1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/ADT/VectorExtras.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/SelectionDAG.h"
33 #include "llvm/CodeGen/SSARegMap.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/StringExtras.h"
39 #include "llvm/ParameterAttributes.h"
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSEf64 = Subtarget->hasSSE2();
46 X86ScalarSSEf32 = Subtarget->hasSSE1();
47 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
50 RegInfo = TM.getRegisterInfo();
52 // Set up the TargetLowering object.
54 // X86 is weird, it always uses i8 for shift amounts and setcc results.
55 setShiftAmountType(MVT::i8);
56 setSetCCResultType(MVT::i8);
57 setSetCCResultContents(ZeroOrOneSetCCResult);
58 setSchedulingPreference(SchedulingForRegPressure);
59 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
60 setStackPointerRegisterToSaveRestore(X86StackPtr);
62 if (Subtarget->isTargetDarwin()) {
63 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
64 setUseUnderscoreSetJmp(false);
65 setUseUnderscoreLongJmp(false);
66 } else if (Subtarget->isTargetMingw()) {
67 // MS runtime is weird: it exports _setjmp, but longjmp!
68 setUseUnderscoreSetJmp(true);
69 setUseUnderscoreLongJmp(false);
71 setUseUnderscoreSetJmp(true);
72 setUseUnderscoreLongJmp(true);
75 // Set up the register classes.
76 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
77 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
78 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
79 if (Subtarget->is64Bit())
80 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
82 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
84 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
86 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
87 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
88 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
90 if (Subtarget->is64Bit()) {
91 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
92 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
95 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
96 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
103 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
104 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
105 // SSE has no i16 to fp conversion, only i32
106 if (X86ScalarSSEf32) {
107 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
108 // f32 and f64 cases are Legal, f80 case is not
109 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
111 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
115 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
116 // are Legal, f80 is custom lowered.
117 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
118 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
120 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
122 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
123 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
125 if (X86ScalarSSEf32) {
126 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
127 // f32 and f64 cases are Legal, f80 case is not
128 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
134 // Handle FP_TO_UINT by promoting the destination to a larger signed
136 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
137 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
138 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
140 if (Subtarget->is64Bit()) {
141 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
142 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
144 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
145 // Expand FP_TO_UINT into a select.
146 // FIXME: We would like to use a Custom expander here eventually to do
147 // the optimal thing for SSE vs. the default expansion in the legalizer.
148 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
150 // With SSE3 we can use fisttpll to convert to a signed i64.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
154 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
155 if (!X86ScalarSSEf64) {
156 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
157 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
160 // Scalar integer multiply, multiply-high, divide, and remainder are
161 // lowered to use operations that produce two results, to match the
162 // available instructions. This exposes the two-result form to trivial
163 // CSE, which is able to combine x/y and x%y into a single instruction,
164 // for example. The single-result multiply instructions are introduced
165 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
167 setOperationAction(ISD::MUL , MVT::i8 , Expand);
168 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
169 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
170 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
171 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
172 setOperationAction(ISD::SREM , MVT::i8 , Expand);
173 setOperationAction(ISD::UREM , MVT::i8 , Expand);
174 setOperationAction(ISD::MUL , MVT::i16 , Expand);
175 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
176 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
177 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
178 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
179 setOperationAction(ISD::SREM , MVT::i16 , Expand);
180 setOperationAction(ISD::UREM , MVT::i16 , Expand);
181 setOperationAction(ISD::MUL , MVT::i32 , Expand);
182 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
183 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
184 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
185 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
186 setOperationAction(ISD::SREM , MVT::i32 , Expand);
187 setOperationAction(ISD::UREM , MVT::i32 , Expand);
188 setOperationAction(ISD::MUL , MVT::i64 , Expand);
189 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
190 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
191 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
192 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
193 setOperationAction(ISD::SREM , MVT::i64 , Expand);
194 setOperationAction(ISD::UREM , MVT::i64 , Expand);
196 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
197 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
198 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
199 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
200 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
201 if (Subtarget->is64Bit())
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
203 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
204 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
205 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
206 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
207 setOperationAction(ISD::FREM , MVT::f64 , Expand);
208 setOperationAction(ISD::FLT_ROUNDS , MVT::i32 , Custom);
210 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
211 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
212 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
213 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
214 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
215 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
216 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
217 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
218 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
219 if (Subtarget->is64Bit()) {
220 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
221 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
222 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
225 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
226 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
228 // These should be promoted to a larger select which is supported.
229 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
230 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
231 // X86 wants to expand cmov itself.
232 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
233 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
234 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
235 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
236 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
237 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
238 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
239 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
240 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
241 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
242 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
243 if (Subtarget->is64Bit()) {
244 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
245 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
247 // X86 ret instruction may pop stack.
248 setOperationAction(ISD::RET , MVT::Other, Custom);
249 if (!Subtarget->is64Bit())
250 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
253 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
254 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
255 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
256 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
257 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
258 if (Subtarget->is64Bit()) {
259 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
260 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
261 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
262 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
264 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
265 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
266 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
267 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
268 // X86 wants to expand memset / memcpy itself.
269 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
270 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
272 // Use the default ISD::LOCATION expansion.
273 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
274 // FIXME - use subtarget debug flags
275 if (!Subtarget->isTargetDarwin() &&
276 !Subtarget->isTargetELF() &&
277 !Subtarget->isTargetCygMing())
278 setOperationAction(ISD::LABEL, MVT::Other, Expand);
280 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
281 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
282 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
283 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
284 if (Subtarget->is64Bit()) {
286 setExceptionPointerRegister(X86::RAX);
287 setExceptionSelectorRegister(X86::RDX);
289 setExceptionPointerRegister(X86::EAX);
290 setExceptionSelectorRegister(X86::EDX);
292 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
294 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
296 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
297 setOperationAction(ISD::VASTART , MVT::Other, Custom);
298 setOperationAction(ISD::VAARG , MVT::Other, Expand);
299 setOperationAction(ISD::VAEND , MVT::Other, Expand);
300 if (Subtarget->is64Bit())
301 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
303 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
305 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
306 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
307 if (Subtarget->is64Bit())
308 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
309 if (Subtarget->isTargetCygMing())
310 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
312 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
314 if (X86ScalarSSEf64) {
315 // f32 and f64 use SSE.
316 // Set up the FP register classes.
317 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
318 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
320 // Use ANDPD to simulate FABS.
321 setOperationAction(ISD::FABS , MVT::f64, Custom);
322 setOperationAction(ISD::FABS , MVT::f32, Custom);
324 // Use XORP to simulate FNEG.
325 setOperationAction(ISD::FNEG , MVT::f64, Custom);
326 setOperationAction(ISD::FNEG , MVT::f32, Custom);
328 // Use ANDPD and ORPD to simulate FCOPYSIGN.
329 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
330 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
332 // We don't support sin/cos/fmod
333 setOperationAction(ISD::FSIN , MVT::f64, Expand);
334 setOperationAction(ISD::FCOS , MVT::f64, Expand);
335 setOperationAction(ISD::FREM , MVT::f64, Expand);
336 setOperationAction(ISD::FSIN , MVT::f32, Expand);
337 setOperationAction(ISD::FCOS , MVT::f32, Expand);
338 setOperationAction(ISD::FREM , MVT::f32, Expand);
340 // Expand FP immediates into loads from the stack, except for the special
342 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
343 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
344 addLegalFPImmediate(APFloat(+0.0)); // xorpd
345 addLegalFPImmediate(APFloat(+0.0f)); // xorps
347 // Conversions to long double (in X87) go through memory.
348 setConvertAction(MVT::f32, MVT::f80, Expand);
349 setConvertAction(MVT::f64, MVT::f80, Expand);
351 // Conversions from long double (in X87) go through memory.
352 setConvertAction(MVT::f80, MVT::f32, Expand);
353 setConvertAction(MVT::f80, MVT::f64, Expand);
354 } else if (X86ScalarSSEf32) {
355 // Use SSE for f32, x87 for f64.
356 // Set up the FP register classes.
357 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
358 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
360 // Use ANDPS to simulate FABS.
361 setOperationAction(ISD::FABS , MVT::f32, Custom);
363 // Use XORP to simulate FNEG.
364 setOperationAction(ISD::FNEG , MVT::f32, Custom);
366 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
368 // Use ANDPS and ORPS to simulate FCOPYSIGN.
369 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
370 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
372 // We don't support sin/cos/fmod
373 setOperationAction(ISD::FSIN , MVT::f32, Expand);
374 setOperationAction(ISD::FCOS , MVT::f32, Expand);
375 setOperationAction(ISD::FREM , MVT::f32, Expand);
377 // Expand FP immediates into loads from the stack, except for the special
379 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
380 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
381 addLegalFPImmediate(APFloat(+0.0f)); // xorps
382 addLegalFPImmediate(APFloat(+0.0)); // FLD0
383 addLegalFPImmediate(APFloat(+1.0)); // FLD1
384 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
385 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
387 // SSE->x87 conversions go through memory.
388 setConvertAction(MVT::f32, MVT::f64, Expand);
389 setConvertAction(MVT::f32, MVT::f80, Expand);
391 // x87->SSE truncations need to go through memory.
392 setConvertAction(MVT::f80, MVT::f32, Expand);
393 setConvertAction(MVT::f64, MVT::f32, Expand);
394 // And x87->x87 truncations also.
395 setConvertAction(MVT::f80, MVT::f64, Expand);
398 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
399 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
402 // f32 and f64 in x87.
403 // Set up the FP register classes.
404 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
405 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
407 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
408 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
409 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
410 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
412 // Floating truncations need to go through memory.
413 setConvertAction(MVT::f80, MVT::f32, Expand);
414 setConvertAction(MVT::f64, MVT::f32, Expand);
415 setConvertAction(MVT::f80, MVT::f64, Expand);
418 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
419 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
422 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
423 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
424 addLegalFPImmediate(APFloat(+0.0)); // FLD0
425 addLegalFPImmediate(APFloat(+1.0)); // FLD1
426 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
427 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
428 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
429 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
430 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
431 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
434 // Long double always uses X87.
435 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
436 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
438 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
440 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
441 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
444 // Always use a library call for pow.
445 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
446 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
447 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
449 // First set operation action for all vector types to expand. Then we
450 // will selectively turn on ones that can be effectively codegen'd.
451 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
452 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
453 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
454 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
455 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
456 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
457 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
458 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
459 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
460 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
461 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
462 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
463 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
464 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
465 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
466 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
467 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
468 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
469 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
470 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
471 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
472 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
473 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
474 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
475 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
476 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
477 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
478 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
479 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
480 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
481 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
482 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
483 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
486 if (Subtarget->hasMMX()) {
487 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
488 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
489 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
490 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
492 // FIXME: add MMX packed arithmetics
494 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
495 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
496 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
497 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
499 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
500 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
501 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
502 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
504 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
505 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
507 setOperationAction(ISD::AND, MVT::v8i8, Promote);
508 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
509 setOperationAction(ISD::AND, MVT::v4i16, Promote);
510 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
511 setOperationAction(ISD::AND, MVT::v2i32, Promote);
512 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
513 setOperationAction(ISD::AND, MVT::v1i64, Legal);
515 setOperationAction(ISD::OR, MVT::v8i8, Promote);
516 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
517 setOperationAction(ISD::OR, MVT::v4i16, Promote);
518 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
519 setOperationAction(ISD::OR, MVT::v2i32, Promote);
520 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
521 setOperationAction(ISD::OR, MVT::v1i64, Legal);
523 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
524 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
525 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
526 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
527 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
528 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
529 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
531 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
532 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
533 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
534 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
535 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
536 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
537 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
539 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
540 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
541 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
542 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
544 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
545 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
546 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
547 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
549 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
550 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
551 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
552 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
555 if (Subtarget->hasSSE1()) {
556 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
558 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
559 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
560 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
561 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
562 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
563 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
564 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
565 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
566 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
567 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
568 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
571 if (Subtarget->hasSSE2()) {
572 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
573 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
574 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
575 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
576 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
578 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
579 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
580 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
581 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
582 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
583 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
584 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
585 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
586 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
587 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
588 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
589 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
590 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
591 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
592 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
594 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
595 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
596 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
597 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
598 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
599 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
601 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
602 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
603 // Do not attempt to custom lower non-power-of-2 vectors
604 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
606 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
607 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
608 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
610 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
611 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
612 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
613 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
614 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
615 if (Subtarget->is64Bit())
616 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
618 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
619 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
620 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
621 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
622 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
623 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
624 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
625 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
626 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
627 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
628 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
629 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
632 // Custom lower v2i64 and v2f64 selects.
633 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
634 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
635 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
636 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
639 // We want to custom lower some of our intrinsics.
640 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
642 // We have target-specific dag combine patterns for the following nodes:
643 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
644 setTargetDAGCombine(ISD::SELECT);
646 computeRegisterProperties();
648 // FIXME: These should be based on subtarget info. Plus, the values should
649 // be smaller when we are in optimizing for size mode.
650 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
651 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
652 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
653 allowUnalignedMemoryAccesses = true; // x86 supports it!
657 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
659 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
660 SelectionDAG &DAG) const {
661 if (usesGlobalOffsetTable())
662 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
663 if (!Subtarget->isPICStyleRIPRel())
664 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
668 //===----------------------------------------------------------------------===//
669 // Return Value Calling Convention Implementation
670 //===----------------------------------------------------------------------===//
672 #include "X86GenCallingConv.inc"
674 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
675 /// exists skip possible ISD:TokenFactor.
676 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
677 if (Chain.getOpcode()==X86ISD::TAILCALL) {
679 } else if (Chain.getOpcode()==ISD::TokenFactor) {
680 if (Chain.getNumOperands() &&
681 Chain.getOperand(0).getOpcode()==X86ISD::TAILCALL)
682 return Chain.getOperand(0);
687 /// LowerRET - Lower an ISD::RET node.
688 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
689 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
691 SmallVector<CCValAssign, 16> RVLocs;
692 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
693 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
694 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
695 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
697 // If this is the first return lowered for this function, add the regs to the
698 // liveout set for the function.
699 if (DAG.getMachineFunction().liveout_empty()) {
700 for (unsigned i = 0; i != RVLocs.size(); ++i)
701 if (RVLocs[i].isRegLoc())
702 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
704 SDOperand Chain = Op.getOperand(0);
706 // Handle tail call return.
707 Chain = GetPossiblePreceedingTailCall(Chain);
708 if (Chain.getOpcode() == X86ISD::TAILCALL) {
709 SDOperand TailCall = Chain;
710 SDOperand TargetAddress = TailCall.getOperand(1);
711 SDOperand StackAdjustment = TailCall.getOperand(2);
712 assert ( ((TargetAddress.getOpcode() == ISD::Register &&
713 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
714 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
715 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
716 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
717 "Expecting an global address, external symbol, or register");
718 assert( StackAdjustment.getOpcode() == ISD::Constant &&
719 "Expecting a const value");
721 SmallVector<SDOperand,8> Operands;
722 Operands.push_back(Chain.getOperand(0));
723 Operands.push_back(TargetAddress);
724 Operands.push_back(StackAdjustment);
725 // Copy registers used by the call. Last operand is a flag so it is not
727 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
728 Operands.push_back(Chain.getOperand(i));
730 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
737 // Copy the result values into the output registers.
738 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
739 RVLocs[0].getLocReg() != X86::ST0) {
740 for (unsigned i = 0; i != RVLocs.size(); ++i) {
741 CCValAssign &VA = RVLocs[i];
742 assert(VA.isRegLoc() && "Can only return in registers!");
743 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
745 Flag = Chain.getValue(1);
748 // We need to handle a destination of ST0 specially, because it isn't really
750 SDOperand Value = Op.getOperand(1);
752 // If this is an FP return with ScalarSSE, we need to move the value from
753 // an XMM register onto the fp-stack.
754 if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
755 (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
758 // If this is a load into a scalarsse value, don't store the loaded value
759 // back to the stack, only to reload it: just replace the scalar-sse load.
760 if (ISD::isNON_EXTLoad(Value.Val) &&
761 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
762 Chain = Value.getOperand(0);
763 MemLoc = Value.getOperand(1);
765 // Spill the value to memory and reload it into top of stack.
766 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
767 MachineFunction &MF = DAG.getMachineFunction();
768 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
769 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
770 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
772 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
773 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
774 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
775 Chain = Value.getValue(1);
778 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
779 SDOperand Ops[] = { Chain, Value };
780 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
781 Flag = Chain.getValue(1);
784 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
786 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
788 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
792 /// LowerCallResult - Lower the result values of an ISD::CALL into the
793 /// appropriate copies out of appropriate physical registers. This assumes that
794 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
795 /// being lowered. The returns a SDNode with the same number of values as the
797 SDNode *X86TargetLowering::
798 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
799 unsigned CallingConv, SelectionDAG &DAG) {
801 // Assign locations to each value returned by this call.
802 SmallVector<CCValAssign, 16> RVLocs;
803 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
804 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
805 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
808 SmallVector<SDOperand, 8> ResultVals;
810 // Copy all of the result registers out of their specified physreg.
811 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
812 for (unsigned i = 0; i != RVLocs.size(); ++i) {
813 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
814 RVLocs[i].getValVT(), InFlag).getValue(1);
815 InFlag = Chain.getValue(2);
816 ResultVals.push_back(Chain.getValue(0));
819 // Copies from the FP stack are special, as ST0 isn't a valid register
820 // before the fp stackifier runs.
822 // Copy ST0 into an RFP register with FP_GET_RESULT.
823 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
824 SDOperand GROps[] = { Chain, InFlag };
825 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
826 Chain = RetVal.getValue(1);
827 InFlag = RetVal.getValue(2);
829 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
831 if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
832 (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
833 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
834 // shouldn't be necessary except that RFP cannot be live across
835 // multiple blocks. When stackifier is fixed, they can be uncoupled.
836 MachineFunction &MF = DAG.getMachineFunction();
837 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
838 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
840 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
842 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
843 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
844 Chain = RetVal.getValue(1);
846 ResultVals.push_back(RetVal);
849 // Merge everything together with a MERGE_VALUES node.
850 ResultVals.push_back(Chain);
851 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
852 &ResultVals[0], ResultVals.size()).Val;
856 //===----------------------------------------------------------------------===//
857 // C & StdCall & Fast Calling Convention implementation
858 //===----------------------------------------------------------------------===//
859 // StdCall calling convention seems to be standard for many Windows' API
860 // routines and around. It differs from C calling convention just a little:
861 // callee should clean up the stack, not caller. Symbols should be also
862 // decorated in some fancy way :) It doesn't support any vector arguments.
863 // For info on fast calling convention see Fast Calling Convention (tail call)
864 // implementation LowerX86_32FastCCCallTo.
866 /// AddLiveIn - This helper function adds the specified physical register to the
867 /// MachineFunction as a live in value. It also creates a corresponding virtual
869 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
870 const TargetRegisterClass *RC) {
871 assert(RC->contains(PReg) && "Not the correct regclass!");
872 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
873 MF.addLiveIn(PReg, VReg);
877 // align stack arguments according to platform alignment needed for tail calls
878 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG& DAG);
880 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
881 const CCValAssign &VA,
882 MachineFrameInfo *MFI,
883 SDOperand Root, unsigned i) {
884 // Create the nodes corresponding to a load from this parameter slot.
885 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
886 VA.getLocMemOffset());
887 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
889 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
891 if (Flags & ISD::ParamFlags::ByVal)
894 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
897 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
899 unsigned NumArgs = Op.Val->getNumValues() - 1;
900 MachineFunction &MF = DAG.getMachineFunction();
901 MachineFrameInfo *MFI = MF.getFrameInfo();
902 SDOperand Root = Op.getOperand(0);
903 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
904 unsigned CC = MF.getFunction()->getCallingConv();
905 // Assign locations to all of the incoming arguments.
906 SmallVector<CCValAssign, 16> ArgLocs;
907 CCState CCInfo(CC, isVarArg,
908 getTargetMachine(), ArgLocs);
909 // Check for possible tail call calling convention.
910 if (CC == CallingConv::Fast && PerformTailCallOpt)
911 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_TailCall);
913 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
915 SmallVector<SDOperand, 8> ArgValues;
916 unsigned LastVal = ~0U;
917 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
918 CCValAssign &VA = ArgLocs[i];
919 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
921 assert(VA.getValNo() != LastVal &&
922 "Don't support value assigned to multiple locs yet");
923 LastVal = VA.getValNo();
926 MVT::ValueType RegVT = VA.getLocVT();
927 TargetRegisterClass *RC;
928 if (RegVT == MVT::i32)
929 RC = X86::GR32RegisterClass;
931 assert(MVT::isVector(RegVT));
932 RC = X86::VR128RegisterClass;
935 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
936 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
938 // If this is an 8 or 16-bit value, it is really passed promoted to 32
939 // bits. Insert an assert[sz]ext to capture this, then truncate to the
941 if (VA.getLocInfo() == CCValAssign::SExt)
942 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
943 DAG.getValueType(VA.getValVT()));
944 else if (VA.getLocInfo() == CCValAssign::ZExt)
945 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
946 DAG.getValueType(VA.getValVT()));
948 if (VA.getLocInfo() != CCValAssign::Full)
949 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
951 ArgValues.push_back(ArgValue);
953 assert(VA.isMemLoc());
954 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
958 unsigned StackSize = CCInfo.getNextStackOffset();
959 // align stack specially for tail calls
960 if (CC==CallingConv::Fast)
961 StackSize = GetAlignedArgumentStackSize(StackSize,DAG);
963 ArgValues.push_back(Root);
965 // If the function takes variable number of arguments, make a frame index for
966 // the start of the first vararg value... for expansion of llvm.va_start.
968 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
970 // Tail call calling convention (CallingConv::Fast) does not support varargs.
971 assert( !(isVarArg && CC == CallingConv::Fast) &&
972 "CallingConv::Fast does not support varargs.");
974 if (isStdCall && !isVarArg &&
975 (CC==CallingConv::Fast && PerformTailCallOpt || CC!=CallingConv::Fast)) {
976 BytesToPopOnReturn = StackSize; // Callee pops everything..
977 BytesCallerReserves = 0;
979 BytesToPopOnReturn = 0; // Callee pops nothing.
981 // If this is an sret function, the return should pop the hidden pointer.
983 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
984 ISD::ParamFlags::StructReturn))
985 BytesToPopOnReturn = 4;
987 BytesCallerReserves = StackSize;
990 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
992 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
993 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
995 // Return the new list of results.
996 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
997 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1000 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
1002 SDOperand Chain = Op.getOperand(0);
1003 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1004 SDOperand Callee = Op.getOperand(4);
1005 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1007 // Analyze operands of the call, assigning locations to each operand.
1008 SmallVector<CCValAssign, 16> ArgLocs;
1009 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1010 if(CC==CallingConv::Fast && PerformTailCallOpt)
1011 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1013 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
1015 // Get a count of how many bytes are to be pushed on the stack.
1016 unsigned NumBytes = CCInfo.getNextStackOffset();
1017 if (CC==CallingConv::Fast)
1018 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1020 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1022 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1023 SmallVector<SDOperand, 8> MemOpChains;
1027 // Walk the register/memloc assignments, inserting copies/loads.
1028 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1029 CCValAssign &VA = ArgLocs[i];
1030 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1032 // Promote the value if needed.
1033 switch (VA.getLocInfo()) {
1034 default: assert(0 && "Unknown loc info!");
1035 case CCValAssign::Full: break;
1036 case CCValAssign::SExt:
1037 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1039 case CCValAssign::ZExt:
1040 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1042 case CCValAssign::AExt:
1043 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1047 if (VA.isRegLoc()) {
1048 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1050 assert(VA.isMemLoc());
1051 if (StackPtr.Val == 0)
1052 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1054 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1059 // If the first argument is an sret pointer, remember it.
1060 bool isSRet = NumOps &&
1061 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
1062 ISD::ParamFlags::StructReturn);
1064 if (!MemOpChains.empty())
1065 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1066 &MemOpChains[0], MemOpChains.size());
1068 // Build a sequence of copy-to-reg nodes chained together with token chain
1069 // and flag operands which copy the outgoing args into registers.
1071 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1072 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1074 InFlag = Chain.getValue(1);
1077 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1079 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1080 Subtarget->isPICStyleGOT()) {
1081 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1082 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1084 InFlag = Chain.getValue(1);
1087 // If the callee is a GlobalAddress node (quite common, every direct call is)
1088 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1089 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1090 // We should use extra load for direct calls to dllimported functions in
1092 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1093 getTargetMachine(), true))
1094 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1095 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1096 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1098 // Returns a chain & a flag for retval copy to use.
1099 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1100 SmallVector<SDOperand, 8> Ops;
1101 Ops.push_back(Chain);
1102 Ops.push_back(Callee);
1104 // Add argument registers to the end of the list so that they are known live
1106 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1107 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1108 RegsToPass[i].second.getValueType()));
1110 // Add an implicit use GOT pointer in EBX.
1111 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1112 Subtarget->isPICStyleGOT())
1113 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1116 Ops.push_back(InFlag);
1118 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1119 InFlag = Chain.getValue(1);
1121 // Create the CALLSEQ_END node.
1122 unsigned NumBytesForCalleeToPush = 0;
1124 if (CC == CallingConv::X86_StdCall ||
1125 (CC == CallingConv::Fast && PerformTailCallOpt)) {
1127 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1129 NumBytesForCalleeToPush = NumBytes;
1130 assert(!(isVarArg && CC==CallingConv::Fast) &&
1131 "CallingConv::Fast does not support varargs.");
1133 // If this is is a call to a struct-return function, the callee
1134 // pops the hidden struct pointer, so we have to push it back.
1135 // This is common for Darwin/X86, Linux & Mingw32 targets.
1136 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1139 Chain = DAG.getCALLSEQ_END(Chain,
1140 DAG.getConstant(NumBytes, getPointerTy()),
1141 DAG.getConstant(NumBytesForCalleeToPush,
1144 InFlag = Chain.getValue(1);
1146 // Handle result values, copying them out of physregs into vregs that we
1148 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1152 //===----------------------------------------------------------------------===//
1153 // FastCall Calling Convention implementation
1154 //===----------------------------------------------------------------------===//
1156 // The X86 'fastcall' calling convention passes up to two integer arguments in
1157 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1158 // and requires that the callee pop its arguments off the stack (allowing proper
1159 // tail calls), and has the same return value conventions as C calling convs.
1161 // This calling convention always arranges for the callee pop value to be 8n+4
1162 // bytes, which is needed for tail recursion elimination and stack alignment
1165 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
1166 MachineFunction &MF = DAG.getMachineFunction();
1167 MachineFrameInfo *MFI = MF.getFrameInfo();
1168 SDOperand Root = Op.getOperand(0);
1169 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1171 // Assign locations to all of the incoming arguments.
1172 SmallVector<CCValAssign, 16> ArgLocs;
1173 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1174 getTargetMachine(), ArgLocs);
1175 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
1177 SmallVector<SDOperand, 8> ArgValues;
1178 unsigned LastVal = ~0U;
1179 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1180 CCValAssign &VA = ArgLocs[i];
1181 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1183 assert(VA.getValNo() != LastVal &&
1184 "Don't support value assigned to multiple locs yet");
1185 LastVal = VA.getValNo();
1187 if (VA.isRegLoc()) {
1188 MVT::ValueType RegVT = VA.getLocVT();
1189 TargetRegisterClass *RC;
1190 if (RegVT == MVT::i32)
1191 RC = X86::GR32RegisterClass;
1193 assert(MVT::isVector(RegVT));
1194 RC = X86::VR128RegisterClass;
1197 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1198 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1200 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1201 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1203 if (VA.getLocInfo() == CCValAssign::SExt)
1204 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1205 DAG.getValueType(VA.getValVT()));
1206 else if (VA.getLocInfo() == CCValAssign::ZExt)
1207 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1208 DAG.getValueType(VA.getValVT()));
1210 if (VA.getLocInfo() != CCValAssign::Full)
1211 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1213 ArgValues.push_back(ArgValue);
1215 assert(VA.isMemLoc());
1216 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1220 ArgValues.push_back(Root);
1222 unsigned StackSize = CCInfo.getNextStackOffset();
1224 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1225 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1226 // arguments and the arguments after the retaddr has been pushed are
1228 if ((StackSize & 7) == 0)
1232 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1233 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1234 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1235 BytesCallerReserves = 0;
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1238 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1240 // Return the new list of results.
1241 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1242 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1246 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1247 const SDOperand &StackPtr,
1248 const CCValAssign &VA,
1251 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1252 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1253 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1254 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1255 if (Flags & ISD::ParamFlags::ByVal) {
1256 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1257 ISD::ParamFlags::ByValAlignOffs);
1259 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1260 ISD::ParamFlags::ByValSizeOffs;
1262 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1263 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1264 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1266 return DAG.getMemcpy(Chain, PtrOff, Arg, SizeNode, AlignNode,
1269 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1273 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1275 SDOperand Chain = Op.getOperand(0);
1276 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1277 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1278 SDOperand Callee = Op.getOperand(4);
1280 // Analyze operands of the call, assigning locations to each operand.
1281 SmallVector<CCValAssign, 16> ArgLocs;
1282 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1283 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1285 // Get a count of how many bytes are to be pushed on the stack.
1286 unsigned NumBytes = CCInfo.getNextStackOffset();
1288 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1289 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1290 // arguments and the arguments after the retaddr has been pushed are
1292 if ((NumBytes & 7) == 0)
1296 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1298 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1299 SmallVector<SDOperand, 8> MemOpChains;
1303 // Walk the register/memloc assignments, inserting copies/loads.
1304 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1305 CCValAssign &VA = ArgLocs[i];
1306 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1308 // Promote the value if needed.
1309 switch (VA.getLocInfo()) {
1310 default: assert(0 && "Unknown loc info!");
1311 case CCValAssign::Full: break;
1312 case CCValAssign::SExt:
1313 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1315 case CCValAssign::ZExt:
1316 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1318 case CCValAssign::AExt:
1319 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1323 if (VA.isRegLoc()) {
1324 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1326 assert(VA.isMemLoc());
1327 if (StackPtr.Val == 0)
1328 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1330 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1335 if (!MemOpChains.empty())
1336 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1337 &MemOpChains[0], MemOpChains.size());
1339 // Build a sequence of copy-to-reg nodes chained together with token chain
1340 // and flag operands which copy the outgoing args into registers.
1342 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1343 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1345 InFlag = Chain.getValue(1);
1348 // If the callee is a GlobalAddress node (quite common, every direct call is)
1349 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1350 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1351 // We should use extra load for direct calls to dllimported functions in
1353 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1354 getTargetMachine(), true))
1355 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1356 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1357 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1359 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1361 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1362 Subtarget->isPICStyleGOT()) {
1363 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1364 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1366 InFlag = Chain.getValue(1);
1369 // Returns a chain & a flag for retval copy to use.
1370 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1371 SmallVector<SDOperand, 8> Ops;
1372 Ops.push_back(Chain);
1373 Ops.push_back(Callee);
1375 // Add argument registers to the end of the list so that they are known live
1377 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1378 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1379 RegsToPass[i].second.getValueType()));
1381 // Add an implicit use GOT pointer in EBX.
1382 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1383 Subtarget->isPICStyleGOT())
1384 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1387 Ops.push_back(InFlag);
1389 assert(isTailCall==false && "no tail call here");
1390 Chain = DAG.getNode(X86ISD::CALL,
1391 NodeTys, &Ops[0], Ops.size());
1392 InFlag = Chain.getValue(1);
1394 // Returns a flag for retval copy to use.
1395 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1397 Ops.push_back(Chain);
1398 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1399 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1400 Ops.push_back(InFlag);
1401 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1402 InFlag = Chain.getValue(1);
1404 // Handle result values, copying them out of physregs into vregs that we
1406 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1409 //===----------------------------------------------------------------------===//
1410 // Fast Calling Convention (tail call) implementation
1411 //===----------------------------------------------------------------------===//
1413 // Like std call, callee cleans arguments, convention except that ECX is
1414 // reserved for storing the tail called function address. Only 2 registers are
1415 // free for argument passing (inreg). Tail call optimization is performed
1417 // * tailcallopt is enabled
1418 // * caller/callee are fastcc
1419 // * elf/pic is disabled OR
1420 // * elf/pic enabled + callee is in module + callee has
1421 // visibility protected or hidden
1422 // To keep the stack aligned according to platform abi the function
1423 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1424 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1425 // If a tail called function callee has more arguments than the caller the
1426 // caller needs to make sure that there is room to move the RETADDR to. This is
1427 // achieved by reserving an area the size of the argument delta right after the
1428 // original REtADDR, but before the saved framepointer or the spilled registers
1429 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1441 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1442 /// for a 16 byte align requirement.
1443 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1444 SelectionDAG& DAG) {
1445 if (PerformTailCallOpt) {
1446 MachineFunction &MF = DAG.getMachineFunction();
1447 const TargetMachine &TM = MF.getTarget();
1448 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1449 unsigned StackAlignment = TFI.getStackAlignment();
1450 uint64_t AlignMask = StackAlignment - 1;
1451 int64_t Offset = StackSize;
1452 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1453 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1454 // Number smaller than 12 so just add the difference.
1455 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1457 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1458 Offset = ((~AlignMask) & Offset) + StackAlignment +
1459 (StackAlignment-SlotSize);
1466 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1467 /// following the call is a return. A function is eligible if caller/callee
1468 /// calling conventions match, currently only fastcc supports tail calls, and
1469 /// the function CALL is immediatly followed by a RET.
1470 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1472 SelectionDAG& DAG) const {
1473 if (!PerformTailCallOpt)
1476 // Check whether CALL node immediatly preceeds the RET node and whether the
1477 // return uses the result of the node or is a void return.
1478 unsigned NumOps = Ret.getNumOperands();
1480 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1481 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1483 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1484 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1485 MachineFunction &MF = DAG.getMachineFunction();
1486 unsigned CallerCC = MF.getFunction()->getCallingConv();
1487 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1488 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1489 SDOperand Callee = Call.getOperand(4);
1490 // On elf/pic %ebx needs to be livein.
1491 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1492 !Subtarget->isPICStyleGOT())
1495 // Can only do local tail calls with PIC.
1496 GlobalValue * GV = 0;
1497 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1499 (GV = G->getGlobal()) &&
1500 (GV->hasHiddenVisibility() || GV->hasProtectedVisibility()))
1508 SDOperand X86TargetLowering::LowerX86_TailCallTo(SDOperand Op,
1511 SDOperand Chain = Op.getOperand(0);
1512 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1513 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1514 SDOperand Callee = Op.getOperand(4);
1515 bool is64Bit = Subtarget->is64Bit();
1517 assert(isTailCall && PerformTailCallOpt && "Should only emit tail calls.");
1519 // Analyze operands of the call, assigning locations to each operand.
1520 SmallVector<CCValAssign, 16> ArgLocs;
1521 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1523 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1525 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1528 // Lower arguments at fp - stackoffset + fpdiff.
1529 MachineFunction &MF = DAG.getMachineFunction();
1531 unsigned NumBytesToBePushed =
1532 GetAlignedArgumentStackSize(CCInfo.getNextStackOffset(), DAG);
1534 unsigned NumBytesCallerPushed =
1535 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1536 int FPDiff = NumBytesCallerPushed - NumBytesToBePushed;
1538 // Set the delta of movement of the returnaddr stackslot.
1539 // But only set if delta is greater than previous delta.
1540 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1541 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1544 getCALLSEQ_START(Chain, DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1546 // Adjust the Return address stack slot.
1547 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1549 MVT::ValueType VT = is64Bit ? MVT::i64 : MVT::i32;
1550 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1551 // Load the "old" Return address.
1553 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1554 // Calculate the new stack slot for the return address.
1555 int SlotSize = is64Bit ? 8 : 4;
1556 int NewReturnAddrFI =
1557 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1558 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1559 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1562 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1563 SmallVector<SDOperand, 8> MemOpChains;
1564 SmallVector<SDOperand, 8> MemOpChains2;
1565 SDOperand FramePtr, StackPtr;
1570 // Walk the register/memloc assignments, inserting copies/loads. Lower
1571 // arguments first to the stack slot where they would normally - in case of a
1572 // normal function call - be.
1573 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1574 CCValAssign &VA = ArgLocs[i];
1575 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1577 // Promote the value if needed.
1578 switch (VA.getLocInfo()) {
1579 default: assert(0 && "Unknown loc info!");
1580 case CCValAssign::Full: break;
1581 case CCValAssign::SExt:
1582 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1584 case CCValAssign::ZExt:
1585 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1587 case CCValAssign::AExt:
1588 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1592 if (VA.isRegLoc()) {
1593 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1595 assert(VA.isMemLoc());
1596 if (StackPtr.Val == 0)
1597 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1599 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1604 if (!MemOpChains.empty())
1605 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1606 &MemOpChains[0], MemOpChains.size());
1608 // Build a sequence of copy-to-reg nodes chained together with token chain
1609 // and flag operands which copy the outgoing args into registers.
1611 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1612 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1614 InFlag = Chain.getValue(1);
1616 InFlag = SDOperand();
1618 // Copy from stack slots to stack slot of a tail called function. This needs
1619 // to be done because if we would lower the arguments directly to their real
1620 // stack slot we might end up overwriting each other.
1621 // TODO: To make this more efficient (sometimes saving a store/load) we could
1622 // analyse the arguments and emit this store/load/store sequence only for
1623 // arguments which would be overwritten otherwise.
1624 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1625 CCValAssign &VA = ArgLocs[i];
1626 if (!VA.isRegLoc()) {
1627 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1628 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1630 // Get source stack slot.
1631 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1632 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1633 // Create frame index.
1634 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1635 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1636 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1637 FIN = DAG.getFrameIndex(FI, MVT::i32);
1638 if (Flags & ISD::ParamFlags::ByVal) {
1639 // Copy relative to framepointer.
1640 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1641 ISD::ParamFlags::ByValAlignOffs);
1643 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1644 ISD::ParamFlags::ByValSizeOffs;
1646 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1647 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1648 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i1);
1650 MemOpChains2.push_back(DAG.getMemcpy(Chain, FIN, PtrOff, SizeNode,
1651 AlignNode,AlwaysInline));
1653 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1654 // Store relative to framepointer.
1655 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1660 if (!MemOpChains2.empty())
1661 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1662 &MemOpChains2[0], MemOpChains.size());
1664 // Store the return address to the appropriate stack slot.
1666 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1668 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1670 // Does not work with tail call since ebx is not restored correctly by
1671 // tailcaller. TODO: at least for x86 - verify for x86-64
1673 // If the callee is a GlobalAddress node (quite common, every direct call is)
1674 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1675 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1676 // We should use extra load for direct calls to dllimported functions in
1678 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1679 getTargetMachine(), true))
1680 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1681 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1682 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1684 assert(Callee.getOpcode() == ISD::LOAD &&
1685 "Function destination must be loaded into virtual register");
1686 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1688 Chain = DAG.getCopyToReg(Chain,
1689 DAG.getRegister(Opc, getPointerTy()) ,
1691 Callee = DAG.getRegister(Opc, getPointerTy());
1692 // Add register as live out.
1693 DAG.getMachineFunction().addLiveOut(Opc);
1696 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1697 SmallVector<SDOperand, 8> Ops;
1699 Ops.push_back(Chain);
1700 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1701 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1703 Ops.push_back(InFlag);
1704 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1705 InFlag = Chain.getValue(1);
1707 // Returns a chain & a flag for retval copy to use.
1708 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1710 Ops.push_back(Chain);
1711 Ops.push_back(Callee);
1712 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1713 // Add argument registers to the end of the list so that they are known live
1715 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1716 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1717 RegsToPass[i].second.getValueType()));
1719 Ops.push_back(InFlag);
1720 assert(InFlag.Val &&
1721 "Flag must be set. Depend on flag being set in LowerRET");
1722 Chain = DAG.getNode(X86ISD::TAILCALL,
1723 Op.Val->getVTList(), &Ops[0], Ops.size());
1725 return SDOperand(Chain.Val, Op.ResNo);
1728 //===----------------------------------------------------------------------===//
1729 // X86-64 C Calling Convention implementation
1730 //===----------------------------------------------------------------------===//
1733 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1734 MachineFunction &MF = DAG.getMachineFunction();
1735 MachineFrameInfo *MFI = MF.getFrameInfo();
1736 SDOperand Root = Op.getOperand(0);
1737 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1738 unsigned CC= MF.getFunction()->getCallingConv();
1740 static const unsigned GPR64ArgRegs[] = {
1741 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1743 static const unsigned XMMArgRegs[] = {
1744 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1745 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1749 // Assign locations to all of the incoming arguments.
1750 SmallVector<CCValAssign, 16> ArgLocs;
1751 CCState CCInfo(CC, isVarArg,
1752 getTargetMachine(), ArgLocs);
1753 if (CC == CallingConv::Fast && PerformTailCallOpt)
1754 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1756 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1758 SmallVector<SDOperand, 8> ArgValues;
1759 unsigned LastVal = ~0U;
1760 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1761 CCValAssign &VA = ArgLocs[i];
1762 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1764 assert(VA.getValNo() != LastVal &&
1765 "Don't support value assigned to multiple locs yet");
1766 LastVal = VA.getValNo();
1768 if (VA.isRegLoc()) {
1769 MVT::ValueType RegVT = VA.getLocVT();
1770 TargetRegisterClass *RC;
1771 if (RegVT == MVT::i32)
1772 RC = X86::GR32RegisterClass;
1773 else if (RegVT == MVT::i64)
1774 RC = X86::GR64RegisterClass;
1775 else if (RegVT == MVT::f32)
1776 RC = X86::FR32RegisterClass;
1777 else if (RegVT == MVT::f64)
1778 RC = X86::FR64RegisterClass;
1780 assert(MVT::isVector(RegVT));
1781 if (MVT::getSizeInBits(RegVT) == 64) {
1782 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1785 RC = X86::VR128RegisterClass;
1788 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1789 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1791 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1792 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1794 if (VA.getLocInfo() == CCValAssign::SExt)
1795 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1796 DAG.getValueType(VA.getValVT()));
1797 else if (VA.getLocInfo() == CCValAssign::ZExt)
1798 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1799 DAG.getValueType(VA.getValVT()));
1801 if (VA.getLocInfo() != CCValAssign::Full)
1802 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1804 // Handle MMX values passed in GPRs.
1805 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1806 MVT::getSizeInBits(RegVT) == 64)
1807 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1809 ArgValues.push_back(ArgValue);
1811 assert(VA.isMemLoc());
1812 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1816 unsigned StackSize = CCInfo.getNextStackOffset();
1817 if (CC==CallingConv::Fast)
1818 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1820 // If the function takes variable number of arguments, make a frame index for
1821 // the start of the first vararg value... for expansion of llvm.va_start.
1823 assert(CC!=CallingConv::Fast
1824 && "Var arg not supported with calling convention fastcc");
1825 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1826 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1828 // For X86-64, if there are vararg parameters that are passed via
1829 // registers, then we must store them to their spots on the stack so they
1830 // may be loaded by deferencing the result of va_next.
1831 VarArgsGPOffset = NumIntRegs * 8;
1832 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1833 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1834 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1836 // Store the integer parameter registers.
1837 SmallVector<SDOperand, 8> MemOps;
1838 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1839 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1840 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1841 for (; NumIntRegs != 6; ++NumIntRegs) {
1842 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1843 X86::GR64RegisterClass);
1844 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1845 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1846 MemOps.push_back(Store);
1847 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1848 DAG.getConstant(8, getPointerTy()));
1851 // Now store the XMM (fp + vector) parameter registers.
1852 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1853 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1854 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1855 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1856 X86::VR128RegisterClass);
1857 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1858 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1859 MemOps.push_back(Store);
1860 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1861 DAG.getConstant(16, getPointerTy()));
1863 if (!MemOps.empty())
1864 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1865 &MemOps[0], MemOps.size());
1868 ArgValues.push_back(Root);
1869 // Tail call convention (fastcc) needs callee pop.
1870 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1871 BytesToPopOnReturn = StackSize; // Callee pops everything.
1872 BytesCallerReserves = 0;
1874 BytesToPopOnReturn = 0; // Callee pops nothing.
1875 BytesCallerReserves = StackSize;
1877 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1878 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1880 // Return the new list of results.
1881 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1882 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1886 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1888 SDOperand Chain = Op.getOperand(0);
1889 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1890 SDOperand Callee = Op.getOperand(4);
1892 // Analyze operands of the call, assigning locations to each operand.
1893 SmallVector<CCValAssign, 16> ArgLocs;
1894 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1895 if (CC==CallingConv::Fast && PerformTailCallOpt)
1896 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1898 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1900 // Get a count of how many bytes are to be pushed on the stack.
1901 unsigned NumBytes = CCInfo.getNextStackOffset();
1902 if (CC == CallingConv::Fast)
1903 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1905 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1907 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1908 SmallVector<SDOperand, 8> MemOpChains;
1912 // Walk the register/memloc assignments, inserting copies/loads.
1913 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1914 CCValAssign &VA = ArgLocs[i];
1915 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1917 // Promote the value if needed.
1918 switch (VA.getLocInfo()) {
1919 default: assert(0 && "Unknown loc info!");
1920 case CCValAssign::Full: break;
1921 case CCValAssign::SExt:
1922 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1924 case CCValAssign::ZExt:
1925 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1927 case CCValAssign::AExt:
1928 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1932 if (VA.isRegLoc()) {
1933 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1935 assert(VA.isMemLoc());
1936 if (StackPtr.Val == 0)
1937 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1939 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1944 if (!MemOpChains.empty())
1945 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1946 &MemOpChains[0], MemOpChains.size());
1948 // Build a sequence of copy-to-reg nodes chained together with token chain
1949 // and flag operands which copy the outgoing args into registers.
1951 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1952 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1954 InFlag = Chain.getValue(1);
1958 assert ( CallingConv::Fast != CC &&
1959 "Var args not supported with calling convention fastcc");
1961 // From AMD64 ABI document:
1962 // For calls that may call functions that use varargs or stdargs
1963 // (prototype-less calls or calls to functions containing ellipsis (...) in
1964 // the declaration) %al is used as hidden argument to specify the number
1965 // of SSE registers used. The contents of %al do not need to match exactly
1966 // the number of registers, but must be an ubound on the number of SSE
1967 // registers used and is in the range 0 - 8 inclusive.
1969 // Count the number of XMM registers allocated.
1970 static const unsigned XMMArgRegs[] = {
1971 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1972 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1974 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1976 Chain = DAG.getCopyToReg(Chain, X86::AL,
1977 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1978 InFlag = Chain.getValue(1);
1981 // If the callee is a GlobalAddress node (quite common, every direct call is)
1982 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1983 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1984 // We should use extra load for direct calls to dllimported functions in
1986 if (getTargetMachine().getCodeModel() != CodeModel::Large
1987 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1988 getTargetMachine(), true))
1989 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1990 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1991 if (getTargetMachine().getCodeModel() != CodeModel::Large)
1992 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1994 // Returns a chain & a flag for retval copy to use.
1995 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1996 SmallVector<SDOperand, 8> Ops;
1997 Ops.push_back(Chain);
1998 Ops.push_back(Callee);
2000 // Add argument registers to the end of the list so that they are known live
2002 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2003 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2004 RegsToPass[i].second.getValueType()));
2007 Ops.push_back(InFlag);
2009 Chain = DAG.getNode(X86ISD::CALL,
2010 NodeTys, &Ops[0], Ops.size());
2011 InFlag = Chain.getValue(1);
2012 int NumBytesForCalleeToPush = 0;
2013 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2014 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2016 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2018 // Returns a flag for retval copy to use.
2019 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2021 Ops.push_back(Chain);
2022 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2023 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2024 Ops.push_back(InFlag);
2025 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2026 InFlag = Chain.getValue(1);
2028 // Handle result values, copying them out of physregs into vregs that we
2030 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2034 //===----------------------------------------------------------------------===//
2035 // Other Lowering Hooks
2036 //===----------------------------------------------------------------------===//
2039 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2040 MachineFunction &MF = DAG.getMachineFunction();
2041 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2042 int ReturnAddrIndex = FuncInfo->getRAIndex();
2044 if (ReturnAddrIndex == 0) {
2045 // Set up a frame object for the return address.
2046 if (Subtarget->is64Bit())
2047 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2049 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2051 FuncInfo->setRAIndex(ReturnAddrIndex);
2054 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2059 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2060 /// specific condition code. It returns a false if it cannot do a direct
2061 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2063 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2064 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2065 SelectionDAG &DAG) {
2066 X86CC = X86::COND_INVALID;
2068 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2069 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2070 // X > -1 -> X == 0, jump !sign.
2071 RHS = DAG.getConstant(0, RHS.getValueType());
2072 X86CC = X86::COND_NS;
2074 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2075 // X < 0 -> X == 0, jump on sign.
2076 X86CC = X86::COND_S;
2078 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2080 RHS = DAG.getConstant(0, RHS.getValueType());
2081 X86CC = X86::COND_LE;
2086 switch (SetCCOpcode) {
2088 case ISD::SETEQ: X86CC = X86::COND_E; break;
2089 case ISD::SETGT: X86CC = X86::COND_G; break;
2090 case ISD::SETGE: X86CC = X86::COND_GE; break;
2091 case ISD::SETLT: X86CC = X86::COND_L; break;
2092 case ISD::SETLE: X86CC = X86::COND_LE; break;
2093 case ISD::SETNE: X86CC = X86::COND_NE; break;
2094 case ISD::SETULT: X86CC = X86::COND_B; break;
2095 case ISD::SETUGT: X86CC = X86::COND_A; break;
2096 case ISD::SETULE: X86CC = X86::COND_BE; break;
2097 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2100 // On a floating point condition, the flags are set as follows:
2102 // 0 | 0 | 0 | X > Y
2103 // 0 | 0 | 1 | X < Y
2104 // 1 | 0 | 0 | X == Y
2105 // 1 | 1 | 1 | unordered
2107 switch (SetCCOpcode) {
2110 case ISD::SETEQ: X86CC = X86::COND_E; break;
2111 case ISD::SETOLT: Flip = true; // Fallthrough
2113 case ISD::SETGT: X86CC = X86::COND_A; break;
2114 case ISD::SETOLE: Flip = true; // Fallthrough
2116 case ISD::SETGE: X86CC = X86::COND_AE; break;
2117 case ISD::SETUGT: Flip = true; // Fallthrough
2119 case ISD::SETLT: X86CC = X86::COND_B; break;
2120 case ISD::SETUGE: Flip = true; // Fallthrough
2122 case ISD::SETLE: X86CC = X86::COND_BE; break;
2124 case ISD::SETNE: X86CC = X86::COND_NE; break;
2125 case ISD::SETUO: X86CC = X86::COND_P; break;
2126 case ISD::SETO: X86CC = X86::COND_NP; break;
2129 std::swap(LHS, RHS);
2132 return X86CC != X86::COND_INVALID;
2135 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2136 /// code. Current x86 isa includes the following FP cmov instructions:
2137 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2138 static bool hasFPCMov(unsigned X86CC) {
2154 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2155 /// true if Op is undef or if its value falls within the specified range (L, H].
2156 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2157 if (Op.getOpcode() == ISD::UNDEF)
2160 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2161 return (Val >= Low && Val < Hi);
2164 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2165 /// true if Op is undef or if its value equal to the specified value.
2166 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2167 if (Op.getOpcode() == ISD::UNDEF)
2169 return cast<ConstantSDNode>(Op)->getValue() == Val;
2172 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2173 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2174 bool X86::isPSHUFDMask(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2177 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2180 // Check if the value doesn't reference the second vector.
2181 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2182 SDOperand Arg = N->getOperand(i);
2183 if (Arg.getOpcode() == ISD::UNDEF) continue;
2184 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2185 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2192 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2193 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2194 bool X86::isPSHUFHWMask(SDNode *N) {
2195 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2197 if (N->getNumOperands() != 8)
2200 // Lower quadword copied in order.
2201 for (unsigned i = 0; i != 4; ++i) {
2202 SDOperand Arg = N->getOperand(i);
2203 if (Arg.getOpcode() == ISD::UNDEF) continue;
2204 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2205 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2209 // Upper quadword shuffled.
2210 for (unsigned i = 4; i != 8; ++i) {
2211 SDOperand Arg = N->getOperand(i);
2212 if (Arg.getOpcode() == ISD::UNDEF) continue;
2213 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2214 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2215 if (Val < 4 || Val > 7)
2222 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2223 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2224 bool X86::isPSHUFLWMask(SDNode *N) {
2225 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2227 if (N->getNumOperands() != 8)
2230 // Upper quadword copied in order.
2231 for (unsigned i = 4; i != 8; ++i)
2232 if (!isUndefOrEqual(N->getOperand(i), i))
2235 // Lower quadword shuffled.
2236 for (unsigned i = 0; i != 4; ++i)
2237 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2243 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2244 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2245 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2246 if (NumElems != 2 && NumElems != 4) return false;
2248 unsigned Half = NumElems / 2;
2249 for (unsigned i = 0; i < Half; ++i)
2250 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2252 for (unsigned i = Half; i < NumElems; ++i)
2253 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2259 bool X86::isSHUFPMask(SDNode *N) {
2260 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2261 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2264 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2265 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2266 /// half elements to come from vector 1 (which would equal the dest.) and
2267 /// the upper half to come from vector 2.
2268 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2269 if (NumOps != 2 && NumOps != 4) return false;
2271 unsigned Half = NumOps / 2;
2272 for (unsigned i = 0; i < Half; ++i)
2273 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2275 for (unsigned i = Half; i < NumOps; ++i)
2276 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2281 static bool isCommutedSHUFP(SDNode *N) {
2282 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2283 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2286 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2287 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2288 bool X86::isMOVHLPSMask(SDNode *N) {
2289 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2291 if (N->getNumOperands() != 4)
2294 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2295 return isUndefOrEqual(N->getOperand(0), 6) &&
2296 isUndefOrEqual(N->getOperand(1), 7) &&
2297 isUndefOrEqual(N->getOperand(2), 2) &&
2298 isUndefOrEqual(N->getOperand(3), 3);
2301 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2302 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2304 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2305 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2307 if (N->getNumOperands() != 4)
2310 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2311 return isUndefOrEqual(N->getOperand(0), 2) &&
2312 isUndefOrEqual(N->getOperand(1), 3) &&
2313 isUndefOrEqual(N->getOperand(2), 2) &&
2314 isUndefOrEqual(N->getOperand(3), 3);
2317 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2318 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2319 bool X86::isMOVLPMask(SDNode *N) {
2320 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2322 unsigned NumElems = N->getNumOperands();
2323 if (NumElems != 2 && NumElems != 4)
2326 for (unsigned i = 0; i < NumElems/2; ++i)
2327 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2330 for (unsigned i = NumElems/2; i < NumElems; ++i)
2331 if (!isUndefOrEqual(N->getOperand(i), i))
2337 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2338 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2340 bool X86::isMOVHPMask(SDNode *N) {
2341 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2343 unsigned NumElems = N->getNumOperands();
2344 if (NumElems != 2 && NumElems != 4)
2347 for (unsigned i = 0; i < NumElems/2; ++i)
2348 if (!isUndefOrEqual(N->getOperand(i), i))
2351 for (unsigned i = 0; i < NumElems/2; ++i) {
2352 SDOperand Arg = N->getOperand(i + NumElems/2);
2353 if (!isUndefOrEqual(Arg, i + NumElems))
2360 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2361 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2362 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2363 bool V2IsSplat = false) {
2364 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2367 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2368 SDOperand BitI = Elts[i];
2369 SDOperand BitI1 = Elts[i+1];
2370 if (!isUndefOrEqual(BitI, j))
2373 if (isUndefOrEqual(BitI1, NumElts))
2376 if (!isUndefOrEqual(BitI1, j + NumElts))
2384 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2385 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2386 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2389 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2390 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2391 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2392 bool V2IsSplat = false) {
2393 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2396 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2397 SDOperand BitI = Elts[i];
2398 SDOperand BitI1 = Elts[i+1];
2399 if (!isUndefOrEqual(BitI, j + NumElts/2))
2402 if (isUndefOrEqual(BitI1, NumElts))
2405 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2413 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2414 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2415 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2418 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2419 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2421 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2422 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2424 unsigned NumElems = N->getNumOperands();
2425 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2428 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2429 SDOperand BitI = N->getOperand(i);
2430 SDOperand BitI1 = N->getOperand(i+1);
2432 if (!isUndefOrEqual(BitI, j))
2434 if (!isUndefOrEqual(BitI1, j))
2441 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2442 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2444 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2445 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2447 unsigned NumElems = N->getNumOperands();
2448 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2451 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2452 SDOperand BitI = N->getOperand(i);
2453 SDOperand BitI1 = N->getOperand(i + 1);
2455 if (!isUndefOrEqual(BitI, j))
2457 if (!isUndefOrEqual(BitI1, j))
2464 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2465 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2466 /// MOVSD, and MOVD, i.e. setting the lowest element.
2467 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2468 if (NumElts != 2 && NumElts != 4)
2471 if (!isUndefOrEqual(Elts[0], NumElts))
2474 for (unsigned i = 1; i < NumElts; ++i) {
2475 if (!isUndefOrEqual(Elts[i], i))
2482 bool X86::isMOVLMask(SDNode *N) {
2483 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2484 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2487 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2488 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2489 /// element of vector 2 and the other elements to come from vector 1 in order.
2490 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2491 bool V2IsSplat = false,
2492 bool V2IsUndef = false) {
2493 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2496 if (!isUndefOrEqual(Ops[0], 0))
2499 for (unsigned i = 1; i < NumOps; ++i) {
2500 SDOperand Arg = Ops[i];
2501 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2502 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2503 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2510 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2511 bool V2IsUndef = false) {
2512 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2513 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2514 V2IsSplat, V2IsUndef);
2517 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2518 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2519 bool X86::isMOVSHDUPMask(SDNode *N) {
2520 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2522 if (N->getNumOperands() != 4)
2525 // Expect 1, 1, 3, 3
2526 for (unsigned i = 0; i < 2; ++i) {
2527 SDOperand Arg = N->getOperand(i);
2528 if (Arg.getOpcode() == ISD::UNDEF) continue;
2529 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2530 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2531 if (Val != 1) return false;
2535 for (unsigned i = 2; i < 4; ++i) {
2536 SDOperand Arg = N->getOperand(i);
2537 if (Arg.getOpcode() == ISD::UNDEF) continue;
2538 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2539 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2540 if (Val != 3) return false;
2544 // Don't use movshdup if it can be done with a shufps.
2548 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2549 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2550 bool X86::isMOVSLDUPMask(SDNode *N) {
2551 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2553 if (N->getNumOperands() != 4)
2556 // Expect 0, 0, 2, 2
2557 for (unsigned i = 0; i < 2; ++i) {
2558 SDOperand Arg = N->getOperand(i);
2559 if (Arg.getOpcode() == ISD::UNDEF) continue;
2560 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2561 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2562 if (Val != 0) return false;
2566 for (unsigned i = 2; i < 4; ++i) {
2567 SDOperand Arg = N->getOperand(i);
2568 if (Arg.getOpcode() == ISD::UNDEF) continue;
2569 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2570 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2571 if (Val != 2) return false;
2575 // Don't use movshdup if it can be done with a shufps.
2579 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2580 /// specifies a identity operation on the LHS or RHS.
2581 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2582 unsigned NumElems = N->getNumOperands();
2583 for (unsigned i = 0; i < NumElems; ++i)
2584 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2589 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2590 /// a splat of a single element.
2591 static bool isSplatMask(SDNode *N) {
2592 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2594 // This is a splat operation if each element of the permute is the same, and
2595 // if the value doesn't reference the second vector.
2596 unsigned NumElems = N->getNumOperands();
2597 SDOperand ElementBase;
2599 for (; i != NumElems; ++i) {
2600 SDOperand Elt = N->getOperand(i);
2601 if (isa<ConstantSDNode>(Elt)) {
2607 if (!ElementBase.Val)
2610 for (; i != NumElems; ++i) {
2611 SDOperand Arg = N->getOperand(i);
2612 if (Arg.getOpcode() == ISD::UNDEF) continue;
2613 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2614 if (Arg != ElementBase) return false;
2617 // Make sure it is a splat of the first vector operand.
2618 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2621 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2622 /// a splat of a single element and it's a 2 or 4 element mask.
2623 bool X86::isSplatMask(SDNode *N) {
2624 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2626 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2627 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2629 return ::isSplatMask(N);
2632 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2633 /// specifies a splat of zero element.
2634 bool X86::isSplatLoMask(SDNode *N) {
2635 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2637 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2638 if (!isUndefOrEqual(N->getOperand(i), 0))
2643 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2644 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2646 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2647 unsigned NumOperands = N->getNumOperands();
2648 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2650 for (unsigned i = 0; i < NumOperands; ++i) {
2652 SDOperand Arg = N->getOperand(NumOperands-i-1);
2653 if (Arg.getOpcode() != ISD::UNDEF)
2654 Val = cast<ConstantSDNode>(Arg)->getValue();
2655 if (Val >= NumOperands) Val -= NumOperands;
2657 if (i != NumOperands - 1)
2664 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2665 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2667 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2669 // 8 nodes, but we only care about the last 4.
2670 for (unsigned i = 7; i >= 4; --i) {
2672 SDOperand Arg = N->getOperand(i);
2673 if (Arg.getOpcode() != ISD::UNDEF)
2674 Val = cast<ConstantSDNode>(Arg)->getValue();
2683 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2684 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2686 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2688 // 8 nodes, but we only care about the first 4.
2689 for (int i = 3; i >= 0; --i) {
2691 SDOperand Arg = N->getOperand(i);
2692 if (Arg.getOpcode() != ISD::UNDEF)
2693 Val = cast<ConstantSDNode>(Arg)->getValue();
2702 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2703 /// specifies a 8 element shuffle that can be broken into a pair of
2704 /// PSHUFHW and PSHUFLW.
2705 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2706 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2708 if (N->getNumOperands() != 8)
2711 // Lower quadword shuffled.
2712 for (unsigned i = 0; i != 4; ++i) {
2713 SDOperand Arg = N->getOperand(i);
2714 if (Arg.getOpcode() == ISD::UNDEF) continue;
2715 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2716 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2721 // Upper quadword shuffled.
2722 for (unsigned i = 4; i != 8; ++i) {
2723 SDOperand Arg = N->getOperand(i);
2724 if (Arg.getOpcode() == ISD::UNDEF) continue;
2725 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2726 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2727 if (Val < 4 || Val > 7)
2734 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2735 /// values in ther permute mask.
2736 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2737 SDOperand &V2, SDOperand &Mask,
2738 SelectionDAG &DAG) {
2739 MVT::ValueType VT = Op.getValueType();
2740 MVT::ValueType MaskVT = Mask.getValueType();
2741 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2742 unsigned NumElems = Mask.getNumOperands();
2743 SmallVector<SDOperand, 8> MaskVec;
2745 for (unsigned i = 0; i != NumElems; ++i) {
2746 SDOperand Arg = Mask.getOperand(i);
2747 if (Arg.getOpcode() == ISD::UNDEF) {
2748 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2751 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2752 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2754 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2756 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2760 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2761 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2764 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2765 /// the two vector operands have swapped position.
2767 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2768 MVT::ValueType MaskVT = Mask.getValueType();
2769 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2770 unsigned NumElems = Mask.getNumOperands();
2771 SmallVector<SDOperand, 8> MaskVec;
2772 for (unsigned i = 0; i != NumElems; ++i) {
2773 SDOperand Arg = Mask.getOperand(i);
2774 if (Arg.getOpcode() == ISD::UNDEF) {
2775 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2778 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2779 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2781 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2783 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2785 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2789 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2790 /// match movhlps. The lower half elements should come from upper half of
2791 /// V1 (and in order), and the upper half elements should come from the upper
2792 /// half of V2 (and in order).
2793 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2794 unsigned NumElems = Mask->getNumOperands();
2797 for (unsigned i = 0, e = 2; i != e; ++i)
2798 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2800 for (unsigned i = 2; i != 4; ++i)
2801 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2806 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2807 /// is promoted to a vector.
2808 static inline bool isScalarLoadToVector(SDNode *N) {
2809 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2810 N = N->getOperand(0).Val;
2811 return ISD::isNON_EXTLoad(N);
2816 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2817 /// match movlp{s|d}. The lower half elements should come from lower half of
2818 /// V1 (and in order), and the upper half elements should come from the upper
2819 /// half of V2 (and in order). And since V1 will become the source of the
2820 /// MOVLP, it must be either a vector load or a scalar load to vector.
2821 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2822 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2824 // Is V2 is a vector load, don't do this transformation. We will try to use
2825 // load folding shufps op.
2826 if (ISD::isNON_EXTLoad(V2))
2829 unsigned NumElems = Mask->getNumOperands();
2830 if (NumElems != 2 && NumElems != 4)
2832 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2833 if (!isUndefOrEqual(Mask->getOperand(i), i))
2835 for (unsigned i = NumElems/2; i != NumElems; ++i)
2836 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2841 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2843 static bool isSplatVector(SDNode *N) {
2844 if (N->getOpcode() != ISD::BUILD_VECTOR)
2847 SDOperand SplatValue = N->getOperand(0);
2848 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2849 if (N->getOperand(i) != SplatValue)
2854 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2856 static bool isUndefShuffle(SDNode *N) {
2857 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2860 SDOperand V1 = N->getOperand(0);
2861 SDOperand V2 = N->getOperand(1);
2862 SDOperand Mask = N->getOperand(2);
2863 unsigned NumElems = Mask.getNumOperands();
2864 for (unsigned i = 0; i != NumElems; ++i) {
2865 SDOperand Arg = Mask.getOperand(i);
2866 if (Arg.getOpcode() != ISD::UNDEF) {
2867 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2868 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2870 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2877 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2879 static inline bool isZeroNode(SDOperand Elt) {
2880 return ((isa<ConstantSDNode>(Elt) &&
2881 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2882 (isa<ConstantFPSDNode>(Elt) &&
2883 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2886 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2887 /// to an zero vector.
2888 static bool isZeroShuffle(SDNode *N) {
2889 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2892 SDOperand V1 = N->getOperand(0);
2893 SDOperand V2 = N->getOperand(1);
2894 SDOperand Mask = N->getOperand(2);
2895 unsigned NumElems = Mask.getNumOperands();
2896 for (unsigned i = 0; i != NumElems; ++i) {
2897 SDOperand Arg = Mask.getOperand(i);
2898 if (Arg.getOpcode() == ISD::UNDEF)
2901 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2902 if (Idx < NumElems) {
2903 unsigned Opc = V1.Val->getOpcode();
2904 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2906 if (Opc != ISD::BUILD_VECTOR ||
2907 !isZeroNode(V1.Val->getOperand(Idx)))
2909 } else if (Idx >= NumElems) {
2910 unsigned Opc = V2.Val->getOpcode();
2911 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2913 if (Opc != ISD::BUILD_VECTOR ||
2914 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2921 /// getZeroVector - Returns a vector of specified type with all zero elements.
2923 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2924 assert(MVT::isVector(VT) && "Expected a vector type");
2926 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2927 // type. This ensures they get CSE'd.
2928 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2930 if (MVT::getSizeInBits(VT) == 64) // MMX
2931 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2933 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2934 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2937 /// getOnesVector - Returns a vector of specified type with all bits set.
2939 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2940 assert(MVT::isVector(VT) && "Expected a vector type");
2942 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2943 // type. This ensures they get CSE'd.
2944 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2946 if (MVT::getSizeInBits(VT) == 64) // MMX
2947 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2949 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2950 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2954 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2955 /// that point to V2 points to its first element.
2956 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2957 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2959 bool Changed = false;
2960 SmallVector<SDOperand, 8> MaskVec;
2961 unsigned NumElems = Mask.getNumOperands();
2962 for (unsigned i = 0; i != NumElems; ++i) {
2963 SDOperand Arg = Mask.getOperand(i);
2964 if (Arg.getOpcode() != ISD::UNDEF) {
2965 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2966 if (Val > NumElems) {
2967 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2971 MaskVec.push_back(Arg);
2975 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2976 &MaskVec[0], MaskVec.size());
2980 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2981 /// operation of specified width.
2982 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2983 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2984 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2986 SmallVector<SDOperand, 8> MaskVec;
2987 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2988 for (unsigned i = 1; i != NumElems; ++i)
2989 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2990 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2993 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2994 /// of specified width.
2995 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2996 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2997 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
2998 SmallVector<SDOperand, 8> MaskVec;
2999 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3000 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3001 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3003 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
3006 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3007 /// of specified width.
3008 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
3009 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3010 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3011 unsigned Half = NumElems/2;
3012 SmallVector<SDOperand, 8> MaskVec;
3013 for (unsigned i = 0; i != Half; ++i) {
3014 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3015 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3017 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
3020 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
3022 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
3023 SDOperand V1 = Op.getOperand(0);
3024 SDOperand Mask = Op.getOperand(2);
3025 MVT::ValueType VT = Op.getValueType();
3026 unsigned NumElems = Mask.getNumOperands();
3027 Mask = getUnpacklMask(NumElems, DAG);
3028 while (NumElems != 4) {
3029 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
3032 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
3034 Mask = getZeroVector(MVT::v4i32, DAG);
3035 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
3036 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
3037 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3040 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3041 /// vector of zero or undef vector. This produces a shuffle where the low
3042 /// element of V2 is swizzled into the zero/undef vector, landing at element
3043 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3044 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
3045 unsigned NumElems, unsigned Idx,
3046 bool isZero, SelectionDAG &DAG) {
3047 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
3048 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3049 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3050 SmallVector<SDOperand, 16> MaskVec;
3051 for (unsigned i = 0; i != NumElems; ++i)
3052 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3053 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3055 MaskVec.push_back(DAG.getConstant(i, EVT));
3056 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3057 &MaskVec[0], MaskVec.size());
3058 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3061 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3063 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
3064 unsigned NumNonZero, unsigned NumZero,
3065 SelectionDAG &DAG, TargetLowering &TLI) {
3071 for (unsigned i = 0; i < 16; ++i) {
3072 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3073 if (ThisIsNonZero && First) {
3075 V = getZeroVector(MVT::v8i16, DAG);
3077 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3082 SDOperand ThisElt(0, 0), LastElt(0, 0);
3083 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3084 if (LastIsNonZero) {
3085 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3087 if (ThisIsNonZero) {
3088 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3089 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3090 ThisElt, DAG.getConstant(8, MVT::i8));
3092 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3097 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3098 DAG.getConstant(i/2, TLI.getPointerTy()));
3102 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3105 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3107 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3108 unsigned NumNonZero, unsigned NumZero,
3109 SelectionDAG &DAG, TargetLowering &TLI) {
3115 for (unsigned i = 0; i < 8; ++i) {
3116 bool isNonZero = (NonZeros & (1 << i)) != 0;
3120 V = getZeroVector(MVT::v8i16, DAG);
3122 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3125 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3126 DAG.getConstant(i, TLI.getPointerTy()));
3134 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3135 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3136 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
3137 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3138 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3139 // eliminated on x86-32 hosts.
3140 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3143 if (ISD::isBuildVectorAllOnes(Op.Val))
3144 return getOnesVector(Op.getValueType(), DAG);
3145 return getZeroVector(Op.getValueType(), DAG);
3148 MVT::ValueType VT = Op.getValueType();
3149 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3150 unsigned EVTBits = MVT::getSizeInBits(EVT);
3152 unsigned NumElems = Op.getNumOperands();
3153 unsigned NumZero = 0;
3154 unsigned NumNonZero = 0;
3155 unsigned NonZeros = 0;
3156 unsigned NumNonZeroImms = 0;
3157 std::set<SDOperand> Values;
3158 for (unsigned i = 0; i < NumElems; ++i) {
3159 SDOperand Elt = Op.getOperand(i);
3160 if (Elt.getOpcode() != ISD::UNDEF) {
3162 if (isZeroNode(Elt))
3165 NonZeros |= (1 << i);
3167 if (Elt.getOpcode() == ISD::Constant ||
3168 Elt.getOpcode() == ISD::ConstantFP)
3174 if (NumNonZero == 0) {
3175 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3176 return DAG.getNode(ISD::UNDEF, VT);
3179 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3180 if (Values.size() == 1)
3183 // Special case for single non-zero element.
3184 if (NumNonZero == 1) {
3185 unsigned Idx = CountTrailingZeros_32(NonZeros);
3186 SDOperand Item = Op.getOperand(Idx);
3187 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3189 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3190 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3193 if (EVTBits == 32) {
3194 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3195 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3197 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3198 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3199 SmallVector<SDOperand, 8> MaskVec;
3200 for (unsigned i = 0; i < NumElems; i++)
3201 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3202 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3203 &MaskVec[0], MaskVec.size());
3204 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3205 DAG.getNode(ISD::UNDEF, VT), Mask);
3209 // A vector full of immediates; various special cases are already
3210 // handled, so this is best done with a single constant-pool load.
3211 if (NumNonZero == NumNonZeroImms)
3214 // Let legalizer expand 2-wide build_vectors.
3218 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3219 if (EVTBits == 8 && NumElems == 16) {
3220 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3222 if (V.Val) return V;
3225 if (EVTBits == 16 && NumElems == 8) {
3226 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3228 if (V.Val) return V;
3231 // If element VT is == 32 bits, turn it into a number of shuffles.
3232 SmallVector<SDOperand, 8> V;
3234 if (NumElems == 4 && NumZero > 0) {
3235 for (unsigned i = 0; i < 4; ++i) {
3236 bool isZero = !(NonZeros & (1 << i));
3238 V[i] = getZeroVector(VT, DAG);
3240 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3243 for (unsigned i = 0; i < 2; ++i) {
3244 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3247 V[i] = V[i*2]; // Must be a zero vector.
3250 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3251 getMOVLMask(NumElems, DAG));
3254 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3255 getMOVLMask(NumElems, DAG));
3258 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3259 getUnpacklMask(NumElems, DAG));
3264 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3265 // clears the upper bits.
3266 // FIXME: we can do the same for v4f32 case when we know both parts of
3267 // the lower half come from scalar_to_vector (loadf32). We should do
3268 // that in post legalizer dag combiner with target specific hooks.
3269 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3271 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3272 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3273 SmallVector<SDOperand, 8> MaskVec;
3274 bool Reverse = (NonZeros & 0x3) == 2;
3275 for (unsigned i = 0; i < 2; ++i)
3277 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3279 MaskVec.push_back(DAG.getConstant(i, EVT));
3280 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3281 for (unsigned i = 0; i < 2; ++i)
3283 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3285 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3286 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3287 &MaskVec[0], MaskVec.size());
3288 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3291 if (Values.size() > 2) {
3292 // Expand into a number of unpckl*.
3294 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3295 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3296 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3297 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3298 for (unsigned i = 0; i < NumElems; ++i)
3299 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3301 while (NumElems != 0) {
3302 for (unsigned i = 0; i < NumElems; ++i)
3303 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3314 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3315 SDOperand PermMask, SelectionDAG &DAG,
3316 TargetLowering &TLI) {
3317 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3318 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3319 if (isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3320 // Handle v8i16 shuffle high / low shuffle node pair.
3321 SmallVector<SDOperand, 8> MaskVec;
3322 for (unsigned i = 0; i != 4; ++i)
3323 MaskVec.push_back(PermMask.getOperand(i));
3324 for (unsigned i = 4; i != 8; ++i)
3325 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3326 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3327 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V2, Mask);
3329 for (unsigned i = 0; i != 4; ++i)
3330 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3331 for (unsigned i = 4; i != 8; ++i)
3332 MaskVec.push_back(PermMask.getOperand(i));
3333 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3334 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V2, Mask);
3337 // Lower than into extracts and inserts but try to do as few as possible.
3338 // First, let's find out how many elements are already in the right order.
3339 unsigned V1InOrder = 0;
3340 unsigned V1FromV1 = 0;
3341 unsigned V2InOrder = 0;
3342 unsigned V2FromV2 = 0;
3343 SmallVector<unsigned, 8> V1Elts;
3344 SmallVector<unsigned, 8> V2Elts;
3345 for (unsigned i = 0; i < 8; ++i) {
3346 SDOperand Elt = PermMask.getOperand(i);
3347 if (Elt.getOpcode() == ISD::UNDEF) {
3348 V1Elts.push_back(i);
3349 V2Elts.push_back(i);
3353 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3355 V1Elts.push_back(i);
3356 V2Elts.push_back(i+8);
3358 } else if (EltIdx == i+8) {
3359 V1Elts.push_back(i+8);
3360 V2Elts.push_back(i);
3363 V1Elts.push_back(EltIdx);
3364 V2Elts.push_back(EltIdx);
3373 if (V2InOrder > V1InOrder) {
3374 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3376 std::swap(V1Elts, V2Elts);
3377 std::swap(V1FromV1, V2FromV2);
3380 MVT::ValueType PtrVT = TLI.getPointerTy();
3382 // If there are elements that are from V1 but out of place,
3383 // then first sort them in place
3384 SmallVector<SDOperand, 8> MaskVec;
3385 for (unsigned i = 0; i < 8; ++i) {
3386 unsigned EltIdx = V1Elts[i];
3388 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3390 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3392 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3393 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3396 // Now let's insert elements from the other vector.
3397 for (unsigned i = 0; i < 8; ++i) {
3398 unsigned EltIdx = V1Elts[i];
3401 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3402 DAG.getConstant(EltIdx - 8, PtrVT));
3403 V1 = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V1, ExtOp,
3404 DAG.getConstant(i, PtrVT));
3410 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3411 SDOperand V1 = Op.getOperand(0);
3412 SDOperand V2 = Op.getOperand(1);
3413 SDOperand PermMask = Op.getOperand(2);
3414 MVT::ValueType VT = Op.getValueType();
3415 unsigned NumElems = PermMask.getNumOperands();
3416 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3417 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3418 bool V1IsSplat = false;
3419 bool V2IsSplat = false;
3421 if (isUndefShuffle(Op.Val))
3422 return DAG.getNode(ISD::UNDEF, VT);
3424 if (isZeroShuffle(Op.Val))
3425 return getZeroVector(VT, DAG);
3427 if (isIdentityMask(PermMask.Val))
3429 else if (isIdentityMask(PermMask.Val, true))
3432 if (isSplatMask(PermMask.Val)) {
3433 if (NumElems <= 4) return Op;
3434 // Promote it to a v4i32 splat.
3435 return PromoteSplat(Op, DAG);
3438 if (X86::isMOVLMask(PermMask.Val))
3439 return (V1IsUndef) ? V2 : Op;
3441 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3442 X86::isMOVSLDUPMask(PermMask.Val) ||
3443 X86::isMOVHLPSMask(PermMask.Val) ||
3444 X86::isMOVHPMask(PermMask.Val) ||
3445 X86::isMOVLPMask(PermMask.Val))
3448 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3449 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3450 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3452 bool Commuted = false;
3453 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3454 // 1,1,1,1 -> v8i16 though.
3455 V1IsSplat = isSplatVector(V1.Val);
3456 V2IsSplat = isSplatVector(V2.Val);
3458 // Canonicalize the splat or undef, if present, to be on the RHS.
3459 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3460 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3461 std::swap(V1IsSplat, V2IsSplat);
3462 std::swap(V1IsUndef, V2IsUndef);
3466 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3467 if (V2IsUndef) return V1;
3468 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3470 // V2 is a splat, so the mask may be malformed. That is, it may point
3471 // to any V2 element. The instruction selectior won't like this. Get
3472 // a corrected mask and commute to form a proper MOVS{S|D}.
3473 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3474 if (NewMask.Val != PermMask.Val)
3475 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3480 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3481 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3482 X86::isUNPCKLMask(PermMask.Val) ||
3483 X86::isUNPCKHMask(PermMask.Val))
3487 // Normalize mask so all entries that point to V2 points to its first
3488 // element then try to match unpck{h|l} again. If match, return a
3489 // new vector_shuffle with the corrected mask.
3490 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3491 if (NewMask.Val != PermMask.Val) {
3492 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3493 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3494 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3495 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3496 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3497 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3502 // Normalize the node to match x86 shuffle ops if needed
3503 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3504 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3507 // Commute is back and try unpck* again.
3508 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3509 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3510 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3511 X86::isUNPCKLMask(PermMask.Val) ||
3512 X86::isUNPCKHMask(PermMask.Val))
3516 // If VT is integer, try PSHUF* first, then SHUFP*.
3517 if (MVT::isInteger(VT)) {
3518 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3519 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3520 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3521 X86::isPSHUFDMask(PermMask.Val)) ||
3522 X86::isPSHUFHWMask(PermMask.Val) ||
3523 X86::isPSHUFLWMask(PermMask.Val)) {
3524 if (V2.getOpcode() != ISD::UNDEF)
3525 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3526 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3530 if (X86::isSHUFPMask(PermMask.Val) &&
3531 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3534 // Floating point cases in the other order.
3535 if (X86::isSHUFPMask(PermMask.Val))
3537 if (X86::isPSHUFDMask(PermMask.Val) ||
3538 X86::isPSHUFHWMask(PermMask.Val) ||
3539 X86::isPSHUFLWMask(PermMask.Val)) {
3540 if (V2.getOpcode() != ISD::UNDEF)
3541 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3542 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3547 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3548 if (VT == MVT::v8i16)
3549 return LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3551 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3552 // Don't do this for MMX.
3553 MVT::ValueType MaskVT = PermMask.getValueType();
3554 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3555 SmallVector<std::pair<int, int>, 8> Locs;
3556 Locs.reserve(NumElems);
3557 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3558 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3561 // If no more than two elements come from either vector. This can be
3562 // implemented with two shuffles. First shuffle gather the elements.
3563 // The second shuffle, which takes the first shuffle as both of its
3564 // vector operands, put the elements into the right order.
3565 for (unsigned i = 0; i != NumElems; ++i) {
3566 SDOperand Elt = PermMask.getOperand(i);
3567 if (Elt.getOpcode() == ISD::UNDEF) {
3568 Locs[i] = std::make_pair(-1, -1);
3570 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3571 if (Val < NumElems) {
3572 Locs[i] = std::make_pair(0, NumLo);
3576 Locs[i] = std::make_pair(1, NumHi);
3577 if (2+NumHi < NumElems)
3578 Mask1[2+NumHi] = Elt;
3583 if (NumLo <= 2 && NumHi <= 2) {
3584 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3585 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3586 &Mask1[0], Mask1.size()));
3587 for (unsigned i = 0; i != NumElems; ++i) {
3588 if (Locs[i].first == -1)
3591 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3592 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3593 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3597 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3598 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3599 &Mask2[0], Mask2.size()));
3602 // Break it into (shuffle shuffle_hi, shuffle_lo).
3604 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3605 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3606 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3607 unsigned MaskIdx = 0;
3609 unsigned HiIdx = NumElems/2;
3610 for (unsigned i = 0; i != NumElems; ++i) {
3611 if (i == NumElems/2) {
3617 SDOperand Elt = PermMask.getOperand(i);
3618 if (Elt.getOpcode() == ISD::UNDEF) {
3619 Locs[i] = std::make_pair(-1, -1);
3620 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3621 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3622 (*MaskPtr)[LoIdx] = Elt;
3625 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3626 (*MaskPtr)[HiIdx] = Elt;
3631 SDOperand LoShuffle =
3632 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3633 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3634 &LoMask[0], LoMask.size()));
3635 SDOperand HiShuffle =
3636 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3637 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3638 &HiMask[0], HiMask.size()));
3639 SmallVector<SDOperand, 8> MaskOps;
3640 for (unsigned i = 0; i != NumElems; ++i) {
3641 if (Locs[i].first == -1) {
3642 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3644 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3645 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3648 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3649 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3650 &MaskOps[0], MaskOps.size()));
3657 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3658 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3661 MVT::ValueType VT = Op.getValueType();
3662 // TODO: handle v16i8.
3663 if (MVT::getSizeInBits(VT) == 16) {
3664 // Transform it so it match pextrw which produces a 32-bit result.
3665 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3666 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3667 Op.getOperand(0), Op.getOperand(1));
3668 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3669 DAG.getValueType(VT));
3670 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3671 } else if (MVT::getSizeInBits(VT) == 32) {
3672 SDOperand Vec = Op.getOperand(0);
3673 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3676 // SHUFPS the element to the lowest double word, then movss.
3677 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3678 SmallVector<SDOperand, 8> IdxVec;
3680 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3682 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3684 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3686 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3687 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3688 &IdxVec[0], IdxVec.size());
3689 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3690 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3691 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3692 DAG.getConstant(0, getPointerTy()));
3693 } else if (MVT::getSizeInBits(VT) == 64) {
3694 SDOperand Vec = Op.getOperand(0);
3695 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3699 // UNPCKHPD the element to the lowest double word, then movsd.
3700 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3701 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3702 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3703 SmallVector<SDOperand, 8> IdxVec;
3704 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3706 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3707 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3708 &IdxVec[0], IdxVec.size());
3709 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3710 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3711 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3712 DAG.getConstant(0, getPointerTy()));
3719 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3720 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3721 // as its second argument.
3722 MVT::ValueType VT = Op.getValueType();
3723 MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
3724 SDOperand N0 = Op.getOperand(0);
3725 SDOperand N1 = Op.getOperand(1);
3726 SDOperand N2 = Op.getOperand(2);
3727 if (MVT::getSizeInBits(BaseVT) == 16) {
3728 if (N1.getValueType() != MVT::i32)
3729 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3730 if (N2.getValueType() != MVT::i32)
3731 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
3732 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3733 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3734 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3737 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3738 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3739 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3740 SmallVector<SDOperand, 8> MaskVec;
3741 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3742 for (unsigned i = 1; i <= 3; ++i)
3743 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3744 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3745 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3746 &MaskVec[0], MaskVec.size()));
3748 // Use two pinsrw instructions to insert a 32 bit value.
3750 if (MVT::isFloatingPoint(N1.getValueType())) {
3751 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3752 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3753 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3754 DAG.getConstant(0, getPointerTy()));
3756 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3757 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3758 DAG.getConstant(Idx, getPointerTy()));
3759 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3760 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3761 DAG.getConstant(Idx+1, getPointerTy()));
3762 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3770 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3771 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3772 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3775 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3776 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3777 // one of the above mentioned nodes. It has to be wrapped because otherwise
3778 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3779 // be used to form addressing mode. These wrapped nodes will be selected
3782 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3783 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3784 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3786 CP->getAlignment());
3787 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3788 // With PIC, the address is actually $g + Offset.
3789 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3790 !Subtarget->isPICStyleRIPRel()) {
3791 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3792 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3800 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3801 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3802 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3803 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3804 // With PIC, the address is actually $g + Offset.
3805 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3806 !Subtarget->isPICStyleRIPRel()) {
3807 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3808 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3812 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3813 // load the value at address GV, not the value of GV itself. This means that
3814 // the GlobalAddress must be in the base or index register of the address, not
3815 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3816 // The same applies for external symbols during PIC codegen
3817 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3818 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3823 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3825 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3826 const MVT::ValueType PtrVT) {
3828 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
3829 DAG.getNode(X86ISD::GlobalBaseReg,
3831 InFlag = Chain.getValue(1);
3833 // emit leal symbol@TLSGD(,%ebx,1), %eax
3834 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
3835 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3836 GA->getValueType(0),
3838 SDOperand Ops[] = { Chain, TGA, InFlag };
3839 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
3840 InFlag = Result.getValue(2);
3841 Chain = Result.getValue(1);
3843 // call ___tls_get_addr. This function receives its argument in
3844 // the register EAX.
3845 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
3846 InFlag = Chain.getValue(1);
3848 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
3849 SDOperand Ops1[] = { Chain,
3850 DAG.getTargetExternalSymbol("___tls_get_addr",
3852 DAG.getRegister(X86::EAX, PtrVT),
3853 DAG.getRegister(X86::EBX, PtrVT),
3855 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
3856 InFlag = Chain.getValue(1);
3858 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
3861 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
3862 // "local exec" model.
3864 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
3865 const MVT::ValueType PtrVT) {
3866 // Get the Thread Pointer
3867 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
3868 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
3870 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
3871 GA->getValueType(0),
3873 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
3875 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
3876 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
3878 // The address of the thread local variable is the add of the thread
3879 // pointer with the offset of the variable.
3880 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
3884 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
3885 // TODO: implement the "local dynamic" model
3886 // TODO: implement the "initial exec"model for pic executables
3887 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
3888 "TLS not implemented for non-ELF and 64-bit targets");
3889 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3890 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
3891 // otherwise use the "Local Exec"TLS Model
3892 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
3893 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
3895 return LowerToTLSExecModel(GA, DAG, getPointerTy());
3899 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3900 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3901 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3902 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3903 // With PIC, the address is actually $g + Offset.
3904 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3905 !Subtarget->isPICStyleRIPRel()) {
3906 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3907 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3914 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3915 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3916 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3917 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3918 // With PIC, the address is actually $g + Offset.
3919 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3920 !Subtarget->isPICStyleRIPRel()) {
3921 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3922 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3929 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
3930 /// take a 2 x i32 value to shift plus a shift amount.
3931 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3932 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3933 "Not an i64 shift!");
3934 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3935 SDOperand ShOpLo = Op.getOperand(0);
3936 SDOperand ShOpHi = Op.getOperand(1);
3937 SDOperand ShAmt = Op.getOperand(2);
3938 SDOperand Tmp1 = isSRA ?
3939 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3940 DAG.getConstant(0, MVT::i32);
3942 SDOperand Tmp2, Tmp3;
3943 if (Op.getOpcode() == ISD::SHL_PARTS) {
3944 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3945 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3947 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3948 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3951 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3952 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3953 DAG.getConstant(32, MVT::i8));
3954 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
3955 AndNode, DAG.getConstant(0, MVT::i8));
3958 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3959 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3960 SmallVector<SDOperand, 4> Ops;
3961 if (Op.getOpcode() == ISD::SHL_PARTS) {
3962 Ops.push_back(Tmp2);
3963 Ops.push_back(Tmp3);
3965 Ops.push_back(Cond);
3966 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3969 Ops.push_back(Tmp3);
3970 Ops.push_back(Tmp1);
3972 Ops.push_back(Cond);
3973 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3975 Ops.push_back(Tmp2);
3976 Ops.push_back(Tmp3);
3978 Ops.push_back(Cond);
3979 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3982 Ops.push_back(Tmp3);
3983 Ops.push_back(Tmp1);
3985 Ops.push_back(Cond);
3986 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
3989 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3993 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3996 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3997 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3998 Op.getOperand(0).getValueType() >= MVT::i16 &&
3999 "Unknown SINT_TO_FP to lower!");
4002 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
4003 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
4004 MachineFunction &MF = DAG.getMachineFunction();
4005 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4006 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4007 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4008 StackSlot, NULL, 0);
4010 // These are really Legal; caller falls through into that case.
4011 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
4013 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
4015 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
4016 Subtarget->is64Bit())
4021 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
4022 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
4024 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4026 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4027 SmallVector<SDOperand, 8> Ops;
4028 Ops.push_back(Chain);
4029 Ops.push_back(StackSlot);
4030 Ops.push_back(DAG.getValueType(SrcVT));
4031 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
4032 Tys, &Ops[0], Ops.size());
4035 Chain = Result.getValue(1);
4036 SDOperand InFlag = Result.getValue(2);
4038 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4039 // shouldn't be necessary except that RFP cannot be live across
4040 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4041 MachineFunction &MF = DAG.getMachineFunction();
4042 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4043 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4044 Tys = DAG.getVTList(MVT::Other);
4045 SmallVector<SDOperand, 8> Ops;
4046 Ops.push_back(Chain);
4047 Ops.push_back(Result);
4048 Ops.push_back(StackSlot);
4049 Ops.push_back(DAG.getValueType(Op.getValueType()));
4050 Ops.push_back(InFlag);
4051 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4052 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
4058 std::pair<SDOperand,SDOperand> X86TargetLowering::
4059 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4060 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4061 "Unknown FP_TO_SINT to lower!");
4063 // These are really Legal.
4064 if (Op.getValueType() == MVT::i32 &&
4065 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
4066 return std::make_pair(SDOperand(), SDOperand());
4067 if (Op.getValueType() == MVT::i32 &&
4068 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
4069 return std::make_pair(SDOperand(), SDOperand());
4070 if (Subtarget->is64Bit() &&
4071 Op.getValueType() == MVT::i64 &&
4072 Op.getOperand(0).getValueType() != MVT::f80)
4073 return std::make_pair(SDOperand(), SDOperand());
4075 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4077 MachineFunction &MF = DAG.getMachineFunction();
4078 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4079 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4080 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4082 switch (Op.getValueType()) {
4083 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4084 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4085 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4086 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4089 SDOperand Chain = DAG.getEntryNode();
4090 SDOperand Value = Op.getOperand(0);
4091 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
4092 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
4093 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4094 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
4095 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4097 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4099 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4100 Chain = Value.getValue(1);
4101 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4102 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4105 // Build the FP_TO_INT*_IN_MEM
4106 SDOperand Ops[] = { Chain, Value, StackSlot };
4107 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4109 return std::make_pair(FIST, StackSlot);
4112 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4113 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4114 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4115 if (FIST.Val == 0) return SDOperand();
4118 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4121 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4122 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4123 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4124 if (FIST.Val == 0) return 0;
4126 // Return an i64 load from the stack slot.
4127 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4129 // Use a MERGE_VALUES node to drop the chain result value.
4130 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4133 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4134 MVT::ValueType VT = Op.getValueType();
4135 MVT::ValueType EltVT = VT;
4136 if (MVT::isVector(VT))
4137 EltVT = MVT::getVectorElementType(VT);
4138 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4139 std::vector<Constant*> CV;
4140 if (EltVT == MVT::f64) {
4141 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4145 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4151 Constant *C = ConstantVector::get(CV);
4152 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4153 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4155 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4158 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4159 MVT::ValueType VT = Op.getValueType();
4160 MVT::ValueType EltVT = VT;
4161 unsigned EltNum = 1;
4162 if (MVT::isVector(VT)) {
4163 EltVT = MVT::getVectorElementType(VT);
4164 EltNum = MVT::getVectorNumElements(VT);
4166 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4167 std::vector<Constant*> CV;
4168 if (EltVT == MVT::f64) {
4169 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4173 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4179 Constant *C = ConstantVector::get(CV);
4180 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4181 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4183 if (MVT::isVector(VT)) {
4184 return DAG.getNode(ISD::BIT_CONVERT, VT,
4185 DAG.getNode(ISD::XOR, MVT::v2i64,
4186 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4187 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4189 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4193 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4194 SDOperand Op0 = Op.getOperand(0);
4195 SDOperand Op1 = Op.getOperand(1);
4196 MVT::ValueType VT = Op.getValueType();
4197 MVT::ValueType SrcVT = Op1.getValueType();
4198 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4200 // If second operand is smaller, extend it first.
4201 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4202 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4204 SrcTy = MVT::getTypeForValueType(SrcVT);
4206 // And if it is bigger, shrink it first.
4207 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4208 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1);
4210 SrcTy = MVT::getTypeForValueType(SrcVT);
4213 // At this point the operands and the result should have the same
4214 // type, and that won't be f80 since that is not custom lowered.
4216 // First get the sign bit of second operand.
4217 std::vector<Constant*> CV;
4218 if (SrcVT == MVT::f64) {
4219 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4220 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4222 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4223 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4224 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4225 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4227 Constant *C = ConstantVector::get(CV);
4228 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4229 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4231 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4233 // Shift sign bit right or left if the two operands have different types.
4234 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4235 // Op0 is MVT::f32, Op1 is MVT::f64.
4236 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4237 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4238 DAG.getConstant(32, MVT::i32));
4239 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4240 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4241 DAG.getConstant(0, getPointerTy()));
4244 // Clear first operand sign bit.
4246 if (VT == MVT::f64) {
4247 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4248 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4250 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4251 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4252 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4253 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4255 C = ConstantVector::get(CV);
4256 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4257 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4259 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4261 // Or the value with the sign bit.
4262 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4265 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4266 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4268 SDOperand Op0 = Op.getOperand(0);
4269 SDOperand Op1 = Op.getOperand(1);
4270 SDOperand CC = Op.getOperand(2);
4271 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4272 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4275 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4277 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4278 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4279 DAG.getConstant(X86CC, MVT::i8), Cond);
4282 assert(isFP && "Illegal integer SetCC!");
4284 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4285 switch (SetCCOpcode) {
4286 default: assert(false && "Illegal floating point SetCC!");
4287 case ISD::SETOEQ: { // !PF & ZF
4288 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4289 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4290 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4291 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4292 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4294 case ISD::SETUNE: { // PF | !ZF
4295 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4296 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4297 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4298 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4299 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4305 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4306 bool addTest = true;
4307 SDOperand Cond = Op.getOperand(0);
4310 if (Cond.getOpcode() == ISD::SETCC)
4311 Cond = LowerSETCC(Cond, DAG);
4313 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4314 // setting operand in place of the X86ISD::SETCC.
4315 if (Cond.getOpcode() == X86ISD::SETCC) {
4316 CC = Cond.getOperand(0);
4318 SDOperand Cmp = Cond.getOperand(1);
4319 unsigned Opc = Cmp.getOpcode();
4320 MVT::ValueType VT = Op.getValueType();
4321 bool IllegalFPCMov = false;
4322 if (VT == MVT::f32 && !X86ScalarSSEf32)
4323 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4324 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4325 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4326 else if (VT == MVT::f80)
4327 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4328 if ((Opc == X86ISD::CMP ||
4329 Opc == X86ISD::COMI ||
4330 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4337 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4338 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4341 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4343 SmallVector<SDOperand, 4> Ops;
4344 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4345 // condition is true.
4346 Ops.push_back(Op.getOperand(2));
4347 Ops.push_back(Op.getOperand(1));
4349 Ops.push_back(Cond);
4350 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4353 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4354 bool addTest = true;
4355 SDOperand Chain = Op.getOperand(0);
4356 SDOperand Cond = Op.getOperand(1);
4357 SDOperand Dest = Op.getOperand(2);
4360 if (Cond.getOpcode() == ISD::SETCC)
4361 Cond = LowerSETCC(Cond, DAG);
4363 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4364 // setting operand in place of the X86ISD::SETCC.
4365 if (Cond.getOpcode() == X86ISD::SETCC) {
4366 CC = Cond.getOperand(0);
4368 SDOperand Cmp = Cond.getOperand(1);
4369 unsigned Opc = Cmp.getOpcode();
4370 if (Opc == X86ISD::CMP ||
4371 Opc == X86ISD::COMI ||
4372 Opc == X86ISD::UCOMI) {
4379 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4380 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4382 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4383 Chain, Op.getOperand(2), CC, Cond);
4386 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4387 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4388 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4390 if (Subtarget->is64Bit())
4391 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4392 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4394 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4396 switch (CallingConv) {
4398 assert(0 && "Unsupported calling convention");
4399 case CallingConv::Fast:
4400 if (isTailCall && PerformTailCallOpt)
4401 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4403 return LowerCCCCallTo(Op,DAG, CallingConv);
4404 case CallingConv::C:
4405 case CallingConv::X86_StdCall:
4406 return LowerCCCCallTo(Op, DAG, CallingConv);
4407 case CallingConv::X86_FastCall:
4408 return LowerFastCCCallTo(Op, DAG, CallingConv);
4413 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4414 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4415 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4416 // that the guard pages used by the OS virtual memory manager are allocated in
4417 // correct sequence.
4419 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4420 SelectionDAG &DAG) {
4421 assert(Subtarget->isTargetCygMing() &&
4422 "This should be used only on Cygwin/Mingw targets");
4425 SDOperand Chain = Op.getOperand(0);
4426 SDOperand Size = Op.getOperand(1);
4427 // FIXME: Ensure alignment here
4431 MVT::ValueType IntPtr = getPointerTy();
4432 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4434 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4435 Flag = Chain.getValue(1);
4437 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4438 SDOperand Ops[] = { Chain,
4439 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4440 DAG.getRegister(X86::EAX, IntPtr),
4442 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4443 Flag = Chain.getValue(1);
4445 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4447 std::vector<MVT::ValueType> Tys;
4448 Tys.push_back(SPTy);
4449 Tys.push_back(MVT::Other);
4450 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4451 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4455 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4456 MachineFunction &MF = DAG.getMachineFunction();
4457 const Function* Fn = MF.getFunction();
4458 if (Fn->hasExternalLinkage() &&
4459 Subtarget->isTargetCygMing() &&
4460 Fn->getName() == "main")
4461 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4463 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4464 if (Subtarget->is64Bit())
4465 return LowerX86_64CCCArguments(Op, DAG);
4469 assert(0 && "Unsupported calling convention");
4470 case CallingConv::Fast:
4471 return LowerCCCArguments(Op,DAG, true);
4473 case CallingConv::C:
4474 return LowerCCCArguments(Op, DAG);
4475 case CallingConv::X86_StdCall:
4476 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4477 return LowerCCCArguments(Op, DAG, true);
4478 case CallingConv::X86_FastCall:
4479 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4480 return LowerFastCCArguments(Op, DAG);
4484 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4485 SDOperand InFlag(0, 0);
4486 SDOperand Chain = Op.getOperand(0);
4488 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4489 if (Align == 0) Align = 1;
4491 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4492 // If not DWORD aligned or size is more than the threshold, call memset.
4493 // The libc version is likely to be faster for these cases. It can use the
4494 // address value and run time information about the CPU.
4495 if ((Align & 3) != 0 ||
4496 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4497 MVT::ValueType IntPtr = getPointerTy();
4498 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4499 TargetLowering::ArgListTy Args;
4500 TargetLowering::ArgListEntry Entry;
4501 Entry.Node = Op.getOperand(1);
4502 Entry.Ty = IntPtrTy;
4503 Args.push_back(Entry);
4504 // Extend the unsigned i8 argument to be an int value for the call.
4505 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4506 Entry.Ty = IntPtrTy;
4507 Args.push_back(Entry);
4508 Entry.Node = Op.getOperand(3);
4509 Args.push_back(Entry);
4510 std::pair<SDOperand,SDOperand> CallResult =
4511 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4512 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4513 return CallResult.second;
4518 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4519 unsigned BytesLeft = 0;
4520 bool TwoRepStos = false;
4523 uint64_t Val = ValC->getValue() & 255;
4525 // If the value is a constant, then we can potentially use larger sets.
4526 switch (Align & 3) {
4527 case 2: // WORD aligned
4530 Val = (Val << 8) | Val;
4532 case 0: // DWORD aligned
4535 Val = (Val << 8) | Val;
4536 Val = (Val << 16) | Val;
4537 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4540 Val = (Val << 32) | Val;
4543 default: // Byte aligned
4546 Count = Op.getOperand(3);
4550 if (AVT > MVT::i8) {
4552 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4553 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4554 BytesLeft = I->getValue() % UBytes;
4556 assert(AVT >= MVT::i32 &&
4557 "Do not use rep;stos if not at least DWORD aligned");
4558 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4559 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4564 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4566 InFlag = Chain.getValue(1);
4569 Count = Op.getOperand(3);
4570 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4571 InFlag = Chain.getValue(1);
4574 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4576 InFlag = Chain.getValue(1);
4577 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4578 Op.getOperand(1), InFlag);
4579 InFlag = Chain.getValue(1);
4581 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4582 SmallVector<SDOperand, 8> Ops;
4583 Ops.push_back(Chain);
4584 Ops.push_back(DAG.getValueType(AVT));
4585 Ops.push_back(InFlag);
4586 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4589 InFlag = Chain.getValue(1);
4590 Count = Op.getOperand(3);
4591 MVT::ValueType CVT = Count.getValueType();
4592 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4593 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4594 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4596 InFlag = Chain.getValue(1);
4597 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4599 Ops.push_back(Chain);
4600 Ops.push_back(DAG.getValueType(MVT::i8));
4601 Ops.push_back(InFlag);
4602 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4603 } else if (BytesLeft) {
4604 // Issue stores for the last 1 - 7 bytes.
4606 unsigned Val = ValC->getValue() & 255;
4607 unsigned Offset = I->getValue() - BytesLeft;
4608 SDOperand DstAddr = Op.getOperand(1);
4609 MVT::ValueType AddrVT = DstAddr.getValueType();
4610 if (BytesLeft >= 4) {
4611 Val = (Val << 8) | Val;
4612 Val = (Val << 16) | Val;
4613 Value = DAG.getConstant(Val, MVT::i32);
4614 Chain = DAG.getStore(Chain, Value,
4615 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4616 DAG.getConstant(Offset, AddrVT)),
4621 if (BytesLeft >= 2) {
4622 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4623 Chain = DAG.getStore(Chain, Value,
4624 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4625 DAG.getConstant(Offset, AddrVT)),
4630 if (BytesLeft == 1) {
4631 Value = DAG.getConstant(Val, MVT::i8);
4632 Chain = DAG.getStore(Chain, Value,
4633 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4634 DAG.getConstant(Offset, AddrVT)),
4642 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4647 SelectionDAG &DAG) {
4649 unsigned BytesLeft = 0;
4650 switch (Align & 3) {
4651 case 2: // WORD aligned
4654 case 0: // DWORD aligned
4656 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4659 default: // Byte aligned
4664 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4665 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4666 BytesLeft = Size % UBytes;
4668 SDOperand InFlag(0, 0);
4669 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4671 InFlag = Chain.getValue(1);
4672 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4674 InFlag = Chain.getValue(1);
4675 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4677 InFlag = Chain.getValue(1);
4679 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4680 SmallVector<SDOperand, 8> Ops;
4681 Ops.push_back(Chain);
4682 Ops.push_back(DAG.getValueType(AVT));
4683 Ops.push_back(InFlag);
4684 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4687 // Issue loads and stores for the last 1 - 7 bytes.
4688 unsigned Offset = Size - BytesLeft;
4689 SDOperand DstAddr = Dest;
4690 MVT::ValueType DstVT = DstAddr.getValueType();
4691 SDOperand SrcAddr = Source;
4692 MVT::ValueType SrcVT = SrcAddr.getValueType();
4694 if (BytesLeft >= 4) {
4695 Value = DAG.getLoad(MVT::i32, Chain,
4696 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4697 DAG.getConstant(Offset, SrcVT)),
4699 Chain = Value.getValue(1);
4700 Chain = DAG.getStore(Chain, Value,
4701 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4702 DAG.getConstant(Offset, DstVT)),
4707 if (BytesLeft >= 2) {
4708 Value = DAG.getLoad(MVT::i16, Chain,
4709 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4710 DAG.getConstant(Offset, SrcVT)),
4712 Chain = Value.getValue(1);
4713 Chain = DAG.getStore(Chain, Value,
4714 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4715 DAG.getConstant(Offset, DstVT)),
4721 if (BytesLeft == 1) {
4722 Value = DAG.getLoad(MVT::i8, Chain,
4723 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4724 DAG.getConstant(Offset, SrcVT)),
4726 Chain = Value.getValue(1);
4727 Chain = DAG.getStore(Chain, Value,
4728 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4729 DAG.getConstant(Offset, DstVT)),
4737 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
4738 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
4739 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4740 SDOperand TheChain = N->getOperand(0);
4741 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
4742 if (Subtarget->is64Bit()) {
4743 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4744 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
4745 MVT::i64, rax.getValue(2));
4746 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
4747 DAG.getConstant(32, MVT::i8));
4749 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
4752 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4753 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4756 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4757 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
4758 MVT::i32, eax.getValue(2));
4759 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
4760 SDOperand Ops[] = { eax, edx };
4761 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
4763 // Use a MERGE_VALUES to return the value and chain.
4764 Ops[1] = edx.getValue(1);
4765 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4766 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
4769 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4770 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4772 if (!Subtarget->is64Bit()) {
4773 // vastart just stores the address of the VarArgsFrameIndex slot into the
4774 // memory location argument.
4775 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4776 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4781 // gp_offset (0 - 6 * 8)
4782 // fp_offset (48 - 48 + 8 * 16)
4783 // overflow_arg_area (point to parameters coming in memory).
4785 SmallVector<SDOperand, 8> MemOps;
4786 SDOperand FIN = Op.getOperand(1);
4788 SDOperand Store = DAG.getStore(Op.getOperand(0),
4789 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4790 FIN, SV->getValue(), SV->getOffset());
4791 MemOps.push_back(Store);
4794 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4795 DAG.getConstant(4, getPointerTy()));
4796 Store = DAG.getStore(Op.getOperand(0),
4797 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4798 FIN, SV->getValue(), SV->getOffset());
4799 MemOps.push_back(Store);
4801 // Store ptr to overflow_arg_area
4802 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4803 DAG.getConstant(4, getPointerTy()));
4804 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4805 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4807 MemOps.push_back(Store);
4809 // Store ptr to reg_save_area.
4810 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4811 DAG.getConstant(8, getPointerTy()));
4812 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4813 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4815 MemOps.push_back(Store);
4816 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4819 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
4820 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
4821 SDOperand Chain = Op.getOperand(0);
4822 SDOperand DstPtr = Op.getOperand(1);
4823 SDOperand SrcPtr = Op.getOperand(2);
4824 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
4825 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
4827 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
4828 SrcSV->getValue(), SrcSV->getOffset());
4829 Chain = SrcPtr.getValue(1);
4830 for (unsigned i = 0; i < 3; ++i) {
4831 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
4832 SrcSV->getValue(), SrcSV->getOffset());
4833 Chain = Val.getValue(1);
4834 Chain = DAG.getStore(Chain, Val, DstPtr,
4835 DstSV->getValue(), DstSV->getOffset());
4838 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
4839 DAG.getConstant(8, getPointerTy()));
4840 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
4841 DAG.getConstant(8, getPointerTy()));
4847 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4848 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4850 default: return SDOperand(); // Don't custom lower most intrinsics.
4851 // Comparison intrinsics.
4852 case Intrinsic::x86_sse_comieq_ss:
4853 case Intrinsic::x86_sse_comilt_ss:
4854 case Intrinsic::x86_sse_comile_ss:
4855 case Intrinsic::x86_sse_comigt_ss:
4856 case Intrinsic::x86_sse_comige_ss:
4857 case Intrinsic::x86_sse_comineq_ss:
4858 case Intrinsic::x86_sse_ucomieq_ss:
4859 case Intrinsic::x86_sse_ucomilt_ss:
4860 case Intrinsic::x86_sse_ucomile_ss:
4861 case Intrinsic::x86_sse_ucomigt_ss:
4862 case Intrinsic::x86_sse_ucomige_ss:
4863 case Intrinsic::x86_sse_ucomineq_ss:
4864 case Intrinsic::x86_sse2_comieq_sd:
4865 case Intrinsic::x86_sse2_comilt_sd:
4866 case Intrinsic::x86_sse2_comile_sd:
4867 case Intrinsic::x86_sse2_comigt_sd:
4868 case Intrinsic::x86_sse2_comige_sd:
4869 case Intrinsic::x86_sse2_comineq_sd:
4870 case Intrinsic::x86_sse2_ucomieq_sd:
4871 case Intrinsic::x86_sse2_ucomilt_sd:
4872 case Intrinsic::x86_sse2_ucomile_sd:
4873 case Intrinsic::x86_sse2_ucomigt_sd:
4874 case Intrinsic::x86_sse2_ucomige_sd:
4875 case Intrinsic::x86_sse2_ucomineq_sd: {
4877 ISD::CondCode CC = ISD::SETCC_INVALID;
4880 case Intrinsic::x86_sse_comieq_ss:
4881 case Intrinsic::x86_sse2_comieq_sd:
4885 case Intrinsic::x86_sse_comilt_ss:
4886 case Intrinsic::x86_sse2_comilt_sd:
4890 case Intrinsic::x86_sse_comile_ss:
4891 case Intrinsic::x86_sse2_comile_sd:
4895 case Intrinsic::x86_sse_comigt_ss:
4896 case Intrinsic::x86_sse2_comigt_sd:
4900 case Intrinsic::x86_sse_comige_ss:
4901 case Intrinsic::x86_sse2_comige_sd:
4905 case Intrinsic::x86_sse_comineq_ss:
4906 case Intrinsic::x86_sse2_comineq_sd:
4910 case Intrinsic::x86_sse_ucomieq_ss:
4911 case Intrinsic::x86_sse2_ucomieq_sd:
4912 Opc = X86ISD::UCOMI;
4915 case Intrinsic::x86_sse_ucomilt_ss:
4916 case Intrinsic::x86_sse2_ucomilt_sd:
4917 Opc = X86ISD::UCOMI;
4920 case Intrinsic::x86_sse_ucomile_ss:
4921 case Intrinsic::x86_sse2_ucomile_sd:
4922 Opc = X86ISD::UCOMI;
4925 case Intrinsic::x86_sse_ucomigt_ss:
4926 case Intrinsic::x86_sse2_ucomigt_sd:
4927 Opc = X86ISD::UCOMI;
4930 case Intrinsic::x86_sse_ucomige_ss:
4931 case Intrinsic::x86_sse2_ucomige_sd:
4932 Opc = X86ISD::UCOMI;
4935 case Intrinsic::x86_sse_ucomineq_ss:
4936 case Intrinsic::x86_sse2_ucomineq_sd:
4937 Opc = X86ISD::UCOMI;
4943 SDOperand LHS = Op.getOperand(1);
4944 SDOperand RHS = Op.getOperand(2);
4945 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4947 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
4948 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
4949 DAG.getConstant(X86CC, MVT::i8), Cond);
4950 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4955 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4956 // Depths > 0 not supported yet!
4957 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4960 // Just load the return address
4961 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4962 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4965 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4966 // Depths > 0 not supported yet!
4967 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4970 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4971 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4972 DAG.getConstant(4, getPointerTy()));
4975 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
4976 SelectionDAG &DAG) {
4977 // Is not yet supported on x86-64
4978 if (Subtarget->is64Bit())
4981 return DAG.getConstant(8, getPointerTy());
4984 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
4986 assert(!Subtarget->is64Bit() &&
4987 "Lowering of eh_return builtin is not supported yet on x86-64");
4989 MachineFunction &MF = DAG.getMachineFunction();
4990 SDOperand Chain = Op.getOperand(0);
4991 SDOperand Offset = Op.getOperand(1);
4992 SDOperand Handler = Op.getOperand(2);
4994 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
4997 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
4998 DAG.getConstant(-4UL, getPointerTy()));
4999 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5000 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5001 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5002 MF.addLiveOut(X86::ECX);
5004 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5005 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5008 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
5009 SelectionDAG &DAG) {
5010 SDOperand Root = Op.getOperand(0);
5011 SDOperand Trmp = Op.getOperand(1); // trampoline
5012 SDOperand FPtr = Op.getOperand(2); // nested function
5013 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
5015 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
5017 if (Subtarget->is64Bit()) {
5018 return SDOperand(); // not yet supported
5020 Function *Func = (Function *)
5021 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5022 unsigned CC = Func->getCallingConv();
5027 assert(0 && "Unsupported calling convention");
5028 case CallingConv::C:
5029 case CallingConv::X86_StdCall: {
5030 // Pass 'nest' parameter in ECX.
5031 // Must be kept in sync with X86CallingConv.td
5034 // Check that ECX wasn't needed by an 'inreg' parameter.
5035 const FunctionType *FTy = Func->getFunctionType();
5036 const ParamAttrsList *Attrs = Func->getParamAttrs();
5038 if (Attrs && !Func->isVarArg()) {
5039 unsigned InRegCount = 0;
5042 for (FunctionType::param_iterator I = FTy->param_begin(),
5043 E = FTy->param_end(); I != E; ++I, ++Idx)
5044 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
5045 // FIXME: should only count parameters that are lowered to integers.
5046 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5048 if (InRegCount > 2) {
5049 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5055 case CallingConv::X86_FastCall:
5056 // Pass 'nest' parameter in EAX.
5057 // Must be kept in sync with X86CallingConv.td
5062 const X86InstrInfo *TII =
5063 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5065 SDOperand OutChains[4];
5066 SDOperand Addr, Disp;
5068 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5069 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5071 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5072 unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
5073 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5074 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
5076 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5077 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
5078 TrmpSV->getOffset() + 1, false, 1);
5080 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5081 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5082 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5083 TrmpSV->getValue() + 5, TrmpSV->getOffset());
5085 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5086 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
5087 TrmpSV->getOffset() + 6, false, 1);
5090 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5091 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5095 SDOperand X86TargetLowering::LowerFLT_ROUNDS(SDOperand Op, SelectionDAG &DAG) {
5097 The rounding mode is in bits 11:10 of FPSR, and has the following
5104 FLT_ROUNDS, on the other hand, expects the following:
5111 To perform the conversion, we do:
5112 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5115 MachineFunction &MF = DAG.getMachineFunction();
5116 const TargetMachine &TM = MF.getTarget();
5117 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5118 unsigned StackAlignment = TFI.getStackAlignment();
5119 MVT::ValueType VT = Op.getValueType();
5121 // Save FP Control Word to stack slot
5122 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5123 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5125 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5126 DAG.getEntryNode(), StackSlot);
5128 // Load FP Control Word from stack slot
5129 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5131 // Transform as necessary
5133 DAG.getNode(ISD::SRL, MVT::i16,
5134 DAG.getNode(ISD::AND, MVT::i16,
5135 CWD, DAG.getConstant(0x800, MVT::i16)),
5136 DAG.getConstant(11, MVT::i8));
5138 DAG.getNode(ISD::SRL, MVT::i16,
5139 DAG.getNode(ISD::AND, MVT::i16,
5140 CWD, DAG.getConstant(0x400, MVT::i16)),
5141 DAG.getConstant(9, MVT::i8));
5144 DAG.getNode(ISD::AND, MVT::i16,
5145 DAG.getNode(ISD::ADD, MVT::i16,
5146 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5147 DAG.getConstant(1, MVT::i16)),
5148 DAG.getConstant(3, MVT::i16));
5151 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5152 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5155 /// LowerOperation - Provide custom lowering hooks for some operations.
5157 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5158 switch (Op.getOpcode()) {
5159 default: assert(0 && "Should not custom lower this!");
5160 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5161 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5162 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5163 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5164 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5165 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5166 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5167 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5168 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5169 case ISD::SHL_PARTS:
5170 case ISD::SRA_PARTS:
5171 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5172 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5173 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5174 case ISD::FABS: return LowerFABS(Op, DAG);
5175 case ISD::FNEG: return LowerFNEG(Op, DAG);
5176 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5177 case ISD::SETCC: return LowerSETCC(Op, DAG);
5178 case ISD::SELECT: return LowerSELECT(Op, DAG);
5179 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5180 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5181 case ISD::CALL: return LowerCALL(Op, DAG);
5182 case ISD::RET: return LowerRET(Op, DAG);
5183 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5184 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5185 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5186 case ISD::VASTART: return LowerVASTART(Op, DAG);
5187 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5188 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5189 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5190 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5191 case ISD::FRAME_TO_ARGS_OFFSET:
5192 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5193 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5194 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5195 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5196 case ISD::FLT_ROUNDS: return LowerFLT_ROUNDS(Op, DAG);
5199 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5200 case ISD::READCYCLECOUNTER:
5201 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5205 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5206 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5207 switch (N->getOpcode()) {
5208 default: assert(0 && "Should not custom lower this!");
5209 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5210 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5214 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5216 default: return NULL;
5217 case X86ISD::SHLD: return "X86ISD::SHLD";
5218 case X86ISD::SHRD: return "X86ISD::SHRD";
5219 case X86ISD::FAND: return "X86ISD::FAND";
5220 case X86ISD::FOR: return "X86ISD::FOR";
5221 case X86ISD::FXOR: return "X86ISD::FXOR";
5222 case X86ISD::FSRL: return "X86ISD::FSRL";
5223 case X86ISD::FILD: return "X86ISD::FILD";
5224 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5225 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5226 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5227 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5228 case X86ISD::FLD: return "X86ISD::FLD";
5229 case X86ISD::FST: return "X86ISD::FST";
5230 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5231 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5232 case X86ISD::CALL: return "X86ISD::CALL";
5233 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5234 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5235 case X86ISD::CMP: return "X86ISD::CMP";
5236 case X86ISD::COMI: return "X86ISD::COMI";
5237 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5238 case X86ISD::SETCC: return "X86ISD::SETCC";
5239 case X86ISD::CMOV: return "X86ISD::CMOV";
5240 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5241 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5242 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5243 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5244 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5245 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5246 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5247 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5248 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5249 case X86ISD::FMAX: return "X86ISD::FMAX";
5250 case X86ISD::FMIN: return "X86ISD::FMIN";
5251 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5252 case X86ISD::FRCP: return "X86ISD::FRCP";
5253 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5254 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5255 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5256 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5257 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5261 // isLegalAddressingMode - Return true if the addressing mode represented
5262 // by AM is legal for this target, for a load/store of the specified type.
5263 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5264 const Type *Ty) const {
5265 // X86 supports extremely general addressing modes.
5267 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5268 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5272 // We can only fold this if we don't need an extra load.
5273 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5276 // X86-64 only supports addr of globals in small code model.
5277 if (Subtarget->is64Bit()) {
5278 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5280 // If lower 4G is not available, then we must use rip-relative addressing.
5281 if (AM.BaseOffs || AM.Scale > 1)
5292 // These scales always work.
5297 // These scales are formed with basereg+scalereg. Only accept if there is
5302 default: // Other stuff never works.
5310 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5311 if (!Ty1->isInteger() || !Ty2->isInteger())
5313 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5314 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5315 if (NumBits1 <= NumBits2)
5317 return Subtarget->is64Bit() || NumBits1 < 64;
5320 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5321 MVT::ValueType VT2) const {
5322 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5324 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5325 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5326 if (NumBits1 <= NumBits2)
5328 return Subtarget->is64Bit() || NumBits1 < 64;
5331 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5332 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5333 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5334 /// are assumed to be legal.
5336 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5337 // Only do shuffles on 128-bit vector types for now.
5338 if (MVT::getSizeInBits(VT) == 64) return false;
5339 return (Mask.Val->getNumOperands() <= 4 ||
5340 isIdentityMask(Mask.Val) ||
5341 isIdentityMask(Mask.Val, true) ||
5342 isSplatMask(Mask.Val) ||
5343 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5344 X86::isUNPCKLMask(Mask.Val) ||
5345 X86::isUNPCKHMask(Mask.Val) ||
5346 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5347 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5350 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5352 SelectionDAG &DAG) const {
5353 unsigned NumElts = BVOps.size();
5354 // Only do shuffles on 128-bit vector types for now.
5355 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5356 if (NumElts == 2) return true;
5358 return (isMOVLMask(&BVOps[0], 4) ||
5359 isCommutedMOVL(&BVOps[0], 4, true) ||
5360 isSHUFPMask(&BVOps[0], 4) ||
5361 isCommutedSHUFP(&BVOps[0], 4));
5366 //===----------------------------------------------------------------------===//
5367 // X86 Scheduler Hooks
5368 //===----------------------------------------------------------------------===//
5371 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5372 MachineBasicBlock *BB) {
5373 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5374 switch (MI->getOpcode()) {
5375 default: assert(false && "Unexpected instr type to insert");
5376 case X86::CMOV_FR32:
5377 case X86::CMOV_FR64:
5378 case X86::CMOV_V4F32:
5379 case X86::CMOV_V2F64:
5380 case X86::CMOV_V2I64: {
5381 // To "insert" a SELECT_CC instruction, we actually have to insert the
5382 // diamond control-flow pattern. The incoming instruction knows the
5383 // destination vreg to set, the condition code register to branch on, the
5384 // true/false values to select between, and a branch opcode to use.
5385 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5386 ilist<MachineBasicBlock>::iterator It = BB;
5392 // cmpTY ccX, r1, r2
5394 // fallthrough --> copy0MBB
5395 MachineBasicBlock *thisMBB = BB;
5396 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5397 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5399 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5400 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5401 MachineFunction *F = BB->getParent();
5402 F->getBasicBlockList().insert(It, copy0MBB);
5403 F->getBasicBlockList().insert(It, sinkMBB);
5404 // Update machine-CFG edges by first adding all successors of the current
5405 // block to the new block which will contain the Phi node for the select.
5406 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5407 e = BB->succ_end(); i != e; ++i)
5408 sinkMBB->addSuccessor(*i);
5409 // Next, remove all successors of the current block, and add the true
5410 // and fallthrough blocks as its successors.
5411 while(!BB->succ_empty())
5412 BB->removeSuccessor(BB->succ_begin());
5413 BB->addSuccessor(copy0MBB);
5414 BB->addSuccessor(sinkMBB);
5417 // %FalseValue = ...
5418 // # fallthrough to sinkMBB
5421 // Update machine-CFG edges
5422 BB->addSuccessor(sinkMBB);
5425 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5428 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5429 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5430 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5432 delete MI; // The pseudo instruction is gone now.
5436 case X86::FP32_TO_INT16_IN_MEM:
5437 case X86::FP32_TO_INT32_IN_MEM:
5438 case X86::FP32_TO_INT64_IN_MEM:
5439 case X86::FP64_TO_INT16_IN_MEM:
5440 case X86::FP64_TO_INT32_IN_MEM:
5441 case X86::FP64_TO_INT64_IN_MEM:
5442 case X86::FP80_TO_INT16_IN_MEM:
5443 case X86::FP80_TO_INT32_IN_MEM:
5444 case X86::FP80_TO_INT64_IN_MEM: {
5445 // Change the floating point control register to use "round towards zero"
5446 // mode when truncating to an integer value.
5447 MachineFunction *F = BB->getParent();
5448 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5449 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5451 // Load the old value of the high byte of the control word...
5453 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5454 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5456 // Set the high part to be round to zero...
5457 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5460 // Reload the modified control word now...
5461 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5463 // Restore the memory image of control word to original value
5464 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5467 // Get the X86 opcode to use.
5469 switch (MI->getOpcode()) {
5470 default: assert(0 && "illegal opcode!");
5471 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5472 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5473 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5474 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5475 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5476 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5477 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5478 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5479 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5483 MachineOperand &Op = MI->getOperand(0);
5484 if (Op.isRegister()) {
5485 AM.BaseType = X86AddressMode::RegBase;
5486 AM.Base.Reg = Op.getReg();
5488 AM.BaseType = X86AddressMode::FrameIndexBase;
5489 AM.Base.FrameIndex = Op.getFrameIndex();
5491 Op = MI->getOperand(1);
5492 if (Op.isImmediate())
5493 AM.Scale = Op.getImm();
5494 Op = MI->getOperand(2);
5495 if (Op.isImmediate())
5496 AM.IndexReg = Op.getImm();
5497 Op = MI->getOperand(3);
5498 if (Op.isGlobalAddress()) {
5499 AM.GV = Op.getGlobal();
5501 AM.Disp = Op.getImm();
5503 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5504 .addReg(MI->getOperand(4).getReg());
5506 // Reload the original control word now.
5507 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5509 delete MI; // The pseudo instruction is gone now.
5515 //===----------------------------------------------------------------------===//
5516 // X86 Optimization Hooks
5517 //===----------------------------------------------------------------------===//
5519 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5521 uint64_t &KnownZero,
5523 const SelectionDAG &DAG,
5524 unsigned Depth) const {
5525 unsigned Opc = Op.getOpcode();
5526 assert((Opc >= ISD::BUILTIN_OP_END ||
5527 Opc == ISD::INTRINSIC_WO_CHAIN ||
5528 Opc == ISD::INTRINSIC_W_CHAIN ||
5529 Opc == ISD::INTRINSIC_VOID) &&
5530 "Should use MaskedValueIsZero if you don't know whether Op"
5531 " is a target node!");
5533 KnownZero = KnownOne = 0; // Don't know anything.
5537 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5542 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5543 /// element of the result of the vector shuffle.
5544 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5545 MVT::ValueType VT = N->getValueType(0);
5546 SDOperand PermMask = N->getOperand(2);
5547 unsigned NumElems = PermMask.getNumOperands();
5548 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5550 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5552 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5553 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5554 SDOperand Idx = PermMask.getOperand(i);
5555 if (Idx.getOpcode() == ISD::UNDEF)
5556 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5557 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5562 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5563 /// node is a GlobalAddress + an offset.
5564 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5565 unsigned Opc = N->getOpcode();
5566 if (Opc == X86ISD::Wrapper) {
5567 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5568 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5571 } else if (Opc == ISD::ADD) {
5572 SDOperand N1 = N->getOperand(0);
5573 SDOperand N2 = N->getOperand(1);
5574 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5575 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5577 Offset += V->getSignExtended();
5580 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5581 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5583 Offset += V->getSignExtended();
5591 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5593 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5594 MachineFrameInfo *MFI) {
5595 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5598 SDOperand Loc = N->getOperand(1);
5599 SDOperand BaseLoc = Base->getOperand(1);
5600 if (Loc.getOpcode() == ISD::FrameIndex) {
5601 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5603 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5604 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5605 int FS = MFI->getObjectSize(FI);
5606 int BFS = MFI->getObjectSize(BFI);
5607 if (FS != BFS || FS != Size) return false;
5608 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5610 GlobalValue *GV1 = NULL;
5611 GlobalValue *GV2 = NULL;
5612 int64_t Offset1 = 0;
5613 int64_t Offset2 = 0;
5614 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5615 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5616 if (isGA1 && isGA2 && GV1 == GV2)
5617 return Offset1 == (Offset2 + Dist*Size);
5623 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5624 const X86Subtarget *Subtarget) {
5627 if (isGAPlusOffset(Base, GV, Offset))
5628 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5630 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5631 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5633 // Fixed objects do not specify alignment, however the offsets are known.
5634 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5635 (MFI->getObjectOffset(BFI) % 16) == 0);
5637 return MFI->getObjectAlignment(BFI) >= 16;
5643 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5644 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5645 /// if the load addresses are consecutive, non-overlapping, and in the right
5647 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5648 const X86Subtarget *Subtarget) {
5649 MachineFunction &MF = DAG.getMachineFunction();
5650 MachineFrameInfo *MFI = MF.getFrameInfo();
5651 MVT::ValueType VT = N->getValueType(0);
5652 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5653 SDOperand PermMask = N->getOperand(2);
5654 int NumElems = (int)PermMask.getNumOperands();
5655 SDNode *Base = NULL;
5656 for (int i = 0; i < NumElems; ++i) {
5657 SDOperand Idx = PermMask.getOperand(i);
5658 if (Idx.getOpcode() == ISD::UNDEF) {
5659 if (!Base) return SDOperand();
5662 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5663 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5667 else if (!isConsecutiveLoad(Arg.Val, Base,
5668 i, MVT::getSizeInBits(EVT)/8,MFI))
5673 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5674 LoadSDNode *LD = cast<LoadSDNode>(Base);
5676 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5677 LD->getSrcValueOffset(), LD->isVolatile());
5679 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5680 LD->getSrcValueOffset(), LD->isVolatile(),
5681 LD->getAlignment());
5685 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5686 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5687 const X86Subtarget *Subtarget) {
5688 SDOperand Cond = N->getOperand(0);
5690 // If we have SSE[12] support, try to form min/max nodes.
5691 if (Subtarget->hasSSE2() &&
5692 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5693 if (Cond.getOpcode() == ISD::SETCC) {
5694 // Get the LHS/RHS of the select.
5695 SDOperand LHS = N->getOperand(1);
5696 SDOperand RHS = N->getOperand(2);
5697 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5699 unsigned Opcode = 0;
5700 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5703 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5706 if (!UnsafeFPMath) break;
5708 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5710 Opcode = X86ISD::FMIN;
5713 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5716 if (!UnsafeFPMath) break;
5718 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5720 Opcode = X86ISD::FMAX;
5723 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5726 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5729 if (!UnsafeFPMath) break;
5731 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5733 Opcode = X86ISD::FMIN;
5736 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5739 if (!UnsafeFPMath) break;
5741 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5743 Opcode = X86ISD::FMAX;
5749 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5758 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5759 DAGCombinerInfo &DCI) const {
5760 SelectionDAG &DAG = DCI.DAG;
5761 switch (N->getOpcode()) {
5763 case ISD::VECTOR_SHUFFLE:
5764 return PerformShuffleCombine(N, DAG, Subtarget);
5766 return PerformSELECTCombine(N, DAG, Subtarget);
5772 //===----------------------------------------------------------------------===//
5773 // X86 Inline Assembly Support
5774 //===----------------------------------------------------------------------===//
5776 /// getConstraintType - Given a constraint letter, return the type of
5777 /// constraint it is for this target.
5778 X86TargetLowering::ConstraintType
5779 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
5780 if (Constraint.size() == 1) {
5781 switch (Constraint[0]) {
5790 return C_RegisterClass;
5795 return TargetLowering::getConstraintType(Constraint);
5798 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5799 /// vector. If it is invalid, don't add anything to Ops.
5800 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
5802 std::vector<SDOperand>&Ops,
5803 SelectionDAG &DAG) {
5804 SDOperand Result(0, 0);
5806 switch (Constraint) {
5809 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5810 if (C->getValue() <= 31) {
5811 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
5818 if (C->getValue() <= 255) {
5819 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
5825 // Literal immediates are always ok.
5826 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
5827 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
5831 // If we are in non-pic codegen mode, we allow the address of a global (with
5832 // an optional displacement) to be used with 'i'.
5833 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
5836 // Match either (GA) or (GA+C)
5838 Offset = GA->getOffset();
5839 } else if (Op.getOpcode() == ISD::ADD) {
5840 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5841 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5843 Offset = GA->getOffset()+C->getValue();
5845 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5846 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
5848 Offset = GA->getOffset()+C->getValue();
5855 // If addressing this global requires a load (e.g. in PIC mode), we can't
5857 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
5861 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5867 // Otherwise, not valid for this mode.
5873 Ops.push_back(Result);
5876 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5879 std::vector<unsigned> X86TargetLowering::
5880 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5881 MVT::ValueType VT) const {
5882 if (Constraint.size() == 1) {
5883 // FIXME: not handling fp-stack yet!
5884 switch (Constraint[0]) { // GCC X86 Constraint Letters
5885 default: break; // Unknown constraint letter
5886 case 'A': // EAX/EDX
5887 if (VT == MVT::i32 || VT == MVT::i64)
5888 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5890 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5893 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5894 else if (VT == MVT::i16)
5895 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5896 else if (VT == MVT::i8)
5897 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5898 else if (VT == MVT::i64)
5899 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
5904 return std::vector<unsigned>();
5907 std::pair<unsigned, const TargetRegisterClass*>
5908 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5909 MVT::ValueType VT) const {
5910 // First, see if this is a constraint that directly corresponds to an LLVM
5912 if (Constraint.size() == 1) {
5913 // GCC Constraint Letters
5914 switch (Constraint[0]) {
5916 case 'r': // GENERAL_REGS
5917 case 'R': // LEGACY_REGS
5918 case 'l': // INDEX_REGS
5919 if (VT == MVT::i64 && Subtarget->is64Bit())
5920 return std::make_pair(0U, X86::GR64RegisterClass);
5922 return std::make_pair(0U, X86::GR32RegisterClass);
5923 else if (VT == MVT::i16)
5924 return std::make_pair(0U, X86::GR16RegisterClass);
5925 else if (VT == MVT::i8)
5926 return std::make_pair(0U, X86::GR8RegisterClass);
5928 case 'y': // MMX_REGS if MMX allowed.
5929 if (!Subtarget->hasMMX()) break;
5930 return std::make_pair(0U, X86::VR64RegisterClass);
5932 case 'Y': // SSE_REGS if SSE2 allowed
5933 if (!Subtarget->hasSSE2()) break;
5935 case 'x': // SSE_REGS if SSE1 allowed
5936 if (!Subtarget->hasSSE1()) break;
5940 // Scalar SSE types.
5943 return std::make_pair(0U, X86::FR32RegisterClass);
5946 return std::make_pair(0U, X86::FR64RegisterClass);
5954 return std::make_pair(0U, X86::VR128RegisterClass);
5960 // Use the default implementation in TargetLowering to convert the register
5961 // constraint into a member of a register class.
5962 std::pair<unsigned, const TargetRegisterClass*> Res;
5963 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5965 // Not found as a standard register?
5966 if (Res.second == 0) {
5967 // GCC calls "st(0)" just plain "st".
5968 if (StringsEqualNoCase("{st}", Constraint)) {
5969 Res.first = X86::ST0;
5970 Res.second = X86::RFP80RegisterClass;
5976 // Otherwise, check to see if this is a register class of the wrong value
5977 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5978 // turn into {ax},{dx}.
5979 if (Res.second->hasType(VT))
5980 return Res; // Correct type already, nothing to do.
5982 // All of the single-register GCC register classes map their values onto
5983 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5984 // really want an 8-bit or 32-bit register, map to the appropriate register
5985 // class and return the appropriate register.
5986 if (Res.second != X86::GR16RegisterClass)
5989 if (VT == MVT::i8) {
5990 unsigned DestReg = 0;
5991 switch (Res.first) {
5993 case X86::AX: DestReg = X86::AL; break;
5994 case X86::DX: DestReg = X86::DL; break;
5995 case X86::CX: DestReg = X86::CL; break;
5996 case X86::BX: DestReg = X86::BL; break;
5999 Res.first = DestReg;
6000 Res.second = Res.second = X86::GR8RegisterClass;
6002 } else if (VT == MVT::i32) {
6003 unsigned DestReg = 0;
6004 switch (Res.first) {
6006 case X86::AX: DestReg = X86::EAX; break;
6007 case X86::DX: DestReg = X86::EDX; break;
6008 case X86::CX: DestReg = X86::ECX; break;
6009 case X86::BX: DestReg = X86::EBX; break;
6010 case X86::SI: DestReg = X86::ESI; break;
6011 case X86::DI: DestReg = X86::EDI; break;
6012 case X86::BP: DestReg = X86::EBP; break;
6013 case X86::SP: DestReg = X86::ESP; break;
6016 Res.first = DestReg;
6017 Res.second = Res.second = X86::GR32RegisterClass;
6019 } else if (VT == MVT::i64) {
6020 unsigned DestReg = 0;
6021 switch (Res.first) {
6023 case X86::AX: DestReg = X86::RAX; break;
6024 case X86::DX: DestReg = X86::RDX; break;
6025 case X86::CX: DestReg = X86::RCX; break;
6026 case X86::BX: DestReg = X86::RBX; break;
6027 case X86::SI: DestReg = X86::RSI; break;
6028 case X86::DI: DestReg = X86::RDI; break;
6029 case X86::BP: DestReg = X86::RBP; break;
6030 case X86::SP: DestReg = X86::RSP; break;
6033 Res.first = DestReg;
6034 Res.second = Res.second = X86::GR64RegisterClass;