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/BitVector.h"
27 #include "llvm/ADT/VectorExtras.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/SelectionDAG.h"
34 #include "llvm/CodeGen/SSARegMap.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/StringExtras.h"
40 #include "llvm/ParameterAttributes.h"
43 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
44 : TargetLowering(TM) {
45 Subtarget = &TM.getSubtarget<X86Subtarget>();
46 X86ScalarSSEf64 = Subtarget->hasSSE2();
47 X86ScalarSSEf32 = Subtarget->hasSSE1();
48 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
51 RegInfo = TM.getRegisterInfo();
53 // Set up the TargetLowering object.
55 // X86 is weird, it always uses i8 for shift amounts and setcc results.
56 setShiftAmountType(MVT::i8);
57 setSetCCResultType(MVT::i8);
58 setSetCCResultContents(ZeroOrOneSetCCResult);
59 setSchedulingPreference(SchedulingForRegPressure);
60 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
61 setStackPointerRegisterToSaveRestore(X86StackPtr);
63 if (Subtarget->isTargetDarwin()) {
64 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
65 setUseUnderscoreSetJmp(false);
66 setUseUnderscoreLongJmp(false);
67 } else if (Subtarget->isTargetMingw()) {
68 // MS runtime is weird: it exports _setjmp, but longjmp!
69 setUseUnderscoreSetJmp(true);
70 setUseUnderscoreLongJmp(false);
72 setUseUnderscoreSetJmp(true);
73 setUseUnderscoreLongJmp(true);
76 // Set up the register classes.
77 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
78 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
79 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
80 if (Subtarget->is64Bit())
81 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
83 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
85 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
87 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
88 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
89 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
91 if (Subtarget->is64Bit()) {
92 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
96 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
97 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
99 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
102 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
104 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
105 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
106 // SSE has no i16 to fp conversion, only i32
107 if (X86ScalarSSEf32) {
108 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
109 // f32 and f64 cases are Legal, f80 case is not
110 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
113 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
116 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
117 // are Legal, f80 is custom lowered.
118 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
119 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
121 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
123 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
124 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
126 if (X86ScalarSSEf32) {
127 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
128 // f32 and f64 cases are Legal, f80 case is not
129 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
132 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
135 // Handle FP_TO_UINT by promoting the destination to a larger signed
137 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
138 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
139 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
141 if (Subtarget->is64Bit()) {
142 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
143 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
145 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
146 // Expand FP_TO_UINT into a select.
147 // FIXME: We would like to use a Custom expander here eventually to do
148 // the optimal thing for SSE vs. the default expansion in the legalizer.
149 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
151 // With SSE3 we can use fisttpll to convert to a signed i64.
152 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
155 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
156 if (!X86ScalarSSEf64) {
157 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
158 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
161 // Scalar integer multiply, multiply-high, divide, and remainder are
162 // lowered to use operations that produce two results, to match the
163 // available instructions. This exposes the two-result form to trivial
164 // CSE, which is able to combine x/y and x%y into a single instruction,
165 // for example. The single-result multiply instructions are introduced
166 // in X86ISelDAGToDAG.cpp, after CSE, for uses where the the high part
168 setOperationAction(ISD::MUL , MVT::i8 , Expand);
169 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
170 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
171 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
172 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
173 setOperationAction(ISD::SREM , MVT::i8 , Expand);
174 setOperationAction(ISD::UREM , MVT::i8 , Expand);
175 setOperationAction(ISD::MUL , MVT::i16 , Expand);
176 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
177 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
178 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
179 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
180 setOperationAction(ISD::SREM , MVT::i16 , Expand);
181 setOperationAction(ISD::UREM , MVT::i16 , Expand);
182 setOperationAction(ISD::MUL , MVT::i32 , Expand);
183 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
184 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
185 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
186 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
187 setOperationAction(ISD::SREM , MVT::i32 , Expand);
188 setOperationAction(ISD::UREM , MVT::i32 , Expand);
189 setOperationAction(ISD::MUL , MVT::i64 , Expand);
190 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
191 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
192 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
193 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
194 setOperationAction(ISD::SREM , MVT::i64 , Expand);
195 setOperationAction(ISD::UREM , MVT::i64 , Expand);
197 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
198 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
199 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
200 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
201 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
202 if (Subtarget->is64Bit())
203 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
204 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
205 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
206 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
207 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
208 setOperationAction(ISD::FREM , MVT::f64 , Expand);
209 setOperationAction(ISD::FLT_ROUNDS , MVT::i32 , Custom);
211 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
212 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
213 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
214 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
215 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
216 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
217 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
218 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
219 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
220 if (Subtarget->is64Bit()) {
221 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
222 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
223 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
226 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
227 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
229 // These should be promoted to a larger select which is supported.
230 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
231 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
232 // X86 wants to expand cmov itself.
233 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
234 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
235 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
236 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
237 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
238 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
239 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
240 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
241 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
242 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
243 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
244 if (Subtarget->is64Bit()) {
245 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
246 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
248 // X86 ret instruction may pop stack.
249 setOperationAction(ISD::RET , MVT::Other, Custom);
250 if (!Subtarget->is64Bit())
251 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
254 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
255 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
256 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
257 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
258 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
259 if (Subtarget->is64Bit()) {
260 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
261 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
262 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
263 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
265 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
266 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
267 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
268 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
269 // X86 wants to expand memset / memcpy itself.
270 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
271 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
273 // Use the default ISD::LOCATION expansion.
274 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
275 // FIXME - use subtarget debug flags
276 if (!Subtarget->isTargetDarwin() &&
277 !Subtarget->isTargetELF() &&
278 !Subtarget->isTargetCygMing())
279 setOperationAction(ISD::LABEL, MVT::Other, Expand);
281 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
282 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
283 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
284 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
285 if (Subtarget->is64Bit()) {
287 setExceptionPointerRegister(X86::RAX);
288 setExceptionSelectorRegister(X86::RDX);
290 setExceptionPointerRegister(X86::EAX);
291 setExceptionSelectorRegister(X86::EDX);
293 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
295 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
297 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
298 setOperationAction(ISD::VASTART , MVT::Other, Custom);
299 setOperationAction(ISD::VAARG , MVT::Other, Expand);
300 setOperationAction(ISD::VAEND , MVT::Other, Expand);
301 if (Subtarget->is64Bit())
302 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
304 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
306 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
307 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
308 if (Subtarget->is64Bit())
309 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
310 if (Subtarget->isTargetCygMing())
311 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
313 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
315 if (X86ScalarSSEf64) {
316 // f32 and f64 use SSE.
317 // Set up the FP register classes.
318 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
319 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
321 // Use ANDPD to simulate FABS.
322 setOperationAction(ISD::FABS , MVT::f64, Custom);
323 setOperationAction(ISD::FABS , MVT::f32, Custom);
325 // Use XORP to simulate FNEG.
326 setOperationAction(ISD::FNEG , MVT::f64, Custom);
327 setOperationAction(ISD::FNEG , MVT::f32, Custom);
329 // Use ANDPD and ORPD to simulate FCOPYSIGN.
330 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
331 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
333 // We don't support sin/cos/fmod
334 setOperationAction(ISD::FSIN , MVT::f64, Expand);
335 setOperationAction(ISD::FCOS , MVT::f64, Expand);
336 setOperationAction(ISD::FREM , MVT::f64, Expand);
337 setOperationAction(ISD::FSIN , MVT::f32, Expand);
338 setOperationAction(ISD::FCOS , MVT::f32, Expand);
339 setOperationAction(ISD::FREM , MVT::f32, Expand);
341 // Expand FP immediates into loads from the stack, except for the special
343 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
344 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
345 addLegalFPImmediate(APFloat(+0.0)); // xorpd
346 addLegalFPImmediate(APFloat(+0.0f)); // xorps
348 // Conversions to long double (in X87) go through memory.
349 setConvertAction(MVT::f32, MVT::f80, Expand);
350 setConvertAction(MVT::f64, MVT::f80, Expand);
352 // Conversions from long double (in X87) go through memory.
353 setConvertAction(MVT::f80, MVT::f32, Expand);
354 setConvertAction(MVT::f80, MVT::f64, Expand);
355 } else if (X86ScalarSSEf32) {
356 // Use SSE for f32, x87 for f64.
357 // Set up the FP register classes.
358 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
359 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
361 // Use ANDPS to simulate FABS.
362 setOperationAction(ISD::FABS , MVT::f32, Custom);
364 // Use XORP to simulate FNEG.
365 setOperationAction(ISD::FNEG , MVT::f32, Custom);
367 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
369 // Use ANDPS and ORPS to simulate FCOPYSIGN.
370 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
371 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
373 // We don't support sin/cos/fmod
374 setOperationAction(ISD::FSIN , MVT::f32, Expand);
375 setOperationAction(ISD::FCOS , MVT::f32, Expand);
376 setOperationAction(ISD::FREM , MVT::f32, Expand);
378 // Expand FP immediates into loads from the stack, except for the special
380 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
381 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
382 addLegalFPImmediate(APFloat(+0.0f)); // xorps
383 addLegalFPImmediate(APFloat(+0.0)); // FLD0
384 addLegalFPImmediate(APFloat(+1.0)); // FLD1
385 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
386 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
388 // SSE->x87 conversions go through memory.
389 setConvertAction(MVT::f32, MVT::f64, Expand);
390 setConvertAction(MVT::f32, MVT::f80, Expand);
392 // x87->SSE truncations need to go through memory.
393 setConvertAction(MVT::f80, MVT::f32, Expand);
394 setConvertAction(MVT::f64, MVT::f32, Expand);
395 // And x87->x87 truncations also.
396 setConvertAction(MVT::f80, MVT::f64, Expand);
399 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
400 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
403 // f32 and f64 in x87.
404 // Set up the FP register classes.
405 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
406 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
408 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
409 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
410 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
411 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
413 // Floating truncations need to go through memory.
414 setConvertAction(MVT::f80, MVT::f32, Expand);
415 setConvertAction(MVT::f64, MVT::f32, Expand);
416 setConvertAction(MVT::f80, MVT::f64, Expand);
419 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
420 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
423 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
424 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
425 addLegalFPImmediate(APFloat(+0.0)); // FLD0
426 addLegalFPImmediate(APFloat(+1.0)); // FLD1
427 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
428 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
429 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
430 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
431 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
432 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
435 // Long double always uses X87.
436 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
437 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
439 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
441 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
442 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
445 // Always use a library call for pow.
446 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
447 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
448 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
450 // First set operation action for all vector types to expand. Then we
451 // will selectively turn on ones that can be effectively codegen'd.
452 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
453 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
454 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
455 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
456 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
457 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
458 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
459 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
460 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
461 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
462 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
463 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
464 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
465 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
466 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
467 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
468 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
469 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
470 setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
471 setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
472 setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
473 setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
474 setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
475 setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
476 setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
477 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
478 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
479 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
480 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
481 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
482 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
483 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
484 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
485 setOperationAction(ISD::SHL, (MVT::ValueType)VT, Expand);
486 setOperationAction(ISD::SRA, (MVT::ValueType)VT, Expand);
487 setOperationAction(ISD::SRL, (MVT::ValueType)VT, Expand);
488 setOperationAction(ISD::ROTL, (MVT::ValueType)VT, Expand);
489 setOperationAction(ISD::ROTR, (MVT::ValueType)VT, Expand);
490 setOperationAction(ISD::BSWAP, (MVT::ValueType)VT, Expand);
493 if (Subtarget->hasMMX()) {
494 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
495 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
496 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
497 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
499 // FIXME: add MMX packed arithmetics
501 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
502 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
503 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
504 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
506 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
507 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
508 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
509 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
511 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
512 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
514 setOperationAction(ISD::AND, MVT::v8i8, Promote);
515 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
516 setOperationAction(ISD::AND, MVT::v4i16, Promote);
517 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
518 setOperationAction(ISD::AND, MVT::v2i32, Promote);
519 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
520 setOperationAction(ISD::AND, MVT::v1i64, Legal);
522 setOperationAction(ISD::OR, MVT::v8i8, Promote);
523 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
524 setOperationAction(ISD::OR, MVT::v4i16, Promote);
525 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
526 setOperationAction(ISD::OR, MVT::v2i32, Promote);
527 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
528 setOperationAction(ISD::OR, MVT::v1i64, Legal);
530 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
531 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
532 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
533 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
534 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
535 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
536 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
538 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
539 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
540 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
541 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
542 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
543 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
544 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
546 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
547 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
548 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
549 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
551 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
552 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
553 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
554 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
556 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
557 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
558 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Custom);
559 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
562 if (Subtarget->hasSSE1()) {
563 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
565 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
566 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
567 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
568 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
569 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
570 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
571 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
572 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
573 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
574 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
575 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
578 if (Subtarget->hasSSE2()) {
579 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
580 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
581 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
582 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
583 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
585 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
586 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
587 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
588 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
589 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
590 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
591 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
592 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
593 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
594 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
595 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
596 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
597 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
598 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
599 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
601 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
602 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
603 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
604 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
605 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
606 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
608 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
609 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
610 // Do not attempt to custom lower non-power-of-2 vectors
611 if (!isPowerOf2_32(MVT::getVectorNumElements(VT)))
613 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
614 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
615 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
617 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
619 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
620 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
621 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
622 if (Subtarget->is64Bit())
623 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
625 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
626 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
627 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
628 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
629 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
630 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
631 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
632 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
633 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
634 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
635 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
636 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
639 // Custom lower v2i64 and v2f64 selects.
640 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
641 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
642 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
643 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
646 // We want to custom lower some of our intrinsics.
647 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
649 // We have target-specific dag combine patterns for the following nodes:
650 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
651 setTargetDAGCombine(ISD::SELECT);
653 computeRegisterProperties();
655 // FIXME: These should be based on subtarget info. Plus, the values should
656 // be smaller when we are in optimizing for size mode.
657 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
658 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
659 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
660 allowUnalignedMemoryAccesses = true; // x86 supports it!
664 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
666 SDOperand X86TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
667 SelectionDAG &DAG) const {
668 if (usesGlobalOffsetTable())
669 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
670 if (!Subtarget->isPICStyleRIPRel())
671 return DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy());
675 //===----------------------------------------------------------------------===//
676 // Return Value Calling Convention Implementation
677 //===----------------------------------------------------------------------===//
679 #include "X86GenCallingConv.inc"
681 /// GetPossiblePreceedingTailCall - Get preceeding X86ISD::TAILCALL node if it
682 /// exists skip possible ISD:TokenFactor.
683 static SDOperand GetPossiblePreceedingTailCall(SDOperand Chain) {
684 if (Chain.getOpcode()==X86ISD::TAILCALL) {
686 } else if (Chain.getOpcode()==ISD::TokenFactor) {
687 if (Chain.getNumOperands() &&
688 Chain.getOperand(0).getOpcode()==X86ISD::TAILCALL)
689 return Chain.getOperand(0);
694 /// LowerRET - Lower an ISD::RET node.
695 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
696 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
698 SmallVector<CCValAssign, 16> RVLocs;
699 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
700 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
701 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
702 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
704 // If this is the first return lowered for this function, add the regs to the
705 // liveout set for the function.
706 if (DAG.getMachineFunction().liveout_empty()) {
707 for (unsigned i = 0; i != RVLocs.size(); ++i)
708 if (RVLocs[i].isRegLoc())
709 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
711 SDOperand Chain = Op.getOperand(0);
713 // Handle tail call return.
714 Chain = GetPossiblePreceedingTailCall(Chain);
715 if (Chain.getOpcode() == X86ISD::TAILCALL) {
716 SDOperand TailCall = Chain;
717 SDOperand TargetAddress = TailCall.getOperand(1);
718 SDOperand StackAdjustment = TailCall.getOperand(2);
719 assert ( ((TargetAddress.getOpcode() == ISD::Register &&
720 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::ECX ||
721 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
722 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
723 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
724 "Expecting an global address, external symbol, or register");
725 assert( StackAdjustment.getOpcode() == ISD::Constant &&
726 "Expecting a const value");
728 SmallVector<SDOperand,8> Operands;
729 Operands.push_back(Chain.getOperand(0));
730 Operands.push_back(TargetAddress);
731 Operands.push_back(StackAdjustment);
732 // Copy registers used by the call. Last operand is a flag so it is not
734 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
735 Operands.push_back(Chain.getOperand(i));
737 return DAG.getNode(X86ISD::TC_RETURN, MVT::Other, &Operands[0],
744 // Copy the result values into the output registers.
745 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
746 RVLocs[0].getLocReg() != X86::ST0) {
747 for (unsigned i = 0; i != RVLocs.size(); ++i) {
748 CCValAssign &VA = RVLocs[i];
749 assert(VA.isRegLoc() && "Can only return in registers!");
750 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
752 Flag = Chain.getValue(1);
755 // We need to handle a destination of ST0 specially, because it isn't really
757 SDOperand Value = Op.getOperand(1);
759 // If this is an FP return with ScalarSSE, we need to move the value from
760 // an XMM register onto the fp-stack.
761 if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
762 (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
765 // If this is a load into a scalarsse value, don't store the loaded value
766 // back to the stack, only to reload it: just replace the scalar-sse load.
767 if (ISD::isNON_EXTLoad(Value.Val) &&
768 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
769 Chain = Value.getOperand(0);
770 MemLoc = Value.getOperand(1);
772 // Spill the value to memory and reload it into top of stack.
773 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
774 MachineFunction &MF = DAG.getMachineFunction();
775 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
776 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
777 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
779 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
780 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
781 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
782 Chain = Value.getValue(1);
785 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
786 SDOperand Ops[] = { Chain, Value };
787 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
788 Flag = Chain.getValue(1);
791 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
793 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
795 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
799 /// LowerCallResult - Lower the result values of an ISD::CALL into the
800 /// appropriate copies out of appropriate physical registers. This assumes that
801 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
802 /// being lowered. The returns a SDNode with the same number of values as the
804 SDNode *X86TargetLowering::
805 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
806 unsigned CallingConv, SelectionDAG &DAG) {
808 // Assign locations to each value returned by this call.
809 SmallVector<CCValAssign, 16> RVLocs;
810 bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
811 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
812 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
814 SmallVector<SDOperand, 8> ResultVals;
816 // Copy all of the result registers out of their specified physreg.
817 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
818 for (unsigned i = 0; i != RVLocs.size(); ++i) {
819 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
820 RVLocs[i].getValVT(), InFlag).getValue(1);
821 InFlag = Chain.getValue(2);
822 ResultVals.push_back(Chain.getValue(0));
825 // Copies from the FP stack are special, as ST0 isn't a valid register
826 // before the fp stackifier runs.
828 // Copy ST0 into an RFP register with FP_GET_RESULT.
829 SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
830 SDOperand GROps[] = { Chain, InFlag };
831 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
832 Chain = RetVal.getValue(1);
833 InFlag = RetVal.getValue(2);
835 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
837 if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
838 (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
840 const Value *SrcVal = 0;
841 int SrcValOffset = 0;
843 // Determine where to store the value. If the call result is directly
844 // used by a store, see if we can store directly into the location. In
845 // this case, we'll end up producing a fst + movss[load] + movss[store] to
846 // the same location, and the two movss's will be nuked as dead. This
847 // optimizes common things like "*D = atof(..)" to not need an
848 // intermediate stack slot.
849 if (SDOperand(TheCall, 0).hasOneUse() &&
850 SDOperand(TheCall, 1).hasOneUse()) {
851 // Ok, we have one use of the value and one use of the chain. See if
852 // they are the same node: a store.
853 if (StoreSDNode *N = dyn_cast<StoreSDNode>(*TheCall->use_begin())) {
854 if (N->getChain().Val == TheCall && N->getValue().Val == TheCall &&
855 !N->isVolatile() && !N->isTruncatingStore() &&
856 N->getAddressingMode() == ISD::UNINDEXED) {
857 StoreLoc = N->getBasePtr();
858 SrcVal = N->getSrcValue();
859 SrcValOffset = N->getSrcValueOffset();
864 // If we weren't able to optimize the result, just create a temporary
866 if (StoreLoc.Val == 0) {
867 MachineFunction &MF = DAG.getMachineFunction();
868 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
869 StoreLoc = DAG.getFrameIndex(SSFI, getPointerTy());
872 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
873 // shouldn't be necessary except that RFP cannot be live across
874 // multiple blocks (which could happen if a select gets lowered into
875 // multiple blocks and scheduled in between them). When stackifier is
876 // fixed, they can be uncoupled.
878 Chain, RetVal, StoreLoc, DAG.getValueType(RVLocs[0].getValVT()), InFlag
880 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
881 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain,
882 StoreLoc, SrcVal, SrcValOffset);
883 Chain = RetVal.getValue(1);
885 ResultVals.push_back(RetVal);
888 // Merge everything together with a MERGE_VALUES node.
889 ResultVals.push_back(Chain);
890 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
891 &ResultVals[0], ResultVals.size()).Val;
895 //===----------------------------------------------------------------------===//
896 // C & StdCall & Fast Calling Convention implementation
897 //===----------------------------------------------------------------------===//
898 // StdCall calling convention seems to be standard for many Windows' API
899 // routines and around. It differs from C calling convention just a little:
900 // callee should clean up the stack, not caller. Symbols should be also
901 // decorated in some fancy way :) It doesn't support any vector arguments.
902 // For info on fast calling convention see Fast Calling Convention (tail call)
903 // implementation LowerX86_32FastCCCallTo.
905 /// AddLiveIn - This helper function adds the specified physical register to the
906 /// MachineFunction as a live in value. It also creates a corresponding virtual
908 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
909 const TargetRegisterClass *RC) {
910 assert(RC->contains(PReg) && "Not the correct regclass!");
911 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
912 MF.addLiveIn(PReg, VReg);
916 // align stack arguments according to platform alignment needed for tail calls
917 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG& DAG);
919 SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
920 const CCValAssign &VA,
921 MachineFrameInfo *MFI,
922 SDOperand Root, unsigned i) {
923 // Create the nodes corresponding to a load from this parameter slot.
924 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
925 VA.getLocMemOffset());
926 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
928 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
930 if (Flags & ISD::ParamFlags::ByVal)
933 return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
936 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
938 unsigned NumArgs = Op.Val->getNumValues() - 1;
939 MachineFunction &MF = DAG.getMachineFunction();
940 MachineFrameInfo *MFI = MF.getFrameInfo();
941 SDOperand Root = Op.getOperand(0);
942 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
943 unsigned CC = MF.getFunction()->getCallingConv();
944 // Assign locations to all of the incoming arguments.
945 SmallVector<CCValAssign, 16> ArgLocs;
946 CCState CCInfo(CC, isVarArg,
947 getTargetMachine(), ArgLocs);
948 // Check for possible tail call calling convention.
949 if (CC == CallingConv::Fast && PerformTailCallOpt)
950 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_TailCall);
952 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
954 SmallVector<SDOperand, 8> ArgValues;
955 unsigned LastVal = ~0U;
956 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
957 CCValAssign &VA = ArgLocs[i];
958 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
960 assert(VA.getValNo() != LastVal &&
961 "Don't support value assigned to multiple locs yet");
962 LastVal = VA.getValNo();
965 MVT::ValueType RegVT = VA.getLocVT();
966 TargetRegisterClass *RC;
967 if (RegVT == MVT::i32)
968 RC = X86::GR32RegisterClass;
970 assert(MVT::isVector(RegVT));
971 RC = X86::VR128RegisterClass;
974 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
975 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
977 // If this is an 8 or 16-bit value, it is really passed promoted to 32
978 // bits. Insert an assert[sz]ext to capture this, then truncate to the
980 if (VA.getLocInfo() == CCValAssign::SExt)
981 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
982 DAG.getValueType(VA.getValVT()));
983 else if (VA.getLocInfo() == CCValAssign::ZExt)
984 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
985 DAG.getValueType(VA.getValVT()));
987 if (VA.getLocInfo() != CCValAssign::Full)
988 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
990 ArgValues.push_back(ArgValue);
992 assert(VA.isMemLoc());
993 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
997 unsigned StackSize = CCInfo.getNextStackOffset();
998 // align stack specially for tail calls
999 if (CC==CallingConv::Fast)
1000 StackSize = GetAlignedArgumentStackSize(StackSize,DAG);
1002 ArgValues.push_back(Root);
1004 // If the function takes variable number of arguments, make a frame index for
1005 // the start of the first vararg value... for expansion of llvm.va_start.
1007 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1009 // Tail call calling convention (CallingConv::Fast) does not support varargs.
1010 assert( !(isVarArg && CC == CallingConv::Fast) &&
1011 "CallingConv::Fast does not support varargs.");
1013 if (isStdCall && !isVarArg &&
1014 (CC==CallingConv::Fast && PerformTailCallOpt || CC!=CallingConv::Fast)) {
1015 BytesToPopOnReturn = StackSize; // Callee pops everything..
1016 BytesCallerReserves = 0;
1018 BytesToPopOnReturn = 0; // Callee pops nothing.
1020 // If this is an sret function, the return should pop the hidden pointer.
1022 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() &
1023 ISD::ParamFlags::StructReturn))
1024 BytesToPopOnReturn = 4;
1026 BytesCallerReserves = StackSize;
1029 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1031 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1032 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1034 // Return the new list of results.
1035 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1036 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1039 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
1041 SDOperand Chain = Op.getOperand(0);
1042 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1043 SDOperand Callee = Op.getOperand(4);
1044 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1046 // Analyze operands of the call, assigning locations to each operand.
1047 SmallVector<CCValAssign, 16> ArgLocs;
1048 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1049 if(CC==CallingConv::Fast && PerformTailCallOpt)
1050 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1052 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
1054 // Get a count of how many bytes are to be pushed on the stack.
1055 unsigned NumBytes = CCInfo.getNextStackOffset();
1056 if (CC==CallingConv::Fast)
1057 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1059 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1061 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1062 SmallVector<SDOperand, 8> MemOpChains;
1066 // Walk the register/memloc assignments, inserting copies/loads.
1067 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1068 CCValAssign &VA = ArgLocs[i];
1069 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1071 // Promote the value if needed.
1072 switch (VA.getLocInfo()) {
1073 default: assert(0 && "Unknown loc info!");
1074 case CCValAssign::Full: break;
1075 case CCValAssign::SExt:
1076 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1078 case CCValAssign::ZExt:
1079 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1081 case CCValAssign::AExt:
1082 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1086 if (VA.isRegLoc()) {
1087 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1089 assert(VA.isMemLoc());
1090 if (StackPtr.Val == 0)
1091 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1093 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1098 // If the first argument is an sret pointer, remember it.
1099 bool isSRet = NumOps &&
1100 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() &
1101 ISD::ParamFlags::StructReturn);
1103 if (!MemOpChains.empty())
1104 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1105 &MemOpChains[0], MemOpChains.size());
1107 // Build a sequence of copy-to-reg nodes chained together with token chain
1108 // and flag operands which copy the outgoing args into registers.
1110 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1111 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1113 InFlag = Chain.getValue(1);
1116 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1118 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1119 Subtarget->isPICStyleGOT()) {
1120 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1121 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1123 InFlag = Chain.getValue(1);
1126 // If the callee is a GlobalAddress node (quite common, every direct call is)
1127 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1128 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1129 // We should use extra load for direct calls to dllimported functions in
1131 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1132 getTargetMachine(), true))
1133 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1134 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1135 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1137 // Returns a chain & a flag for retval copy to use.
1138 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1139 SmallVector<SDOperand, 8> Ops;
1140 Ops.push_back(Chain);
1141 Ops.push_back(Callee);
1143 // Add argument registers to the end of the list so that they are known live
1145 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1146 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1147 RegsToPass[i].second.getValueType()));
1149 // Add an implicit use GOT pointer in EBX.
1150 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1151 Subtarget->isPICStyleGOT())
1152 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1155 Ops.push_back(InFlag);
1157 Chain = DAG.getNode(X86ISD::CALL, NodeTys, &Ops[0], Ops.size());
1158 InFlag = Chain.getValue(1);
1160 // Create the CALLSEQ_END node.
1161 unsigned NumBytesForCalleeToPush = 0;
1163 if (CC == CallingConv::X86_StdCall ||
1164 (CC == CallingConv::Fast && PerformTailCallOpt)) {
1166 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1168 NumBytesForCalleeToPush = NumBytes;
1169 assert(!(isVarArg && CC==CallingConv::Fast) &&
1170 "CallingConv::Fast does not support varargs.");
1172 // If this is is a call to a struct-return function, the callee
1173 // pops the hidden struct pointer, so we have to push it back.
1174 // This is common for Darwin/X86, Linux & Mingw32 targets.
1175 NumBytesForCalleeToPush = isSRet ? 4 : 0;
1178 Chain = DAG.getCALLSEQ_END(Chain,
1179 DAG.getConstant(NumBytes, getPointerTy()),
1180 DAG.getConstant(NumBytesForCalleeToPush,
1183 InFlag = Chain.getValue(1);
1185 // Handle result values, copying them out of physregs into vregs that we
1187 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1191 //===----------------------------------------------------------------------===//
1192 // FastCall Calling Convention implementation
1193 //===----------------------------------------------------------------------===//
1195 // The X86 'fastcall' calling convention passes up to two integer arguments in
1196 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1197 // and requires that the callee pop its arguments off the stack (allowing proper
1198 // tail calls), and has the same return value conventions as C calling convs.
1200 // This calling convention always arranges for the callee pop value to be 8n+4
1201 // bytes, which is needed for tail recursion elimination and stack alignment
1204 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
1205 MachineFunction &MF = DAG.getMachineFunction();
1206 MachineFrameInfo *MFI = MF.getFrameInfo();
1207 SDOperand Root = Op.getOperand(0);
1208 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1210 // Assign locations to all of the incoming arguments.
1211 SmallVector<CCValAssign, 16> ArgLocs;
1212 CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
1213 getTargetMachine(), ArgLocs);
1214 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
1216 SmallVector<SDOperand, 8> ArgValues;
1217 unsigned LastVal = ~0U;
1218 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1219 CCValAssign &VA = ArgLocs[i];
1220 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1222 assert(VA.getValNo() != LastVal &&
1223 "Don't support value assigned to multiple locs yet");
1224 LastVal = VA.getValNo();
1226 if (VA.isRegLoc()) {
1227 MVT::ValueType RegVT = VA.getLocVT();
1228 TargetRegisterClass *RC;
1229 if (RegVT == MVT::i32)
1230 RC = X86::GR32RegisterClass;
1232 assert(MVT::isVector(RegVT));
1233 RC = X86::VR128RegisterClass;
1236 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1237 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1239 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1240 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1242 if (VA.getLocInfo() == CCValAssign::SExt)
1243 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1244 DAG.getValueType(VA.getValVT()));
1245 else if (VA.getLocInfo() == CCValAssign::ZExt)
1246 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1247 DAG.getValueType(VA.getValVT()));
1249 if (VA.getLocInfo() != CCValAssign::Full)
1250 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1252 ArgValues.push_back(ArgValue);
1254 assert(VA.isMemLoc());
1255 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1259 ArgValues.push_back(Root);
1261 unsigned StackSize = CCInfo.getNextStackOffset();
1263 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1264 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1265 // arguments and the arguments after the retaddr has been pushed are
1267 if ((StackSize & 7) == 0)
1271 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1272 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1273 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
1274 BytesCallerReserves = 0;
1276 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1277 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1279 // Return the new list of results.
1280 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1281 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1285 X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
1286 const SDOperand &StackPtr,
1287 const CCValAssign &VA,
1290 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1291 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1292 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1293 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1294 if (Flags & ISD::ParamFlags::ByVal) {
1295 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1296 ISD::ParamFlags::ByValAlignOffs);
1298 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1299 ISD::ParamFlags::ByValSizeOffs;
1301 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1302 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1303 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i32);
1305 return DAG.getMemcpy(Chain, PtrOff, Arg, SizeNode, AlignNode,
1308 return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1312 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1314 SDOperand Chain = Op.getOperand(0);
1315 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1316 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1317 SDOperand Callee = Op.getOperand(4);
1319 // Analyze operands of the call, assigning locations to each operand.
1320 SmallVector<CCValAssign, 16> ArgLocs;
1321 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1322 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
1324 // Get a count of how many bytes are to be pushed on the stack.
1325 unsigned NumBytes = CCInfo.getNextStackOffset();
1327 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
1328 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1329 // arguments and the arguments after the retaddr has been pushed are
1331 if ((NumBytes & 7) == 0)
1335 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1337 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1338 SmallVector<SDOperand, 8> MemOpChains;
1342 // Walk the register/memloc assignments, inserting copies/loads.
1343 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1344 CCValAssign &VA = ArgLocs[i];
1345 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1347 // Promote the value if needed.
1348 switch (VA.getLocInfo()) {
1349 default: assert(0 && "Unknown loc info!");
1350 case CCValAssign::Full: break;
1351 case CCValAssign::SExt:
1352 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1354 case CCValAssign::ZExt:
1355 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1357 case CCValAssign::AExt:
1358 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1362 if (VA.isRegLoc()) {
1363 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1365 assert(VA.isMemLoc());
1366 if (StackPtr.Val == 0)
1367 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1369 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1374 if (!MemOpChains.empty())
1375 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1376 &MemOpChains[0], MemOpChains.size());
1378 // Build a sequence of copy-to-reg nodes chained together with token chain
1379 // and flag operands which copy the outgoing args into registers.
1381 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1382 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1384 InFlag = Chain.getValue(1);
1387 // If the callee is a GlobalAddress node (quite common, every direct call is)
1388 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1389 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1390 // We should use extra load for direct calls to dllimported functions in
1392 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1393 getTargetMachine(), true))
1394 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1395 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1396 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1398 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1400 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1401 Subtarget->isPICStyleGOT()) {
1402 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1403 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1405 InFlag = Chain.getValue(1);
1408 // Returns a chain & a flag for retval copy to use.
1409 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1410 SmallVector<SDOperand, 8> Ops;
1411 Ops.push_back(Chain);
1412 Ops.push_back(Callee);
1414 // Add argument registers to the end of the list so that they are known live
1416 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1417 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1418 RegsToPass[i].second.getValueType()));
1420 // Add an implicit use GOT pointer in EBX.
1421 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1422 Subtarget->isPICStyleGOT())
1423 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1426 Ops.push_back(InFlag);
1428 assert(isTailCall==false && "no tail call here");
1429 Chain = DAG.getNode(X86ISD::CALL,
1430 NodeTys, &Ops[0], Ops.size());
1431 InFlag = Chain.getValue(1);
1433 // Returns a flag for retval copy to use.
1434 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1436 Ops.push_back(Chain);
1437 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1438 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1439 Ops.push_back(InFlag);
1440 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1441 InFlag = Chain.getValue(1);
1443 // Handle result values, copying them out of physregs into vregs that we
1445 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1448 //===----------------------------------------------------------------------===//
1449 // Fast Calling Convention (tail call) implementation
1450 //===----------------------------------------------------------------------===//
1452 // Like std call, callee cleans arguments, convention except that ECX is
1453 // reserved for storing the tail called function address. Only 2 registers are
1454 // free for argument passing (inreg). Tail call optimization is performed
1456 // * tailcallopt is enabled
1457 // * caller/callee are fastcc
1458 // * elf/pic is disabled OR
1459 // * elf/pic enabled + callee is in module + callee has
1460 // visibility protected or hidden
1461 // To keep the stack aligned according to platform abi the function
1462 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1463 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1464 // If a tail called function callee has more arguments than the caller the
1465 // caller needs to make sure that there is room to move the RETADDR to. This is
1466 // achieved by reserving an area the size of the argument delta right after the
1467 // original REtADDR, but before the saved framepointer or the spilled registers
1468 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1480 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1481 /// for a 16 byte align requirement.
1482 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1483 SelectionDAG& DAG) {
1484 if (PerformTailCallOpt) {
1485 MachineFunction &MF = DAG.getMachineFunction();
1486 const TargetMachine &TM = MF.getTarget();
1487 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1488 unsigned StackAlignment = TFI.getStackAlignment();
1489 uint64_t AlignMask = StackAlignment - 1;
1490 int64_t Offset = StackSize;
1491 unsigned SlotSize = Subtarget->is64Bit() ? 8 : 4;
1492 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1493 // Number smaller than 12 so just add the difference.
1494 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1496 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1497 Offset = ((~AlignMask) & Offset) + StackAlignment +
1498 (StackAlignment-SlotSize);
1505 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1506 /// following the call is a return. A function is eligible if caller/callee
1507 /// calling conventions match, currently only fastcc supports tail calls, and
1508 /// the function CALL is immediatly followed by a RET.
1509 bool X86TargetLowering::IsEligibleForTailCallOptimization(SDOperand Call,
1511 SelectionDAG& DAG) const {
1512 if (!PerformTailCallOpt)
1515 // Check whether CALL node immediatly preceeds the RET node and whether the
1516 // return uses the result of the node or is a void return.
1517 unsigned NumOps = Ret.getNumOperands();
1519 (Ret.getOperand(0) == SDOperand(Call.Val,1) ||
1520 Ret.getOperand(0) == SDOperand(Call.Val,0))) ||
1522 Ret.getOperand(0) == SDOperand(Call.Val,Call.Val->getNumValues()-1) &&
1523 Ret.getOperand(1) == SDOperand(Call.Val,0))) {
1524 MachineFunction &MF = DAG.getMachineFunction();
1525 unsigned CallerCC = MF.getFunction()->getCallingConv();
1526 unsigned CalleeCC = cast<ConstantSDNode>(Call.getOperand(1))->getValue();
1527 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1528 SDOperand Callee = Call.getOperand(4);
1529 // On elf/pic %ebx needs to be livein.
1530 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1531 !Subtarget->isPICStyleGOT())
1534 // Can only do local tail calls with PIC.
1535 GlobalValue * GV = 0;
1536 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1538 (GV = G->getGlobal()) &&
1539 (GV->hasHiddenVisibility() || GV->hasProtectedVisibility()))
1547 SDOperand X86TargetLowering::LowerX86_TailCallTo(SDOperand Op,
1550 SDOperand Chain = Op.getOperand(0);
1551 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1552 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1553 SDOperand Callee = Op.getOperand(4);
1554 bool is64Bit = Subtarget->is64Bit();
1556 assert(isTailCall && PerformTailCallOpt && "Should only emit tail calls.");
1558 // Analyze operands of the call, assigning locations to each operand.
1559 SmallVector<CCValAssign, 16> ArgLocs;
1560 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1562 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1564 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_TailCall);
1567 // Lower arguments at fp - stackoffset + fpdiff.
1568 MachineFunction &MF = DAG.getMachineFunction();
1570 unsigned NumBytesToBePushed =
1571 GetAlignedArgumentStackSize(CCInfo.getNextStackOffset(), DAG);
1573 unsigned NumBytesCallerPushed =
1574 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1575 int FPDiff = NumBytesCallerPushed - NumBytesToBePushed;
1577 // Set the delta of movement of the returnaddr stackslot.
1578 // But only set if delta is greater than previous delta.
1579 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1580 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1583 getCALLSEQ_START(Chain, DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1585 // Adjust the Return address stack slot.
1586 SDOperand RetAddrFrIdx, NewRetAddrFrIdx;
1588 MVT::ValueType VT = is64Bit ? MVT::i64 : MVT::i32;
1589 RetAddrFrIdx = getReturnAddressFrameIndex(DAG);
1590 // Load the "old" Return address.
1592 DAG.getLoad(VT, Chain,RetAddrFrIdx, NULL, 0);
1593 // Calculate the new stack slot for the return address.
1594 int SlotSize = is64Bit ? 8 : 4;
1595 int NewReturnAddrFI =
1596 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1597 NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1598 Chain = SDOperand(RetAddrFrIdx.Val, 1);
1601 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1602 SmallVector<SDOperand, 8> MemOpChains;
1603 SmallVector<SDOperand, 8> MemOpChains2;
1604 SDOperand FramePtr, StackPtr;
1609 // Walk the register/memloc assignments, inserting copies/loads. Lower
1610 // arguments first to the stack slot where they would normally - in case of a
1611 // normal function call - be.
1612 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1613 CCValAssign &VA = ArgLocs[i];
1614 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1616 // Promote the value if needed.
1617 switch (VA.getLocInfo()) {
1618 default: assert(0 && "Unknown loc info!");
1619 case CCValAssign::Full: break;
1620 case CCValAssign::SExt:
1621 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1623 case CCValAssign::ZExt:
1624 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1626 case CCValAssign::AExt:
1627 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1631 if (VA.isRegLoc()) {
1632 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1634 assert(VA.isMemLoc());
1635 if (StackPtr.Val == 0)
1636 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1638 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1643 if (!MemOpChains.empty())
1644 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1645 &MemOpChains[0], MemOpChains.size());
1647 // Build a sequence of copy-to-reg nodes chained together with token chain
1648 // and flag operands which copy the outgoing args into registers.
1650 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1651 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1653 InFlag = Chain.getValue(1);
1655 InFlag = SDOperand();
1657 // Copy from stack slots to stack slot of a tail called function. This needs
1658 // to be done because if we would lower the arguments directly to their real
1659 // stack slot we might end up overwriting each other.
1660 // TODO: To make this more efficient (sometimes saving a store/load) we could
1661 // analyse the arguments and emit this store/load/store sequence only for
1662 // arguments which would be overwritten otherwise.
1663 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1664 CCValAssign &VA = ArgLocs[i];
1665 if (!VA.isRegLoc()) {
1666 SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
1667 unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
1669 // Get source stack slot.
1670 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1671 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1672 // Create frame index.
1673 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1674 uint32_t OpSize = (MVT::getSizeInBits(VA.getLocVT())+7)/8;
1675 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1676 FIN = DAG.getFrameIndex(FI, MVT::i32);
1677 if (Flags & ISD::ParamFlags::ByVal) {
1678 // Copy relative to framepointer.
1679 unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
1680 ISD::ParamFlags::ByValAlignOffs);
1682 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1683 ISD::ParamFlags::ByValSizeOffs;
1685 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1686 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1687 SDOperand AlwaysInline = DAG.getConstant(1, MVT::i1);
1689 MemOpChains2.push_back(DAG.getMemcpy(Chain, FIN, PtrOff, SizeNode,
1690 AlignNode,AlwaysInline));
1692 SDOperand LoadedArg = DAG.getLoad(VA.getValVT(), Chain, PtrOff, NULL,0);
1693 // Store relative to framepointer.
1694 MemOpChains2.push_back(DAG.getStore(Chain, LoadedArg, FIN, NULL, 0));
1699 if (!MemOpChains2.empty())
1700 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1701 &MemOpChains2[0], MemOpChains.size());
1703 // Store the return address to the appropriate stack slot.
1705 Chain = DAG.getStore(Chain,RetAddrFrIdx, NewRetAddrFrIdx, NULL, 0);
1707 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1709 // Does not work with tail call since ebx is not restored correctly by
1710 // tailcaller. TODO: at least for x86 - verify for x86-64
1712 // If the callee is a GlobalAddress node (quite common, every direct call is)
1713 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1714 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1715 // We should use extra load for direct calls to dllimported functions in
1717 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1718 getTargetMachine(), true))
1719 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1720 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1721 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1723 assert(Callee.getOpcode() == ISD::LOAD &&
1724 "Function destination must be loaded into virtual register");
1725 unsigned Opc = is64Bit ? X86::R9 : X86::ECX;
1727 Chain = DAG.getCopyToReg(Chain,
1728 DAG.getRegister(Opc, getPointerTy()) ,
1730 Callee = DAG.getRegister(Opc, getPointerTy());
1731 // Add register as live out.
1732 DAG.getMachineFunction().addLiveOut(Opc);
1735 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1736 SmallVector<SDOperand, 8> Ops;
1738 Ops.push_back(Chain);
1739 Ops.push_back(DAG.getConstant(NumBytesToBePushed, getPointerTy()));
1740 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1742 Ops.push_back(InFlag);
1743 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1744 InFlag = Chain.getValue(1);
1746 // Returns a chain & a flag for retval copy to use.
1747 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1749 Ops.push_back(Chain);
1750 Ops.push_back(Callee);
1751 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1752 // Add argument registers to the end of the list so that they are known live
1754 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1755 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1756 RegsToPass[i].second.getValueType()));
1758 Ops.push_back(InFlag);
1759 assert(InFlag.Val &&
1760 "Flag must be set. Depend on flag being set in LowerRET");
1761 Chain = DAG.getNode(X86ISD::TAILCALL,
1762 Op.Val->getVTList(), &Ops[0], Ops.size());
1764 return SDOperand(Chain.Val, Op.ResNo);
1767 //===----------------------------------------------------------------------===//
1768 // X86-64 C Calling Convention implementation
1769 //===----------------------------------------------------------------------===//
1772 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1773 MachineFunction &MF = DAG.getMachineFunction();
1774 MachineFrameInfo *MFI = MF.getFrameInfo();
1775 SDOperand Root = Op.getOperand(0);
1776 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1777 unsigned CC= MF.getFunction()->getCallingConv();
1779 static const unsigned GPR64ArgRegs[] = {
1780 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1782 static const unsigned XMMArgRegs[] = {
1783 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1784 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1788 // Assign locations to all of the incoming arguments.
1789 SmallVector<CCValAssign, 16> ArgLocs;
1790 CCState CCInfo(CC, isVarArg,
1791 getTargetMachine(), ArgLocs);
1792 if (CC == CallingConv::Fast && PerformTailCallOpt)
1793 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_TailCall);
1795 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1797 SmallVector<SDOperand, 8> ArgValues;
1798 unsigned LastVal = ~0U;
1799 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1800 CCValAssign &VA = ArgLocs[i];
1801 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1803 assert(VA.getValNo() != LastVal &&
1804 "Don't support value assigned to multiple locs yet");
1805 LastVal = VA.getValNo();
1807 if (VA.isRegLoc()) {
1808 MVT::ValueType RegVT = VA.getLocVT();
1809 TargetRegisterClass *RC;
1810 if (RegVT == MVT::i32)
1811 RC = X86::GR32RegisterClass;
1812 else if (RegVT == MVT::i64)
1813 RC = X86::GR64RegisterClass;
1814 else if (RegVT == MVT::f32)
1815 RC = X86::FR32RegisterClass;
1816 else if (RegVT == MVT::f64)
1817 RC = X86::FR64RegisterClass;
1819 assert(MVT::isVector(RegVT));
1820 if (MVT::getSizeInBits(RegVT) == 64) {
1821 RC = X86::GR64RegisterClass; // MMX values are passed in GPRs.
1824 RC = X86::VR128RegisterClass;
1827 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1828 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1830 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1831 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1833 if (VA.getLocInfo() == CCValAssign::SExt)
1834 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1835 DAG.getValueType(VA.getValVT()));
1836 else if (VA.getLocInfo() == CCValAssign::ZExt)
1837 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1838 DAG.getValueType(VA.getValVT()));
1840 if (VA.getLocInfo() != CCValAssign::Full)
1841 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1843 // Handle MMX values passed in GPRs.
1844 if (RegVT != VA.getLocVT() && RC == X86::GR64RegisterClass &&
1845 MVT::getSizeInBits(RegVT) == 64)
1846 ArgValue = DAG.getNode(ISD::BIT_CONVERT, VA.getLocVT(), ArgValue);
1848 ArgValues.push_back(ArgValue);
1850 assert(VA.isMemLoc());
1851 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
1855 unsigned StackSize = CCInfo.getNextStackOffset();
1856 if (CC==CallingConv::Fast)
1857 StackSize =GetAlignedArgumentStackSize(StackSize, DAG);
1859 // If the function takes variable number of arguments, make a frame index for
1860 // the start of the first vararg value... for expansion of llvm.va_start.
1862 assert(CC!=CallingConv::Fast
1863 && "Var arg not supported with calling convention fastcc");
1864 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1865 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1867 // For X86-64, if there are vararg parameters that are passed via
1868 // registers, then we must store them to their spots on the stack so they
1869 // may be loaded by deferencing the result of va_next.
1870 VarArgsGPOffset = NumIntRegs * 8;
1871 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1872 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1873 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1875 // Store the integer parameter registers.
1876 SmallVector<SDOperand, 8> MemOps;
1877 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1878 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1879 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1880 for (; NumIntRegs != 6; ++NumIntRegs) {
1881 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1882 X86::GR64RegisterClass);
1883 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1884 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1885 MemOps.push_back(Store);
1886 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1887 DAG.getConstant(8, getPointerTy()));
1890 // Now store the XMM (fp + vector) parameter registers.
1891 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1892 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1893 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1894 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1895 X86::VR128RegisterClass);
1896 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1897 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1898 MemOps.push_back(Store);
1899 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1900 DAG.getConstant(16, getPointerTy()));
1902 if (!MemOps.empty())
1903 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1904 &MemOps[0], MemOps.size());
1907 ArgValues.push_back(Root);
1908 // Tail call convention (fastcc) needs callee pop.
1909 if (CC == CallingConv::Fast && PerformTailCallOpt) {
1910 BytesToPopOnReturn = StackSize; // Callee pops everything.
1911 BytesCallerReserves = 0;
1913 BytesToPopOnReturn = 0; // Callee pops nothing.
1914 BytesCallerReserves = StackSize;
1916 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1917 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1919 // Return the new list of results.
1920 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1921 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1925 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1927 SDOperand Chain = Op.getOperand(0);
1928 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1929 SDOperand Callee = Op.getOperand(4);
1931 // Analyze operands of the call, assigning locations to each operand.
1932 SmallVector<CCValAssign, 16> ArgLocs;
1933 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1934 if (CC==CallingConv::Fast && PerformTailCallOpt)
1935 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_TailCall);
1937 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1939 // Get a count of how many bytes are to be pushed on the stack.
1940 unsigned NumBytes = CCInfo.getNextStackOffset();
1941 if (CC == CallingConv::Fast)
1942 NumBytes = GetAlignedArgumentStackSize(NumBytes,DAG);
1944 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1946 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1947 SmallVector<SDOperand, 8> MemOpChains;
1951 // Walk the register/memloc assignments, inserting copies/loads.
1952 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1953 CCValAssign &VA = ArgLocs[i];
1954 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1956 // Promote the value if needed.
1957 switch (VA.getLocInfo()) {
1958 default: assert(0 && "Unknown loc info!");
1959 case CCValAssign::Full: break;
1960 case CCValAssign::SExt:
1961 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1963 case CCValAssign::ZExt:
1964 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1966 case CCValAssign::AExt:
1967 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1971 if (VA.isRegLoc()) {
1972 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1974 assert(VA.isMemLoc());
1975 if (StackPtr.Val == 0)
1976 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1978 MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
1983 if (!MemOpChains.empty())
1984 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1985 &MemOpChains[0], MemOpChains.size());
1987 // Build a sequence of copy-to-reg nodes chained together with token chain
1988 // and flag operands which copy the outgoing args into registers.
1990 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1991 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1993 InFlag = Chain.getValue(1);
1997 assert ( CallingConv::Fast != CC &&
1998 "Var args not supported with calling convention fastcc");
2000 // From AMD64 ABI document:
2001 // For calls that may call functions that use varargs or stdargs
2002 // (prototype-less calls or calls to functions containing ellipsis (...) in
2003 // the declaration) %al is used as hidden argument to specify the number
2004 // of SSE registers used. The contents of %al do not need to match exactly
2005 // the number of registers, but must be an ubound on the number of SSE
2006 // registers used and is in the range 0 - 8 inclusive.
2008 // Count the number of XMM registers allocated.
2009 static const unsigned XMMArgRegs[] = {
2010 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2011 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2013 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2015 Chain = DAG.getCopyToReg(Chain, X86::AL,
2016 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2017 InFlag = Chain.getValue(1);
2020 // If the callee is a GlobalAddress node (quite common, every direct call is)
2021 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
2022 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2023 // We should use extra load for direct calls to dllimported functions in
2025 if (getTargetMachine().getCodeModel() != CodeModel::Large
2026 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(),
2027 getTargetMachine(), true))
2028 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
2029 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2030 if (getTargetMachine().getCodeModel() != CodeModel::Large)
2031 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
2033 // Returns a chain & a flag for retval copy to use.
2034 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2035 SmallVector<SDOperand, 8> Ops;
2036 Ops.push_back(Chain);
2037 Ops.push_back(Callee);
2039 // Add argument registers to the end of the list so that they are known live
2041 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2042 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2043 RegsToPass[i].second.getValueType()));
2046 Ops.push_back(InFlag);
2048 Chain = DAG.getNode(X86ISD::CALL,
2049 NodeTys, &Ops[0], Ops.size());
2050 InFlag = Chain.getValue(1);
2051 int NumBytesForCalleeToPush = 0;
2052 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2053 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2055 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2057 // Returns a flag for retval copy to use.
2058 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2060 Ops.push_back(Chain);
2061 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
2062 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
2063 Ops.push_back(InFlag);
2064 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
2065 InFlag = Chain.getValue(1);
2067 // Handle result values, copying them out of physregs into vregs that we
2069 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
2073 //===----------------------------------------------------------------------===//
2074 // Other Lowering Hooks
2075 //===----------------------------------------------------------------------===//
2078 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2079 MachineFunction &MF = DAG.getMachineFunction();
2080 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2081 int ReturnAddrIndex = FuncInfo->getRAIndex();
2083 if (ReturnAddrIndex == 0) {
2084 // Set up a frame object for the return address.
2085 if (Subtarget->is64Bit())
2086 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
2088 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
2090 FuncInfo->setRAIndex(ReturnAddrIndex);
2093 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2098 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
2099 /// specific condition code. It returns a false if it cannot do a direct
2100 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
2102 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2103 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
2104 SelectionDAG &DAG) {
2105 X86CC = X86::COND_INVALID;
2107 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2108 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2109 // X > -1 -> X == 0, jump !sign.
2110 RHS = DAG.getConstant(0, RHS.getValueType());
2111 X86CC = X86::COND_NS;
2113 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2114 // X < 0 -> X == 0, jump on sign.
2115 X86CC = X86::COND_S;
2117 } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
2119 RHS = DAG.getConstant(0, RHS.getValueType());
2120 X86CC = X86::COND_LE;
2125 switch (SetCCOpcode) {
2127 case ISD::SETEQ: X86CC = X86::COND_E; break;
2128 case ISD::SETGT: X86CC = X86::COND_G; break;
2129 case ISD::SETGE: X86CC = X86::COND_GE; break;
2130 case ISD::SETLT: X86CC = X86::COND_L; break;
2131 case ISD::SETLE: X86CC = X86::COND_LE; break;
2132 case ISD::SETNE: X86CC = X86::COND_NE; break;
2133 case ISD::SETULT: X86CC = X86::COND_B; break;
2134 case ISD::SETUGT: X86CC = X86::COND_A; break;
2135 case ISD::SETULE: X86CC = X86::COND_BE; break;
2136 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2139 // On a floating point condition, the flags are set as follows:
2141 // 0 | 0 | 0 | X > Y
2142 // 0 | 0 | 1 | X < Y
2143 // 1 | 0 | 0 | X == Y
2144 // 1 | 1 | 1 | unordered
2146 switch (SetCCOpcode) {
2149 case ISD::SETEQ: X86CC = X86::COND_E; break;
2150 case ISD::SETOLT: Flip = true; // Fallthrough
2152 case ISD::SETGT: X86CC = X86::COND_A; break;
2153 case ISD::SETOLE: Flip = true; // Fallthrough
2155 case ISD::SETGE: X86CC = X86::COND_AE; break;
2156 case ISD::SETUGT: Flip = true; // Fallthrough
2158 case ISD::SETLT: X86CC = X86::COND_B; break;
2159 case ISD::SETUGE: Flip = true; // Fallthrough
2161 case ISD::SETLE: X86CC = X86::COND_BE; break;
2163 case ISD::SETNE: X86CC = X86::COND_NE; break;
2164 case ISD::SETUO: X86CC = X86::COND_P; break;
2165 case ISD::SETO: X86CC = X86::COND_NP; break;
2168 std::swap(LHS, RHS);
2171 return X86CC != X86::COND_INVALID;
2174 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2175 /// code. Current x86 isa includes the following FP cmov instructions:
2176 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2177 static bool hasFPCMov(unsigned X86CC) {
2193 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2194 /// true if Op is undef or if its value falls within the specified range (L, H].
2195 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2196 if (Op.getOpcode() == ISD::UNDEF)
2199 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2200 return (Val >= Low && Val < Hi);
2203 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2204 /// true if Op is undef or if its value equal to the specified value.
2205 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2206 if (Op.getOpcode() == ISD::UNDEF)
2208 return cast<ConstantSDNode>(Op)->getValue() == Val;
2211 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2212 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2213 bool X86::isPSHUFDMask(SDNode *N) {
2214 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2216 if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
2219 // Check if the value doesn't reference the second vector.
2220 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2221 SDOperand Arg = N->getOperand(i);
2222 if (Arg.getOpcode() == ISD::UNDEF) continue;
2223 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2224 if (cast<ConstantSDNode>(Arg)->getValue() >= e)
2231 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2232 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2233 bool X86::isPSHUFHWMask(SDNode *N) {
2234 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2236 if (N->getNumOperands() != 8)
2239 // Lower quadword copied in order.
2240 for (unsigned i = 0; i != 4; ++i) {
2241 SDOperand Arg = N->getOperand(i);
2242 if (Arg.getOpcode() == ISD::UNDEF) continue;
2243 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2244 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2248 // Upper quadword shuffled.
2249 for (unsigned i = 4; i != 8; ++i) {
2250 SDOperand Arg = N->getOperand(i);
2251 if (Arg.getOpcode() == ISD::UNDEF) continue;
2252 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2253 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2254 if (Val < 4 || Val > 7)
2261 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2262 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2263 bool X86::isPSHUFLWMask(SDNode *N) {
2264 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2266 if (N->getNumOperands() != 8)
2269 // Upper quadword copied in order.
2270 for (unsigned i = 4; i != 8; ++i)
2271 if (!isUndefOrEqual(N->getOperand(i), i))
2274 // Lower quadword shuffled.
2275 for (unsigned i = 0; i != 4; ++i)
2276 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2282 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2283 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2284 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2285 if (NumElems != 2 && NumElems != 4) return false;
2287 unsigned Half = NumElems / 2;
2288 for (unsigned i = 0; i < Half; ++i)
2289 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2291 for (unsigned i = Half; i < NumElems; ++i)
2292 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2298 bool X86::isSHUFPMask(SDNode *N) {
2299 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2300 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2303 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2304 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2305 /// half elements to come from vector 1 (which would equal the dest.) and
2306 /// the upper half to come from vector 2.
2307 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2308 if (NumOps != 2 && NumOps != 4) return false;
2310 unsigned Half = NumOps / 2;
2311 for (unsigned i = 0; i < Half; ++i)
2312 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2314 for (unsigned i = Half; i < NumOps; ++i)
2315 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2320 static bool isCommutedSHUFP(SDNode *N) {
2321 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2322 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2325 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2326 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2327 bool X86::isMOVHLPSMask(SDNode *N) {
2328 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2330 if (N->getNumOperands() != 4)
2333 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2334 return isUndefOrEqual(N->getOperand(0), 6) &&
2335 isUndefOrEqual(N->getOperand(1), 7) &&
2336 isUndefOrEqual(N->getOperand(2), 2) &&
2337 isUndefOrEqual(N->getOperand(3), 3);
2340 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2341 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2343 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2344 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2346 if (N->getNumOperands() != 4)
2349 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2350 return isUndefOrEqual(N->getOperand(0), 2) &&
2351 isUndefOrEqual(N->getOperand(1), 3) &&
2352 isUndefOrEqual(N->getOperand(2), 2) &&
2353 isUndefOrEqual(N->getOperand(3), 3);
2356 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2357 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2358 bool X86::isMOVLPMask(SDNode *N) {
2359 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2361 unsigned NumElems = N->getNumOperands();
2362 if (NumElems != 2 && NumElems != 4)
2365 for (unsigned i = 0; i < NumElems/2; ++i)
2366 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2369 for (unsigned i = NumElems/2; i < NumElems; ++i)
2370 if (!isUndefOrEqual(N->getOperand(i), i))
2376 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2377 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2379 bool X86::isMOVHPMask(SDNode *N) {
2380 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2382 unsigned NumElems = N->getNumOperands();
2383 if (NumElems != 2 && NumElems != 4)
2386 for (unsigned i = 0; i < NumElems/2; ++i)
2387 if (!isUndefOrEqual(N->getOperand(i), i))
2390 for (unsigned i = 0; i < NumElems/2; ++i) {
2391 SDOperand Arg = N->getOperand(i + NumElems/2);
2392 if (!isUndefOrEqual(Arg, i + NumElems))
2399 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2400 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2401 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2402 bool V2IsSplat = false) {
2403 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2406 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2407 SDOperand BitI = Elts[i];
2408 SDOperand BitI1 = Elts[i+1];
2409 if (!isUndefOrEqual(BitI, j))
2412 if (isUndefOrEqual(BitI1, NumElts))
2415 if (!isUndefOrEqual(BitI1, j + NumElts))
2423 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2424 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2425 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2428 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2429 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2430 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2431 bool V2IsSplat = false) {
2432 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2435 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2436 SDOperand BitI = Elts[i];
2437 SDOperand BitI1 = Elts[i+1];
2438 if (!isUndefOrEqual(BitI, j + NumElts/2))
2441 if (isUndefOrEqual(BitI1, NumElts))
2444 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2452 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2453 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2454 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2457 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2458 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2460 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2461 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2463 unsigned NumElems = N->getNumOperands();
2464 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2467 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2468 SDOperand BitI = N->getOperand(i);
2469 SDOperand BitI1 = N->getOperand(i+1);
2471 if (!isUndefOrEqual(BitI, j))
2473 if (!isUndefOrEqual(BitI1, j))
2480 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2481 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2483 bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
2484 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2486 unsigned NumElems = N->getNumOperands();
2487 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2490 for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2491 SDOperand BitI = N->getOperand(i);
2492 SDOperand BitI1 = N->getOperand(i + 1);
2494 if (!isUndefOrEqual(BitI, j))
2496 if (!isUndefOrEqual(BitI1, j))
2503 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2504 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2505 /// MOVSD, and MOVD, i.e. setting the lowest element.
2506 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2507 if (NumElts != 2 && NumElts != 4)
2510 if (!isUndefOrEqual(Elts[0], NumElts))
2513 for (unsigned i = 1; i < NumElts; ++i) {
2514 if (!isUndefOrEqual(Elts[i], i))
2521 bool X86::isMOVLMask(SDNode *N) {
2522 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2523 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2526 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2527 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2528 /// element of vector 2 and the other elements to come from vector 1 in order.
2529 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2530 bool V2IsSplat = false,
2531 bool V2IsUndef = false) {
2532 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2535 if (!isUndefOrEqual(Ops[0], 0))
2538 for (unsigned i = 1; i < NumOps; ++i) {
2539 SDOperand Arg = Ops[i];
2540 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2541 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2542 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2549 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2550 bool V2IsUndef = false) {
2551 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2552 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2553 V2IsSplat, V2IsUndef);
2556 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2557 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2558 bool X86::isMOVSHDUPMask(SDNode *N) {
2559 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2561 if (N->getNumOperands() != 4)
2564 // Expect 1, 1, 3, 3
2565 for (unsigned i = 0; i < 2; ++i) {
2566 SDOperand Arg = N->getOperand(i);
2567 if (Arg.getOpcode() == ISD::UNDEF) continue;
2568 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2569 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2570 if (Val != 1) return false;
2574 for (unsigned i = 2; i < 4; ++i) {
2575 SDOperand Arg = N->getOperand(i);
2576 if (Arg.getOpcode() == ISD::UNDEF) continue;
2577 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2578 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2579 if (Val != 3) return false;
2583 // Don't use movshdup if it can be done with a shufps.
2587 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2588 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2589 bool X86::isMOVSLDUPMask(SDNode *N) {
2590 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2592 if (N->getNumOperands() != 4)
2595 // Expect 0, 0, 2, 2
2596 for (unsigned i = 0; i < 2; ++i) {
2597 SDOperand Arg = N->getOperand(i);
2598 if (Arg.getOpcode() == ISD::UNDEF) continue;
2599 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2600 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2601 if (Val != 0) return false;
2605 for (unsigned i = 2; i < 4; ++i) {
2606 SDOperand Arg = N->getOperand(i);
2607 if (Arg.getOpcode() == ISD::UNDEF) continue;
2608 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2609 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2610 if (Val != 2) return false;
2614 // Don't use movshdup if it can be done with a shufps.
2618 /// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
2619 /// specifies a identity operation on the LHS or RHS.
2620 static bool isIdentityMask(SDNode *N, bool RHS = false) {
2621 unsigned NumElems = N->getNumOperands();
2622 for (unsigned i = 0; i < NumElems; ++i)
2623 if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
2628 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2629 /// a splat of a single element.
2630 static bool isSplatMask(SDNode *N) {
2631 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2633 // This is a splat operation if each element of the permute is the same, and
2634 // if the value doesn't reference the second vector.
2635 unsigned NumElems = N->getNumOperands();
2636 SDOperand ElementBase;
2638 for (; i != NumElems; ++i) {
2639 SDOperand Elt = N->getOperand(i);
2640 if (isa<ConstantSDNode>(Elt)) {
2646 if (!ElementBase.Val)
2649 for (; i != NumElems; ++i) {
2650 SDOperand Arg = N->getOperand(i);
2651 if (Arg.getOpcode() == ISD::UNDEF) continue;
2652 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2653 if (Arg != ElementBase) return false;
2656 // Make sure it is a splat of the first vector operand.
2657 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2660 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2661 /// a splat of a single element and it's a 2 or 4 element mask.
2662 bool X86::isSplatMask(SDNode *N) {
2663 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2665 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2666 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2668 return ::isSplatMask(N);
2671 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2672 /// specifies a splat of zero element.
2673 bool X86::isSplatLoMask(SDNode *N) {
2674 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2676 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2677 if (!isUndefOrEqual(N->getOperand(i), 0))
2682 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2683 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2685 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2686 unsigned NumOperands = N->getNumOperands();
2687 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2689 for (unsigned i = 0; i < NumOperands; ++i) {
2691 SDOperand Arg = N->getOperand(NumOperands-i-1);
2692 if (Arg.getOpcode() != ISD::UNDEF)
2693 Val = cast<ConstantSDNode>(Arg)->getValue();
2694 if (Val >= NumOperands) Val -= NumOperands;
2696 if (i != NumOperands - 1)
2703 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2704 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2706 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2708 // 8 nodes, but we only care about the last 4.
2709 for (unsigned i = 7; i >= 4; --i) {
2711 SDOperand Arg = N->getOperand(i);
2712 if (Arg.getOpcode() != ISD::UNDEF)
2713 Val = cast<ConstantSDNode>(Arg)->getValue();
2722 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2723 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2725 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2727 // 8 nodes, but we only care about the first 4.
2728 for (int i = 3; i >= 0; --i) {
2730 SDOperand Arg = N->getOperand(i);
2731 if (Arg.getOpcode() != ISD::UNDEF)
2732 Val = cast<ConstantSDNode>(Arg)->getValue();
2741 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2742 /// specifies a 8 element shuffle that can be broken into a pair of
2743 /// PSHUFHW and PSHUFLW.
2744 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2745 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2747 if (N->getNumOperands() != 8)
2750 // Lower quadword shuffled.
2751 for (unsigned i = 0; i != 4; ++i) {
2752 SDOperand Arg = N->getOperand(i);
2753 if (Arg.getOpcode() == ISD::UNDEF) continue;
2754 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2755 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2760 // Upper quadword shuffled.
2761 for (unsigned i = 4; i != 8; ++i) {
2762 SDOperand Arg = N->getOperand(i);
2763 if (Arg.getOpcode() == ISD::UNDEF) continue;
2764 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2765 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2766 if (Val < 4 || Val > 7)
2773 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as
2774 /// values in ther permute mask.
2775 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2776 SDOperand &V2, SDOperand &Mask,
2777 SelectionDAG &DAG) {
2778 MVT::ValueType VT = Op.getValueType();
2779 MVT::ValueType MaskVT = Mask.getValueType();
2780 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2781 unsigned NumElems = Mask.getNumOperands();
2782 SmallVector<SDOperand, 8> MaskVec;
2784 for (unsigned i = 0; i != NumElems; ++i) {
2785 SDOperand Arg = Mask.getOperand(i);
2786 if (Arg.getOpcode() == ISD::UNDEF) {
2787 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2790 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2791 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2793 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2795 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2799 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2800 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2803 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2804 /// the two vector operands have swapped position.
2806 SDOperand CommuteVectorShuffleMask(SDOperand Mask, SelectionDAG &DAG) {
2807 MVT::ValueType MaskVT = Mask.getValueType();
2808 MVT::ValueType EltVT = MVT::getVectorElementType(MaskVT);
2809 unsigned NumElems = Mask.getNumOperands();
2810 SmallVector<SDOperand, 8> MaskVec;
2811 for (unsigned i = 0; i != NumElems; ++i) {
2812 SDOperand Arg = Mask.getOperand(i);
2813 if (Arg.getOpcode() == ISD::UNDEF) {
2814 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2817 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2818 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2820 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2822 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2824 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], NumElems);
2828 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2829 /// match movhlps. The lower half elements should come from upper half of
2830 /// V1 (and in order), and the upper half elements should come from the upper
2831 /// half of V2 (and in order).
2832 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2833 unsigned NumElems = Mask->getNumOperands();
2836 for (unsigned i = 0, e = 2; i != e; ++i)
2837 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2839 for (unsigned i = 2; i != 4; ++i)
2840 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2845 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2846 /// is promoted to a vector.
2847 static inline bool isScalarLoadToVector(SDNode *N) {
2848 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2849 N = N->getOperand(0).Val;
2850 return ISD::isNON_EXTLoad(N);
2855 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2856 /// match movlp{s|d}. The lower half elements should come from lower half of
2857 /// V1 (and in order), and the upper half elements should come from the upper
2858 /// half of V2 (and in order). And since V1 will become the source of the
2859 /// MOVLP, it must be either a vector load or a scalar load to vector.
2860 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2861 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2863 // Is V2 is a vector load, don't do this transformation. We will try to use
2864 // load folding shufps op.
2865 if (ISD::isNON_EXTLoad(V2))
2868 unsigned NumElems = Mask->getNumOperands();
2869 if (NumElems != 2 && NumElems != 4)
2871 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2872 if (!isUndefOrEqual(Mask->getOperand(i), i))
2874 for (unsigned i = NumElems/2; i != NumElems; ++i)
2875 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2880 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2882 static bool isSplatVector(SDNode *N) {
2883 if (N->getOpcode() != ISD::BUILD_VECTOR)
2886 SDOperand SplatValue = N->getOperand(0);
2887 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2888 if (N->getOperand(i) != SplatValue)
2893 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2895 static bool isUndefShuffle(SDNode *N) {
2896 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2899 SDOperand V1 = N->getOperand(0);
2900 SDOperand V2 = N->getOperand(1);
2901 SDOperand Mask = N->getOperand(2);
2902 unsigned NumElems = Mask.getNumOperands();
2903 for (unsigned i = 0; i != NumElems; ++i) {
2904 SDOperand Arg = Mask.getOperand(i);
2905 if (Arg.getOpcode() != ISD::UNDEF) {
2906 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2907 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2909 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2916 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2918 static inline bool isZeroNode(SDOperand Elt) {
2919 return ((isa<ConstantSDNode>(Elt) &&
2920 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2921 (isa<ConstantFPSDNode>(Elt) &&
2922 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2925 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2926 /// to an zero vector.
2927 static bool isZeroShuffle(SDNode *N) {
2928 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2931 SDOperand V1 = N->getOperand(0);
2932 SDOperand V2 = N->getOperand(1);
2933 SDOperand Mask = N->getOperand(2);
2934 unsigned NumElems = Mask.getNumOperands();
2935 for (unsigned i = 0; i != NumElems; ++i) {
2936 SDOperand Arg = Mask.getOperand(i);
2937 if (Arg.getOpcode() == ISD::UNDEF)
2940 unsigned Idx = cast<ConstantSDNode>(Arg)->getValue();
2941 if (Idx < NumElems) {
2942 unsigned Opc = V1.Val->getOpcode();
2943 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.Val))
2945 if (Opc != ISD::BUILD_VECTOR ||
2946 !isZeroNode(V1.Val->getOperand(Idx)))
2948 } else if (Idx >= NumElems) {
2949 unsigned Opc = V2.Val->getOpcode();
2950 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.Val))
2952 if (Opc != ISD::BUILD_VECTOR ||
2953 !isZeroNode(V2.Val->getOperand(Idx - NumElems)))
2960 /// getZeroVector - Returns a vector of specified type with all zero elements.
2962 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2963 assert(MVT::isVector(VT) && "Expected a vector type");
2965 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2966 // type. This ensures they get CSE'd.
2967 SDOperand Cst = DAG.getTargetConstant(0, MVT::i32);
2969 if (MVT::getSizeInBits(VT) == 64) // MMX
2970 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2972 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2973 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2976 /// getOnesVector - Returns a vector of specified type with all bits set.
2978 static SDOperand getOnesVector(MVT::ValueType VT, SelectionDAG &DAG) {
2979 assert(MVT::isVector(VT) && "Expected a vector type");
2981 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2982 // type. This ensures they get CSE'd.
2983 SDOperand Cst = DAG.getTargetConstant(~0U, MVT::i32);
2985 if (MVT::getSizeInBits(VT) == 64) // MMX
2986 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
2988 Vec = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst, Cst, Cst, Cst);
2989 return DAG.getNode(ISD::BIT_CONVERT, VT, Vec);
2993 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2994 /// that point to V2 points to its first element.
2995 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2996 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2998 bool Changed = false;
2999 SmallVector<SDOperand, 8> MaskVec;
3000 unsigned NumElems = Mask.getNumOperands();
3001 for (unsigned i = 0; i != NumElems; ++i) {
3002 SDOperand Arg = Mask.getOperand(i);
3003 if (Arg.getOpcode() != ISD::UNDEF) {
3004 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
3005 if (Val > NumElems) {
3006 Arg = DAG.getConstant(NumElems, Arg.getValueType());
3010 MaskVec.push_back(Arg);
3014 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
3015 &MaskVec[0], MaskVec.size());
3019 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3020 /// operation of specified width.
3021 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
3022 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3023 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3025 SmallVector<SDOperand, 8> MaskVec;
3026 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
3027 for (unsigned i = 1; i != NumElems; ++i)
3028 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3029 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
3032 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
3033 /// of specified width.
3034 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
3035 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3036 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3037 SmallVector<SDOperand, 8> MaskVec;
3038 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3039 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3040 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
3042 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
3045 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
3046 /// of specified width.
3047 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
3048 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3049 MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
3050 unsigned Half = NumElems/2;
3051 SmallVector<SDOperand, 8> MaskVec;
3052 for (unsigned i = 0; i != Half; ++i) {
3053 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
3054 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
3056 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
3059 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
3061 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
3062 SDOperand V1 = Op.getOperand(0);
3063 SDOperand Mask = Op.getOperand(2);
3064 MVT::ValueType VT = Op.getValueType();
3065 unsigned NumElems = Mask.getNumOperands();
3066 Mask = getUnpacklMask(NumElems, DAG);
3067 while (NumElems != 4) {
3068 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
3071 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
3073 Mask = getZeroVector(MVT::v4i32, DAG);
3074 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
3075 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
3076 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
3079 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3080 /// vector of zero or undef vector. This produces a shuffle where the low
3081 /// element of V2 is swizzled into the zero/undef vector, landing at element
3082 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3083 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
3084 unsigned NumElems, unsigned Idx,
3085 bool isZero, SelectionDAG &DAG) {
3086 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
3087 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3088 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3089 SmallVector<SDOperand, 16> MaskVec;
3090 for (unsigned i = 0; i != NumElems; ++i)
3091 if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
3092 MaskVec.push_back(DAG.getConstant(NumElems, EVT));
3094 MaskVec.push_back(DAG.getConstant(i, EVT));
3095 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3096 &MaskVec[0], MaskVec.size());
3097 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3100 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3102 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
3103 unsigned NumNonZero, unsigned NumZero,
3104 SelectionDAG &DAG, TargetLowering &TLI) {
3110 for (unsigned i = 0; i < 16; ++i) {
3111 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3112 if (ThisIsNonZero && First) {
3114 V = getZeroVector(MVT::v8i16, DAG);
3116 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3121 SDOperand ThisElt(0, 0), LastElt(0, 0);
3122 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3123 if (LastIsNonZero) {
3124 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
3126 if (ThisIsNonZero) {
3127 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
3128 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
3129 ThisElt, DAG.getConstant(8, MVT::i8));
3131 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
3136 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
3137 DAG.getConstant(i/2, TLI.getPointerTy()));
3141 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
3144 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3146 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
3147 unsigned NumNonZero, unsigned NumZero,
3148 SelectionDAG &DAG, TargetLowering &TLI) {
3154 for (unsigned i = 0; i < 8; ++i) {
3155 bool isNonZero = (NonZeros & (1 << i)) != 0;
3159 V = getZeroVector(MVT::v8i16, DAG);
3161 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
3164 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
3165 DAG.getConstant(i, TLI.getPointerTy()));
3173 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3174 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3175 if (ISD::isBuildVectorAllZeros(Op.Val) || ISD::isBuildVectorAllOnes(Op.Val)) {
3176 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3177 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3178 // eliminated on x86-32 hosts.
3179 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3182 if (ISD::isBuildVectorAllOnes(Op.Val))
3183 return getOnesVector(Op.getValueType(), DAG);
3184 return getZeroVector(Op.getValueType(), DAG);
3187 MVT::ValueType VT = Op.getValueType();
3188 MVT::ValueType EVT = MVT::getVectorElementType(VT);
3189 unsigned EVTBits = MVT::getSizeInBits(EVT);
3191 unsigned NumElems = Op.getNumOperands();
3192 unsigned NumZero = 0;
3193 unsigned NumNonZero = 0;
3194 unsigned NonZeros = 0;
3195 bool HasNonImms = false;
3196 SmallSet<SDOperand, 8> Values;
3197 for (unsigned i = 0; i < NumElems; ++i) {
3198 SDOperand Elt = Op.getOperand(i);
3199 if (Elt.getOpcode() == ISD::UNDEF)
3202 if (Elt.getOpcode() != ISD::Constant &&
3203 Elt.getOpcode() != ISD::ConstantFP)
3205 if (isZeroNode(Elt))
3208 NonZeros |= (1 << i);
3213 if (NumNonZero == 0) {
3214 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3215 return DAG.getNode(ISD::UNDEF, VT);
3218 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3219 if (Values.size() == 1)
3222 // Special case for single non-zero element.
3223 if (NumNonZero == 1 && NumElems <= 4) {
3224 unsigned Idx = CountTrailingZeros_32(NonZeros);
3225 SDOperand Item = Op.getOperand(Idx);
3226 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
3228 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3229 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
3231 else if (!HasNonImms) // Otherwise, it's better to do a constpool load.
3234 if (EVTBits == 32) {
3235 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3236 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
3238 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3239 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3240 SmallVector<SDOperand, 8> MaskVec;
3241 for (unsigned i = 0; i < NumElems; i++)
3242 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
3243 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3244 &MaskVec[0], MaskVec.size());
3245 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
3246 DAG.getNode(ISD::UNDEF, VT), Mask);
3250 // A vector full of immediates; various special cases are already
3251 // handled, so this is best done with a single constant-pool load.
3255 // Let legalizer expand 2-wide build_vectors.
3259 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3260 if (EVTBits == 8 && NumElems == 16) {
3261 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3263 if (V.Val) return V;
3266 if (EVTBits == 16 && NumElems == 8) {
3267 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3269 if (V.Val) return V;
3272 // If element VT is == 32 bits, turn it into a number of shuffles.
3273 SmallVector<SDOperand, 8> V;
3275 if (NumElems == 4 && NumZero > 0) {
3276 for (unsigned i = 0; i < 4; ++i) {
3277 bool isZero = !(NonZeros & (1 << i));
3279 V[i] = getZeroVector(VT, DAG);
3281 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3284 for (unsigned i = 0; i < 2; ++i) {
3285 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3288 V[i] = V[i*2]; // Must be a zero vector.
3291 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3292 getMOVLMask(NumElems, DAG));
3295 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3296 getMOVLMask(NumElems, DAG));
3299 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3300 getUnpacklMask(NumElems, DAG));
3305 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3306 // clears the upper bits.
3307 // FIXME: we can do the same for v4f32 case when we know both parts of
3308 // the lower half come from scalar_to_vector (loadf32). We should do
3309 // that in post legalizer dag combiner with target specific hooks.
3310 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3312 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3313 MVT::ValueType EVT = MVT::getVectorElementType(MaskVT);
3314 SmallVector<SDOperand, 8> MaskVec;
3315 bool Reverse = (NonZeros & 0x3) == 2;
3316 for (unsigned i = 0; i < 2; ++i)
3318 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3320 MaskVec.push_back(DAG.getConstant(i, EVT));
3321 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3322 for (unsigned i = 0; i < 2; ++i)
3324 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3326 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3327 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3328 &MaskVec[0], MaskVec.size());
3329 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3332 if (Values.size() > 2) {
3333 // Expand into a number of unpckl*.
3335 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3336 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3337 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3338 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3339 for (unsigned i = 0; i < NumElems; ++i)
3340 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3342 while (NumElems != 0) {
3343 for (unsigned i = 0; i < NumElems; ++i)
3344 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3355 SDOperand LowerVECTOR_SHUFFLEv8i16(SDOperand V1, SDOperand V2,
3356 SDOperand PermMask, SelectionDAG &DAG,
3357 TargetLowering &TLI) {
3359 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(8);
3360 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3361 MVT::ValueType PtrVT = TLI.getPointerTy();
3362 SmallVector<SDOperand, 8> MaskElts(PermMask.Val->op_begin(),
3363 PermMask.Val->op_end());
3365 // First record which half of which vector the low elements come from.
3366 SmallVector<unsigned, 4> LowQuad(4);
3367 for (unsigned i = 0; i < 4; ++i) {
3368 SDOperand Elt = MaskElts[i];
3369 if (Elt.getOpcode() == ISD::UNDEF)
3371 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3372 int QuadIdx = EltIdx / 4;
3375 int BestLowQuad = -1;
3376 unsigned MaxQuad = 1;
3377 for (unsigned i = 0; i < 4; ++i) {
3378 if (LowQuad[i] > MaxQuad) {
3380 MaxQuad = LowQuad[i];
3384 // Record which half of which vector the high elements come from.
3385 SmallVector<unsigned, 4> HighQuad(4);
3386 for (unsigned i = 4; i < 8; ++i) {
3387 SDOperand Elt = MaskElts[i];
3388 if (Elt.getOpcode() == ISD::UNDEF)
3390 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3391 int QuadIdx = EltIdx / 4;
3392 ++HighQuad[QuadIdx];
3394 int BestHighQuad = -1;
3396 for (unsigned i = 0; i < 4; ++i) {
3397 if (HighQuad[i] > MaxQuad) {
3399 MaxQuad = HighQuad[i];
3403 // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
3404 if (BestLowQuad != -1 || BestHighQuad != -1) {
3405 // First sort the 4 chunks in order using shufpd.
3406 SmallVector<SDOperand, 8> MaskVec;
3407 if (BestLowQuad != -1)
3408 MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
3410 MaskVec.push_back(DAG.getConstant(0, MVT::i32));
3411 if (BestHighQuad != -1)
3412 MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
3414 MaskVec.push_back(DAG.getConstant(1, MVT::i32));
3415 SDOperand Mask= DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, &MaskVec[0],2);
3416 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
3417 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V1),
3418 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, V2), Mask);
3419 NewV = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, NewV);
3421 // Now sort high and low parts separately.
3422 BitVector InOrder(8);
3423 if (BestLowQuad != -1) {
3424 // Sort lower half in order using PSHUFLW.
3426 bool AnyOutOrder = false;
3427 for (unsigned i = 0; i != 4; ++i) {
3428 SDOperand Elt = MaskElts[i];
3429 if (Elt.getOpcode() == ISD::UNDEF) {
3430 MaskVec.push_back(Elt);
3433 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3436 MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
3437 // If this element is in the right place after this shuffle, then
3439 if ((int)(EltIdx / 4) == BestLowQuad)
3444 for (unsigned i = 4; i != 8; ++i)
3445 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3446 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3447 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3451 if (BestHighQuad != -1) {
3452 // Sort high half in order using PSHUFHW if possible.
3454 for (unsigned i = 0; i != 4; ++i)
3455 MaskVec.push_back(DAG.getConstant(i, MaskEVT));
3456 bool AnyOutOrder = false;
3457 for (unsigned i = 4; i != 8; ++i) {
3458 SDOperand Elt = MaskElts[i];
3459 if (Elt.getOpcode() == ISD::UNDEF) {
3460 MaskVec.push_back(Elt);
3463 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3466 MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
3467 // If this element is in the right place after this shuffle, then
3469 if ((int)(EltIdx / 4) == BestHighQuad)
3474 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3475 NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, NewV, NewV, Mask);
3479 // The other elements are put in the right place using pextrw and pinsrw.
3480 for (unsigned i = 0; i != 8; ++i) {
3483 SDOperand Elt = MaskElts[i];
3484 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3487 SDOperand ExtOp = (EltIdx < 8)
3488 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3489 DAG.getConstant(EltIdx, PtrVT))
3490 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3491 DAG.getConstant(EltIdx - 8, PtrVT));
3492 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3493 DAG.getConstant(i, PtrVT));
3498 // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use
3499 ///as few as possible.
3500 // First, let's find out how many elements are already in the right order.
3501 unsigned V1InOrder = 0;
3502 unsigned V1FromV1 = 0;
3503 unsigned V2InOrder = 0;
3504 unsigned V2FromV2 = 0;
3505 SmallVector<SDOperand, 8> V1Elts;
3506 SmallVector<SDOperand, 8> V2Elts;
3507 for (unsigned i = 0; i < 8; ++i) {
3508 SDOperand Elt = MaskElts[i];
3509 if (Elt.getOpcode() == ISD::UNDEF) {
3510 V1Elts.push_back(Elt);
3511 V2Elts.push_back(Elt);
3516 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3518 V1Elts.push_back(Elt);
3519 V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
3521 } else if (EltIdx == i+8) {
3522 V1Elts.push_back(Elt);
3523 V2Elts.push_back(DAG.getConstant(i, MaskEVT));
3525 } else if (EltIdx < 8) {
3526 V1Elts.push_back(Elt);
3529 V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
3534 if (V2InOrder > V1InOrder) {
3535 PermMask = CommuteVectorShuffleMask(PermMask, DAG);
3537 std::swap(V1Elts, V2Elts);
3538 std::swap(V1FromV1, V2FromV2);
3541 if ((V1FromV1 + V1InOrder) != 8) {
3542 // Some elements are from V2.
3544 // If there are elements that are from V1 but out of place,
3545 // then first sort them in place
3546 SmallVector<SDOperand, 8> MaskVec;
3547 for (unsigned i = 0; i < 8; ++i) {
3548 SDOperand Elt = V1Elts[i];
3549 if (Elt.getOpcode() == ISD::UNDEF) {
3550 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3553 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3555 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3557 MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
3559 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], 8);
3560 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, V1, V1, Mask);
3564 for (unsigned i = 0; i < 8; ++i) {
3565 SDOperand Elt = V1Elts[i];
3566 if (Elt.getOpcode() == ISD::UNDEF)
3568 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3571 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V2,
3572 DAG.getConstant(EltIdx - 8, PtrVT));
3573 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3574 DAG.getConstant(i, PtrVT));
3578 // All elements are from V1.
3580 for (unsigned i = 0; i < 8; ++i) {
3581 SDOperand Elt = V1Elts[i];
3582 if (Elt.getOpcode() == ISD::UNDEF)
3584 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3585 SDOperand ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, V1,
3586 DAG.getConstant(EltIdx, PtrVT));
3587 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, NewV, ExtOp,
3588 DAG.getConstant(i, PtrVT));
3594 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3595 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3596 /// done when every pair / quad of shuffle mask elements point to elements in
3597 /// the right sequence. e.g.
3598 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3600 SDOperand RewriteAsNarrowerShuffle(SDOperand V1, SDOperand V2,
3602 SDOperand PermMask, SelectionDAG &DAG,
3603 TargetLowering &TLI) {
3604 unsigned NumElems = PermMask.getNumOperands();
3605 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3606 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3607 MVT::ValueType NewVT = MaskVT;
3609 case MVT::v4f32: NewVT = MVT::v2f64; break;
3610 case MVT::v4i32: NewVT = MVT::v2i64; break;
3611 case MVT::v8i16: NewVT = MVT::v4i32; break;
3612 case MVT::v16i8: NewVT = MVT::v4i32; break;
3613 default: assert(false && "Unexpected!");
3617 if (MVT::isInteger(VT))
3621 unsigned Scale = NumElems / NewWidth;
3622 SmallVector<SDOperand, 8> MaskVec;
3623 for (unsigned i = 0; i < NumElems; i += Scale) {
3624 unsigned StartIdx = ~0U;
3625 for (unsigned j = 0; j < Scale; ++j) {
3626 SDOperand Elt = PermMask.getOperand(i+j);
3627 if (Elt.getOpcode() == ISD::UNDEF)
3629 unsigned EltIdx = cast<ConstantSDNode>(Elt)->getValue();
3630 if (StartIdx == ~0U)
3631 StartIdx = EltIdx - (EltIdx % Scale);
3632 if (EltIdx != StartIdx + j)
3635 if (StartIdx == ~0U)
3636 MaskVec.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3638 MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MVT::i32));
3641 V1 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V1);
3642 V2 = DAG.getNode(ISD::BIT_CONVERT, NewVT, V2);
3643 return DAG.getNode(ISD::VECTOR_SHUFFLE, NewVT, V1, V2,
3644 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3645 &MaskVec[0], MaskVec.size()));
3649 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3650 SDOperand V1 = Op.getOperand(0);
3651 SDOperand V2 = Op.getOperand(1);
3652 SDOperand PermMask = Op.getOperand(2);
3653 MVT::ValueType VT = Op.getValueType();
3654 unsigned NumElems = PermMask.getNumOperands();
3655 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3656 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3657 bool V1IsSplat = false;
3658 bool V2IsSplat = false;
3660 if (isUndefShuffle(Op.Val))
3661 return DAG.getNode(ISD::UNDEF, VT);
3663 if (isZeroShuffle(Op.Val))
3664 return getZeroVector(VT, DAG);
3666 if (isIdentityMask(PermMask.Val))
3668 else if (isIdentityMask(PermMask.Val, true))
3671 if (isSplatMask(PermMask.Val)) {
3672 if (NumElems <= 4) return Op;
3673 // Promote it to a v4i32 splat.
3674 return PromoteSplat(Op, DAG);
3677 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3679 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3680 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3682 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3683 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3684 // FIXME: Figure out a cleaner way to do this.
3685 // Try to make use of movq to zero out the top part.
3686 if (ISD::isBuildVectorAllZeros(V2.Val)) {
3687 SDOperand NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3689 SDOperand NewV1 = NewOp.getOperand(0);
3690 SDOperand NewV2 = NewOp.getOperand(1);
3691 SDOperand NewMask = NewOp.getOperand(2);
3692 if (isCommutedMOVL(NewMask.Val, true, false)) {
3693 NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
3694 NewOp = DAG.getNode(ISD::VECTOR_SHUFFLE, NewOp.getValueType(),
3695 NewV1, NewV2, getMOVLMask(2, DAG));
3696 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3699 } else if (ISD::isBuildVectorAllZeros(V1.Val)) {
3700 SDOperand NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG, *this);
3701 if (NewOp.Val && X86::isMOVLMask(NewOp.getOperand(2).Val))
3702 return DAG.getNode(ISD::BIT_CONVERT, VT, LowerVECTOR_SHUFFLE(NewOp, DAG));
3706 if (X86::isMOVLMask(PermMask.Val))
3707 return (V1IsUndef) ? V2 : Op;
3709 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3710 X86::isMOVSLDUPMask(PermMask.Val) ||
3711 X86::isMOVHLPSMask(PermMask.Val) ||
3712 X86::isMOVHPMask(PermMask.Val) ||
3713 X86::isMOVLPMask(PermMask.Val))
3716 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3717 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3718 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3720 bool Commuted = false;
3721 // FIXME: This should also accept a bitcast of a splat? Be careful, not
3722 // 1,1,1,1 -> v8i16 though.
3723 V1IsSplat = isSplatVector(V1.Val);
3724 V2IsSplat = isSplatVector(V2.Val);
3726 // Canonicalize the splat or undef, if present, to be on the RHS.
3727 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3728 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3729 std::swap(V1IsSplat, V2IsSplat);
3730 std::swap(V1IsUndef, V2IsUndef);
3734 // FIXME: Figure out a cleaner way to do this.
3735 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3736 if (V2IsUndef) return V1;
3737 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3739 // V2 is a splat, so the mask may be malformed. That is, it may point
3740 // to any V2 element. The instruction selectior won't like this. Get
3741 // a corrected mask and commute to form a proper MOVS{S|D}.
3742 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3743 if (NewMask.Val != PermMask.Val)
3744 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3749 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3750 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3751 X86::isUNPCKLMask(PermMask.Val) ||
3752 X86::isUNPCKHMask(PermMask.Val))
3756 // Normalize mask so all entries that point to V2 points to its first
3757 // element then try to match unpck{h|l} again. If match, return a
3758 // new vector_shuffle with the corrected mask.
3759 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3760 if (NewMask.Val != PermMask.Val) {
3761 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3762 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3763 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3764 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3765 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3766 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3771 // Normalize the node to match x86 shuffle ops if needed
3772 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3773 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3776 // Commute is back and try unpck* again.
3777 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3778 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3779 X86::isUNPCKH_v_undef_Mask(PermMask.Val) ||
3780 X86::isUNPCKLMask(PermMask.Val) ||
3781 X86::isUNPCKHMask(PermMask.Val))
3785 // If VT is integer, try PSHUF* first, then SHUFP*.
3786 if (MVT::isInteger(VT)) {
3787 // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
3788 // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
3789 if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
3790 X86::isPSHUFDMask(PermMask.Val)) ||
3791 X86::isPSHUFHWMask(PermMask.Val) ||
3792 X86::isPSHUFLWMask(PermMask.Val)) {
3793 if (V2.getOpcode() != ISD::UNDEF)
3794 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3795 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3799 if (X86::isSHUFPMask(PermMask.Val) &&
3800 MVT::getSizeInBits(VT) != 64) // Don't do this for MMX.
3803 // Floating point cases in the other order.
3804 if (X86::isSHUFPMask(PermMask.Val))
3806 if (X86::isPSHUFDMask(PermMask.Val) ||
3807 X86::isPSHUFHWMask(PermMask.Val) ||
3808 X86::isPSHUFLWMask(PermMask.Val)) {
3809 if (V2.getOpcode() != ISD::UNDEF)
3810 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3811 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3816 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
3817 if (VT == MVT::v8i16) {
3818 SDOperand NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
3823 // Handle all 4 wide cases with a number of shuffles.
3824 if (NumElems == 4 && MVT::getSizeInBits(VT) != 64) {
3825 // Don't do this for MMX.
3826 MVT::ValueType MaskVT = PermMask.getValueType();
3827 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
3828 SmallVector<std::pair<int, int>, 8> Locs;
3829 Locs.reserve(NumElems);
3830 SmallVector<SDOperand, 8> Mask1(NumElems,
3831 DAG.getNode(ISD::UNDEF, MaskEVT));
3832 SmallVector<SDOperand, 8> Mask2(NumElems,
3833 DAG.getNode(ISD::UNDEF, MaskEVT));
3836 // If no more than two elements come from either vector. This can be
3837 // implemented with two shuffles. First shuffle gather the elements.
3838 // The second shuffle, which takes the first shuffle as both of its
3839 // vector operands, put the elements into the right order.
3840 for (unsigned i = 0; i != NumElems; ++i) {
3841 SDOperand Elt = PermMask.getOperand(i);
3842 if (Elt.getOpcode() == ISD::UNDEF) {
3843 Locs[i] = std::make_pair(-1, -1);
3845 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3846 if (Val < NumElems) {
3847 Locs[i] = std::make_pair(0, NumLo);
3851 Locs[i] = std::make_pair(1, NumHi);
3852 if (2+NumHi < NumElems)
3853 Mask1[2+NumHi] = Elt;
3858 if (NumLo <= 2 && NumHi <= 2) {
3859 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3860 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3861 &Mask1[0], Mask1.size()));
3862 for (unsigned i = 0; i != NumElems; ++i) {
3863 if (Locs[i].first == -1)
3866 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3867 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3868 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3872 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3873 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3874 &Mask2[0], Mask2.size()));
3877 // Break it into (shuffle shuffle_hi, shuffle_lo).
3879 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3880 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3881 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3882 unsigned MaskIdx = 0;
3884 unsigned HiIdx = NumElems/2;
3885 for (unsigned i = 0; i != NumElems; ++i) {
3886 if (i == NumElems/2) {
3892 SDOperand Elt = PermMask.getOperand(i);
3893 if (Elt.getOpcode() == ISD::UNDEF) {
3894 Locs[i] = std::make_pair(-1, -1);
3895 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3896 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3897 (*MaskPtr)[LoIdx] = Elt;
3900 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3901 (*MaskPtr)[HiIdx] = Elt;
3906 SDOperand LoShuffle =
3907 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3908 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3909 &LoMask[0], LoMask.size()));
3910 SDOperand HiShuffle =
3911 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3912 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3913 &HiMask[0], HiMask.size()));
3914 SmallVector<SDOperand, 8> MaskOps;
3915 for (unsigned i = 0; i != NumElems; ++i) {
3916 if (Locs[i].first == -1) {
3917 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3919 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3920 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3923 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3924 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3925 &MaskOps[0], MaskOps.size()));
3932 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3933 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3936 MVT::ValueType VT = Op.getValueType();
3937 // TODO: handle v16i8.
3938 if (MVT::getSizeInBits(VT) == 16) {
3939 SDOperand Vec = Op.getOperand(0);
3940 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3942 return DAG.getNode(ISD::TRUNCATE, MVT::i16,
3943 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
3944 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
3946 // Transform it so it match pextrw which produces a 32-bit result.
3947 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3948 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3949 Op.getOperand(0), Op.getOperand(1));
3950 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3951 DAG.getValueType(VT));
3952 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3953 } else if (MVT::getSizeInBits(VT) == 32) {
3954 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3957 // SHUFPS the element to the lowest double word, then movss.
3958 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3959 SmallVector<SDOperand, 8> IdxVec;
3961 push_back(DAG.getConstant(Idx, MVT::getVectorElementType(MaskVT)));
3963 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3965 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3967 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3968 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3969 &IdxVec[0], IdxVec.size());
3970 SDOperand Vec = Op.getOperand(0);
3971 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3972 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3973 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3974 DAG.getConstant(0, getPointerTy()));
3975 } else if (MVT::getSizeInBits(VT) == 64) {
3976 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3980 // UNPCKHPD the element to the lowest double word, then movsd.
3981 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3982 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3983 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3984 SmallVector<SDOperand, 8> IdxVec;
3985 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorElementType(MaskVT)));
3987 push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(MaskVT)));
3988 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3989 &IdxVec[0], IdxVec.size());
3990 SDOperand Vec = Op.getOperand(0);
3991 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3992 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3993 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3994 DAG.getConstant(0, getPointerTy()));
4001 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
4002 MVT::ValueType VT = Op.getValueType();
4003 MVT::ValueType EVT = MVT::getVectorElementType(VT);
4007 SDOperand N0 = Op.getOperand(0);
4008 SDOperand N1 = Op.getOperand(1);
4009 SDOperand N2 = Op.getOperand(2);
4011 if (MVT::getSizeInBits(EVT) == 16) {
4012 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4013 // as its second argument.
4014 if (N1.getValueType() != MVT::i32)
4015 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
4016 if (N2.getValueType() != MVT::i32)
4017 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
4018 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
4021 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
4022 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
4023 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
4024 MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
4025 SmallVector<SDOperand, 4> MaskVec;
4026 for (unsigned i = 0; i < 4; ++i)
4027 MaskVec.push_back(DAG.getConstant((i == Idx) ? i+4 : i, MaskEVT));
4028 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
4029 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
4030 &MaskVec[0], MaskVec.size()));
4034 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
4035 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
4036 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
4039 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4040 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4041 // one of the above mentioned nodes. It has to be wrapped because otherwise
4042 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4043 // be used to form addressing mode. These wrapped nodes will be selected
4046 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
4047 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4048 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
4050 CP->getAlignment());
4051 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4052 // With PIC, the address is actually $g + Offset.
4053 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4054 !Subtarget->isPICStyleRIPRel()) {
4055 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4056 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4064 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
4065 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4066 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
4067 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4068 // With PIC, the address is actually $g + Offset.
4069 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4070 !Subtarget->isPICStyleRIPRel()) {
4071 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4072 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4076 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4077 // load the value at address GV, not the value of GV itself. This means that
4078 // the GlobalAddress must be in the base or index register of the address, not
4079 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4080 // The same applies for external symbols during PIC codegen
4081 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
4082 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
4087 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
4089 LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4090 const MVT::ValueType PtrVT) {
4092 SDOperand Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
4093 DAG.getNode(X86ISD::GlobalBaseReg,
4095 InFlag = Chain.getValue(1);
4097 // emit leal symbol@TLSGD(,%ebx,1), %eax
4098 SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
4099 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4100 GA->getValueType(0),
4102 SDOperand Ops[] = { Chain, TGA, InFlag };
4103 SDOperand Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
4104 InFlag = Result.getValue(2);
4105 Chain = Result.getValue(1);
4107 // call ___tls_get_addr. This function receives its argument in
4108 // the register EAX.
4109 Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
4110 InFlag = Chain.getValue(1);
4112 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4113 SDOperand Ops1[] = { Chain,
4114 DAG.getTargetExternalSymbol("___tls_get_addr",
4116 DAG.getRegister(X86::EAX, PtrVT),
4117 DAG.getRegister(X86::EBX, PtrVT),
4119 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
4120 InFlag = Chain.getValue(1);
4122 return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
4125 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4126 // "local exec" model.
4128 LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4129 const MVT::ValueType PtrVT) {
4130 // Get the Thread Pointer
4131 SDOperand ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
4132 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4134 SDOperand TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4135 GA->getValueType(0),
4137 SDOperand Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);
4139 if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
4140 Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset, NULL, 0);
4142 // The address of the thread local variable is the add of the thread
4143 // pointer with the offset of the variable.
4144 return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
4148 X86TargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
4149 // TODO: implement the "local dynamic" model
4150 // TODO: implement the "initial exec"model for pic executables
4151 assert(!Subtarget->is64Bit() && Subtarget->isTargetELF() &&
4152 "TLS not implemented for non-ELF and 64-bit targets");
4153 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4154 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
4155 // otherwise use the "Local Exec"TLS Model
4156 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4157 return LowerToTLSGeneralDynamicModel(GA, DAG, getPointerTy());
4159 return LowerToTLSExecModel(GA, DAG, getPointerTy());
4163 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
4164 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4165 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4166 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4167 // With PIC, the address is actually $g + Offset.
4168 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4169 !Subtarget->isPICStyleRIPRel()) {
4170 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4171 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4178 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
4179 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4180 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4181 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
4182 // With PIC, the address is actually $g + Offset.
4183 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4184 !Subtarget->isPICStyleRIPRel()) {
4185 Result = DAG.getNode(ISD::ADD, getPointerTy(),
4186 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
4193 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4194 /// take a 2 x i32 value to shift plus a shift amount.
4195 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
4196 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
4197 "Not an i64 shift!");
4198 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4199 SDOperand ShOpLo = Op.getOperand(0);
4200 SDOperand ShOpHi = Op.getOperand(1);
4201 SDOperand ShAmt = Op.getOperand(2);
4202 SDOperand Tmp1 = isSRA ?
4203 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
4204 DAG.getConstant(0, MVT::i32);
4206 SDOperand Tmp2, Tmp3;
4207 if (Op.getOpcode() == ISD::SHL_PARTS) {
4208 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
4209 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
4211 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
4212 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
4215 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4216 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
4217 DAG.getConstant(32, MVT::i8));
4218 SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::i32,
4219 AndNode, DAG.getConstant(0, MVT::i8));
4222 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4223 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
4224 SmallVector<SDOperand, 4> Ops;
4225 if (Op.getOpcode() == ISD::SHL_PARTS) {
4226 Ops.push_back(Tmp2);
4227 Ops.push_back(Tmp3);
4229 Ops.push_back(Cond);
4230 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4233 Ops.push_back(Tmp3);
4234 Ops.push_back(Tmp1);
4236 Ops.push_back(Cond);
4237 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4239 Ops.push_back(Tmp2);
4240 Ops.push_back(Tmp3);
4242 Ops.push_back(Cond);
4243 Lo = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4246 Ops.push_back(Tmp3);
4247 Ops.push_back(Tmp1);
4249 Ops.push_back(Cond);
4250 Hi = DAG.getNode(X86ISD::CMOV, MVT::i32, &Ops[0], Ops.size());
4253 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
4257 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
4260 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
4261 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
4262 Op.getOperand(0).getValueType() >= MVT::i16 &&
4263 "Unknown SINT_TO_FP to lower!");
4266 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
4267 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
4268 MachineFunction &MF = DAG.getMachineFunction();
4269 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4270 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4271 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
4272 StackSlot, NULL, 0);
4274 // These are really Legal; caller falls through into that case.
4275 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
4277 if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
4279 if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
4280 Subtarget->is64Bit())
4285 bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
4286 (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
4288 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4290 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4291 SmallVector<SDOperand, 8> Ops;
4292 Ops.push_back(Chain);
4293 Ops.push_back(StackSlot);
4294 Ops.push_back(DAG.getValueType(SrcVT));
4295 Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
4296 Tys, &Ops[0], Ops.size());
4299 Chain = Result.getValue(1);
4300 SDOperand InFlag = Result.getValue(2);
4302 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4303 // shouldn't be necessary except that RFP cannot be live across
4304 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4305 MachineFunction &MF = DAG.getMachineFunction();
4306 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4307 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4308 Tys = DAG.getVTList(MVT::Other);
4309 SmallVector<SDOperand, 8> Ops;
4310 Ops.push_back(Chain);
4311 Ops.push_back(Result);
4312 Ops.push_back(StackSlot);
4313 Ops.push_back(DAG.getValueType(Op.getValueType()));
4314 Ops.push_back(InFlag);
4315 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
4316 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
4322 std::pair<SDOperand,SDOperand> X86TargetLowering::
4323 FP_TO_SINTHelper(SDOperand Op, SelectionDAG &DAG) {
4324 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
4325 "Unknown FP_TO_SINT to lower!");
4327 // These are really Legal.
4328 if (Op.getValueType() == MVT::i32 &&
4329 X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
4330 return std::make_pair(SDOperand(), SDOperand());
4331 if (Op.getValueType() == MVT::i32 &&
4332 X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
4333 return std::make_pair(SDOperand(), SDOperand());
4334 if (Subtarget->is64Bit() &&
4335 Op.getValueType() == MVT::i64 &&
4336 Op.getOperand(0).getValueType() != MVT::f80)
4337 return std::make_pair(SDOperand(), SDOperand());
4339 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4341 MachineFunction &MF = DAG.getMachineFunction();
4342 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
4343 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4344 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4346 switch (Op.getValueType()) {
4347 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4348 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4349 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4350 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4353 SDOperand Chain = DAG.getEntryNode();
4354 SDOperand Value = Op.getOperand(0);
4355 if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
4356 (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
4357 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4358 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
4359 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4361 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4363 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4364 Chain = Value.getValue(1);
4365 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4366 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4369 // Build the FP_TO_INT*_IN_MEM
4370 SDOperand Ops[] = { Chain, Value, StackSlot };
4371 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
4373 return std::make_pair(FIST, StackSlot);
4376 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
4377 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(Op, DAG);
4378 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4379 if (FIST.Val == 0) return SDOperand();
4382 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
4385 SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
4386 std::pair<SDOperand,SDOperand> Vals = FP_TO_SINTHelper(SDOperand(N, 0), DAG);
4387 SDOperand FIST = Vals.first, StackSlot = Vals.second;
4388 if (FIST.Val == 0) return 0;
4390 // Return an i64 load from the stack slot.
4391 SDOperand Res = DAG.getLoad(MVT::i64, FIST, StackSlot, NULL, 0);
4393 // Use a MERGE_VALUES node to drop the chain result value.
4394 return DAG.getNode(ISD::MERGE_VALUES, MVT::i64, Res).Val;
4397 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
4398 MVT::ValueType VT = Op.getValueType();
4399 MVT::ValueType EltVT = VT;
4400 if (MVT::isVector(VT))
4401 EltVT = MVT::getVectorElementType(VT);
4402 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4403 std::vector<Constant*> CV;
4404 if (EltVT == MVT::f64) {
4405 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
4409 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
4415 Constant *C = ConstantVector::get(CV);
4416 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4417 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4419 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
4422 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
4423 MVT::ValueType VT = Op.getValueType();
4424 MVT::ValueType EltVT = VT;
4425 unsigned EltNum = 1;
4426 if (MVT::isVector(VT)) {
4427 EltVT = MVT::getVectorElementType(VT);
4428 EltNum = MVT::getVectorNumElements(VT);
4430 const Type *OpNTy = MVT::getTypeForValueType(EltVT);
4431 std::vector<Constant*> CV;
4432 if (EltVT == MVT::f64) {
4433 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
4437 Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
4443 Constant *C = ConstantVector::get(CV);
4444 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4445 SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4447 if (MVT::isVector(VT)) {
4448 return DAG.getNode(ISD::BIT_CONVERT, VT,
4449 DAG.getNode(ISD::XOR, MVT::v2i64,
4450 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
4451 DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
4453 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
4457 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
4458 SDOperand Op0 = Op.getOperand(0);
4459 SDOperand Op1 = Op.getOperand(1);
4460 MVT::ValueType VT = Op.getValueType();
4461 MVT::ValueType SrcVT = Op1.getValueType();
4462 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
4464 // If second operand is smaller, extend it first.
4465 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
4466 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
4468 SrcTy = MVT::getTypeForValueType(SrcVT);
4470 // And if it is bigger, shrink it first.
4471 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4472 Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1);
4474 SrcTy = MVT::getTypeForValueType(SrcVT);
4477 // At this point the operands and the result should have the same
4478 // type, and that won't be f80 since that is not custom lowered.
4480 // First get the sign bit of second operand.
4481 std::vector<Constant*> CV;
4482 if (SrcVT == MVT::f64) {
4483 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
4484 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4486 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
4487 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4488 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4489 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4491 Constant *C = ConstantVector::get(CV);
4492 SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4493 SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
4495 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
4497 // Shift sign bit right or left if the two operands have different types.
4498 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
4499 // Op0 is MVT::f32, Op1 is MVT::f64.
4500 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
4501 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
4502 DAG.getConstant(32, MVT::i32));
4503 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
4504 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
4505 DAG.getConstant(0, getPointerTy()));
4508 // Clear first operand sign bit.
4510 if (VT == MVT::f64) {
4511 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
4512 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
4514 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
4515 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4516 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4517 CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
4519 C = ConstantVector::get(CV);
4520 CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
4521 SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
4523 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
4525 // Or the value with the sign bit.
4526 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
4529 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
4530 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
4532 SDOperand Op0 = Op.getOperand(0);
4533 SDOperand Op1 = Op.getOperand(1);
4534 SDOperand CC = Op.getOperand(2);
4535 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4536 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
4539 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
4541 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4542 return DAG.getNode(X86ISD::SETCC, MVT::i8,
4543 DAG.getConstant(X86CC, MVT::i8), Cond);
4546 assert(isFP && "Illegal integer SetCC!");
4548 Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
4549 switch (SetCCOpcode) {
4550 default: assert(false && "Illegal floating point SetCC!");
4551 case ISD::SETOEQ: { // !PF & ZF
4552 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4553 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
4554 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4555 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
4556 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
4558 case ISD::SETUNE: { // PF | !ZF
4559 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4560 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
4561 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
4562 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
4563 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
4569 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
4570 bool addTest = true;
4571 SDOperand Cond = Op.getOperand(0);
4574 if (Cond.getOpcode() == ISD::SETCC)
4575 Cond = LowerSETCC(Cond, DAG);
4577 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4578 // setting operand in place of the X86ISD::SETCC.
4579 if (Cond.getOpcode() == X86ISD::SETCC) {
4580 CC = Cond.getOperand(0);
4582 SDOperand Cmp = Cond.getOperand(1);
4583 unsigned Opc = Cmp.getOpcode();
4584 MVT::ValueType VT = Op.getValueType();
4585 bool IllegalFPCMov = false;
4586 if (VT == MVT::f32 && !X86ScalarSSEf32)
4587 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4588 else if (VT == MVT::f64 && !X86ScalarSSEf64)
4589 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4590 else if (VT == MVT::f80)
4591 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
4592 if ((Opc == X86ISD::CMP ||
4593 Opc == X86ISD::COMI ||
4594 Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
4601 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4602 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4605 const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
4607 SmallVector<SDOperand, 4> Ops;
4608 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
4609 // condition is true.
4610 Ops.push_back(Op.getOperand(2));
4611 Ops.push_back(Op.getOperand(1));
4613 Ops.push_back(Cond);
4614 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
4617 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
4618 bool addTest = true;
4619 SDOperand Chain = Op.getOperand(0);
4620 SDOperand Cond = Op.getOperand(1);
4621 SDOperand Dest = Op.getOperand(2);
4624 if (Cond.getOpcode() == ISD::SETCC)
4625 Cond = LowerSETCC(Cond, DAG);
4627 // If condition flag is set by a X86ISD::CMP, then use it as the condition
4628 // setting operand in place of the X86ISD::SETCC.
4629 if (Cond.getOpcode() == X86ISD::SETCC) {
4630 CC = Cond.getOperand(0);
4632 SDOperand Cmp = Cond.getOperand(1);
4633 unsigned Opc = Cmp.getOpcode();
4634 if (Opc == X86ISD::CMP ||
4635 Opc == X86ISD::COMI ||
4636 Opc == X86ISD::UCOMI) {
4643 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4644 Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
4646 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
4647 Chain, Op.getOperand(2), CC, Cond);
4650 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
4651 unsigned CallingConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4652 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
4654 if (Subtarget->is64Bit())
4655 if(CallingConv==CallingConv::Fast && isTailCall && PerformTailCallOpt)
4656 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4658 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
4660 switch (CallingConv) {
4662 assert(0 && "Unsupported calling convention");
4663 case CallingConv::Fast:
4664 if (isTailCall && PerformTailCallOpt)
4665 return LowerX86_TailCallTo(Op, DAG, CallingConv);
4667 return LowerCCCCallTo(Op,DAG, CallingConv);
4668 case CallingConv::C:
4669 case CallingConv::X86_StdCall:
4670 return LowerCCCCallTo(Op, DAG, CallingConv);
4671 case CallingConv::X86_FastCall:
4672 return LowerFastCCCallTo(Op, DAG, CallingConv);
4677 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
4678 // Calls to _alloca is needed to probe the stack when allocating more than 4k
4679 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
4680 // that the guard pages used by the OS virtual memory manager are allocated in
4681 // correct sequence.
4683 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
4684 SelectionDAG &DAG) {
4685 assert(Subtarget->isTargetCygMing() &&
4686 "This should be used only on Cygwin/Mingw targets");
4689 SDOperand Chain = Op.getOperand(0);
4690 SDOperand Size = Op.getOperand(1);
4691 // FIXME: Ensure alignment here
4695 MVT::ValueType IntPtr = getPointerTy();
4696 MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
4698 Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
4699 Flag = Chain.getValue(1);
4701 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4702 SDOperand Ops[] = { Chain,
4703 DAG.getTargetExternalSymbol("_alloca", IntPtr),
4704 DAG.getRegister(X86::EAX, IntPtr),
4706 Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
4707 Flag = Chain.getValue(1);
4709 Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
4711 std::vector<MVT::ValueType> Tys;
4712 Tys.push_back(SPTy);
4713 Tys.push_back(MVT::Other);
4714 SDOperand Ops1[2] = { Chain.getValue(0), Chain };
4715 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
4719 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4720 MachineFunction &MF = DAG.getMachineFunction();
4721 const Function* Fn = MF.getFunction();
4722 if (Fn->hasExternalLinkage() &&
4723 Subtarget->isTargetCygMing() &&
4724 Fn->getName() == "main")
4725 MF.getInfo<X86MachineFunctionInfo>()->setForceFramePointer(true);
4727 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4728 if (Subtarget->is64Bit())
4729 return LowerX86_64CCCArguments(Op, DAG);
4733 assert(0 && "Unsupported calling convention");
4734 case CallingConv::Fast:
4735 return LowerCCCArguments(Op,DAG, true);
4737 case CallingConv::C:
4738 return LowerCCCArguments(Op, DAG);
4739 case CallingConv::X86_StdCall:
4740 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(StdCall);
4741 return LowerCCCArguments(Op, DAG, true);
4742 case CallingConv::X86_FastCall:
4743 MF.getInfo<X86MachineFunctionInfo>()->setDecorationStyle(FastCall);
4744 return LowerFastCCArguments(Op, DAG);
4748 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4749 SDOperand InFlag(0, 0);
4750 SDOperand Chain = Op.getOperand(0);
4752 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4753 if (Align == 0) Align = 1;
4755 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4756 // If not DWORD aligned or size is more than the threshold, call memset.
4757 // The libc version is likely to be faster for these cases. It can use the
4758 // address value and run time information about the CPU.
4759 if ((Align & 3) != 0 ||
4760 (I && I->getValue() > Subtarget->getMaxInlineSizeThreshold())) {
4761 MVT::ValueType IntPtr = getPointerTy();
4762 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4763 TargetLowering::ArgListTy Args;
4764 TargetLowering::ArgListEntry Entry;
4765 Entry.Node = Op.getOperand(1);
4766 Entry.Ty = IntPtrTy;
4767 Args.push_back(Entry);
4768 // Extend the unsigned i8 argument to be an int value for the call.
4769 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4770 Entry.Ty = IntPtrTy;
4771 Args.push_back(Entry);
4772 Entry.Node = Op.getOperand(3);
4773 Args.push_back(Entry);
4774 std::pair<SDOperand,SDOperand> CallResult =
4775 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4776 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4777 return CallResult.second;
4782 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4783 unsigned BytesLeft = 0;
4784 bool TwoRepStos = false;
4787 uint64_t Val = ValC->getValue() & 255;
4789 // If the value is a constant, then we can potentially use larger sets.
4790 switch (Align & 3) {
4791 case 2: // WORD aligned
4794 Val = (Val << 8) | Val;
4796 case 0: // DWORD aligned
4799 Val = (Val << 8) | Val;
4800 Val = (Val << 16) | Val;
4801 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4804 Val = (Val << 32) | Val;
4807 default: // Byte aligned
4810 Count = Op.getOperand(3);
4814 if (AVT > MVT::i8) {
4816 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4817 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4818 BytesLeft = I->getValue() % UBytes;
4820 assert(AVT >= MVT::i32 &&
4821 "Do not use rep;stos if not at least DWORD aligned");
4822 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4823 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4828 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4830 InFlag = Chain.getValue(1);
4833 Count = Op.getOperand(3);
4834 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4835 InFlag = Chain.getValue(1);
4838 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4840 InFlag = Chain.getValue(1);
4841 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4842 Op.getOperand(1), InFlag);
4843 InFlag = Chain.getValue(1);
4845 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4846 SmallVector<SDOperand, 8> Ops;
4847 Ops.push_back(Chain);
4848 Ops.push_back(DAG.getValueType(AVT));
4849 Ops.push_back(InFlag);
4850 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4853 InFlag = Chain.getValue(1);
4854 Count = Op.getOperand(3);
4855 MVT::ValueType CVT = Count.getValueType();
4856 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4857 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4858 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4860 InFlag = Chain.getValue(1);
4861 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4863 Ops.push_back(Chain);
4864 Ops.push_back(DAG.getValueType(MVT::i8));
4865 Ops.push_back(InFlag);
4866 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4867 } else if (BytesLeft) {
4868 // Issue stores for the last 1 - 7 bytes.
4870 unsigned Val = ValC->getValue() & 255;
4871 unsigned Offset = I->getValue() - BytesLeft;
4872 SDOperand DstAddr = Op.getOperand(1);
4873 MVT::ValueType AddrVT = DstAddr.getValueType();
4874 if (BytesLeft >= 4) {
4875 Val = (Val << 8) | Val;
4876 Val = (Val << 16) | Val;
4877 Value = DAG.getConstant(Val, MVT::i32);
4878 Chain = DAG.getStore(Chain, Value,
4879 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4880 DAG.getConstant(Offset, AddrVT)),
4885 if (BytesLeft >= 2) {
4886 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4887 Chain = DAG.getStore(Chain, Value,
4888 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4889 DAG.getConstant(Offset, AddrVT)),
4894 if (BytesLeft == 1) {
4895 Value = DAG.getConstant(Val, MVT::i8);
4896 Chain = DAG.getStore(Chain, Value,
4897 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4898 DAG.getConstant(Offset, AddrVT)),
4906 SDOperand X86TargetLowering::LowerMEMCPYInline(SDOperand Chain,
4911 SelectionDAG &DAG) {
4913 unsigned BytesLeft = 0;
4914 switch (Align & 3) {
4915 case 2: // WORD aligned
4918 case 0: // DWORD aligned
4920 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4923 default: // Byte aligned
4928 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4929 SDOperand Count = DAG.getConstant(Size / UBytes, getPointerTy());
4930 BytesLeft = Size % UBytes;
4932 SDOperand InFlag(0, 0);
4933 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4935 InFlag = Chain.getValue(1);
4936 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4938 InFlag = Chain.getValue(1);
4939 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4941 InFlag = Chain.getValue(1);
4943 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4944 SmallVector<SDOperand, 8> Ops;
4945 Ops.push_back(Chain);
4946 Ops.push_back(DAG.getValueType(AVT));
4947 Ops.push_back(InFlag);
4948 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4951 // Issue loads and stores for the last 1 - 7 bytes.
4952 unsigned Offset = Size - BytesLeft;
4953 SDOperand DstAddr = Dest;
4954 MVT::ValueType DstVT = DstAddr.getValueType();
4955 SDOperand SrcAddr = Source;
4956 MVT::ValueType SrcVT = SrcAddr.getValueType();
4958 if (BytesLeft >= 4) {
4959 Value = DAG.getLoad(MVT::i32, Chain,
4960 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4961 DAG.getConstant(Offset, SrcVT)),
4963 Chain = Value.getValue(1);
4964 Chain = DAG.getStore(Chain, Value,
4965 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4966 DAG.getConstant(Offset, DstVT)),
4971 if (BytesLeft >= 2) {
4972 Value = DAG.getLoad(MVT::i16, Chain,
4973 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4974 DAG.getConstant(Offset, SrcVT)),
4976 Chain = Value.getValue(1);
4977 Chain = DAG.getStore(Chain, Value,
4978 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4979 DAG.getConstant(Offset, DstVT)),
4985 if (BytesLeft == 1) {
4986 Value = DAG.getLoad(MVT::i8, Chain,
4987 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4988 DAG.getConstant(Offset, SrcVT)),
4990 Chain = Value.getValue(1);
4991 Chain = DAG.getStore(Chain, Value,
4992 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4993 DAG.getConstant(Offset, DstVT)),
5001 /// Expand the result of: i64,outchain = READCYCLECOUNTER inchain
5002 SDNode *X86TargetLowering::ExpandREADCYCLECOUNTER(SDNode *N, SelectionDAG &DAG){
5003 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5004 SDOperand TheChain = N->getOperand(0);
5005 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheChain, 1);
5006 if (Subtarget->is64Bit()) {
5007 SDOperand rax = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
5008 SDOperand rdx = DAG.getCopyFromReg(rax.getValue(1), X86::RDX,
5009 MVT::i64, rax.getValue(2));
5010 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, rdx,
5011 DAG.getConstant(32, MVT::i8));
5013 DAG.getNode(ISD::OR, MVT::i64, rax, Tmp), rdx.getValue(1)
5016 Tys = DAG.getVTList(MVT::i64, MVT::Other);
5017 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
5020 SDOperand eax = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
5021 SDOperand edx = DAG.getCopyFromReg(eax.getValue(1), X86::EDX,
5022 MVT::i32, eax.getValue(2));
5023 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
5024 SDOperand Ops[] = { eax, edx };
5025 Ops[0] = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Ops, 2);
5027 // Use a MERGE_VALUES to return the value and chain.
5028 Ops[1] = edx.getValue(1);
5029 Tys = DAG.getVTList(MVT::i64, MVT::Other);
5030 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2).Val;
5033 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
5034 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
5036 if (!Subtarget->is64Bit()) {
5037 // vastart just stores the address of the VarArgsFrameIndex slot into the
5038 // memory location argument.
5039 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5040 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
5045 // gp_offset (0 - 6 * 8)
5046 // fp_offset (48 - 48 + 8 * 16)
5047 // overflow_arg_area (point to parameters coming in memory).
5049 SmallVector<SDOperand, 8> MemOps;
5050 SDOperand FIN = Op.getOperand(1);
5052 SDOperand Store = DAG.getStore(Op.getOperand(0),
5053 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5054 FIN, SV->getValue(), SV->getOffset());
5055 MemOps.push_back(Store);
5058 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
5059 DAG.getConstant(4, getPointerTy()));
5060 Store = DAG.getStore(Op.getOperand(0),
5061 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5062 FIN, SV->getValue(), SV->getOffset());
5063 MemOps.push_back(Store);
5065 // Store ptr to overflow_arg_area
5066 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
5067 DAG.getConstant(4, getPointerTy()));
5068 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5069 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
5071 MemOps.push_back(Store);
5073 // Store ptr to reg_save_area.
5074 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
5075 DAG.getConstant(8, getPointerTy()));
5076 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5077 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
5079 MemOps.push_back(Store);
5080 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
5083 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
5084 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5085 SDOperand Chain = Op.getOperand(0);
5086 SDOperand DstPtr = Op.getOperand(1);
5087 SDOperand SrcPtr = Op.getOperand(2);
5088 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
5089 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
5091 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
5092 SrcSV->getValue(), SrcSV->getOffset());
5093 Chain = SrcPtr.getValue(1);
5094 for (unsigned i = 0; i < 3; ++i) {
5095 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
5096 SrcSV->getValue(), SrcSV->getOffset());
5097 Chain = Val.getValue(1);
5098 Chain = DAG.getStore(Chain, Val, DstPtr,
5099 DstSV->getValue(), DstSV->getOffset());
5102 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
5103 DAG.getConstant(8, getPointerTy()));
5104 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
5105 DAG.getConstant(8, getPointerTy()));
5111 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
5112 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
5114 default: return SDOperand(); // Don't custom lower most intrinsics.
5115 // Comparison intrinsics.
5116 case Intrinsic::x86_sse_comieq_ss:
5117 case Intrinsic::x86_sse_comilt_ss:
5118 case Intrinsic::x86_sse_comile_ss:
5119 case Intrinsic::x86_sse_comigt_ss:
5120 case Intrinsic::x86_sse_comige_ss:
5121 case Intrinsic::x86_sse_comineq_ss:
5122 case Intrinsic::x86_sse_ucomieq_ss:
5123 case Intrinsic::x86_sse_ucomilt_ss:
5124 case Intrinsic::x86_sse_ucomile_ss:
5125 case Intrinsic::x86_sse_ucomigt_ss:
5126 case Intrinsic::x86_sse_ucomige_ss:
5127 case Intrinsic::x86_sse_ucomineq_ss:
5128 case Intrinsic::x86_sse2_comieq_sd:
5129 case Intrinsic::x86_sse2_comilt_sd:
5130 case Intrinsic::x86_sse2_comile_sd:
5131 case Intrinsic::x86_sse2_comigt_sd:
5132 case Intrinsic::x86_sse2_comige_sd:
5133 case Intrinsic::x86_sse2_comineq_sd:
5134 case Intrinsic::x86_sse2_ucomieq_sd:
5135 case Intrinsic::x86_sse2_ucomilt_sd:
5136 case Intrinsic::x86_sse2_ucomile_sd:
5137 case Intrinsic::x86_sse2_ucomigt_sd:
5138 case Intrinsic::x86_sse2_ucomige_sd:
5139 case Intrinsic::x86_sse2_ucomineq_sd: {
5141 ISD::CondCode CC = ISD::SETCC_INVALID;
5144 case Intrinsic::x86_sse_comieq_ss:
5145 case Intrinsic::x86_sse2_comieq_sd:
5149 case Intrinsic::x86_sse_comilt_ss:
5150 case Intrinsic::x86_sse2_comilt_sd:
5154 case Intrinsic::x86_sse_comile_ss:
5155 case Intrinsic::x86_sse2_comile_sd:
5159 case Intrinsic::x86_sse_comigt_ss:
5160 case Intrinsic::x86_sse2_comigt_sd:
5164 case Intrinsic::x86_sse_comige_ss:
5165 case Intrinsic::x86_sse2_comige_sd:
5169 case Intrinsic::x86_sse_comineq_ss:
5170 case Intrinsic::x86_sse2_comineq_sd:
5174 case Intrinsic::x86_sse_ucomieq_ss:
5175 case Intrinsic::x86_sse2_ucomieq_sd:
5176 Opc = X86ISD::UCOMI;
5179 case Intrinsic::x86_sse_ucomilt_ss:
5180 case Intrinsic::x86_sse2_ucomilt_sd:
5181 Opc = X86ISD::UCOMI;
5184 case Intrinsic::x86_sse_ucomile_ss:
5185 case Intrinsic::x86_sse2_ucomile_sd:
5186 Opc = X86ISD::UCOMI;
5189 case Intrinsic::x86_sse_ucomigt_ss:
5190 case Intrinsic::x86_sse2_ucomigt_sd:
5191 Opc = X86ISD::UCOMI;
5194 case Intrinsic::x86_sse_ucomige_ss:
5195 case Intrinsic::x86_sse2_ucomige_sd:
5196 Opc = X86ISD::UCOMI;
5199 case Intrinsic::x86_sse_ucomineq_ss:
5200 case Intrinsic::x86_sse2_ucomineq_sd:
5201 Opc = X86ISD::UCOMI;
5207 SDOperand LHS = Op.getOperand(1);
5208 SDOperand RHS = Op.getOperand(2);
5209 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
5211 SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
5212 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
5213 DAG.getConstant(X86CC, MVT::i8), Cond);
5214 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
5219 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
5220 // Depths > 0 not supported yet!
5221 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5224 // Just load the return address
5225 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5226 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
5229 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
5230 // Depths > 0 not supported yet!
5231 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
5234 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
5235 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
5236 DAG.getConstant(4, getPointerTy()));
5239 SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
5240 SelectionDAG &DAG) {
5241 // Is not yet supported on x86-64
5242 if (Subtarget->is64Bit())
5245 return DAG.getConstant(8, getPointerTy());
5248 SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
5250 assert(!Subtarget->is64Bit() &&
5251 "Lowering of eh_return builtin is not supported yet on x86-64");
5253 MachineFunction &MF = DAG.getMachineFunction();
5254 SDOperand Chain = Op.getOperand(0);
5255 SDOperand Offset = Op.getOperand(1);
5256 SDOperand Handler = Op.getOperand(2);
5258 SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
5261 SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
5262 DAG.getConstant(-4UL, getPointerTy()));
5263 StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
5264 Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
5265 Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
5266 MF.addLiveOut(X86::ECX);
5268 return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
5269 Chain, DAG.getRegister(X86::ECX, getPointerTy()));
5272 SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
5273 SelectionDAG &DAG) {
5274 SDOperand Root = Op.getOperand(0);
5275 SDOperand Trmp = Op.getOperand(1); // trampoline
5276 SDOperand FPtr = Op.getOperand(2); // nested function
5277 SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
5279 SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
5281 if (Subtarget->is64Bit()) {
5282 return SDOperand(); // not yet supported
5284 Function *Func = (Function *)
5285 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
5286 unsigned CC = Func->getCallingConv();
5291 assert(0 && "Unsupported calling convention");
5292 case CallingConv::C:
5293 case CallingConv::X86_StdCall: {
5294 // Pass 'nest' parameter in ECX.
5295 // Must be kept in sync with X86CallingConv.td
5298 // Check that ECX wasn't needed by an 'inreg' parameter.
5299 const FunctionType *FTy = Func->getFunctionType();
5300 const ParamAttrsList *Attrs = Func->getParamAttrs();
5302 if (Attrs && !Func->isVarArg()) {
5303 unsigned InRegCount = 0;
5306 for (FunctionType::param_iterator I = FTy->param_begin(),
5307 E = FTy->param_end(); I != E; ++I, ++Idx)
5308 if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
5309 // FIXME: should only count parameters that are lowered to integers.
5310 InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
5312 if (InRegCount > 2) {
5313 cerr << "Nest register in use - reduce number of inreg parameters!\n";
5319 case CallingConv::X86_FastCall:
5320 // Pass 'nest' parameter in EAX.
5321 // Must be kept in sync with X86CallingConv.td
5326 const X86InstrInfo *TII =
5327 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
5329 SDOperand OutChains[4];
5330 SDOperand Addr, Disp;
5332 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
5333 Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
5335 unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
5336 unsigned char N86Reg = ((X86RegisterInfo*)RegInfo)->getX86RegNum(NestReg);
5337 OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
5338 Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
5340 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
5341 OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
5342 TrmpSV->getOffset() + 1, false, 1);
5344 unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
5345 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
5346 OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
5347 TrmpSV->getValue() + 5, TrmpSV->getOffset());
5349 Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
5350 OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
5351 TrmpSV->getOffset() + 6, false, 1);
5354 { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
5355 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
5359 SDOperand X86TargetLowering::LowerFLT_ROUNDS(SDOperand Op, SelectionDAG &DAG) {
5361 The rounding mode is in bits 11:10 of FPSR, and has the following
5368 FLT_ROUNDS, on the other hand, expects the following:
5375 To perform the conversion, we do:
5376 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
5379 MachineFunction &MF = DAG.getMachineFunction();
5380 const TargetMachine &TM = MF.getTarget();
5381 const TargetFrameInfo &TFI = *TM.getFrameInfo();
5382 unsigned StackAlignment = TFI.getStackAlignment();
5383 MVT::ValueType VT = Op.getValueType();
5385 // Save FP Control Word to stack slot
5386 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
5387 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5389 SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
5390 DAG.getEntryNode(), StackSlot);
5392 // Load FP Control Word from stack slot
5393 SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
5395 // Transform as necessary
5397 DAG.getNode(ISD::SRL, MVT::i16,
5398 DAG.getNode(ISD::AND, MVT::i16,
5399 CWD, DAG.getConstant(0x800, MVT::i16)),
5400 DAG.getConstant(11, MVT::i8));
5402 DAG.getNode(ISD::SRL, MVT::i16,
5403 DAG.getNode(ISD::AND, MVT::i16,
5404 CWD, DAG.getConstant(0x400, MVT::i16)),
5405 DAG.getConstant(9, MVT::i8));
5408 DAG.getNode(ISD::AND, MVT::i16,
5409 DAG.getNode(ISD::ADD, MVT::i16,
5410 DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
5411 DAG.getConstant(1, MVT::i16)),
5412 DAG.getConstant(3, MVT::i16));
5415 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
5416 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
5419 SDOperand X86TargetLowering::LowerCTLZ(SDOperand Op, SelectionDAG &DAG) {
5420 MVT::ValueType VT = Op.getValueType();
5421 MVT::ValueType OpVT = VT;
5422 unsigned NumBits = MVT::getSizeInBits(VT);
5424 Op = Op.getOperand(0);
5425 if (VT == MVT::i8) {
5426 // Zero extend to i32 since there is not an i8 bsr.
5428 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5431 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
5432 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5433 Op = DAG.getNode(X86ISD::BSR, VTs, Op);
5435 // If src is zero (i.e. bsr sets ZF), returns NumBits.
5436 SmallVector<SDOperand, 4> Ops;
5438 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
5439 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5440 Ops.push_back(Op.getValue(1));
5441 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5443 // Finally xor with NumBits-1.
5444 Op = DAG.getNode(ISD::XOR, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
5447 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5451 SDOperand X86TargetLowering::LowerCTTZ(SDOperand Op, SelectionDAG &DAG) {
5452 MVT::ValueType VT = Op.getValueType();
5453 MVT::ValueType OpVT = VT;
5454 unsigned NumBits = MVT::getSizeInBits(VT);
5456 Op = Op.getOperand(0);
5457 if (VT == MVT::i8) {
5459 Op = DAG.getNode(ISD::ZERO_EXTEND, OpVT, Op);
5462 // Issue a bsf (scan bits forward) which also sets EFLAGS.
5463 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
5464 Op = DAG.getNode(X86ISD::BSF, VTs, Op);
5466 // If src is zero (i.e. bsf sets ZF), returns NumBits.
5467 SmallVector<SDOperand, 4> Ops;
5469 Ops.push_back(DAG.getConstant(NumBits, OpVT));
5470 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
5471 Ops.push_back(Op.getValue(1));
5472 Op = DAG.getNode(X86ISD::CMOV, OpVT, &Ops[0], 4);
5475 Op = DAG.getNode(ISD::TRUNCATE, MVT::i8, Op);
5479 /// LowerOperation - Provide custom lowering hooks for some operations.
5481 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
5482 switch (Op.getOpcode()) {
5483 default: assert(0 && "Should not custom lower this!");
5484 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5485 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5486 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5487 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5488 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5489 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5490 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5491 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5492 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
5493 case ISD::SHL_PARTS:
5494 case ISD::SRA_PARTS:
5495 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
5496 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5497 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
5498 case ISD::FABS: return LowerFABS(Op, DAG);
5499 case ISD::FNEG: return LowerFNEG(Op, DAG);
5500 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5501 case ISD::SETCC: return LowerSETCC(Op, DAG);
5502 case ISD::SELECT: return LowerSELECT(Op, DAG);
5503 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
5504 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5505 case ISD::CALL: return LowerCALL(Op, DAG);
5506 case ISD::RET: return LowerRET(Op, DAG);
5507 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
5508 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
5509 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
5510 case ISD::VASTART: return LowerVASTART(Op, DAG);
5511 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
5512 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5513 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5514 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5515 case ISD::FRAME_TO_ARGS_OFFSET:
5516 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
5517 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
5518 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
5519 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
5520 case ISD::FLT_ROUNDS: return LowerFLT_ROUNDS(Op, DAG);
5521 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
5522 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
5524 // FIXME: REMOVE THIS WHEN LegalizeDAGTypes lands.
5525 case ISD::READCYCLECOUNTER:
5526 return SDOperand(ExpandREADCYCLECOUNTER(Op.Val, DAG), 0);
5530 /// ExpandOperation - Provide custom lowering hooks for expanding operations.
5531 SDNode *X86TargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
5532 switch (N->getOpcode()) {
5533 default: assert(0 && "Should not custom lower this!");
5534 case ISD::FP_TO_SINT: return ExpandFP_TO_SINT(N, DAG);
5535 case ISD::READCYCLECOUNTER: return ExpandREADCYCLECOUNTER(N, DAG);
5539 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
5541 default: return NULL;
5542 case X86ISD::BSF: return "X86ISD::BSF";
5543 case X86ISD::BSR: return "X86ISD::BSR";
5544 case X86ISD::SHLD: return "X86ISD::SHLD";
5545 case X86ISD::SHRD: return "X86ISD::SHRD";
5546 case X86ISD::FAND: return "X86ISD::FAND";
5547 case X86ISD::FOR: return "X86ISD::FOR";
5548 case X86ISD::FXOR: return "X86ISD::FXOR";
5549 case X86ISD::FSRL: return "X86ISD::FSRL";
5550 case X86ISD::FILD: return "X86ISD::FILD";
5551 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
5552 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
5553 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
5554 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
5555 case X86ISD::FLD: return "X86ISD::FLD";
5556 case X86ISD::FST: return "X86ISD::FST";
5557 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
5558 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
5559 case X86ISD::CALL: return "X86ISD::CALL";
5560 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
5561 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
5562 case X86ISD::CMP: return "X86ISD::CMP";
5563 case X86ISD::COMI: return "X86ISD::COMI";
5564 case X86ISD::UCOMI: return "X86ISD::UCOMI";
5565 case X86ISD::SETCC: return "X86ISD::SETCC";
5566 case X86ISD::CMOV: return "X86ISD::CMOV";
5567 case X86ISD::BRCOND: return "X86ISD::BRCOND";
5568 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
5569 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
5570 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
5571 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
5572 case X86ISD::Wrapper: return "X86ISD::Wrapper";
5573 case X86ISD::S2VEC: return "X86ISD::S2VEC";
5574 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
5575 case X86ISD::PINSRW: return "X86ISD::PINSRW";
5576 case X86ISD::FMAX: return "X86ISD::FMAX";
5577 case X86ISD::FMIN: return "X86ISD::FMIN";
5578 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
5579 case X86ISD::FRCP: return "X86ISD::FRCP";
5580 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
5581 case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
5582 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
5583 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
5584 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
5588 // isLegalAddressingMode - Return true if the addressing mode represented
5589 // by AM is legal for this target, for a load/store of the specified type.
5590 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5591 const Type *Ty) const {
5592 // X86 supports extremely general addressing modes.
5594 // X86 allows a sign-extended 32-bit immediate field as a displacement.
5595 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
5599 // We can only fold this if we don't need an extra load.
5600 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
5603 // X86-64 only supports addr of globals in small code model.
5604 if (Subtarget->is64Bit()) {
5605 if (getTargetMachine().getCodeModel() != CodeModel::Small)
5607 // If lower 4G is not available, then we must use rip-relative addressing.
5608 if (AM.BaseOffs || AM.Scale > 1)
5619 // These scales always work.
5624 // These scales are formed with basereg+scalereg. Only accept if there is
5629 default: // Other stuff never works.
5637 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
5638 if (!Ty1->isInteger() || !Ty2->isInteger())
5640 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5641 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5642 if (NumBits1 <= NumBits2)
5644 return Subtarget->is64Bit() || NumBits1 < 64;
5647 bool X86TargetLowering::isTruncateFree(MVT::ValueType VT1,
5648 MVT::ValueType VT2) const {
5649 if (!MVT::isInteger(VT1) || !MVT::isInteger(VT2))
5651 unsigned NumBits1 = MVT::getSizeInBits(VT1);
5652 unsigned NumBits2 = MVT::getSizeInBits(VT2);
5653 if (NumBits1 <= NumBits2)
5655 return Subtarget->is64Bit() || NumBits1 < 64;
5658 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5659 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5660 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5661 /// are assumed to be legal.
5663 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
5664 // Only do shuffles on 128-bit vector types for now.
5665 if (MVT::getSizeInBits(VT) == 64) return false;
5666 return (Mask.Val->getNumOperands() <= 4 ||
5667 isIdentityMask(Mask.Val) ||
5668 isIdentityMask(Mask.Val, true) ||
5669 isSplatMask(Mask.Val) ||
5670 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
5671 X86::isUNPCKLMask(Mask.Val) ||
5672 X86::isUNPCKHMask(Mask.Val) ||
5673 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
5674 X86::isUNPCKH_v_undef_Mask(Mask.Val));
5677 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
5679 SelectionDAG &DAG) const {
5680 unsigned NumElts = BVOps.size();
5681 // Only do shuffles on 128-bit vector types for now.
5682 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
5683 if (NumElts == 2) return true;
5685 return (isMOVLMask(&BVOps[0], 4) ||
5686 isCommutedMOVL(&BVOps[0], 4, true) ||
5687 isSHUFPMask(&BVOps[0], 4) ||
5688 isCommutedSHUFP(&BVOps[0], 4));
5693 //===----------------------------------------------------------------------===//
5694 // X86 Scheduler Hooks
5695 //===----------------------------------------------------------------------===//
5698 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
5699 MachineBasicBlock *BB) {
5700 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5701 switch (MI->getOpcode()) {
5702 default: assert(false && "Unexpected instr type to insert");
5703 case X86::CMOV_FR32:
5704 case X86::CMOV_FR64:
5705 case X86::CMOV_V4F32:
5706 case X86::CMOV_V2F64:
5707 case X86::CMOV_V2I64: {
5708 // To "insert" a SELECT_CC instruction, we actually have to insert the
5709 // diamond control-flow pattern. The incoming instruction knows the
5710 // destination vreg to set, the condition code register to branch on, the
5711 // true/false values to select between, and a branch opcode to use.
5712 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5713 ilist<MachineBasicBlock>::iterator It = BB;
5719 // cmpTY ccX, r1, r2
5721 // fallthrough --> copy0MBB
5722 MachineBasicBlock *thisMBB = BB;
5723 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
5724 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
5726 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
5727 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
5728 MachineFunction *F = BB->getParent();
5729 F->getBasicBlockList().insert(It, copy0MBB);
5730 F->getBasicBlockList().insert(It, sinkMBB);
5731 // Update machine-CFG edges by first adding all successors of the current
5732 // block to the new block which will contain the Phi node for the select.
5733 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
5734 e = BB->succ_end(); i != e; ++i)
5735 sinkMBB->addSuccessor(*i);
5736 // Next, remove all successors of the current block, and add the true
5737 // and fallthrough blocks as its successors.
5738 while(!BB->succ_empty())
5739 BB->removeSuccessor(BB->succ_begin());
5740 BB->addSuccessor(copy0MBB);
5741 BB->addSuccessor(sinkMBB);
5744 // %FalseValue = ...
5745 // # fallthrough to sinkMBB
5748 // Update machine-CFG edges
5749 BB->addSuccessor(sinkMBB);
5752 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5755 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
5756 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5757 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5759 delete MI; // The pseudo instruction is gone now.
5763 case X86::FP32_TO_INT16_IN_MEM:
5764 case X86::FP32_TO_INT32_IN_MEM:
5765 case X86::FP32_TO_INT64_IN_MEM:
5766 case X86::FP64_TO_INT16_IN_MEM:
5767 case X86::FP64_TO_INT32_IN_MEM:
5768 case X86::FP64_TO_INT64_IN_MEM:
5769 case X86::FP80_TO_INT16_IN_MEM:
5770 case X86::FP80_TO_INT32_IN_MEM:
5771 case X86::FP80_TO_INT64_IN_MEM: {
5772 // Change the floating point control register to use "round towards zero"
5773 // mode when truncating to an integer value.
5774 MachineFunction *F = BB->getParent();
5775 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
5776 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
5778 // Load the old value of the high byte of the control word...
5780 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
5781 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
5783 // Set the high part to be round to zero...
5784 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
5787 // Reload the modified control word now...
5788 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5790 // Restore the memory image of control word to original value
5791 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
5794 // Get the X86 opcode to use.
5796 switch (MI->getOpcode()) {
5797 default: assert(0 && "illegal opcode!");
5798 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
5799 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
5800 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
5801 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
5802 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
5803 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
5804 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
5805 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
5806 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
5810 MachineOperand &Op = MI->getOperand(0);
5811 if (Op.isRegister()) {
5812 AM.BaseType = X86AddressMode::RegBase;
5813 AM.Base.Reg = Op.getReg();
5815 AM.BaseType = X86AddressMode::FrameIndexBase;
5816 AM.Base.FrameIndex = Op.getFrameIndex();
5818 Op = MI->getOperand(1);
5819 if (Op.isImmediate())
5820 AM.Scale = Op.getImm();
5821 Op = MI->getOperand(2);
5822 if (Op.isImmediate())
5823 AM.IndexReg = Op.getImm();
5824 Op = MI->getOperand(3);
5825 if (Op.isGlobalAddress()) {
5826 AM.GV = Op.getGlobal();
5828 AM.Disp = Op.getImm();
5830 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
5831 .addReg(MI->getOperand(4).getReg());
5833 // Reload the original control word now.
5834 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
5836 delete MI; // The pseudo instruction is gone now.
5842 //===----------------------------------------------------------------------===//
5843 // X86 Optimization Hooks
5844 //===----------------------------------------------------------------------===//
5846 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
5848 uint64_t &KnownZero,
5850 const SelectionDAG &DAG,
5851 unsigned Depth) const {
5852 unsigned Opc = Op.getOpcode();
5853 assert((Opc >= ISD::BUILTIN_OP_END ||
5854 Opc == ISD::INTRINSIC_WO_CHAIN ||
5855 Opc == ISD::INTRINSIC_W_CHAIN ||
5856 Opc == ISD::INTRINSIC_VOID) &&
5857 "Should use MaskedValueIsZero if you don't know whether Op"
5858 " is a target node!");
5860 KnownZero = KnownOne = 0; // Don't know anything.
5864 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
5869 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5870 /// element of the result of the vector shuffle.
5871 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
5872 MVT::ValueType VT = N->getValueType(0);
5873 SDOperand PermMask = N->getOperand(2);
5874 unsigned NumElems = PermMask.getNumOperands();
5875 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
5877 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5879 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5880 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
5881 SDOperand Idx = PermMask.getOperand(i);
5882 if (Idx.getOpcode() == ISD::UNDEF)
5883 return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
5884 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
5889 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
5890 /// node is a GlobalAddress + an offset.
5891 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
5892 unsigned Opc = N->getOpcode();
5893 if (Opc == X86ISD::Wrapper) {
5894 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
5895 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
5898 } else if (Opc == ISD::ADD) {
5899 SDOperand N1 = N->getOperand(0);
5900 SDOperand N2 = N->getOperand(1);
5901 if (isGAPlusOffset(N1.Val, GA, Offset)) {
5902 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
5904 Offset += V->getSignExtended();
5907 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
5908 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
5910 Offset += V->getSignExtended();
5918 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
5920 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
5921 MachineFrameInfo *MFI) {
5922 if (N->getOperand(0).Val != Base->getOperand(0).Val)
5925 SDOperand Loc = N->getOperand(1);
5926 SDOperand BaseLoc = Base->getOperand(1);
5927 if (Loc.getOpcode() == ISD::FrameIndex) {
5928 if (BaseLoc.getOpcode() != ISD::FrameIndex)
5930 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
5931 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
5932 int FS = MFI->getObjectSize(FI);
5933 int BFS = MFI->getObjectSize(BFI);
5934 if (FS != BFS || FS != Size) return false;
5935 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
5937 GlobalValue *GV1 = NULL;
5938 GlobalValue *GV2 = NULL;
5939 int64_t Offset1 = 0;
5940 int64_t Offset2 = 0;
5941 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
5942 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
5943 if (isGA1 && isGA2 && GV1 == GV2)
5944 return Offset1 == (Offset2 + Dist*Size);
5950 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
5951 const X86Subtarget *Subtarget) {
5954 if (isGAPlusOffset(Base, GV, Offset))
5955 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5957 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5958 int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
5960 // Fixed objects do not specify alignment, however the offsets are known.
5961 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5962 (MFI->getObjectOffset(BFI) % 16) == 0);
5964 return MFI->getObjectAlignment(BFI) >= 16;
5970 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5971 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5972 /// if the load addresses are consecutive, non-overlapping, and in the right
5974 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5975 const X86Subtarget *Subtarget) {
5976 MachineFunction &MF = DAG.getMachineFunction();
5977 MachineFrameInfo *MFI = MF.getFrameInfo();
5978 MVT::ValueType VT = N->getValueType(0);
5979 MVT::ValueType EVT = MVT::getVectorElementType(VT);
5980 SDOperand PermMask = N->getOperand(2);
5981 int NumElems = (int)PermMask.getNumOperands();
5982 SDNode *Base = NULL;
5983 for (int i = 0; i < NumElems; ++i) {
5984 SDOperand Idx = PermMask.getOperand(i);
5985 if (Idx.getOpcode() == ISD::UNDEF) {
5986 if (!Base) return SDOperand();
5989 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5990 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5994 else if (!isConsecutiveLoad(Arg.Val, Base,
5995 i, MVT::getSizeInBits(EVT)/8,MFI))
6000 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
6001 LoadSDNode *LD = cast<LoadSDNode>(Base);
6003 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6004 LD->getSrcValueOffset(), LD->isVolatile());
6006 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
6007 LD->getSrcValueOffset(), LD->isVolatile(),
6008 LD->getAlignment());
6012 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
6013 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
6014 const X86Subtarget *Subtarget) {
6015 SDOperand Cond = N->getOperand(0);
6017 // If we have SSE[12] support, try to form min/max nodes.
6018 if (Subtarget->hasSSE2() &&
6019 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
6020 if (Cond.getOpcode() == ISD::SETCC) {
6021 // Get the LHS/RHS of the select.
6022 SDOperand LHS = N->getOperand(1);
6023 SDOperand RHS = N->getOperand(2);
6024 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
6026 unsigned Opcode = 0;
6027 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
6030 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
6033 if (!UnsafeFPMath) break;
6035 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
6037 Opcode = X86ISD::FMIN;
6040 case ISD::SETOGT: // (X > Y) ? X : Y -> max
6043 if (!UnsafeFPMath) break;
6045 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
6047 Opcode = X86ISD::FMAX;
6050 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
6053 case ISD::SETOGT: // (X > Y) ? Y : X -> min
6056 if (!UnsafeFPMath) break;
6058 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
6060 Opcode = X86ISD::FMIN;
6063 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
6066 if (!UnsafeFPMath) break;
6068 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
6070 Opcode = X86ISD::FMAX;
6076 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
6085 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
6086 DAGCombinerInfo &DCI) const {
6087 SelectionDAG &DAG = DCI.DAG;
6088 switch (N->getOpcode()) {
6090 case ISD::VECTOR_SHUFFLE:
6091 return PerformShuffleCombine(N, DAG, Subtarget);
6093 return PerformSELECTCombine(N, DAG, Subtarget);
6099 //===----------------------------------------------------------------------===//
6100 // X86 Inline Assembly Support
6101 //===----------------------------------------------------------------------===//
6103 /// getConstraintType - Given a constraint letter, return the type of
6104 /// constraint it is for this target.
6105 X86TargetLowering::ConstraintType
6106 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
6107 if (Constraint.size() == 1) {
6108 switch (Constraint[0]) {
6117 return C_RegisterClass;
6122 return TargetLowering::getConstraintType(Constraint);
6125 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6126 /// vector. If it is invalid, don't add anything to Ops.
6127 void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
6129 std::vector<SDOperand>&Ops,
6130 SelectionDAG &DAG) {
6131 SDOperand Result(0, 0);
6133 switch (Constraint) {
6136 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6137 if (C->getValue() <= 31) {
6138 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6144 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
6145 if (C->getValue() <= 255) {
6146 Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
6152 // Literal immediates are always ok.
6153 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
6154 Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
6158 // If we are in non-pic codegen mode, we allow the address of a global (with
6159 // an optional displacement) to be used with 'i'.
6160 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
6163 // Match either (GA) or (GA+C)
6165 Offset = GA->getOffset();
6166 } else if (Op.getOpcode() == ISD::ADD) {
6167 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6168 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6170 Offset = GA->getOffset()+C->getValue();
6172 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6173 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
6175 Offset = GA->getOffset()+C->getValue();
6182 // If addressing this global requires a load (e.g. in PIC mode), we can't
6184 if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
6188 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
6194 // Otherwise, not valid for this mode.
6200 Ops.push_back(Result);
6203 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6206 std::vector<unsigned> X86TargetLowering::
6207 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6208 MVT::ValueType VT) const {
6209 if (Constraint.size() == 1) {
6210 // FIXME: not handling fp-stack yet!
6211 switch (Constraint[0]) { // GCC X86 Constraint Letters
6212 default: break; // Unknown constraint letter
6213 case 'A': // EAX/EDX
6214 if (VT == MVT::i32 || VT == MVT::i64)
6215 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
6217 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
6220 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
6221 else if (VT == MVT::i16)
6222 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
6223 else if (VT == MVT::i8)
6224 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
6225 else if (VT == MVT::i64)
6226 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
6231 return std::vector<unsigned>();
6234 std::pair<unsigned, const TargetRegisterClass*>
6235 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6236 MVT::ValueType VT) const {
6237 // First, see if this is a constraint that directly corresponds to an LLVM
6239 if (Constraint.size() == 1) {
6240 // GCC Constraint Letters
6241 switch (Constraint[0]) {
6243 case 'r': // GENERAL_REGS
6244 case 'R': // LEGACY_REGS
6245 case 'l': // INDEX_REGS
6246 if (VT == MVT::i64 && Subtarget->is64Bit())
6247 return std::make_pair(0U, X86::GR64RegisterClass);
6249 return std::make_pair(0U, X86::GR32RegisterClass);
6250 else if (VT == MVT::i16)
6251 return std::make_pair(0U, X86::GR16RegisterClass);
6252 else if (VT == MVT::i8)
6253 return std::make_pair(0U, X86::GR8RegisterClass);
6255 case 'y': // MMX_REGS if MMX allowed.
6256 if (!Subtarget->hasMMX()) break;
6257 return std::make_pair(0U, X86::VR64RegisterClass);
6259 case 'Y': // SSE_REGS if SSE2 allowed
6260 if (!Subtarget->hasSSE2()) break;
6262 case 'x': // SSE_REGS if SSE1 allowed
6263 if (!Subtarget->hasSSE1()) break;
6267 // Scalar SSE types.
6270 return std::make_pair(0U, X86::FR32RegisterClass);
6273 return std::make_pair(0U, X86::FR64RegisterClass);
6281 return std::make_pair(0U, X86::VR128RegisterClass);
6287 // Use the default implementation in TargetLowering to convert the register
6288 // constraint into a member of a register class.
6289 std::pair<unsigned, const TargetRegisterClass*> Res;
6290 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6292 // Not found as a standard register?
6293 if (Res.second == 0) {
6294 // GCC calls "st(0)" just plain "st".
6295 if (StringsEqualNoCase("{st}", Constraint)) {
6296 Res.first = X86::ST0;
6297 Res.second = X86::RFP80RegisterClass;
6303 // Otherwise, check to see if this is a register class of the wrong value
6304 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
6305 // turn into {ax},{dx}.
6306 if (Res.second->hasType(VT))
6307 return Res; // Correct type already, nothing to do.
6309 // All of the single-register GCC register classes map their values onto
6310 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
6311 // really want an 8-bit or 32-bit register, map to the appropriate register
6312 // class and return the appropriate register.
6313 if (Res.second != X86::GR16RegisterClass)
6316 if (VT == MVT::i8) {
6317 unsigned DestReg = 0;
6318 switch (Res.first) {
6320 case X86::AX: DestReg = X86::AL; break;
6321 case X86::DX: DestReg = X86::DL; break;
6322 case X86::CX: DestReg = X86::CL; break;
6323 case X86::BX: DestReg = X86::BL; break;
6326 Res.first = DestReg;
6327 Res.second = Res.second = X86::GR8RegisterClass;
6329 } else if (VT == MVT::i32) {
6330 unsigned DestReg = 0;
6331 switch (Res.first) {
6333 case X86::AX: DestReg = X86::EAX; break;
6334 case X86::DX: DestReg = X86::EDX; break;
6335 case X86::CX: DestReg = X86::ECX; break;
6336 case X86::BX: DestReg = X86::EBX; break;
6337 case X86::SI: DestReg = X86::ESI; break;
6338 case X86::DI: DestReg = X86::EDI; break;
6339 case X86::BP: DestReg = X86::EBP; break;
6340 case X86::SP: DestReg = X86::ESP; break;
6343 Res.first = DestReg;
6344 Res.second = Res.second = X86::GR32RegisterClass;
6346 } else if (VT == MVT::i64) {
6347 unsigned DestReg = 0;
6348 switch (Res.first) {
6350 case X86::AX: DestReg = X86::RAX; break;
6351 case X86::DX: DestReg = X86::RDX; break;
6352 case X86::CX: DestReg = X86::RCX; break;
6353 case X86::BX: DestReg = X86::RBX; break;
6354 case X86::SI: DestReg = X86::RSI; break;
6355 case X86::DI: DestReg = X86::RDI; break;
6356 case X86::BP: DestReg = X86::RBP; break;
6357 case X86::SP: DestReg = X86::RSP; break;
6360 Res.first = DestReg;
6361 Res.second = Res.second = X86::GR64RegisterClass;