(0);
3261 SDOperand V2 = Op.getOperand(1);
3262 SDOperand PermMask = Op.getOperand(2);
3263 MVT::ValueType VT = Op.getValueType();
3264 unsigned NumElems = PermMask.getNumOperands();
3265 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3266 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3267 bool V1IsSplat = false;
3268 bool V2IsSplat = false;
3270 if (isUndefShuffle(Op.Val))
3271 return DAG.getNode(ISD::UNDEF, VT);
3273 if (isZeroShuffle(Op.Val))
3274 return getZeroVector(VT, DAG);
3276 if (isIdentityMask(PermMask.Val))
3278 else if (isIdentityMask(PermMask.Val, true))
3281 if (isSplatMask(PermMask.Val)) {
3282 if (NumElems <= 4) return Op;
3283 // Promote it to a v4i32 splat.
3284 return PromoteSplat(Op, DAG);
3287 if (X86::isMOVLMask(PermMask.Val))
3288 return (V1IsUndef) ? V2 : Op;
3290 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3291 X86::isMOVSLDUPMask(PermMask.Val) ||
3292 X86::isMOVHLPSMask(PermMask.Val) ||
3293 X86::isMOVHPMask(PermMask.Val) ||
3294 X86::isMOVLPMask(PermMask.Val))
3297 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3298 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3299 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3301 bool Commuted = false;
3302 V1IsSplat = isSplatVector(V1.Val);
3303 V2IsSplat = isSplatVector(V2.Val);
3304 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3305 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3306 std::swap(V1IsSplat, V2IsSplat);
3307 std::swap(V1IsUndef, V2IsUndef);
3311 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3312 if (V2IsUndef) return V1;
3313 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3315 // V2 is a splat, so the mask may be malformed. That is, it may point
3316 // to any V2 element. The instruction selectior won't like this. Get
3317 // a corrected mask and commute to form a proper MOVS{S|D}.
3318 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3319 if (NewMask.Val != PermMask.Val)
3320 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3325 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3326 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3327 X86::isUNPCKLMask(PermMask.Val) ||
3328 X86::isUNPCKHMask(PermMask.Val))
3332 // Normalize mask so all entries that point to V2 points to its first
3333 // element then try to match unpck{h|l} again. If match, return a
3334 // new vector_shuffle with the corrected mask.
3335 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3336 if (NewMask.Val != PermMask.Val) {
3337 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3338 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3339 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3340 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3341 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3342 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3347 // Normalize the node to match x86 shuffle ops if needed
3348 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3349 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3352 // Commute is back and try unpck* again.
3353 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3354 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3355 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3356 X86::isUNPCKLMask(PermMask.Val) ||
3357 X86::isUNPCKHMask(PermMask.Val))
3361 // If VT is integer, try PSHUF* first, then SHUFP*.
3362 if (MVT::isInteger(VT)) {
3363 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3364 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3365 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3366 X86::isPSHUFDMask(PermMask.Val)) ||
3367 X86::isPSHUFHWMask(PermMask.Val) ||
3368 X86::isPSHUFLWMask(PermMask.Val)) {
3369 if (V2.getOpcode() != ISD::UNDEF)
3370 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3371 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3375 if (X86::isSHUFPMask(PermMask.Val) &&
3376 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3379 // Handle v8i16 shuffle high / low shuffle node pair.
3380 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3381 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3382 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3383 SmallVector<SDOperand, 8> MaskVec;
3384 for (unsigned i = 0; i != 4; ++i)
3385 MaskVec.push_back(PermMask.getOperand(i));
3386 for (unsigned i = 4; i != 8; ++i)
3387 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3388 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3389 &MaskVec[0], MaskVec.size());
3390 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3392 for (unsigned i = 0; i != 4; ++i)
3393 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3394 for (unsigned i = 4; i != 8; ++i)
3395 MaskVec.push_back(PermMask.getOperand(i));
3396 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3397 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3400 // Floating point cases in the other order.
3401 if (X86::isSHUFPMask(PermMask.Val))
3403 if (X86::isPSHUFDMask(PermMask.Val) ||
3404 X86::isPSHUFHWMask(PermMask.Val) ||
3405 X86::isPSHUFLWMask(PermMask.Val)) {
3406 if (V2.getOpcode() != ISD::UNDEF)
3407 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3408 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3413 if (NumElems == 4 &&
3414 // Don't do this for MMX.
3415 MVT::getSizeInBits(VT) != 64) {
3416 MVT::ValueType MaskVT = PermMask.getValueType();
3417 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3418 SmallVector<std::pair<int, int>, 8> Locs;
3419 Locs.reserve(NumElems);
3420 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3421 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3424 // If no more than two elements come from either vector. This can be
3425 // implemented with two shuffles. First shuffle gather the elements.
3426 // The second shuffle, which takes the first shuffle as both of its
3427 // vector operands, put the elements into the right order.
3428 for (unsigned i = 0; i != NumElems; ++i) {
3429 SDOperand Elt = PermMask.getOperand(i);
3430 if (Elt.getOpcode() == ISD::UNDEF) {
3431 Locs[i] = std::make_pair(-1, -1);
3433 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3434 if (Val < NumElems) {
3435 Locs[i] = std::make_pair(0, NumLo);
3439 Locs[i] = std::make_pair(1, NumHi);
3440 if (2+NumHi < NumElems)
3441 Mask1[2+NumHi] = Elt;
3446 if (NumLo <= 2 && NumHi <= 2) {
3447 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3448 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3449 &Mask1[0], Mask1.size()));
3450 for (unsigned i = 0; i != NumElems; ++i) {
3451 if (Locs[i].first == -1)
3454 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3455 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3456 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3460 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3461 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3462 &Mask2[0], Mask2.size()));
3465 // Break it into (shuffle shuffle_hi, shuffle_lo).
3467 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3468 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3469 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3470 unsigned MaskIdx = 0;
3472 unsigned HiIdx = NumElems/2;
3473 for (unsigned i = 0; i != NumElems; ++i) {
3474 if (i == NumElems/2) {
3480 SDOperand Elt = PermMask.getOperand(i);
3481 if (Elt.getOpcode() == ISD::UNDEF) {
3482 Locs[i] = std::make_pair(-1, -1);
3483 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3484 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3485 (*MaskPtr)[LoIdx] = Elt;
3488 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3489 (*MaskPtr)[HiIdx] = Elt;
3494 SDOperand LoShuffle =
3495 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3496 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3497 &LoMask[0], LoMask.size()));
3498 SDOperand HiShuffle =
3499 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3500 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3501 &HiMask[0], HiMask.size()));
3502 SmallVector<SDOperand, 8> MaskOps;
3503 for (unsigned i = 0; i != NumElems; ++i) {
3504 if (Locs[i].first == -1) {
3505 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3507 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3508 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3511 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3512 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3513 &MaskOps[0], MaskOps.size()));
3520 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3521 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3524 MVT::ValueType VT = Op.getValueType();
3525 // TODO: handle v16i8.
3526 if (MVT::getSizeInBits(VT) == 16) {
3527 // Transform it so it match pextrw which produces a 32-bit result.
3528 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3529 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3530 Op.getOperand(0), Op.getOperand(1));
3531 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3532 DAG.getValueType(VT));
3533 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3534 } else if (MVT::getSizeInBits(VT) == 32) {
3535 SDOperand Vec = Op.getOperand(0);
3536 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3539 // SHUFPS the element to the lowest double word, then movss.
3540 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3541 SmallVector<SDOperand, 8> IdxVec;
3543 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3545 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3547 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3549 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3550 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3551 &IdxVec[0], IdxVec.size());
3552 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3553 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3554 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3555 DAG.getConstant(0, getPointerTy()));
3556 } else if (MVT::getSizeInBits(VT) == 64) {
3557 SDOperand Vec = Op.getOperand(0);
3558 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3562 // UNPCKHPD the element to the lowest double word, then movsd.
3563 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3564 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3565 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3566 SmallVector<SDOperand, 8> IdxVec;
3567 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3569 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3570 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3571 &IdxVec[0], IdxVec.size());
3572 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3573 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3574 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3575 DAG.getConstant(0, getPointerTy()));
3582 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3583 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3584 // as its second argument.
3585 MVT::ValueType VT = Op.getValueType();
3586 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3587 SDOperand N0 = Op.getOperand(0);
3588 SDOperand N1 = Op.getOperand(1);
3589 SDOperand N2 = Op.getOperand(2);
3590 if (MVT::getSizeInBits(BaseVT) == 16) {
3591 if (N1.getValueType() != MVT::i32)
3592 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3593 if (N2.getValueType() != MVT::i32)
3594 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3595 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3596 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3597 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3600 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3601 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3602 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3603 SmallVector<SDOperand, 8> MaskVec;
3604 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3605 for (unsigned i = 1; i <= 3; ++i)
3606 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3607 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3608 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3609 &MaskVec[0], MaskVec.size()));
3611 // Use two pinsrw instructions to insert a 32 bit value.
3613 if (MVT::isFloatingPoint(N1.getValueType())) {
3614 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3615 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3616 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3617 DAG.getConstant(0, getPointerTy()));
3619 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3620 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3621 DAG.getConstant(Idx, getPointerTy()));
3622 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3623 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3624 DAG.getConstant(Idx+1, getPointerTy()));
3625 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3633 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3634 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3635 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3638 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3639 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3640 // one of the above mentioned nodes. It has to be wrapped because otherwise
3641 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3642 // be used to form addressing mode. These wrapped nodes will be selected
3645 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3646 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3647 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3649 CP->getAlignment());
3650 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3651 // With PIC, the address is actually $g + Offset.
3652 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3653 !Subtarget->isPICStyleRIPRel()) {
3654 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3655 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3663 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3664 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3665 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3666 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3667 // With PIC, the address is actually $g + Offset.
3668 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3669 !Subtarget->isPICStyleRIPRel()) {
3670 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3671 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3675 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3676 // load the value at address GV, not the value of GV itself. This means that
3677 // the GlobalAddress must be in the base or index register of the address, not
3678 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3679 // The same applies for external symbols during PIC codegen
3680 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3681 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3686 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3688 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3689 const MVT::ValueType PtrVT) {
3691 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3692 DAG.getNode(X86ISD::GlobalBaseReg,
3694 InFlag = Chain.getValue(1);
3696 // emit leal symbol@TLSGD(,%ebx,1), %eax
3697 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3698 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3699 GA->getValueType(0),
3701 SDOperand Ops[] = { Chain, TGA, InFlag };
3702 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3703 InFlag = Result.getValue(2);
3704 Chain = Result.getValue(1);
3706 // call ___tls_get_addr. This function receives its argument in
3707 // the register EAX.
3708 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3709 InFlag = Chain.getValue(1);
3711 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3712 SDOperand Ops1[] = { Chain,
3713 DAG.getTargetExternalSymbol("___tls_get_addr",
3715 DAG.getRegister(X86::EAX, PtrVT),
3716 DAG.getRegister(X86::EBX, PtrVT),
3718 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3719 InFlag = Chain.getValue(1);
3721 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3724 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3725 // "local exec" model.
3727 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3728 const MVT::ValueType PtrVT) {
3729 // Get the Thread Pointer
3730 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3731 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3733 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3734 GA->getValueType(0),
3736 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3738 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3739 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3741 // The address of the thread local variable is the add of the thread
3742 // pointer with the offset of the variable.
3743 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3747 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3748 // TODO: implement the "local dynamic" model
3749 // TODO: implement the "initial exec"model for pic executables
3750 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3751 "TLS not implemented for non-ELF and 64-bit targets");
3752 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3753 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3754 // otherwise use the "Local Exec"TLS Model
3755 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3756 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3758 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3762 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3763 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3764 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3765 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3766 // With PIC, the address is actually $g + Offset.
3767 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3768 !Subtarget->isPICStyleRIPRel()) {
3769 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3770 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3777 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3778 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3779 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3780 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3781 // With PIC, the address is actually $g + Offset.
3782 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3783 !Subtarget->isPICStyleRIPRel()) {
3784 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3785 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3792 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3793 /// take a 2 x i32 value to shift plus a shift amount.
3794 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3795 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3796 "Not an i64 shift!");
3797 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3798 SDOperand ShOpLo = Op.getOperand(0);
3799 SDOperand ShOpHi = Op.getOperand(1);
3800 SDOperand ShAmt = Op.getOperand(2);
3801 SDOperand Tmp1 = isSRA ?
3802 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3803 DAG.getConstant(0, MVT::i32);
3805 SDOperand Tmp2, Tmp3;
3806 if (Op.getOpcode() == ISD::SHL_PARTS) {
3807 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3808 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3810 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3811 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3814 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3815 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3816 DAG.getConstant(32, MVT::i8));
3817 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3818 AndNode, DAG.getConstant(0, MVT::i8));
3821 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3822 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3823 SmallVector<SDOperand, 4> Ops;
3824 if (Op.getOpcode() == ISD::SHL_PARTS) {
3825 Ops.push_back(Tmp2);
3826 Ops.push_back(Tmp3);
3828 Ops.push_back(Cond);
3829 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3832 Ops.push_back(Tmp3);
3833 Ops.push_back(Tmp1);
3835 Ops.push_back(Cond);
3836 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3838 Ops.push_back(Tmp2);
3839 Ops.push_back(Tmp3);
3841 Ops.push_back(Cond);
3842 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3845 Ops.push_back(Tmp3);
3846 Ops.push_back(Tmp1);
3848 Ops.push_back(Cond);
3849 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3852 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3856 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3859 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3860 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3861 Op.getOperand(0).getValueType() >= MVT::i16 &&
3862 "Unknown SINT_TO_FP to lower!");
3865 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3866 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3867 MachineFunction &MF = DAG.getMachineFunction();
3868 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3869 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3870 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3871 StackSlot, NULL, 0);
3873 // These are really Legal; caller falls through into that case.
3874 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
3876 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
3878 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
3879 Subtarget->is64Bit())
3884 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
3885 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
3887 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3889 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3890 SmallVector<SDOperand, 8> Ops;
3891 Ops.push_back(Chain);
3892 Ops.push_back(StackSlot);
3893 Ops.push_back(DAG.getValueType(SrcVT));
3894 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3895 Tys, &Ops[0], Ops.size());
3898 Chain = Result.getValue(1);
3899 SDOperand InFlag = Result.getValue(2);
3901 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3902 // shouldn't be necessary except that RFP cannot be live across
3903 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3904 MachineFunction &MF = DAG.getMachineFunction();
3905 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3906 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3907 Tys = DAG.getVTList(MVT::Other);
3908 SmallVector<SDOperand, 8> Ops;
3909 Ops.push_back(Chain);
3910 Ops.push_back(Result);
3911 Ops.push_back(StackSlot);
3912 Ops.push_back(DAG.getValueType(Op.getValueType()));
3913 Ops.push_back(InFlag);
3914 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3915 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3921 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3922 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3923 "Unknown FP_TO_SINT to lower!");
3926 // These are really Legal.
3927 if (Op.getValueType() == MVT::i32 &&
3928 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
3930 if (Op.getValueType() == MVT::i32 &&
3931 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
3933 if (Subtarget->is64Bit() &&
3934 Op.getValueType() == MVT::i64 &&
3935 Op.getOperand(0).getValueType() != MVT::f80)
3938 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3940 MachineFunction &MF = DAG.getMachineFunction();
3941 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3942 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3943 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3945 switch (Op.getValueType()) {
3946 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3947 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3948 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3949 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3952 SDOperand Chain = DAG.getEntryNode();
3953 SDOperand Value = Op.getOperand(0);
3954 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
3955 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
3956 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3957 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3958 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3960 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3962 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3963 Chain = Value.getValue(1);
3964 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3965 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3968 // Build the FP_TO_INT*_IN_MEM
3969 SDOperand Ops[] = { Chain, Value, StackSlot };
3970 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3972 // Load the result. If this is an i64 load on an x86-32 host, expand the
3974 if (Op.getValueType() != MVT::i64 || Subtarget->is64Bit())
3975 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3977 SDOperand Lo = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3978 StackSlot = DAG.getNode(ISD::ADD, StackSlot.getValueType(), StackSlot,
3979 DAG.getConstant(StackSlot.getValueType(), 4));
3980 SDOperand Hi = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3983 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
3986 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3987 MVT::ValueType VT = Op.getValueType();
3988 MVT::ValueType EltVT = VT;
3989 if (MVT::isVector(VT))
3990 EltVT = MVT::getVectorElementType(VT);
3991 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3992 std::vector<Constant*> CV;
3993 if (EltVT == MVT::f64) {
3994 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
3998 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4004 Constant *C = ConstantVector::get(CV);
4005 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4006 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4008 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4011 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4012 MVT::ValueType VT = Op.getValueType();
4013 MVT::ValueType EltVT = VT;
4014 unsigned EltNum = 1;
4015 if (MVT::isVector(VT)) {
4016 EltVT = MVT::getVectorElementType(VT);
4017 EltNum = MVT::getVectorNumElements(VT);
4019 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4020 std::vector<Constant*> CV;
4021 if (EltVT == MVT::f64) {
4022 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4026 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4032 Constant *C = ConstantVector::get(CV);
4033 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4034 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4036 if (MVT::isVector(VT)) {
4037 return DAG.getNode(ISD::BIT_CONVERT, VT,
4038 DAG.getNode(ISD::XOR, MVT::v2i64,
4039 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4040 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4042 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4046 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4047 SDOperand Op0 = Op.getOperand(0);
4048 SDOperand Op1 = Op.getOperand(1);
4049 MVT::ValueType VT = Op.getValueType();
4050 MVT::ValueType SrcVT = Op1.getValueType();
4051 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4053 // If second operand is smaller, extend it first.
4054 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4055 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4057 SrcTy = MVT::getTypeForValueType(SrcVT);
4059 // And if it is bigger, shrink it first.
4060 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4061 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1);
4063 SrcTy = MVT::getTypeForValueType(SrcVT);
4066 // At this point the operands and the result should have the same
4067 // type, and that won't be f80 since that is not custom lowered.
4069 // First get the sign bit of second operand.
4070 std::vector<Constant*> CV;
4071 if (SrcVT == MVT::f64) {
4072 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4073 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4075 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4076 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4077 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4078 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4080 Constant *C = ConstantVector::get(CV);
4081 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4082 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4084 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4086 // Shift sign bit right or left if the two operands have different types.
4087 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4088 // Op0 is MVT::f32, Op1 is MVT::f64.
4089 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4090 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4091 DAG.getConstant(32, MVT::i32));
4092 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4093 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4094 DAG.getConstant(0, getPointerTy()));
4097 // Clear first operand sign bit.
4099 if (VT == MVT::f64) {
4100 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4101 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4103 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4104 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4105 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4106 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4108 C = ConstantVector::get(CV);
4109 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4110 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4112 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4114 // Or the value with the sign bit.
4115 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4118 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4119 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4121 SDOperand Op0 = Op.getOperand(0);
4122 SDOperand Op1 = Op.getOperand(1);
4123 SDOperand CC = Op.getOperand(2);
4124 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4125 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4128 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4130 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4131 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4132 DAG.getConstant(X86CC, MVT::i8), Cond);
4135 assert(isFP && "Illegal integer SetCC!");
4137 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4138 switch (SetCCOpcode) {
4139 default: assert(false && "Illegal floating point SetCC!");
4140 case ISD::SETOEQ: { // !PF & ZF
4141 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4142 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4143 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4144 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4145 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4147 case ISD::SETUNE: { // PF | !ZF
4148 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4149 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4150 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4151 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4152 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4158 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4159 bool addTest = true;
4160 SDOperand Cond = Op.getOperand(0);
4163 if (Cond.getOpcode() == ISD::SETCC)
4164 Cond = LowerSETCC(Cond, DAG);
4166 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4167 // setting operand in place of the X86ISD::SETCC.
4168 if (Cond.getOpcode() == X86ISD::SETCC) {
4169 CC = Cond.getOperand(0);
4171 SDOperand Cmp = Cond.getOperand(1);
4172 unsigned Opc = Cmp.getOpcode();
4173 MVT::ValueType VT = Op.getValueType();
4174 bool IllegalFPCMov = false;
4175 if (VT == MVT::f32 && !X86ScalarSSEf32)
4176 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4177 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4178 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4179 else if (VT == MVT::f80)
4180 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4181 if ((Opc == X86ISD::CMP ||
4182 Opc == X86ISD::COMI ||
4183 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4190 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4191 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4194 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4196 SmallVector<SDOperand, 4> Ops;
4197 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4198 // condition is true.
4199 Ops.push_back(Op.getOperand(2));
4200 Ops.push_back(Op.getOperand(1));
4202 Ops.push_back(Cond);
4203 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4206 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4207 bool addTest = true;
4208 SDOperand Chain = Op.getOperand(0);
4209 SDOperand Cond = Op.getOperand(1);
4210 SDOperand Dest = Op.getOperand(2);
4213 if (Cond.getOpcode() == ISD::SETCC)
4214 Cond = LowerSETCC(Cond, DAG);
4216 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4217 // setting operand in place of the X86ISD::SETCC.
4218 if (Cond.getOpcode() == X86ISD::SETCC) {
4219 CC = Cond.getOperand(0);
4221 SDOperand Cmp = Cond.getOperand(1);
4222 unsigned Opc = Cmp.getOpcode();
4223 if (Opc == X86ISD::CMP ||
4224 Opc == X86ISD::COMI ||
4225 Opc == X86ISD::UCOMI) {
4232 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4233 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4235 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4236 Chain, Op.getOperand(2), CC, Cond);
4239 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4240 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4241 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4243 if (Subtarget->is64Bit())
4244 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4245 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4247 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4249 switch (CallingConv) {
4251 assert(0 && "Unsupported calling convention");
4252 case CallingConv::Fast:
4253 if (isTailCall && PerformTailCallOpt)
4254 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4256 return LowerCCCCallTo(Op,DAG, CallingConv);
4257 case CallingConv::C:
4258 case CallingConv::X86_StdCall:
4259 return LowerCCCCallTo(Op, DAG, CallingConv);
4260 case CallingConv::X86_FastCall:
4261 return LowerFastCCCallTo(Op, DAG, CallingConv);
4266 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4267 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4268 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4269 // that the guard pages used by the OS virtual memory manager are allocated in
4270 // correct sequence.
4272 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4273 SelectionDAG &DAG) {
4274 assert(Subtarget->isTargetCygMing() &&
4275 "This should be used only on Cygwin/Mingw targets");
4278 SDOperand Chain = Op.getOperand(0);
4279 SDOperand Size = Op.getOperand(1);
4280 // FIXME: Ensure alignment here
4284 MVT::ValueType IntPtr = getPointerTy();
4285 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4287 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4288 Flag = Chain.getValue(1);
4290 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4291 SDOperand Ops[] = { Chain,
4292 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4293 DAG.getRegister(X86::EAX, IntPtr),
4295 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4296 Flag = Chain.getValue(1);
4298 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4300 std::vector<MVT::ValueType> Tys;
4301 Tys.push_back(SPTy);
4302 Tys.push_back(MVT::Other);
4303 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4304 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4308 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4309 MachineFunction &MF = DAG.getMachineFunction();
4310 const Function* Fn = MF.getFunction();
4311 if (Fn->hasExternalLinkage() &&
4312 Subtarget->isTargetCygMing() &&
4313 Fn->getName() == "main")
4314 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4316 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4317 if (Subtarget->is64Bit())
4318 return LowerX86_64CCCArguments(Op, DAG);
4322 assert(0 && "Unsupported calling convention");
4323 case CallingConv::Fast:
4324 return LowerCCCArguments(Op,DAG, true);
4326 case CallingConv::C:
4327 return LowerCCCArguments(Op, DAG);
4328 case CallingConv::X86_StdCall:
4329 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4330 return LowerCCCArguments(Op, DAG, true);
4331 case CallingConv::X86_FastCall:
4332 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4333 return LowerFastCCArguments(Op, DAG);
4337 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4338 SDOperand InFlag(0, 0);
4339 SDOperand Chain = Op.getOperand(0);
4341 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4342 if (Align == 0) Align = 1;
4344 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4345 // If not DWORD aligned or size is more than the threshold, call memset.
4346 // The libc version is likely to be faster for these cases. It can use the
4347 // address value and run time information about the CPU.
4348 if ((Align & 3) != 0 ||
4349 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4350 MVT::ValueType IntPtr = getPointerTy();
4351 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4352 TargetLowering::ArgListTy Args;
4353 TargetLowering::ArgListEntry Entry;
4354 Entry.Node = Op.getOperand(1);
4355 Entry.Ty = IntPtrTy;
4356 Args.push_back(Entry);
4357 // Extend the unsigned i8 argument to be an int value for the call.
4358 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4359 Entry.Ty = IntPtrTy;
4360 Args.push_back(Entry);
4361 Entry.Node = Op.getOperand(3);
4362 Args.push_back(Entry);
4363 std::pair<SDOperand,SDOperand> CallResult =
4364 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4365 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4366 return CallResult.second;
4371 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4372 unsigned BytesLeft = 0;
4373 bool TwoRepStos = false;
4376 uint64_t Val = ValC->getValue() & 255;
4378 // If the value is a constant, then we can potentially use larger sets.
4379 switch (Align & 3) {
4380 case 2: // WORD aligned
4383 Val = (Val << 8) | Val;
4385 case 0: // DWORD aligned
4388 Val = (Val << 8) | Val;
4389 Val = (Val << 16) | Val;
4390 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4393 Val = (Val << 32) | Val;
4396 default: // Byte aligned
4399 Count = Op.getOperand(3);
4403 if (AVT > MVT::i8) {
4405 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4406 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4407 BytesLeft = I->getValue() % UBytes;
4409 assert(AVT >= MVT::i32 &&
4410 "Do not use rep;stos if not at least DWORD aligned");
4411 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4412 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4417 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4419 InFlag = Chain.getValue(1);
4422 Count = Op.getOperand(3);
4423 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4424 InFlag = Chain.getValue(1);
4427 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4429 InFlag = Chain.getValue(1);
4430 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4431 Op.getOperand(1), InFlag);
4432 InFlag = Chain.getValue(1);
4434 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4435 SmallVector<SDOperand, 8> Ops;
4436 Ops.push_back(Chain);
4437 Ops.push_back(DAG.getValueType(AVT));
4438 Ops.push_back(InFlag);
4439 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4442 InFlag = Chain.getValue(1);
4443 Count = Op.getOperand(3);
4444 MVT::ValueType CVT = Count.getValueType();
4445 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4446 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4447 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4449 InFlag = Chain.getValue(1);
4450 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4452 Ops.push_back(Chain);
4453 Ops.push_back(DAG.getValueType(MVT::i8));
4454 Ops.push_back(InFlag);
4455 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4456 } else if (BytesLeft) {
4457 // Issue stores for the last 1 - 7 bytes.
4459 unsigned Val = ValC->getValue() & 255;
4460 unsigned Offset = I->getValue() - BytesLeft;
4461 SDOperand DstAddr = Op.getOperand(1);
4462 MVT::ValueType AddrVT = DstAddr.getValueType();
4463 if (BytesLeft >= 4) {
4464 Val = (Val << 8) | Val;
4465 Val = (Val << 16) | Val;
4466 Value = DAG.getConstant(Val, MVT::i32);
4467 Chain = DAG.getStore(Chain, Value,
4468 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4469 DAG.getConstant(Offset, AddrVT)),
4474 if (BytesLeft >= 2) {
4475 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4476 Chain = DAG.getStore(Chain, Value,
4477 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4478 DAG.getConstant(Offset, AddrVT)),
4483 if (BytesLeft == 1) {
4484 Value = DAG.getConstant(Val, MVT::i8);
4485 Chain = DAG.getStore(Chain, Value,
4486 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4487 DAG.getConstant(Offset, AddrVT)),
4495 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4500 SelectionDAG &DAG) {
4502 unsigned BytesLeft = 0;
4503 switch (Align & 3) {
4504 case 2: // WORD aligned
4507 case 0: // DWORD aligned
4509 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4512 default: // Byte aligned
4517 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4518 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4519 BytesLeft = Size % UBytes;
4521 SDOperand InFlag(0, 0);
4522 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4524 InFlag = Chain.getValue(1);
4525 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4527 InFlag = Chain.getValue(1);
4528 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4530 InFlag = Chain.getValue(1);
4532 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4533 SmallVector<SDOperand, 8> Ops;
4534 Ops.push_back(Chain);
4535 Ops.push_back(DAG.getValueType(AVT));
4536 Ops.push_back(InFlag);
4537 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4540 // Issue loads and stores for the last 1 - 7 bytes.
4541 unsigned Offset = Size - BytesLeft;
4542 SDOperand DstAddr = Dest;
4543 MVT::ValueType DstVT = DstAddr.getValueType();
4544 SDOperand SrcAddr = Source;
4545 MVT::ValueType SrcVT = SrcAddr.getValueType();
4547 if (BytesLeft >= 4) {
4548 Value = DAG.getLoad(MVT::i32, Chain,
4549 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4550 DAG.getConstant(Offset, SrcVT)),
4552 Chain = Value.getValue(1);
4553 Chain = DAG.getStore(Chain, Value,
4554 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4555 DAG.getConstant(Offset, DstVT)),
4560 if (BytesLeft >= 2) {
4561 Value = DAG.getLoad(MVT::i16, Chain,
4562 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4563 DAG.getConstant(Offset, SrcVT)),
4565 Chain = Value.getValue(1);
4566 Chain = DAG.getStore(Chain, Value,
4567 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4568 DAG.getConstant(Offset, DstVT)),
4574 if (BytesLeft == 1) {
4575 Value = DAG.getLoad(MVT::i8, Chain,
4576 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4577 DAG.getConstant(Offset, SrcVT)),
4579 Chain = Value.getValue(1);
4580 Chain = DAG.getStore(Chain, Value,
4581 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4582 DAG.getConstant(Offset, DstVT)),
4591 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4592 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4593 SDOperand TheOp = Op.getOperand(0);
4594 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4595 if (Subtarget->is64Bit()) {
4597 DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4598 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4599 MVT::i64, Copy1.getValue(2));
4600 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4601 DAG.getConstant(32, MVT::i8));
4603 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4606 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4607 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4610 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4611 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4612 MVT::i32, Copy1.getValue(2));
4613 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4614 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4615 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4618 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4619 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4621 if (!Subtarget->is64Bit()) {
4622 // vastart just stores the address of the VarArgsFrameIndex slot into the
4623 // memory location argument.
4624 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4625 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4630 // gp_offset (0 - 6 * 8)
4631 // fp_offset (48 - 48 + 8 * 16)
4632 // overflow_arg_area (point to parameters coming in memory).
4634 SmallVector<SDOperand, 8> MemOps;
4635 SDOperand FIN = Op.getOperand(1);
4637 SDOperand Store = DAG.getStore(Op.getOperand(0),
4638 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4639 FIN, SV->getValue(), SV->getOffset());
4640 MemOps.push_back(Store);
4643 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4644 DAG.getConstant(4, getPointerTy()));
4645 Store = DAG.getStore(Op.getOperand(0),
4646 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4647 FIN, SV->getValue(), SV->getOffset());
4648 MemOps.push_back(Store);
4650 // Store ptr to overflow_arg_area
4651 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4652 DAG.getConstant(4, getPointerTy()));
4653 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4654 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4656 MemOps.push_back(Store);
4658 // Store ptr to reg_save_area.
4659 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4660 DAG.getConstant(8, getPointerTy()));
4661 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4662 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4664 MemOps.push_back(Store);
4665 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4668 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4669 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4670 SDOperand Chain = Op.getOperand(0);
4671 SDOperand DstPtr = Op.getOperand(1);
4672 SDOperand SrcPtr = Op.getOperand(2);
4673 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4674 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4676 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4677 SrcSV->getValue(), SrcSV->getOffset());
4678 Chain = SrcPtr.getValue(1);
4679 for (unsigned i = 0; i < 3; ++i) {
4680 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4681 SrcSV->getValue(), SrcSV->getOffset());
4682 Chain = Val.getValue(1);
4683 Chain = DAG.getStore(Chain, Val, DstPtr,
4684 DstSV->getValue(), DstSV->getOffset());
4687 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4688 DAG.getConstant(8, getPointerTy()));
4689 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4690 DAG.getConstant(8, getPointerTy()));
4696 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4697 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4699 default: return SDOperand(); // Don't custom lower most intrinsics.
4700 // Comparison intrinsics.
4701 case Intrinsic::x86_sse_comieq_ss:
4702 case Intrinsic::x86_sse_comilt_ss:
4703 case Intrinsic::x86_sse_comile_ss:
4704 case Intrinsic::x86_sse_comigt_ss:
4705 case Intrinsic::x86_sse_comige_ss:
4706 case Intrinsic::x86_sse_comineq_ss:
4707 case Intrinsic::x86_sse_ucomieq_ss:
4708 case Intrinsic::x86_sse_ucomilt_ss:
4709 case Intrinsic::x86_sse_ucomile_ss:
4710 case Intrinsic::x86_sse_ucomigt_ss:
4711 case Intrinsic::x86_sse_ucomige_ss:
4712 case Intrinsic::x86_sse_ucomineq_ss:
4713 case Intrinsic::x86_sse2_comieq_sd:
4714 case Intrinsic::x86_sse2_comilt_sd:
4715 case Intrinsic::x86_sse2_comile_sd:
4716 case Intrinsic::x86_sse2_comigt_sd:
4717 case Intrinsic::x86_sse2_comige_sd:
4718 case Intrinsic::x86_sse2_comineq_sd:
4719 case Intrinsic::x86_sse2_ucomieq_sd:
4720 case Intrinsic::x86_sse2_ucomilt_sd:
4721 case Intrinsic::x86_sse2_ucomile_sd:
4722 case Intrinsic::x86_sse2_ucomigt_sd:
4723 case Intrinsic::x86_sse2_ucomige_sd:
4724 case Intrinsic::x86_sse2_ucomineq_sd: {
4726 ISD::CondCode CC = ISD::SETCC_INVALID;
4729 case Intrinsic::x86_sse_comieq_ss:
4730 case Intrinsic::x86_sse2_comieq_sd:
4734 case Intrinsic::x86_sse_comilt_ss:
4735 case Intrinsic::x86_sse2_comilt_sd:
4739 case Intrinsic::x86_sse_comile_ss:
4740 case Intrinsic::x86_sse2_comile_sd:
4744 case Intrinsic::x86_sse_comigt_ss:
4745 case Intrinsic::x86_sse2_comigt_sd:
4749 case Intrinsic::x86_sse_comige_ss:
4750 case Intrinsic::x86_sse2_comige_sd:
4754 case Intrinsic::x86_sse_comineq_ss:
4755 case Intrinsic::x86_sse2_comineq_sd:
4759 case Intrinsic::x86_sse_ucomieq_ss:
4760 case Intrinsic::x86_sse2_ucomieq_sd:
4761 Opc = X86ISD::UCOMI;
4764 case Intrinsic::x86_sse_ucomilt_ss:
4765 case Intrinsic::x86_sse2_ucomilt_sd:
4766 Opc = X86ISD::UCOMI;
4769 case Intrinsic::x86_sse_ucomile_ss:
4770 case Intrinsic::x86_sse2_ucomile_sd:
4771 Opc = X86ISD::UCOMI;
4774 case Intrinsic::x86_sse_ucomigt_ss:
4775 case Intrinsic::x86_sse2_ucomigt_sd:
4776 Opc = X86ISD::UCOMI;
4779 case Intrinsic::x86_sse_ucomige_ss:
4780 case Intrinsic::x86_sse2_ucomige_sd:
4781 Opc = X86ISD::UCOMI;
4784 case Intrinsic::x86_sse_ucomineq_ss:
4785 case Intrinsic::x86_sse2_ucomineq_sd:
4786 Opc = X86ISD::UCOMI;
4792 SDOperand LHS = Op.getOperand(1);
4793 SDOperand RHS = Op.getOperand(2);
4794 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4796 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4797 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4798 DAG.getConstant(X86CC, MVT::i8), Cond);
4799 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4804 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4805 // Depths > 0 not supported yet!
4806 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4809 // Just load the return address
4810 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4811 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4814 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4815 // Depths > 0 not supported yet!
4816 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4819 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4820 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4821 DAG.getConstant(4, getPointerTy()));
4824 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4825 SelectionDAG &DAG) {
4826 // Is not yet supported on x86-64
4827 if (Subtarget->is64Bit())
4830 return DAG.getConstant(8, getPointerTy());
4833 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4835 assert(!Subtarget->is64Bit() &&
4836 "Lowering of eh_return builtin is not supported yet on x86-64");
4838 MachineFunction &MF = DAG.getMachineFunction();
4839 SDOperand Chain = Op.getOperand(0);
4840 SDOperand Offset = Op.getOperand(1);
4841 SDOperand Handler = Op.getOperand(2);
4843 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4846 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4847 DAG.getConstant(-4UL, getPointerTy()));
4848 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4849 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4850 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4851 MF.addLiveOut(X86::ECX);
4853 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4854 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4857 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4858 SelectionDAG &DAG) {
4859 SDOperand Root = Op.getOperand(0);
4860 SDOperand Trmp = Op.getOperand(1); // trampoline
4861 SDOperand FPtr = Op.getOperand(2); // nested function
4862 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4864 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4866 if (Subtarget->is64Bit()) {
4867 return SDOperand(); // not yet supported
4869 Function *Func = (Function *)
4870 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4871 unsigned CC = Func->getCallingConv();
4876 assert(0 && "Unsupported calling convention");
4877 case CallingConv::C:
4878 case CallingConv::X86_StdCall: {
4879 // Pass 'nest' parameter in ECX.
4880 // Must be kept in sync with X86CallingConv.td
4883 // Check that ECX wasn't needed by an 'inreg' parameter.
4884 const FunctionType *FTy = Func->getFunctionType();
4885 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4887 if (Attrs && !Func->isVarArg()) {
4888 unsigned InRegCount = 0;
4891 for (FunctionType::param_iterator I = FTy->param_begin(),
4892 E = FTy->param_end(); I != E; ++I, ++Idx)
4893 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4894 // FIXME: should only count parameters that are lowered to integers.
4895 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4897 if (InRegCount > 2) {
4898 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4904 case CallingConv::X86_FastCall:
4905 // Pass 'nest' parameter in EAX.
4906 // Must be kept in sync with X86CallingConv.td
4911 const X86InstrInfo *TII =
4912 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4914 SDOperand OutChains[4];
4915 SDOperand Addr, Disp;
4917 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4918 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4920 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
4921 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
4922 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
4923 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4925 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4926 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4927 TrmpSV->getOffset() + 1, false, 1);
4929 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
4930 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4931 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4932 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4934 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4935 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4936 TrmpSV->getOffset() + 6, false, 1);
4939 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
4940 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4944 /// LowerOperation - Provide custom lowering hooks for some operations.
4946 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4947 switch (Op.getOpcode()) {
4948 default: assert(0 && "Should not custom lower this!");
4949 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4950 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4951 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4952 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4953 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4954 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4955 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4956 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4957 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4958 case ISD::SHL_PARTS:
4959 case ISD::SRA_PARTS:
4960 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4961 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4962 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4963 case ISD::FABS: return LowerFABS(Op, DAG);
4964 case ISD::FNEG: return LowerFNEG(Op, DAG);
4965 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4966 case ISD::SETCC: return LowerSETCC(Op, DAG);
4967 case ISD::SELECT: return LowerSELECT(Op, DAG);
4968 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4969 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4970 case ISD::CALL: return LowerCALL(Op, DAG);
4971 case ISD::RET: return LowerRET(Op, DAG);
4972 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4973 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4974 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4975 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4976 case ISD::VASTART: return LowerVASTART(Op, DAG);
4977 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4978 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4979 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4980 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4981 case ISD::FRAME_TO_ARGS_OFFSET:
4982 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
4983 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4984 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
4985 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4990 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4992 default: return NULL;
4993 case X86ISD::SHLD: return "X86ISD::SHLD";
4994 case X86ISD::SHRD: return "X86ISD::SHRD";
4995 case X86ISD::FAND: return "X86ISD::FAND";
4996 case X86ISD::FOR: return "X86ISD::FOR";
4997 case X86ISD::FXOR: return "X86ISD::FXOR";
4998 case X86ISD::FSRL: return "X86ISD::FSRL";
4999 case X86ISD::FILD: return "X86ISD::FILD";
5000 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5001 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5002 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5003 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5004 case X86ISD::FLD: return "X86ISD::FLD";
5005 case X86ISD::FST: return "X86ISD::FST";
5006 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5007 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5008 case X86ISD::CALL: return "X86ISD::CALL";
5009 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5010 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5011 case X86ISD::CMP: return "X86ISD::CMP";
5012 case X86ISD::COMI: return "X86ISD::COMI";
5013 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5014 case X86ISD::SETCC: return "X86ISD::SETCC";
5015 case X86ISD::CMOV: return "X86ISD::CMOV";
5016 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5017 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5018 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5019 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5020 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5021 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5022 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5023 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5024 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5025 case X86ISD::FMAX: return "X86ISD::FMAX";
5026 case X86ISD::FMIN: return "X86ISD::FMIN";
5027 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5028 case X86ISD::FRCP: return "X86ISD::FRCP";
5029 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5030 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5031 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5032 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5036 // isLegalAddressingMode - Return true if the addressing mode represented
5037 // by AM is legal for this target, for a load/store of the specified type.
5038 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5039 const Type *Ty) const {
5040 // X86 supports extremely general addressing modes.
5042 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5043 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5047 // We can only fold this if we don't need an extra load.
5048 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5051 // X86-64 only supports addr of globals in small code model.
5052 if (Subtarget->is64Bit()) {
5053 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5055 // If lower 4G is not available, then we must use rip-relative addressing.
5056 if (AM.BaseOffs || AM.Scale > 1)
5067 // These scales always work.
5072 // These scales are formed with basereg+scalereg. Only accept if there is
5077 default: // Other stuff never works.
5085 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5086 if (!Ty1->isInteger() || !Ty2->isInteger())
5088 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5089 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5090 if (NumBits1 <= NumBits2)
5092 return Subtarget->is64Bit() || NumBits1 < 64;
5095 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5096 MVT::ValueType VT2) const {
5097 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5099 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5100 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5101 if (NumBits1 <= NumBits2)
5103 return Subtarget->is64Bit() || NumBits1 < 64;
5106 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5107 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5108 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5109 /// are assumed to be legal.
5111 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5112 // Only do shuffles on 128-bit vector types for now.
5113 if (MVT::getSizeInBits(VT) == 64) return false;
5114 return (Mask.Val->getNumOperands() <= 4 ||
5115 isIdentityMask(Mask.Val) ||
5116 isIdentityMask(Mask.Val, true) ||
5117 isSplatMask(Mask.Val) ||
5118 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5119 X86::isUNPCKLMask(Mask.Val) ||
5120 X86::isUNPCKHMask(Mask.Val) ||
5121 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5122 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5125 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5127 SelectionDAG &DAG) const {
5128 unsigned NumElts = BVOps.size();
5129 // Only do shuffles on 128-bit vector types for now.
5130 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5131 if (NumElts == 2) return true;
5133 return (isMOVLMask(&BVOps[0], 4) ||
5134 isCommutedMOVL(&BVOps[0], 4, true) ||
5135 isSHUFPMask(&BVOps[0], 4) ||
5136 isCommutedSHUFP(&BVOps[0], 4));
5141 //===----------------------------------------------------------------------===//
5142 // X86 Scheduler Hooks
5143 //===----------------------------------------------------------------------===//
5146 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5147 MachineBasicBlock *BB) {
5148 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5149 switch (MI->getOpcode()) {
5150 default: assert(false && "Unexpected instr type to insert");
5151 case X86::CMOV_FR32:
5152 case X86::CMOV_FR64:
5153 case X86::CMOV_V4F32:
5154 case X86::CMOV_V2F64:
5155 case X86::CMOV_V2I64: {
5156 // To "insert" a SELECT_CC instruction, we actually have to insert the
5157 // diamond control-flow pattern. The incoming instruction knows the
5158 // destination vreg to set, the condition code register to branch on, the
5159 // true/false values to select between, and a branch opcode to use.
5160 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5161 ilist<MachineBasicBlock>::iterator It = BB;
5167 // cmpTY ccX, r1, r2
5169 // fallthrough --> copy0MBB
5170 MachineBasicBlock *thisMBB = BB;
5171 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5172 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5174 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5175 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5176 MachineFunction *F = BB->getParent();
5177 F->getBasicBlockList().insert(It, copy0MBB);
5178 F->getBasicBlockList().insert(It, sinkMBB);
5179 // Update machine-CFG edges by first adding all successors of the current
5180 // block to the new block which will contain the Phi node for the select.
5181 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5182 e = BB->succ_end(); i != e; ++i)
5183 sinkMBB->addSuccessor(*i);
5184 // Next, remove all successors of the current block, and add the true
5185 // and fallthrough blocks as its successors.
5186 while(!BB->succ_empty())
5187 BB->removeSuccessor(BB->succ_begin());
5188 BB->addSuccessor(copy0MBB);
5189 BB->addSuccessor(sinkMBB);
5192 // %FalseValue = ...
5193 // # fallthrough to sinkMBB
5196 // Update machine-CFG edges
5197 BB->addSuccessor(sinkMBB);
5200 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5203 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5204 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5205 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5207 delete MI; // The pseudo instruction is gone now.
5211 case X86::FP32_TO_INT16_IN_MEM:
5212 case X86::FP32_TO_INT32_IN_MEM:
5213 case X86::FP32_TO_INT64_IN_MEM:
5214 case X86::FP64_TO_INT16_IN_MEM:
5215 case X86::FP64_TO_INT32_IN_MEM:
5216 case X86::FP64_TO_INT64_IN_MEM:
5217 case X86::FP80_TO_INT16_IN_MEM:
5218 case X86::FP80_TO_INT32_IN_MEM:
5219 case X86::FP80_TO_INT64_IN_MEM: {
5220 // Change the floating point control register to use "round towards zero"
5221 // mode when truncating to an integer value.
5222 MachineFunction *F = BB->getParent();
5223 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5224 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5226 // Load the old value of the high byte of the control word...
5228 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5229 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5231 // Set the high part to be round to zero...
5232 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5235 // Reload the modified control word now...
5236 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5238 // Restore the memory image of control word to original value
5239 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5242 // Get the X86 opcode to use.
5244 switch (MI->getOpcode()) {
5245 default: assert(0 && "illegal opcode!");
5246 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5247 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5248 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5249 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5250 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5251 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5252 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5253 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5254 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5258 MachineOperand &Op = MI->getOperand(0);
5259 if (Op.isRegister()) {
5260 AM.BaseType = X86AddressMode::RegBase;
5261 AM.Base.Reg = Op.getReg();
5263 AM.BaseType = X86AddressMode::FrameIndexBase;
5264 AM.Base.FrameIndex = Op.getFrameIndex();
5266 Op = MI->getOperand(1);
5267 if (Op.isImmediate())
5268 AM.Scale = Op.getImm();
5269 Op = MI->getOperand(2);
5270 if (Op.isImmediate())
5271 AM.IndexReg = Op.getImm();
5272 Op = MI->getOperand(3);
5273 if (Op.isGlobalAddress()) {
5274 AM.GV = Op.getGlobal();
5276 AM.Disp = Op.getImm();
5278 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5279 .addReg(MI->getOperand(4).getReg());
5281 // Reload the original control word now.
5282 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5284 delete MI; // The pseudo instruction is gone now.
5290 //===----------------------------------------------------------------------===//
5291 // X86 Optimization Hooks
5292 //===----------------------------------------------------------------------===//
5294 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5296 uint64_t &KnownZero,
5298 const SelectionDAG &DAG,
5299 unsigned Depth) const {
5300 unsigned Opc = Op.getOpcode();
5301 assert((Opc >= ISD::BUILTIN_OP_END ||
5302 Opc == ISD::INTRINSIC_WO_CHAIN ||
5303 Opc == ISD::INTRINSIC_W_CHAIN ||
5304 Opc == ISD::INTRINSIC_VOID) &&
5305 "Should use MaskedValueIsZero if you don't know whether Op"
5306 " is a target node!");
5308 KnownZero = KnownOne = 0; // Don't know anything.
5312 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5317 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5318 /// element of the result of the vector shuffle.
5319 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5320 MVT::ValueType VT = N->getValueType(0);
5321 SDOperand PermMask = N->getOperand(2);
5322 unsigned NumElems = PermMask.getNumOperands();
5323 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5325 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5327 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5328 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5329 SDOperand Idx = PermMask.getOperand(i);
5330 if (Idx.getOpcode() == ISD::UNDEF)
5331 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5332 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5337 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5338 /// node is a GlobalAddress + an offset.
5339 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5340 unsigned Opc = N->getOpcode();
5341 if (Opc == X86ISD::Wrapper) {
5342 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5343 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5346 } else if (Opc == ISD::ADD) {
5347 SDOperand N1 = N->getOperand(0);
5348 SDOperand N2 = N->getOperand(1);
5349 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5350 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5352 Offset += V->getSignExtended();
5355 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5356 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5358 Offset += V->getSignExtended();
5366 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5368 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5369 MachineFrameInfo *MFI) {
5370 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5373 SDOperand Loc = N->getOperand(1);
5374 SDOperand BaseLoc = Base->getOperand(1);
5375 if (Loc.getOpcode() == ISD::FrameIndex) {
5376 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5378 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5379 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5380 int FS = MFI->getObjectSize(FI);
5381 int BFS = MFI->getObjectSize(BFI);
5382 if (FS != BFS || FS != Size) return false;
5383 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5385 GlobalValue *GV1 = NULL;
5386 GlobalValue *GV2 = NULL;
5387 int64_t Offset1 = 0;
5388 int64_t Offset2 = 0;
5389 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5390 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5391 if (isGA1 && isGA2 && GV1 == GV2)
5392 return Offset1 == (Offset2 + Dist*Size);
5398 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5399 const X86Subtarget *Subtarget) {
5402 if (isGAPlusOffset(Base, GV, Offset))
5403 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5405 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5406 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5408 // Fixed objects do not specify alignment, however the offsets are known.
5409 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5410 (MFI->getObjectOffset(BFI) % 16) == 0);
5412 return MFI->getObjectAlignment(BFI) >= 16;
5418 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5419 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5420 /// if the load addresses are consecutive, non-overlapping, and in the right
5422 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5423 const X86Subtarget *Subtarget) {
5424 MachineFunction &MF = DAG.getMachineFunction();
5425 MachineFrameInfo *MFI = MF.getFrameInfo();
5426 MVT::ValueType VT = N->getValueType(0);
5427 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5428 SDOperand PermMask = N->getOperand(2);
5429 int NumElems = (int)PermMask.getNumOperands();
5430 SDNode *Base = NULL;
5431 for (int i = 0; i < NumElems; ++i) {
5432 SDOperand Idx = PermMask.getOperand(i);
5433 if (Idx.getOpcode() == ISD::UNDEF) {
5434 if (!Base) return SDOperand();
5437 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5438 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5442 else if (!isConsecutiveLoad(Arg.Val, Base,
5443 i, MVT::getSizeInBits(EVT)/8,MFI))
5448 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5449 LoadSDNode *LD = cast<LoadSDNode>(Base);
5451 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5452 LD->getSrcValueOffset(), LD->isVolatile());
5454 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5455 LD->getSrcValueOffset(), LD->isVolatile(),
5456 LD->getAlignment());
5460 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5461 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5462 const X86Subtarget *Subtarget) {
5463 SDOperand Cond = N->getOperand(0);
5465 // If we have SSE[12] support, try to form min/max nodes.
5466 if (Subtarget->hasSSE2() &&
5467 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5468 if (Cond.getOpcode() == ISD::SETCC) {
5469 // Get the LHS/RHS of the select.
5470 SDOperand LHS = N->getOperand(1);
5471 SDOperand RHS = N->getOperand(2);
5472 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5474 unsigned Opcode = 0;
5475 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5478 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5481 if (!UnsafeFPMath) break;
5483 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5485 Opcode = X86ISD::FMIN;
5488 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5491 if (!UnsafeFPMath) break;
5493 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5495 Opcode = X86ISD::FMAX;
5498 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5501 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5504 if (!UnsafeFPMath) break;
5506 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5508 Opcode = X86ISD::FMIN;
5511 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5514 if (!UnsafeFPMath) break;
5516 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5518 Opcode = X86ISD::FMAX;
5524 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5533 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5534 DAGCombinerInfo &DCI) const {
5535 SelectionDAG &DAG = DCI.DAG;
5536 switch (N->getOpcode()) {
5538 case ISD::VECTOR_SHUFFLE:
5539 return PerformShuffleCombine(N, DAG, Subtarget);
5541 return PerformSELECTCombine(N, DAG, Subtarget);
5547 //===----------------------------------------------------------------------===//
5548 // X86 Inline Assembly Support
5549 //===----------------------------------------------------------------------===//
5551 /// getConstraintType - Given a constraint letter, return the type of
5552 /// constraint it is for this target.
5553 X86TargetLowering::ConstraintType
5554 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5555 if (Constraint.size() == 1) {
5556 switch (Constraint[0]) {
5565 return C_RegisterClass;
5570 return TargetLowering::getConstraintType(Constraint);
5573 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5574 /// vector. If it is invalid, don't add anything to Ops.
5575 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5577 std::vector<SDOperand>&Ops,
5578 SelectionDAG &DAG) {
5579 SDOperand Result(0, 0);
5581 switch (Constraint) {
5584 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5585 if (C->getValue() <= 31) {
5586 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5592 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5593 if (C->getValue() <= 255) {
5594 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5600 // Literal immediates are always ok.
5601 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5602 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5606 // If we are in non-pic codegen mode, we allow the address of a global (with
5607 // an optional displacement) to be used with 'i'.
5608 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5611 // Match either (GA) or (GA+C)
5613 Offset = GA->getOffset();
5614 } else if (Op.getOpcode() == ISD::ADD) {
5615 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5616 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5618 Offset = GA->getOffset()+C->getValue();
5620 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5621 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5623 Offset = GA->getOffset()+C->getValue();
5630 // If addressing this global requires a load (e.g. in PIC mode), we can't
5632 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5636 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5642 // Otherwise, not valid for this mode.
5648 Ops.push_back(Result);
5651 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5654 std::vector<unsigned> X86TargetLowering::
5655 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5656 MVT::ValueType VT) const {
5657 if (Constraint.size() == 1) {
5658 // FIXME: not handling fp-stack yet!
5659 switch (Constraint[0]) { // GCC X86 Constraint Letters
5660 default: break; // Unknown constraint letter
5661 case 'A': // EAX/EDX
5662 if (VT == MVT::i32 || VT == MVT::i64)
5663 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5665 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5668 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5669 else if (VT == MVT::i16)
5670 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5671 else if (VT == MVT::i8)
5672 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5673 else if (VT == MVT::i64)
5674 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
5679 return std::vector<unsigned>();
5682 std::pair<unsigned, const TargetRegisterClass*>
5683 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5684 MVT::ValueType VT) const {
5685 // First, see if this is a constraint that directly corresponds to an LLVM
5687 if (Constraint.size() == 1) {
5688 // GCC Constraint Letters
5689 switch (Constraint[0]) {
5691 case 'r': // GENERAL_REGS
5692 case 'R': // LEGACY_REGS
5693 case 'l': // INDEX_REGS
5694 if (VT == MVT::i64 && Subtarget->is64Bit())
5695 return std::make_pair(0U, X86::GR64RegisterClass);
5697 return std::make_pair(0U, X86::GR32RegisterClass);
5698 else if (VT == MVT::i16)
5699 return std::make_pair(0U, X86::GR16RegisterClass);
5700 else if (VT == MVT::i8)
5701 return std::make_pair(0U, X86::GR8RegisterClass);
5703 case 'y': // MMX_REGS if MMX allowed.
5704 if (!Subtarget->hasMMX()) break;
5705 return std::make_pair(0U, X86::VR64RegisterClass);
5707 case 'Y': // SSE_REGS if SSE2 allowed
5708 if (!Subtarget->hasSSE2()) break;
5710 case 'x': // SSE_REGS if SSE1 allowed
5711 if (!Subtarget->hasSSE1()) break;
5715 // Scalar SSE types.
5718 return std::make_pair(0U, X86::FR32RegisterClass);
5721 return std::make_pair(0U, X86::FR64RegisterClass);
5729 return std::make_pair(0U, X86::VR128RegisterClass);
5735 // Use the default implementation in TargetLowering to convert the register
5736 // constraint into a member of a register class.
5737 std::pair<unsigned, const TargetRegisterClass*> Res;
5738 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5740 // Not found as a standard register?
5741 if (Res.second == 0) {
5742 // GCC calls "st(0)" just plain "st".
5743 if (StringsEqualNoCase("{st}", Constraint)) {
5744 Res.first = X86::ST0;
5745 Res.second = X86::RFP80RegisterClass;
5751 // Otherwise, check to see if this is a register class of the wrong value
5752 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5753 // turn into {ax},{dx}.
5754 if (Res.second->hasType(VT))
5755 return Res; // Correct type already, nothing to do.
5757 // All of the single-register GCC register classes map their values onto
5758 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5759 // really want an 8-bit or 32-bit register, map to the appropriate register
5760 // class and return the appropriate register.
5761 if (Res.second != X86::GR16RegisterClass)
5764 if (VT == MVT::i8) {
5765 unsigned DestReg = 0;
5766 switch (Res.first) {
5768 case X86::AX: DestReg = X86::AL; break;
5769 case X86::DX: DestReg = X86::DL; break;
5770 case X86::CX: DestReg = X86::CL; break;
5771 case X86::BX: DestReg = X86::BL; break;
5774 Res.first = DestReg;
5775 Res.second = Res.second = X86::GR8RegisterClass;
5777 } else if (VT == MVT::i32) {
5778 unsigned DestReg = 0;
5779 switch (Res.first) {
5781 case X86::AX: DestReg = X86::EAX; break;
5782 case X86::DX: DestReg = X86::EDX; break;
5783 case X86::CX: DestReg = X86::ECX; break;
5784 case X86::BX: DestReg = X86::EBX; break;
5785 case X86::SI: DestReg = X86::ESI; break;
5786 case X86::DI: DestReg = X86::EDI; break;
5787 case X86::BP: DestReg = X86::EBP; break;
5788 case X86::SP: DestReg = X86::ESP; break;
5791 Res.first = DestReg;
5792 Res.second = Res.second = X86::GR32RegisterClass;
5794 } else if (VT == MVT::i64) {
5795 unsigned DestReg = 0;
5796 switch (Res.first) {
5798 case X86::AX: DestReg = X86::RAX; break;
5799 case X86::DX: DestReg = X86::RDX; break;
5800 case X86::CX: DestReg = X86::RCX; break;
5801 case X86::BX: DestReg = X86::RBX; break;
5802 case X86::SI: DestReg = X86::RSI; break;
5803 case X86::DI: DestReg = X86::RDI; break;
5804 case X86::BP: DestReg = X86::RBP; break;
5805 case X86::SP: DestReg = X86::RSP; break;
5808 Res.first = DestReg;
5809 Res.second = Res.second = X86::GR64RegisterClass;