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/Target/TargetOptions.h"
36 #include "llvm/ADT/StringExtras.h"
37 #include "llvm/ParameterAttributes.h"
40 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
41 : TargetLowering(TM) {
42 Subtarget = &TM.getSubtarget<X86Subtarget>();
43 X86ScalarSSE = Subtarget->hasSSE2();
44 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
46 RegInfo = TM.getRegisterInfo();
48 // Set up the TargetLowering object.
50 // X86 is weird, it always uses i8 for shift amounts and setcc results.
51 setShiftAmountType(MVT::i8);
52 setSetCCResultType(MVT::i8);
53 setSetCCResultContents(ZeroOrOneSetCCResult);
54 setSchedulingPreference(SchedulingForRegPressure);
55 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
56 setStackPointerRegisterToSaveRestore(X86StackPtr);
58 if (Subtarget->isTargetDarwin()) {
59 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
60 setUseUnderscoreSetJmp(false);
61 setUseUnderscoreLongJmp(false);
62 } else if (Subtarget->isTargetMingw()) {
63 // MS runtime is weird: it exports _setjmp, but longjmp!
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(false);
67 setUseUnderscoreSetJmp(true);
68 setUseUnderscoreLongJmp(true);
71 // Set up the register classes.
72 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
73 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
74 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
75 if (Subtarget->is64Bit())
76 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
78 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
80 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
82 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
83 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
84 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
86 if (Subtarget->is64Bit()) {
87 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
88 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
91 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
92 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
97 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
99 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
100 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
101 // SSE has no i16 to fp conversion, only i32
103 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
105 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
106 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
109 if (!Subtarget->is64Bit()) {
110 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
111 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
112 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
115 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
117 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
118 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
121 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
123 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
124 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
127 // Handle FP_TO_UINT by promoting the destination to a larger signed
129 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
130 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
131 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
133 if (Subtarget->is64Bit()) {
134 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
135 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
137 if (X86ScalarSSE && !Subtarget->hasSSE3())
138 // Expand FP_TO_UINT into a select.
139 // FIXME: We would like to use a Custom expander here eventually to do
140 // the optimal thing for SSE vs. the default expansion in the legalizer.
141 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
143 // With SSE3 we can use fisttpll to convert to a signed i64.
144 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
147 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
149 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
150 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
153 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
154 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
155 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
156 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
157 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
158 if (Subtarget->is64Bit())
159 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
160 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
162 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
163 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
164 setOperationAction(ISD::FREM , MVT::f64 , Expand);
166 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
167 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
168 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
169 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
170 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
171 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
172 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
173 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
174 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
175 if (Subtarget->is64Bit()) {
176 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
177 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
178 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
181 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
182 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
184 // These should be promoted to a larger select which is supported.
185 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
186 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
187 // X86 wants to expand cmov itself.
188 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
189 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
190 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
191 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
192 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
193 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
194 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
195 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
196 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
197 if (Subtarget->is64Bit()) {
198 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
199 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
201 // X86 ret instruction may pop stack.
202 setOperationAction(ISD::RET , MVT::Other, Custom);
203 if (!Subtarget->is64Bit())
204 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
207 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
208 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
209 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
210 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
211 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
212 if (Subtarget->is64Bit()) {
213 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
214 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
215 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
216 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
218 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
219 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
220 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
221 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
222 // X86 wants to expand memset / memcpy itself.
223 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
224 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
226 // We don't have line number support yet.
227 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
228 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
229 // FIXME - use subtarget debug flags
230 if (!Subtarget->isTargetDarwin() &&
231 !Subtarget->isTargetELF() &&
232 !Subtarget->isTargetCygMing())
233 setOperationAction(ISD::LABEL, MVT::Other, Expand);
235 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
236 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
237 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
238 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
239 if (Subtarget->is64Bit()) {
241 setExceptionPointerRegister(X86::RAX);
242 setExceptionSelectorRegister(X86::RDX);
244 setExceptionPointerRegister(X86::EAX);
245 setExceptionSelectorRegister(X86::EDX);
248 setOperationAction(ISD::ADJUST_TRAMP, MVT::i32, Expand);
249 setOperationAction(ISD::ADJUST_TRAMP, MVT::i64, Expand);
250 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
252 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
253 setOperationAction(ISD::VASTART , MVT::Other, Custom);
254 setOperationAction(ISD::VAARG , MVT::Other, Expand);
255 setOperationAction(ISD::VAEND , MVT::Other, Expand);
256 if (Subtarget->is64Bit())
257 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
259 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
261 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
262 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
263 if (Subtarget->is64Bit())
264 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
265 if (Subtarget->isTargetCygMing())
266 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
268 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
271 // Set up the FP register classes.
272 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
273 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
275 // Use ANDPD to simulate FABS.
276 setOperationAction(ISD::FABS , MVT::f64, Custom);
277 setOperationAction(ISD::FABS , MVT::f32, Custom);
279 // Use XORP to simulate FNEG.
280 setOperationAction(ISD::FNEG , MVT::f64, Custom);
281 setOperationAction(ISD::FNEG , MVT::f32, Custom);
283 // Use ANDPD and ORPD to simulate FCOPYSIGN.
284 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
287 // We don't support sin/cos/fmod
288 setOperationAction(ISD::FSIN , MVT::f64, Expand);
289 setOperationAction(ISD::FCOS , MVT::f64, Expand);
290 setOperationAction(ISD::FREM , MVT::f64, Expand);
291 setOperationAction(ISD::FSIN , MVT::f32, Expand);
292 setOperationAction(ISD::FCOS , MVT::f32, Expand);
293 setOperationAction(ISD::FREM , MVT::f32, Expand);
295 // Expand FP immediates into loads from the stack, except for the special
297 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
298 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
299 addLegalFPImmediate(APFloat(+0.0)); // xorps / xorpd
301 // Conversions to long double (in X87) go through memory.
302 setConvertAction(MVT::f32, MVT::f80, Expand);
303 setConvertAction(MVT::f64, MVT::f80, Expand);
305 // Conversions from long double (in X87) go through memory.
306 setConvertAction(MVT::f80, MVT::f32, Expand);
307 setConvertAction(MVT::f80, MVT::f64, Expand);
309 // Set up the FP register classes.
310 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
311 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
313 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
314 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
315 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
316 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
318 // Floating truncations need to go through memory.
319 setConvertAction(MVT::f80, MVT::f32, Expand);
320 setConvertAction(MVT::f64, MVT::f32, Expand);
321 setConvertAction(MVT::f80, MVT::f64, Expand);
324 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
325 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
328 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
329 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
330 addLegalFPImmediate(APFloat(+0.0)); // FLD0
331 addLegalFPImmediate(APFloat(+1.0)); // FLD1
332 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
333 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
336 // Long double always uses X87.
337 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
339 // First set operation action for all vector types to expand. Then we
340 // will selectively turn on ones that can be effectively codegen'd.
341 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
342 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
343 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
344 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
345 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
346 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
347 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
348 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
349 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
350 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
351 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
352 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
353 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
354 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
355 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
356 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
357 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
358 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
359 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
360 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
361 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
362 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
363 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
364 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
365 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
368 if (Subtarget->hasMMX()) {
369 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
370 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
371 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
372 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
374 // FIXME: add MMX packed arithmetics
376 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
377 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
378 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
379 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
381 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
382 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
383 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
385 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
386 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
388 setOperationAction(ISD::AND, MVT::v8i8, Promote);
389 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
390 setOperationAction(ISD::AND, MVT::v4i16, Promote);
391 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
392 setOperationAction(ISD::AND, MVT::v2i32, Promote);
393 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
394 setOperationAction(ISD::AND, MVT::v1i64, Legal);
396 setOperationAction(ISD::OR, MVT::v8i8, Promote);
397 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
398 setOperationAction(ISD::OR, MVT::v4i16, Promote);
399 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
400 setOperationAction(ISD::OR, MVT::v2i32, Promote);
401 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
402 setOperationAction(ISD::OR, MVT::v1i64, Legal);
404 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
405 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
406 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
407 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
408 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
409 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
410 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
412 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
413 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
414 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
415 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
416 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
417 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
418 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
420 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
421 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
422 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
423 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
425 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
426 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
427 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
428 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
430 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
431 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
432 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
433 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
436 if (Subtarget->hasSSE1()) {
437 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
439 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
440 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
441 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
442 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
443 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
444 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
445 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
446 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
447 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
448 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
449 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
452 if (Subtarget->hasSSE2()) {
453 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
454 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
455 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
456 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
457 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
459 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
460 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
461 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
462 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
463 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
464 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
465 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
466 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
467 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
468 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
469 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
470 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
471 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
472 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
473 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
475 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
476 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
477 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
478 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
479 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
480 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
482 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
483 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
484 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
485 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
486 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
488 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
489 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
490 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
491 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
492 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
493 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
495 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
496 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
497 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
498 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
499 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
500 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
501 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
502 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
503 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
504 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
505 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
506 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
509 // Custom lower v2i64 and v2f64 selects.
510 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
511 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
512 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
513 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
516 // We want to custom lower some of our intrinsics.
517 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
519 // We have target-specific dag combine patterns for the following nodes:
520 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
521 setTargetDAGCombine(ISD::SELECT);
523 computeRegisterProperties();
525 // FIXME: These should be based on subtarget info. Plus, the values should
526 // be smaller when we are in optimizing for size mode.
527 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
528 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
529 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
530 allowUnalignedMemoryAccesses = true; // x86 supports it!
534 //===----------------------------------------------------------------------===//
535 // Return Value Calling Convention Implementation
536 //===----------------------------------------------------------------------===//
538 #include "X86GenCallingConv.inc"
540 /// LowerRET - Lower an ISD::RET node.
541 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
542 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
544 SmallVector<CCValAssign, 16> RVLocs;
545 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
546 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
547 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
548 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
551 // If this is the first return lowered for this function, add the regs to the
552 // liveout set for the function.
553 if (DAG.getMachineFunction().liveout_empty()) {
554 for (unsigned i = 0; i != RVLocs.size(); ++i)
555 if (RVLocs[i].isRegLoc())
556 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
559 SDOperand Chain = Op.getOperand(0);
562 // Copy the result values into the output registers.
563 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
564 RVLocs[0].getLocReg() != X86::ST0) {
565 for (unsigned i = 0; i != RVLocs.size(); ++i) {
566 CCValAssign &VA = RVLocs[i];
567 assert(VA.isRegLoc() && "Can only return in registers!");
568 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
570 Flag = Chain.getValue(1);
573 // We need to handle a destination of ST0 specially, because it isn't really
575 SDOperand Value = Op.getOperand(1);
577 // If this is an FP return with ScalarSSE, we need to move the value from
578 // an XMM register onto the fp-stack.
582 // If this is a load into a scalarsse value, don't store the loaded value
583 // back to the stack, only to reload it: just replace the scalar-sse load.
584 if (ISD::isNON_EXTLoad(Value.Val) &&
585 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
586 Chain = Value.getOperand(0);
587 MemLoc = Value.getOperand(1);
589 // Spill the value to memory and reload it into top of stack.
590 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
591 MachineFunction &MF = DAG.getMachineFunction();
592 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
593 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
594 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
596 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
597 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
598 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
599 Chain = Value.getValue(1);
602 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
603 SDOperand Ops[] = { Chain, Value };
604 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
605 Flag = Chain.getValue(1);
608 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
610 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
612 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
616 /// LowerCallResult - Lower the result values of an ISD::CALL into the
617 /// appropriate copies out of appropriate physical registers. This assumes that
618 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
619 /// being lowered. The returns a SDNode with the same number of values as the
621 SDNode *X86TargetLowering::
622 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
623 unsigned CallingConv, SelectionDAG &DAG) {
625 // Assign locations to each value returned by this call.
626 SmallVector<CCValAssign, 16> RVLocs;
627 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
628 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
629 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
632 SmallVector<SDOperand, 8> ResultVals;
634 // Copy all of the result registers out of their specified physreg.
635 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
636 for (unsigned i = 0; i != RVLocs.size(); ++i) {
637 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
638 RVLocs[i].getValVT(), InFlag).getValue(1);
639 InFlag = Chain.getValue(2);
640 ResultVals.push_back(Chain.getValue(0));
643 // Copies from the FP stack are special, as ST0 isn't a valid register
644 // before the fp stackifier runs.
646 // Copy ST0 into an RFP register with FP_GET_RESULT.
647 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
648 SDOperand GROps[] = { Chain, InFlag };
649 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
650 Chain = RetVal.getValue(1);
651 InFlag = RetVal.getValue(2);
653 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
656 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
657 // shouldn't be necessary except that RFP cannot be live across
658 // multiple blocks. When stackifier is fixed, they can be uncoupled.
659 MachineFunction &MF = DAG.getMachineFunction();
660 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
661 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
663 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
665 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
666 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
667 Chain = RetVal.getValue(1);
669 ResultVals.push_back(RetVal);
672 // Merge everything together with a MERGE_VALUES node.
673 ResultVals.push_back(Chain);
674 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
675 &ResultVals[0], ResultVals.size()).Val;
679 //===----------------------------------------------------------------------===//
680 // C & StdCall Calling Convention implementation
681 //===----------------------------------------------------------------------===//
682 // StdCall calling convention seems to be standard for many Windows' API
683 // routines and around. It differs from C calling convention just a little:
684 // callee should clean up the stack, not caller. Symbols should be also
685 // decorated in some fancy way :) It doesn't support any vector arguments.
687 /// AddLiveIn - This helper function adds the specified physical register to the
688 /// MachineFunction as a live in value. It also creates a corresponding virtual
690 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
691 const TargetRegisterClass *RC) {
692 assert(RC->contains(PReg) && "Not the correct regclass!");
693 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
694 MF.addLiveIn(PReg, VReg);
698 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
700 unsigned NumArgs = Op.Val->getNumValues() - 1;
701 MachineFunction &MF = DAG.getMachineFunction();
702 MachineFrameInfo *MFI = MF.getFrameInfo();
703 SDOperand Root = Op.getOperand(0);
704 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
706 // Assign locations to all of the incoming arguments.
707 SmallVector<CCValAssign, 16> ArgLocs;
708 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
709 getTargetMachine(), ArgLocs);
710 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
712 SmallVector<SDOperand, 8> ArgValues;
713 unsigned LastVal = ~0U;
714 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
715 CCValAssign &VA = ArgLocs[i];
716 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
718 assert(VA.getValNo() != LastVal &&
719 "Don't support value assigned to multiple locs yet");
720 LastVal = VA.getValNo();
723 MVT::ValueType RegVT = VA.getLocVT();
724 TargetRegisterClass *RC;
725 if (RegVT == MVT::i32)
726 RC = X86::GR32RegisterClass;
728 assert(MVT::isVector(RegVT));
729 RC = X86::VR128RegisterClass;
732 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
733 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
735 // If this is an 8 or 16-bit value, it is really passed promoted to 32
736 // bits. Insert an assert[sz]ext to capture this, then truncate to the
738 if (VA.getLocInfo() == CCValAssign::SExt)
739 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
740 DAG.getValueType(VA.getValVT()));
741 else if (VA.getLocInfo() == CCValAssign::ZExt)
742 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
743 DAG.getValueType(VA.getValVT()));
745 if (VA.getLocInfo() != CCValAssign::Full)
746 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
748 ArgValues.push_back(ArgValue);
750 assert(VA.isMemLoc());
752 // Create the nodes corresponding to a load from this parameter slot.
753 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
754 VA.getLocMemOffset());
755 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
756 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
760 unsigned StackSize = CCInfo.getNextStackOffset();
762 ArgValues.push_back(Root);
764 // If the function takes variable number of arguments, make a frame index for
765 // the start of the first vararg value... for expansion of llvm.va_start.
767 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
769 if (isStdCall && !isVarArg) {
770 BytesToPopOnReturn = StackSize; // Callee pops everything..
771 BytesCallerReserves = 0;
773 BytesToPopOnReturn = 0; // Callee pops nothing.
775 // If this is an sret function, the return should pop the hidden pointer.
777 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
778 ISD::ParamFlags::StructReturn))
779 BytesToPopOnReturn = 4;
781 BytesCallerReserves = StackSize;
784 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
786 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
787 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
789 // Return the new list of results.
790 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
791 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
794 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
796 SDOperand Chain = Op.getOperand(0);
797 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
798 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
799 SDOperand Callee = Op.getOperand(4);
800 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
802 // Analyze operands of the call, assigning locations to each operand.
803 SmallVector<CCValAssign, 16> ArgLocs;
804 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
805 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
807 // Get a count of how many bytes are to be pushed on the stack.
808 unsigned NumBytes = CCInfo.getNextStackOffset();
810 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
812 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
813 SmallVector<SDOperand, 8> MemOpChains;
817 // Walk the register/memloc assignments, inserting copies/loads.
818 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
819 CCValAssign &VA = ArgLocs[i];
820 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
822 // Promote the value if needed.
823 switch (VA.getLocInfo()) {
824 default: assert(0 && "Unknown loc info!");
825 case CCValAssign::Full: break;
826 case CCValAssign::SExt:
827 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
829 case CCValAssign::ZExt:
830 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
832 case CCValAssign::AExt:
833 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
838 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
840 assert(VA.isMemLoc());
841 if (StackPtr.Val == 0)
842 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
843 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
844 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
845 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
849 // If the first argument is an sret pointer, remember it.
850 bool isSRet = NumOps &&
851 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
852 ISD::ParamFlags::StructReturn);
854 if (!MemOpChains.empty())
855 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
856 &MemOpChains[0], MemOpChains.size());
858 // Build a sequence of copy-to-reg nodes chained together with token chain
859 // and flag operands which copy the outgoing args into registers.
861 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
862 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
864 InFlag = Chain.getValue(1);
867 // ELF / PIC requires GOT in the EBX register before function calls via PLT
869 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
870 Subtarget->isPICStyleGOT()) {
871 Chain = DAG.getCopyToReg(Chain, X86::EBX,
872 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
874 InFlag = Chain.getValue(1);
877 // If the callee is a GlobalAddress node (quite common, every direct call is)
878 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
879 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
880 // We should use extra load for direct calls to dllimported functions in
882 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
883 getTargetMachine(), true))
884 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
885 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
886 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
888 // Returns a chain & a flag for retval copy to use.
889 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
890 SmallVector<SDOperand, 8> Ops;
891 Ops.push_back(Chain);
892 Ops.push_back(Callee);
894 // Add argument registers to the end of the list so that they are known live
896 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
897 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
898 RegsToPass[i].second.getValueType()));
900 // Add an implicit use GOT pointer in EBX.
901 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
902 Subtarget->isPICStyleGOT())
903 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
906 Ops.push_back(InFlag);
908 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
909 NodeTys, &Ops[0], Ops.size());
910 InFlag = Chain.getValue(1);
912 // Create the CALLSEQ_END node.
913 unsigned NumBytesForCalleeToPush = 0;
915 if (CC == CallingConv::X86_StdCall) {
917 NumBytesForCalleeToPush = isSRet ? 4 : 0;
919 NumBytesForCalleeToPush = NumBytes;
921 // If this is is a call to a struct-return function, the callee
922 // pops the hidden struct pointer, so we have to push it back.
923 // This is common for Darwin/X86, Linux & Mingw32 targets.
924 NumBytesForCalleeToPush = isSRet ? 4 : 0;
927 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
929 Ops.push_back(Chain);
930 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
931 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
932 Ops.push_back(InFlag);
933 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
934 InFlag = Chain.getValue(1);
936 // Handle result values, copying them out of physregs into vregs that we
938 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
942 //===----------------------------------------------------------------------===//
943 // FastCall Calling Convention implementation
944 //===----------------------------------------------------------------------===//
946 // The X86 'fastcall' calling convention passes up to two integer arguments in
947 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
948 // and requires that the callee pop its arguments off the stack (allowing proper
949 // tail calls), and has the same return value conventions as C calling convs.
951 // This calling convention always arranges for the callee pop value to be 8n+4
952 // bytes, which is needed for tail recursion elimination and stack alignment
955 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
956 MachineFunction &MF = DAG.getMachineFunction();
957 MachineFrameInfo *MFI = MF.getFrameInfo();
958 SDOperand Root = Op.getOperand(0);
959 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
961 // Assign locations to all of the incoming arguments.
962 SmallVector<CCValAssign, 16> ArgLocs;
963 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
964 getTargetMachine(), ArgLocs);
965 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
967 SmallVector<SDOperand, 8> ArgValues;
968 unsigned LastVal = ~0U;
969 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
970 CCValAssign &VA = ArgLocs[i];
971 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
973 assert(VA.getValNo() != LastVal &&
974 "Don't support value assigned to multiple locs yet");
975 LastVal = VA.getValNo();
978 MVT::ValueType RegVT = VA.getLocVT();
979 TargetRegisterClass *RC;
980 if (RegVT == MVT::i32)
981 RC = X86::GR32RegisterClass;
983 assert(MVT::isVector(RegVT));
984 RC = X86::VR128RegisterClass;
987 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
988 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
990 // If this is an 8 or 16-bit value, it is really passed promoted to 32
991 // bits. Insert an assert[sz]ext to capture this, then truncate to the
993 if (VA.getLocInfo() == CCValAssign::SExt)
994 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
995 DAG.getValueType(VA.getValVT()));
996 else if (VA.getLocInfo() == CCValAssign::ZExt)
997 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
998 DAG.getValueType(VA.getValVT()));
1000 if (VA.getLocInfo() != CCValAssign::Full)
1001 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1003 ArgValues.push_back(ArgValue);
1005 assert(VA.isMemLoc());
1007 // Create the nodes corresponding to a load from this parameter slot.
1008 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1009 VA.getLocMemOffset());
1010 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1011 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1015 ArgValues.push_back(Root);
1017 unsigned StackSize = CCInfo.getNextStackOffset();
1019 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1020 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1021 // arguments and the arguments after the retaddr has been pushed are aligned.
1022 if ((StackSize & 7) == 0)
1026 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1027 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1028 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1029 BytesCallerReserves = 0;
1031 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1032 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1034 // Return the new list of results.
1035 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1036 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1040 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1041 const SDOperand &StackPtr,
1042 const CCValAssign &VA,
1045 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1046 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1047 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1048 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1049 if (Flags & ISD::ParamFlags::ByVal) {
1050 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1051 ISD::ParamFlags::ByValAlignOffs);
1053 assert (Align >= 8);
1054 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1055 ISD::ParamFlags::ByValSizeOffs;
1057 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1058 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1060 return DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, PtrOff, Arg, SizeNode,
1063 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1067 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1069 SDOperand Chain = Op.getOperand(0);
1070 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1071 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1072 SDOperand Callee = Op.getOperand(4);
1074 // Analyze operands of the call, assigning locations to each operand.
1075 SmallVector<CCValAssign, 16> ArgLocs;
1076 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1077 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1079 // Get a count of how many bytes are to be pushed on the stack.
1080 unsigned NumBytes = CCInfo.getNextStackOffset();
1082 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1083 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1084 // arguments and the arguments after the retaddr has been pushed are aligned.
1085 if ((NumBytes & 7) == 0)
1089 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1091 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1092 SmallVector<SDOperand, 8> MemOpChains;
1096 // Walk the register/memloc assignments, inserting copies/loads.
1097 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1098 CCValAssign &VA = ArgLocs[i];
1099 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1101 // Promote the value if needed.
1102 switch (VA.getLocInfo()) {
1103 default: assert(0 && "Unknown loc info!");
1104 case CCValAssign::Full: break;
1105 case CCValAssign::SExt:
1106 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1108 case CCValAssign::ZExt:
1109 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1111 case CCValAssign::AExt:
1112 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1116 if (VA.isRegLoc()) {
1117 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1119 assert(VA.isMemLoc());
1120 if (StackPtr.Val == 0)
1121 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1122 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1123 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1124 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1128 if (!MemOpChains.empty())
1129 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1130 &MemOpChains[0], MemOpChains.size());
1132 // Build a sequence of copy-to-reg nodes chained together with token chain
1133 // and flag operands which copy the outgoing args into registers.
1135 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1136 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1138 InFlag = Chain.getValue(1);
1141 // If the callee is a GlobalAddress node (quite common, every direct call is)
1142 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1143 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1144 // We should use extra load for direct calls to dllimported functions in
1146 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1147 getTargetMachine(), true))
1148 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1149 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1150 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1152 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1154 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1155 Subtarget->isPICStyleGOT()) {
1156 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1157 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1159 InFlag = Chain.getValue(1);
1162 // Returns a chain & a flag for retval copy to use.
1163 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1164 SmallVector<SDOperand, 8> Ops;
1165 Ops.push_back(Chain);
1166 Ops.push_back(Callee);
1168 // Add argument registers to the end of the list so that they are known live
1170 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1171 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1172 RegsToPass[i].second.getValueType()));
1174 // Add an implicit use GOT pointer in EBX.
1175 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1176 Subtarget->isPICStyleGOT())
1177 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1180 Ops.push_back(InFlag);
1182 // FIXME: Do not generate X86ISD::TAILCALL for now.
1183 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1184 NodeTys, &Ops[0], Ops.size());
1185 InFlag = Chain.getValue(1);
1187 // Returns a flag for retval copy to use.
1188 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1190 Ops.push_back(Chain);
1191 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1192 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1193 Ops.push_back(InFlag);
1194 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1195 InFlag = Chain.getValue(1);
1197 // Handle result values, copying them out of physregs into vregs that we
1199 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1203 //===----------------------------------------------------------------------===//
1204 // X86-64 C Calling Convention implementation
1205 //===----------------------------------------------------------------------===//
1208 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1209 MachineFunction &MF = DAG.getMachineFunction();
1210 MachineFrameInfo *MFI = MF.getFrameInfo();
1211 SDOperand Root = Op.getOperand(0);
1212 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1214 static const unsigned GPR64ArgRegs[] = {
1215 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1217 static const unsigned XMMArgRegs[] = {
1218 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1219 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1223 // Assign locations to all of the incoming arguments.
1224 SmallVector<CCValAssign, 16> ArgLocs;
1225 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1226 getTargetMachine(), ArgLocs);
1227 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1229 SmallVector<SDOperand, 8> ArgValues;
1230 unsigned LastVal = ~0U;
1231 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1232 CCValAssign &VA = ArgLocs[i];
1233 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1235 assert(VA.getValNo() != LastVal &&
1236 "Don't support value assigned to multiple locs yet");
1237 LastVal = VA.getValNo();
1239 if (VA.isRegLoc()) {
1240 MVT::ValueType RegVT = VA.getLocVT();
1241 TargetRegisterClass *RC;
1242 if (RegVT == MVT::i32)
1243 RC = X86::GR32RegisterClass;
1244 else if (RegVT == MVT::i64)
1245 RC = X86::GR64RegisterClass;
1246 else if (RegVT == MVT::f32)
1247 RC = X86::FR32RegisterClass;
1248 else if (RegVT == MVT::f64)
1249 RC = X86::FR64RegisterClass;
1251 assert(MVT::isVector(RegVT));
1252 if (MVT::getSizeInBits(RegVT) == 64) {
1253 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1256 RC = X86::VR128RegisterClass;
1259 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1260 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1262 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1263 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1265 if (VA.getLocInfo() == CCValAssign::SExt)
1266 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1267 DAG.getValueType(VA.getValVT()));
1268 else if (VA.getLocInfo() == CCValAssign::ZExt)
1269 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1270 DAG.getValueType(VA.getValVT()));
1272 if (VA.getLocInfo() != CCValAssign::Full)
1273 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1275 // Handle MMX values passed in GPRs.
1276 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1277 MVT::getSizeInBits(RegVT) == 64)
1278 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1280 ArgValues.push_back(ArgValue);
1282 assert(VA.isMemLoc());
1284 // Create the nodes corresponding to a load from this parameter slot.
1285 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1286 VA.getLocMemOffset());
1287 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1289 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
1290 if (Flags & ISD::ParamFlags::ByVal)
1291 ArgValues.push_back(FIN);
1293 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1297 unsigned StackSize = CCInfo.getNextStackOffset();
1299 // If the function takes variable number of arguments, make a frame index for
1300 // the start of the first vararg value... for expansion of llvm.va_start.
1302 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1303 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1305 // For X86-64, if there are vararg parameters that are passed via
1306 // registers, then we must store them to their spots on the stack so they
1307 // may be loaded by deferencing the result of va_next.
1308 VarArgsGPOffset = NumIntRegs * 8;
1309 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1310 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1311 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1313 // Store the integer parameter registers.
1314 SmallVector<SDOperand, 8> MemOps;
1315 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1316 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1317 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1318 for (; NumIntRegs != 6; ++NumIntRegs) {
1319 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1320 X86::GR64RegisterClass);
1321 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1322 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1323 MemOps.push_back(Store);
1324 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1325 DAG.getConstant(8, getPointerTy()));
1328 // Now store the XMM (fp + vector) parameter registers.
1329 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1330 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1331 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1332 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1333 X86::VR128RegisterClass);
1334 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1335 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1336 MemOps.push_back(Store);
1337 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1338 DAG.getConstant(16, getPointerTy()));
1340 if (!MemOps.empty())
1341 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1342 &MemOps[0], MemOps.size());
1345 ArgValues.push_back(Root);
1347 BytesToPopOnReturn = 0; // Callee pops nothing.
1348 BytesCallerReserves = StackSize;
1350 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1351 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1353 // Return the new list of results.
1354 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1355 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1359 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1361 SDOperand Chain = Op.getOperand(0);
1362 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1363 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1364 SDOperand Callee = Op.getOperand(4);
1366 // Analyze operands of the call, assigning locations to each operand.
1367 SmallVector<CCValAssign, 16> ArgLocs;
1368 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1369 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1371 // Get a count of how many bytes are to be pushed on the stack.
1372 unsigned NumBytes = CCInfo.getNextStackOffset();
1373 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1375 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1376 SmallVector<SDOperand, 8> MemOpChains;
1380 // Walk the register/memloc assignments, inserting copies/loads.
1381 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1382 CCValAssign &VA = ArgLocs[i];
1383 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1385 // Promote the value if needed.
1386 switch (VA.getLocInfo()) {
1387 default: assert(0 && "Unknown loc info!");
1388 case CCValAssign::Full: break;
1389 case CCValAssign::SExt:
1390 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1392 case CCValAssign::ZExt:
1393 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1395 case CCValAssign::AExt:
1396 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1400 if (VA.isRegLoc()) {
1401 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1403 assert(VA.isMemLoc());
1404 if (StackPtr.Val == 0)
1405 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1407 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1412 if (!MemOpChains.empty())
1413 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1414 &MemOpChains[0], MemOpChains.size());
1416 // Build a sequence of copy-to-reg nodes chained together with token chain
1417 // and flag operands which copy the outgoing args into registers.
1419 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1420 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1422 InFlag = Chain.getValue(1);
1426 // From AMD64 ABI document:
1427 // For calls that may call functions that use varargs or stdargs
1428 // (prototype-less calls or calls to functions containing ellipsis (...) in
1429 // the declaration) %al is used as hidden argument to specify the number
1430 // of SSE registers used. The contents of %al do not need to match exactly
1431 // the number of registers, but must be an ubound on the number of SSE
1432 // registers used and is in the range 0 - 8 inclusive.
1434 // Count the number of XMM registers allocated.
1435 static const unsigned XMMArgRegs[] = {
1436 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1437 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1439 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1441 Chain = DAG.getCopyToReg(Chain, X86::AL,
1442 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1443 InFlag = Chain.getValue(1);
1446 // If the callee is a GlobalAddress node (quite common, every direct call is)
1447 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1448 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1449 // We should use extra load for direct calls to dllimported functions in
1451 if (getTargetMachine().getCodeModel() != CodeModel::Large
1452 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1453 getTargetMachine(), true))
1454 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1455 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1456 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1457 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1459 // Returns a chain & a flag for retval copy to use.
1460 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1461 SmallVector<SDOperand, 8> Ops;
1462 Ops.push_back(Chain);
1463 Ops.push_back(Callee);
1465 // Add argument registers to the end of the list so that they are known live
1467 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1468 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1469 RegsToPass[i].second.getValueType()));
1472 Ops.push_back(InFlag);
1474 // FIXME: Do not generate X86ISD::TAILCALL for now.
1475 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1476 NodeTys, &Ops[0], Ops.size());
1477 InFlag = Chain.getValue(1);
1479 // Returns a flag for retval copy to use.
1480 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1482 Ops.push_back(Chain);
1483 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1484 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1485 Ops.push_back(InFlag);
1486 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1487 InFlag = Chain.getValue(1);
1489 // Handle result values, copying them out of physregs into vregs that we
1491 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1495 //===----------------------------------------------------------------------===//
1496 // Other Lowering Hooks
1497 //===----------------------------------------------------------------------===//
1500 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1501 MachineFunction &MF = DAG.getMachineFunction();
1502 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1503 int ReturnAddrIndex = FuncInfo->getRAIndex();
1505 if (ReturnAddrIndex == 0) {
1506 // Set up a frame object for the return address.
1507 if (Subtarget->is64Bit())
1508 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1510 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1512 FuncInfo->setRAIndex(ReturnAddrIndex);
1515 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1520 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1521 /// specific condition code. It returns a false if it cannot do a direct
1522 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1524 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1525 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1526 SelectionDAG &DAG) {
1527 X86CC = X86::COND_INVALID;
1529 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1530 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1531 // X > -1 -> X == 0, jump !sign.
1532 RHS = DAG.getConstant(0, RHS.getValueType());
1533 X86CC = X86::COND_NS;
1535 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1536 // X < 0 -> X == 0, jump on sign.
1537 X86CC = X86::COND_S;
1542 switch (SetCCOpcode) {
1544 case ISD::SETEQ: X86CC = X86::COND_E; break;
1545 case ISD::SETGT: X86CC = X86::COND_G; break;
1546 case ISD::SETGE: X86CC = X86::COND_GE; break;
1547 case ISD::SETLT: X86CC = X86::COND_L; break;
1548 case ISD::SETLE: X86CC = X86::COND_LE; break;
1549 case ISD::SETNE: X86CC = X86::COND_NE; break;
1550 case ISD::SETULT: X86CC = X86::COND_B; break;
1551 case ISD::SETUGT: X86CC = X86::COND_A; break;
1552 case ISD::SETULE: X86CC = X86::COND_BE; break;
1553 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1556 // On a floating point condition, the flags are set as follows:
1558 // 0 | 0 | 0 | X > Y
1559 // 0 | 0 | 1 | X < Y
1560 // 1 | 0 | 0 | X == Y
1561 // 1 | 1 | 1 | unordered
1563 switch (SetCCOpcode) {
1566 case ISD::SETEQ: X86CC = X86::COND_E; break;
1567 case ISD::SETOLT: Flip = true; // Fallthrough
1569 case ISD::SETGT: X86CC = X86::COND_A; break;
1570 case ISD::SETOLE: Flip = true; // Fallthrough
1572 case ISD::SETGE: X86CC = X86::COND_AE; break;
1573 case ISD::SETUGT: Flip = true; // Fallthrough
1575 case ISD::SETLT: X86CC = X86::COND_B; break;
1576 case ISD::SETUGE: Flip = true; // Fallthrough
1578 case ISD::SETLE: X86CC = X86::COND_BE; break;
1580 case ISD::SETNE: X86CC = X86::COND_NE; break;
1581 case ISD::SETUO: X86CC = X86::COND_P; break;
1582 case ISD::SETO: X86CC = X86::COND_NP; break;
1585 std::swap(LHS, RHS);
1588 return X86CC != X86::COND_INVALID;
1591 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1592 /// code. Current x86 isa includes the following FP cmov instructions:
1593 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1594 static bool hasFPCMov(unsigned X86CC) {
1610 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1611 /// true if Op is undef or if its value falls within the specified range (L, H].
1612 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1613 if (Op.getOpcode() == ISD::UNDEF)
1616 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1617 return (Val >= Low && Val < Hi);
1620 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1621 /// true if Op is undef or if its value equal to the specified value.
1622 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1623 if (Op.getOpcode() == ISD::UNDEF)
1625 return cast<ConstantSDNode>(Op)->getValue() == Val;
1628 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1629 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1630 bool X86::isPSHUFDMask(SDNode *N) {
1631 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1633 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
1636 // Check if the value doesn't reference the second vector.
1637 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1638 SDOperand Arg = N->getOperand(i);
1639 if (Arg.getOpcode() == ISD::UNDEF) continue;
1640 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1641 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
1648 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1649 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1650 bool X86::isPSHUFHWMask(SDNode *N) {
1651 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1653 if (N->getNumOperands() != 8)
1656 // Lower quadword copied in order.
1657 for (unsigned i = 0; i != 4; ++i) {
1658 SDOperand Arg = N->getOperand(i);
1659 if (Arg.getOpcode() == ISD::UNDEF) continue;
1660 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1661 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1665 // Upper quadword shuffled.
1666 for (unsigned i = 4; i != 8; ++i) {
1667 SDOperand Arg = N->getOperand(i);
1668 if (Arg.getOpcode() == ISD::UNDEF) continue;
1669 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1670 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1671 if (Val < 4 || Val > 7)
1678 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1679 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1680 bool X86::isPSHUFLWMask(SDNode *N) {
1681 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1683 if (N->getNumOperands() != 8)
1686 // Upper quadword copied in order.
1687 for (unsigned i = 4; i != 8; ++i)
1688 if (!isUndefOrEqual(N->getOperand(i), i))
1691 // Lower quadword shuffled.
1692 for (unsigned i = 0; i != 4; ++i)
1693 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1699 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1700 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1701 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1702 if (NumElems != 2 && NumElems != 4) return false;
1704 unsigned Half = NumElems / 2;
1705 for (unsigned i = 0; i < Half; ++i)
1706 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1708 for (unsigned i = Half; i < NumElems; ++i)
1709 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1715 bool X86::isSHUFPMask(SDNode *N) {
1716 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1717 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1720 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
1721 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1722 /// half elements to come from vector 1 (which would equal the dest.) and
1723 /// the upper half to come from vector 2.
1724 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1725 if (NumOps != 2 && NumOps != 4) return false;
1727 unsigned Half = NumOps / 2;
1728 for (unsigned i = 0; i < Half; ++i)
1729 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1731 for (unsigned i = Half; i < NumOps; ++i)
1732 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1737 static bool isCommutedSHUFP(SDNode *N) {
1738 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1739 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1742 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1743 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1744 bool X86::isMOVHLPSMask(SDNode *N) {
1745 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1747 if (N->getNumOperands() != 4)
1750 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1751 return isUndefOrEqual(N->getOperand(0), 6) &&
1752 isUndefOrEqual(N->getOperand(1), 7) &&
1753 isUndefOrEqual(N->getOperand(2), 2) &&
1754 isUndefOrEqual(N->getOperand(3), 3);
1757 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1758 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1760 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1761 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1763 if (N->getNumOperands() != 4)
1766 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1767 return isUndefOrEqual(N->getOperand(0), 2) &&
1768 isUndefOrEqual(N->getOperand(1), 3) &&
1769 isUndefOrEqual(N->getOperand(2), 2) &&
1770 isUndefOrEqual(N->getOperand(3), 3);
1773 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1774 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1775 bool X86::isMOVLPMask(SDNode *N) {
1776 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1778 unsigned NumElems = N->getNumOperands();
1779 if (NumElems != 2 && NumElems != 4)
1782 for (unsigned i = 0; i < NumElems/2; ++i)
1783 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1786 for (unsigned i = NumElems/2; i < NumElems; ++i)
1787 if (!isUndefOrEqual(N->getOperand(i), i))
1793 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1794 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1796 bool X86::isMOVHPMask(SDNode *N) {
1797 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1799 unsigned NumElems = N->getNumOperands();
1800 if (NumElems != 2 && NumElems != 4)
1803 for (unsigned i = 0; i < NumElems/2; ++i)
1804 if (!isUndefOrEqual(N->getOperand(i), i))
1807 for (unsigned i = 0; i < NumElems/2; ++i) {
1808 SDOperand Arg = N->getOperand(i + NumElems/2);
1809 if (!isUndefOrEqual(Arg, i + NumElems))
1816 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1817 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1818 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1819 bool V2IsSplat = false) {
1820 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1823 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1824 SDOperand BitI = Elts[i];
1825 SDOperand BitI1 = Elts[i+1];
1826 if (!isUndefOrEqual(BitI, j))
1829 if (isUndefOrEqual(BitI1, NumElts))
1832 if (!isUndefOrEqual(BitI1, j + NumElts))
1840 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1841 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1842 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1845 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1846 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1847 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1848 bool V2IsSplat = false) {
1849 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1852 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1853 SDOperand BitI = Elts[i];
1854 SDOperand BitI1 = Elts[i+1];
1855 if (!isUndefOrEqual(BitI, j + NumElts/2))
1858 if (isUndefOrEqual(BitI1, NumElts))
1861 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1869 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1870 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1871 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1874 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1875 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1877 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1878 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1880 unsigned NumElems = N->getNumOperands();
1881 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1884 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1885 SDOperand BitI = N->getOperand(i);
1886 SDOperand BitI1 = N->getOperand(i+1);
1888 if (!isUndefOrEqual(BitI, j))
1890 if (!isUndefOrEqual(BitI1, j))
1897 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
1898 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
1900 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
1901 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1903 unsigned NumElems = N->getNumOperands();
1904 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1907 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
1908 SDOperand BitI = N->getOperand(i);
1909 SDOperand BitI1 = N->getOperand(i + 1);
1911 if (!isUndefOrEqual(BitI, j))
1913 if (!isUndefOrEqual(BitI1, j))
1920 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1921 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1922 /// MOVSD, and MOVD, i.e. setting the lowest element.
1923 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1924 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1927 if (!isUndefOrEqual(Elts[0], NumElts))
1930 for (unsigned i = 1; i < NumElts; ++i) {
1931 if (!isUndefOrEqual(Elts[i], i))
1938 bool X86::isMOVLMask(SDNode *N) {
1939 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1940 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1943 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1944 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1945 /// element of vector 2 and the other elements to come from vector 1 in order.
1946 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1947 bool V2IsSplat = false,
1948 bool V2IsUndef = false) {
1949 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1952 if (!isUndefOrEqual(Ops[0], 0))
1955 for (unsigned i = 1; i < NumOps; ++i) {
1956 SDOperand Arg = Ops[i];
1957 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1958 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1959 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1966 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1967 bool V2IsUndef = false) {
1968 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1969 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1970 V2IsSplat, V2IsUndef);
1973 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1974 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1975 bool X86::isMOVSHDUPMask(SDNode *N) {
1976 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1978 if (N->getNumOperands() != 4)
1981 // Expect 1, 1, 3, 3
1982 for (unsigned i = 0; i < 2; ++i) {
1983 SDOperand Arg = N->getOperand(i);
1984 if (Arg.getOpcode() == ISD::UNDEF) continue;
1985 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1986 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1987 if (Val != 1) return false;
1991 for (unsigned i = 2; i < 4; ++i) {
1992 SDOperand Arg = N->getOperand(i);
1993 if (Arg.getOpcode() == ISD::UNDEF) continue;
1994 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1995 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1996 if (Val != 3) return false;
2000 // Don't use movshdup if it can be done with a shufps.
2004 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2005 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2006 bool X86::isMOVSLDUPMask(SDNode *N) {
2007 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2009 if (N->getNumOperands() != 4)
2012 // Expect 0, 0, 2, 2
2013 for (unsigned i = 0; i < 2; ++i) {
2014 SDOperand Arg = N->getOperand(i);
2015 if (Arg.getOpcode() == ISD::UNDEF) continue;
2016 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2017 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2018 if (Val != 0) return false;
2022 for (unsigned i = 2; i < 4; ++i) {
2023 SDOperand Arg = N->getOperand(i);
2024 if (Arg.getOpcode() == ISD::UNDEF) continue;
2025 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2026 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2027 if (Val != 2) return false;
2031 // Don't use movshdup if it can be done with a shufps.
2035 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2036 /// specifies a identity operation on the LHS or RHS.
2037 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2038 unsigned NumElems = N->getNumOperands();
2039 for (unsigned i = 0; i < NumElems; ++i)
2040 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2045 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2046 /// a splat of a single element.
2047 static bool isSplatMask(SDNode *N) {
2048 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2050 // This is a splat operation if each element of the permute is the same, and
2051 // if the value doesn't reference the second vector.
2052 unsigned NumElems = N->getNumOperands();
2053 SDOperand ElementBase;
2055 for (; i != NumElems; ++i) {
2056 SDOperand Elt = N->getOperand(i);
2057 if (isa<ConstantSDNode>(Elt)) {
2063 if (!ElementBase.Val)
2066 for (; i != NumElems; ++i) {
2067 SDOperand Arg = N->getOperand(i);
2068 if (Arg.getOpcode() == ISD::UNDEF) continue;
2069 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2070 if (Arg != ElementBase) return false;
2073 // Make sure it is a splat of the first vector operand.
2074 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2077 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2078 /// a splat of a single element and it's a 2 or 4 element mask.
2079 bool X86::isSplatMask(SDNode *N) {
2080 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2082 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2083 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2085 return ::isSplatMask(N);
2088 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2089 /// specifies a splat of zero element.
2090 bool X86::isSplatLoMask(SDNode *N) {
2091 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2093 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2094 if (!isUndefOrEqual(N->getOperand(i), 0))
2099 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2100 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2102 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2103 unsigned NumOperands = N->getNumOperands();
2104 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2106 for (unsigned i = 0; i < NumOperands; ++i) {
2108 SDOperand Arg = N->getOperand(NumOperands-i-1);
2109 if (Arg.getOpcode() != ISD::UNDEF)
2110 Val = cast<ConstantSDNode>(Arg)->getValue();
2111 if (Val >= NumOperands) Val -= NumOperands;
2113 if (i != NumOperands - 1)
2120 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2121 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2123 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2125 // 8 nodes, but we only care about the last 4.
2126 for (unsigned i = 7; i >= 4; --i) {
2128 SDOperand Arg = N->getOperand(i);
2129 if (Arg.getOpcode() != ISD::UNDEF)
2130 Val = cast<ConstantSDNode>(Arg)->getValue();
2139 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2140 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2142 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2144 // 8 nodes, but we only care about the first 4.
2145 for (int i = 3; i >= 0; --i) {
2147 SDOperand Arg = N->getOperand(i);
2148 if (Arg.getOpcode() != ISD::UNDEF)
2149 Val = cast<ConstantSDNode>(Arg)->getValue();
2158 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2159 /// specifies a 8 element shuffle that can be broken into a pair of
2160 /// PSHUFHW and PSHUFLW.
2161 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2162 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2164 if (N->getNumOperands() != 8)
2167 // Lower quadword shuffled.
2168 for (unsigned i = 0; i != 4; ++i) {
2169 SDOperand Arg = N->getOperand(i);
2170 if (Arg.getOpcode() == ISD::UNDEF) continue;
2171 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2172 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2177 // Upper quadword shuffled.
2178 for (unsigned i = 4; i != 8; ++i) {
2179 SDOperand Arg = N->getOperand(i);
2180 if (Arg.getOpcode() == ISD::UNDEF) continue;
2181 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2182 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2183 if (Val < 4 || Val > 7)
2190 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2191 /// values in ther permute mask.
2192 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2193 SDOperand &V2, SDOperand &Mask,
2194 SelectionDAG &DAG) {
2195 MVT::ValueType VT = Op.getValueType();
2196 MVT::ValueType MaskVT = Mask.getValueType();
2197 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2198 unsigned NumElems = Mask.getNumOperands();
2199 SmallVector<SDOperand, 8> MaskVec;
2201 for (unsigned i = 0; i != NumElems; ++i) {
2202 SDOperand Arg = Mask.getOperand(i);
2203 if (Arg.getOpcode() == ISD::UNDEF) {
2204 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2207 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2208 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2210 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2212 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2216 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2217 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2220 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2221 /// match movhlps. The lower half elements should come from upper half of
2222 /// V1 (and in order), and the upper half elements should come from the upper
2223 /// half of V2 (and in order).
2224 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2225 unsigned NumElems = Mask->getNumOperands();
2228 for (unsigned i = 0, e = 2; i != e; ++i)
2229 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2231 for (unsigned i = 2; i != 4; ++i)
2232 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2237 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2238 /// is promoted to a vector.
2239 static inline bool isScalarLoadToVector(SDNode *N) {
2240 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2241 N = N->getOperand(0).Val;
2242 return ISD::isNON_EXTLoad(N);
2247 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2248 /// match movlp{s|d}. The lower half elements should come from lower half of
2249 /// V1 (and in order), and the upper half elements should come from the upper
2250 /// half of V2 (and in order). And since V1 will become the source of the
2251 /// MOVLP, it must be either a vector load or a scalar load to vector.
2252 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2253 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2255 // Is V2 is a vector load, don't do this transformation. We will try to use
2256 // load folding shufps op.
2257 if (ISD::isNON_EXTLoad(V2))
2260 unsigned NumElems = Mask->getNumOperands();
2261 if (NumElems != 2 && NumElems != 4)
2263 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2264 if (!isUndefOrEqual(Mask->getOperand(i), i))
2266 for (unsigned i = NumElems/2; i != NumElems; ++i)
2267 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2272 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2274 static bool isSplatVector(SDNode *N) {
2275 if (N->getOpcode() != ISD::BUILD_VECTOR)
2278 SDOperand SplatValue = N->getOperand(0);
2279 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2280 if (N->getOperand(i) != SplatValue)
2285 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2287 static bool isUndefShuffle(SDNode *N) {
2288 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2291 SDOperand V1 = N->getOperand(0);
2292 SDOperand V2 = N->getOperand(1);
2293 SDOperand Mask = N->getOperand(2);
2294 unsigned NumElems = Mask.getNumOperands();
2295 for (unsigned i = 0; i != NumElems; ++i) {
2296 SDOperand Arg = Mask.getOperand(i);
2297 if (Arg.getOpcode() != ISD::UNDEF) {
2298 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2299 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2301 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2308 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2310 static inline bool isZeroNode(SDOperand Elt) {
2311 return ((isa<ConstantSDNode>(Elt) &&
2312 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2313 (isa<ConstantFPSDNode>(Elt) &&
2314 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2317 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2318 /// to an zero vector.
2319 static bool isZeroShuffle(SDNode *N) {
2320 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2323 SDOperand V1 = N->getOperand(0);
2324 SDOperand V2 = N->getOperand(1);
2325 SDOperand Mask = N->getOperand(2);
2326 unsigned NumElems = Mask.getNumOperands();
2327 for (unsigned i = 0; i != NumElems; ++i) {
2328 SDOperand Arg = Mask.getOperand(i);
2329 if (Arg.getOpcode() != ISD::UNDEF) {
2330 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2331 if (Idx < NumElems) {
2332 unsigned Opc = V1.Val->getOpcode();
2333 if (Opc == ISD::UNDEF)
2335 if (Opc != ISD::BUILD_VECTOR ||
2336 !isZeroNode(V1.Val->getOperand(Idx)))
2338 } else if (Idx >= NumElems) {
2339 unsigned Opc = V2.Val->getOpcode();
2340 if (Opc == ISD::UNDEF)
2342 if (Opc != ISD::BUILD_VECTOR ||
2343 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2351 /// getZeroVector - Returns a vector of specified type with all zero elements.
2353 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2354 assert(MVT::isVector(VT) && "Expected a vector type");
2355 unsigned NumElems = MVT::getVectorNumElements(VT);
2356 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2357 bool isFP = MVT::isFloatingPoint(EVT);
2358 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2359 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2360 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2363 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2364 /// that point to V2 points to its first element.
2365 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2366 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2368 bool Changed = false;
2369 SmallVector<SDOperand, 8> MaskVec;
2370 unsigned NumElems = Mask.getNumOperands();
2371 for (unsigned i = 0; i != NumElems; ++i) {
2372 SDOperand Arg = Mask.getOperand(i);
2373 if (Arg.getOpcode() != ISD::UNDEF) {
2374 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2375 if (Val > NumElems) {
2376 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2380 MaskVec.push_back(Arg);
2384 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2385 &MaskVec[0], MaskVec.size());
2389 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2390 /// operation of specified width.
2391 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2392 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2393 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2395 SmallVector<SDOperand, 8> MaskVec;
2396 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2397 for (unsigned i = 1; i != NumElems; ++i)
2398 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2399 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2402 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2403 /// of specified width.
2404 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2405 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2406 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2407 SmallVector<SDOperand, 8> MaskVec;
2408 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2409 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2410 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2412 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2415 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2416 /// of specified width.
2417 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2418 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2419 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2420 unsigned Half = NumElems/2;
2421 SmallVector<SDOperand, 8> MaskVec;
2422 for (unsigned i = 0; i != Half; ++i) {
2423 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2424 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2426 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2429 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2431 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2432 SDOperand V1 = Op.getOperand(0);
2433 SDOperand Mask = Op.getOperand(2);
2434 MVT::ValueType VT = Op.getValueType();
2435 unsigned NumElems = Mask.getNumOperands();
2436 Mask = getUnpacklMask(NumElems, DAG);
2437 while (NumElems != 4) {
2438 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2441 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2443 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2444 Mask = getZeroVector(MaskVT, DAG);
2445 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2446 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2447 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2450 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2451 /// vector of zero or undef vector.
2452 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2453 unsigned NumElems, unsigned Idx,
2454 bool isZero, SelectionDAG &DAG) {
2455 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2456 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2457 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2458 SDOperand Zero = DAG.getConstant(0, EVT);
2459 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2460 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2461 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2462 &MaskVec[0], MaskVec.size());
2463 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2466 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2468 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2469 unsigned NumNonZero, unsigned NumZero,
2470 SelectionDAG &DAG, TargetLowering &TLI) {
2476 for (unsigned i = 0; i < 16; ++i) {
2477 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2478 if (ThisIsNonZero && First) {
2480 V = getZeroVector(MVT::v8i16, DAG);
2482 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2487 SDOperand ThisElt(0, 0), LastElt(0, 0);
2488 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2489 if (LastIsNonZero) {
2490 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2492 if (ThisIsNonZero) {
2493 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2494 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2495 ThisElt, DAG.getConstant(8, MVT::i8));
2497 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2502 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2503 DAG.getConstant(i/2, TLI.getPointerTy()));
2507 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2510 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2512 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2513 unsigned NumNonZero, unsigned NumZero,
2514 SelectionDAG &DAG, TargetLowering &TLI) {
2520 for (unsigned i = 0; i < 8; ++i) {
2521 bool isNonZero = (NonZeros & (1 << i)) != 0;
2525 V = getZeroVector(MVT::v8i16, DAG);
2527 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2530 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2531 DAG.getConstant(i, TLI.getPointerTy()));
2539 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2540 // All zero's are handled with pxor.
2541 if (ISD::isBuildVectorAllZeros(Op.Val))
2544 // All one's are handled with pcmpeqd.
2545 if (ISD::isBuildVectorAllOnes(Op.Val))
2548 MVT::ValueType VT = Op.getValueType();
2549 MVT::ValueType EVT = MVT::getVectorElementType(VT);
2550 unsigned EVTBits = MVT::getSizeInBits(EVT);
2552 unsigned NumElems = Op.getNumOperands();
2553 unsigned NumZero = 0;
2554 unsigned NumNonZero = 0;
2555 unsigned NonZeros = 0;
2556 unsigned NumNonZeroImms = 0;
2557 std::set<SDOperand> Values;
2558 for (unsigned i = 0; i < NumElems; ++i) {
2559 SDOperand Elt = Op.getOperand(i);
2560 if (Elt.getOpcode() != ISD::UNDEF) {
2562 if (isZeroNode(Elt))
2565 NonZeros |= (1 << i);
2567 if (Elt.getOpcode() == ISD::Constant ||
2568 Elt.getOpcode() == ISD::ConstantFP)
2574 if (NumNonZero == 0) {
2576 // All undef vector. Return an UNDEF.
2577 return DAG.getNode(ISD::UNDEF, VT);
2579 // A mix of zero and undef. Return a zero vector.
2580 return getZeroVector(VT, DAG);
2583 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2584 if (Values.size() == 1)
2587 // Special case for single non-zero element.
2588 if (NumNonZero == 1) {
2589 unsigned Idx = CountTrailingZeros_32(NonZeros);
2590 SDOperand Item = Op.getOperand(Idx);
2591 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2593 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2594 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2597 if (EVTBits == 32) {
2598 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2599 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2601 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2602 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2603 SmallVector<SDOperand, 8> MaskVec;
2604 for (unsigned i = 0; i < NumElems; i++)
2605 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2606 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2607 &MaskVec[0], MaskVec.size());
2608 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2609 DAG.getNode(ISD::UNDEF, VT), Mask);
2613 // A vector full of immediates; various special cases are already
2614 // handled, so this is best done with a single constant-pool load.
2615 if (NumNonZero == NumNonZeroImms)
2618 // Let legalizer expand 2-wide build_vectors.
2622 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2623 if (EVTBits == 8 && NumElems == 16) {
2624 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2626 if (V.Val) return V;
2629 if (EVTBits == 16 && NumElems == 8) {
2630 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2632 if (V.Val) return V;
2635 // If element VT is == 32 bits, turn it into a number of shuffles.
2636 SmallVector<SDOperand, 8> V;
2638 if (NumElems == 4 && NumZero > 0) {
2639 for (unsigned i = 0; i < 4; ++i) {
2640 bool isZero = !(NonZeros & (1 << i));
2642 V[i] = getZeroVector(VT, DAG);
2644 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2647 for (unsigned i = 0; i < 2; ++i) {
2648 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2651 V[i] = V[i*2]; // Must be a zero vector.
2654 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2655 getMOVLMask(NumElems, DAG));
2658 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2659 getMOVLMask(NumElems, DAG));
2662 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2663 getUnpacklMask(NumElems, DAG));
2668 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2669 // clears the upper bits.
2670 // FIXME: we can do the same for v4f32 case when we know both parts of
2671 // the lower half come from scalar_to_vector (loadf32). We should do
2672 // that in post legalizer dag combiner with target specific hooks.
2673 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2675 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2676 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
2677 SmallVector<SDOperand, 8> MaskVec;
2678 bool Reverse = (NonZeros & 0x3) == 2;
2679 for (unsigned i = 0; i < 2; ++i)
2681 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2683 MaskVec.push_back(DAG.getConstant(i, EVT));
2684 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2685 for (unsigned i = 0; i < 2; ++i)
2687 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2689 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2690 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2691 &MaskVec[0], MaskVec.size());
2692 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2695 if (Values.size() > 2) {
2696 // Expand into a number of unpckl*.
2698 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2699 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2700 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2701 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2702 for (unsigned i = 0; i < NumElems; ++i)
2703 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2705 while (NumElems != 0) {
2706 for (unsigned i = 0; i < NumElems; ++i)
2707 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2718 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2719 SDOperand V1 = Op.getOperand(0);
2720 SDOperand V2 = Op.getOperand(1);
2721 SDOperand PermMask = Op.getOperand(2);
2722 MVT::ValueType VT = Op.getValueType();
2723 unsigned NumElems = PermMask.getNumOperands();
2724 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2725 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2726 bool V1IsSplat = false;
2727 bool V2IsSplat = false;
2729 if (isUndefShuffle(Op.Val))
2730 return DAG.getNode(ISD::UNDEF, VT);
2732 if (isZeroShuffle(Op.Val))
2733 return getZeroVector(VT, DAG);
2735 if (isIdentityMask(PermMask.Val))
2737 else if (isIdentityMask(PermMask.Val, true))
2740 if (isSplatMask(PermMask.Val)) {
2741 if (NumElems <= 4) return Op;
2742 // Promote it to a v4i32 splat.
2743 return PromoteSplat(Op, DAG);
2746 if (X86::isMOVLMask(PermMask.Val))
2747 return (V1IsUndef) ? V2 : Op;
2749 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2750 X86::isMOVSLDUPMask(PermMask.Val) ||
2751 X86::isMOVHLPSMask(PermMask.Val) ||
2752 X86::isMOVHPMask(PermMask.Val) ||
2753 X86::isMOVLPMask(PermMask.Val))
2756 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2757 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2758 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2760 bool Commuted = false;
2761 V1IsSplat = isSplatVector(V1.Val);
2762 V2IsSplat = isSplatVector(V2.Val);
2763 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2764 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2765 std::swap(V1IsSplat, V2IsSplat);
2766 std::swap(V1IsUndef, V2IsUndef);
2770 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2771 if (V2IsUndef) return V1;
2772 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2774 // V2 is a splat, so the mask may be malformed. That is, it may point
2775 // to any V2 element. The instruction selectior won't like this. Get
2776 // a corrected mask and commute to form a proper MOVS{S|D}.
2777 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2778 if (NewMask.Val != PermMask.Val)
2779 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2784 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2785 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2786 X86::isUNPCKLMask(PermMask.Val) ||
2787 X86::isUNPCKHMask(PermMask.Val))
2791 // Normalize mask so all entries that point to V2 points to its first
2792 // element then try to match unpck{h|l} again. If match, return a
2793 // new vector_shuffle with the corrected mask.
2794 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2795 if (NewMask.Val != PermMask.Val) {
2796 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2797 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2798 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2799 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2800 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2801 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2806 // Normalize the node to match x86 shuffle ops if needed
2807 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2808 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2811 // Commute is back and try unpck* again.
2812 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2813 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2814 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
2815 X86::isUNPCKLMask(PermMask.Val) ||
2816 X86::isUNPCKHMask(PermMask.Val))
2820 // If VT is integer, try PSHUF* first, then SHUFP*.
2821 if (MVT::isInteger(VT)) {
2822 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
2823 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
2824 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
2825 X86::isPSHUFDMask(PermMask.Val)) ||
2826 X86::isPSHUFHWMask(PermMask.Val) ||
2827 X86::isPSHUFLWMask(PermMask.Val)) {
2828 if (V2.getOpcode() != ISD::UNDEF)
2829 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2830 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2834 if (X86::isSHUFPMask(PermMask.Val) &&
2835 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
2838 // Handle v8i16 shuffle high / low shuffle node pair.
2839 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2840 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2841 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2842 SmallVector<SDOperand, 8> MaskVec;
2843 for (unsigned i = 0; i != 4; ++i)
2844 MaskVec.push_back(PermMask.getOperand(i));
2845 for (unsigned i = 4; i != 8; ++i)
2846 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2847 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2848 &MaskVec[0], MaskVec.size());
2849 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2851 for (unsigned i = 0; i != 4; ++i)
2852 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2853 for (unsigned i = 4; i != 8; ++i)
2854 MaskVec.push_back(PermMask.getOperand(i));
2855 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2856 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2859 // Floating point cases in the other order.
2860 if (X86::isSHUFPMask(PermMask.Val))
2862 if (X86::isPSHUFDMask(PermMask.Val) ||
2863 X86::isPSHUFHWMask(PermMask.Val) ||
2864 X86::isPSHUFLWMask(PermMask.Val)) {
2865 if (V2.getOpcode() != ISD::UNDEF)
2866 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2867 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2872 if (NumElems == 4 &&
2873 // Don't do this for MMX.
2874 MVT::getSizeInBits(VT) != 64) {
2875 MVT::ValueType MaskVT = PermMask.getValueType();
2876 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
2877 SmallVector<std::pair<int, int>, 8> Locs;
2878 Locs.reserve(NumElems);
2879 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2880 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2883 // If no more than two elements come from either vector. This can be
2884 // implemented with two shuffles. First shuffle gather the elements.
2885 // The second shuffle, which takes the first shuffle as both of its
2886 // vector operands, put the elements into the right order.
2887 for (unsigned i = 0; i != NumElems; ++i) {
2888 SDOperand Elt = PermMask.getOperand(i);
2889 if (Elt.getOpcode() == ISD::UNDEF) {
2890 Locs[i] = std::make_pair(-1, -1);
2892 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2893 if (Val < NumElems) {
2894 Locs[i] = std::make_pair(0, NumLo);
2898 Locs[i] = std::make_pair(1, NumHi);
2899 if (2+NumHi < NumElems)
2900 Mask1[2+NumHi] = Elt;
2905 if (NumLo <= 2 && NumHi <= 2) {
2906 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2907 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2908 &Mask1[0], Mask1.size()));
2909 for (unsigned i = 0; i != NumElems; ++i) {
2910 if (Locs[i].first == -1)
2913 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2914 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2915 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2919 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2920 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2921 &Mask2[0], Mask2.size()));
2924 // Break it into (shuffle shuffle_hi, shuffle_lo).
2926 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2927 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2928 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2929 unsigned MaskIdx = 0;
2931 unsigned HiIdx = NumElems/2;
2932 for (unsigned i = 0; i != NumElems; ++i) {
2933 if (i == NumElems/2) {
2939 SDOperand Elt = PermMask.getOperand(i);
2940 if (Elt.getOpcode() == ISD::UNDEF) {
2941 Locs[i] = std::make_pair(-1, -1);
2942 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2943 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2944 (*MaskPtr)[LoIdx] = Elt;
2947 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2948 (*MaskPtr)[HiIdx] = Elt;
2953 SDOperand LoShuffle =
2954 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2955 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2956 &LoMask[0], LoMask.size()));
2957 SDOperand HiShuffle =
2958 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2959 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2960 &HiMask[0], HiMask.size()));
2961 SmallVector<SDOperand, 8> MaskOps;
2962 for (unsigned i = 0; i != NumElems; ++i) {
2963 if (Locs[i].first == -1) {
2964 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2966 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2967 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2970 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2971 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2972 &MaskOps[0], MaskOps.size()));
2979 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2980 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2983 MVT::ValueType VT = Op.getValueType();
2984 // TODO: handle v16i8.
2985 if (MVT::getSizeInBits(VT) == 16) {
2986 // Transform it so it match pextrw which produces a 32-bit result.
2987 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2988 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2989 Op.getOperand(0), Op.getOperand(1));
2990 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2991 DAG.getValueType(VT));
2992 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2993 } else if (MVT::getSizeInBits(VT) == 32) {
2994 SDOperand Vec = Op.getOperand(0);
2995 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2998 // SHUFPS the element to the lowest double word, then movss.
2999 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3000 SmallVector<SDOperand, 8> IdxVec;
3001 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3002 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3003 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3004 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3005 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3006 &IdxVec[0], IdxVec.size());
3007 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3008 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3009 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3010 DAG.getConstant(0, getPointerTy()));
3011 } else if (MVT::getSizeInBits(VT) == 64) {
3012 SDOperand Vec = Op.getOperand(0);
3013 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3017 // UNPCKHPD the element to the lowest double word, then movsd.
3018 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3019 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3020 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3021 SmallVector<SDOperand, 8> IdxVec;
3022 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3023 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3024 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3025 &IdxVec[0], IdxVec.size());
3026 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3027 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3028 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3029 DAG.getConstant(0, getPointerTy()));
3036 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3037 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3038 // as its second argument.
3039 MVT::ValueType VT = Op.getValueType();
3040 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3041 SDOperand N0 = Op.getOperand(0);
3042 SDOperand N1 = Op.getOperand(1);
3043 SDOperand N2 = Op.getOperand(2);
3044 if (MVT::getSizeInBits(BaseVT) == 16) {
3045 if (N1.getValueType() != MVT::i32)
3046 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3047 if (N2.getValueType() != MVT::i32)
3048 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3049 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3050 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3051 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3054 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3055 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3056 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3057 SmallVector<SDOperand, 8> MaskVec;
3058 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3059 for (unsigned i = 1; i <= 3; ++i)
3060 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3061 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3062 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3063 &MaskVec[0], MaskVec.size()));
3065 // Use two pinsrw instructions to insert a 32 bit value.
3067 if (MVT::isFloatingPoint(N1.getValueType())) {
3068 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3069 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3070 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3071 DAG.getConstant(0, getPointerTy()));
3073 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3074 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3075 DAG.getConstant(Idx, getPointerTy()));
3076 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3077 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3078 DAG.getConstant(Idx+1, getPointerTy()));
3079 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3087 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3088 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3089 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3092 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3093 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3094 // one of the above mentioned nodes. It has to be wrapped because otherwise
3095 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3096 // be used to form addressing mode. These wrapped nodes will be selected
3099 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3100 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3101 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3103 CP->getAlignment());
3104 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3105 // With PIC, the address is actually $g + Offset.
3106 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3107 !Subtarget->isPICStyleRIPRel()) {
3108 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3109 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3117 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3118 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3119 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3120 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3121 // With PIC, the address is actually $g + Offset.
3122 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3123 !Subtarget->isPICStyleRIPRel()) {
3124 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3125 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3129 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3130 // load the value at address GV, not the value of GV itself. This means that
3131 // the GlobalAddress must be in the base or index register of the address, not
3132 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3133 // The same applies for external symbols during PIC codegen
3134 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3135 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3140 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3142 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3143 const MVT::ValueType PtrVT) {
3145 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3146 DAG.getNode(X86ISD::GlobalBaseReg,
3148 InFlag = Chain.getValue(1);
3150 // emit leal symbol@TLSGD(,%ebx,1), %eax
3151 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3152 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3153 GA->getValueType(0),
3155 SDOperand Ops[] = { Chain, TGA, InFlag };
3156 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3157 InFlag = Result.getValue(2);
3158 Chain = Result.getValue(1);
3160 // call ___tls_get_addr. This function receives its argument in
3161 // the register EAX.
3162 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3163 InFlag = Chain.getValue(1);
3165 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3166 SDOperand Ops1[] = { Chain,
3167 DAG.getTargetExternalSymbol("___tls_get_addr",
3169 DAG.getRegister(X86::EAX, PtrVT),
3170 DAG.getRegister(X86::EBX, PtrVT),
3172 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3173 InFlag = Chain.getValue(1);
3175 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3178 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3179 // "local exec" model.
3181 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3182 const MVT::ValueType PtrVT) {
3183 // Get the Thread Pointer
3184 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3185 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3187 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3188 GA->getValueType(0),
3190 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3192 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3193 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3195 // The address of the thread local variable is the add of the thread
3196 // pointer with the offset of the variable.
3197 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3201 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3202 // TODO: implement the "local dynamic" model
3203 // TODO: implement the "initial exec"model for pic executables
3204 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3205 "TLS not implemented for non-ELF and 64-bit targets");
3206 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3207 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3208 // otherwise use the "Local Exec"TLS Model
3209 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3210 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3212 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3216 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3217 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3218 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3219 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3220 // With PIC, the address is actually $g + Offset.
3221 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3222 !Subtarget->isPICStyleRIPRel()) {
3223 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3224 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3231 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3232 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3233 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3234 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3235 // With PIC, the address is actually $g + Offset.
3236 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3237 !Subtarget->isPICStyleRIPRel()) {
3238 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3239 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3246 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3247 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3248 "Not an i64 shift!");
3249 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3250 SDOperand ShOpLo = Op.getOperand(0);
3251 SDOperand ShOpHi = Op.getOperand(1);
3252 SDOperand ShAmt = Op.getOperand(2);
3253 SDOperand Tmp1 = isSRA ?
3254 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3255 DAG.getConstant(0, MVT::i32);
3257 SDOperand Tmp2, Tmp3;
3258 if (Op.getOpcode() == ISD::SHL_PARTS) {
3259 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3260 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3262 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3263 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3266 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3267 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3268 DAG.getConstant(32, MVT::i8));
3269 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3270 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3273 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3275 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3276 SmallVector<SDOperand, 4> Ops;
3277 if (Op.getOpcode() == ISD::SHL_PARTS) {
3278 Ops.push_back(Tmp2);
3279 Ops.push_back(Tmp3);
3281 Ops.push_back(InFlag);
3282 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3283 InFlag = Hi.getValue(1);
3286 Ops.push_back(Tmp3);
3287 Ops.push_back(Tmp1);
3289 Ops.push_back(InFlag);
3290 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3292 Ops.push_back(Tmp2);
3293 Ops.push_back(Tmp3);
3295 Ops.push_back(InFlag);
3296 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3297 InFlag = Lo.getValue(1);
3300 Ops.push_back(Tmp3);
3301 Ops.push_back(Tmp1);
3303 Ops.push_back(InFlag);
3304 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3307 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3311 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3314 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3315 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3316 Op.getOperand(0).getValueType() >= MVT::i16 &&
3317 "Unknown SINT_TO_FP to lower!");
3320 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3321 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3322 MachineFunction &MF = DAG.getMachineFunction();
3323 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3324 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3325 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3326 StackSlot, NULL, 0);
3331 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3333 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
3334 SmallVector<SDOperand, 8> Ops;
3335 Ops.push_back(Chain);
3336 Ops.push_back(StackSlot);
3337 Ops.push_back(DAG.getValueType(SrcVT));
3338 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3339 Tys, &Ops[0], Ops.size());
3342 Chain = Result.getValue(1);
3343 SDOperand InFlag = Result.getValue(2);
3345 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3346 // shouldn't be necessary except that RFP cannot be live across
3347 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3348 MachineFunction &MF = DAG.getMachineFunction();
3349 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3350 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3351 Tys = DAG.getVTList(MVT::Other);
3352 SmallVector<SDOperand, 8> Ops;
3353 Ops.push_back(Chain);
3354 Ops.push_back(Result);
3355 Ops.push_back(StackSlot);
3356 Ops.push_back(DAG.getValueType(Op.getValueType()));
3357 Ops.push_back(InFlag);
3358 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3359 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3365 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3366 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3367 "Unknown FP_TO_SINT to lower!");
3368 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3370 MachineFunction &MF = DAG.getMachineFunction();
3371 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3372 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3373 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3376 switch (Op.getValueType()) {
3377 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3378 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3379 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3380 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3383 SDOperand Chain = DAG.getEntryNode();
3384 SDOperand Value = Op.getOperand(0);
3386 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3387 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3388 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
3390 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3392 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3393 Chain = Value.getValue(1);
3394 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3395 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3398 // Build the FP_TO_INT*_IN_MEM
3399 SDOperand Ops[] = { Chain, Value, StackSlot };
3400 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3403 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3406 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3407 MVT::ValueType VT = Op.getValueType();
3408 MVT::ValueType EltVT = VT;
3409 if (MVT::isVector(VT))
3410 EltVT = MVT::getVectorElementType(VT);
3411 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3412 std::vector<Constant*> CV;
3413 if (EltVT == MVT::f64) {
3414 Constant *C = ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63)));
3418 Constant *C = ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31)));
3424 Constant *C = ConstantVector::get(CV);
3425 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3426 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3428 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3431 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3432 MVT::ValueType VT = Op.getValueType();
3433 MVT::ValueType EltVT = VT;
3434 unsigned EltNum = 1;
3435 if (MVT::isVector(VT)) {
3436 EltVT = MVT::getVectorElementType(VT);
3437 EltNum = MVT::getVectorNumElements(VT);
3439 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
3440 std::vector<Constant*> CV;
3441 if (EltVT == MVT::f64) {
3442 Constant *C = ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63));
3446 Constant *C = ConstantFP::get(OpNTy, BitsToFloat(1U << 31));
3452 Constant *C = ConstantVector::get(CV);
3453 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3454 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3456 if (MVT::isVector(VT)) {
3457 return DAG.getNode(ISD::BIT_CONVERT, VT,
3458 DAG.getNode(ISD::XOR, MVT::v2i64,
3459 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
3460 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
3462 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3466 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3467 SDOperand Op0 = Op.getOperand(0);
3468 SDOperand Op1 = Op.getOperand(1);
3469 MVT::ValueType VT = Op.getValueType();
3470 MVT::ValueType SrcVT = Op1.getValueType();
3471 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3473 // If second operand is smaller, extend it first.
3474 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3475 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3479 // First get the sign bit of second operand.
3480 std::vector<Constant*> CV;
3481 if (SrcVT == MVT::f64) {
3482 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3483 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3485 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3486 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3487 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3488 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3490 Constant *C = ConstantVector::get(CV);
3491 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3492 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
3494 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3496 // Shift sign bit right or left if the two operands have different types.
3497 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3498 // Op0 is MVT::f32, Op1 is MVT::f64.
3499 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3500 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3501 DAG.getConstant(32, MVT::i32));
3502 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3503 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3504 DAG.getConstant(0, getPointerTy()));
3507 // Clear first operand sign bit.
3509 if (VT == MVT::f64) {
3510 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3511 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3513 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3514 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3515 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3516 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3518 C = ConstantVector::get(CV);
3519 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
3520 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
3522 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3524 // Or the value with the sign bit.
3525 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3528 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3530 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3532 SDOperand Op0 = Op.getOperand(0);
3533 SDOperand Op1 = Op.getOperand(1);
3534 SDOperand CC = Op.getOperand(2);
3535 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3536 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3537 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3538 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3541 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3543 SDOperand Ops1[] = { Chain, Op0, Op1 };
3544 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3545 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3546 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3549 assert(isFP && "Illegal integer SetCC!");
3551 SDOperand COps[] = { Chain, Op0, Op1 };
3552 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3554 switch (SetCCOpcode) {
3555 default: assert(false && "Illegal floating point SetCC!");
3556 case ISD::SETOEQ: { // !PF & ZF
3557 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3558 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3559 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3561 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3562 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3564 case ISD::SETUNE: { // PF | !ZF
3565 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3566 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3567 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3569 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3570 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3575 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3576 bool addTest = true;
3577 SDOperand Chain = DAG.getEntryNode();
3578 SDOperand Cond = Op.getOperand(0);
3580 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3582 if (Cond.getOpcode() == ISD::SETCC)
3583 Cond = LowerSETCC(Cond, DAG, Chain);
3585 if (Cond.getOpcode() == X86ISD::SETCC) {
3586 CC = Cond.getOperand(0);
3588 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3589 // (since flag operand cannot be shared). Use it as the condition setting
3590 // operand in place of the X86ISD::SETCC.
3591 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3592 // to use a test instead of duplicating the X86ISD::CMP (for register
3593 // pressure reason)?
3594 SDOperand Cmp = Cond.getOperand(1);
3595 unsigned Opc = Cmp.getOpcode();
3596 bool IllegalFPCMov = !X86ScalarSSE &&
3597 MVT::isFloatingPoint(Op.getValueType()) &&
3598 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3599 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3601 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3602 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3608 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3609 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3610 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3613 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3614 SmallVector<SDOperand, 4> Ops;
3615 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3616 // condition is true.
3617 Ops.push_back(Op.getOperand(2));
3618 Ops.push_back(Op.getOperand(1));
3620 Ops.push_back(Cond.getValue(1));
3621 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3624 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3625 bool addTest = true;
3626 SDOperand Chain = Op.getOperand(0);
3627 SDOperand Cond = Op.getOperand(1);
3628 SDOperand Dest = Op.getOperand(2);
3630 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3632 if (Cond.getOpcode() == ISD::SETCC)
3633 Cond = LowerSETCC(Cond, DAG, Chain);
3635 if (Cond.getOpcode() == X86ISD::SETCC) {
3636 CC = Cond.getOperand(0);
3638 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3639 // (since flag operand cannot be shared). Use it as the condition setting
3640 // operand in place of the X86ISD::SETCC.
3641 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3642 // to use a test instead of duplicating the X86ISD::CMP (for register
3643 // pressure reason)?
3644 SDOperand Cmp = Cond.getOperand(1);
3645 unsigned Opc = Cmp.getOpcode();
3646 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3647 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3648 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3654 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3655 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3656 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3658 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3659 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3662 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3663 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3665 if (Subtarget->is64Bit())
3666 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3668 switch (CallingConv) {
3670 assert(0 && "Unsupported calling convention");
3671 case CallingConv::Fast:
3672 // TODO: Implement fastcc
3674 case CallingConv::C:
3675 case CallingConv::X86_StdCall:
3676 return LowerCCCCallTo(Op, DAG, CallingConv);
3677 case CallingConv::X86_FastCall:
3678 return LowerFastCCCallTo(Op, DAG, CallingConv);
3683 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
3684 // Calls to _alloca is needed to probe the stack when allocating more than 4k
3685 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
3686 // that the guard pages used by the OS virtual memory manager are allocated in
3687 // correct sequence.
3689 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
3690 SelectionDAG &DAG) {
3691 assert(Subtarget->isTargetCygMing() &&
3692 "This should be used only on Cygwin/Mingw targets");
3695 SDOperand Chain = Op.getOperand(0);
3696 SDOperand Size = Op.getOperand(1);
3697 // FIXME: Ensure alignment here
3701 MVT::ValueType IntPtr = getPointerTy();
3702 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
3704 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
3705 Flag = Chain.getValue(1);
3707 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3708 SDOperand Ops[] = { Chain,
3709 DAG.getTargetExternalSymbol("_alloca", IntPtr),
3710 DAG.getRegister(X86::EAX, IntPtr),
3712 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
3713 Flag = Chain.getValue(1);
3715 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
3717 std::vector<MVT::ValueType> Tys;
3718 Tys.push_back(SPTy);
3719 Tys.push_back(MVT::Other);
3720 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
3721 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
3725 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3726 MachineFunction &MF = DAG.getMachineFunction();
3727 const Function* Fn = MF.getFunction();
3728 if (Fn->hasExternalLinkage() &&
3729 Subtarget->isTargetCygMing() &&
3730 Fn->getName() == "main")
3731 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
3733 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3734 if (Subtarget->is64Bit())
3735 return LowerX86_64CCCArguments(Op, DAG);
3739 assert(0 && "Unsupported calling convention");
3740 case CallingConv::Fast:
3741 // TODO: implement fastcc.
3744 case CallingConv::C:
3745 return LowerCCCArguments(Op, DAG);
3746 case CallingConv::X86_StdCall:
3747 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
3748 return LowerCCCArguments(Op, DAG, true);
3749 case CallingConv::X86_FastCall:
3750 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
3751 return LowerFastCCArguments(Op, DAG);
3755 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3756 SDOperand InFlag(0, 0);
3757 SDOperand Chain = Op.getOperand(0);
3759 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3760 if (Align == 0) Align = 1;
3762 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3763 // If not DWORD aligned or size is more than the threshold, call memset.
3764 // The libc version is likely to be faster for these cases. It can use the
3765 // address value and run time information about the CPU.
3766 if ((Align & 3) != 0 ||
3767 (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
3768 MVT::ValueType IntPtr = getPointerTy();
3769 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3770 TargetLowering::ArgListTy Args;
3771 TargetLowering::ArgListEntry Entry;
3772 Entry.Node = Op.getOperand(1);
3773 Entry.Ty = IntPtrTy;
3774 Args.push_back(Entry);
3775 // Extend the unsigned i8 argument to be an int value for the call.
3776 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3777 Entry.Ty = IntPtrTy;
3778 Args.push_back(Entry);
3779 Entry.Node = Op.getOperand(3);
3780 Args.push_back(Entry);
3781 std::pair<SDOperand,SDOperand> CallResult =
3782 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3783 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3784 return CallResult.second;
3789 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3790 unsigned BytesLeft = 0;
3791 bool TwoRepStos = false;
3794 uint64_t Val = ValC->getValue() & 255;
3796 // If the value is a constant, then we can potentially use larger sets.
3797 switch (Align & 3) {
3798 case 2: // WORD aligned
3801 Val = (Val << 8) | Val;
3803 case 0: // DWORD aligned
3806 Val = (Val << 8) | Val;
3807 Val = (Val << 16) | Val;
3808 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3811 Val = (Val << 32) | Val;
3814 default: // Byte aligned
3817 Count = Op.getOperand(3);
3821 if (AVT > MVT::i8) {
3823 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3824 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3825 BytesLeft = I->getValue() % UBytes;
3827 assert(AVT >= MVT::i32 &&
3828 "Do not use rep;stos if not at least DWORD aligned");
3829 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3830 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3835 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3837 InFlag = Chain.getValue(1);
3840 Count = Op.getOperand(3);
3841 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3842 InFlag = Chain.getValue(1);
3845 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3847 InFlag = Chain.getValue(1);
3848 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3849 Op.getOperand(1), InFlag);
3850 InFlag = Chain.getValue(1);
3852 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3853 SmallVector<SDOperand, 8> Ops;
3854 Ops.push_back(Chain);
3855 Ops.push_back(DAG.getValueType(AVT));
3856 Ops.push_back(InFlag);
3857 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3860 InFlag = Chain.getValue(1);
3861 Count = Op.getOperand(3);
3862 MVT::ValueType CVT = Count.getValueType();
3863 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3864 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3865 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3867 InFlag = Chain.getValue(1);
3868 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3870 Ops.push_back(Chain);
3871 Ops.push_back(DAG.getValueType(MVT::i8));
3872 Ops.push_back(InFlag);
3873 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3874 } else if (BytesLeft) {
3875 // Issue stores for the last 1 - 7 bytes.
3877 unsigned Val = ValC->getValue() & 255;
3878 unsigned Offset = I->getValue() - BytesLeft;
3879 SDOperand DstAddr = Op.getOperand(1);
3880 MVT::ValueType AddrVT = DstAddr.getValueType();
3881 if (BytesLeft >= 4) {
3882 Val = (Val << 8) | Val;
3883 Val = (Val << 16) | Val;
3884 Value = DAG.getConstant(Val, MVT::i32);
3885 Chain = DAG.getStore(Chain, Value,
3886 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3887 DAG.getConstant(Offset, AddrVT)),
3892 if (BytesLeft >= 2) {
3893 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3894 Chain = DAG.getStore(Chain, Value,
3895 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3896 DAG.getConstant(Offset, AddrVT)),
3901 if (BytesLeft == 1) {
3902 Value = DAG.getConstant(Val, MVT::i8);
3903 Chain = DAG.getStore(Chain, Value,
3904 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3905 DAG.getConstant(Offset, AddrVT)),
3913 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3914 SDOperand Chain = Op.getOperand(0);
3916 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3917 if (Align == 0) Align = 1;
3919 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3920 // If not DWORD aligned or size is more than the threshold, call memcpy.
3921 // The libc version is likely to be faster for these cases. It can use the
3922 // address value and run time information about the CPU.
3923 // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
3924 if ((Align & 3) != 0 ||
3925 (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
3926 MVT::ValueType IntPtr = getPointerTy();
3927 TargetLowering::ArgListTy Args;
3928 TargetLowering::ArgListEntry Entry;
3929 Entry.Ty = getTargetData()->getIntPtrType();
3930 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3931 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3932 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3933 std::pair<SDOperand,SDOperand> CallResult =
3934 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3935 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3936 return CallResult.second;
3941 unsigned BytesLeft = 0;
3942 bool TwoRepMovs = false;
3943 switch (Align & 3) {
3944 case 2: // WORD aligned
3947 case 0: // DWORD aligned
3949 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3952 default: // Byte aligned
3954 Count = Op.getOperand(3);
3958 if (AVT > MVT::i8) {
3960 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3961 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3962 BytesLeft = I->getValue() % UBytes;
3964 assert(AVT >= MVT::i32 &&
3965 "Do not use rep;movs if not at least DWORD aligned");
3966 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3967 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3972 SDOperand InFlag(0, 0);
3973 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3975 InFlag = Chain.getValue(1);
3976 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3977 Op.getOperand(1), InFlag);
3978 InFlag = Chain.getValue(1);
3979 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3980 Op.getOperand(2), InFlag);
3981 InFlag = Chain.getValue(1);
3983 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3984 SmallVector<SDOperand, 8> Ops;
3985 Ops.push_back(Chain);
3986 Ops.push_back(DAG.getValueType(AVT));
3987 Ops.push_back(InFlag);
3988 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3991 InFlag = Chain.getValue(1);
3992 Count = Op.getOperand(3);
3993 MVT::ValueType CVT = Count.getValueType();
3994 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3995 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3996 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3998 InFlag = Chain.getValue(1);
3999 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4001 Ops.push_back(Chain);
4002 Ops.push_back(DAG.getValueType(MVT::i8));
4003 Ops.push_back(InFlag);
4004 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4005 } else if (BytesLeft) {
4006 // Issue loads and stores for the last 1 - 7 bytes.
4007 unsigned Offset = I->getValue() - BytesLeft;
4008 SDOperand DstAddr = Op.getOperand(1);
4009 MVT::ValueType DstVT = DstAddr.getValueType();
4010 SDOperand SrcAddr = Op.getOperand(2);
4011 MVT::ValueType SrcVT = SrcAddr.getValueType();
4013 if (BytesLeft >= 4) {
4014 Value = DAG.getLoad(MVT::i32, Chain,
4015 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4016 DAG.getConstant(Offset, SrcVT)),
4018 Chain = Value.getValue(1);
4019 Chain = DAG.getStore(Chain, Value,
4020 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4021 DAG.getConstant(Offset, DstVT)),
4026 if (BytesLeft >= 2) {
4027 Value = DAG.getLoad(MVT::i16, Chain,
4028 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4029 DAG.getConstant(Offset, SrcVT)),
4031 Chain = Value.getValue(1);
4032 Chain = DAG.getStore(Chain, Value,
4033 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4034 DAG.getConstant(Offset, DstVT)),
4040 if (BytesLeft == 1) {
4041 Value = DAG.getLoad(MVT::i8, Chain,
4042 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4043 DAG.getConstant(Offset, SrcVT)),
4045 Chain = Value.getValue(1);
4046 Chain = DAG.getStore(Chain, Value,
4047 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4048 DAG.getConstant(Offset, DstVT)),
4057 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4058 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4059 SDOperand TheOp = Op.getOperand(0);
4060 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4061 if (Subtarget->is64Bit()) {
4062 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4063 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4064 MVT::i64, Copy1.getValue(2));
4065 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4066 DAG.getConstant(32, MVT::i8));
4068 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4071 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4072 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4075 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4076 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4077 MVT::i32, Copy1.getValue(2));
4078 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4079 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4080 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4083 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4084 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4086 if (!Subtarget->is64Bit()) {
4087 // vastart just stores the address of the VarArgsFrameIndex slot into the
4088 // memory location argument.
4089 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4090 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4095 // gp_offset (0 - 6 * 8)
4096 // fp_offset (48 - 48 + 8 * 16)
4097 // overflow_arg_area (point to parameters coming in memory).
4099 SmallVector<SDOperand, 8> MemOps;
4100 SDOperand FIN = Op.getOperand(1);
4102 SDOperand Store = DAG.getStore(Op.getOperand(0),
4103 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4104 FIN, SV->getValue(), SV->getOffset());
4105 MemOps.push_back(Store);
4108 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4109 DAG.getConstant(4, getPointerTy()));
4110 Store = DAG.getStore(Op.getOperand(0),
4111 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4112 FIN, SV->getValue(), SV->getOffset());
4113 MemOps.push_back(Store);
4115 // Store ptr to overflow_arg_area
4116 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4117 DAG.getConstant(4, getPointerTy()));
4118 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4119 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4121 MemOps.push_back(Store);
4123 // Store ptr to reg_save_area.
4124 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4125 DAG.getConstant(8, getPointerTy()));
4126 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4127 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4129 MemOps.push_back(Store);
4130 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4133 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4134 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4135 SDOperand Chain = Op.getOperand(0);
4136 SDOperand DstPtr = Op.getOperand(1);
4137 SDOperand SrcPtr = Op.getOperand(2);
4138 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4139 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4141 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4142 SrcSV->getValue(), SrcSV->getOffset());
4143 Chain = SrcPtr.getValue(1);
4144 for (unsigned i = 0; i < 3; ++i) {
4145 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4146 SrcSV->getValue(), SrcSV->getOffset());
4147 Chain = Val.getValue(1);
4148 Chain = DAG.getStore(Chain, Val, DstPtr,
4149 DstSV->getValue(), DstSV->getOffset());
4152 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4153 DAG.getConstant(8, getPointerTy()));
4154 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4155 DAG.getConstant(8, getPointerTy()));
4161 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4162 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4164 default: return SDOperand(); // Don't custom lower most intrinsics.
4165 // Comparison intrinsics.
4166 case Intrinsic::x86_sse_comieq_ss:
4167 case Intrinsic::x86_sse_comilt_ss:
4168 case Intrinsic::x86_sse_comile_ss:
4169 case Intrinsic::x86_sse_comigt_ss:
4170 case Intrinsic::x86_sse_comige_ss:
4171 case Intrinsic::x86_sse_comineq_ss:
4172 case Intrinsic::x86_sse_ucomieq_ss:
4173 case Intrinsic::x86_sse_ucomilt_ss:
4174 case Intrinsic::x86_sse_ucomile_ss:
4175 case Intrinsic::x86_sse_ucomigt_ss:
4176 case Intrinsic::x86_sse_ucomige_ss:
4177 case Intrinsic::x86_sse_ucomineq_ss:
4178 case Intrinsic::x86_sse2_comieq_sd:
4179 case Intrinsic::x86_sse2_comilt_sd:
4180 case Intrinsic::x86_sse2_comile_sd:
4181 case Intrinsic::x86_sse2_comigt_sd:
4182 case Intrinsic::x86_sse2_comige_sd:
4183 case Intrinsic::x86_sse2_comineq_sd:
4184 case Intrinsic::x86_sse2_ucomieq_sd:
4185 case Intrinsic::x86_sse2_ucomilt_sd:
4186 case Intrinsic::x86_sse2_ucomile_sd:
4187 case Intrinsic::x86_sse2_ucomigt_sd:
4188 case Intrinsic::x86_sse2_ucomige_sd:
4189 case Intrinsic::x86_sse2_ucomineq_sd: {
4191 ISD::CondCode CC = ISD::SETCC_INVALID;
4194 case Intrinsic::x86_sse_comieq_ss:
4195 case Intrinsic::x86_sse2_comieq_sd:
4199 case Intrinsic::x86_sse_comilt_ss:
4200 case Intrinsic::x86_sse2_comilt_sd:
4204 case Intrinsic::x86_sse_comile_ss:
4205 case Intrinsic::x86_sse2_comile_sd:
4209 case Intrinsic::x86_sse_comigt_ss:
4210 case Intrinsic::x86_sse2_comigt_sd:
4214 case Intrinsic::x86_sse_comige_ss:
4215 case Intrinsic::x86_sse2_comige_sd:
4219 case Intrinsic::x86_sse_comineq_ss:
4220 case Intrinsic::x86_sse2_comineq_sd:
4224 case Intrinsic::x86_sse_ucomieq_ss:
4225 case Intrinsic::x86_sse2_ucomieq_sd:
4226 Opc = X86ISD::UCOMI;
4229 case Intrinsic::x86_sse_ucomilt_ss:
4230 case Intrinsic::x86_sse2_ucomilt_sd:
4231 Opc = X86ISD::UCOMI;
4234 case Intrinsic::x86_sse_ucomile_ss:
4235 case Intrinsic::x86_sse2_ucomile_sd:
4236 Opc = X86ISD::UCOMI;
4239 case Intrinsic::x86_sse_ucomigt_ss:
4240 case Intrinsic::x86_sse2_ucomigt_sd:
4241 Opc = X86ISD::UCOMI;
4244 case Intrinsic::x86_sse_ucomige_ss:
4245 case Intrinsic::x86_sse2_ucomige_sd:
4246 Opc = X86ISD::UCOMI;
4249 case Intrinsic::x86_sse_ucomineq_ss:
4250 case Intrinsic::x86_sse2_ucomineq_sd:
4251 Opc = X86ISD::UCOMI;
4257 SDOperand LHS = Op.getOperand(1);
4258 SDOperand RHS = Op.getOperand(2);
4259 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4261 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4262 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4263 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4264 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4265 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4266 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4267 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4272 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4273 // Depths > 0 not supported yet!
4274 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4277 // Just load the return address
4278 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4279 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4282 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4283 // Depths > 0 not supported yet!
4284 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4287 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4288 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4289 DAG.getConstant(4, getPointerTy()));
4292 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4293 SelectionDAG &DAG) {
4294 // Is not yet supported on x86-64
4295 if (Subtarget->is64Bit())
4298 return DAG.getConstant(8, getPointerTy());
4301 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4303 assert(!Subtarget->is64Bit() &&
4304 "Lowering of eh_return builtin is not supported yet on x86-64");
4306 MachineFunction &MF = DAG.getMachineFunction();
4307 SDOperand Chain = Op.getOperand(0);
4308 SDOperand Offset = Op.getOperand(1);
4309 SDOperand Handler = Op.getOperand(2);
4311 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4314 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4315 DAG.getConstant(-4UL, getPointerTy()));
4316 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
4317 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
4318 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
4319 MF.addLiveOut(X86::ECX);
4321 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
4322 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
4325 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
4326 SelectionDAG &DAG) {
4327 SDOperand Root = Op.getOperand(0);
4328 SDOperand Trmp = Op.getOperand(1); // trampoline
4329 SDOperand FPtr = Op.getOperand(2); // nested function
4330 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
4332 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
4334 if (Subtarget->is64Bit()) {
4335 return SDOperand(); // not yet supported
4337 Function *Func = (Function *)
4338 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
4339 unsigned CC = Func->getCallingConv();
4344 assert(0 && "Unsupported calling convention");
4345 case CallingConv::C:
4346 case CallingConv::Fast:
4347 case CallingConv::X86_StdCall: {
4348 // Pass 'nest' parameter in ECX.
4349 // Must be kept in sync with X86CallingConv.td
4352 // Check that ECX wasn't needed by an 'inreg' parameter.
4353 const FunctionType *FTy = Func->getFunctionType();
4354 const ParamAttrsList *Attrs = FTy->getParamAttrs();
4356 if (Attrs && !Func->isVarArg()) {
4357 unsigned InRegCount = 0;
4360 for (FunctionType::param_iterator I = FTy->param_begin(),
4361 E = FTy->param_end(); I != E; ++I, ++Idx)
4362 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
4363 // FIXME: should only count parameters that are lowered to integers.
4364 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
4366 if (InRegCount > 2) {
4367 cerr << "Nest register in use - reduce number of inreg parameters!\n";
4373 case CallingConv::X86_FastCall:
4374 // Pass 'nest' parameter in EAX.
4375 // Must be kept in sync with X86CallingConv.td
4380 const X86InstrInfo *TII =
4381 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
4383 SDOperand OutChains[4];
4384 SDOperand Addr, Disp;
4386 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
4387 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
4389 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
4390 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
4391 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
4392 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
4394 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
4395 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
4396 TrmpSV->getOffset() + 1, false, 1);
4398 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
4399 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
4400 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
4401 TrmpSV->getValue() + 5, TrmpSV->getOffset());
4403 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
4404 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
4405 TrmpSV->getOffset() + 6, false, 1);
4407 return DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4);
4411 /// LowerOperation - Provide custom lowering hooks for some operations.
4413 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4414 switch (Op.getOpcode()) {
4415 default: assert(0 && "Should not custom lower this!");
4416 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4417 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4418 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4419 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4420 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4421 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4422 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4423 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4424 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4425 case ISD::SHL_PARTS:
4426 case ISD::SRA_PARTS:
4427 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4428 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4429 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4430 case ISD::FABS: return LowerFABS(Op, DAG);
4431 case ISD::FNEG: return LowerFNEG(Op, DAG);
4432 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4433 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4434 case ISD::SELECT: return LowerSELECT(Op, DAG);
4435 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4436 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4437 case ISD::CALL: return LowerCALL(Op, DAG);
4438 case ISD::RET: return LowerRET(Op, DAG);
4439 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4440 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4441 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4442 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4443 case ISD::VASTART: return LowerVASTART(Op, DAG);
4444 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
4445 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4446 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4447 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4448 case ISD::FRAME_TO_ARGS_OFFSET:
4449 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
4450 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
4451 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
4452 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4457 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4459 default: return NULL;
4460 case X86ISD::SHLD: return "X86ISD::SHLD";
4461 case X86ISD::SHRD: return "X86ISD::SHRD";
4462 case X86ISD::FAND: return "X86ISD::FAND";
4463 case X86ISD::FOR: return "X86ISD::FOR";
4464 case X86ISD::FXOR: return "X86ISD::FXOR";
4465 case X86ISD::FSRL: return "X86ISD::FSRL";
4466 case X86ISD::FILD: return "X86ISD::FILD";
4467 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4468 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4469 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4470 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4471 case X86ISD::FLD: return "X86ISD::FLD";
4472 case X86ISD::FST: return "X86ISD::FST";
4473 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4474 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4475 case X86ISD::CALL: return "X86ISD::CALL";
4476 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4477 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4478 case X86ISD::CMP: return "X86ISD::CMP";
4479 case X86ISD::COMI: return "X86ISD::COMI";
4480 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4481 case X86ISD::SETCC: return "X86ISD::SETCC";
4482 case X86ISD::CMOV: return "X86ISD::CMOV";
4483 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4484 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4485 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4486 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4487 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4488 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4489 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4490 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4491 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4492 case X86ISD::FMAX: return "X86ISD::FMAX";
4493 case X86ISD::FMIN: return "X86ISD::FMIN";
4494 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
4495 case X86ISD::FRCP: return "X86ISD::FRCP";
4496 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
4497 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
4498 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
4502 // isLegalAddressingMode - Return true if the addressing mode represented
4503 // by AM is legal for this target, for a load/store of the specified type.
4504 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
4505 const Type *Ty) const {
4506 // X86 supports extremely general addressing modes.
4508 // X86 allows a sign-extended 32-bit immediate field as a displacement.
4509 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
4513 // We can only fold this if we don't need an extra load.
4514 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
4517 // X86-64 only supports addr of globals in small code model.
4518 if (Subtarget->is64Bit()) {
4519 if (getTargetMachine().getCodeModel() != CodeModel::Small)
4521 // If lower 4G is not available, then we must use rip-relative addressing.
4522 if (AM.BaseOffs || AM.Scale > 1)
4533 // These scales always work.
4538 // These scales are formed with basereg+scalereg. Only accept if there is
4543 default: // Other stuff never works.
4551 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4552 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4553 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4554 /// are assumed to be legal.
4556 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4557 // Only do shuffles on 128-bit vector types for now.
4558 if (MVT::getSizeInBits(VT) == 64) return false;
4559 return (Mask.Val->getNumOperands() <= 4 ||
4560 isIdentityMask(Mask.Val) ||
4561 isIdentityMask(Mask.Val, true) ||
4562 isSplatMask(Mask.Val) ||
4563 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4564 X86::isUNPCKLMask(Mask.Val) ||
4565 X86::isUNPCKHMask(Mask.Val) ||
4566 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4567 X86::isUNPCKH_v_undef_Mask(Mask.Val));
4570 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4572 SelectionDAG &DAG) const {
4573 unsigned NumElts = BVOps.size();
4574 // Only do shuffles on 128-bit vector types for now.
4575 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4576 if (NumElts == 2) return true;
4578 return (isMOVLMask(&BVOps[0], 4) ||
4579 isCommutedMOVL(&BVOps[0], 4, true) ||
4580 isSHUFPMask(&BVOps[0], 4) ||
4581 isCommutedSHUFP(&BVOps[0], 4));
4586 //===----------------------------------------------------------------------===//
4587 // X86 Scheduler Hooks
4588 //===----------------------------------------------------------------------===//
4591 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4592 MachineBasicBlock *BB) {
4593 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4594 switch (MI->getOpcode()) {
4595 default: assert(false && "Unexpected instr type to insert");
4596 case X86::CMOV_FR32:
4597 case X86::CMOV_FR64:
4598 case X86::CMOV_V4F32:
4599 case X86::CMOV_V2F64:
4600 case X86::CMOV_V2I64: {
4601 // To "insert" a SELECT_CC instruction, we actually have to insert the
4602 // diamond control-flow pattern. The incoming instruction knows the
4603 // destination vreg to set, the condition code register to branch on, the
4604 // true/false values to select between, and a branch opcode to use.
4605 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4606 ilist<MachineBasicBlock>::iterator It = BB;
4612 // cmpTY ccX, r1, r2
4614 // fallthrough --> copy0MBB
4615 MachineBasicBlock *thisMBB = BB;
4616 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4617 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4619 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4620 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4621 MachineFunction *F = BB->getParent();
4622 F->getBasicBlockList().insert(It, copy0MBB);
4623 F->getBasicBlockList().insert(It, sinkMBB);
4624 // Update machine-CFG edges by first adding all successors of the current
4625 // block to the new block which will contain the Phi node for the select.
4626 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4627 e = BB->succ_end(); i != e; ++i)
4628 sinkMBB->addSuccessor(*i);
4629 // Next, remove all successors of the current block, and add the true
4630 // and fallthrough blocks as its successors.
4631 while(!BB->succ_empty())
4632 BB->removeSuccessor(BB->succ_begin());
4633 BB->addSuccessor(copy0MBB);
4634 BB->addSuccessor(sinkMBB);
4637 // %FalseValue = ...
4638 // # fallthrough to sinkMBB
4641 // Update machine-CFG edges
4642 BB->addSuccessor(sinkMBB);
4645 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4648 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4649 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4650 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4652 delete MI; // The pseudo instruction is gone now.
4656 case X86::FP32_TO_INT16_IN_MEM:
4657 case X86::FP32_TO_INT32_IN_MEM:
4658 case X86::FP32_TO_INT64_IN_MEM:
4659 case X86::FP64_TO_INT16_IN_MEM:
4660 case X86::FP64_TO_INT32_IN_MEM:
4661 case X86::FP64_TO_INT64_IN_MEM:
4662 case X86::FP80_TO_INT16_IN_MEM:
4663 case X86::FP80_TO_INT32_IN_MEM:
4664 case X86::FP80_TO_INT64_IN_MEM: {
4665 // Change the floating point control register to use "round towards zero"
4666 // mode when truncating to an integer value.
4667 MachineFunction *F = BB->getParent();
4668 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4669 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4671 // Load the old value of the high byte of the control word...
4673 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4674 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4676 // Set the high part to be round to zero...
4677 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4680 // Reload the modified control word now...
4681 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4683 // Restore the memory image of control word to original value
4684 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4687 // Get the X86 opcode to use.
4689 switch (MI->getOpcode()) {
4690 default: assert(0 && "illegal opcode!");
4691 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
4692 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
4693 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
4694 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
4695 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
4696 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
4697 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
4698 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
4699 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
4703 MachineOperand &Op = MI->getOperand(0);
4704 if (Op.isRegister()) {
4705 AM.BaseType = X86AddressMode::RegBase;
4706 AM.Base.Reg = Op.getReg();
4708 AM.BaseType = X86AddressMode::FrameIndexBase;
4709 AM.Base.FrameIndex = Op.getFrameIndex();
4711 Op = MI->getOperand(1);
4712 if (Op.isImmediate())
4713 AM.Scale = Op.getImm();
4714 Op = MI->getOperand(2);
4715 if (Op.isImmediate())
4716 AM.IndexReg = Op.getImm();
4717 Op = MI->getOperand(3);
4718 if (Op.isGlobalAddress()) {
4719 AM.GV = Op.getGlobal();
4721 AM.Disp = Op.getImm();
4723 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4724 .addReg(MI->getOperand(4).getReg());
4726 // Reload the original control word now.
4727 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4729 delete MI; // The pseudo instruction is gone now.
4735 //===----------------------------------------------------------------------===//
4736 // X86 Optimization Hooks
4737 //===----------------------------------------------------------------------===//
4739 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4741 uint64_t &KnownZero,
4743 const SelectionDAG &DAG,
4744 unsigned Depth) const {
4745 unsigned Opc = Op.getOpcode();
4746 assert((Opc >= ISD::BUILTIN_OP_END ||
4747 Opc == ISD::INTRINSIC_WO_CHAIN ||
4748 Opc == ISD::INTRINSIC_W_CHAIN ||
4749 Opc == ISD::INTRINSIC_VOID) &&
4750 "Should use MaskedValueIsZero if you don't know whether Op"
4751 " is a target node!");
4753 KnownZero = KnownOne = 0; // Don't know anything.
4757 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4762 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4763 /// element of the result of the vector shuffle.
4764 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4765 MVT::ValueType VT = N->getValueType(0);
4766 SDOperand PermMask = N->getOperand(2);
4767 unsigned NumElems = PermMask.getNumOperands();
4768 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4770 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4772 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
4773 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4774 SDOperand Idx = PermMask.getOperand(i);
4775 if (Idx.getOpcode() == ISD::UNDEF)
4776 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
4777 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4782 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4783 /// node is a GlobalAddress + an offset.
4784 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4785 unsigned Opc = N->getOpcode();
4786 if (Opc == X86ISD::Wrapper) {
4787 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4788 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4791 } else if (Opc == ISD::ADD) {
4792 SDOperand N1 = N->getOperand(0);
4793 SDOperand N2 = N->getOperand(1);
4794 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4795 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4797 Offset += V->getSignExtended();
4800 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4801 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4803 Offset += V->getSignExtended();
4811 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4813 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4814 MachineFrameInfo *MFI) {
4815 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4818 SDOperand Loc = N->getOperand(1);
4819 SDOperand BaseLoc = Base->getOperand(1);
4820 if (Loc.getOpcode() == ISD::FrameIndex) {
4821 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4823 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
4824 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4825 int FS = MFI->getObjectSize(FI);
4826 int BFS = MFI->getObjectSize(BFI);
4827 if (FS != BFS || FS != Size) return false;
4828 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4830 GlobalValue *GV1 = NULL;
4831 GlobalValue *GV2 = NULL;
4832 int64_t Offset1 = 0;
4833 int64_t Offset2 = 0;
4834 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4835 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4836 if (isGA1 && isGA2 && GV1 == GV2)
4837 return Offset1 == (Offset2 + Dist*Size);
4843 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4844 const X86Subtarget *Subtarget) {
4847 if (isGAPlusOffset(Base, GV, Offset))
4848 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4850 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4851 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
4853 // Fixed objects do not specify alignment, however the offsets are known.
4854 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4855 (MFI->getObjectOffset(BFI) % 16) == 0);
4857 return MFI->getObjectAlignment(BFI) >= 16;
4863 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4864 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4865 /// if the load addresses are consecutive, non-overlapping, and in the right
4867 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4868 const X86Subtarget *Subtarget) {
4869 MachineFunction &MF = DAG.getMachineFunction();
4870 MachineFrameInfo *MFI = MF.getFrameInfo();
4871 MVT::ValueType VT = N->getValueType(0);
4872 MVT::ValueType EVT = MVT::getVectorElementType(VT);
4873 SDOperand PermMask = N->getOperand(2);
4874 int NumElems = (int)PermMask.getNumOperands();
4875 SDNode *Base = NULL;
4876 for (int i = 0; i < NumElems; ++i) {
4877 SDOperand Idx = PermMask.getOperand(i);
4878 if (Idx.getOpcode() == ISD::UNDEF) {
4879 if (!Base) return SDOperand();
4882 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4883 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4887 else if (!isConsecutiveLoad(Arg.Val, Base,
4888 i, MVT::getSizeInBits(EVT)/8,MFI))
4893 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4894 LoadSDNode *LD = cast<LoadSDNode>(Base);
4896 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4897 LD->getSrcValueOffset(), LD->isVolatile());
4899 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4900 LD->getSrcValueOffset(), LD->isVolatile(),
4901 LD->getAlignment());
4905 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4906 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4907 const X86Subtarget *Subtarget) {
4908 SDOperand Cond = N->getOperand(0);
4910 // If we have SSE[12] support, try to form min/max nodes.
4911 if (Subtarget->hasSSE2() &&
4912 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4913 if (Cond.getOpcode() == ISD::SETCC) {
4914 // Get the LHS/RHS of the select.
4915 SDOperand LHS = N->getOperand(1);
4916 SDOperand RHS = N->getOperand(2);
4917 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4919 unsigned Opcode = 0;
4920 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4923 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4926 if (!UnsafeFPMath) break;
4928 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4930 Opcode = X86ISD::FMIN;
4933 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4936 if (!UnsafeFPMath) break;
4938 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4940 Opcode = X86ISD::FMAX;
4943 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4946 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4949 if (!UnsafeFPMath) break;
4951 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4953 Opcode = X86ISD::FMIN;
4956 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4959 if (!UnsafeFPMath) break;
4961 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4963 Opcode = X86ISD::FMAX;
4969 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4978 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4979 DAGCombinerInfo &DCI) const {
4980 SelectionDAG &DAG = DCI.DAG;
4981 switch (N->getOpcode()) {
4983 case ISD::VECTOR_SHUFFLE:
4984 return PerformShuffleCombine(N, DAG, Subtarget);
4986 return PerformSELECTCombine(N, DAG, Subtarget);
4992 //===----------------------------------------------------------------------===//
4993 // X86 Inline Assembly Support
4994 //===----------------------------------------------------------------------===//
4996 /// getConstraintType - Given a constraint letter, return the type of
4997 /// constraint it is for this target.
4998 X86TargetLowering::ConstraintType
4999 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5000 if (Constraint.size() == 1) {
5001 switch (Constraint[0]) {
5010 return C_RegisterClass;
5015 return TargetLowering::getConstraintType(Constraint);
5018 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5019 /// vector. If it is invalid, don't add anything to Ops.
5020 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5022 std::vector<SDOperand>&Ops,
5023 SelectionDAG &DAG) {
5024 SDOperand Result(0, 0);
5026 switch (Constraint) {
5029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5030 if (C->getValue() <= 31) {
5031 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5037 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5038 if (C->getValue() <= 255) {
5039 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5045 // Literal immediates are always ok.
5046 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5047 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5051 // If we are in non-pic codegen mode, we allow the address of a global (with
5052 // an optional displacement) to be used with 'i'.
5053 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5056 // Match either (GA) or (GA+C)
5058 Offset = GA->getOffset();
5059 } else if (Op.getOpcode() == ISD::ADD) {
5060 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5061 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5063 Offset = GA->getOffset()+C->getValue();
5065 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5066 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5068 Offset = GA->getOffset()+C->getValue();
5075 // If addressing this global requires a load (e.g. in PIC mode), we can't
5077 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5081 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5087 // Otherwise, not valid for this mode.
5093 Ops.push_back(Result);
5096 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5099 std::vector<unsigned> X86TargetLowering::
5100 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5101 MVT::ValueType VT) const {
5102 if (Constraint.size() == 1) {
5103 // FIXME: not handling fp-stack yet!
5104 switch (Constraint[0]) { // GCC X86 Constraint Letters
5105 default: break; // Unknown constraint letter
5106 case 'A': // EAX/EDX
5107 if (VT == MVT::i32 || VT == MVT::i64)
5108 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5110 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5113 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5114 else if (VT == MVT::i16)
5115 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5116 else if (VT == MVT::i8)
5117 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5122 return std::vector<unsigned>();
5125 std::pair<unsigned, const TargetRegisterClass*>
5126 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5127 MVT::ValueType VT) const {
5128 // First, see if this is a constraint that directly corresponds to an LLVM
5130 if (Constraint.size() == 1) {
5131 // GCC Constraint Letters
5132 switch (Constraint[0]) {
5134 case 'r': // GENERAL_REGS
5135 case 'R': // LEGACY_REGS
5136 case 'l': // INDEX_REGS
5137 if (VT == MVT::i64 && Subtarget->is64Bit())
5138 return std::make_pair(0U, X86::GR64RegisterClass);
5140 return std::make_pair(0U, X86::GR32RegisterClass);
5141 else if (VT == MVT::i16)
5142 return std::make_pair(0U, X86::GR16RegisterClass);
5143 else if (VT == MVT::i8)
5144 return std::make_pair(0U, X86::GR8RegisterClass);
5146 case 'y': // MMX_REGS if MMX allowed.
5147 if (!Subtarget->hasMMX()) break;
5148 return std::make_pair(0U, X86::VR64RegisterClass);
5150 case 'Y': // SSE_REGS if SSE2 allowed
5151 if (!Subtarget->hasSSE2()) break;
5153 case 'x': // SSE_REGS if SSE1 allowed
5154 if (!Subtarget->hasSSE1()) break;
5158 // Scalar SSE types.
5161 return std::make_pair(0U, X86::FR32RegisterClass);
5164 return std::make_pair(0U, X86::FR64RegisterClass);
5172 return std::make_pair(0U, X86::VR128RegisterClass);
5178 // Use the default implementation in TargetLowering to convert the register
5179 // constraint into a member of a register class.
5180 std::pair<unsigned, const TargetRegisterClass*> Res;
5181 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5183 // Not found as a standard register?
5184 if (Res.second == 0) {
5185 // GCC calls "st(0)" just plain "st".
5186 if (StringsEqualNoCase("{st}", Constraint)) {
5187 Res.first = X86::ST0;
5188 Res.second = X86::RSTRegisterClass;
5194 // Otherwise, check to see if this is a register class of the wrong value
5195 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5196 // turn into {ax},{dx}.
5197 if (Res.second->hasType(VT))
5198 return Res; // Correct type already, nothing to do.
5200 // All of the single-register GCC register classes map their values onto
5201 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5202 // really want an 8-bit or 32-bit register, map to the appropriate register
5203 // class and return the appropriate register.
5204 if (Res.second != X86::GR16RegisterClass)
5207 if (VT == MVT::i8) {
5208 unsigned DestReg = 0;
5209 switch (Res.first) {
5211 case X86::AX: DestReg = X86::AL; break;
5212 case X86::DX: DestReg = X86::DL; break;
5213 case X86::CX: DestReg = X86::CL; break;
5214 case X86::BX: DestReg = X86::BL; break;
5217 Res.first = DestReg;
5218 Res.second = Res.second = X86::GR8RegisterClass;
5220 } else if (VT == MVT::i32) {
5221 unsigned DestReg = 0;
5222 switch (Res.first) {
5224 case X86::AX: DestReg = X86::EAX; break;
5225 case X86::DX: DestReg = X86::EDX; break;
5226 case X86::CX: DestReg = X86::ECX; break;
5227 case X86::BX: DestReg = X86::EBX; break;
5228 case X86::SI: DestReg = X86::ESI; break;
5229 case X86::DI: DestReg = X86::EDI; break;
5230 case X86::BP: DestReg = X86::EBP; break;
5231 case X86::SP: DestReg = X86::ESP; break;
5234 Res.first = DestReg;
5235 Res.second = Res.second = X86::GR32RegisterClass;
5237 } else if (VT == MVT::i64) {
5238 unsigned DestReg = 0;
5239 switch (Res.first) {
5241 case X86::AX: DestReg = X86::RAX; break;
5242 case X86::DX: DestReg = X86::RDX; break;
5243 case X86::CX: DestReg = X86::RCX; break;
5244 case X86::BX: DestReg = X86::RBX; break;
5245 case X86::SI: DestReg = X86::RSI; break;
5246 case X86::DI: DestReg = X86::RDI; break;
5247 case X86::BP: DestReg = X86::RBP; break;
5248 case X86::SP: DestReg = X86::RSP; break;
5251 Res.first = DestReg;
5252 Res.second = Res.second = X86::GR64RegisterClass;