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 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i1);
1648 MemOpChains2.push_back(DAG.getMemcpy(Chain, FIN, PtrOff, SizeNode,
1649 AlignNode,AlwaysInline));
1651 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1652 // Store relative to framepointer.
1653 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1658 if (!MemOpChains2.empty())
1659 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1660 &MemOpChains2[0], MemOpChains.size());
1662 // Store the return address to the appropriate stack slot.
1664 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1666 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1668 // Does not work with tail call since ebx is not restored correctly by
1669 // tailcaller. TODO: at least for x86 - verify for x86-64
1671 // If the callee is a GlobalAddress node (quite common, every direct call is)
1672 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1673 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1674 // We should use extra load for direct calls to dllimported functions in
1676 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1677 getTargetMachine(), true))
1678 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1679 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1680 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1682 assert(Callee.getOpcode() == ISD::LOAD &&
1683 "Function destination must be loaded into virtual register");
1684 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1686 Chain = DAG.getCopyToReg(Chain,
1687 DAG.getRegister(Opc, getPointerTy()) ,
1689 Callee = DAG.getRegister(Opc, getPointerTy());
1690 // Add register as live out.
1691 DAG.getMachineFunction().addLiveOut(Opc);
1694 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1695 SmallVector<SDOperand, 8> Ops;
1697 Ops.push_back(Chain);
1698 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1699 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1701 Ops.push_back(InFlag);
1702 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1703 InFlag = Chain.getValue(1);
1705 // Returns a chain & a flag for retval copy to use.
1706 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1708 Ops.push_back(Chain);
1709 Ops.push_back(Callee);
1710 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1711 // Add argument registers to the end of the list so that they are known live
1713 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1714 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1715 RegsToPass[i].second.getValueType()));
1717 Ops.push_back(InFlag);
1718 assert(InFlag.Val &&
1719 "Flag must be set. Depend on flag being set in LowerRET");
1720 Chain = DAG.getNode(X86ISD::TAILCALL,
1721 Op.Val->getVTList(), &Ops[0], Ops.size());
1723 return SDOperand(Chain.Val, Op.ResNo);
1726 //===----------------------------------------------------------------------===//
1727 // X86-64 C Calling Convention implementation
1728 //===----------------------------------------------------------------------===//
1731 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1732 MachineFunction &MF = DAG.getMachineFunction();
1733 MachineFrameInfo *MFI = MF.getFrameInfo();
1734 SDOperand Root = Op.getOperand(0);
1735 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1736 unsigned CC= MF.getFunction()->getCallingConv();
1738 static const unsigned GPR64ArgRegs[] = {
1739 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1741 static const unsigned XMMArgRegs[] = {
1742 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1743 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1747 // Assign locations to all of the incoming arguments.
1748 SmallVector<CCValAssign, 16> ArgLocs;
1749 CCState CCInfo(CC, isVarArg,
1750 getTargetMachine(), ArgLocs);
1751 if (CC == CallingConv::Fast && PerformTailCallOpt)
1752 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1754 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1756 SmallVector<SDOperand, 8> ArgValues;
1757 unsigned LastVal = ~0U;
1758 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1759 CCValAssign &VA = ArgLocs[i];
1760 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1762 assert(VA.getValNo() != LastVal &&
1763 "Don't support value assigned to multiple locs yet");
1764 LastVal = VA.getValNo();
1766 if (VA.isRegLoc()) {
1767 MVT::ValueType RegVT = VA.getLocVT();
1768 TargetRegisterClass *RC;
1769 if (RegVT == MVT::i32)
1770 RC = X86::GR32RegisterClass;
1771 else if (RegVT == MVT::i64)
1772 RC = X86::GR64RegisterClass;
1773 else if (RegVT == MVT::f32)
1774 RC = X86::FR32RegisterClass;
1775 else if (RegVT == MVT::f64)
1776 RC = X86::FR64RegisterClass;
1778 assert(MVT::isVector(RegVT));
1779 if (MVT::getSizeInBits(RegVT) == 64) {
1780 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1783 RC = X86::VR128RegisterClass;
1786 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1787 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1789 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1790 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1792 if (VA.getLocInfo() == CCValAssign::SExt)
1793 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1794 DAG.getValueType(VA.getValVT()));
1795 else if (VA.getLocInfo() == CCValAssign::ZExt)
1796 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1797 DAG.getValueType(VA.getValVT()));
1799 if (VA.getLocInfo() != CCValAssign::Full)
1800 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1802 // Handle MMX values passed in GPRs.
1803 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1804 MVT::getSizeInBits(RegVT) == 64)
1805 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1807 ArgValues.push_back(ArgValue);
1809 assert(VA.isMemLoc());
1810 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1814 unsigned StackSize = CCInfo.getNextStackOffset();
1815 if (CC==CallingConv::Fast)
1816 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1818 // If the function takes variable number of arguments, make a frame index for
1819 // the start of the first vararg value... for expansion of llvm.va_start.
1821 assert(CC!=CallingConv::Fast
1822 && "Var arg not supported with calling convention fastcc");
1823 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1824 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1826 // For X86-64, if there are vararg parameters that are passed via
1827 // registers, then we must store them to their spots on the stack so they
1828 // may be loaded by deferencing the result of va_next.
1829 VarArgsGPOffset = NumIntRegs * 8;
1830 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1831 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1832 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1834 // Store the integer parameter registers.
1835 SmallVector<SDOperand, 8> MemOps;
1836 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1837 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1838 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1839 for (; NumIntRegs != 6; ++NumIntRegs) {
1840 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1841 X86::GR64RegisterClass);
1842 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1843 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1844 MemOps.push_back(Store);
1845 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1846 DAG.getConstant(8, getPointerTy()));
1849 // Now store the XMM (fp + vector) parameter registers.
1850 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1851 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1852 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1853 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1854 X86::VR128RegisterClass);
1855 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1856 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1857 MemOps.push_back(Store);
1858 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1859 DAG.getConstant(16, getPointerTy()));
1861 if (!MemOps.empty())
1862 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1863 &MemOps[0], MemOps.size());
1866 ArgValues.push_back(Root);
1867 // Tail call convention (fastcc) needs callee pop.
1868 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1869 BytesToPopOnReturn = StackSize; // Callee pops everything.
1870 BytesCallerReserves = 0;
1872 BytesToPopOnReturn = 0; // Callee pops nothing.
1873 BytesCallerReserves = StackSize;
1875 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1876 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1878 // Return the new list of results.
1879 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1880 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1884 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1886 SDOperand Chain = Op.getOperand(0);
1887 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1888 SDOperand Callee = Op.getOperand(4);
1890 // Analyze operands of the call, assigning locations to each operand.
1891 SmallVector<CCValAssign, 16> ArgLocs;
1892 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1893 if (CC==CallingConv::Fast && PerformTailCallOpt)
1894 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1896 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1898 // Get a count of how many bytes are to be pushed on the stack.
1899 unsigned NumBytes = CCInfo.getNextStackOffset();
1900 if (CC == CallingConv::Fast)
1901 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1903 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1905 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1906 SmallVector<SDOperand, 8> MemOpChains;
1910 // Walk the register/memloc assignments, inserting copies/loads.
1911 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1912 CCValAssign &VA = ArgLocs[i];
1913 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1915 // Promote the value if needed.
1916 switch (VA.getLocInfo()) {
1917 default: assert(0 && "Unknown loc info!");
1918 case CCValAssign::Full: break;
1919 case CCValAssign::SExt:
1920 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1922 case CCValAssign::ZExt:
1923 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1925 case CCValAssign::AExt:
1926 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1930 if (VA.isRegLoc()) {
1931 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1933 assert(VA.isMemLoc());
1934 if (StackPtr.Val == 0)
1935 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1937 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1942 if (!MemOpChains.empty())
1943 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1944 &MemOpChains[0], MemOpChains.size());
1946 // Build a sequence of copy-to-reg nodes chained together with token chain
1947 // and flag operands which copy the outgoing args into registers.
1949 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1950 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1952 InFlag = Chain.getValue(1);
1956 assert ( CallingConv::Fast != CC &&
1957 "Var args not supported with calling convention fastcc");
1959 // From AMD64 ABI document:
1960 // For calls that may call functions that use varargs or stdargs
1961 // (prototype-less calls or calls to functions containing ellipsis (...) in
1962 // the declaration) %al is used as hidden argument to specify the number
1963 // of SSE registers used. The contents of %al do not need to match exactly
1964 // the number of registers, but must be an ubound on the number of SSE
1965 // registers used and is in the range 0 - 8 inclusive.
1967 // Count the number of XMM registers allocated.
1968 static const unsigned XMMArgRegs[] = {
1969 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1970 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1972 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1974 Chain = DAG.getCopyToReg(Chain, X86::AL,
1975 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1976 InFlag = Chain.getValue(1);
1979 // If the callee is a GlobalAddress node (quite common, every direct call is)
1980 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1981 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1982 // We should use extra load for direct calls to dllimported functions in
1984 if (getTargetMachine().getCodeModel() != CodeModel::Large
1985 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1986 getTargetMachine(), true))
1987 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1988 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1989 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1990 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1992 // Returns a chain & a flag for retval copy to use.
1993 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1994 SmallVector<SDOperand, 8> Ops;
1995 Ops.push_back(Chain);
1996 Ops.push_back(Callee);
1998 // Add argument registers to the end of the list so that they are known live
2000 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2001 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2002 RegsToPass[i].second.getValueType()));
2005 Ops.push_back(InFlag);
2007 Chain = DAG.getNode(X86ISD::CALL,
2008 NodeTys, &Ops[0], Ops.size());
2009 InFlag = Chain.getValue(1);
2010 int NumBytesForCalleeToPush = 0;
2011 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2012 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2014 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2016 // Returns a flag for retval copy to use.
2017 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2019 Ops.push_back(Chain);
2020 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2021 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2022 Ops.push_back(InFlag);
2023 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2024 InFlag = Chain.getValue(1);
2026 // Handle result values, copying them out of physregs into vregs that we
2028 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2032 //===----------------------------------------------------------------------===//
2033 // Other Lowering Hooks
2034 //===----------------------------------------------------------------------===//
2037 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2038 MachineFunction &MF = DAG.getMachineFunction();
2039 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2040 int ReturnAddrIndex = FuncInfo->getRAIndex();
2042 if (ReturnAddrIndex == 0) {
2043 // Set up a frame object for the return address.
2044 if (Subtarget->is64Bit())
2045 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2047 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2049 FuncInfo->setRAIndex(ReturnAddrIndex);
2052 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2057 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2058 /// specific condition code. It returns a false if it cannot do a direct
2059 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2061 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2062 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2063 SelectionDAG &DAG) {
2064 X86CC = X86::COND_INVALID;
2066 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2067 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2068 // X > -1 -> X == 0, jump !sign.
2069 RHS = DAG.getConstant(0, RHS.getValueType());
2070 X86CC = X86::COND_NS;
2072 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2073 // X < 0 -> X == 0, jump on sign.
2074 X86CC = X86::COND_S;
2076 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2078 RHS = DAG.getConstant(0, RHS.getValueType());
2079 X86CC = X86::COND_LE;
2084 switch (SetCCOpcode) {
2086 case ISD::SETEQ: X86CC = X86::COND_E; break;
2087 case ISD::SETGT: X86CC = X86::COND_G; break;
2088 case ISD::SETGE: X86CC = X86::COND_GE; break;
2089 case ISD::SETLT: X86CC = X86::COND_L; break;
2090 case ISD::SETLE: X86CC = X86::COND_LE; break;
2091 case ISD::SETNE: X86CC = X86::COND_NE; break;
2092 case ISD::SETULT: X86CC = X86::COND_B; break;
2093 case ISD::SETUGT: X86CC = X86::COND_A; break;
2094 case ISD::SETULE: X86CC = X86::COND_BE; break;
2095 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2098 // On a floating point condition, the flags are set as follows:
2100 // 0 | 0 | 0 | X > Y
2101 // 0 | 0 | 1 | X < Y
2102 // 1 | 0 | 0 | X == Y
2103 // 1 | 1 | 1 | unordered
2105 switch (SetCCOpcode) {
2108 case ISD::SETEQ: X86CC = X86::COND_E; break;
2109 case ISD::SETOLT: Flip = true; // Fallthrough
2111 case ISD::SETGT: X86CC = X86::COND_A; break;
2112 case ISD::SETOLE: Flip = true; // Fallthrough
2114 case ISD::SETGE: X86CC = X86::COND_AE; break;
2115 case ISD::SETUGT: Flip = true; // Fallthrough
2117 case ISD::SETLT: X86CC = X86::COND_B; break;
2118 case ISD::SETUGE: Flip = true; // Fallthrough
2120 case ISD::SETLE: X86CC = X86::COND_BE; break;
2122 case ISD::SETNE: X86CC = X86::COND_NE; break;
2123 case ISD::SETUO: X86CC = X86::COND_P; break;
2124 case ISD::SETO: X86CC = X86::COND_NP; break;
2127 std::swap(LHS, RHS);
2130 return X86CC != X86::COND_INVALID;
2133 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2134 /// code. Current x86 isa includes the following FP cmov instructions:
2135 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2136 static bool hasFPCMov(unsigned X86CC) {
2152 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2153 /// true if Op is undef or if its value falls within the specified range (L, H].
2154 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2155 if (Op.getOpcode() == ISD::UNDEF)
2158 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2159 return (Val >= Low && Val < Hi);
2162 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2163 /// true if Op is undef or if its value equal to the specified value.
2164 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2165 if (Op.getOpcode() == ISD::UNDEF)
2167 return cast<ConstantSDNode>(Op)->getValue() == Val;
2170 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2171 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2172 bool X86::isPSHUFDMask(SDNode *N) {
2173 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2175 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2178 // Check if the value doesn't reference the second vector.
2179 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2180 SDOperand Arg = N->getOperand(i);
2181 if (Arg.getOpcode() == ISD::UNDEF) continue;
2182 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2183 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2190 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2191 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2192 bool X86::isPSHUFHWMask(SDNode *N) {
2193 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2195 if (N->getNumOperands() != 8)
2198 // Lower quadword copied in order.
2199 for (unsigned i = 0; i != 4; ++i) {
2200 SDOperand Arg = N->getOperand(i);
2201 if (Arg.getOpcode() == ISD::UNDEF) continue;
2202 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2203 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2207 // Upper quadword shuffled.
2208 for (unsigned i = 4; i != 8; ++i) {
2209 SDOperand Arg = N->getOperand(i);
2210 if (Arg.getOpcode() == ISD::UNDEF) continue;
2211 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2212 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2213 if (Val < 4 || Val > 7)
2220 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2221 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2222 bool X86::isPSHUFLWMask(SDNode *N) {
2223 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2225 if (N->getNumOperands() != 8)
2228 // Upper quadword copied in order.
2229 for (unsigned i = 4; i != 8; ++i)
2230 if (!isUndefOrEqual(N->getOperand(i), i))
2233 // Lower quadword shuffled.
2234 for (unsigned i = 0; i != 4; ++i)
2235 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2241 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2242 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2243 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2244 if (NumElems != 2 && NumElems != 4) return false;
2246 unsigned Half = NumElems / 2;
2247 for (unsigned i = 0; i < Half; ++i)
2248 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2250 for (unsigned i = Half; i < NumElems; ++i)
2251 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2257 bool X86::isSHUFPMask(SDNode *N) {
2258 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2259 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2262 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2263 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2264 /// half elements to come from vector 1 (which would equal the dest.) and
2265 /// the upper half to come from vector 2.
2266 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2267 if (NumOps != 2 && NumOps != 4) return false;
2269 unsigned Half = NumOps / 2;
2270 for (unsigned i = 0; i < Half; ++i)
2271 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2273 for (unsigned i = Half; i < NumOps; ++i)
2274 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2279 static bool isCommutedSHUFP(SDNode *N) {
2280 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2281 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2284 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2285 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2286 bool X86::isMOVHLPSMask(SDNode *N) {
2287 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2289 if (N->getNumOperands() != 4)
2292 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2293 return isUndefOrEqual(N->getOperand(0), 6) &&
2294 isUndefOrEqual(N->getOperand(1), 7) &&
2295 isUndefOrEqual(N->getOperand(2), 2) &&
2296 isUndefOrEqual(N->getOperand(3), 3);
2299 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2300 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2302 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2303 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2305 if (N->getNumOperands() != 4)
2308 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2309 return isUndefOrEqual(N->getOperand(0), 2) &&
2310 isUndefOrEqual(N->getOperand(1), 3) &&
2311 isUndefOrEqual(N->getOperand(2), 2) &&
2312 isUndefOrEqual(N->getOperand(3), 3);
2315 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2316 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2317 bool X86::isMOVLPMask(SDNode *N) {
2318 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2320 unsigned NumElems = N->getNumOperands();
2321 if (NumElems != 2 && NumElems != 4)
2324 for (unsigned i = 0; i < NumElems/2; ++i)
2325 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2328 for (unsigned i = NumElems/2; i < NumElems; ++i)
2329 if (!isUndefOrEqual(N->getOperand(i), i))
2335 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2336 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2338 bool X86::isMOVHPMask(SDNode *N) {
2339 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2341 unsigned NumElems = N->getNumOperands();
2342 if (NumElems != 2 && NumElems != 4)
2345 for (unsigned i = 0; i < NumElems/2; ++i)
2346 if (!isUndefOrEqual(N->getOperand(i), i))
2349 for (unsigned i = 0; i < NumElems/2; ++i) {
2350 SDOperand Arg = N->getOperand(i + NumElems/2);
2351 if (!isUndefOrEqual(Arg, i + NumElems))
2358 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2359 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2360 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2361 bool V2IsSplat = false) {
2362 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2365 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2366 SDOperand BitI = Elts[i];
2367 SDOperand BitI1 = Elts[i+1];
2368 if (!isUndefOrEqual(BitI, j))
2371 if (isUndefOrEqual(BitI1, NumElts))
2374 if (!isUndefOrEqual(BitI1, j + NumElts))
2382 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2383 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2384 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2387 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2388 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2389 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2390 bool V2IsSplat = false) {
2391 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2394 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2395 SDOperand BitI = Elts[i];
2396 SDOperand BitI1 = Elts[i+1];
2397 if (!isUndefOrEqual(BitI, j + NumElts/2))
2400 if (isUndefOrEqual(BitI1, NumElts))
2403 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2411 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2412 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2413 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2416 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2417 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2419 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2420 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2422 unsigned NumElems = N->getNumOperands();
2423 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2426 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2427 SDOperand BitI = N->getOperand(i);
2428 SDOperand BitI1 = N->getOperand(i+1);
2430 if (!isUndefOrEqual(BitI, j))
2432 if (!isUndefOrEqual(BitI1, j))
2439 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2440 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2442 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2443 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2445 unsigned NumElems = N->getNumOperands();
2446 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2449 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2450 SDOperand BitI = N->getOperand(i);
2451 SDOperand BitI1 = N->getOperand(i + 1);
2453 if (!isUndefOrEqual(BitI, j))
2455 if (!isUndefOrEqual(BitI1, j))
2462 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2463 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2464 /// MOVSD, and MOVD, i.e. setting the lowest element.
2465 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2466 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2469 if (!isUndefOrEqual(Elts[0], NumElts))
2472 for (unsigned i = 1; i < NumElts; ++i) {
2473 if (!isUndefOrEqual(Elts[i], i))
2480 bool X86::isMOVLMask(SDNode *N) {
2481 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2482 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2485 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2486 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2487 /// element of vector 2 and the other elements to come from vector 1 in order.
2488 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2489 bool V2IsSplat = false,
2490 bool V2IsUndef = false) {
2491 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2494 if (!isUndefOrEqual(Ops[0], 0))
2497 for (unsigned i = 1; i < NumOps; ++i) {
2498 SDOperand Arg = Ops[i];
2499 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2500 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2501 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2508 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2509 bool V2IsUndef = false) {
2510 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2511 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2512 V2IsSplat, V2IsUndef);
2515 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2516 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2517 bool X86::isMOVSHDUPMask(SDNode *N) {
2518 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2520 if (N->getNumOperands() != 4)
2523 // Expect 1, 1, 3, 3
2524 for (unsigned i = 0; i < 2; ++i) {
2525 SDOperand Arg = N->getOperand(i);
2526 if (Arg.getOpcode() == ISD::UNDEF) continue;
2527 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2528 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2529 if (Val != 1) return false;
2533 for (unsigned i = 2; i < 4; ++i) {
2534 SDOperand Arg = N->getOperand(i);
2535 if (Arg.getOpcode() == ISD::UNDEF) continue;
2536 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2537 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2538 if (Val != 3) return false;
2542 // Don't use movshdup if it can be done with a shufps.
2546 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2547 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2548 bool X86::isMOVSLDUPMask(SDNode *N) {
2549 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2551 if (N->getNumOperands() != 4)
2554 // Expect 0, 0, 2, 2
2555 for (unsigned i = 0; i < 2; ++i) {
2556 SDOperand Arg = N->getOperand(i);
2557 if (Arg.getOpcode() == ISD::UNDEF) continue;
2558 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2559 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2560 if (Val != 0) return false;
2564 for (unsigned i = 2; i < 4; ++i) {
2565 SDOperand Arg = N->getOperand(i);
2566 if (Arg.getOpcode() == ISD::UNDEF) continue;
2567 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2568 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2569 if (Val != 2) return false;
2573 // Don't use movshdup if it can be done with a shufps.
2577 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2578 /// specifies a identity operation on the LHS or RHS.
2579 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2580 unsigned NumElems = N->getNumOperands();
2581 for (unsigned i = 0; i < NumElems; ++i)
2582 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2587 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2588 /// a splat of a single element.
2589 static bool isSplatMask(SDNode *N) {
2590 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2592 // This is a splat operation if each element of the permute is the same, and
2593 // if the value doesn't reference the second vector.
2594 unsigned NumElems = N->getNumOperands();
2595 SDOperand ElementBase;
2597 for (; i != NumElems; ++i) {
2598 SDOperand Elt = N->getOperand(i);
2599 if (isa<ConstantSDNode>(Elt)) {
2605 if (!ElementBase.Val)
2608 for (; i != NumElems; ++i) {
2609 SDOperand Arg = N->getOperand(i);
2610 if (Arg.getOpcode() == ISD::UNDEF) continue;
2611 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2612 if (Arg != ElementBase) return false;
2615 // Make sure it is a splat of the first vector operand.
2616 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2619 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2620 /// a splat of a single element and it's a 2 or 4 element mask.
2621 bool X86::isSplatMask(SDNode *N) {
2622 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2624 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2625 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2627 return ::isSplatMask(N);
2630 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2631 /// specifies a splat of zero element.
2632 bool X86::isSplatLoMask(SDNode *N) {
2633 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2635 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2636 if (!isUndefOrEqual(N->getOperand(i), 0))
2641 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2642 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2644 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2645 unsigned NumOperands = N->getNumOperands();
2646 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2648 for (unsigned i = 0; i < NumOperands; ++i) {
2650 SDOperand Arg = N->getOperand(NumOperands-i-1);
2651 if (Arg.getOpcode() != ISD::UNDEF)
2652 Val = cast<ConstantSDNode>(Arg)->getValue();
2653 if (Val >= NumOperands) Val -= NumOperands;
2655 if (i != NumOperands - 1)
2662 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2663 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2665 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2667 // 8 nodes, but we only care about the last 4.
2668 for (unsigned i = 7; i >= 4; --i) {
2670 SDOperand Arg = N->getOperand(i);
2671 if (Arg.getOpcode() != ISD::UNDEF)
2672 Val = cast<ConstantSDNode>(Arg)->getValue();
2681 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2682 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2684 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2686 // 8 nodes, but we only care about the first 4.
2687 for (int i = 3; i >= 0; --i) {
2689 SDOperand Arg = N->getOperand(i);
2690 if (Arg.getOpcode() != ISD::UNDEF)
2691 Val = cast<ConstantSDNode>(Arg)->getValue();
2700 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2701 /// specifies a 8 element shuffle that can be broken into a pair of
2702 /// PSHUFHW and PSHUFLW.
2703 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2704 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2706 if (N->getNumOperands() != 8)
2709 // Lower quadword shuffled.
2710 for (unsigned i = 0; i != 4; ++i) {
2711 SDOperand Arg = N->getOperand(i);
2712 if (Arg.getOpcode() == ISD::UNDEF) continue;
2713 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2714 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2719 // Upper quadword shuffled.
2720 for (unsigned i = 4; i != 8; ++i) {
2721 SDOperand Arg = N->getOperand(i);
2722 if (Arg.getOpcode() == ISD::UNDEF) continue;
2723 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2724 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2725 if (Val < 4 || Val > 7)
2732 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2733 /// values in ther permute mask.
2734 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2735 SDOperand &V2, SDOperand &Mask,
2736 SelectionDAG &DAG) {
2737 MVT::ValueType VT = Op.getValueType();
2738 MVT::ValueType MaskVT = Mask.getValueType();
2739 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2740 unsigned NumElems = Mask.getNumOperands();
2741 SmallVector<SDOperand, 8> MaskVec;
2743 for (unsigned i = 0; i != NumElems; ++i) {
2744 SDOperand Arg = Mask.getOperand(i);
2745 if (Arg.getOpcode() == ISD::UNDEF) {
2746 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2749 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2750 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2752 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2754 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2758 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2759 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2762 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2763 /// match movhlps. The lower half elements should come from upper half of
2764 /// V1 (and in order), and the upper half elements should come from the upper
2765 /// half of V2 (and in order).
2766 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2767 unsigned NumElems = Mask->getNumOperands();
2770 for (unsigned i = 0, e = 2; i != e; ++i)
2771 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2773 for (unsigned i = 2; i != 4; ++i)
2774 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2779 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2780 /// is promoted to a vector.
2781 static inline bool isScalarLoadToVector(SDNode *N) {
2782 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2783 N = N->getOperand(0).Val;
2784 return ISD::isNON_EXTLoad(N);
2789 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2790 /// match movlp{s|d}. The lower half elements should come from lower half of
2791 /// V1 (and in order), and the upper half elements should come from the upper
2792 /// half of V2 (and in order). And since V1 will become the source of the
2793 /// MOVLP, it must be either a vector load or a scalar load to vector.
2794 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2795 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2797 // Is V2 is a vector load, don't do this transformation. We will try to use
2798 // load folding shufps op.
2799 if (ISD::isNON_EXTLoad(V2))
2802 unsigned NumElems = Mask->getNumOperands();
2803 if (NumElems != 2 && NumElems != 4)
2805 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2806 if (!isUndefOrEqual(Mask->getOperand(i), i))
2808 for (unsigned i = NumElems/2; i != NumElems; ++i)
2809 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2814 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2816 static bool isSplatVector(SDNode *N) {
2817 if (N->getOpcode() != ISD::BUILD_VECTOR)
2820 SDOperand SplatValue = N->getOperand(0);
2821 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2822 if (N->getOperand(i) != SplatValue)
2827 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2829 static bool isUndefShuffle(SDNode *N) {
2830 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2833 SDOperand V1 = N->getOperand(0);
2834 SDOperand V2 = N->getOperand(1);
2835 SDOperand Mask = N->getOperand(2);
2836 unsigned NumElems = Mask.getNumOperands();
2837 for (unsigned i = 0; i != NumElems; ++i) {
2838 SDOperand Arg = Mask.getOperand(i);
2839 if (Arg.getOpcode() != ISD::UNDEF) {
2840 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2841 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2843 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2850 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2852 static inline bool isZeroNode(SDOperand Elt) {
2853 return ((isa<ConstantSDNode>(Elt) &&
2854 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2855 (isa<ConstantFPSDNode>(Elt) &&
2856 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2859 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2860 /// to an zero vector.
2861 static bool isZeroShuffle(SDNode *N) {
2862 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2865 SDOperand V1 = N->getOperand(0);
2866 SDOperand V2 = N->getOperand(1);
2867 SDOperand Mask = N->getOperand(2);
2868 unsigned NumElems = Mask.getNumOperands();
2869 for (unsigned i = 0; i != NumElems; ++i) {
2870 SDOperand Arg = Mask.getOperand(i);
2871 if (Arg.getOpcode() != ISD::UNDEF) {
2872 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2873 if (Idx < NumElems) {
2874 unsigned Opc = V1.Val->getOpcode();
2875 if (Opc == ISD::UNDEF)
2877 if (Opc != ISD::BUILD_VECTOR ||
2878 !isZeroNode(V1.Val->getOperand(Idx)))
2880 } else if (Idx >= NumElems) {
2881 unsigned Opc = V2.Val->getOpcode();
2882 if (Opc == ISD::UNDEF)
2884 if (Opc != ISD::BUILD_VECTOR ||
2885 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2893 /// getZeroVector - Returns a vector of specified type with all zero elements.
2895 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2896 assert(MVT::isVector(VT) && "Expected a vector type");
2897 unsigned NumElems = MVT::getVectorNumElements(VT);
2898 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2899 bool isFP = MVT::isFloatingPoint(EVT);
2900 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2901 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2902 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2905 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2906 /// that point to V2 points to its first element.
2907 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2908 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2910 bool Changed = false;
2911 SmallVector<SDOperand, 8> MaskVec;
2912 unsigned NumElems = Mask.getNumOperands();
2913 for (unsigned i = 0; i != NumElems; ++i) {
2914 SDOperand Arg = Mask.getOperand(i);
2915 if (Arg.getOpcode() != ISD::UNDEF) {
2916 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2917 if (Val > NumElems) {
2918 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2922 MaskVec.push_back(Arg);
2926 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2927 &MaskVec[0], MaskVec.size());
2931 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2932 /// operation of specified width.
2933 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2934 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2935 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2937 SmallVector<SDOperand, 8> MaskVec;
2938 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2939 for (unsigned i = 1; i != NumElems; ++i)
2940 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2941 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2944 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2945 /// of specified width.
2946 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2947 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2948 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2949 SmallVector<SDOperand, 8> MaskVec;
2950 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2951 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2952 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2954 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2957 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2958 /// of specified width.
2959 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2960 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2961 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2962 unsigned Half = NumElems/2;
2963 SmallVector<SDOperand, 8> MaskVec;
2964 for (unsigned i = 0; i != Half; ++i) {
2965 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2966 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2968 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2971 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2973 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2974 SDOperand V1 = Op.getOperand(0);
2975 SDOperand Mask = Op.getOperand(2);
2976 MVT::ValueType VT = Op.getValueType();
2977 unsigned NumElems = Mask.getNumOperands();
2978 Mask = getUnpacklMask(NumElems, DAG);
2979 while (NumElems != 4) {
2980 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2983 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2985 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2986 Mask = getZeroVector(MaskVT, DAG);
2987 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2988 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2989 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2992 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2993 /// vector of zero or undef vector.
2994 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2995 unsigned NumElems, unsigned Idx,
2996 bool isZero, SelectionDAG &DAG) {
2997 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2998 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2999 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3000 SDOperand Zero = DAG.getConstant(0, EVT);
3001 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
3002 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
3003 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3004 &MaskVec[0], MaskVec.size());
3005 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3008 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3010 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
3011 unsigned NumNonZero, unsigned NumZero,
3012 SelectionDAG &DAG, TargetLowering &TLI) {
3018 for (unsigned i = 0; i < 16; ++i) {
3019 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3020 if (ThisIsNonZero && First) {
3022 V = getZeroVector(MVT::v8i16, DAG);
3024 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3029 SDOperand ThisElt(0, 0), LastElt(0, 0);
3030 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3031 if (LastIsNonZero) {
3032 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3034 if (ThisIsNonZero) {
3035 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3036 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3037 ThisElt, DAG.getConstant(8, MVT::i8));
3039 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3044 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3045 DAG.getConstant(i/2, TLI.getPointerTy()));
3049 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3052 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3054 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3055 unsigned NumNonZero, unsigned NumZero,
3056 SelectionDAG &DAG, TargetLowering &TLI) {
3062 for (unsigned i = 0; i < 8; ++i) {
3063 bool isNonZero = (NonZeros & (1 << i)) != 0;
3067 V = getZeroVector(MVT::v8i16, DAG);
3069 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3072 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3073 DAG.getConstant(i, TLI.getPointerTy()));
3081 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3082 // All zero's are handled with pxor.
3083 if (ISD::isBuildVectorAllZeros(Op.Val))
3086 // All one's are handled with pcmpeqd.
3087 if (ISD::isBuildVectorAllOnes(Op.Val))
3090 MVT::ValueType VT = Op.getValueType();
3091 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3092 unsigned EVTBits = MVT::getSizeInBits(EVT);
3094 unsigned NumElems = Op.getNumOperands();
3095 unsigned NumZero = 0;
3096 unsigned NumNonZero = 0;
3097 unsigned NonZeros = 0;
3098 unsigned NumNonZeroImms = 0;
3099 std::set<SDOperand> Values;
3100 for (unsigned i = 0; i < NumElems; ++i) {
3101 SDOperand Elt = Op.getOperand(i);
3102 if (Elt.getOpcode() != ISD::UNDEF) {
3104 if (isZeroNode(Elt))
3107 NonZeros |= (1 << i);
3109 if (Elt.getOpcode() == ISD::Constant ||
3110 Elt.getOpcode() == ISD::ConstantFP)
3116 if (NumNonZero == 0) {
3118 // All undef vector. Return an UNDEF.
3119 return DAG.getNode(ISD::UNDEF, VT);
3121 // A mix of zero and undef. Return a zero vector.
3122 return getZeroVector(VT, DAG);
3125 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3126 if (Values.size() == 1)
3129 // Special case for single non-zero element.
3130 if (NumNonZero == 1) {
3131 unsigned Idx = CountTrailingZeros_32(NonZeros);
3132 SDOperand Item = Op.getOperand(Idx);
3133 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3135 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3136 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3139 if (EVTBits == 32) {
3140 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3141 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3143 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3144 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3145 SmallVector<SDOperand, 8> MaskVec;
3146 for (unsigned i = 0; i < NumElems; i++)
3147 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3148 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3149 &MaskVec[0], MaskVec.size());
3150 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3151 DAG.getNode(ISD::UNDEF, VT), Mask);
3155 // A vector full of immediates; various special cases are already
3156 // handled, so this is best done with a single constant-pool load.
3157 if (NumNonZero == NumNonZeroImms)
3160 // Let legalizer expand 2-wide build_vectors.
3164 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3165 if (EVTBits == 8 && NumElems == 16) {
3166 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3168 if (V.Val) return V;
3171 if (EVTBits == 16 && NumElems == 8) {
3172 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3174 if (V.Val) return V;
3177 // If element VT is == 32 bits, turn it into a number of shuffles.
3178 SmallVector<SDOperand, 8> V;
3180 if (NumElems == 4 && NumZero > 0) {
3181 for (unsigned i = 0; i < 4; ++i) {
3182 bool isZero = !(NonZeros & (1 << i));
3184 V[i] = getZeroVector(VT, DAG);
3186 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3189 for (unsigned i = 0; i < 2; ++i) {
3190 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3193 V[i] = V[i*2]; // Must be a zero vector.
3196 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3197 getMOVLMask(NumElems, DAG));
3200 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3201 getMOVLMask(NumElems, DAG));
3204 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3205 getUnpacklMask(NumElems, DAG));
3210 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3211 // clears the upper bits.
3212 // FIXME: we can do the same for v4f32 case when we know both parts of
3213 // the lower half come from scalar_to_vector (loadf32). We should do
3214 // that in post legalizer dag combiner with target specific hooks.
3215 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3217 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3218 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3219 SmallVector<SDOperand, 8> MaskVec;
3220 bool Reverse = (NonZeros & 0x3) == 2;
3221 for (unsigned i = 0; i < 2; ++i)
3223 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3225 MaskVec.push_back(DAG.getConstant(i, EVT));
3226 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3227 for (unsigned i = 0; i < 2; ++i)
3229 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3231 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3232 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3233 &MaskVec[0], MaskVec.size());
3234 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3237 if (Values.size() > 2) {
3238 // Expand into a number of unpckl*.
3240 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3241 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3242 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3243 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3244 for (unsigned i = 0; i < NumElems; ++i)
3245 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3247 while (NumElems != 0) {
3248 for (unsigned i = 0; i < NumElems; ++i)
3249 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3260 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3261 SDOperand V1 = Op.getOperand(0);
3262 SDOperand V2 = Op.getOperand(1);
3263 SDOperand PermMask = Op.getOperand(2);
3264 MVT::ValueType VT = Op.getValueType();
3265 unsigned NumElems = PermMask.getNumOperands();
3266 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3267 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3268 bool V1IsSplat = false;
3269 bool V2IsSplat = false;
3271 if (isUndefShuffle(Op.Val))
3272 return DAG.getNode(ISD::UNDEF, VT);
3274 if (isZeroShuffle(Op.Val))
3275 return getZeroVector(VT, DAG);
3277 if (isIdentityMask(PermMask.Val))
3279 else if (isIdentityMask(PermMask.Val, true))
3282 if (isSplatMask(PermMask.Val)) {
3283 if (NumElems <= 4) return Op;
3284 // Promote it to a v4i32 splat.
3285 return PromoteSplat(Op, DAG);
3288 if (X86::isMOVLMask(PermMask.Val))
3289 return (V1IsUndef) ? V2 : Op;
3291 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3292 X86::isMOVSLDUPMask(PermMask.Val) ||
3293 X86::isMOVHLPSMask(PermMask.Val) ||
3294 X86::isMOVHPMask(PermMask.Val) ||
3295 X86::isMOVLPMask(PermMask.Val))
3298 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3299 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3300 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3302 bool Commuted = false;
3303 V1IsSplat = isSplatVector(V1.Val);
3304 V2IsSplat = isSplatVector(V2.Val);
3305 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3306 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3307 std::swap(V1IsSplat, V2IsSplat);
3308 std::swap(V1IsUndef, V2IsUndef);
3312 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3313 if (V2IsUndef) return V1;
3314 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3316 // V2 is a splat, so the mask may be malformed. That is, it may point
3317 // to any V2 element. The instruction selectior won't like this. Get
3318 // a corrected mask and commute to form a proper MOVS{S|D}.
3319 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3320 if (NewMask.Val != PermMask.Val)
3321 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3326 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3327 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3328 X86::isUNPCKLMask(PermMask.Val) ||
3329 X86::isUNPCKHMask(PermMask.Val))
3333 // Normalize mask so all entries that point to V2 points to its first
3334 // element then try to match unpck{h|l} again. If match, return a
3335 // new vector_shuffle with the corrected mask.
3336 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3337 if (NewMask.Val != PermMask.Val) {
3338 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3339 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3340 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3341 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3342 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3343 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3348 // Normalize the node to match x86 shuffle ops if needed
3349 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3350 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3353 // Commute is back and try unpck* again.
3354 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3355 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3356 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3357 X86::isUNPCKLMask(PermMask.Val) ||
3358 X86::isUNPCKHMask(PermMask.Val))
3362 // If VT is integer, try PSHUF* first, then SHUFP*.
3363 if (MVT::isInteger(VT)) {
3364 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3365 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3366 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3367 X86::isPSHUFDMask(PermMask.Val)) ||
3368 X86::isPSHUFHWMask(PermMask.Val) ||
3369 X86::isPSHUFLWMask(PermMask.Val)) {
3370 if (V2.getOpcode() != ISD::UNDEF)
3371 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3372 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3376 if (X86::isSHUFPMask(PermMask.Val) &&
3377 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3380 // Handle v8i16 shuffle high / low shuffle node pair.
3381 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3382 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3383 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3384 SmallVector<SDOperand, 8> MaskVec;
3385 for (unsigned i = 0; i != 4; ++i)
3386 MaskVec.push_back(PermMask.getOperand(i));
3387 for (unsigned i = 4; i != 8; ++i)
3388 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3389 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3390 &MaskVec[0], MaskVec.size());
3391 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3393 for (unsigned i = 0; i != 4; ++i)
3394 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3395 for (unsigned i = 4; i != 8; ++i)
3396 MaskVec.push_back(PermMask.getOperand(i));
3397 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3398 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3401 // Floating point cases in the other order.
3402 if (X86::isSHUFPMask(PermMask.Val))
3404 if (X86::isPSHUFDMask(PermMask.Val) ||
3405 X86::isPSHUFHWMask(PermMask.Val) ||
3406 X86::isPSHUFLWMask(PermMask.Val)) {
3407 if (V2.getOpcode() != ISD::UNDEF)
3408 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3409 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3414 if (NumElems == 4 &&
3415 // Don't do this for MMX.
3416 MVT::getSizeInBits(VT) != 64) {
3417 MVT::ValueType MaskVT = PermMask.getValueType();
3418 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3419 SmallVector<std::pair<int, int>, 8> Locs;
3420 Locs.reserve(NumElems);
3421 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3422 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3425 // If no more than two elements come from either vector. This can be
3426 // implemented with two shuffles. First shuffle gather the elements.
3427 // The second shuffle, which takes the first shuffle as both of its
3428 // vector operands, put the elements into the right order.
3429 for (unsigned i = 0; i != NumElems; ++i) {
3430 SDOperand Elt = PermMask.getOperand(i);
3431 if (Elt.getOpcode() == ISD::UNDEF) {
3432 Locs[i] = std::make_pair(-1, -1);
3434 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3435 if (Val < NumElems) {
3436 Locs[i] = std::make_pair(0, NumLo);
3440 Locs[i] = std::make_pair(1, NumHi);
3441 if (2+NumHi < NumElems)
3442 Mask1[2+NumHi] = Elt;
3447 if (NumLo <= 2 && NumHi <= 2) {
3448 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3449 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3450 &Mask1[0], Mask1.size()));
3451 for (unsigned i = 0; i != NumElems; ++i) {
3452 if (Locs[i].first == -1)
3455 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3456 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3457 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3461 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3462 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3463 &Mask2[0], Mask2.size()));
3466 // Break it into (shuffle shuffle_hi, shuffle_lo).
3468 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3469 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3470 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3471 unsigned MaskIdx = 0;
3473 unsigned HiIdx = NumElems/2;
3474 for (unsigned i = 0; i != NumElems; ++i) {
3475 if (i == NumElems/2) {
3481 SDOperand Elt = PermMask.getOperand(i);
3482 if (Elt.getOpcode() == ISD::UNDEF) {
3483 Locs[i] = std::make_pair(-1, -1);
3484 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3485 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3486 (*MaskPtr)[LoIdx] = Elt;
3489 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3490 (*MaskPtr)[HiIdx] = Elt;
3495 SDOperand LoShuffle =
3496 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3497 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3498 &LoMask[0], LoMask.size()));
3499 SDOperand HiShuffle =
3500 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3501 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3502 &HiMask[0], HiMask.size()));
3503 SmallVector<SDOperand, 8> MaskOps;
3504 for (unsigned i = 0; i != NumElems; ++i) {
3505 if (Locs[i].first == -1) {
3506 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3508 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3509 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3512 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3513 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3514 &MaskOps[0], MaskOps.size()));
3521 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3522 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3525 MVT::ValueType VT = Op.getValueType();
3526 // TODO: handle v16i8.
3527 if (MVT::getSizeInBits(VT) == 16) {
3528 // Transform it so it match pextrw which produces a 32-bit result.
3529 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3530 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3531 Op.getOperand(0), Op.getOperand(1));
3532 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3533 DAG.getValueType(VT));
3534 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3535 } else if (MVT::getSizeInBits(VT) == 32) {
3536 SDOperand Vec = Op.getOperand(0);
3537 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3540 // SHUFPS the element to the lowest double word, then movss.
3541 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3542 SmallVector<SDOperand, 8> IdxVec;
3544 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3546 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3548 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3550 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3551 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3552 &IdxVec[0], IdxVec.size());
3553 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3554 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3555 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3556 DAG.getConstant(0, getPointerTy()));
3557 } else if (MVT::getSizeInBits(VT) == 64) {
3558 SDOperand Vec = Op.getOperand(0);
3559 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3563 // UNPCKHPD the element to the lowest double word, then movsd.
3564 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3565 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3566 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3567 SmallVector<SDOperand, 8> IdxVec;
3568 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3570 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3571 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3572 &IdxVec[0], IdxVec.size());
3573 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3574 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3575 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3576 DAG.getConstant(0, getPointerTy()));
3583 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3584 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3585 // as its second argument.
3586 MVT::ValueType VT = Op.getValueType();
3587 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3588 SDOperand N0 = Op.getOperand(0);
3589 SDOperand N1 = Op.getOperand(1);
3590 SDOperand N2 = Op.getOperand(2);
3591 if (MVT::getSizeInBits(BaseVT) == 16) {
3592 if (N1.getValueType() != MVT::i32)
3593 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3594 if (N2.getValueType() != MVT::i32)
3595 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3596 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3597 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3598 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3601 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3602 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3603 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3604 SmallVector<SDOperand, 8> MaskVec;
3605 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3606 for (unsigned i = 1; i <= 3; ++i)
3607 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3608 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3609 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3610 &MaskVec[0], MaskVec.size()));
3612 // Use two pinsrw instructions to insert a 32 bit value.
3614 if (MVT::isFloatingPoint(N1.getValueType())) {
3615 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3616 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3617 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3618 DAG.getConstant(0, getPointerTy()));
3620 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3621 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3622 DAG.getConstant(Idx, getPointerTy()));
3623 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3624 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3625 DAG.getConstant(Idx+1, getPointerTy()));
3626 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3634 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3635 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3636 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3639 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3640 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3641 // one of the above mentioned nodes. It has to be wrapped because otherwise
3642 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3643 // be used to form addressing mode. These wrapped nodes will be selected
3646 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3647 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3648 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3650 CP->getAlignment());
3651 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3652 // With PIC, the address is actually $g + Offset.
3653 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3654 !Subtarget->isPICStyleRIPRel()) {
3655 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3656 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3664 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3665 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3666 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3667 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3668 // With PIC, the address is actually $g + Offset.
3669 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3670 !Subtarget->isPICStyleRIPRel()) {
3671 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3672 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3676 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3677 // load the value at address GV, not the value of GV itself. This means that
3678 // the GlobalAddress must be in the base or index register of the address, not
3679 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3680 // The same applies for external symbols during PIC codegen
3681 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3682 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3687 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3689 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3690 const MVT::ValueType PtrVT) {
3692 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3693 DAG.getNode(X86ISD::GlobalBaseReg,
3695 InFlag = Chain.getValue(1);
3697 // emit leal symbol@TLSGD(,%ebx,1), %eax
3698 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3699 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3700 GA->getValueType(0),
3702 SDOperand Ops[] = { Chain, TGA, InFlag };
3703 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3704 InFlag = Result.getValue(2);
3705 Chain = Result.getValue(1);
3707 // call ___tls_get_addr. This function receives its argument in
3708 // the register EAX.
3709 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3710 InFlag = Chain.getValue(1);
3712 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3713 SDOperand Ops1[] = { Chain,
3714 DAG.getTargetExternalSymbol("___tls_get_addr",
3716 DAG.getRegister(X86::EAX, PtrVT),
3717 DAG.getRegister(X86::EBX, PtrVT),
3719 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3720 InFlag = Chain.getValue(1);
3722 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3725 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3726 // "local exec" model.
3728 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3729 const MVT::ValueType PtrVT) {
3730 // Get the Thread Pointer
3731 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3732 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3734 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3735 GA->getValueType(0),
3737 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3739 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3740 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3742 // The address of the thread local variable is the add of the thread
3743 // pointer with the offset of the variable.
3744 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3748 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3749 // TODO: implement the "local dynamic" model
3750 // TODO: implement the "initial exec"model for pic executables
3751 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3752 "TLS not implemented for non-ELF and 64-bit targets");
3753 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3754 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3755 // otherwise use the "Local Exec"TLS Model
3756 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3757 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3759 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3763 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3764 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3765 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3766 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3767 // With PIC, the address is actually $g + Offset.
3768 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3769 !Subtarget->isPICStyleRIPRel()) {
3770 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3771 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3778 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3779 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3780 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3781 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3782 // With PIC, the address is actually $g + Offset.
3783 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3784 !Subtarget->isPICStyleRIPRel()) {
3785 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3786 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3793 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3794 /// take a 2 x i32 value to shift plus a shift amount.
3795 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3796 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3797 "Not an i64 shift!");
3798 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3799 SDOperand ShOpLo = Op.getOperand(0);
3800 SDOperand ShOpHi = Op.getOperand(1);
3801 SDOperand ShAmt = Op.getOperand(2);
3802 SDOperand Tmp1 = isSRA ?
3803 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3804 DAG.getConstant(0, MVT::i32);
3806 SDOperand Tmp2, Tmp3;
3807 if (Op.getOpcode() == ISD::SHL_PARTS) {
3808 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3809 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3811 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3812 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3815 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3816 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3817 DAG.getConstant(32, MVT::i8));
3818 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3819 AndNode, DAG.getConstant(0, MVT::i8));
3822 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3823 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3824 SmallVector<SDOperand, 4> Ops;
3825 if (Op.getOpcode() == ISD::SHL_PARTS) {
3826 Ops.push_back(Tmp2);
3827 Ops.push_back(Tmp3);
3829 Ops.push_back(Cond);
3830 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3833 Ops.push_back(Tmp3);
3834 Ops.push_back(Tmp1);
3836 Ops.push_back(Cond);
3837 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3839 Ops.push_back(Tmp2);
3840 Ops.push_back(Tmp3);
3842 Ops.push_back(Cond);
3843 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3846 Ops.push_back(Tmp3);
3847 Ops.push_back(Tmp1);
3849 Ops.push_back(Cond);
3850 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3853 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3857 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3860 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3861 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3862 Op.getOperand(0).getValueType() >= MVT::i16 &&
3863 "Unknown SINT_TO_FP to lower!");
3866 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3867 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3868 MachineFunction &MF = DAG.getMachineFunction();
3869 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3870 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3871 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3872 StackSlot, NULL, 0);
3874 // These are really Legal; caller falls through into that case.
3875 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
3877 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
3879 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
3880 Subtarget->is64Bit())
3885 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
3886 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
3888 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3890 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3891 SmallVector<SDOperand, 8> Ops;
3892 Ops.push_back(Chain);
3893 Ops.push_back(StackSlot);
3894 Ops.push_back(DAG.getValueType(SrcVT));
3895 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3896 Tys, &Ops[0], Ops.size());
3899 Chain = Result.getValue(1);
3900 SDOperand InFlag = Result.getValue(2);
3902 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3903 // shouldn't be necessary except that RFP cannot be live across
3904 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3905 MachineFunction &MF = DAG.getMachineFunction();
3906 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3907 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3908 Tys = DAG.getVTList(MVT::Other);
3909 SmallVector<SDOperand, 8> Ops;
3910 Ops.push_back(Chain);
3911 Ops.push_back(Result);
3912 Ops.push_back(StackSlot);
3913 Ops.push_back(DAG.getValueType(Op.getValueType()));
3914 Ops.push_back(InFlag);
3915 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3916 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3922 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3923 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3924 "Unknown FP_TO_SINT to lower!");
3927 // These are really Legal.
3928 if (Op.getValueType() == MVT::i32 &&
3929 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
3931 if (Op.getValueType() == MVT::i32 &&
3932 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
3934 if (Subtarget->is64Bit() &&
3935 Op.getValueType() == MVT::i64 &&
3936 Op.getOperand(0).getValueType() != MVT::f80)
3939 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3941 MachineFunction &MF = DAG.getMachineFunction();
3942 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3943 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3944 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3946 switch (Op.getValueType()) {
3947 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3948 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3949 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3950 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3953 SDOperand Chain = DAG.getEntryNode();
3954 SDOperand Value = Op.getOperand(0);
3955 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
3956 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
3957 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3958 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3959 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3961 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3963 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3964 Chain = Value.getValue(1);
3965 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3966 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3969 // Build the FP_TO_INT*_IN_MEM
3970 SDOperand Ops[] = { Chain, Value, StackSlot };
3971 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3973 // Load the result. If this is an i64 load on an x86-32 host, expand the
3975 if (Op.getValueType() != MVT::i64 || Subtarget->is64Bit())
3976 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3978 SDOperand Lo = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3979 StackSlot = DAG.getNode(ISD::ADD, StackSlot.getValueType(), StackSlot,
3980 DAG.getConstant(StackSlot.getValueType(), 4));
3981 SDOperand Hi = DAG.getLoad(MVT::i32, FIST, StackSlot, NULL, 0);
3984 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
3987 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3988 MVT::ValueType VT = Op.getValueType();
3989 MVT::ValueType EltVT = VT;
3990 if (MVT::isVector(VT))
3991 EltVT = MVT::getVectorElementType(VT);
3992 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3993 std::vector<Constant*> CV;
3994 if (EltVT == MVT::f64) {
3995 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
3999 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4005 Constant *C = ConstantVector::get(CV);
4006 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4007 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4009 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4012 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4013 MVT::ValueType VT = Op.getValueType();
4014 MVT::ValueType EltVT = VT;
4015 unsigned EltNum = 1;
4016 if (MVT::isVector(VT)) {
4017 EltVT = MVT::getVectorElementType(VT);
4018 EltNum = MVT::getVectorNumElements(VT);
4020 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4021 std::vector<Constant*> CV;
4022 if (EltVT == MVT::f64) {
4023 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4027 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4033 Constant *C = ConstantVector::get(CV);
4034 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4035 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4037 if (MVT::isVector(VT)) {
4038 return DAG.getNode(ISD::BIT_CONVERT, VT,
4039 DAG.getNode(ISD::XOR, MVT::v2i64,
4040 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4041 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4043 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4047 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4048 SDOperand Op0 = Op.getOperand(0);
4049 SDOperand Op1 = Op.getOperand(1);
4050 MVT::ValueType VT = Op.getValueType();
4051 MVT::ValueType SrcVT = Op1.getValueType();
4052 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4054 // If second operand is smaller, extend it first.
4055 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4056 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4058 SrcTy = MVT::getTypeForValueType(SrcVT);
4060 // And if it is bigger, shrink it first.
4061 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4062 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1);
4064 SrcTy = MVT::getTypeForValueType(SrcVT);
4067 // At this point the operands and the result should have the same
4068 // type, and that won't be f80 since that is not custom lowered.
4070 // First get the sign bit of second operand.
4071 std::vector<Constant*> CV;
4072 if (SrcVT == MVT::f64) {
4073 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4074 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4076 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4077 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4078 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4079 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4081 Constant *C = ConstantVector::get(CV);
4082 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4083 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4085 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4087 // Shift sign bit right or left if the two operands have different types.
4088 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4089 // Op0 is MVT::f32, Op1 is MVT::f64.
4090 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4091 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4092 DAG.getConstant(32, MVT::i32));
4093 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4094 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4095 DAG.getConstant(0, getPointerTy()));
4098 // Clear first operand sign bit.
4100 if (VT == MVT::f64) {
4101 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4102 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4104 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4105 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4106 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4107 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4109 C = ConstantVector::get(CV);
4110 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4111 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4113 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4115 // Or the value with the sign bit.
4116 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4119 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4120 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4122 SDOperand Op0 = Op.getOperand(0);
4123 SDOperand Op1 = Op.getOperand(1);
4124 SDOperand CC = Op.getOperand(2);
4125 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4126 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4129 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4131 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4132 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4133 DAG.getConstant(X86CC, MVT::i8), Cond);
4136 assert(isFP && "Illegal integer SetCC!");
4138 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4139 switch (SetCCOpcode) {
4140 default: assert(false && "Illegal floating point SetCC!");
4141 case ISD::SETOEQ: { // !PF & ZF
4142 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4143 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4144 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4145 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4146 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4148 case ISD::SETUNE: { // PF | !ZF
4149 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4150 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4151 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4152 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4153 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4159 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4160 bool addTest = true;
4161 SDOperand Cond = Op.getOperand(0);
4164 if (Cond.getOpcode() == ISD::SETCC)
4165 Cond = LowerSETCC(Cond, DAG);
4167 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4168 // setting operand in place of the X86ISD::SETCC.
4169 if (Cond.getOpcode() == X86ISD::SETCC) {
4170 CC = Cond.getOperand(0);
4172 SDOperand Cmp = Cond.getOperand(1);
4173 unsigned Opc = Cmp.getOpcode();
4174 MVT::ValueType VT = Op.getValueType();
4175 bool IllegalFPCMov = false;
4176 if (VT == MVT::f32 && !X86ScalarSSEf32)
4177 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4178 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4179 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4180 else if (VT == MVT::f80)
4181 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4182 if ((Opc == X86ISD::CMP ||
4183 Opc == X86ISD::COMI ||
4184 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4191 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4192 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4195 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4197 SmallVector<SDOperand, 4> Ops;
4198 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4199 // condition is true.
4200 Ops.push_back(Op.getOperand(2));
4201 Ops.push_back(Op.getOperand(1));
4203 Ops.push_back(Cond);
4204 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4207 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4208 bool addTest = true;
4209 SDOperand Chain = Op.getOperand(0);
4210 SDOperand Cond = Op.getOperand(1);
4211 SDOperand Dest = Op.getOperand(2);
4214 if (Cond.getOpcode() == ISD::SETCC)
4215 Cond = LowerSETCC(Cond, DAG);
4217 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4218 // setting operand in place of the X86ISD::SETCC.
4219 if (Cond.getOpcode() == X86ISD::SETCC) {
4220 CC = Cond.getOperand(0);
4222 SDOperand Cmp = Cond.getOperand(1);
4223 unsigned Opc = Cmp.getOpcode();
4224 if (Opc == X86ISD::CMP ||
4225 Opc == X86ISD::COMI ||
4226 Opc == X86ISD::UCOMI) {
4233 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4234 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4236 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4237 Chain, Op.getOperand(2), CC, Cond);
4240 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4241 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4242 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4244 if (Subtarget->is64Bit())
4245 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4246 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4248 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4250 switch (CallingConv) {
4252 assert(0 && "Unsupported calling convention");
4253 case CallingConv::Fast:
4254 if (isTailCall && PerformTailCallOpt)
4255 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4257 return LowerCCCCallTo(Op,DAG, CallingConv);
4258 case CallingConv::C:
4259 case CallingConv::X86_StdCall:
4260 return LowerCCCCallTo(Op, DAG, CallingConv);
4261 case CallingConv::X86_FastCall:
4262 return LowerFastCCCallTo(Op, DAG, CallingConv);
4267 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4268 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4269 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4270 // that the guard pages used by the OS virtual memory manager are allocated in
4271 // correct sequence.
4273 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4274 SelectionDAG &DAG) {
4275 assert(Subtarget->isTargetCygMing() &&
4276 "This should be used only on Cygwin/Mingw targets");
4279 SDOperand Chain = Op.getOperand(0);
4280 SDOperand Size = Op.getOperand(1);
4281 // FIXME: Ensure alignment here
4285 MVT::ValueType IntPtr = getPointerTy();
4286 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4288 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4289 Flag = Chain.getValue(1);
4291 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4292 SDOperand Ops[] = { Chain,
4293 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4294 DAG.getRegister(X86::EAX, IntPtr),
4296 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4297 Flag = Chain.getValue(1);
4299 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4301 std::vector<MVT::ValueType> Tys;
4302 Tys.push_back(SPTy);
4303 Tys.push_back(MVT::Other);
4304 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4305 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4309 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4310 MachineFunction &MF = DAG.getMachineFunction();
4311 const Function* Fn = MF.getFunction();
4312 if (Fn->hasExternalLinkage() &&
4313 Subtarget->isTargetCygMing() &&
4314 Fn->getName() == "main")
4315 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4317 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4318 if (Subtarget->is64Bit())
4319 return LowerX86_64CCCArguments(Op, DAG);
4323 assert(0 && "Unsupported calling convention");
4324 case CallingConv::Fast:
4325 return LowerCCCArguments(Op,DAG, true);
4327 case CallingConv::C:
4328 return LowerCCCArguments(Op, DAG);
4329 case CallingConv::X86_StdCall:
4330 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4331 return LowerCCCArguments(Op, DAG, true);
4332 case CallingConv::X86_FastCall:
4333 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4334 return LowerFastCCArguments(Op, DAG);
4338 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4339 SDOperand InFlag(0, 0);
4340 SDOperand Chain = Op.getOperand(0);
4342 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4343 if (Align == 0) Align = 1;
4345 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4346 // If not DWORD aligned or size is more than the threshold, call memset.
4347 // The libc version is likely to be faster for these cases. It can use the
4348 // address value and run time information about the CPU.
4349 if ((Align & 3) != 0 ||
4350 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4351 MVT::ValueType IntPtr = getPointerTy();
4352 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4353 TargetLowering::ArgListTy Args;
4354 TargetLowering::ArgListEntry Entry;
4355 Entry.Node = Op.getOperand(1);
4356 Entry.Ty = IntPtrTy;
4357 Args.push_back(Entry);
4358 // Extend the unsigned i8 argument to be an int value for the call.
4359 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4360 Entry.Ty = IntPtrTy;
4361 Args.push_back(Entry);
4362 Entry.Node = Op.getOperand(3);
4363 Args.push_back(Entry);
4364 std::pair<SDOperand,SDOperand> CallResult =
4365 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4366 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4367 return CallResult.second;
4372 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4373 unsigned BytesLeft = 0;
4374 bool TwoRepStos = false;
4377 uint64_t Val = ValC->getValue() & 255;
4379 // If the value is a constant, then we can potentially use larger sets.
4380 switch (Align & 3) {
4381 case 2: // WORD aligned
4384 Val = (Val << 8) | Val;
4386 case 0: // DWORD aligned
4389 Val = (Val << 8) | Val;
4390 Val = (Val << 16) | Val;
4391 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4394 Val = (Val << 32) | Val;
4397 default: // Byte aligned
4400 Count = Op.getOperand(3);
4404 if (AVT > MVT::i8) {
4406 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4407 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4408 BytesLeft = I->getValue() % UBytes;
4410 assert(AVT >= MVT::i32 &&
4411 "Do not use rep;stos if not at least DWORD aligned");
4412 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4413 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4418 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4420 InFlag = Chain.getValue(1);
4423 Count = Op.getOperand(3);
4424 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4425 InFlag = Chain.getValue(1);
4428 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4430 InFlag = Chain.getValue(1);
4431 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4432 Op.getOperand(1), InFlag);
4433 InFlag = Chain.getValue(1);
4435 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4436 SmallVector<SDOperand, 8> Ops;
4437 Ops.push_back(Chain);
4438 Ops.push_back(DAG.getValueType(AVT));
4439 Ops.push_back(InFlag);
4440 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4443 InFlag = Chain.getValue(1);
4444 Count = Op.getOperand(3);
4445 MVT::ValueType CVT = Count.getValueType();
4446 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4447 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4448 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4450 InFlag = Chain.getValue(1);
4451 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4453 Ops.push_back(Chain);
4454 Ops.push_back(DAG.getValueType(MVT::i8));
4455 Ops.push_back(InFlag);
4456 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4457 } else if (BytesLeft) {
4458 // Issue stores for the last 1 - 7 bytes.
4460 unsigned Val = ValC->getValue() & 255;
4461 unsigned Offset = I->getValue() - BytesLeft;
4462 SDOperand DstAddr = Op.getOperand(1);
4463 MVT::ValueType AddrVT = DstAddr.getValueType();
4464 if (BytesLeft >= 4) {
4465 Val = (Val << 8) | Val;
4466 Val = (Val << 16) | Val;
4467 Value = DAG.getConstant(Val, MVT::i32);
4468 Chain = DAG.getStore(Chain, Value,
4469 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4470 DAG.getConstant(Offset, AddrVT)),
4475 if (BytesLeft >= 2) {
4476 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4477 Chain = DAG.getStore(Chain, Value,
4478 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4479 DAG.getConstant(Offset, AddrVT)),
4484 if (BytesLeft == 1) {
4485 Value = DAG.getConstant(Val, MVT::i8);
4486 Chain = DAG.getStore(Chain, Value,
4487 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4488 DAG.getConstant(Offset, AddrVT)),
4496 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4501 SelectionDAG &DAG) {
4503 unsigned BytesLeft = 0;
4504 switch (Align & 3) {
4505 case 2: // WORD aligned
4508 case 0: // DWORD aligned
4510 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4513 default: // Byte aligned
4518 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4519 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4520 BytesLeft = Size % UBytes;
4522 SDOperand InFlag(0, 0);
4523 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4525 InFlag = Chain.getValue(1);
4526 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4528 InFlag = Chain.getValue(1);
4529 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4531 InFlag = Chain.getValue(1);
4533 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4534 SmallVector<SDOperand, 8> Ops;
4535 Ops.push_back(Chain);
4536 Ops.push_back(DAG.getValueType(AVT));
4537 Ops.push_back(InFlag);
4538 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4541 // Issue loads and stores for the last 1 - 7 bytes.
4542 unsigned Offset = Size - BytesLeft;
4543 SDOperand DstAddr = Dest;
4544 MVT::ValueType DstVT = DstAddr.getValueType();
4545 SDOperand SrcAddr = Source;
4546 MVT::ValueType SrcVT = SrcAddr.getValueType();
4548 if (BytesLeft >= 4) {
4549 Value = DAG.getLoad(MVT::i32, Chain,
4550 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4551 DAG.getConstant(Offset, SrcVT)),
4553 Chain = Value.getValue(1);
4554 Chain = DAG.getStore(Chain, Value,
4555 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4556 DAG.getConstant(Offset, DstVT)),
4561 if (BytesLeft >= 2) {
4562 Value = DAG.getLoad(MVT::i16, Chain,
4563 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4564 DAG.getConstant(Offset, SrcVT)),
4566 Chain = Value.getValue(1);
4567 Chain = DAG.getStore(Chain, Value,
4568 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4569 DAG.getConstant(Offset, DstVT)),
4575 if (BytesLeft == 1) {
4576 Value = DAG.getLoad(MVT::i8, Chain,
4577 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4578 DAG.getConstant(Offset, SrcVT)),
4580 Chain = Value.getValue(1);
4581 Chain = DAG.getStore(Chain, Value,
4582 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4583 DAG.getConstant(Offset, DstVT)),
4592 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4593 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4594 SDOperand TheOp = Op.getOperand(0);
4595 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4596 if (Subtarget->is64Bit()) {
4598 DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4599 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4600 MVT::i64, Copy1.getValue(2));
4601 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4602 DAG.getConstant(32, MVT::i8));
4604 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4607 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4608 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4611 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4612 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4613 MVT::i32, Copy1.getValue(2));
4614 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4615 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4616 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4619 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4620 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4622 if (!Subtarget->is64Bit()) {
4623 // vastart just stores the address of the VarArgsFrameIndex slot into the
4624 // memory location argument.
4625 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4626 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4631 // gp_offset (0 - 6 * 8)
4632 // fp_offset (48 - 48 + 8 * 16)
4633 // overflow_arg_area (point to parameters coming in memory).
4635 SmallVector<SDOperand, 8> MemOps;
4636 SDOperand FIN = Op.getOperand(1);
4638 SDOperand Store = DAG.getStore(Op.getOperand(0),
4639 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4640 FIN, SV->getValue(), SV->getOffset());
4641 MemOps.push_back(Store);
4644 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4645 DAG.getConstant(4, getPointerTy()));
4646 Store = DAG.getStore(Op.getOperand(0),
4647 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4648 FIN, SV->getValue(), SV->getOffset());
4649 MemOps.push_back(Store);
4651 // Store ptr to overflow_arg_area
4652 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4653 DAG.getConstant(4, getPointerTy()));
4654 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4655 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4657 MemOps.push_back(Store);
4659 // Store ptr to reg_save_area.
4660 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4661 DAG.getConstant(8, getPointerTy()));
4662 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4663 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4665 MemOps.push_back(Store);
4666 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4669 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4670 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4671 SDOperand Chain = Op.getOperand(0);
4672 SDOperand DstPtr = Op.getOperand(1);
4673 SDOperand SrcPtr = Op.getOperand(2);
4674 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4675 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4677 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4678 SrcSV->getValue(), SrcSV->getOffset());
4679 Chain = SrcPtr.getValue(1);
4680 for (unsigned i = 0; i < 3; ++i) {
4681 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4682 SrcSV->getValue(), SrcSV->getOffset());
4683 Chain = Val.getValue(1);
4684 Chain = DAG.getStore(Chain, Val, DstPtr,
4685 DstSV->getValue(), DstSV->getOffset());
4688 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4689 DAG.getConstant(8, getPointerTy()));
4690 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4691 DAG.getConstant(8, getPointerTy()));
4697 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4698 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4700 default: return SDOperand(); // Don't custom lower most intrinsics.
4701 // Comparison intrinsics.
4702 case Intrinsic::x86_sse_comieq_ss:
4703 case Intrinsic::x86_sse_comilt_ss:
4704 case Intrinsic::x86_sse_comile_ss:
4705 case Intrinsic::x86_sse_comigt_ss:
4706 case Intrinsic::x86_sse_comige_ss:
4707 case Intrinsic::x86_sse_comineq_ss:
4708 case Intrinsic::x86_sse_ucomieq_ss:
4709 case Intrinsic::x86_sse_ucomilt_ss:
4710 case Intrinsic::x86_sse_ucomile_ss:
4711 case Intrinsic::x86_sse_ucomigt_ss:
4712 case Intrinsic::x86_sse_ucomige_ss:
4713 case Intrinsic::x86_sse_ucomineq_ss:
4714 case Intrinsic::x86_sse2_comieq_sd:
4715 case Intrinsic::x86_sse2_comilt_sd:
4716 case Intrinsic::x86_sse2_comile_sd:
4717 case Intrinsic::x86_sse2_comigt_sd:
4718 case Intrinsic::x86_sse2_comige_sd:
4719 case Intrinsic::x86_sse2_comineq_sd:
4720 case Intrinsic::x86_sse2_ucomieq_sd:
4721 case Intrinsic::x86_sse2_ucomilt_sd:
4722 case Intrinsic::x86_sse2_ucomile_sd:
4723 case Intrinsic::x86_sse2_ucomigt_sd:
4724 case Intrinsic::x86_sse2_ucomige_sd:
4725 case Intrinsic::x86_sse2_ucomineq_sd: {
4727 ISD::CondCode CC = ISD::SETCC_INVALID;
4730 case Intrinsic::x86_sse_comieq_ss:
4731 case Intrinsic::x86_sse2_comieq_sd:
4735 case Intrinsic::x86_sse_comilt_ss:
4736 case Intrinsic::x86_sse2_comilt_sd:
4740 case Intrinsic::x86_sse_comile_ss:
4741 case Intrinsic::x86_sse2_comile_sd:
4745 case Intrinsic::x86_sse_comigt_ss:
4746 case Intrinsic::x86_sse2_comigt_sd:
4750 case Intrinsic::x86_sse_comige_ss:
4751 case Intrinsic::x86_sse2_comige_sd:
4755 case Intrinsic::x86_sse_comineq_ss:
4756 case Intrinsic::x86_sse2_comineq_sd:
4760 case Intrinsic::x86_sse_ucomieq_ss:
4761 case Intrinsic::x86_sse2_ucomieq_sd:
4762 Opc = X86ISD::UCOMI;
4765 case Intrinsic::x86_sse_ucomilt_ss:
4766 case Intrinsic::x86_sse2_ucomilt_sd:
4767 Opc = X86ISD::UCOMI;
4770 case Intrinsic::x86_sse_ucomile_ss:
4771 case Intrinsic::x86_sse2_ucomile_sd:
4772 Opc = X86ISD::UCOMI;
4775 case Intrinsic::x86_sse_ucomigt_ss:
4776 case Intrinsic::x86_sse2_ucomigt_sd:
4777 Opc = X86ISD::UCOMI;
4780 case Intrinsic::x86_sse_ucomige_ss:
4781 case Intrinsic::x86_sse2_ucomige_sd:
4782 Opc = X86ISD::UCOMI;
4785 case Intrinsic::x86_sse_ucomineq_ss:
4786 case Intrinsic::x86_sse2_ucomineq_sd:
4787 Opc = X86ISD::UCOMI;
4793 SDOperand LHS = Op.getOperand(1);
4794 SDOperand RHS = Op.getOperand(2);
4795 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4797 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4798 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4799 DAG.getConstant(X86CC, MVT::i8), Cond);
4800 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4805 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4806 // Depths > 0 not supported yet!
4807 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4810 // Just load the return address
4811 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4812 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4815 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4816 // Depths > 0 not supported yet!
4817 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4820 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4821 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4822 DAG.getConstant(4, getPointerTy()));
4825 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4826 SelectionDAG &DAG) {
4827 // Is not yet supported on x86-64
4828 if (Subtarget->is64Bit())
4831 return DAG.getConstant(8, getPointerTy());
4834 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4836 assert(!Subtarget->is64Bit() &&
4837 "Lowering of eh_return builtin is not supported yet on x86-64");
4839 MachineFunction &MF = DAG.getMachineFunction();
4840 SDOperand Chain = Op.getOperand(0);
4841 SDOperand Offset = Op.getOperand(1);
4842 SDOperand Handler = Op.getOperand(2);
4844 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4847 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4848 DAG.getConstant(-4UL, getPointerTy()));
4849 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4850 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4851 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4852 MF.addLiveOut(X86::ECX);
4854 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4855 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4858 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4859 SelectionDAG &DAG) {
4860 SDOperand Root = Op.getOperand(0);
4861 SDOperand Trmp = Op.getOperand(1); // trampoline
4862 SDOperand FPtr = Op.getOperand(2); // nested function
4863 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4865 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4867 if (Subtarget->is64Bit()) {
4868 return SDOperand(); // not yet supported
4870 Function *Func = (Function *)
4871 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4872 unsigned CC = Func->getCallingConv();
4877 assert(0 && "Unsupported calling convention");
4878 case CallingConv::C:
4879 case CallingConv::X86_StdCall: {
4880 // Pass 'nest' parameter in ECX.
4881 // Must be kept in sync with X86CallingConv.td
4884 // Check that ECX wasn't needed by an 'inreg' parameter.
4885 const FunctionType *FTy = Func->getFunctionType();
4886 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4888 if (Attrs && !Func->isVarArg()) {
4889 unsigned InRegCount = 0;
4892 for (FunctionType::param_iterator I = FTy->param_begin(),
4893 E = FTy->param_end(); I != E; ++I, ++Idx)
4894 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4895 // FIXME: should only count parameters that are lowered to integers.
4896 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4898 if (InRegCount > 2) {
4899 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4905 case CallingConv::X86_FastCall:
4906 // Pass 'nest' parameter in EAX.
4907 // Must be kept in sync with X86CallingConv.td
4912 const X86InstrInfo *TII =
4913 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4915 SDOperand OutChains[4];
4916 SDOperand Addr, Disp;
4918 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4919 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4921 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
4922 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
4923 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
4924 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4926 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4927 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4928 TrmpSV->getOffset() + 1, false, 1);
4930 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
4931 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4932 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4933 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4935 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4936 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4937 TrmpSV->getOffset() + 6, false, 1);
4940 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
4941 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
4945 /// LowerOperation - Provide custom lowering hooks for some operations.
4947 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4948 switch (Op.getOpcode()) {
4949 default: assert(0 && "Should not custom lower this!");
4950 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4951 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4952 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4953 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4954 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4955 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4956 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4957 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4958 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4959 case ISD::SHL_PARTS:
4960 case ISD::SRA_PARTS:
4961 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4962 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4963 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4964 case ISD::FABS: return LowerFABS(Op, DAG);
4965 case ISD::FNEG: return LowerFNEG(Op, DAG);
4966 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4967 case ISD::SETCC: return LowerSETCC(Op, DAG);
4968 case ISD::SELECT: return LowerSELECT(Op, DAG);
4969 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4970 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4971 case ISD::CALL: return LowerCALL(Op, DAG);
4972 case ISD::RET: return LowerRET(Op, DAG);
4973 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4974 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4975 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4976 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4977 case ISD::VASTART: return LowerVASTART(Op, DAG);
4978 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4979 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4980 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4981 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4982 case ISD::FRAME_TO_ARGS_OFFSET:
4983 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
4984 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4985 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
4986 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4991 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4993 default: return NULL;
4994 case X86ISD::SHLD: return "X86ISD::SHLD";
4995 case X86ISD::SHRD: return "X86ISD::SHRD";
4996 case X86ISD::FAND: return "X86ISD::FAND";
4997 case X86ISD::FOR: return "X86ISD::FOR";
4998 case X86ISD::FXOR: return "X86ISD::FXOR";
4999 case X86ISD::FSRL: return "X86ISD::FSRL";
5000 case X86ISD::FILD: return "X86ISD::FILD";
5001 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5002 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5003 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5004 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5005 case X86ISD::FLD: return "X86ISD::FLD";
5006 case X86ISD::FST: return "X86ISD::FST";
5007 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5008 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5009 case X86ISD::CALL: return "X86ISD::CALL";
5010 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5011 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5012 case X86ISD::CMP: return "X86ISD::CMP";
5013 case X86ISD::COMI: return "X86ISD::COMI";
5014 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5015 case X86ISD::SETCC: return "X86ISD::SETCC";
5016 case X86ISD::CMOV: return "X86ISD::CMOV";
5017 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5018 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5019 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5020 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5021 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5022 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5023 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5024 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5025 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5026 case X86ISD::FMAX: return "X86ISD::FMAX";
5027 case X86ISD::FMIN: return "X86ISD::FMIN";
5028 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5029 case X86ISD::FRCP: return "X86ISD::FRCP";
5030 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5031 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5032 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5033 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5037 // isLegalAddressingMode - Return true if the addressing mode represented
5038 // by AM is legal for this target, for a load/store of the specified type.
5039 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5040 const Type *Ty) const {
5041 // X86 supports extremely general addressing modes.
5043 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5044 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5048 // We can only fold this if we don't need an extra load.
5049 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5052 // X86-64 only supports addr of globals in small code model.
5053 if (Subtarget->is64Bit()) {
5054 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5056 // If lower 4G is not available, then we must use rip-relative addressing.
5057 if (AM.BaseOffs || AM.Scale > 1)
5068 // These scales always work.
5073 // These scales are formed with basereg+scalereg. Only accept if there is
5078 default: // Other stuff never works.
5086 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5087 if (!Ty1->isInteger() || !Ty2->isInteger())
5089 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5090 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5091 if (NumBits1 <= NumBits2)
5093 return Subtarget->is64Bit() || NumBits1 < 64;
5096 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5097 MVT::ValueType VT2) const {
5098 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5100 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5101 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5102 if (NumBits1 <= NumBits2)
5104 return Subtarget->is64Bit() || NumBits1 < 64;
5107 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5108 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5109 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5110 /// are assumed to be legal.
5112 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5113 // Only do shuffles on 128-bit vector types for now.
5114 if (MVT::getSizeInBits(VT) == 64) return false;
5115 return (Mask.Val->getNumOperands() <= 4 ||
5116 isIdentityMask(Mask.Val) ||
5117 isIdentityMask(Mask.Val, true) ||
5118 isSplatMask(Mask.Val) ||
5119 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5120 X86::isUNPCKLMask(Mask.Val) ||
5121 X86::isUNPCKHMask(Mask.Val) ||
5122 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5123 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5126 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5128 SelectionDAG &DAG) const {
5129 unsigned NumElts = BVOps.size();
5130 // Only do shuffles on 128-bit vector types for now.
5131 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5132 if (NumElts == 2) return true;
5134 return (isMOVLMask(&BVOps[0], 4) ||
5135 isCommutedMOVL(&BVOps[0], 4, true) ||
5136 isSHUFPMask(&BVOps[0], 4) ||
5137 isCommutedSHUFP(&BVOps[0], 4));
5142 //===----------------------------------------------------------------------===//
5143 // X86 Scheduler Hooks
5144 //===----------------------------------------------------------------------===//
5147 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5148 MachineBasicBlock *BB) {
5149 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5150 switch (MI->getOpcode()) {
5151 default: assert(false && "Unexpected instr type to insert");
5152 case X86::CMOV_FR32:
5153 case X86::CMOV_FR64:
5154 case X86::CMOV_V4F32:
5155 case X86::CMOV_V2F64:
5156 case X86::CMOV_V2I64: {
5157 // To "insert" a SELECT_CC instruction, we actually have to insert the
5158 // diamond control-flow pattern. The incoming instruction knows the
5159 // destination vreg to set, the condition code register to branch on, the
5160 // true/false values to select between, and a branch opcode to use.
5161 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5162 ilist<MachineBasicBlock>::iterator It = BB;
5168 // cmpTY ccX, r1, r2
5170 // fallthrough --> copy0MBB
5171 MachineBasicBlock *thisMBB = BB;
5172 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5173 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5175 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5176 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5177 MachineFunction *F = BB->getParent();
5178 F->getBasicBlockList().insert(It, copy0MBB);
5179 F->getBasicBlockList().insert(It, sinkMBB);
5180 // Update machine-CFG edges by first adding all successors of the current
5181 // block to the new block which will contain the Phi node for the select.
5182 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5183 e = BB->succ_end(); i != e; ++i)
5184 sinkMBB->addSuccessor(*i);
5185 // Next, remove all successors of the current block, and add the true
5186 // and fallthrough blocks as its successors.
5187 while(!BB->succ_empty())
5188 BB->removeSuccessor(BB->succ_begin());
5189 BB->addSuccessor(copy0MBB);
5190 BB->addSuccessor(sinkMBB);
5193 // %FalseValue = ...
5194 // # fallthrough to sinkMBB
5197 // Update machine-CFG edges
5198 BB->addSuccessor(sinkMBB);
5201 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5204 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5205 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5206 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5208 delete MI; // The pseudo instruction is gone now.
5212 case X86::FP32_TO_INT16_IN_MEM:
5213 case X86::FP32_TO_INT32_IN_MEM:
5214 case X86::FP32_TO_INT64_IN_MEM:
5215 case X86::FP64_TO_INT16_IN_MEM:
5216 case X86::FP64_TO_INT32_IN_MEM:
5217 case X86::FP64_TO_INT64_IN_MEM:
5218 case X86::FP80_TO_INT16_IN_MEM:
5219 case X86::FP80_TO_INT32_IN_MEM:
5220 case X86::FP80_TO_INT64_IN_MEM: {
5221 // Change the floating point control register to use "round towards zero"
5222 // mode when truncating to an integer value.
5223 MachineFunction *F = BB->getParent();
5224 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5225 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5227 // Load the old value of the high byte of the control word...
5229 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5230 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5232 // Set the high part to be round to zero...
5233 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5236 // Reload the modified control word now...
5237 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5239 // Restore the memory image of control word to original value
5240 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5243 // Get the X86 opcode to use.
5245 switch (MI->getOpcode()) {
5246 default: assert(0 && "illegal opcode!");
5247 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5248 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5249 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5250 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5251 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5252 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5253 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5254 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5255 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5259 MachineOperand &Op = MI->getOperand(0);
5260 if (Op.isRegister()) {
5261 AM.BaseType = X86AddressMode::RegBase;
5262 AM.Base.Reg = Op.getReg();
5264 AM.BaseType = X86AddressMode::FrameIndexBase;
5265 AM.Base.FrameIndex = Op.getFrameIndex();
5267 Op = MI->getOperand(1);
5268 if (Op.isImmediate())
5269 AM.Scale = Op.getImm();
5270 Op = MI->getOperand(2);
5271 if (Op.isImmediate())
5272 AM.IndexReg = Op.getImm();
5273 Op = MI->getOperand(3);
5274 if (Op.isGlobalAddress()) {
5275 AM.GV = Op.getGlobal();
5277 AM.Disp = Op.getImm();
5279 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5280 .addReg(MI->getOperand(4).getReg());
5282 // Reload the original control word now.
5283 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5285 delete MI; // The pseudo instruction is gone now.
5291 //===----------------------------------------------------------------------===//
5292 // X86 Optimization Hooks
5293 //===----------------------------------------------------------------------===//
5295 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5297 uint64_t &KnownZero,
5299 const SelectionDAG &DAG,
5300 unsigned Depth) const {
5301 unsigned Opc = Op.getOpcode();
5302 assert((Opc >= ISD::BUILTIN_OP_END ||
5303 Opc == ISD::INTRINSIC_WO_CHAIN ||
5304 Opc == ISD::INTRINSIC_W_CHAIN ||
5305 Opc == ISD::INTRINSIC_VOID) &&
5306 "Should use MaskedValueIsZero if you don't know whether Op"
5307 " is a target node!");
5309 KnownZero = KnownOne = 0; // Don't know anything.
5313 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5318 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5319 /// element of the result of the vector shuffle.
5320 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5321 MVT::ValueType VT = N->getValueType(0);
5322 SDOperand PermMask = N->getOperand(2);
5323 unsigned NumElems = PermMask.getNumOperands();
5324 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5326 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5328 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5329 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5330 SDOperand Idx = PermMask.getOperand(i);
5331 if (Idx.getOpcode() == ISD::UNDEF)
5332 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5333 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5338 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5339 /// node is a GlobalAddress + an offset.
5340 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5341 unsigned Opc = N->getOpcode();
5342 if (Opc == X86ISD::Wrapper) {
5343 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5344 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5347 } else if (Opc == ISD::ADD) {
5348 SDOperand N1 = N->getOperand(0);
5349 SDOperand N2 = N->getOperand(1);
5350 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5351 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5353 Offset += V->getSignExtended();
5356 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5357 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5359 Offset += V->getSignExtended();
5367 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5369 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5370 MachineFrameInfo *MFI) {
5371 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5374 SDOperand Loc = N->getOperand(1);
5375 SDOperand BaseLoc = Base->getOperand(1);
5376 if (Loc.getOpcode() == ISD::FrameIndex) {
5377 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5379 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5380 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5381 int FS = MFI->getObjectSize(FI);
5382 int BFS = MFI->getObjectSize(BFI);
5383 if (FS != BFS || FS != Size) return false;
5384 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5386 GlobalValue *GV1 = NULL;
5387 GlobalValue *GV2 = NULL;
5388 int64_t Offset1 = 0;
5389 int64_t Offset2 = 0;
5390 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5391 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5392 if (isGA1 && isGA2 && GV1 == GV2)
5393 return Offset1 == (Offset2 + Dist*Size);
5399 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5400 const X86Subtarget *Subtarget) {
5403 if (isGAPlusOffset(Base, GV, Offset))
5404 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5406 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5407 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5409 // Fixed objects do not specify alignment, however the offsets are known.
5410 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5411 (MFI->getObjectOffset(BFI) % 16) == 0);
5413 return MFI->getObjectAlignment(BFI) >= 16;
5419 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5420 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5421 /// if the load addresses are consecutive, non-overlapping, and in the right
5423 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5424 const X86Subtarget *Subtarget) {
5425 MachineFunction &MF = DAG.getMachineFunction();
5426 MachineFrameInfo *MFI = MF.getFrameInfo();
5427 MVT::ValueType VT = N->getValueType(0);
5428 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5429 SDOperand PermMask = N->getOperand(2);
5430 int NumElems = (int)PermMask.getNumOperands();
5431 SDNode *Base = NULL;
5432 for (int i = 0; i < NumElems; ++i) {
5433 SDOperand Idx = PermMask.getOperand(i);
5434 if (Idx.getOpcode() == ISD::UNDEF) {
5435 if (!Base) return SDOperand();
5438 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5439 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5443 else if (!isConsecutiveLoad(Arg.Val, Base,
5444 i, MVT::getSizeInBits(EVT)/8,MFI))
5449 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5450 LoadSDNode *LD = cast<LoadSDNode>(Base);
5452 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5453 LD->getSrcValueOffset(), LD->isVolatile());
5455 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5456 LD->getSrcValueOffset(), LD->isVolatile(),
5457 LD->getAlignment());
5461 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5462 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5463 const X86Subtarget *Subtarget) {
5464 SDOperand Cond = N->getOperand(0);
5466 // If we have SSE[12] support, try to form min/max nodes.
5467 if (Subtarget->hasSSE2() &&
5468 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5469 if (Cond.getOpcode() == ISD::SETCC) {
5470 // Get the LHS/RHS of the select.
5471 SDOperand LHS = N->getOperand(1);
5472 SDOperand RHS = N->getOperand(2);
5473 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5475 unsigned Opcode = 0;
5476 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5479 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5482 if (!UnsafeFPMath) break;
5484 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5486 Opcode = X86ISD::FMIN;
5489 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5492 if (!UnsafeFPMath) break;
5494 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5496 Opcode = X86ISD::FMAX;
5499 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5502 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5505 if (!UnsafeFPMath) break;
5507 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5509 Opcode = X86ISD::FMIN;
5512 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5515 if (!UnsafeFPMath) break;
5517 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5519 Opcode = X86ISD::FMAX;
5525 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5534 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5535 DAGCombinerInfo &DCI) const {
5536 SelectionDAG &DAG = DCI.DAG;
5537 switch (N->getOpcode()) {
5539 case ISD::VECTOR_SHUFFLE:
5540 return PerformShuffleCombine(N, DAG, Subtarget);
5542 return PerformSELECTCombine(N, DAG, Subtarget);
5548 //===----------------------------------------------------------------------===//
5549 // X86 Inline Assembly Support
5550 //===----------------------------------------------------------------------===//
5552 /// getConstraintType - Given a constraint letter, return the type of
5553 /// constraint it is for this target.
5554 X86TargetLowering::ConstraintType
5555 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5556 if (Constraint.size() == 1) {
5557 switch (Constraint[0]) {
5566 return C_RegisterClass;
5571 return TargetLowering::getConstraintType(Constraint);
5574 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5575 /// vector. If it is invalid, don't add anything to Ops.
5576 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5578 std::vector<SDOperand>&Ops,
5579 SelectionDAG &DAG) {
5580 SDOperand Result(0, 0);
5582 switch (Constraint) {
5585 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5586 if (C->getValue() <= 31) {
5587 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5593 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5594 if (C->getValue() <= 255) {
5595 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5601 // Literal immediates are always ok.
5602 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5603 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5607 // If we are in non-pic codegen mode, we allow the address of a global (with
5608 // an optional displacement) to be used with 'i'.
5609 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5612 // Match either (GA) or (GA+C)
5614 Offset = GA->getOffset();
5615 } else if (Op.getOpcode() == ISD::ADD) {
5616 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5617 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5619 Offset = GA->getOffset()+C->getValue();
5621 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5622 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5624 Offset = GA->getOffset()+C->getValue();
5631 // If addressing this global requires a load (e.g. in PIC mode), we can't
5633 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5637 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5643 // Otherwise, not valid for this mode.
5649 Ops.push_back(Result);
5652 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5655 std::vector<unsigned> X86TargetLowering::
5656 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5657 MVT::ValueType VT) const {
5658 if (Constraint.size() == 1) {
5659 // FIXME: not handling fp-stack yet!
5660 switch (Constraint[0]) { // GCC X86 Constraint Letters
5661 default: break; // Unknown constraint letter
5662 case 'A': // EAX/EDX
5663 if (VT == MVT::i32 || VT == MVT::i64)
5664 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5666 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5669 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5670 else if (VT == MVT::i16)
5671 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5672 else if (VT == MVT::i8)
5673 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5674 else if (VT == MVT::i64)
5675 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
5680 return std::vector<unsigned>();
5683 std::pair<unsigned, const TargetRegisterClass*>
5684 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5685 MVT::ValueType VT) const {
5686 // First, see if this is a constraint that directly corresponds to an LLVM
5688 if (Constraint.size() == 1) {
5689 // GCC Constraint Letters
5690 switch (Constraint[0]) {
5692 case 'r': // GENERAL_REGS
5693 case 'R': // LEGACY_REGS
5694 case 'l': // INDEX_REGS
5695 if (VT == MVT::i64 && Subtarget->is64Bit())
5696 return std::make_pair(0U, X86::GR64RegisterClass);
5698 return std::make_pair(0U, X86::GR32RegisterClass);
5699 else if (VT == MVT::i16)
5700 return std::make_pair(0U, X86::GR16RegisterClass);
5701 else if (VT == MVT::i8)
5702 return std::make_pair(0U, X86::GR8RegisterClass);
5704 case 'y': // MMX_REGS if MMX allowed.
5705 if (!Subtarget->hasMMX()) break;
5706 return std::make_pair(0U, X86::VR64RegisterClass);
5708 case 'Y': // SSE_REGS if SSE2 allowed
5709 if (!Subtarget->hasSSE2()) break;
5711 case 'x': // SSE_REGS if SSE1 allowed
5712 if (!Subtarget->hasSSE1()) break;
5716 // Scalar SSE types.
5719 return std::make_pair(0U, X86::FR32RegisterClass);
5722 return std::make_pair(0U, X86::FR64RegisterClass);
5730 return std::make_pair(0U, X86::VR128RegisterClass);
5736 // Use the default implementation in TargetLowering to convert the register
5737 // constraint into a member of a register class.
5738 std::pair<unsigned, const TargetRegisterClass*> Res;
5739 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5741 // Not found as a standard register?
5742 if (Res.second == 0) {
5743 // GCC calls "st(0)" just plain "st".
5744 if (StringsEqualNoCase("{st}", Constraint)) {
5745 Res.first = X86::ST0;
5746 Res.second = X86::RFP80RegisterClass;
5752 // Otherwise, check to see if this is a register class of the wrong value
5753 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5754 // turn into {ax},{dx}.
5755 if (Res.second->hasType(VT))
5756 return Res; // Correct type already, nothing to do.
5758 // All of the single-register GCC register classes map their values onto
5759 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5760 // really want an 8-bit or 32-bit register, map to the appropriate register
5761 // class and return the appropriate register.
5762 if (Res.second != X86::GR16RegisterClass)
5765 if (VT == MVT::i8) {
5766 unsigned DestReg = 0;
5767 switch (Res.first) {
5769 case X86::AX: DestReg = X86::AL; break;
5770 case X86::DX: DestReg = X86::DL; break;
5771 case X86::CX: DestReg = X86::CL; break;
5772 case X86::BX: DestReg = X86::BL; break;
5775 Res.first = DestReg;
5776 Res.second = Res.second = X86::GR8RegisterClass;
5778 } else if (VT == MVT::i32) {
5779 unsigned DestReg = 0;
5780 switch (Res.first) {
5782 case X86::AX: DestReg = X86::EAX; break;
5783 case X86::DX: DestReg = X86::EDX; break;
5784 case X86::CX: DestReg = X86::ECX; break;
5785 case X86::BX: DestReg = X86::EBX; break;
5786 case X86::SI: DestReg = X86::ESI; break;
5787 case X86::DI: DestReg = X86::EDI; break;
5788 case X86::BP: DestReg = X86::EBP; break;
5789 case X86::SP: DestReg = X86::ESP; break;
5792 Res.first = DestReg;
5793 Res.second = Res.second = X86::GR32RegisterClass;
5795 } else if (VT == MVT::i64) {
5796 unsigned DestReg = 0;
5797 switch (Res.first) {
5799 case X86::AX: DestReg = X86::RAX; break;
5800 case X86::DX: DestReg = X86::RDX; break;
5801 case X86::CX: DestReg = X86::RCX; break;
5802 case X86::BX: DestReg = X86::RBX; break;
5803 case X86::SI: DestReg = X86::RSI; break;
5804 case X86::DI: DestReg = X86::RDI; break;
5805 case X86::BP: DestReg = X86::RBP; break;
5806 case X86::SP: DestReg = X86::RSP; break;
5809 Res.first = DestReg;
5810 Res.second = Res.second = X86::GR64RegisterClass;